Automatic diagnosis and alarm device for vibration fault of steam turbine generator unit

Abstract

The invention discloses an automatic diagnosis and alarm device for a vibration fault of a steam turbine generator unit. According to the invention, through the synergistic effect of the multi-modal fusion self-adaptive diagnosis module, the accuracy and adaptability of the vibration fault diagnosis of the steam turbine generator unit are significantly improved. The module integrates time domain and frequency domain characteristics of vibration signals and auxiliary parameters such as temperature and rotating speed to form a multi-dimensional input matrix, and realizes a dual verification mechanism by combining a rule base constructed based on a typical fault atlas base and a CNN-LSTM deep learning network. The rule base generates quantifiable diagnosis rules based on historical fault data, and faults with clear characteristics such as rotor imbalance and misalignment are rapidly matched; and the deep learning model optimizes the diagnosis result through a dynamic weight fusion algorithm, so that the system still keeps stable diagnosis capability under the scenes of equipment aging, working condition change and the like, and subjective errors and missed diagnosis risks of traditional manual judgment are effectively reduced.

Description

Technical Field

[0001] This invention belongs to the technical field of the field of steam turbine generator set vibration fault automatic diagnosis and alarm device. Background Technology

[0002] With the continuous development of automation technology in the power industry, building "smart power plants" is a major trend in my country's power plant development. Steam turbine generator sets are the most important equipment in coal-fired power plants. They are core power equipment that converts heat energy into electrical energy and are widely used in thermal power generation, nuclear power generation, and industrial waste heat utilization. They mainly consist of a steam turbine and a generator connected by a rigid coupling. The steam turbine utilizes the expansion of high-temperature, high-pressure steam to drive the impeller to rotate at high speed, converting heat energy into mechanical energy; the generator, driven by this mechanical energy, generates electrical energy through electromagnetic induction. As a key component of modern power systems, steam turbine generator sets have large single-unit capacity, high operating efficiency, and strong stability. Their technical performance directly determines the economic efficiency and power supply reliability of the power plant, serving as the cornerstone of ensuring social energy supply.

[0003] However, existing technologies mainly rely on manual inspection and experience-based judgment, and the diagnostic results are greatly affected by the personnel's ability, which can easily lead to missed or incorrect diagnoses. Fault handling is mostly done after the fact or during regular maintenance, making it difficult to predict potential problems in advance, resulting in high equipment downtime risks and high maintenance costs. At the same time, traditional monitoring methods have limited ability to identify complex faults and are not adaptable enough to equipment aging or changes in operating conditions, failing to meet the needs of smart power plants for precise and proactive operation and maintenance. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic diagnosis and alarm device for vibration faults in steam turbine generator sets in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: an automatic diagnosis and alarm device for vibration faults of steam turbine generator sets, wherein the device is equipped with an automatic diagnosis and alarm system for vibration faults of steam turbine generator sets, the system including: a vibration data acquisition module, a feature extraction and signal analysis module, a multimodal fusion adaptive diagnosis module, an alarm and graded response module, a fault map library knowledge management module and a preventive maintenance decision module; The multimodal fusion adaptive diagnostic module is internally configured with: a multi-source data fusion submodule, an adaptive diagnostic model submodule, and a fault evolution prediction submodule; The data output terminal of the vibration data acquisition module is connected to the raw signal input terminal of the feature extraction and signal analysis module; The feature parameter output of the feature extraction and signal analysis module is connected to the multi-source feature input of the multimodal fusion adaptive diagnosis module; The diagnostic result output terminal of the multimodal fusion adaptive diagnostic module is connected to the fault information input terminal of the alarm and graded response module and the diagnostic conclusion input terminal of the preventive maintenance decision module, respectively. At the same time, its graph query request terminal is bidirectionally connected to the graph data interface of the fault graph library knowledge management module. The maintenance plan output terminal of the fault map knowledge management module is connected to the plan reference input terminal of the preventive maintenance decision module. The alarm log storage terminal of the alarm and graded response module is connected to the log recording input terminal of the fault map library knowledge management module, and the maintenance linkage output terminal is connected to the emergency maintenance request input terminal of the preventive maintenance decision module; the maintenance status feedback terminal of the preventive maintenance decision module is connected to the processing status input terminal of the alarm and graded response module.

[0006] In a preferred embodiment of the present invention, the vibration data acquisition module internally comprises a sensor deployment layer, a signal conditioning unit, a data preprocessing submodule, and a data transmission interface. The sensor deployment layer determines the measurement point locations based on the characteristics of key components of the turbine generator set, such as bearing housings and rotor shaft systems, combined with layout specifications and technical requirements. Piezoelectric accelerometers are selected, covering a measurement range of 0 to 50g, and the sampling frequency is configured to be at least 20 times the highest operating frequency of the unit to ensure the capture of high-frequency fault characteristics. The signal conditioning unit includes a charge amplifier and a low-pass filter. The charge amplifier converts the charge signal output by the sensor into a standard voltage signal, and the low-pass filter sets the cutoff frequency to half the sampling frequency to filter out high-frequency interference. The data preprocessing submodule performs segmented smoothing of the signal, with each segment lasting five times the unit's rotation period. An amplitude-limiting filter is used to remove abnormal jump values, and normalization is performed to provide standardized data for subsequent generation of spectrum and waveform diagrams. The data transmission interface adopts the industrial Ethernet protocol, packaging and transmitting preprocessed data at fixed time intervals, and ensuring integrity through CRC verification to support real-time monitoring requirements.

