Steam turbine vibration fault diagnosis system fused with deep learning

Abstract

The invention relates to the field of power equipment data processing, in particular to a steam turbine vibration fault diagnosis system fused with deep learning. Comprising a dynamic knowledge base construction module, a working condition adaptive data synchronization module, a knowledge-guided heterogeneous feature fusion module, a dynamic structure neural network module, an online self-optimization weight distribution module and a knowledge-enhanced coupling fault reasoning module. The dynamic knowledge base construction module updates the multi-modal knowledge graph through an incremental knowledge distillation mechanism, and generates an interpolation strategy template and a frequency band sensitivity matrix; the working condition self-adaptive data synchronization module dynamically calls an interpolation algorithm based on a rotating speed fluctuation mode to realize time sequence alignment optimization of multi-source sensor data; the system effectively solves the problems of time scale asynchronism and feature heterogeneity in multi-source heterogeneous data fusion through a knowledge-driven and data-driven closed-loop interaction mechanism, and realizes accurate diagnosis and early warning of steam turbine vibration faults under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the field of power equipment data processing, and particularly to a steam turbine vibration fault diagnosis system integrating deep learning. Background Art

[0002] In the vibration fault diagnosis of power plant steam turbines, vibration signal analysis is the core method. By collecting shafting vibration acceleration, displacement, and phase data, and combining time-domain waveform and frequency-domain spectrum characteristics, fault modes can be identified. When the vibration amplitude exceeds the standard threshold, attribution analysis needs to be carried out based on the vibration energy distribution and harmonic components: if the fundamental frequency component in the spectrum is significant and the phase is stable, it usually indicates rotor mass imbalance; the prominence of the second harmonic may be caused by misalignment of the coupling; high-order harmonics or fractional frequency components are related to bearing wear or oil film oscillation. By establishing a rotor dynamics model, the fault mechanism can be further verified, such as the influence of critical speed deviation or support stiffness change on vibration characteristics. During the diagnosis process, historical operation data and real-time state parameters need to be integrated, and pattern recognition algorithms are used to distinguish coupled faults, and finally guide dynamic balance correction or shafting alignment adjustment to restore the vibration index of the unit to the safe range.

[0003] In the vibration fault diagnosis of power plant steam turbines, during the data processing process, it is necessary to collect sensor signals such as vibration acceleration, displacement, and phase with a millisecond-level sampling rate, which has a time scale mismatch with the minute-level operating parameters (such as steam pressure and speed); at the same time, it is difficult to uniformly represent the heterogeneous data of time-domain waveforms, frequency-domain spectrum characteristics, and equipment historical maintenance records as an associated feature space that can be parsed by a deep neural network, resulting in the failure of implicit association of fault modes, weakening the cross-domain feature extraction ability of coupled faults, and restricting the generalization of the diagnosis model to multi-fault concurrent scenarios under complex working conditions. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a steam turbine vibration fault diagnosis system integrating deep learning, which solves the problem of difficult extraction of cross-domain association features caused by time-scale asynchrony and feature space heterogeneity in multi-source heterogeneous data fusion.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: The steam turbine vibration fault diagnosis system integrating deep learning provided by the present invention includes: A dynamic knowledge base construction module, which is used to integrate historical fault cases, rotor dynamics equations, and expert rules, construct a multi-modal knowledge graph including the relationship between fault modes - features - maintenance strategies, and update the edge weights of the graph through an incremental knowledge distillation mechanism. The knowledge graph outputs an interpolation strategy template to the data synchronization module; The working condition adaptive data synchronization module receives real-time sensor data, calls the interpolation strategy template generated by the dynamic knowledge base construction module for time series alignment, transmits the synchronized data with timestamp metadata to the feature fusion module, and feeds back the timestamp error distribution to the dynamic knowledge base construction module to optimize the interpolation strategy template; The knowledge-guided heterogeneous feature fusion module receives the synchronized data output by the working condition adaptive data synchronization module, inputs the time-frequency analysis features of the vibration signal into a convolutional neural network, adjusts the channel attention weights based on the frequency band sensitivity matrix provided by the dynamic knowledge base construction module, generates fusion features under physical constraints, and transmits them to the dynamic structure neural network module; The dynamic structure neural network module activates the corresponding network branch and loads the pre-trained model parameters according to the matching result of the fusion features output by the knowledge-guided heterogeneous feature fusion module and the fault mode in the dynamic knowledge base construction module, extracts fault features, and transmits them to the online self-optimizing weight allocation module; The online self-optimizing weight allocation module receives the fault features extracted by the dynamic structure neural network module, combines the real-time working condition parameters with the feature contribution degree reference value in the dynamic knowledge base construction module, dynamically allocates feature importance weights through a gated recurrent unit, and feeds back the weight deviation information to the dynamic knowledge base construction module to trigger model parameter tuning; The knowledge-enhanced coupled fault reasoning module receives the feature weights allocated by the online self-optimizing weight allocation module and the fault propagation logic of the dynamic knowledge base construction module, performs confidence fusion on the data-driven classification result and physical rules, generates a diagnostic conclusion, and triggers the incremental update of the knowledge graph in the dynamic knowledge base construction module, forming a closed-loop feedback mechanism.

[0006] Furthermore, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the incremental knowledge distillation mechanism in the dynamic knowledge base construction module specifically includes: The online diagnosis result generated by the knowledge-enhanced coupled fault reasoning module triggers the update of the edge weights of the knowledge graph, automatically extracts new fault feature vectors and association rules to generate adversarial samples, and injects the adversarial samples into the training data stream of the dynamic structure neural network module; The updated edge weights of the knowledge graph are transmitted to the online self-optimizing weight allocation module to constrain the regularization term calculation process of the gated recurrent unit, so that the feature weight allocation is synchronized with the updated physical rules.

[0007] Furthermore, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the working condition adaptive data synchronization module performs time series alignment: Call the interpolation coefficient template from the dynamic knowledge base construction module according to the real-time rotational speed fluctuation mode. When a sudden change in steam pressure is detected, link the dynamic knowledge base construction module to retrieve the sensor response mode under similar working conditions; The corrected timestamp matching tolerance threshold is fed back to the dynamic knowledge base construction module by the working condition adaptive data synchronization module, triggering the gradient update of the interpolation strategy template, and forming the adaptive collaborative optimization of data synchronization accuracy and knowledge graph parameters.

[0008] Furthermore, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, in the knowledge-guided heterogeneous feature fusion module: The feature map output by the convolutional neural network carries physical rule constraint information, and generates a feature importance ranking after being verified by the fault classifier; The feature importance ranking is fed back to the dynamic knowledge base construction module, triggering the gradient update of the frequency band sensitivity matrix, and making the frequency band weight distribution dynamically adapt to the current equipment fault mode.

[0009] Furthermore, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the dynamic structure neural network module activates the network branch: The proportion of the fundamental frequency component extracted by the frequency spectrum analysis subnet is input to the dynamic knowledge base construction module for fault mode retrieval, generating a branch switching instruction and loading the pre-trained model parameters; The output feature map of the activated network branch is sent to the online self-optimizing weight allocation module, synchronously triggering the adversarial sample verification process in the dynamic knowledge base construction module, and suppressing the abnormal feature distribution caused by the deviation of model parameters.

