The hydropower unit condition monitoring system, through real-time detection and monitoring of various parameters of the unit and comprehensive analysis of historical equipment conditions, can diagnose faults and predict trends, promptly assess equipment performance, and facilitate the development of reasonable equipment maintenance and repair systems for hydropower plants, thereby extending maintenance cycles and shortening maintenance time. 1. Characteristics of Hydropower Unit Fault Diagnosis Hydropower unit fault diagnosis mainly includes turbine diagnosis and generator diagnosis, which are interconnected. Turbine diagnosis includes diagnosing faults in the shaft system, impeller, blades, guide bearings, lubrication system, support system, and control system. Key fault characteristics include: increases and decreases in oil tank level, increases in oil temperature, increases in average bearing temperature, rate of increase in bearing temperature, increases in runout, increases in bearing vibration, trends in vibration and runout with changes in speed and load, and vibration conditions with and without load. Generator diagnosis includes diagnosing faults in the stator coil, stator core, rotor, and shaft system. Stator coil diagnosis uses partial discharge method (pulse high-frequency capacitance) to detect the insulation of the coil, bar, and bar support, as well as coil open circuits. The stator core and rotor are diagnosed using a set of air gap measurement sensors to monitor the dynamic air gap caused by the stator aperture, rotor circumference, and eccentricity. The parameters detected generally include: frame vibration, runout, temperature, electrical quantity, air gap, insulation monitoring, cavitation, and pressure pulsation. For diagnosing guide bearings, thrust bearings, and other complex faults, the diagnostic system must comprehensively analyze all detected parameters. [b]2 Fault Diagnosis System Structure[/b] The unit fault diagnosis system includes data preprocessing and data normalization, information processing, fault diagnosis knowledge modeling, fault detection, fault isolation and estimation, fault classification evaluation and decision-making, knowledge model base, database, intelligent decision support system, human-machine interface, and monitoring center. The fault diagnosis system structure is shown in Figure 1. The data preprocessing and data normalization module preprocesses the measurable variables sampled online to eliminate noise, and then converts them into standardized digital signals for input into the database. The information processing and fault diagnosis knowledge modeling module reprocesses the vibration, runout, and water pressure pulsation signals input from the unit and builds a knowledge model base. The knowledge model base and database play a real-time supporting role in the fault detection, fault isolation and estimation, fault classification evaluation and decision-making modules. The intelligent decision support system module is responsible for selecting which diagnostic methods or methods to use for different operating states of the unit, as well as the integration between various diagnostic methods, providing corresponding fault handling solutions, optimizing operation decisions and maintenance plan decisions, and coordinating the use of knowledge in the knowledge model base. The human-machine interface and monitoring center are responsible for the input and output of unit information, including the input of deep and shallow knowledge, the output of fault information, and the activation of protection measures when a fault is determined, which is accomplished by the actuator. In addition, there are data report output and printing functions. [b]3 Knowledge Model Base[/b] How to specifically realize the function of fault diagnosis knowledge modeling, acquire diagnostic knowledge models, form a knowledge model base, and thus implement the functions of online fault detection, fault isolation and estimation, fault classification evaluation and decision-making modules is the difficulty of the system. In reference [1], Yang Jie et al. proposed a fault diagnosis modeling and reasoning method based on a comprehensive model, namely artificial neural network, case, rule and object model, to effectively acquire diagnostic knowledge models. The knowledge model of a hydropower unit can be represented in four ways: unit model, diagnostic experience rules, diagnostic neural network model, and diagnostic cases. The diagnostic knowledge model is organized into four layers from general to specific: The first layer describes the most general diagnostic knowledge, consisting of unit models of the turbine and generator. The operation model describes the state of the unit during normal operation, and the fault model describes the state of the unit during faults; these are used for model-based diagnosis, truth maintenance, and interpretation. The second layer describes general diagnostic knowledge, consisting of diagnostic rules summarized from the technical standards, regulations, and expert diagnostic experience of hydropower units, used for rule-based diagnosis. The third layer is an artificial neural network model constructed based on similar diagnostic cases, used for neural network model-based diagnosis. The fourth layer consists of a case-sub-case hierarchical framework, describing specific diagnostic case knowledge for each power plant unit or between power plant units, used for case-based diagnosis. 3.1 The first layer of quantitative information description generally includes parameter description and state description. Parameter description refers to describing the occurrence of faults by significant changes in unit parameters, such as threshold crossings or abrupt changes in the values ​​of temperature, electricity, etc. State description refers to the fault information contained in the description of the unit's start-up and shutdown processes and operating status. The operating model and fault model formed by the quantitative information description of the unit constitute the unit model sub-library, forming the first layer of the model knowledge base. The fault diagnosis methods based on the unit model supported by this layer include: parameter estimation diagnosis and state estimation diagnosis. In the parameter estimation diagnosis method, when performing fault detection, fault isolation and estimation, fault classification evaluation, and decision-making, the process parameter is the comparison value between the simulated quantity of the unit model and the actual operating parameters of the unit, and the resulting residual deviation is the relative change value between the two. Various fault information is contained in the series of residual deviations. Combined with the corresponding models in the unit model library, statistical testing is used on the basic residual sequence to detect the location and cause of the unit's faults, and further isolate, estimate, and make decisions. The state estimation diagnosis method determines whether the unit has a fault state based on the control logic of the hydropower production process and can be directly supported by the unit monitoring system. 