Power plant equipment intelligent diagnosis method and system based on man-machine experience fusion

Abstract

The invention provides a power plant equipment intelligent diagnosis method and system based on man-machine experience fusion, and belongs to the technical field of fault diagnosis. The method comprises the following steps: inputting equipment operation data into a fault diagnosis knowledge base, executing fault symptom feature matching, and determining weights of matched fault symptom features under corresponding fault types; inputting the equipment operation data and the weights of the matched fault symptom features under the corresponding fault types into the intelligent diagnosis model to obtain a fault diagnosis result; and the fault diagnosis result is obtained based on fusion of fault occurrence probability prediction results of the machine learning diagnosis unit and the expert rule diagnosis unit on the fault types of the matched fault symptom features. Intelligent diagnosis of the power plant equipment is realized by fusing man-machine experience, the accuracy, the real-time performance and the reliability of fault diagnosis of the gas turbine power plant equipment are improved, early warning and accurate diagnosis of equipment faults are realized, non-planned shutdown is reduced, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to an intelligent diagnostic method for power plant equipment based on human-machine experience fusion and an intelligent diagnostic system for power plant equipment based on human-machine experience fusion. Background Technology

[0002] Currently, gas turbine power plants are complex, comprising various main and auxiliary equipment such as gas turbines, steam turbines, boilers, and electrical systems, generating a large amount of heterogeneous data from multiple sources during operation. Traditional equipment fault diagnosis methods, such as those based on analytical models, rely on precise mathematical models. However, gas turbine power plant systems are complex, making it difficult to establish accurate models, and their ability to handle situations not covered by the models is limited. Methods based on single data sources, such as relying solely on vibration signals or temperature data, are prone to missed or misdiagnosed cases when early fault characteristics are not obvious. Furthermore, pure expert systems have low diagnostic efficiency and cannot meet the needs of real-time monitoring and rapid decision-making.

[0003] Traditional maintenance methods at present include: 1. Post-failure maintenance: Also known as fault diagnosis, this method involves maintenance only after a failure has occurred. This approach relies on on-site technical personnel to obtain the current status information of the operating equipment and diagnose recurring functional faults. Therefore, it only addresses issues after they occur and cannot guarantee maintenance during operation. Consequently, this "repair after it breaks" approach cannot ensure the healthy operation of gas turbine equipment. Even minor equipment malfunctions can affect the normal operation of critical components or the entire gas turbine system, leading to significant economic losses.

[0004] 2. Time-Based Maintenance (TBM): Also known as scheduled maintenance (SM), this method analyzes and statistically summarizes the operating performance and historical trends of gas turbine equipment. Based on long-term statistical analysis, a reasonable normal operation and maintenance cycle for the gas turbine equipment can be set to achieve regular preventive maintenance. Currently, TBM remains the mainstream maintenance method. It can directly prevent or delay the occurrence of related faults in ensuring the normal operation of major rotating machinery, further reducing the failure rate of gas turbine equipment and thus reducing the cost of gas turbine equipment downtime. Although TBM can prevent equipment failures to a certain extent, its maintenance costs are relatively high, and it requires extensive expert experience, leading to the serious drawback of over-maintenance. Therefore, TBM does not fundamentally solve the problem.

[0005] Therefore, improving the accuracy, real-time performance, and reliability of fault diagnosis in gas turbine power plants is an urgent problem to be solved in order to achieve early warning and accurate diagnosis of equipment faults, reduce unplanned downtime, and lower operation and maintenance costs. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for intelligent diagnosis of power plant equipment based on human-machine experience fusion, so as to at least solve the above-mentioned problems of how to improve the accuracy, real-time performance and reliability of fault diagnosis of gas turbine power plant equipment.

[0007] To achieve the above objectives, the first aspect of the present invention provides an intelligent diagnostic method for power plant equipment based on human-machine experience fusion. The method includes: inputting equipment operation data of a target gas turbine power plant within a target period into a fault diagnosis knowledge base; sequentially performing fault symptom feature matching corresponding to the changing trends of the equipment operation data within the target period and determining the weights of the matched fault symptom features under each corresponding fault type; inputting the equipment operation data within the target period and the weights of the matched fault symptom features under each corresponding fault type into an intelligent diagnostic model containing a machine learning diagnostic unit and an expert rule diagnostic unit to obtain a fault diagnosis result; wherein the fault diagnosis result is obtained by fusing the results of the machine learning diagnostic unit and the expert rule diagnostic unit predicting the probability of fault occurrence for each fault type of the matched fault symptom features.

[0008] A second aspect of this invention provides an intelligent diagnostic system for power plant equipment based on human-machine experience fusion. The system includes: a fault symptom feature matching module, used to input equipment operation data of a target gas turbine power plant within a target period into a fault diagnosis knowledge base, and sequentially perform fault symptom feature matching corresponding to the changing trends of the equipment operation data within the target period, and determine the weights of the matched fault symptom features under each corresponding fault type; and a fault diagnosis module, used to input the equipment operation data within the target period and the weights of the matched fault symptom features under each corresponding fault type into an intelligent diagnostic model containing a machine learning diagnostic unit and an expert rule diagnostic unit to obtain fault diagnosis results; wherein the fault diagnosis results are obtained by fusing the results of the machine learning diagnostic unit and the expert rule diagnostic unit predicting the probability of fault occurrence for each fault type of the matched fault symptom features.

[0009] The above technical solution provides a method and system for intelligent diagnosis of power plant equipment based on human-machine experience fusion. It achieves fault symptom feature matching by comparing the changing trends of equipment operating data of a target gas turbine power plant within a target period with fault symptom features in a fault diagnosis knowledge base used to characterize different data changing trends. The system then obtains the weights of the matched fault symptom features under the corresponding fault types from the fault diagnosis knowledge base. The machine learning diagnostic unit in the intelligent diagnostic model processes the input equipment operating data within the target period and the weights of the matched fault symptom features under each corresponding fault type from the perspective of machine learning algorithms, outputting the fault type determined by the machine learning algorithm. Furthermore, the expert rule diagnostic unit in the intelligent diagnostic model infers the weights of the input equipment operating data within the target period and the matched fault symptom features under each corresponding fault type from the perspective of a logical rule base established by domain experts, outputting the determined fault type. By fusing the fault types determined by the machine learning diagnostic unit and the expert rule diagnostic unit, a comprehensive fault diagnosis result is formed. This method and system integrate human-machine experience to achieve intelligent diagnosis of power plant equipment operation. It uses machine learning algorithms to process equipment operation data and combines the rich experience of domain experts to improve the accuracy, real-time performance, and reliability of fault diagnosis for gas turbine power plant equipment. This enables early warning and accurate diagnosis of equipment faults, reduces unplanned downtime, and lowers operation and maintenance costs.

[0010] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0011] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of an intelligent diagnostic method for power plant equipment based on human-machine experience fusion, provided by one embodiment of the present invention. Figure 2 This is a schematic diagram of a data acquisition device provided in one embodiment of the present invention; Figure 3 This is a schematic diagram of the fault analysis of compressor surge in a target gas turbine power plant provided by one embodiment of the present invention; Figure 4 This is a block diagram of an intelligent diagnostic system for power plant equipment based on human-machine experience fusion, provided by one embodiment of the present invention. Detailed Implementation

[0012] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0013] It should be noted that, in this application, human-computer experience refers to the actions of combining machine learning algorithms with the user's actual experience to make judgments about the target object.

[0014] Figure 1 This is a flowchart of an intelligent diagnostic method for power plant equipment based on human-machine experience fusion, provided by one embodiment of the present invention. Figure 1 As shown, this invention provides an intelligent diagnostic method for power plant equipment based on human-machine experience fusion, the method comprising: S110: Input the equipment operation data of the target gas turbine power plant within the target period into the fault diagnosis knowledge base, and perform fault symptom feature matching with the change trend of equipment operation data within the target period one by one, and determine the weight of the matched fault symptom features under each corresponding fault type. Among them, the target gas turbine power plant can be the 9H gas turbine power plant.

[0015] The fault diagnosis knowledge base needs to pre-store fault symptom features corresponding to different fault types, as well as the weight data of each fault symptom feature under the corresponding fault type. The trend of equipment operation data is matched with the preset feature templates containing fault symptom features in the fault diagnosis knowledge base to achieve fault symptom feature matching.