[0010] Furthermore, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the online self-optimizing weight allocation module performs dynamic weight allocation: Call the benchmark value of the feature contribution degree of the same type of equipment stored in the dynamic knowledge base construction module, and calculate the deviation amount between the real-time weight and the benchmark value; When the deviation amount exceeds the preset threshold, trigger the online fine-tuning instruction of the dynamic structure neural network module, and update the adversarial sample generation strategy in the dynamic knowledge base construction module, and the adversarial sample generation strategy is used for the training data stream optimization of the dynamic structure neural network module.

[0011] Furthermore, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the knowledge-enhanced coupled fault reasoning module performs confidence fusion: Jointly reason about the fault probability output by the data-driven classifier and the logical rules of the fault propagation directed graph in the dynamic knowledge base construction module; The inference result triggers the numerical simulation verification process. For the cases that pass the verification, incremental training data is generated and sent back to the dynamic knowledge base construction module to update the edge weight parameters of the fault coupling relationship in the fault propagation directed graph.

[0012] Further, for the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the system exception handling process includes: when a branch heartbeat of the dynamic structure neural network module is lost, the control center sends a downgrade instruction to the working condition adaptive data synchronization module to switch to the basic interpolation mode; The rule engine and the fault safety weight template in the dynamic knowledge base construction module are synchronously activated to maintain the diagnostic function and record the exception event in the dynamic knowledge base construction module, generating a training data set for the network branch health prediction model.

[0013] Further, for the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the cooperation between the dynamic knowledge base construction module and the feature fusion module includes: The dynamic knowledge base construction module outputs a frequency band sensitivity matrix to the knowledge-guided heterogeneous feature fusion module to constrain the channel attention calculation process of the convolutional neural network; The feature importance ranking fed back by the knowledge-guided heterogeneous feature fusion module triggers the update of the historical fault case spectrum fingerprint library in the dynamic knowledge base construction module, forming a two-way optimization link between physical rule constraints and data-driven features.

[0014] Further, for the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the interaction between the dynamic structure neural network module and the weight allocation module includes: The branch switching instruction of the dynamic structure neural network module triggers the online self-optimizing weight allocation module to call the feature contribution degree reference value to constrain the regularization term calculation of the gated recurrent unit; The feature weight deviation information output by the online self-optimizing weight allocation module triggers the online fine-tuning instruction of the dynamic structure neural network module and starts the parameter version rollback mechanism of the dynamic knowledge base construction module to achieve the dynamic balance of model stability and diagnostic accuracy.

[0015] Advantages of the present invention; The beneficial effects of the present invention are as follows: through the deep fusion architecture of knowledge-driven and data-driven, it effectively overcomes the problems of time-scale asynchrony and feature-space heterogeneity in multi-source heterogeneous data fusion. The interpolation strategy template and band sensitivity matrix provided by the dynamic knowledge base construction module guide the working condition adaptive data synchronization module to achieve time-series alignment optimization based on physical mechanisms, and at the same time constrain the cross-domain feature extraction process of the feature fusion module, significantly improving the spatio-temporal correlation accuracy of multi-modal data such as vibration signals, temperature gradients, and maintenance records; the closed-loop feedback mechanism of the dynamic structure neural network module and the online self-optimizing weight allocation module realizes the dynamic balance of the model generalization ability and working condition adaptability through parameter fine-tuning triggered by the feature contribution degree reference value constraint and weight deviation; the knowledge-enhanced coupled fault inference module fuses data-driven probabilities and fault propagation logic rules, generates physically interpretable diagnostic conclusions after numerical simulation verification, and continuously optimizes the edge weights of the knowledge graph through incremental knowledge distillation, forming a collaborative iterative improvement of diagnostic accuracy and knowledge completeness, and finally realizing the accurate positioning and early warning of steam turbine vibration faults under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0017] Figure 1 FIG. is the system architecture diagram of the steam turbine vibration fault diagnosis system integrating deep learning provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below with reference to the drawings. To better understand the objectives of the present invention, the present invention is further described in detail below.

[0019] Please refer to Figure 1 , the present invention provides a steam turbine vibration fault diagnosis system integrating deep learning, including: The dynamic knowledge base construction module is used to integrate historical fault cases, rotor dynamics equations, and expert rules, construct a multi-modal knowledge graph containing the relationship between fault modes - features - maintenance strategies, and update the edge weights of the graph through an incremental knowledge distillation mechanism. The knowledge graph outputs an interpolation strategy template to the data synchronization module; The working condition adaptive data synchronization module receives real-time sensor data and calls the interpolation strategy template generated by the dynamic knowledge base construction module for time series alignment. After attaching timestamp metadata to the synchronized data, it transmits the data to the feature fusion module and feeds back the timestamp error distribution to the dynamic knowledge base construction module to optimize the interpolation strategy template; The knowledge-guided heterogeneous feature fusion module receives the synchronized data output by the working condition adaptive data synchronization module. It inputs the time-frequency analysis features of the vibration signal into a convolutional neural network, adjusts the channel attention weights based on the frequency band sensitivity matrix provided by the dynamic knowledge base construction module, generates fusion features under physical constraints, and transmits them to the dynamic structure neural network module; The dynamic structure neural network module, according to the matching result of the fusion features output by the knowledge-guided heterogeneous feature fusion module and the fault modes in the dynamic knowledge base construction module, activates the corresponding network branches and loads the pre-trained model parameters, extracts fault features, and then transmits them to the online self-optimizing weight allocation module; The online self-optimizing weight allocation module receives the fault features extracted by the dynamic structure neural network module, combines the real-time working condition parameters with the feature contribution degree benchmark value in the dynamic knowledge base construction module, dynamically allocates feature importance weights through a gated recurrent unit, and feeds back the weight deviation information to the dynamic knowledge base construction module to trigger model parameter tuning; The knowledge-enhanced coupled fault reasoning module receives the feature weights allocated by the online self-optimizing weight allocation module and the fault propagation logic of the dynamic knowledge base construction module, fuses the confidence levels of the data-driven classification results and physical rules, generates a diagnostic conclusion, and triggers an incremental update of the knowledge graph in the dynamic knowledge base construction module, forming a closed-loop feedback mechanism.

[0020] The steam turbine vibration fault diagnosis system integrating deep learning provided by the present invention realizes the closed-loop interaction between data and knowledge through a modular architecture. The dynamic knowledge base construction module first constructs a multi-modal knowledge graph with fault modes as nodes and feature associations as edges by analyzing the vibration spectrum features, maintenance record texts, and rotor dynamics simulation data in historical fault cases. This module adopts an incremental knowledge distillation mechanism, matches the newly identified fault feature vectors with the graph nodes during the online diagnosis process, updates the edge weights through a graph convolutional network to reflect the fault evolution law, and at the same time pushes the updated interpolation strategy template to the working condition adaptive data synchronization module to provide domain knowledge guidance for data preprocessing.

[0021] After receiving the vibration acceleration, rotational speed, and temperature signals from multi-source sensors, the working condition adaptive data synchronization module selects a time series alignment algorithm based on the interpolation strategy template provided by the dynamic knowledge base construction module. When it detects that the rotational speed fluctuation enters the critical range, it automatically calls the cubic spline interpolation template to compensate for the sampling interval difference; if there is a step change in the steam pressure, it switches to the adaptive interpolation mode based on the device response characteristics. The synchronized data with time stamp accuracy index after alignment is used as metadata and transmitted to the downstream module. At the same time, the interpolation error distribution is fed back to the knowledge base to trigger the optimization of template parameters, forming an iterative improvement mechanism for data preprocessing accuracy and knowledge base strategy.