3.2 The second layer is based on diagnostic rules summarized from standards and experience. Faults can be divided into two types: deterministic faults and uncertain faults. For deterministic faults, i.e., general production faults, a knowledge model base for logical reasoning can be established. For nondeterministic faults, fuzzy production rules are generally used to represent fault diagnosis knowledge, that is, fuzzy relation matrices are used to represent the causal relationship between preconditions and conclusions. In addition, credibility methods, probabilistic methods, etc., can be used to describe their uncertainty. These methods can be fully collected, organized, and optimized to form a relatively complete diagnostic reasoning mechanism. 3.3 The third layer is an artificial neural network model constructed based on similar fault diagnosis cases. Its essence is a fault classification and recognition process. Here, the artificial neural network acts as an adaptive pattern recognition technology. It uses its own learning mechanism to automatically form the corresponding decision region by learning from case samples. Moreover, when the samples change, such as when the number of cases increases, the mapping relationship obtained by the neural network training can be adaptive, achieving a further approximation of accurate diagnosis. 3.4 The fourth layer consists of a case-sub-case hierarchical framework, forming a diagnostic case sub-base of the most specific knowledge. The information diagnosis of the unit's online status is matched with the case descriptions in the case sub-base to obtain a solution strategy. The above four-level knowledge model sub-bases are both independent and closely related. Once a new type of fault in the unit is diagnosed, it can be described and added to the case-sub-case level framework. The new type of fault and the existing similar faults in the framework can be used to construct and train new neural network models, which are then added to the third-layer artificial neural network model sub-library. The diagnostic method rules for the new type of fault and the existing similar faults are added to the second-layer rule sub-library. If the knowledge representation of the above three layers can be represented by quantitative information, it can be modeled and added to the first-layer unit model library. [b]4 Fusion Diagnostic Reasoning[/b] For the diagnosis of complex faults, it is not possible to simply diagnose them using only one method. An effective method is the fusion diagnostic method that effectively combines and judges various detection information. In reference [2], Peng Tao et al. proposed an artificial neural network diagnostic method based on signal processing, namely, feature extraction based on wavelet transform, feature selection based on genetic algorithm, and state recognition theory based on neural network. This method can use a weighted method to achieve primary fusion of information from multiple sensor signals of the unit, such as oscillation, cavitation, and water pressure pulsation signals. Wavelet transform is then performed according to a given wavelet function to extract its feature components. A genetic algorithm is used to search and select the most important feature parameters from the input parameters, which, along with known target feature information, are used as training samples for neural network training to achieve state recognition and fault diagnosis. Wavelet transform is also used for some complex detection information, such as the processing of partial discharge data measured in insulation monitoring. Given the weak and noisy characteristics of the partial discharge signal, wavelet transform is used for analysis, fully utilizing the excellent time-frequency analysis characteristics of wavelet analysis. The method employs a rational strategy until the optimal processing scheme is given, resulting in the most accurate predictive control and diagnostic results. [b]5 Conclusion[/b] In the design, manufacturing, installation, and overhaul of hydropower units, regulations and requirements are put forward for various state parameters of the unit. However, due to irregular hydraulic interference during the operation of hydropower units, not only are the actual operating parameter values ​​and their variation patterns different for units of different models, capacities, and structures, but the actual operating parameters of several units of the same model in the same power station are also difficult to be consistent. For example, the bearing temperature of a typical generator set is 3-5°C higher under full load than under no-load, while some units do not experience a temperature increase under load, and may even experience a slight decrease. Furthermore, some units can operate for extended periods despite frame vibrations of up to 1 mm, while others have water guide oscillations of up to 0.8 mm, yet their bearing temperatures remain normal and their operation is stable. Therefore, establishing a unified standard within a fault diagnosis reasoning system is very difficult, and to date, there is no international standard for condition monitoring. From the perspective of actual overhaul work, the main causes of component damage and loosening are cavitation, wear, corrosion, aging, and fatigue. However, direct measurement, especially online measurement of the aging degree of rubber seals, is extremely difficult, if not impossible. Therefore, it is necessary to include a manual intervention function in the monitoring center of the fault diagnosis reasoning system to supplement case-level knowledge in a timely manner, achieving human-machine collaboration and intelligent complementarity. Integrating humans into the entire system also allows for the resetting or modification of certain diagnoses based on the characteristics of each unit, forming an intelligent and efficient reasoning mechanism tailored to the specific characteristics of each unit. This enables the most accurate prediction and diagnosis of faults for each unit.