[0016] S120: Input the equipment operation data within the target period and the weights of the matched fault symptom features under each corresponding fault type into the intelligent diagnostic model containing the machine learning diagnostic unit and the expert rule diagnostic unit to obtain the fault diagnosis result; wherein, the fault diagnosis result is obtained by fusing the results of the machine learning diagnostic unit and the expert rule diagnostic unit in predicting the probability of fault occurrence for each fault type of the matched fault symptom features.

[0017] Specifically, this method achieves fault symptom feature matching by comparing the changing trends of equipment operating data of the target gas turbine power plant within the target period with fault symptom features in the fault diagnosis knowledge base used to characterize different data changing trends. The weights of the matched fault symptom features under the corresponding fault types are then obtained from the fault diagnosis knowledge base. The machine learning diagnostic unit in the intelligent diagnostic model processes the weights of the input equipment operating data within the target period and the matched fault symptom features under each corresponding fault type from the perspective of machine learning algorithms, outputting the fault type determined by the machine learning algorithm. Furthermore, the expert rule diagnostic unit in the intelligent diagnostic model infers the weights of the input equipment operating data within the target period and the matched fault symptom features under each corresponding fault type from the perspective of a logical rule base established by domain experts, outputting the determined fault type. By fusing the fault types determined by the machine learning diagnostic unit and the expert rule diagnostic unit, a comprehensive fault diagnosis result is formed. This method integrates human and machine experience to achieve intelligent diagnosis of power plant equipment operation. It uses machine learning algorithms to process equipment operation data and combines the rich experience of domain experts to improve the accuracy, real-time performance, and reliability of fault diagnosis for gas turbine power plant equipment. This enables early warning and accurate diagnosis of equipment faults, reduces unplanned downtime, and lowers operation and maintenance costs.

[0018] In some embodiments of this example, the process of fusing the fault types determined by the machine learning diagnostic unit and the fault types determined by the expert rule diagnostic unit can adjust the weight ratio of the two units according to the actual application scenario to improve diagnostic accuracy.

[0019] In some embodiments of this example, the rules for collecting the above-mentioned equipment operation data include: determining the equipment operation data collection scheme for each collection system based on the joint operation process of the target gas turbine power plant, and executing the equipment operation data collection of each collection system based on the determined equipment operation data collection scheme; wherein, the collection system includes a DCS system, a SIS system, and an auxiliary system; Specifically, data acquisition requires comprehensive and highly reliable real-time access. Data acquisition needs to cover the entire combined cycle process of the target gas turbine power plant, from gas turbine to waste heat boiler to steam turbine generator, focusing on solving three core issues: "multi-system protocol compatibility," "high-frequency capture of key parameters," and "no data loss."

[0020] The equipment operation data acquisition scheme includes the acquisition priority and parameter dimensions of the corresponding acquisition systems to ensure that no key parameters are missed and no secondary parameters are redundant. The acquisition priority and parameter dimensions of each acquisition system are shown in Table 1: Table 1. Acquisition Priority and Parameter Dimensions of Each Acquisition System

[0021] Based on the transmission protocol of each acquisition system, corresponding data access rules are matched, and based on the matched data access rules, the equipment operation data collected by each acquisition system is obtained through the edge gateway deployed at the target gas turbine power plant.

[0022] Specifically, the data acquisition scheme in this application achieves protocol compatibility and edge access. More specifically, considering the heterogeneity of multi-system protocols in the target gas turbine power plant (DCS mostly uses vendor-specific protocols, while SIS mostly uses standard industrial protocols), an "edge gateway + standardized protocol" architecture is adopted to achieve unified cross-system data access.

[0023] The specific deployment methods for the protocol adaptation layer are as follows: 1. DCS access: Private protocols are parsed through vendor-authorized interfaces (such as Profinet for Siemens DCS and OvationLink for Emerson Ovation) or edge gateways to avoid security risks caused by directly cracking the DCS core program; 2. SIS access: The OPCUA protocol (Industrial Internet Standard) is used first to directly read the real-time database of SIS (such as PI and eDNA) and support the synchronous transmission of data context (such as parameter units, alarm thresholds, and device numbers); 3. Auxiliary device access: Access is supplemented through ModbusTCP (such as lubricating oil pumps) and 4G / 5G (sensors in remote areas) to achieve "no dead zones" coverage.

[0024] Furthermore, this method also includes edge computing gateway deployment. Specifically, industrial-grade edge gateways (such as Advantech and Huawei edge computing boxes) are deployed in the control building of the target gas turbine power plant to collect DCS / SIS data locally, reducing latency and bandwidth consumption in direct transmission between "field devices and the cloud". The gateway supports "local caching + breakpoint resumption": when the network is interrupted (such as a power plant intranet failure), data is cached locally for 1-7 days, and automatically retransmitted after recovery to ensure no data loss.

[0025] In some implementations of this embodiment, the method also includes a time synchronization function, specifically including: all acquisition devices (gateway, DCS controller, SIS server) are synchronized with the time through an NTP server (accuracy ≤ 1ms) to solve the problem of "inconsistent timestamps for different parameters of the same event" (e.g., when the gas turbine trips, the timestamps of T3 temperature and speed need to be fully synchronized).

[0026] In some implementations of this embodiment, the method also includes a data acquisition guarantee mechanism, as follows: 1. Redundant acquisition: Key parameters (such as T3 temperature and shaft vibration) are acquired using a "dual gateway + dual link" approach. If any gateway or link fails, the method automatically switches to the backup channel, with an acquisition interruption time of ≤100ms; 2. Acquisition status monitoring: The acquisition success rate of each parameter is monitored in real time (target ≥99.99%). When a parameter acquisition fails (such as sensor failure), an alarm is immediately triggered (pushed to the operation and maintenance platform), and the data status is marked as "failed".

[0027] Please refer to Figure 2 , Figure 2 This is a schematic diagram of a data acquisition device according to one embodiment of the present invention. It collects data in real time from the DCS system, SIS system, and auxiliary systems of the target gas turbine power plant, including multi-dimensional data such as temperature, pressure, vibration, and current. The data undergoes preprocessing such as cleaning and filtering to remove noise and outliers.

[0028] In some embodiments of this example, after collecting equipment operation data, the method further includes: performing one or more preprocessing steps on the collected equipment operation data, namely data cleaning, noise filtering, data standardization, and normalization, and performing preprocessing quality verification on the preprocessed equipment operation data, including: performing data cleaning on the corresponding equipment operation data based on the data cleaning rules matched to the business logic of the target gas turbine power plant; performing noise filtering on the corresponding equipment operation data based on the noise type matched to the noise filtering rules in the equipment operation data; and performing integrity verification, consistency verification, and data change trend verification before / after preprocessing on each piece of preprocessed equipment operation data.