[0022] The knowledge-guided heterogeneous feature fusion module performs time-frequency transformation on the synchronized data to generate a mixed feature set containing Gram angular field images and spectral energy entropy. The dynamic knowledge base construction module pushes the frequency band sensitivity matrix according to the current device model. This matrix dynamically adjusts the weight allocation of each channel of the convolutional neural network through the attention mechanism, strengthening the ability to extract frequency band features with high relevance to historical faults. The feature map constrained by physical rules is output to the dynamic structure neural network module, and at the same time, the feature importance ranking is fed back to the knowledge base to drive the online optimization of the frequency band sensitivity matrix.

[0023] Based on the output vector of the feature fusion module, the dynamic structure neural network module calculates the energy proportion of the fundamental frequency and harmonic components in the spectrum analysis subnet in real time. It matches the analysis result with the spectrum fingerprint of the fault mode in the knowledge base. When the confidence level of a specific fault type continuously exceeds the threshold, it activates the corresponding pre-trained network branch. When loading the branch parameters, the adversarial samples stored in the knowledge base are injected synchronously to verify the model robustness and then perform feature extraction. The activated fault type identification code is transmitted to the online self-optimizing weight allocation module to constrain the weight calculation logic.

[0024] After receiving the fault feature vector, the online self-optimizing weight allocation module calculates the importance weights of each feature dimension through a gated recurrent unit in combination with the real-time working condition parameters and the feature contribution degree benchmark value of similar devices in the knowledge base. When it detects a significant deviation between the real-time weight distribution and the benchmark value, it triggers the online fine-tuning instruction of the dynamic structure neural network branch. At the same time, it feeds back the deviation information to the knowledge base to drive the update of the adversarial sample generation strategy, forming a collaborative optimization of feature selection and model generalization ability.

[0025] The knowledge-enhanced coupled fault inference module synthesizes the probability distribution output by the data-driven classifier and the logical rules of the fault propagation directed graph in the knowledge base, and uses the evidence theory for confidence fusion. For the conflicting inference results, the rotor dynamics equation is called for numerical simulation verification, and the diagnostic conclusion is corrected through physical mechanism constraints. The cases that pass the verification generate incremental training samples. After extracting the feature association rules through knowledge distillation, the edge weight parameters of the graph are updated, completing the transformation of the diagnostic results into domain knowledge and realizing the closed-loop evolution of the system's diagnostic ability and knowledge completeness.

[0026] The data flow and feedback mechanism between modules constitute a multi-level optimization loop. The timestamp error feedback optimizes the interpolation strategy, the feature importance ranking drives the update of the frequency band sensitive matrix, the weight deviation triggers the model parameter tuning, and the diagnostic verification results generate knowledge increment, forming the full-link adaptive ability from data preprocessing, feature engineering to decision-making inference. This architecture design based on the deep integration of knowledge guidance and data-driven effectively solves the technical bottlenecks of difficult embedding of physical rules and insufficient working condition adaptability in traditional methods.

[0027] Specifically, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the incremental knowledge distillation mechanism in the dynamic knowledge base construction module specifically includes: The online diagnostic results generated by the knowledge-enhanced coupled fault inference module trigger the update of the edge weights of the knowledge graph, automatically extract new fault feature vectors and association rules to generate adversarial samples, and inject the adversarial samples into the training data stream of the dynamic structure neural network module; The updated edge weights of the knowledge graph are transmitted to the online self-optimizing weight allocation module to constrain the regularization term calculation process of the gated recurrent unit, so that the feature weight allocation is synchronized with the updated physical rules.

[0028] In the present invention, the incremental knowledge distillation mechanism of the dynamic knowledge base construction module realizes continuous optimization through the synergistic effect of knowledge evolution and model training. The diagnostic results output by the knowledge-enhanced coupled fault inference module trigger the knowledge graph update process. First, the fault features in the diagnostic conclusion are vectorized and encoded, and the similarity is matched with the existing nodes of the knowledge graph. When a fault mode that is not fully matched is detected, new feature vectors and association rules of adjacent nodes are extracted through the graph convolutional network, and an edge weight update instruction including the fault coupling relationship is generated. The updated knowledge graph quantifies the logical connection strength between the newly added fault mode nodes and the existing nodes as edge weight parameters, forming a dynamic representation of the fault propagation path.

[0029] After feature engineering, the new fault feature vectors and association rules are processed to generate adversarial samples with physical significance. This process retrieves the critical state data before the occurrence of similar faults in the knowledge base, extracts the time-frequency domain disturbance pattern of its vibration waveform, and generates adversarial input samples after Gram angular field transformation. The generated adversarial samples are injected into the training data stream of the dynamic structure neural network module to enhance the sensitivity to early fault features during the forward propagation of the model. During the adversarial training process, the feature activation pattern output by the model is compared and verified with the fault propagation path in the knowledge graph, and effective disturbance samples are screened to update the adversarial sample library, forming a two-way verification mechanism for training data and domain knowledge.

[0030] The updated knowledge graph edge weight parameters are mapped to regularization constraints of the gated recurrent unit through the feature contribution conversion algorithm. Specifically, the edge weight distribution of the fault mode node in the knowledge graph is quantified into the baseline weight of the feature dimension, and the deviation of the real-time weight distribution from the physical rules is limited by the regularization term in the loss function. When calculating the feature importance weight, the gated recurrent unit simultaneously integrates the real-time operating condition parameters and the regularization constraints. When it is detected that the deviation of the weight distribution from the baseline value exceeds the threshold, the rule verification process of the dynamic knowledge base construction module is triggered, the abnormal weight distribution strategy is corrected, and a new adversarial sample injection training iteration is generated to form a knowledge-driven feature selection optimization closed loop.

[0031] Specifically, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the operating condition adaptive data synchronization module performs timing alignment: The interpolation coefficient template is called from the dynamic knowledge base construction module according to the real-time speed fluctuation pattern. When a sudden change in steam pressure is detected, the dynamic knowledge base construction module is linked to retrieve the sensor response pattern under similar working conditions. The corrected timestamp matching tolerance threshold is fed back to the dynamic knowledge base construction module through the working condition adaptive data synchronization module, triggering the gradient update of the interpolation strategy template, forming an adaptive collaborative optimization of data synchronization accuracy and knowledge graph parameters.

[0032] The timing alignment operation of the working condition adaptive data synchronization module in the present invention realizes high-precision synchronization of sensor data through a dynamic strategy driven by a knowledge base. The dynamic knowledge base construction module pre-stores interpolation coefficient templates corresponding to different speed intervals, including cubic spline interpolation parameters for the critical speed interval and a set of linear interpolation coefficients for the steady-state operation area. When the real-time speed fluctuation enters the preset critical speed threshold range, the data synchronization module calls the interpolation template for the speed interval in the knowledge base, and dynamically adjusts the interpolation window width based on the speed-temperature correlation matrix to compensate for the sensor sampling interval drift caused by changes in centrifugal force. After the interpolation operation is completed, the generated timestamp metadata records the alignment error distribution of each channel data, providing a quantitative basis for subsequent knowledge base optimization.

[0033] When a sudden step change in steam pressure is detected, the data synchronization module initiates a similar operating condition retrieval request to the dynamic knowledge base construction module. The dynamic knowledge base construction module extracts the current pressure change rate, temperature gradient, and equipment response mode feature vectors in historical cases, calculates the similarity using a dynamic time warping algorithm, and returns the operating condition interpolation strategy with the highest matching degree. The data synchronization module switches to a special interpolation mode for pressure mutations based on the retrieval results, uses a piecewise polynomial interpolation algorithm to eliminate the phase distortion of sensor data caused by pressure shock, and relaxes the timestamp matching tolerance threshold to adapt to transient operating condition characteristics.