[0007] As a preferred embodiment of the present invention, the feature extraction and signal analysis module internally comprises a time-domain feature extraction unit, a frequency-domain feature analysis unit, a time-frequency domain feature fusion unit, and a spectrum generation submodule. The time-domain feature extraction unit calculates indicators such as peak value, kurtosis, and root mean square value, and updates them in real time using a sliding window technique. The frequency-domain feature analysis unit uses Fast Fourier Transform to generate a spectrum, extracts the amplitude and phase of 1X, 2X, and 3X harmonic components, and calculates the spectral centroid and spectral kurtosis. The time-frequency domain feature fusion unit generates an envelope diagram using wavelet transform, extracts high-frequency impulse features, and analyzes the amplitude change rate using a trend graph to form a multi-dimensional feature matrix. The spectrum generation submodule transforms the processed feature parameters into four types of visual graphs: spectrum, waveform, trend, and envelope, which are stored according to fault type, providing an intuitive analytical basis for fault diagnosis.

[0008] As a preferred embodiment of the present invention, the multi-source data fusion submodule internally comprises a data access layer, a data cleaning unit, a feature fusion algorithm module, and a fusion result storage unit. The data access layer supports the access of multiple data types, including vibration sensor data, process parameters (temperature, pressure, speed), and equipment ledger information. It uses protocols such as Industrial Ethernet, Modbus, and OPC UA to connect to different data sources and establishes a unified data index based on equipment number and timestamp. The data cleaning unit addresses the noise characteristics of data from different sources by using a moving average method to process vibration signal drift, removing outliers in process parameters using the 3σ criterion, and completing missing data using an interpolation algorithm based on similar operating conditions to ensure data integrity. The feature fusion algorithm module uses a weighted average to fuse time-domain and frequency-domain features, dynamically adjusting the weight ratio of vibration data and auxiliary parameters through an attention mechanism, and generating a fused feature matrix by combining typical feature vectors from a fault map library. The fusion result storage unit stores fused data categorized by equipment type and fault type. It uses a relational database to record structured feature parameters and a file system to store unstructured map data such as spectrum diagrams and trend charts, providing a unified data interface for the diagnostic model.

[0009] As a preferred embodiment of the present invention, the adaptive diagnostic model submodule uses a rule-based prior matching + deep learning dynamic optimization as its core framework to achieve high-precision diagnosis of vibration faults in steam turbine generator sets. The rule base is constructed based on historical fault data, focusing on typical faults such as rotor imbalance, misalignment, and oil film whirl, transforming their features into directly matchable quantitative rules. For example, for rotor imbalance faults, by analyzing the vibration signal spectrum characteristics corresponding to this fault in historical cases, it is determined that when the amplitude of the 1X frequency exceeds a specific threshold and the phase remains stable within the monitoring period, the "rotor imbalance" rule matching can be triggered. For rotor misalignment faults, it is set that when the ratio of the 2X frequency amplitude to the 1X frequency amplitude is greater than 0.5, the system automatically matches the "rotor misalignment" rule. The key parameters of these rules are uniformly stored in a feature threshold matrix, supporting real-time invocation and static feature matching, providing a basic judgment basis for fault diagnosis.

[0010] The deep learning model employs a CNN-LSTM hybrid network structure, using multi-dimensional features as input. These input features encompass the time-domain and frequency-domain features of the vibration signal, as well as auxiliary parameters during unit operation; these features collectively constitute a multimodal feature vector. The model's convolutional layers process the input features through multiple convolutional kernels, extracting key local features and generating high-dimensional feature maps. The LSTM layers further capture the dynamic changes of these features over time, ultimately outputting probability vectors for various fault types.

[0011] The fusion and optimization mechanism combines the fault matching results output by the rule base with the fault probability vector output by the deep learning model through a dynamic weight fusion algorithm to obtain the final fault diagnosis result. Simultaneously, the model possesses adaptive evolution capabilities: based on real-time fault feedback data, it continuously adjusts the parameters in the feature threshold matrix and optimizes the network weights and biases of the deep learning model, enabling the diagnostic rules and model parameters to dynamically adapt to changes in the random group's operating state, continuously improving diagnostic accuracy.

[0012] As a preferred embodiment of the present invention, the rule base of the adaptive diagnostic model submodule has a confidence level for type II faults. It is obtained by weighting feature matching degree and rule priority, and the calculation formula is: ; In the formula: m represents the number of features of the i-th type of fault (e.g., rotor imbalance includes two features: amplitude and phase). α i,k The weight of the k-th feature of the i-th type of fault (set by expert experience) ; xk represents the k-th feature value monitored in real time (such as 1X frequency amplitude). This represents the mean and variance of the k-th feature of the ⅂ type of fault in the rule base (obtained through statistical analysis of historical fault data). The input feature vector X of CNN-LSTM achieves multimodal fusion through normalization and dynamic weight concatenation: ; In the formula: This represents the standardized vibration eigenvector (* indicates normalization). Represents a standardized auxiliary parameter vector (temperature T, rotational speed n); β represents the vibration characteristic weighting coefficient (dynamically adjusted; when the vibration signal-to-noise ratio (SNR) < 10dB, β = 0.3 to enhance the weighting of auxiliary parameters). Final failure probability P final,i The calculation formula is: ; ; In the formula: P DL,i This represents the probability of type II failure output by the deep learning model; γ t This represents the rule base weight coefficient at time t (initial value γ0 = 0.7, adaptively decaying as the deep learning model error accumulates). y j The label represents the actual fault label for the j-th diagnosis (1 indicates occurrence, 0 indicates non-occurrence).