[0029] Specifically, data preprocessing enables precise noise reduction and anomaly repair, ensuring data usability. Data from target gas turbine power plants is affected by electromagnetic interference (such as strong electromagnetic environments of generators), equipment fluctuations (such as instantaneous changes in fuel pressure), and sensor malfunctions (such as temperature sensor drift). Therefore, multi-step preprocessing is required to transform the raw data into usable data. The core objectives are: noise removal without disrupting trends, anomaly repair without deviating from reality, and a standardized format for easy analysis. The specific process is as follows: 1. First step: Data cleaning to solve the "dirty data" problem: Data cleaning rules are formulated based on the business logic of gas turbine power plants, addressing common issues such as missing values, outliers, and duplicate values ​​in the collected data. 2. Second step: Noise filtering to eliminate interference and restore the true trend: High-frequency parameters of the target gas turbine power plant (such as shaft vibration and T3 temperature) are susceptible to electromagnetic interference, generating "high-frequency noise." Appropriate filtering algorithms need to be selected based on the noise type. The core principle is to "retain useful signals (such as the fault characteristic frequency of shaft vibration) and filter out useless noise." The matching table for filtering algorithm processing methods is shown in Table 2. 3. Third step: Data standardization and normalization to unify the format and facilitate cross-dimensional analysis: The dimensions and ranges of different parameters vary greatly (such as temperature in °C, pressure in MPa, and current in A). Standardization / normalization is required to transform the data into a unified range, laying the foundation for subsequent modeling (such as unit performance calculation and fault prediction). Standardization (Z-Score standardization) is suitable for scenarios where the original distribution characteristics of parameters need to be preserved (such as unit thermal efficiency calculation). The formula is: Z = σX μ (where X is the original data, μ is the historical mean of the parameter, and σ is the standard deviation). Example: After standardizing the T3 temperature (mean 850℃, standard deviation 20℃), 870℃ corresponds to Z=1, and 830℃ corresponds to Z=-1. Normalization (Min-Max normalization): suitable for scenarios where data needs to be compressed to a fixed range (e.g., 0-1) (e.g., machine learning model input), the formula is: Xnorm = (X... (Xmin) / (Xmax) Xmin (where Xmin / Xmax are the historical minimum / maximum values ​​of the parameter X). Example: After normalization of the gas turbine speed (range 2950-3050rpm), 3000rpm corresponds to 0.5, and 3050rpm corresponds to 1. 4. Fourth step: Preprocessing quality verification to ensure that the preprocessed data is "usable and reliable": After preprocessing, it needs to pass the triple verification of "completeness, consistency, and rationality" to avoid introducing new errors in the preprocessing process, as follows: Completeness verification: Statistically calculate the effectiveness of the preprocessed data (target ≥99.9%), and the proportion of missing / abnormal data should be <0.1%; Consistency verification: Based on the 9H gas turbine operation logic, verify the correlation between parameters (e.g., when the gas turbine power increases, the fuel flow should increase synchronously. If "power increases and flow decreases" occurs, it is determined that the preprocessing is abnormal); Trend verification: Compare the data trends before and after preprocessing (e.g., during the gas turbine startup process, the T3 temperature should gradually increase) to ensure that filtering / cleaning does not destroy the true trend (e.g., avoid excessive filtering that causes "temperature rise trend to become flat").

[0030] Table 2 Matching Table of Filtering Algorithm Processing Methods

[0031] In some implementations of this embodiment, the core applications of the preprocessed data are as follows: 1. The preprocessed "clean data" will directly support the core business scenarios of the target gas turbine power plant; 2. Real-time monitoring: Pushed to the central control room screen to display the real-time operating status of the unit (such as dynamic curves of key parameters such as T3 temperature and shaft vibration), triggering audible and visual alarms when limits are exceeded; 3. Performance analysis: Calculate the unit's thermal efficiency and gas consumption rate (based on accurate temperature, pressure, and flow data), identify performance degradation trends (such as a 0.1% monthly decrease in thermal efficiency, requiring investigation of compressor fouling); 4. Fault prediction and health management (PHM): Provide input for AI models (such as neural networks and random forests) to predict equipment failures (such as predicting bearing wear based on shaft vibration and bearing temperature data, issuing maintenance recommendations 30 days in advance); 5. Historical data tracing: Stored in a time-series database (such as InfluxDB and TimescaleDB) for accident review (such as retrieving the complete preprocessed data chain to analyze the cause after a gas turbine trip).

[0032] In some embodiments of this example, the construction rules of the above-mentioned fault diagnosis knowledge base include: based on production rules, using pre-collected expert experience data and historical fault data to identify fault symptom features under each fault type and determine the weights of the identified fault symptom features under each corresponding fault type, so as to comprehensively determine the weights of each fault symptom feature under each corresponding fault type and establish a fault diagnosis knowledge base; wherein, the rules for determining the weights of each fault symptom feature under each corresponding fault type include: comparing the importance of fault symptom features under the same fault type pairwise, so as to establish a corresponding fault diagnosis knowledge base based on the pairwise comparison results. A fault type judgment matrix is ​​established. Based on the established judgment matrix, initial weights are assigned to each fault symptom feature under the same fault type, and a consistency check is performed on the assigned fault symptom features. Based on the statistical weights of the fault symptom features under each corresponding fault type, determined by the support and confidence between the fault symptom features and each corresponding fault type, the weights of the fault symptom features after the consistency check are calibrated to obtain the final weights of the fault symptom features under each corresponding fault type. The support and confidence between the fault symptom features and each corresponding fault type are obtained based on the statistical analysis of pre-collected historical fault data.

[0033] Specifically, firstly, expert experience and historical fault data are collected, and a fault diagnosis knowledge base is established using production rules and other methods. Fault symptom characteristics are categorized into types such as constant, slow increase, sudden increase, slow decrease, and sudden decrease, and the weight of each fault symptom characteristic is clarified. The specific process of establishing the fault diagnosis knowledge base is as follows: 1. Determine the knowledge base construction goals and principles: The core goals include: achieving accurate mapping between fault symptoms and fault types, quantifying the degree of influence of different symptoms on fault judgment, supporting the interpretability and maintainability of diagnostic logic, and providing an efficient rule matching basis for real-time diagnosis; The basic principles include: Systematicity: covering typical and potential faults throughout the entire life cycle of equipment; Accuracy: based on authoritative expert experience and real fault cases; Operability: clear rule definitions, easy for computer parsing and human understanding; Dynamism: supporting continuous iterative optimization of rules and weights; 2. Determine the knowledge sources and collection methods: (1) Expert experience collection: systematically collect domain expert knowledge through structured interviews, fault analysis workshops, Delphi method, etc. The collection targets include equipment design experts, operation and maintenance engineers, maintenance technicians, manufacturer technical support, etc. The collection content includes: the characteristic symptoms and changing trends of typical faults, the strength of the causal relationship between faults and symptoms, the identification points of similar faults, and the fault manifestations under special working conditions. (2) Historical fault data mining: Based on the equipment operation database (such as SIS, DCS system) and fault record system, knowledge is extracted. The data range includes fault records, operating parameter curves and maintenance reports for the past 5 years. Mining methods include: time series data analysis: identifying the parameter change patterns before and after the fault occurs; association rule mining: discovering the implicit association between symptom combinations and fault types; statistical analysis: calculating the frequency of each symptom and its contribution to the fault; 3. Using knowledge representation methods for knowledge representation: production rules are used as the main knowledge representation form, and the basic structure is: Plaintext: IF [Sign 1 (Trend Type)] AND [Sign 2 (Trend Type)] AND ... AND [Sign n (Trend Type)] AND NOT [Negative Sign]; THEN [Fault Type], Confidence Level = [0-100%]; Here are some examples of rules: Plaintext: Rule ID: R-001: IF Turbine outlet temperature drops sharply (>100°C within 10 seconds) AND fuel pressure drops sharply (>0.5MPa within 5 seconds) AND unit speed drops sharply (>300rpm within 5 seconds) AND NOT ignition system fault signal (unchanged); THEN Fault Type = Combustion Chamber Flameout, Confidence Level = 95%.

[0034] 4. Establish a fault symptom classification system: Fault symptoms are divided into five categories based on parameter change trends, and clear quantitative standards are established for each category. The fault symptom classification system is shown in Table 3. Table 3 Classification System of Fault Symptoms

[0035] 5. Method for determining the weight of symptoms: The weight of symptoms is determined by combining the Analytic Hierarchy Process (AHP) with data statistics. The steps are as follows: (1) Establish a judgment matrix: 5-7 experts compare the importance of each symptom under the same fault pairwise and score them using the 1-9 scale: 1 point means that the two symptoms are equally important; 3 points means that the former is slightly more important than the latter; 5 points means that the former is significantly more important than the latter; 7 points means that the former is strongly more important than the latter; 9 points means that the former is extremely more important than the latter; even numbers indicate that the importance is between adjacent odd numbers. (2) Calculate the weight value: Calculate the initial weight of each symptom through matrix operation and perform a consistency test (CR<0.1); (3) Data calibration: Combine historical fault data to calculate the support and confidence of each symptom. Support represents the probability that a certain symptom and a specific fault occur at the same time; confidence represents the probability that a specific fault occurs when the symptom occurs; (4) Obtain the final weight = expert weight × 0.6 + data statistics weight × 0.4. Table 4 provides an example of the weight of bearing wear fault: Table 4. Examples of Weights for Bearing Wear Failures