[0034] The corrected timestamp matching tolerance threshold carries the equipment operation status label and is fed back to the dynamic knowledge base construction module through the data synchronization module. The feedback data packet contains the error statistics, operating condition feature vectors and strategy application effect indicators of each interpolation interval, triggering the gradient update of the interpolation strategy template in the knowledge base. During the gradient update process, the contribution of the interpolation coefficient to the timestamp accuracy is calculated based on the error statistics, and the weight distribution of the interpolation window parameters and the speed association matrix is ​​adjusted through the back propagation algorithm. The updated interpolation strategy template is redeployed to the data synchronization module to form an adaptive evolution mechanism for interpolation parameters for specific equipment, so that the data synchronization accuracy and the operating condition mode characterization capability in the knowledge base are improved simultaneously.

[0035] Specifically, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, in the knowledge-guided heterogeneous feature fusion module: The feature map output by the convolutional neural network carries the physical rule constraint information, and after verification by the fault classifier, the feature importance ranking is generated; The feature importance ranking is fed back to the dynamic knowledge base construction module, triggering the gradient update of the frequency band sensitivity matrix, so that the frequency band weight distribution is dynamically adapted to the current equipment failure mode.

[0036] The knowledge-guided heterogeneous feature fusion module in the present invention realizes feature optimization through the dynamic interaction of physical rules and data features. After the convolutional neural network receives the vibration signal features after time-frequency transformation, the frequency band sensitivity matrix provided by the dynamic knowledge base construction module is introduced in the convolution layer calculation process. The matrix is ​​generated based on the frequency domain energy distribution of historical equipment failures, and physical constraints are imposed on the convolution kernel weights through the channel attention mechanism to strengthen the frequency band response that is highly correlated with the current fault mode. For example, for rotor unbalance faults, the matrix assigns a higher attention weight to the 1x frequency channel to suppress noise interference in non-related frequency bands. When the feature map constrained by physical rules is output to the fault classifier, it carries the feature activation mode of the fault-sensitive frequency band, providing prior knowledge guidance for subsequent classification.

[0037] After the fault classifier performs multi-label classification on the feature map, the contribution degree of each frequency band feature to the classification result is calculated based on the gradient backpropagation algorithm. The contribution degree quantization index constitutes the feature importance ranking, where the high contribution degree frequency band reflects the core feature distribution of the current fault mode. This ranking data is encapsulated as a feedback vector and transmitted to the frequency band sensitivity matrix update interface of the dynamic knowledge base construction module. The dynamic knowledge base construction module analyzes the difference in the frequency band weight distribution in the feedback vector and adjusts the sensitivity coefficients of each frequency band in the matrix through the gradient descent algorithm to make the matrix weights tend to be consistent with the real-time fault feature importance distribution.

[0038] After the frequency band sensitivity matrix is updated, the dynamic knowledge base construction module pushes the adjusted matrix parameters to the channel attention layer of the feature fusion module. The updated matrix dynamically adjusts the frequency domain focusing range of the convolution kernel in the next round of feature extraction. For example, when it is detected that the contribution degree of the second harmonic feature increases significantly, the response intensity of the convolution kernel corresponding to the corresponding frequency band is enhanced. This closed-loop feedback mechanism enables the dynamic coordination of physical rule constraints and data-driven features, while retaining expert experience, adaptively tracking the evolution trend of the device fault mode, and improving the working condition adaptability of feature representation.

[0039] Specifically, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the dynamic structure neural network module activates a network branch: The proportion of the fundamental frequency component extracted by the spectrum analysis subnet is input to the dynamic knowledge base construction module for fault mode retrieval, generating a branch switching instruction and loading the pre-trained model parameters; The activated network branch outputs a feature map to the online self-optimizing weight allocation module, synchronously triggering the adversarial sample verification process in the dynamic knowledge base construction module to suppress the abnormal feature distribution caused by the deviation of model parameters.

[0040] In the present invention, the dynamic structure neural network module realizes the adaptive extraction of fault features through a knowledge-guided branch switching mechanism. The spectrum analysis subnet performs a fast Fourier transform on the input vibration signal and calculates the proportion distribution of the fundamental frequency and harmonic components in the total energy. During the analysis process, a sliding window strategy is adopted to dynamically track the rotational speed fluctuation, and the energy features of the target frequency band are extracted through a bandpass filter bank to generate a spectrum feature vector containing the proportion of the fundamental frequency component. This feature vector carries the device operating state label and is input to the fault mode retrieval interface of the dynamic knowledge base construction module for similarity matching with the pre-stored typical fault spectrum fingerprints in the knowledge graph.

[0041] The dynamic knowledge base construction module uses the cosine similarity algorithm to calculate the matching degree between the real-time spectrum features and the historical fault patterns. When the matching degree of a specific fault type continuously exceeds the activation threshold, a network branch switching instruction is generated. The switching instruction includes the target fault type code and the storage address of the corresponding pre-trained model parameters, triggering the dynamic structure neural network module to load the dedicated network branch parameters from the knowledge base. During the loading process, the batch normalization layer parameters are dynamically adjusted according to the statistical characteristics of the current batch of data to eliminate the influence of device individual differences on the feature distribution, and at the same time, the adversarial samples stored in the knowledge base are injected to verify the model robustness.

[0042] The activated network branch performs convolution operations and pooling operations on the input feature map, extracts fault-sensitive features and outputs them to the online self-optimizing weight allocation module. Band focusing degree metadata is attached during the transmission of the feature map, reflecting the attention distribution of each convolutional layer to the key fault frequency bands. The adversarial sample verification process of the dynamic knowledge base construction module is synchronously triggered. This process retrieves the critical state data before the occurrence of the same type of fault in the knowledge base, generates perturbation samples containing early fault features and injects them into the current branch for forward inference to verify the model output stability.

[0043] After the analysis of the adversarial sample verification results by the dynamic knowledge base construction module, the parameter rollback strategy of the dedicated branch is updated. When it is detected that the perturbation sample causes abnormal fluctuations in the classification confidence, the model parameters are automatically rolled back to the previous stable version, and new adversarial samples are generated and injected into the training process to optimize the network robustness. The verified adversarial samples are stored in the adversarial sample pool of the knowledge base for subsequent model iterative training, forming a collaborative mechanism for branch stability maintenance and knowledge base defense ability improvement.

[0044] Specifically, for the steam turbine vibration fault diagnosis system integrating deep learning described in the present invention, when the online self-optimizing weight allocation module performs dynamic weight allocation: Call the benchmark value of the feature contribution degree of the same type of device stored in the dynamic knowledge base construction module, and calculate the deviation amount between the real-time weight and the benchmark value; When the deviation amount exceeds the preset threshold, trigger the online fine-tuning instruction of the dynamic structure neural network module, and update the adversarial sample generation strategy in the dynamic knowledge base construction module. The adversarial sample generation strategy is used for the training data stream optimization of the dynamic structure neural network module.