[0013] Based on diagnostic error feedback, the weights W of the deep learning model and the rule base threshold Θ are updated according to the following formula: ;

[0014] In the formula: ηt represents the adaptive learning rate. ; L(·) represents the cross-entropy loss function; λ represents the rule base threshold update step size (λ=0.005); N represents the number of historical diagnostic samples.

[0015] As a preferred embodiment of the present invention, the fault evolution prediction submodule internally includes a historical data training unit, a real-time status assessment unit, a trend prediction model, and a maintenance timing recommendation unit. The historical data training unit extracts full lifecycle data of typical faults such as rotor imbalance, misalignment, and oil film whirl from the fault map library, classifying them into four levels—normal, minor, moderate, and severe—based on fault severity, and constructs a fault evolution sample set. The real-time status assessment unit calculates the equipment health index by comparing the Euclidean distance between the current fused features and historical samples; when the index falls below a threshold, the prediction process is triggered. The trend prediction model uses an LSTM neural network, taking the fused feature matrix and the time-series health index as input, and outputting the fault probability curve and confidence interval for the next 30 days, focusing on capturing the evolution patterns of key indicators such as the 1X frequency amplitude growth rate and the 2X / 1X amplitude ratio change rate. The maintenance timing recommendation unit combines the equipment operation plan and predicted fault time to generate a maintenance suggestion plan including the optimal maintenance window, required spare parts models, and estimated working hours. The plan is associated with standardized processes in the preventive maintenance mechanism and can be exported as an executable work order.

[0016] In a preferred embodiment of the present invention, the alarm and graded response module internally includes a fault information receiving unit, a multi-level threshold configuration submodule, a response strategy generation unit, and an alarm log storage unit. The fault information receiving unit receives fault types, characteristic parameters, and graph data output by the diagnostic module and stores them according to fault codes. The multi-level threshold configuration submodule sets three levels based on historical data from the fault graph library: Level 1 warning (characteristic parameters exceeding normal range), Level 2 alarm (approaching critical value), and Level 3 emergency response (exceeding critical value). The threshold standards are dynamically associated with typical fault characteristics in the graph library. The response strategy generation unit triggers corresponding processes according to the fault level: Level 1 warnings are pushed to employee computer terminals and automatically recorded; Level 2 alarms initiate team-level maintenance preparation; and Level 3 emergency responses trigger shutdown protection signals and notify the technical supervisor. The alarm log storage unit records alarm information, response measures, and processing status in timestamp order, stores them in CSV format, and supports retrieval by fault type and time range.

[0017] As a preferred embodiment of the present invention, the fault spectrum knowledge management module internally includes a typical fault spectrum storage unit, a rule base management submodule, a fault case input interface, and a threshold parameter maintenance unit. The typical fault spectrum storage unit stores spectrum diagrams, waveform diagrams, trend diagrams, and envelope diagrams of typical faults such as rotor imbalance, misalignment, and oil film whirl in a structured format. Each spectrum is associated with a fault name, characteristic parameter range, and probability of occurrence. The rule base management submodule stores diagnostic rules based on spectrum features, such as "1X frequency amplitude exceeds threshold T". 1x "If the phase is stable, it is determined to be rotor imbalance." The rule is stored in a parsable text format and supports editing via a visual interface. The fault case entry interface receives the characteristic parameters, diagnostic results, and maintenance records of new fault cases. After format verification, it is stored in the case library. The case includes four types of graph data at the time of the fault. The threshold parameter maintenance unit periodically adjusts the characteristic threshold matrix based on real-time diagnostic data to ensure dynamic matching with the unit's operating status. The adjustment record includes the values ​​before and after the adjustment and the basis for the adjustment.

[0018] As a preferred embodiment of the present invention, the preventive maintenance decision-making module internally includes a diagnostic conclusion receiving unit, a maintenance plan matching submodule, a resource scheduling and coordination unit, and a maintenance plan generation unit. The diagnostic conclusion receiving unit receives fault type, location, and severity data, converting it into a standardized decision input format. The maintenance plan matching submodule retrieves the corresponding preventive maintenance process from the plan library based on the fault type. The plan includes fault handling steps, required tools, materials, and component models, and is associated with typical cases in the fault diagram library. The resource scheduling and coordination unit interfaces with the equipment management system to query spare parts inventory, storage locations, and maintenance team scheduling, assessing resource availability. The maintenance plan generation unit, combining fault severity and unit operation plan, formulates a maintenance plan including start time and responsible personnel, outputs it in Gantt chart format, and synchronizes it to the production management system, implementing a predictive maintenance mechanism and improving equipment continuous operation cycle.