[0036] 6. Knowledge Base Structure Design: A three-layer architecture is adopted to achieve structured storage of knowledge: 1) Fault Information Layer, including: Basic Fault Information: Fault ID, Name, System, Severity, Scope of Impact; Fault Description: Fault Phenomenon, Mechanism, Typical Consequences; Handling Suggestions: Emergency Measures, Maintenance Plan, Preventive Measures; 2) Symptom Information Layer, including: Basic Symptom Information: Symptom ID, Name, Associated Parameters, Equipment Location; Feature Description: Data Type, Unit, Normal Range, Trend Type; Detection Method: Acquisition Frequency, Detection Accuracy, Sensor Location; 3) Rule Information Layer, including: Basic Rule Information: Rule ID, Associated Fault, Credibility, Effective Status; Prerequisites: Symptom Combination, Logical Relationship, Weight Allocation; Conclusion Information: Fault Type, Probability of Occurrence, Diagnostic Suggestions. 7. Knowledge base management and maintenance, including: 1) Knowledge verification mechanism, including: new rules must undergo expert review (logical rationality), historical data verification (accuracy ≥ 85%), and on-site testing (3-5 case verifications) before being added to the knowledge base; 2) Dynamic update mechanism, including: regular updates: rules are updated quarterly based on new fault cases; triggered updates: emergency updates are initiated when the diagnostic accuracy is below the threshold (< 80%); version management: historical versions of the knowledge base are retained, and retrospection is supported; 3) Knowledge base optimization, including: rule merging: merging duplicate or highly similar rules; weight adjustment: weights are calibrated regularly based on new operational data; redundancy cleanup: invalid rules that have not been matched for a long time are deleted.

[0037] In some embodiments of this example, the collection rules for the aforementioned historical fault data include: obtaining fault data within a historical time period from the equipment operation database of the target gas turbine power plant; and, based on the obtained fault data, performing corresponding fault data analysis according to multiple data mining rules to extract historical fault data used to characterize the correlation between fault symptom features and fault types. The data mining rules include at least one or more of the following: time-series data analysis rules for identifying patterns of equipment operation data changes before / after a fault occurs; association mining rules for mining the correlation between combined features of fault symptom features and fault types; and statistical analysis rules for determining the frequency of occurrence and contribution of each fault symptom feature to the fault.

[0038] In some embodiments of this example, the rules for obtaining the above-mentioned fault diagnosis results include: comparing the confidence levels of the fault type prediction results output by the machine learning diagnosis unit and the fault type prediction results output by the expert rule diagnosis unit; when the confidence levels of the fault type prediction results output by both are higher than a preset confidence threshold and the fault types in the fault type prediction results output by both are consistent, the output fault type is determined as the fault diagnosis result; otherwise, the equipment operation data within the target period and the fault type prediction results output by both are input to the expert diagnosis workbench, and the fault diagnosis results fed back by the expert diagnosis workbench are received.

[0039] In some implementations of this embodiment, the rules for predicting the probability of failure for each fault type based on the matched fault symptom features by the machine learning diagnostic unit include: using a CNN-LSTM hybrid network to extract latent features from the equipment operation data within the target period, and using a Transformer model with attention mechanism to match fault types based on the extracted latent features, outputting the fault type and corresponding confidence level; the rules for predicting the probability of failure for each fault type based on the matched fault symptom features by the expert rule diagnostic unit include: extracting explicit features from the equipment operation data within the target period based on predefined key explicit features, and inputting the extracted explicit features into the fault diagnosis knowledge base to match fault types, outputting the fault type and corresponding confidence level; wherein, the key explicit features are defined based on the operating mechanism of the target gas turbine power plant.

[0040] Specifically, intelligent diagnostic models combine machine learning algorithms and expert rules. For example, they use deep neural networks to extract features and recognize patterns from data, while expert rules are used to verify and supplement the results of the neural networks. For some complex fault scenarios, human-computer interaction allows experts to correct and improve the diagnostic results based on their experience.

[0041] Furthermore, the overall architecture of the intelligent diagnostic model is a three-layer fusion diagnostic logic. The intelligent diagnostic model adopts a progressive architecture of "feature extraction - dual-track diagnosis - result fusion", which takes into account both the depth of data mining and the reliability of expert experience, as follows: Input layer: preprocessed data (time series parameters such as temperature, pressure, vibration, etc. + trend labels); Feature extraction layer: deep neural network (automatically extracts latent features) + expert feature engineering (extracts explicit features); Dual-track diagnosis layer: machine learning branch: deep neural network (outputs fault type + confidence level), and expert rule branch: rule engine (outputs fault type + matching degree); Result fusion layer: high confidence results (direct output) and low confidence / conflict results (trigger human-computer interaction); Output layer: final diagnostic results (including fault location, cause analysis, and handling suggestions); Feedback iteration: diagnostic results + expert correction → update model parameters and rule base.

[0042] Furthermore, the core components of the intelligent diagnostic model are designed and technically implemented as follows: 1. Feature extraction layer: dual feature complementarity, specifically including: (1) Deep neural network automatic feature extraction: for high-frequency time series data of 9H gas turbine (such as shaft vibration at the 100ms level, T3 temperature), CNN-LSTM hybrid network is used to extract latent features. CNN module: captures local mutation features (such as instantaneous spikes of shaft vibration, sudden drops in pressure) through 1D convolution kernel (size 5-10) and filters noise interference; LSTM module: captures long-term time-dependent features (such as the slow increase trend of bearing temperature in 2 hours, the gradual degradation of compressor efficiency) through gating mechanism; output: 256-dimensional feature vector, including the time domain, frequency domain and trend correlation features of parameters (such as "the coupling relationship between the slow increase of T3 temperature and the slow increase of fuel flow"). (2) Expert feature engineering to extract explicit features: combined with the operating mechanism of 9H gas turbine, key explicit features are manually defined and complemented by neural network features. Trend features: such as “shaft vibration increases slowly by 8μm in the X direction within 1 hour” and “fuel pressure drops sharply by 0.6MPa within 5 seconds” (based on pre-processed trend labels); Correlation features: such as “turbine outlet temperature deviation > 5℃ and steam pressure deviation > 0.2MPa” (reflecting uneven combustion); Threshold features: such as “shaft vibration > 30μm (alarm value)” and “battery temperature > 95℃ (warning value)”. 2. Dual-track diagnostic layer: collaborative reasoning of data and knowledge, specifically including: (1) Machine learning branch: deep neural network diagnosis. Based on the 256-dimensional features of the feature extraction layer, the Transformer model with attention mechanism is used for fault classification. Attention mechanism: focus on key features (such as automatically increasing the weight of the feature “sudden drop in outlet pressure” when diagnosing “compressor surge”); Output: fault type (such as “combustion chamber flameout” and “bearing wear”) and confidence (0-100%), supporting concurrent diagnosis of multiple faults (such as simultaneously outputting “fuel valve jam (85%) + ignition system abnormality (60%)”). Advantages: It is good at handling complex faults with "multi-parameter coupling and ambiguous symptoms" (such as the combined cycle mode of waste heat boiler and gas turbine collaborative faults), and can identify hidden fault modes that have not been clearly summarized by experts. (2) Expert rule branch: rule engine verification and supplementation. Based on the constructed fault diagnosis knowledge base, the Drools rule engine is used to perform expert rule reasoning. Reasoning logic: Match real-time features with the "IF-THEN" rules in the rule base (such as "IFT3 sudden drop and fuel pressure sudden drop THEN combustion chamber extinguishing"); Output: the matched fault type and rule matching degree (0-100%, calculated based on the weight of symptoms). Core function: Verify the neural network results, such as the neural network diagnosis of "bearing wear", the rule engine checks whether it meets "shaft vibration slow increase + bearing temperature slow increase", if it does not meet the requirements, it marks the conflict; Supplement small sample faults: For rare faults with insufficient historical data (such as "fuel nozzle blockage"), the expert rule output results are directly relied upon. 3. Result fusion layer: dynamic decision-making and conflict resolution.The final diagnostic results are output using a fusion strategy of confidence weighting and rule verification, as follows: (1) High confidence consistent results (direct output): When the fault types of the dual-track outputs are consistent, and the neural network confidence is ≥80% and the rule matching degree is ≥75%, the results are output directly. For example, the diagnostic result is slight bearing wear, based on the following: the neural network output confidence is 88% (based on the trend characteristics of shaft vibration and bearing temperature), and the expert rule output matching degree is 82% (shaft vibration gradual increase weight 0.63 + bearing temperature gradual increase weight 0.42). (2) Low confidence / conflict results (triggering human-computer interaction): When the following situations occur, the results are automatically pushed to the expert diagnostic interface: ① The confidence level of the neural network is less than 70% (the model is uncertain); ② The fault type of the dual-track output is conflicting (such as the network diagnosis of "compressor surge" and the rule diagnosis of "combustion chamber flameout"); ③ The rule matching degree is less than 60% and the network confidence level is 70%-80% (manual confirmation is required). 4. Human-computer interaction layer: expert intervention and knowledge feedback: For cases that trigger interaction, an expert diagnosis workbench is designed to realize the closed loop of "manual correction - result interpretation - knowledge accumulation", which includes: (1) expert intervention process: First, the system displays the parameter curves at the time of the fault (such as the shaft vibration time domain diagram, the T3 temperature change trend), the dual-track diagnosis results and basis, and historical similar cases; then the expert operates, including: correcting the fault type (such as correcting "unknown fault" to "fuel nozzle blockage"), supplementing new symptoms (such as "flame detector signal abnormality" not being captured by the model), adjusting the weight (such as increasing the weight of "sudden increase in nozzle pressure difference") and outputting the corrected diagnosis report. (2) Feedback and iteration mechanism, including: expert correction results are automatically labeled as "gold samples" for incremental training of neural networks (a model fine-tuning is triggered every 50 samples accumulated); new symptoms and weight adjustments are automatically updated to the knowledge base to expand the rule coverage (such as the new rule "fuel nozzle blockage" is added to the database).