[0045] In the present invention, the online self-optimizing weight allocation module realizes feature selection optimization through a knowledge-guided weight dynamic adjustment mechanism. The dynamic knowledge base construction module stores the benchmark values of feature contribution degrees of different types of steam turbines under typical fault scenarios, and the benchmark values are statistically generated based on the contribution rates of feature dimensions to the classification results in historical diagnosis cases. The weight allocation module calls the set of benchmark values corresponding to the current equipment model, calculates the theoretical weight distribution of each feature dimension in combination with real-time operating condition parameters, and makes a one-by-one comparison with the real-time weights output by the gated recurrent unit to generate a quantization index vector containing the deviation direction and amplitude.

[0046] When it is detected that the weight deviation of a specific feature dimension exceeds the preset threshold, an online fine-tuning instruction for the dynamic structure neural network module is triggered. The fine-tuning process adopts a sliding window training strategy, intercepts time series segments from the current data stream to generate a mini-batch data set, and locally optimizes the parameters of the last layer of the network based on the deviation direction to calculate the gradient correction amount. At the same time, a deviation type identification code and a fine-tuning effect index are sent to the dynamic knowledge base construction module to trigger the iterative update of the adversarial sample generation strategy.

[0047] The dynamic knowledge base construction module retrieves the perturbation mode features of historical adversarial samples under the same type of fault mode according to the received deviation type identification code. Combining the current equipment operating state parameters, it generates a perturbation waveform template containing rotational speed fluctuation and temperature gradient features, and constructs a targeted adversarial sample after time-frequency transformation. The updated adversarial sample is injected into the training data stream of the dynamic structure neural network module to enhance the sensitivity of the model to the deviation feature dimension during the forward propagation process and suppress the weight allocation deviation caused by operating condition drift.

[0048] The fine-tuned network parameters and the updated adversarial samples act together to form an optimization closed-loop for feature weight stability. The gated recurrent unit fuses the corrected network output features and the robustness constraint conditions obtained from adversarial training in the next round of weight allocation, so that the real-time weight distribution gradually approaches the knowledge base benchmark value. When the deviation amount is continuously lower than the threshold, a knowledge base benchmark value calibration process is triggered, and the set of feature contribution degree benchmark values is updated based on the optimized weight distribution to complete the dynamic alignment of knowledge experience and data-driven results.

[0049] Specifically, for the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, when the knowledge-enhanced coupled fault reasoning module performs confidence fusion: The fault probability output by the data-driven classifier and the logical rules of the fault propagation directed graph in the dynamic knowledge base construction module are jointly inferred; The inference result triggers a numerical simulation verification process. The cases that pass the verification generate incremental training data and are transmitted back to the dynamic knowledge base construction module to update the edge weight parameters of the fault coupling relationship in the fault propagation directed graph.

[0050] In the present invention, the knowledge-enhanced coupled fault inference module improves the diagnostic reliability through a collaborative verification mechanism of physical rules and data-driven methods. After the fault probability distribution output by the data-driven classifier is normalized, it is fused with the logical rules of the fault propagation directed graph in the dynamic knowledge base construction module using the Dempster-Shafer theory. During the fusion process, the edge weights of the fault propagation directed graph represent the causal association strength between fault modes. The data-driven probability and the rule constraint confidence are combined through a weighted summation algorithm to generate a joint inference result. For inference conclusions with logical conflicts, such as when the data-driven classifier determines a single fault while the rule base indicates a coupling risk, a multi-dimensional verification process is triggered.

[0051] The verification process calls the rotor dynamics equation in the dynamic knowledge base construction module to build a numerical simulation model based on the current device structure parameters and operating status. During the simulation process, the excitation signals of candidate fault modes are injected, and the correlation coefficient between the shafting vibration response and the measured data is calculated to verify the physical rationality of fault coupling. For cases that pass the verification, the fault feature vectors, operating condition parameters, and simulation verification results are extracted, packaged into an incremental training data packet, and transmitted to the dynamic knowledge base construction module with a time stamp and the device identification code.

[0052] After receiving the incremental data, the dynamic knowledge base construction module analyzes the feature association rules of the fault coupling relationship and updates the edge weight parameters of the fault propagation directed graph through a graph convolutional network. During the update process, based on the co-occurrence frequency of fault modes and the credibility of simulation verification results in the incremental data, the gradient update step size of the edge weights is adjusted. For example, for strongly coupled fault pairs confirmed by simulation, the weight increase of their connecting edges is increased to strengthen the logical constraint effect. The updated fault propagation directed graph is redeployed to the inference module to optimize the rule constraint strength of subsequent joint inferences.

[0053] The incremental training data also triggers the knowledge distillation process of the dynamic knowledge base construction module to extract the key feature combinations and rule association patterns in the fault coupling cases. The distilled feature association rules are encoded and injected into the training loss function of the data-driven classifier, and the consistency between the model output and physical rules is constrained through a regularization term. The updated classifier reduces the misjudgment probability that conflicts with the fault propagation logic in the next round of inference, forming a two-way verification closed-loop between the data-driven results and physical mechanisms.

[0054] The above process screens physically feasible fault coupling hypotheses through simulation verification, dynamically optimizes the knowledge base logical rules based on incremental data, and feeds back the rule constraints to model training, realizing a spiral improvement in the credibility of diagnostic conclusions and the completeness of the knowledge base. The synergistic effect of rule update and model optimization effectively suppresses the misjudgment risk caused by training bias in a single data-driven method, and at the same time avoids the missed detection defect of new fault modes in pure rule-based reasoning.

[0055] Specifically, for the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the abnormal handling process of the system includes: when the heartbeat of a branch of the dynamic structure neural network module is lost, the control center sends a downgrade instruction to the working condition adaptive data synchronization module to switch to the basic interpolation mode; Synchronously activate the rule engine and the fault-safe weight template in the dynamic knowledge base construction module, maintain the diagnostic function, record the abnormal event in the dynamic knowledge base construction module, and generate a training data set for the network branch health prediction model.

[0056] In the present invention, the abnormal handling process of the system ensures the continuous reliability of the diagnostic system through multi-level redundancy design. The dynamic structure neural network module is built-in with a heartbeat monitoring unit, which periodically detects the computing resource occupancy rate and response delay of each network branch. When a specific branch fails to return a valid heartbeat signal for three consecutive monitoring cycles, it is determined as a branch heartbeat loss event, triggering the abnormal handling protocol of the control center. The control center sends a downgrade instruction code to the working condition adaptive data synchronization module. The instruction contains the abnormal branch identifier and the current operating state label of the device, driving the data synchronization module to switch to the preset basic interpolation mode.

[0057] After the basic interpolation mode is enabled, the working condition adaptive data synchronization module calls the device basic parameter set in the dynamic knowledge base construction module, including static parameters such as the rated rotor speed and bearing clearance threshold, and uses the linear interpolation algorithm to perform data alignment. During the interpolation process, the requirement for the timestamp matching accuracy is relaxed, sacrificing part of the time series resolution in exchange for the release of computing resources to maintain the continuity of the basic diagnostic function. Synchronously activate the rule engine in the dynamic knowledge base construction module, load the expert experience rule set in the historical fault cases, and perform logical verification on the real-time feature weights in combination with the fault-safe weight template to suppress the feature selection deviation caused by branch anomalies.

[0058] The activated rule engine retrieves the historical abnormal records of similar devices in the knowledge base, matches the current working condition parameters and fault feature combinations, and generates auxiliary diagnostic suggestions based on physical rules. The fault-safe weight template forcibly increases the weight ratio of time-domain statistical features and temperature correlation features, and reduces the weight of frequency-domain features that are highly dependent on abnormal branches to prevent sudden changes in the model output. The full-life cycle data of abnormal events, including the feature map snapshot at the time of heartbeat loss and the diagnostic result deviation amount in the downgraded mode, are desensitized, appended with timestamps and device identification codes, and stored in the abnormal case library of the dynamic knowledge base construction module.