[0019] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, the accuracy and adaptability of vibration fault diagnosis for steam turbine generator sets are significantly improved through the synergistic effect of a multimodal fusion adaptive diagnostic module. The module integrates the time-domain and frequency-domain features of vibration signals with auxiliary parameters such as temperature and speed to form a multi-dimensional input matrix. This matrix, combined with a rule base built on a typical fault map library and a CNN-LSTM deep learning network, achieves a dual verification mechanism. The rule base generates quantifiable diagnostic rules based on historical fault data, quickly matching faults with clear characteristics such as rotor imbalance and misalignment. The deep learning model optimizes the diagnostic results through a dynamic weight fusion algorithm, achieving higher accuracy, especially for complex faults such as oil film eddy. The adaptive evolution mechanism updates the rule base thresholds and network parameters based on real-time feedback data, ensuring stable diagnostic capabilities even under scenarios such as equipment aging and changing operating conditions, effectively reducing the subjective errors and missed diagnoses associated with traditional manual judgment.

[0020] 2. In this invention, the system promotes a shift in maintenance strategies from passive response to proactive prevention. Through real-time monitoring and intelligent diagnostics, the system can detect fault signs in advance and issue tiered alarms, providing maintenance personnel with sufficient time to handle the situation and avoiding sudden downtime. The fault evolution prediction function, combined with equipment lifecycle data, generates trend analysis reports, supporting the development of precise preventative maintenance plans and reducing resource waste caused by blind maintenance. Simultaneously, the system replaces traditional manual inspection methods, freeing maintenance personnel from repetitive tasks, reducing the impact of human factors on monitoring effectiveness, and improving the standardization of equipment management. Over the long term, the continuous operating cycle of equipment is extended, maintenance costs are reduced, providing solid technical support for the safe and stable operation of power plants, aligning with the digital and intelligent development direction of smart power plant construction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an overall system block diagram of the present invention; Figure 2 This is a system block diagram of the multimodal fusion adaptive diagnostic module in this invention. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Example: Reference Figure 1-2 An automatic diagnosis and alarm device for vibration faults of steam turbine generator sets is provided. The device is equipped with an automatic diagnosis and alarm system for vibration faults of steam turbine generator sets. The system includes: a vibration data acquisition module, a feature extraction and signal analysis module, a multimodal fusion adaptive diagnosis module, an alarm and graded response module, a fault map library knowledge management module, and a preventive maintenance decision module. The multimodal fusion adaptive diagnosis module is internally configured with: a multi-source data fusion submodule, an adaptive diagnosis model submodule, and a fault evolution prediction submodule; The data output terminal of the vibration data acquisition module is connected to the raw signal input terminal of the feature extraction and signal analysis module, transmitting the preprocessed vibration signal to the feature extraction module; The feature parameter output of the feature extraction and signal analysis module is connected to the multi-source feature input of the multimodal fusion adaptive diagnosis module, transmitting time-domain, frequency-domain, and time-frequency-domain feature parameters; The diagnostic result output of the multimodal fusion adaptive diagnostic module is connected to the fault information input of the alarm and graded response module and the diagnostic conclusion input of the preventive maintenance decision module, respectively. At the same time, its graph query request end is bidirectionally connected to the graph data interface of the fault graph library knowledge management module (for reading standard graph data), and the case update end is connected to the case entry interface of the fault graph library knowledge management module (for writing new fault cases). The maintenance plan output of the fault map library knowledge management module is connected to the plan reference input of the preventive maintenance decision module, and the threshold parameter interface is bidirectionally connected to the threshold configuration end of the alarm and graded response module (used to read historical threshold standards and feedback adjustment results). The alarm log storage end of the alarm and graded response module is connected to the log recording input end of the fault map library knowledge management module, and the maintenance linkage output end is connected to the emergency maintenance request input end of the preventive maintenance decision module; the maintenance status feedback end of the preventive maintenance decision module is connected to the processing status input end of the alarm and graded response module, forming a closed-loop management.

[0027] The vibration data acquisition module internally comprises a sensor deployment layer, a signal conditioning unit, a data preprocessing submodule, and a data transmission interface. The sensor deployment layer determines the measurement point locations based on the characteristics of key components such as the turbine generator set's bearing housing and rotor shaft system, combined with site layout specifications and technical requirements. Piezoelectric accelerometers are selected, covering a measurement range of 0 to 50g, and the sampling frequency is configured to be at least 20 times the unit's highest operating frequency to ensure the capture of high-frequency fault characteristics. The signal conditioning unit includes a charge amplifier and a low-pass filter. The charge amplifier converts the sensor's output charge signal into a standard voltage signal, and the low-pass filter has its cutoff frequency set at half the sampling frequency to filter out high-frequency interference.

[0028] The data preprocessing submodule performs segmented smoothing on the signal, with each segment lasting five times the unit's rotation cycle. It employs amplitude-limiting filtering to eliminate abnormal jump values ​​and performs normalization processing to provide standardized data for subsequent generation of spectrum and waveform diagrams. The data transmission interface uses the industrial Ethernet protocol, packaging and transmitting preprocessed data at fixed time intervals. CRC checksums ensure integrity, supporting real-time monitoring requirements.