[0043] In some embodiments of this example, the training process of the above-mentioned intelligent diagnostic model includes: acquiring multiple training samples of the target gas turbine power plant; taking training samples whose number of samples belonging to the same training sample is less than a preset number as small samples; using the SMOTE algorithm to generate synthetic samples based on the small samples; and adding the synthetic samples to the training samples; wherein the training samples include pre-collected historical fault data and edge fault scenario simulation data; labeling each training sample; using the pre-trained intelligent diagnostic model to predict the labeled training samples; and performing multi-objective optimization on the pre-trained intelligent diagnostic model based on the prediction results and the labels of the training samples; wherein the intelligent diagnostic model is pre-trained based on equipment operation data of gas turbines of the same type as the target gas turbine power plant.

[0044] Specifically, the training and optimization strategies for the intelligent diagnostic model include: 1. Dataset construction. The sample source is 5 years of historical fault data from the 9H gas turbine (1000+ labeled cases) + expert-simulated edge fault scenarios (200+ human cases); data augmentation, including: using the SMOTE algorithm to generate synthetic samples for small sample faults (such as "turbine blade cooling hole blockage" with only 10 cases) to avoid model overfitting; using a labeling system: using multi-label labeling (such as "combustion chamber flameout" may also be labeled "fuel system fault" and "ignition system abnormality"). 2. Model training. Utilizing transfer learning: first pre-training the model with publicly available data from similar gas turbines (such as 9F), and then fine-tuning it with private data from 9H to accelerate convergence; multi-objective optimization: the loss function simultaneously optimizes "classification accuracy" and "rule matching degree" (such as L=α classification loss + β rule conflict loss); dynamic threshold adjustment: setting different confidence thresholds according to the fault risk level (such as increasing the "emergency fault" threshold to 85% to avoid false alarms leading to shutdown). 3. The performance indicators of the intelligent diagnostic model are evaluated, and the performance indicator evaluation table is shown in Table 5.

[0045] Table 5 Performance Index Evaluation Table

[0046] Please refer to Figure 3 , Figure 3 This is a schematic diagram of a fault analysis of compressor surge in a target gas turbine power plant provided by one embodiment of the present invention. Taking the "compressor surge" fault in the 9H gas turbine as an example, the workflow of the intelligent diagnostic model is demonstrated as follows: First, data input: compressor outlet pressure (sudden drop of 0.4MPa / 3 seconds), inlet flow rate (sudden drop of 12% / 3 seconds), shaft vibration (sudden increase of 30μm / 2 seconds), IGV opening (unchanged, fluctuation <1%); then, feature extraction: neural network: extracts latent features such as "pressure-flow rate decrease in the same step" and "vibration peak frequency coupled with speed"; expert features: "sudden drop in outlet pressure (trend type)" and "unchanged IGV opening (exclusionary sign)"; next, dual-track diagnosis: neural network: outputs "compressor surge" with a confidence level of 89%; rule engine: matches rule R-GT-003 and outputs "compressor surge" with a matching degree of 92%; finally, fusion output: the results are consistent and have high confidence, directly outputting "compressor surge" with a handling suggestion ("emergency load reduction to 50%, check IGV actuator").

[0047] In some embodiments of this example, after obtaining the fault diagnosis result, the method further includes: matching a fault handling strategy for processing the fault diagnosis result and a standardized processing procedure for controlling the equipment operation process of the target gas turbine power plant based on the current operating conditions of the target gas turbine power plant and the fault diagnosis result.

[0048] Specifically, the method also provides a decision support module, which provides detailed fault handling suggestions and maintenance strategies based on the fault diagnosis results, including the severity level of the fault, suggested repair time, repair steps, etc., to help maintenance personnel make quick decisions.

[0049] Furthermore, the core objectives and design principles of the aforementioned decision support module are as follows: Core objectives include: automatically generating standardized processing procedures based on fault diagnosis results (fault type, severity, and scope of impact), shortening fault response time (from an average of 30 minutes to within 10 minutes); providing a tiered strategy selection of "emergency handling - short-term maintenance - long-term cure" in conjunction with the current operating conditions of the unit (such as load factor and grid dispatch requirements); integrating data such as equipment manuals, historical maintenance records, and spare parts inventory to ensure the feasibility and economy of the recommendations; supporting multi-role collaborative decision-making, providing emergency operation guidance for operators, and providing resource allocation suggestions for managers. The design principles include: Safety first: All recommendations must comply with the 9H gas turbine safety operation procedures (such as the GE original operation manual), prioritizing the avoidance of "secondary failures caused by handling measures"; Scenario adaptability: The handling strategies for the same failure under different operating conditions (such as full load operation vs. start-up and shutdown phases) need to be differentiated (such as peak-shaving power plants focusing more on "temporary measures without shutting down"); Layered operational granularity: Providing "step-by-step instructions" for on-site operators (such as "closing the #2 fuel valve manual shut-off valve") and providing "resource coordination solutions" for management (such as "coordinating 2 turbine technicians + spare nozzles"); Dynamic updates: Continuously optimizing the recommendation library based on historical handling effect feedback (such as "the success rate of a certain measure is 80%) to improve the effectiveness of strategies.

[0050] Furthermore, the aforementioned decision support module needs to integrate multi-source data to support decision-making recommendations and avoid unrealistic, impractical solutions. The core data tables for the decision support module are shown in Table 6. Table 6 Core Data Tables for the Decision Support Module

[0051] Furthermore, the aforementioned decision support module also includes a fault classification and response strategy system. Specifically, based on the degree of impact of faults on unit safety and power generation plans, 9H gas turbine faults are divided into three levels: emergency, important, and general. Each level corresponds to a differentiated response mechanism and processing time limit, as follows: 1. Emergency Fault (Level I): Definition: Faults that may lead to unit tripping, equipment damage, or safety accidents (such as combustion chamber flameout, compressor surge, or lubricating oil system leakage); Processing Time Limit: Second-level to minute-level response (maximum not exceeding 5 minutes); Core Strategy: Prioritize equipment safety, take emergency shutdown or isolation measures, and prevent the fault from escalating; Output Content: Emergency operation steps (arranged in chronological order, each step clearly defining "operation object + parameter target + verification method"); Emergency personnel deployment (such as "immediately notify the shift leader + main operator + safety officer to arrive on site"); Risk warning (such as "after shutdown, prevent rotor thermal bending and maintain turning gear operation"). Taking combustion chamber flameout as an example:

Emergency Handling Steps

Resource Allocation

Temporary Measures

Maintenance Plan

[0052] Furthermore, the decision support module needs to continuously evolve through feedback on processing results to avoid the static suggestion library becoming disconnected from actual needs, and to achieve a closed-loop optimization and knowledge accumulation mechanism. Specifically, this includes: 1. Feedback data collection, including: after each fault handling is completed, automatically collecting the deviation between the actual measures implemented and the system suggestions (e.g., no load reduction measures were adopted, and the system was shut down directly), processing results (e.g., "successfully resolved", "partially resolved", "unresolved"), time and cost (e.g., "actual maintenance time was 8 hours, 2 hours longer than estimated") and evaluations from maintenance personnel (e.g., "step 3 is vaguely described, it is recommended to add accompanying pictures"). 2. Strategy optimization logic includes: Success rate calibration: For cases that are "successfully resolved," the recommendation weight of the corresponding strategy is increased; for "failed" cases, strategy defects are marked and expert review is triggered (e.g., "a temporary measure caused the fault to worsen and needs to be deleted from the rule base"); Parameter correction: Based on actual time consumption data, the "maintenance time estimate" is dynamically adjusted (e.g., "the actual average time for bearing replacement is 7 hours, the original estimate is 6 hours, corrected to 7 ± 0.5 hours"); Knowledge supplementation: Manually corrected solutions (e.g., "new disassembly techniques" invented by operators) are included in the case base after expert review, enriching the diversity of strategies. This decision support module achieves seamless integration from fault diagnosis results to execution solutions through four mechanisms: "fault graded response, multi-dimensional strategy generation, hierarchical information presentation, and closed-loop optimization." For the 9H gas turbine power plant, the core value of this decision support module lies in: shortening response time: reducing the "decision hesitation period" for fault handling from 30 minutes to within 10 minutes; reducing operational risks: standardizing procedures to avoid human error (such as incorrect sequence of emergency shutdowns); optimizing resource allocation: accurately matching spare parts, personnel, and time windows to avoid "over-preparation" or "insufficient resources"; and accumulating operational knowledge: transforming individual experience into organizational knowledge to improve overall operational capabilities.

[0053] In summary, this application has the following advantages and positive effects: 1. Improves diagnostic accuracy and solves the problem of "misdiagnosis and missed diagnosis under complex operating conditions". The faults of the 9H gas turbine are characterized by "multi-factor coupling, ambiguous symptoms, and different causes for the same symptoms" (e.g., "increased shaft vibration" may be caused by multiple reasons such as bearing wear, rotor imbalance, and airflow excitation). Relying solely on expert experience or data models has limitations, while the fusion mechanism of "big data + expert experience" can achieve complementary advantages and fundamentally improve diagnostic accuracy. Specifically: 1. Technical implementation of the fusion mechanism: Objectivity support of the data model: Through machine learning (such as random forest, LSTM), "hidden correlations" in historical data are mined. For example, based on 10 years of operating data, it was found that "when the ambient temperature is >35℃ + compressor inlet guide vane opening >85% + fuel calorific value fluctuation >5%, the combustion chamber flameout probability increases by 3 times". Such patterns hidden in massive amounts of data often exceed the scope of human experience. Subjectivity calibration of expert experience: For "small sample edge faults" (such as rare faults in the early stage of new engine operation), the data model output is corrected through expert rules. For example, a data model might classify a 2°C deviation in turbine outlet temperature as normal fluctuation, but expert rules state that "if this deviation is accompanied by fluctuations in waste heat boiler steam pressure during the combined cycle mode of the 9H gas turbine, it should be considered a precursor to uneven combustion," thus avoiding missed diagnoses. A dynamic weighted fusion strategy is employed: the weights of both are dynamically adjusted based on the fault type. For common faults (such as low lubricating oil pressure), the data model output takes precedence (70% weight), with expert experience providing supplementary verification. For complex faults (such as multi-system cascading faults), expert rules take precedence (60% weight), with the data model providing trend evidence. 2. Quantitative Implementation of Practical Benefits: Reduced False Diagnosis Rate of Typical Faults: For example, the false diagnosis rate of "compressor surge" and "combustion chamber flameout" has decreased from 15%-20% using traditional methods to below 5%; Improved Early Minor Fault Identification Rate: For example, the accuracy rate of identifying minor bearing wear has increased from 60% to over 90%, allowing for early warnings before vibration values ​​reach alarm thresholds; Reduced Missed Cases: Through "model output + expert rule cross-validation," the number of missed major fault cases per year has decreased from an average of 3-5 to 0-1. II. Enhanced Real-Time Performance, Constructing a Millisecond-Level Fault Early Warning Closed Loop. 9H gas turbine faults develop rapidly (e.g., compressor surge takes only 2-3 seconds from occurrence to unit tripping), and insufficient real-time performance can lead to a passive situation of "early warning followed by tripping." The system uses "edge computing + streaming processing" technology to compress the end-to-end latency of data acquisition, analysis, diagnosis, and early warning to the second or even millisecond level, gaining critical time windows for operation and maintenance intervention. Specifically, the technical architecture for ensuring real-time performance is as follows: Real-time analysis at the edge: Deploy lightweight diagnostic models (such as rule engines and simplified neural networks) at the edge gateway close to the DCS / SIS system to perform local real-time calculations on high-frequency parameters (such as shaft vibration and T3 temperature) to avoid delays in data uploading to the cloud (single parameter analysis delay ≤ 100ms).Streaming data processing: Employing streaming computing frameworks such as Apache Flink, preprocessed time-series data undergoes "sliding window analysis" (window size can be dynamically adjusted, e.g., a 1-second window for critical parameters and a 5-second window for minor parameters) to capture real-time trend changes (e.g., "shaft vibration suddenly increases by 15μm within 100ms"). Multi-level early warning mechanism: Based on the severity and speed of fault development, a three-level response system of "early warning - alarm - emergency shutdown" is set, including: Early warning (e.g., bearing bearing temperature increases slowly, expected to reach alarm value in 24 hours): pushed to the maintenance terminal, prompting planned inspection; Alarm (e.g., shaft vibration suddenly increases to 25μm, approaching the trip value of 30μm): triggers audible and visual alarms and pushes emergency handling plans; Emergency shutdown early warning (e.g., compressor surge symptoms appear): issues shutdown suggestions 0.5-1 second in advance, 200-300ms faster than traditional DCS interlock response. 2. Time Value Conversion: Gaining intervention time for maintenance: For example, combustion chamber flameout warnings can be issued 3-5 seconds in advance, allowing operators sufficient time to perform "fuel quantity fine-tuning + ignition enhancement" to avoid downtime; Reducing fault propagation time: For example, lubricating oil system leak warnings can be issued 10-15 minutes in advance, preventing "minor leaks" from developing into "bearing burnout"; Improving accident handling efficiency: Real-time diagnostic results include "fault location maps" (e.g., indicating "abnormal vibration in the X direction of bearing #2"), shortening maintenance personnel's troubleshooting time (from an average of 30 minutes to within 10 minutes). III. Knowledge Inheritance and Accumulation, Building a Dynamically Growing Diagnostic Wisdom Base. The experience of 9H gas turbine maintenance experts is characterized by "implicit and individualized" features (e.g., the ability of senior engineers to diagnose faults by "listening to sounds" is difficult to standardize), while the iterative mechanism of the knowledge base can achieve the digital accumulation and continuous evolution of experience, solving the industry pain point of "expert retirement = experience loss." Specifically: 1. Implementation path of knowledge inheritance: Structuring expert experience: Implicit experience is transformed into computable rules through "Fault Tree Analysis (FTA) + Symptom-Fault Matrix." For example, the rule "Hearing a 'humming' sound during gas turbine startup may be a precursor to compressor surge" is transformed into a quantitative rule: "Compressor outlet pressure fluctuation > 5% and inlet flow fluctuation > 3% during startup → surge warning." New cases are automatically added to the database: After each fault handling, the system automatically associates the parameter curves before and after the fault, the handling process, and the result verification to form a "fault case package," which is then included in the knowledge base after expert review. For example, after handling a new type of "fuel nozzle blockage" fault, its "characteristic symptoms (fuel flow deviation + T3 temperature field inhomogeneity) + handling solution" automatically becomes a new rule. The knowledge graph is dynamically expanded: A "equipment-parameter-fault-handling solution" knowledge graph is constructed, and new associations are automatically discovered as cases accumulate (such as the indirect association between "scaling of the economizer in the waste heat boiler" and "slow increase in gas turbine exhaust temperature"), enriching the diagnostic dimensions.2. Long-term Value Realization: Improved Diagnostic Coverage: The knowledge base initially covers 80% of typical faults in 9H gas turbines, and this coverage can be increased to over 95% within 3 years through the accumulation of new cases each year; Shortened Newcomer Training Cycle: New maintenance personnel can shorten their independent onboarding time from 12 months to 6 months by querying "fault cases + handling videos" in the knowledge base; Cross-Plant Experience Sharing: The knowledge bases of multiple power plants within the group are interconnected, allowing "special operating condition fault handling experience" from one power plant to quickly empower other power plants (e.g., experience on "salt spray corrosion causing sensor drift" from coastal power plants can be shared with inland power plants). IV. Reducing Maintenance Costs: Economic Transformation from Reactive Emergency Repair to Precise Maintenance. In the maintenance costs of 9H gas turbine power plants, unplanned downtime losses, excessive maintenance waste, and spare parts inventory backlog account for over 60%. The system can achieve structured cost optimization through precise diagnosis and early warning. Specifically: Reducing Unplanned Downtime Losses: The system can reduce unplanned downtime by 3-5 times per year through early warning. Avoid over-maintenance: In traditional scheduled overhauls, approximately 20% of component replacements are unnecessary (e.g., bearings in good condition are replaced simply because their service life has expired). The system, based on condition monitoring and remaining life prediction, can reduce the over-maintenance rate to below 5%. Optimize spare parts inventory: Through precise fault diagnosis (e.g., pinpointing the fault as "bearing #1" rather than "the entire shaft system"), spare parts inventory turnover can be increased by 30%. Extend equipment life: Timely handling of early-stage faults prevents minor issues from accumulating into major damage; for example, addressing minor bearing wear promptly can extend its lifespan by 2-3 years.