[0059] The data in the abnormal case library is regularly input into the training process of the network branch health prediction model. During the training process, early warning signals such as the resource occupancy trend and feature weight distribution before the occurrence of abnormal events are extracted to construct an early identification feature set for branch faults. The trained prediction model is deployed to the dynamic structure neural network module to monitor the running status of each branch in real time, predict potential faulty branches before the heartbeat is lost, and initiate hot backup switching, forming an iterative enhancement mechanism for abnormal defense capabilities.

[0060] The above processing flow maintains the basic functions through the degradation mode, provides decision redundancy through the rule engine, and optimizes the prediction model driven by abnormal data. It ensures the minimum availability of the system in the scenarios of hardware anomalies or software failures, and at the same time accumulates abnormal data to improve the self-healing ability of the system. The closed-loop feedback between the abnormal event data and the health prediction model enables the system to have the technical characteristics of evolving from single-fault handling to predictive maintenance.

[0061] Specifically, for the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the cooperation between the dynamic knowledge base construction module and the feature fusion module includes: The dynamic knowledge base construction module outputs a frequency band sensitivity matrix to the knowledge-guided heterogeneous feature fusion module to constrain the channel attention calculation process of the convolutional neural network; The feature importance ranking fed back by the knowledge-guided heterogeneous feature fusion module triggers the update of the historical fault case spectrum fingerprint library in the dynamic knowledge base construction module, forming a two-way optimization link between physical rule constraints and data-driven features.

[0062] In the present invention, the cooperation mechanism between the dynamic knowledge base construction module and the knowledge-guided heterogeneous feature fusion module realizes the dynamic improvement of the feature representation ability through the interactive optimization of physical rules and data features. The dynamic knowledge base construction module extracts the energy ratio statistics of sensitive frequency bands in different fault modes based on the frequency domain energy distribution characteristics of device historical fault cases, and generates a frequency band sensitivity matrix through normalization processing. The row vectors of this matrix correspond to the fault type codes, the column vectors represent the sensitive coefficients of each frequency band, and the matrix element values reflect the dependence degree of specific faults on the target frequency band. After the matrix is generated, it is pushed to the channel attention layer of the knowledge-guided heterogeneous feature fusion module as a physical constraint benchmark for the channel weights of the convolutional neural network.

[0063] During the calculation of channel attention in the knowledge-guided heterogeneous feature fusion module, the frequency band sensitivity matrix is multiplied by the frequency domain energy distribution of the real-time feature map to generate an attention weight distribution guided by physical rules. The convolutional neural network strengthens the response intensity of high-sensitivity frequency bands and suppresses interference signals in noise frequency bands during the feature extraction stage. After the physically constrained feature map is input into the fault classifier, the contribution of each frequency band feature to the decision is calculated based on the backpropagation of the classification result's reverse gradient, generating a feature importance index containing the frequency band weight ranking.

[0064] The feature importance ranking index is encapsulated as a feedback data packet and transmitted to the spectrum fingerprint update interface of the dynamic knowledge base construction module. The dynamic knowledge base construction module analyzes the frequency band weight distribution in the feedback data, compares it with the spectrum fingerprint features of historical fault cases, and calculates the frequency domain evolution trend of the current device's fault mode through the sliding window statistical method. For the frequency band weights that continuously deviate from the historical distribution, the gradient descent algorithm is used to adjust the corresponding coefficients of the frequency band sensitivity matrix to keep the matrix parameters dynamically adapted to the real-time fault features.

[0065] The updated frequency band sensitivity matrix is redeployed to the channel attention layer of the feature fusion module, imposing the optimized physical rule constraints in the next round of feature extraction. During the matrix parameter adjustment process, the dynamic knowledge base construction module synchronously retrieves the frequency band sensitivity evolution records of similar devices, extracts common adjustment rules, and updates the benchmark parameters of the spectrum fingerprint library. The updated spectrum fingerprint library provides cross-device generalization experience for subsequent matrix generation, forming a collaborative mechanism for device-specific optimization and group knowledge transfer.

[0066] The above collaborative process realizes the dynamic balance between physical prior knowledge and data-driven features through the closed-loop interaction of rule injection, feature importance feedback, and matrix gradient optimization of the frequency band sensitivity matrix. The feature fusion module extracts highly discriminative features under physical constraints, and the feedback data drives the knowledge base to update the rule parameters, forming a two-way optimization link, effectively overcoming the problem of insufficient working condition adaptability caused by rule solidification in traditional methods.

[0067] Specifically, in the steam turbine vibration fault diagnosis system integrating deep learning of the present invention, the interaction between the dynamic structure neural network module and the weight allocation module includes: The branch switching instruction of the dynamic structure neural network module triggers the online self-optimizing weight allocation module to call the feature contribution degree reference value, constraining the regularization term calculation of the gated recurrent unit; The feature weight deviation information output by the online self-optimizing weight allocation module triggers the online fine-tuning instruction of the dynamic structure neural network module and starts the parameter version rollback mechanism of the dynamic knowledge base construction module, realizing the dynamic balance between model stability and diagnosis accuracy.

[0068] In the present invention, the interaction between the dynamic structure neural network module and the online self-optimizing weight allocation module realizes the dynamic optimization of model stability and diagnostic accuracy through a multi-level constraint mechanism. During the generation process of the branch switching instruction of the dynamic structure neural network module, the fundamental frequency harmonic features extracted by the spectrum analysis subnet are input into the dynamic knowledge base construction module for fault mode matching. The matching result is appended with the device operation status label to generate a switching instruction containing the target fault type code. After the switching instruction is transmitted to the online self-optimizing weight allocation module, it triggers the feature contribution degree reference value retrieval process, and calls the reference weight distribution data of the current device model in the corresponding fault mode from the dynamic knowledge base construction module.

[0069] The online self-optimizing weight allocation module encodes the reference weight distribution data into the regularization constraint term of the gated recurrent unit, and restricts the deviation amplitude of the real-time weight from the reference value during the weight calculation stage. The gated recurrent unit fuses the real-time working condition parameters and the regularization constraint conditions to generate the feature importance weights that take into account both data-driven features and physical rules. The weight allocation result is appended with the deviation amount statistical indicators of each dimension to form a quantization feedback vector containing the deviation direction and amplitude, which is transmitted to the dynamic structure neural network module.

[0070] The dynamic structure neural network module analyzes the deviation amount statistical indicators in the feedback vector. When a continuous deviation in the key feature dimension is detected, it triggers the online fine-tuning instruction generation process. The fine-tuning instruction carries the deviation feature dimension identification code and the correction gradient direction, and drives the sliding window type incremental training of the parameters of the last layer of the target network branch. During the training process, the time series segments in the current data stream are used to construct a mini-batch data set, and the convolution kernel weight distribution is adjusted based on the gradient correction amount to suppress the decrease in diagnostic accuracy caused by feature drift.

[0071] The online fine-tuning operation synchronously triggers the parameter version management process of the dynamic knowledge base construction module, and records the current branch parameter snapshot and the fine-tuning operation log. When the confidence level of the output of the fine-tuned model fluctuates abnormally after being verified by adversarial samples, the parameter version rollback mechanism automatically rolls back to the network parameters of the previous stable version according to the time stamp recorded in the log and the device status label. After the rollback operation is completed, new adversarial samples are generated and injected into the training process to optimize the network robustness and update the adversarial sample generation strategy in the knowledge base.