[0029] The feature extraction and signal analysis module internally comprises a time-domain feature extraction unit, a frequency-domain feature analysis unit, a time-frequency domain feature fusion unit, and a spectrum generation submodule. The time-domain feature extraction unit calculates metrics such as peak value, kurtosis, and root mean square value, and updates them in real time using a sliding window technique. The frequency-domain feature analysis unit uses Fast Fourier Transform to generate a spectrum, extracting the amplitude and phase of 1X, 2X, and 3X harmonic components, and calculating the spectral centroid and kurtosis. The time-frequency domain feature fusion unit generates an envelope graph using wavelet transform, extracts high-frequency impulse features, and analyzes the amplitude change rate using a trend graph, forming a multi-dimensional feature matrix.

[0030] The spectrum generation submodule transforms the processed feature parameters into four types of visual spectra: spectrum, waveform, trend, and envelope. These spectra are then stored according to fault type, providing an intuitive analytical basis for fault diagnosis.

[0031] The multi-source data fusion submodule internally comprises a data access layer, a data cleaning unit, a feature fusion algorithm module, and a fusion result storage unit. The data access layer supports various data types, including vibration sensor data, process parameters (temperature, pressure, speed), and equipment ledger information. It uses protocols such as Industrial Ethernet, Modbus, and OPC UA to connect to different data sources and establishes a unified data index based on equipment number and timestamp. The data cleaning unit addresses the noise characteristics of data from different sources by using a moving average method to handle vibration signal drift, removing outliers in process parameters using the 3σ criterion, and completing missing data using interpolation algorithms based on similar operating conditions to ensure data integrity.

[0032] The feature fusion algorithm module employs a weighted average to fuse time-domain and frequency-domain features. It dynamically adjusts the weight ratios of vibration data and auxiliary parameters through an attention mechanism, and generates a fused feature matrix by combining typical feature vectors from the fault map library. The fusion result storage unit stores the fused data categorized by equipment type and fault category. A relational database records structured feature parameters, while a file system stores unstructured map data such as spectrum diagrams and trend charts, providing a unified data interface for the diagnostic model.

[0033] The adaptive diagnostic model submodule uses a rule-based prior matching and deep learning dynamic optimization as its core framework to achieve high-precision diagnosis of vibration faults in steam turbine generator sets. The rule base is constructed based on historical fault data, focusing on typical faults such as rotor imbalance, misalignment, and oil film whirl, transforming their features into directly matchable quantitative rules. For example, for rotor imbalance faults, by analyzing the vibration signal spectrum characteristics corresponding to this fault in historical cases, it is determined that when the amplitude of the 1X frequency exceeds a specific threshold and the phase remains stable within the monitoring period, the "rotor imbalance" rule matching can be triggered. For rotor misalignment faults, the system automatically matches the "rotor misalignment" rule when the ratio of the 2X frequency amplitude to the 1X frequency amplitude is greater than 0.5. The key parameters of these rules are uniformly stored in a feature threshold matrix, supporting real-time invocation and static feature matching, providing a basic judgment basis for fault diagnosis.

[0034] The deep learning model employs a CNN-LSTM hybrid network structure, using multi-dimensional features as input. These input features encompass the time-domain and frequency-domain features of the vibration signal, as well as auxiliary parameters during unit operation; these features collectively constitute a multimodal feature vector. The model's convolutional layers process the input features through multiple convolutional kernels, extracting key local features and generating high-dimensional feature maps. The LSTM layers further capture the dynamic changes of these features over time, ultimately outputting probability vectors for various fault types.

[0035] The fusion and optimization mechanism combines the fault matching results output by the rule base with the fault probability vector output by the deep learning model through a dynamic weight fusion algorithm to obtain the final fault diagnosis result. Simultaneously, the model possesses adaptive evolution capabilities: based on real-time fault feedback data, it continuously adjusts the parameters in the feature threshold matrix and optimizes the network weights and biases of the deep learning model, enabling the diagnostic rules and model parameters to dynamically adapt to changes in the random group's operating state, continuously improving diagnostic accuracy.

[0036] The confidence level of the rule base of the adaptive diagnostic model submodule for type II faults. It is obtained by weighting feature matching degree and rule priority, and the calculation formula is: ; In the formula: m represents the number of features of the i-th type of fault (e.g., rotor imbalance includes two features: amplitude and phase). α i,k The weight of the k-th feature of the i-th type of fault (set by expert experience) ; xk represents the k-th feature value monitored in real time (such as 1X frequency amplitude). This represents the mean and variance of the k-th feature of the ⅂ type of fault in the rule base (obtained through statistical analysis of historical fault data). The input feature vector X of CNN-LSTM achieves multimodal fusion through normalization and dynamic weight concatenation: ; In the formula: This represents the standardized vibration eigenvector (* indicates normalization). Represents a standardized auxiliary parameter vector (temperature T, rotational speed n); β represents the vibration characteristic weighting coefficient (dynamically adjusted; when the vibration signal-to-noise ratio (SNR) < 10dB, β = 0.3 to enhance the weighting of auxiliary parameters). Final failure probability P final,i The calculation formula is: ; ; In the formula: P DL,i This represents the probability of type II failure output by the deep learning model; γ t This represents the rule base weight coefficient at time t (initial value γ0 = 0.7, adaptively decaying as the deep learning model error accumulates). y j The label represents the actual fault label for the j-th diagnosis (1 indicates occurrence, 0 indicates non-occurrence).