[0054] In some implementations of this embodiment, taking the compressor fault diagnosis of a 9H gas turbine power plant as an example, the system collects real-time data on compressor temperature, pressure, vibration, etc., and analyzes this data through a deep neural network. It discovers that the vibration frequency of the compressor blades fluctuates abnormally over a certain period, but has not yet exceeded the traditional threshold. At this point, the system automatically triggers the human-machine interface, prompting experts to make further judgments. Based on their experience, combined with current operating conditions and historical fault data, the experts determine that the abnormal vibration may be caused by blade fouling. Subsequently, based on the expert's judgment, the system further analyzes relevant data, ultimately determining the cause of the compressor blade fouling fault and providing maintenance suggestions for cleaning the blades. Through the application of this system, the further development of the compressor fault was successfully prevented, ensuring the safe and stable operation of the unit. The specific process is as follows: First, data acquisition and real-time monitoring: The power plant collects real-time compressor operating data through high-precision sensors distributed throughout the compressor, including inlet temperature, pressure, flow rate, vibration frequency and displacement of each stage of blades, and multi-dimensional data such as compressor outlet temperature and pressure. These data are rapidly transmitted to the Data Acquisition and Monitoring System (DCS) at second-level intervals, and then aggregated to the Data Preprocessing Module (IDCS) of the Intelligent Diagnostic System. For example, the intake air temperature sensor collects data every second to ensure high-frequency monitoring of the compressor's operating status. Then, data preprocessing and feature extraction occur: the data preprocessing module uses a Kalman filter algorithm to denoise the collected raw data, removing noise data caused by electromagnetic interference, sensor errors, and other factors. Simultaneously, a normalization method is used to standardize the data, enabling analysis of data with different dimensions on a unified scale. In processing vibration data, Fourier transform is used to convert time-domain vibration data into frequency-domain data, extracting key feature parameters such as the dominant frequency and harmonics of the vibration. These feature parameters will serve as important inputs for subsequent diagnostic models. Then, the intelligent diagnostic model performs preliminary analysis: the machine learning models in the intelligent diagnostic system, such as deep neural networks (DNNs), are pre-trained based on a large amount of historical operating data and known fault cases. When the real-time collected compressor data, after preprocessing and feature extraction, is input into the DNN model, the model performs complex pattern recognition and analysis on the data. At a certain point, the DNN model analysis detected abnormal fluctuations in the vibration frequency of a compressor stage blade. The dominant frequency amplitude exceeded the historical data characteristic range corresponding to normal operation, and the vibration frequency showed a gradually increasing trend. However, the DNN model analysis alone was insufficient to definitively determine the type and cause of the fault. Then, human-machine interaction and expert experience were introduced: upon detecting the abnormal vibration signal, the system immediately triggered the human-machine interaction mechanism. On one hand, the system presented the abnormal data and the preliminary analysis results of the DNN model to senior equipment experts at the power plant through an intuitive visual interface, including the vibration frequency variation curve over time, and comparisons of real-time values ​​and historical averages of relevant parameters.On the other hand, experts queried the compressor's historical operating records, recent maintenance status, and similar fault case database through the system. Based on years of accumulated field experience and the current operating conditions, the experts determined that the compressor was operating in a high-humidity environment and had been running continuously for an extended period without cleaning or maintenance. This strongly suggested that blade fouling was causing uneven mass distribution, leading to abnormal vibration. Then, a comprehensive diagnosis and fault confirmation were performed: based on the expert's judgment, the intelligent diagnostic system further utilized relevant rules and cases from the knowledge base to verify and supplement the results of the DNN model. By comparing historical cases of abnormal vibration caused by blade fouling, the system found a high degree of similarity between the current compressor's vibration characteristics and historical cases. Simultaneously, the system used Fault Tree Analysis (FTA) to analyze the possible causes of the abnormal vibration from the top event, eliminating other possible factors such as bearing failure and airflow excitation. Through collaborative analysis between the system and experts, the abnormal compressor vibration was ultimately confirmed to be caused by blade fouling. Finally, maintenance decisions and recommendations were output: based on the fault diagnosis results, the diagnostic system generated a detailed maintenance decision and recommendation report. The report explicitly states that the compressor blades need cleaning and maintenance to restore their normal aerodynamic performance and mass distribution. It also provides a suggested maintenance timeframe, recommending that cleaning be carried out during a suitable shutdown window in the near future, considering that the current compressor vibration does not pose a serious threat to the safe operation of the unit, to avoid significant impact on power generation due to unplanned shutdowns. The report also provides specific maintenance steps, including recommended cleaning methods (such as a combination of chemical cleaning and high-pressure water flushing), a list of required tools and materials, and post-maintenance acceptance criteria, such as the blade vibration frequency returning to the normal operating range and the compressor's inlet flow rate and pressure returning to design values. Finally, the report details the maintenance implementation and effectiveness verification: Based on the maintenance recommendations provided by the diagnostic system, the power plant maintenance team conducted a comprehensive cleaning of the compressor blades during the scheduled shutdown time. After the maintenance was completed, the compressor was started and the unit resumed operation. The intelligent diagnostic system continuously monitored the compressor's operating data, and the results showed that the blade vibration frequency quickly returned to the normal range, the inlet flow rate and pressure stabilized near the design values, and all performance indicators returned to normal. This maintenance not only successfully resolved potential compressor malfunctions and ensured the safe and stable operation of the unit, but also verified the effectiveness and accuracy of the intelligent diagnostic method based on human-machine experience integration in equipment fault diagnosis at the 9H gas turbine power plant.

[0055] Figure 4 This is a block diagram of an intelligent diagnostic system for power plant equipment based on human-machine experience fusion, provided by one embodiment of the present invention. Figure 4As shown, this invention provides an intelligent diagnostic system for power plant equipment based on human-machine experience fusion. The system includes: a fault symptom feature matching module, used to input equipment operation data of a target gas turbine power plant within a target period into a fault diagnosis knowledge base, and sequentially perform fault symptom feature matching corresponding to the changing trends of the equipment operation data within the target period, and determine the weights of the matched fault symptom features under each corresponding fault type; and a fault diagnosis module, used to input the equipment operation data within the target period and the weights of the matched fault symptom features under each corresponding fault type into an intelligent diagnostic model containing a machine learning diagnostic unit and an expert rule diagnostic unit to obtain fault diagnosis results; wherein, the fault diagnosis results are obtained by fusing the results of the machine learning diagnostic unit and the expert rule diagnostic unit predicting the probability of fault occurrence for each fault type of the matched fault symptom features.