[0072] The above interaction process forms a dynamic balance between model stability maintenance and diagnostic accuracy optimization through a closed-loop design of reference value constraint, weight deviation feedback, online fine-tuning and parameter rollback. The reference value call driven by the branch switching instruction ensures the effectiveness of physical rule constraints, the fine-tuning mechanism triggered by weight deviation realizes feature adaptability optimization, and the parameter version rollback provides the ability to quickly recover under abnormal conditions. The data flow and operation coordination between modules effectively solve the technical contradiction that it is difficult to be compatible between model solidification and working condition drift in traditional methods.

[0073] In the specific implementation of the present invention, first, a temperature sensor array, a wind and flue gas pressure transmitter, a flue gas composition analyzer, and an on-line coal quality detection device are deployed in the boiler combustion system to form a multi-source data acquisition network. The temperature sensors are installed at the four corners and the middle of the furnace in a cross shape, and the temperature field distribution data is collected every 10 seconds; the air volume measuring device uses a Venturi tube combined with a differential pressure transmitter to collect the primary air and secondary air flow rates at a second-level frequency; the coal quality detector is used in combination with a belt scale and a laser particle size analyzer to output a histogram of the coal powder particle size distribution every minute. The collected original data is transmitted to the edge computing node through the industrial Ethernet, and mean filtering noise reduction processing is performed. The length of the filtering window is set to 3 times the coal powder conveying cycle to eliminate the pulse interference of the fly ash carbon content signal.

[0074] In the data preprocessing stage, principal component analysis is performed on the denoised air-coal ratio and oxygen concentration parameters, and the first 3 principal components with a cumulative variance contribution rate of up to 85% are retained, and the excess air coefficient and the adjustment sensitivity of the secondary air damper opening are extracted as key sensitive factors. Aiming at the time series alignment problem between the minute-level coal powder flow data and the second-level air volume data, the dynamic time warping algorithm is used to set a path constraint with a horizontal offset of no more than 5 seconds, and cubic spline interpolation is performed on the air volume data to generate a standardized feature matrix with a unified time baseline. This matrix contains 28-dimensional features such as coal quality characteristics, air volume ratio, and temperature field standard deviation, which are used as the input of the dynamic coupling model.

[0075] When constructing the dynamic coupling model, the gradient boosting decision tree is set to a maximum depth of 7 layers, and the influence weight of each bin interval of the coal powder particle size distribution on the air-coal ratio is analyzed based on the Gini coefficient splitting criterion to generate a feature importance ranking table. The long short-term memory network adopts a 3-layer 128-unit structure, and a time delay compensation layer is embedded in the second layer. According to the difference matrix between the air damper adjustment timestamp and the temperature response timestamp, the forgetting gate weight coefficient is dynamically adjusted to compensate for the combustion thermodynamics inertia delay. When the coal quality detector recognizes a coal type change or a load fluctuation exceeding 15%, sliding window incremental training is triggered, and the working condition data of the most recent 2 hours is retained, and the Adam optimizer is used to update the model parameters at a learning rate of 0.001.

[0076] When the multi-objective optimization module runs, the weight of the combustion efficiency sub-problem is set in a dynamic range of 0-0.6 based on the volatile content in the real-time coal quality characteristics, and the ash softening temperature parameter determines that the weight of the temperature field uniformity fluctuates between 0.2-0.4. The Chebyshev decomposition method is used to decompose the three-objective problem into 12 scalar sub-problems, and the Pareto front solution set is generated through neighborhood search. The fuzzy membership function sets the steam pressure fluctuation threshold of ±0.5 MPa. After screening the feasible solution domain, a burner tilt-temperature field gradient influence matrix is constructed in combination with historical operation data to generate a hierarchical control instruction set with priority ranking.

[0077] During the instruction execution phase, the distributed control system decomposes the secondary air damper coarse adjustment instruction into 3 sub - instructions with a 5% opening step size, each with an interval of 30 seconds; the fine adjustment instruction is issued at a second - level frequency with an accuracy of 0.5%. The burner tilt compensation adopts PID closed - loop control. The stepping motor divides each 15 - degree range into an adjustment interval, and simultaneously monitors the change in the standard deviation of the temperature field. The actual combustion data after execution is uploaded via the OPC protocol. When the thermal efficiency deviation exceeds 2% continuously for 3 times, the model self - calibration process is triggered: the long - short - term memory network updates the weights using the RMSProp optimizer, the gradient - boosting decision tree adjusts the splitting threshold to above 0.01 of the information gain rate, and the updated parameters are injected into the online model after MD5 verification.

[0078] For data closed - loop management, a circular buffer is used to store the operating condition data of the most recent 72 hours. Data rolling update is performed every 15 minutes, and expired data is automatically archived after being marked. The incremental training module loads data in 32 batches using the mini - batch method, and the learning rate decay coefficient is set to 0.95 to prevent model oscillation. The weight dynamic adjustment module calculates the emission concentration deviation rate every hour. When the NOx exceedance duration exceeds 10 minutes, the weight of the pollutant sub - problem is increased by 0.1, and the weight fluctuation is smoothed through a moving average algorithm to maintain the stability of multi - objective optimization.

[0079] The above - mentioned implementation method aligns multi - resolution data through the dynamic time warping algorithm to solve the problem of time - series mismatch; combines the gradient - boosting decision tree with the long - short - term memory network to analyze the non - linear relationship of the pulverized coal particle size distribution and capture the combustion time - lag effect; the multi - objective optimization algorithm generates hierarchical control instructions based on the real - time coal quality characteristics to achieve the balance between efficiency and environmental protection; the closed - loop feedback mechanism suppresses the influence of operating condition drift through model self - calibration and data rolling update, forming a complete control link from data acquisition to execution feedback.

[0080] The present invention solves the problems of time - scale asynchrony and feature heterogeneity in multi - source heterogeneous data fusion through a knowledge - guided time - series alignment and feature space mapping mechanism. The operating condition adaptive data synchronization module dynamically selects an interpolation algorithm for different rotational speed intervals and pressure mutation scenarios according to the interpolation strategy template provided by the dynamic knowledge base construction module. When it is detected that the rotational speed enters the critical interval, the cubic spline interpolation template is called to compensate for the sensor sampling drift; when the steam pressure mutates, it switches to an adaptive interpolation mode based on the device response characteristics and adjusts the timestamp matching tolerance threshold. The synchronized data with timestamp metadata is fed back to the knowledge base to drive the gradient update of the interpolation strategy parameters, forming a collaborative optimization of the time - series alignment accuracy and the operating condition mode representation ability of the knowledge base.

[0081] Regarding the heterogeneity of the feature space, the knowledge-guided heterogeneous feature fusion module realizes cross-domain feature mapping through physical rule constraints. The frequency band sensitivity matrix output by the dynamic knowledge base construction module encodes the historical fault frequency domain energy distribution as channel attention weights, guiding the convolutional neural network to strengthen the feature extraction of fault-sensitive frequency bands. After being verified by the fault classifier, the feature map generates a feature importance ranking, which is fed back to the knowledge base to trigger the gradient adjustment of the frequency band sensitivity matrix, forming a dynamic adaptation mechanism between physical rules and data features. This process establishes a unified knowledge-driven feature space, realizing the cross-domain association of time-frequency features of vibration signals, temperature gradient features, and maintenance record semantic features.