[0037] Based on diagnostic error feedback, the weights W of the deep learning model and the rule base threshold Θ are updated according to the following formula: ;

[0038] In the formula: ηt represents the adaptive learning rate. ; L(·) represents the cross-entropy loss function; λ represents the rule base threshold update step size (λ=0.005); N represents the number of historical diagnostic samples.

[0039] The fault evolution prediction submodule internally includes a historical data training unit, a real-time status assessment unit, a trend prediction model, and a maintenance timing recommendation unit. The historical data training unit extracts full lifecycle data of typical faults such as rotor imbalance, misalignment, and oil film whirl from the fault map library, classifying them into four levels—normal, minor, moderate, and severe—to construct a fault evolution sample set. The real-time status assessment unit calculates the equipment health index by comparing the Euclidean distance between the current fused features and historical samples; when the index falls below a threshold, the prediction process is triggered. The trend prediction model uses an LSTM neural network, taking the fused feature matrix and time-series health index as input, and outputting the fault probability curve and confidence interval for the next 30 days, focusing on capturing the evolution patterns of key indicators such as the 1X frequency amplitude growth rate and the 2X / 1X amplitude ratio change rate. The maintenance timing recommendation unit combines the equipment operation plan and predicted fault time to generate maintenance suggestion plans including the optimal maintenance window, required spare parts models, and estimated working hours. These plans are linked to standardized processes in the preventative maintenance mechanism and can be exported as executable work orders.

[0040] The alarm and graded response module internally includes a fault information receiving unit, a multi-level threshold configuration submodule, a response strategy generation unit, and an alarm log storage unit. The fault information receiving unit receives fault types, characteristic parameters, and graph data output by the diagnostic module and stores them categorized by fault code. The multi-level threshold configuration submodule sets three levels based on historical data from the fault graph library: Level 1 warning (characteristic parameters exceeding normal range), Level 2 alarm (approaching critical value), and Level 3 emergency response (exceeding critical value). The threshold standards are dynamically associated with typical fault characteristics in the graph library. The response strategy generation unit triggers corresponding processes according to the fault level: Level 1 warnings are pushed to employee computer terminals and automatically recorded; Level 2 alarms initiate team-level maintenance preparation; and Level 3 emergency responses trigger shutdown protection signals and notify the technical supervisor. The alarm log storage unit records alarm information, response measures, and processing status in timestamp order, stored in CSV format, and supports retrieval by fault type and time range.

[0041] The fault spectrum knowledge management module internally includes a typical fault spectrum storage unit, a rule base management submodule, a fault case entry interface, and a threshold parameter maintenance unit. The typical fault spectrum storage unit stores spectrum diagrams, waveform diagrams, trend diagrams, and envelope diagrams of typical faults such as rotor imbalance, misalignment, and oil film whirl in a structured format. Each spectrum is associated with a fault name, characteristic parameter range, and probability of occurrence. The rule base management submodule stores diagnostic rules based on spectrum features. Rules are stored in a parsable text format and support editing via a visual interface. The fault case entry interface receives the characteristic parameters, diagnostic results, and maintenance records of new fault cases. After format verification, the cases are stored in the case library. Each case includes four types of spectrum data at the time of the fault occurrence. The threshold parameter maintenance unit periodically adjusts the characteristic threshold matrix based on real-time diagnostic data to ensure dynamic matching with the unit's operating status. The adjustment records include the preceding and following values ​​and their basis.

[0042] The preventative maintenance decision-making module internally comprises a diagnostic conclusion receiving unit, a maintenance plan matching submodule, a resource scheduling and coordination unit, and a maintenance plan generation unit. The diagnostic conclusion receiving unit receives fault type, location, and severity data, converting it into a standardized decision input format. The maintenance plan matching submodule retrieves the corresponding preventative maintenance process from the plan library based on the fault type. The plan includes fault handling steps, required tools, materials, and component models, and is linked to typical cases in the fault diagram library. The resource scheduling and coordination unit interfaces with the equipment management system to query spare parts inventory, storage locations, and maintenance team scheduling, assessing resource availability. The maintenance plan generation unit, combining fault severity and unit operation plan, generates a maintenance plan including start time and responsible personnel, outputting it in Gantt chart format and synchronizing it to the production management system to implement predictive maintenance mechanisms and improve equipment continuous operation cycles.

[0043] From the above, we can conclude that: In this application, the accuracy and adaptability of vibration fault diagnosis for steam turbine generator sets are significantly improved through the synergistic effect of a multimodal fusion adaptive diagnostic module. The module integrates the time-domain and frequency-domain features of vibration signals with auxiliary parameters such as temperature and speed to form a multi-dimensional input matrix. This matrix, combined with a rule base built on a typical fault map library and a CNN-LSTM deep learning network, achieves a dual verification mechanism. The rule base generates quantifiable diagnostic rules based on historical fault data, quickly matching faults with clear characteristics such as rotor imbalance and misalignment. The deep learning model optimizes the diagnostic results through a dynamic weight fusion algorithm, achieving higher accuracy, especially for complex faults such as oil film eddy. The adaptive evolution mechanism updates the rule base thresholds and network parameters based on real-time feedback data, ensuring stable diagnostic capabilities even under scenarios such as equipment aging and changing operating conditions, effectively reducing the subjective errors and missed diagnoses associated with traditional manual judgment.