[0056] Specifically, the system achieves fault symptom feature matching by comparing the changing trends of equipment operating data of the target gas turbine power plant within a target period with fault symptom features in a fault diagnosis knowledge base used to characterize different data changing trends. It then obtains the weights of the matched fault symptom features under the corresponding fault types from the fault diagnosis knowledge base. The machine learning diagnostic unit in the intelligent diagnostic model processes the input equipment operating data within the target period and the weights of the matched fault symptom features under each corresponding fault type from the perspective of machine learning algorithms, outputting the fault type determined by the machine learning algorithm. Furthermore, the expert rule diagnostic unit in the intelligent diagnostic model infers the weights of the input equipment operating data within the target period and the matched fault symptom features under each corresponding fault type from the perspective of a logical rule base established by domain experts, outputting the determined fault type. By fusing the fault types determined by the machine learning diagnostic unit and the expert rule diagnostic unit, a comprehensive fault diagnosis result is formed. This system integrates human and machine experience to achieve intelligent diagnosis of power plant equipment operation. It uses machine learning algorithms to process equipment operation data and combines the rich experience of domain experts to improve the accuracy, real-time performance, and reliability of fault diagnosis for gas turbine power plant equipment. This enables early warning and accurate diagnosis of equipment faults, reduces unplanned downtime, and lowers operation and maintenance costs.

[0057] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0058] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0059] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0060] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for intelligent diagnosis of power plant equipment based on human-machine experience fusion, characterized in that, The method includes: Input the equipment operation data of the target gas turbine power plant within the target period into the fault diagnosis knowledge base, and perform fault symptom feature matching with the change trend of equipment operation data within the target period one by one, and determine the weight of the matched fault symptom features under each corresponding fault type. The equipment operation data within the target period and the weights of the matched fault symptom features under each corresponding fault type are input into an intelligent diagnostic model that includes a machine learning diagnostic unit and an expert rule diagnostic unit to obtain fault diagnosis results. The fault diagnosis results are obtained by fusing the results of the machine learning diagnosis unit and the expert rule diagnosis unit in predicting the probability of occurrence of each fault type based on the matched fault symptom features.

2. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 1, characterized in that, The rules for collecting the device operation data include: Based on the joint operation process of the target gas turbine power plant, a data acquisition scheme for each data acquisition system is determined, and the data acquisition of each data acquisition system is executed based on the determined data acquisition scheme; wherein, the data acquisition system includes a DCS system, a SIS system, and an auxiliary system; Based on the transmission protocol of each acquisition system, corresponding data access rules are matched, and based on the matched data access rules, the equipment operation data collected by each acquisition system is obtained through the edge gateway deployed at the target gas turbine power plant.

3. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 2, characterized in that, After collecting device operation data, the method further includes: The collected equipment operation data undergoes one or more preprocessing steps, including data cleaning, noise filtering, data standardization, and normalization. The preprocessed equipment operation data is then subjected to preprocessing quality verification, including: Based on the business logic of the target gas turbine power plant, data cleaning rules are matched with the equipment operation data to perform data cleaning of the corresponding equipment operation data; based on the type of interference noise in the equipment operation data, noise filtering rules are matched with the corresponding equipment operation data to perform noise filtering of the corresponding equipment operation data. The preprocessed equipment operation data is subjected to integrity verification, consistency verification, and data change trend verification before / after preprocessing.

4. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 1, characterized in that, The construction rules for the fault diagnosis knowledge base include: Based on production rules, this method utilizes pre-collected expert experience data and historical fault data to identify fault symptom features for each fault type and determine the weights of these identified features for each fault type. By comprehensively considering the determined weights of each fault symptom feature for each fault type, a fault diagnosis knowledge base is established. The rules for determining the weights of each fault symptom feature under each corresponding fault type include: The importance of fault symptom features under the same fault type is compared pairwise, and a judgment matrix for the corresponding fault type is established based on the pairwise comparison results. Based on the established judgment matrix, initial weights are assigned to the fault symptom features under the same fault type, and consistency checks are performed on the assigned fault symptom features. Based on the statistical weights of the fault symptom features under each corresponding fault type, determined by the support and confidence of the fault symptom features with respect to each corresponding fault type, the weights of the fault symptom features after consistency verification are calibrated to obtain the final weights of the fault symptom features under each corresponding fault type; wherein, the support and confidence of the fault symptom features with respect to each corresponding fault type are obtained based on the statistical analysis of historical fault data collected in advance.

5. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 4, characterized in that, The rules for collecting historical fault data include: Fault data within a historical time period is obtained from the equipment operation database of the target gas turbine power plant. Based on the obtained fault data, corresponding fault data analysis is performed according to multiple data mining rules to extract historical fault data used to characterize the correlation between fault symptom features and fault types. The data mining rules include at least one or more of the following: time-series data analysis rules for identifying patterns of equipment operation data changes before / after a fault occurs; association mining rules for mining the correlation between combined features of fault symptom features and fault types; and statistical analysis rules for determining the frequency of occurrence of each fault symptom feature and its contribution to the fault.

6. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 1, characterized in that, The rules for obtaining the fault diagnosis results include: The confidence levels of the fault type prediction results output by the machine learning diagnostic unit and the fault type prediction results output by the expert rule diagnostic unit are compared. When the confidence levels of the fault type prediction results output by both are higher than the preset confidence threshold and the fault types in the fault type prediction results output by both are consistent, the output fault type is determined as the fault diagnosis result. Otherwise, the equipment operation data within the target period and the fault type prediction results output by both are input into the expert diagnostic workbench, and the fault diagnosis results fed back by the expert diagnostic workbench are received.

7. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 1, characterized in that, The rules for predicting the probability of failure occurrence for each fault type based on the matched fault symptom features using machine learning diagnostic units include: A CNN-LSTM hybrid network is used to extract latent features from the equipment operation data within the target period. Based on the extracted latent features, a Transformer model with attention mechanism is used to match fault types and output the fault type and corresponding confidence level. The rules for predicting the probability of failure occurrence for each fault type based on the matched fault symptom features, according to the expert rule diagnostic unit, include: Based on predefined key explicit features, explicit features are extracted from the equipment operation data within the target period, and the extracted explicit features are input into the fault diagnosis knowledge base for fault type matching, outputting the fault type and corresponding confidence level; wherein, the key explicit features are based on the operating mechanism definition of the target gas turbine power plant.

8. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 1, characterized in that, The training process of the intelligent diagnostic model includes: Multiple training samples of the target gas turbine power plant are obtained. Training samples with fewer than a preset number of samples belonging to the same training sample are regarded as small samples. The SMOTE algorithm is used to generate synthetic samples based on the small samples, and the synthetic samples are added to the training samples. The training samples include pre-collected historical fault data and edge fault scenario simulation data. Each training sample is labeled, and a pre-trained intelligent diagnostic model is used to predict the labeled training samples. Based on the prediction results and the labels of the training samples, the pre-trained intelligent diagnostic model is optimized for multiple objectives. The intelligent diagnostic model is pre-trained based on the equipment operation data of the same type of gas turbine as the target gas turbine power plant.

9. The intelligent diagnostic method for power plant equipment based on human-machine experience fusion according to claim 1, characterized in that, After obtaining the fault diagnosis results, the method further includes: Based on the current operating conditions and fault diagnosis results of the target gas turbine power plant, a fault handling strategy for processing the fault diagnosis results and a standardized processing procedure for controlling the equipment operation process of the target gas turbine power plant are matched.

10. A power plant equipment intelligent diagnostic system based on human-machine experience fusion, characterized in that, The system includes: The fault symptom feature matching module is used to input the equipment operation data of the target gas turbine power plant within the target period into the fault diagnosis knowledge base, and perform fault symptom feature matching corresponding to the changing trend of equipment operation data within the target period, and determine the weight of the matched fault symptom features under each corresponding fault type. The fault diagnosis module is used to input the equipment operation data within the target period and the weights of the matched fault symptom features under each corresponding fault type into the intelligent diagnosis model, which includes a machine learning diagnosis unit and an expert rule diagnosis unit, to obtain the fault diagnosis results. The fault diagnosis results are obtained by fusing the results of the machine learning diagnosis unit and the expert rule diagnosis unit in predicting the probability of occurrence of each fault type based on the matched fault symptom features.

Images