[0082] The cooperation mechanism between the dynamic structure neural network module and the online self-optimizing weight allocation module further optimizes the processing of feature heterogeneity. The network branch switching instruction triggers the weight allocation module to call the feature contribution benchmark value in the knowledge base, and the real-time weight is constrained within the deviation range allowed by the physical rules through the gated recurrent unit. The weight deviation information triggers the online fine-tuning of network parameters and the rollback of knowledge base parameters, suppressing the model degradation caused by feature space drift. The knowledge-enhanced coupled fault inference module fuses the data-driven results with the fault propagation logic rules, verifies the physical relevance of cross-domain features through numerical simulation, screens out effective association patterns and injects them into the knowledge base, and finally realizes the reliable fusion and accurate extraction of multi-source heterogeneous features.

Claims

1. A steam turbine vibration fault diagnosis system integrating deep learning, characterized in that: include: A dynamic knowledge base construction module is used to integrate historical fault cases, rotor dynamics equations and expert rules, construct a multimodal knowledge graph containing fault mode-feature-maintenance strategy relationships, and update graph edge weights through an incremental knowledge distillation mechanism. The knowledge graph outputs an interpolation strategy template to the data synchronization module; A working condition adaptive data synchronization module receives real-time sensor data and calls the interpolation strategy template generated by the dynamic knowledge base construction module to perform time alignment, transmits the synchronized data to the feature fusion module after attaching timestamp metadata, and feeds back the timestamp error distribution to the dynamic knowledge base construction module to optimize the interpolation strategy template; The knowledge-guided heterogeneous feature fusion module receives the synchronized data output by the working condition adaptive data synchronization module, inputs the time-frequency analysis features of the vibration signal into the convolutional neural network, adjusts the channel attention weight based on the frequency band sensitivity matrix provided by the dynamic knowledge base construction module, generates fusion features under physical constraints and transmits them to the dynamic structure neural network module; The dynamic structure neural network module activates the corresponding network branch and loads the pre-trained model parameters according to the fusion features output by the knowledge-guided heterogeneous feature fusion module and the fault mode matching results of the dynamic knowledge base construction module, extracts the fault features and transmits them to the online self-optimization weight distribution module; An online self-optimizing weight allocation module receives the fault features extracted by the dynamic structure neural network module, combines the real-time operating condition parameters with the feature contribution benchmark value in the dynamic knowledge base construction module, dynamically allocates feature importance weights through a gated recurrent unit, and feeds back weight deviation information to the dynamic knowledge base construction module to trigger model parameter tuning; The knowledge-enhanced coupled fault reasoning module receives the feature weights assigned by the online self-optimizing weight assignment module and the fault propagation logic of the dynamic knowledge base construction module, performs confidence fusion on the data-driven classification results and the physical rules, generates a diagnostic conclusion and triggers an incremental update of the knowledge graph of the dynamic knowledge base construction module, thus forming a closed-loop feedback mechanism.

2. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1 is characterized in that: The incremental knowledge distillation mechanism in the dynamic knowledge base construction module specifically includes: The online diagnosis results generated by the knowledge-enhanced coupled fault reasoning module trigger the update of the edge weights of the knowledge graph, automatically extract new fault feature vectors and association rules to generate adversarial samples, and inject the adversarial samples into the training data stream of the dynamic structure neural network module; The updated knowledge graph edge weights are transmitted to the online self-optimizing weight allocation module to constrain the regularization term calculation process of the gated recurrent unit so that the feature weight allocation is synchronized with the updated physical rules.

3. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1 is characterized in that: When the working condition adaptive data synchronization module performs timing alignment: The interpolation coefficient template is called from the dynamic knowledge base construction module according to the real-time speed fluctuation pattern. When a sudden change in steam pressure is detected, the dynamic knowledge base construction module is linked to retrieve the sensor response pattern under similar working conditions. The corrected timestamp matching tolerance threshold is fed back to the dynamic knowledge base construction module through the working condition adaptive data synchronization module, triggering the gradient update of the interpolation strategy template, forming an adaptive collaborative optimization of data synchronization accuracy and knowledge graph parameters.

4. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1 is characterized in that: In the knowledge-guided heterogeneous feature fusion module: The feature map output by the convolutional neural network carries the physical rule constraint information, and after verification by the fault classifier, the feature importance ranking is generated; The feature importance ranking is fed back to the dynamic knowledge base construction module, triggering the gradient update of the frequency band sensitivity matrix, so that the frequency band weight distribution is dynamically adapted to the current equipment failure mode.

5. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1 is characterized in that: When the dynamic structure neural network module activates the network branch: The fundamental frequency component ratio extracted by the spectrum analysis subnet is input into the dynamic knowledge base construction module for fault mode retrieval, generating branch switching instructions and loading pre-trained model parameters; The activated network branch outputs the feature map to the online self-optimization weight allocation module, which simultaneously triggers the adversarial sample verification process in the dynamic knowledge base construction module to suppress the feature distribution anomaly caused by model parameter offset.

6. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1 is characterized in that: When the online self-optimizing weight allocation module performs dynamic weight allocation: Calling the characteristic contribution benchmark value of the same type of equipment stored in the dynamic knowledge base construction module, and calculating the deviation between the real-time weight and the benchmark value; When the deviation exceeds a preset threshold, the online fine-tuning instruction of the dynamic structure neural network module is triggered, and the adversarial sample generation strategy in the dynamic knowledge base construction module is updated. The adversarial sample generation strategy is used to optimize the training data flow of the dynamic structure neural network module.

7. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1 is characterized in that: When the knowledge-enhanced coupled fault reasoning module performs confidence fusion: The fault probability output by the data-driven classifier is jointly inferred with the logical rules of the fault propagation directed graph in the dynamic knowledge base building module; The inference result triggers the numerical simulation verification process, and the verified cases generate incremental training data and are transmitted back to the dynamic knowledge base construction module to update the edge weight parameters of the fault coupling relationship in the fault propagation directed graph.

8. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1, characterized in that: The system exception handling process includes: when the branch heartbeat of the dynamic structure neural network module is lost, the control center sends a downgrade instruction to the working condition adaptive data synchronization module to switch to the basic interpolation mode; The rule engine and the fail-safe weight template in the dynamic knowledge base building module are activated synchronously to maintain the diagnostic function and record abnormal events to the dynamic knowledge base building module to generate a training data set for the network branch health prediction model.

9. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1, characterized in that: The collaboration between the dynamic knowledge base construction module and the feature fusion module includes: The dynamic knowledge base construction module outputs a frequency band sensitivity matrix to the knowledge-guided heterogeneous feature fusion module to constrain the channel attention calculation process of the convolutional neural network; The feature importance ranking fed back by the knowledge-guided heterogeneous feature fusion module triggers the update of the historical fault case spectrum fingerprint library in the dynamic knowledge base construction module, forming a two-way optimization link between physical rule constraints and data-driven features.

10. The steam turbine vibration fault diagnosis system integrating deep learning according to claim 1, characterized in that: The interaction between the dynamic structure neural network module and the weight distribution module includes: The branch switching instruction of the dynamic structure neural network module triggers the online self-optimization weight allocation module to call the feature contribution benchmark value to constrain the regularization term calculation of the gated recurrent unit; The feature weight deviation information output by the online self-optimizing weight allocation module triggers the online fine-tuning instructions of the dynamic structure neural network module and starts the parameter version rollback mechanism of the dynamic knowledge base construction module to achieve a dynamic balance between model stability and diagnostic accuracy.

Images