[0044] In this application, the system promotes a shift in maintenance strategies from passive response to proactive prevention. Through real-time monitoring and intelligent diagnostics, the system can detect fault signs in advance and issue tiered alarms, providing maintenance personnel with sufficient time to handle the situation and avoiding sudden downtime. The fault evolution prediction function, combined with equipment lifecycle data, generates trend analysis reports, supporting the development of precise preventative maintenance plans and reducing resource waste caused by blind maintenance. Simultaneously, the system replaces traditional manual inspection methods, freeing maintenance personnel from repetitive tasks, reducing the impact of human factors on monitoring effectiveness, and improving the standardization of equipment management. Over the long term, the continuous operating cycle of equipment is extended, maintenance costs are reduced, providing solid technical support for the safe and stable operation of power plants, aligning with the digital and intelligent development direction of smart power plant construction.

[0045] This application conducts research on online equipment monitoring and fault diagnosis to predict equipment malfunctions based on operational symptoms, forming an experience database. It proposes predictive maintenance solutions adapted to the actual conditions of power plants, improving employee work efficiency and equipment management quality, training employees to enhance their ability to predict equipment faults, and achieving full control over equipment operation. This extends equipment lifespan, reduces maintenance rates, and lowers maintenance costs; simultaneously, it ensures safe production operation of the unit, avoids temporary shutdowns due to equipment failures, and increases unit uptime. This is of great significance for ensuring the safe, stable, long-term, and high-quality operation of equipment.

[0046] In this application, the development and application of the online equipment fault diagnosis system technology achieves three main benefits: First, it replaces the traditional manual inspection method for monitoring key equipment, ensuring the personal safety of inspection personnel and eliminating the influence of human factors such as personnel experience and dedication on inspection effectiveness. The operating status of equipment no longer depends on on-site human experience judgment, improving the enterprise's equipment management level. Second, it achieves the goals of proactive protection and predictive maintenance, thereby increasing the continuous operation cycle of equipment and reducing downtime for maintenance. This lowers equipment maintenance costs and increases equipment uptime. Third, it ensures production safety and prevents serious accidents.

[0047] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic diagnosis and alarm device for vibration faults in steam turbine generator sets, characterized in that: The device is equipped with an automatic diagnosis and alarm system for vibration faults of steam turbine generator sets. The system includes: a vibration data acquisition module, a feature extraction and signal analysis module, a multimodal fusion adaptive diagnosis module, an alarm and graded response module, a fault map database knowledge management module, and a preventive maintenance decision module. The multimodal fusion adaptive diagnostic module is internally configured with: a multi-source data fusion submodule, an adaptive diagnostic model submodule, and a fault evolution prediction submodule; The data output terminal of the vibration data acquisition module is connected to the raw signal input terminal of the feature extraction and signal analysis module; The feature parameter output of the feature extraction and signal analysis module is connected to the multi-source feature input of the multimodal fusion adaptive diagnosis module; The diagnostic result output terminal of the multimodal fusion adaptive diagnostic module is connected to the fault information input terminal of the alarm and graded response module and the diagnostic conclusion input terminal of the preventive maintenance decision module, respectively. At the same time, its graph query request terminal is bidirectionally connected to the graph data interface of the fault graph library knowledge management module. The maintenance plan output terminal of the fault map knowledge management module is connected to the plan reference input terminal of the preventive maintenance decision module. The alarm log storage terminal of the alarm and graded response module is connected to the log recording input terminal of the fault map library knowledge management module, and the maintenance linkage output terminal is connected to the emergency maintenance request input terminal of the preventive maintenance decision module; the maintenance status feedback terminal of the preventive maintenance decision module is connected to the processing status input terminal of the alarm and graded response module.

2. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The vibration data acquisition module is internally equipped with a sensor deployment layer, a signal conditioning unit, a data preprocessing submodule, and a data transmission interface.

3. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The feature extraction and signal analysis module is internally configured with a time-domain feature extraction unit, a frequency-domain feature analysis unit, a time-frequency domain feature fusion unit, and a spectrum generation submodule. The temporal feature extraction unit calculates indicators such as peak value, kurtosis, and root mean square value, and updates them in real time through sliding window technology. The frequency domain feature analysis unit uses Fast Fourier Transform to generate the spectrum. The time-frequency domain feature fusion unit generates an envelope map through wavelet transform, extracts high-frequency impact features, and analyzes the amplitude change rate in conjunction with the trend map to form a multi-dimensional feature matrix; the spectrum generation submodule transforms the processed feature parameters into four types of visualization maps: spectrum map, waveform map, trend map, and envelope map.

4. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The multi-source data fusion submodule is internally configured with a data access layer, a data cleaning unit, a feature fusion algorithm module, and a fusion result storage unit. The data cleaning unit uses the moving average method to process vibration signal drift based on the noise characteristics of data from different sources, removes outlier values ​​of process parameters using the 3σ criterion, and completes missing data using an interpolation algorithm based on similar working conditions to ensure data integrity.

5. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The adaptive diagnostic model submodule uses rule base prior matching + deep learning dynamic optimization as its core framework to achieve high-precision diagnosis of vibration faults in steam turbine generator sets. The rule base is constructed based on historical fault data, focusing on typical faults such as rotor imbalance, misalignment, and oil film whirl, and transforms their features into quantitative rules that can be directly matched. The deep learning model uses a CNN-LSTM hybrid network structure and takes multi-dimensional features as input; The input features cover the time-domain and frequency-domain features of the vibration signal, as well as auxiliary parameters during unit operation. These features together constitute a multimodal feature vector. The fusion and optimization mechanism combines the fault matching results output by the rule base with the fault probability vector output by the deep learning model through a dynamic weight fusion algorithm to obtain the final fault diagnosis result. At the same time, the model has adaptive evolution capability: based on real-time fault feedback data, it continuously adjusts the parameters in the feature threshold matrix and optimizes the network weights and biases of the deep learning model, so that the diagnostic rules and model parameters can dynamically adapt to changes in the random group's operating state, continuously improving diagnostic accuracy.

6. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The rule base of the adaptive diagnostic model submodule has a confidence level for type II faults. It is obtained by weighting feature matching degree and rule priority, and the calculation formula is: ； In the formula: m represents the number of features of the i-th type of fault (e.g., rotor imbalance includes two features: amplitude and phase). α i,k The weight of the k-th feature of the i-th type of fault is represented (set by expert experience). ; xk represents the k-th feature value monitored in real time; This represents the mean and variance of the kk-th feature of the ii-th fault in the rule base; The input feature vector X of CNN-LSTM achieves multimodal fusion through normalization and dynamic weight concatenation: ； In the formula: This represents the standardized vibration eigenvector (* indicates normalization). Represents a standardized auxiliary parameter vector (temperature T, rotational speed n); β represents the vibration characteristic weighting coefficient (dynamically adjusted; when the vibration signal-to-noise ratio (SNR) < 10dB, β = 0.3 to enhance the weighting of auxiliary parameters). Final failure probability P final,i The calculation formula is: ； ； In the formula: P DL,i This represents the probability of type II failure output by the deep learning model; This represents the rule base weight coefficient at time t (initial value γ0 = 0.7, adaptively decaying as the deep learning model error accumulates). y j The actual fault label for the j-th diagnosis (1 indicates occurrence, 0 indicates non-occurrence); Based on diagnostic error feedback, the weights W of the deep learning model and the rule base threshold Θ are updated according to the following formula: ； In the formula: ηt represents the adaptive learning rate. ; L(·) represents the cross-entropy loss function; This indicates the rule base threshold update step size (λ=0.005). N represents the number of historical diagnostic samples.

7. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The fault evolution prediction submodule is internally equipped with a historical data training unit, a real-time status assessment unit, a trend prediction model, and a maintenance timing recommendation unit. The historical data training unit extracts the full life cycle data of rotor imbalance, misalignment, and oil film whirl faults from the fault map library, and divides them into four levels according to the severity of the fault: normal, minor, moderate, and severe, to construct a fault evolution sample set. The real-time status assessment unit calculates the device health index by comparing the current fused features with the Euclidean distance of historical samples. When the index is lower than the threshold, the prediction process is triggered.

8. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The alarm and graded response module is internally equipped with a fault information receiving unit, a multi-level threshold configuration submodule, a response strategy generation unit, and an alarm log storage unit. The fault information receiving unit receives the fault type, characteristic parameters, and graph data output by the diagnostic module and stores them according to the fault code. The alarm log storage unit records alarm information, response measures, and processing status in timestamp order, stores them in CSV format, and supports retrieval by fault type and time range.

9. The automatic diagnosis and alarm device for vibration faults of steam turbine generator sets as described in claim 1, characterized in that: The fault map knowledge management module is internally equipped with a typical fault map storage unit, a rule base management sub-module, a fault case input interface, and a threshold parameter maintenance unit. The typical fault spectrum storage unit stores the spectrum, waveform, trend and envelope diagrams of rotor imbalance, misalignment and oil film whirl in a structured form. Each spectrum is associated with the fault name, characteristic parameter range and occurrence probability.

10. The automatic diagnosis and alarm device for vibration faults of a steam turbine generator set as described in claim 1, characterized in that: The preventive maintenance decision-making module is internally equipped with a diagnostic conclusion receiving unit, a maintenance plan matching submodule, a resource scheduling and coordination unit, and a maintenance plan generation unit; the diagnostic conclusion receiving unit receives fault type, location, and severity data, and converts them into a standardized decision input format; The maintenance plan matching submodule retrieves the corresponding preventive maintenance process from the plan library based on the fault type. The plan includes fault handling steps, required tools, materials, and component models, and is associated with typical cases in the fault diagram library. The resource scheduling and coordination unit connects to the equipment management system to query spare parts inventory, storage location, and maintenance team scheduling, and assesses resource availability. The maintenance plan generation unit combines the severity of the fault and the unit's operation plan to formulate a maintenance plan that includes start time and responsible person, outputs it in Gantt chart format, and synchronizes it to the production management system to implement the predictive maintenance mechanism and improve the continuous operation cycle of equipment.

Images