1. Introduction Power plant boiler equipment is a complex system. Its fault diagnosis requires high real-time performance, high reasoning efficiency, and accurate location. Once fault symptoms appear, an accurate judgment must be made immediately, and rapid measures taken to avoid catastrophic accidents. Fault diagnosis involves many issues such as diagnostic knowledge processing mechanisms, real-time databases, and knowledge bases. Traditional knowledge processing methods have advantages such as intuitive knowledge representation, strong modularity, and clear logical reasoning. However, these methods suffer from limitations such as bottlenecks in knowledge acquisition, low reasoning efficiency, poor adaptability, and poor real-time performance. This is a critical obstacle for a diagnostic system that must be accurate, practical, real-time, and continuously improve and expand. Therefore, for many years, people have been actively seeking various solutions, such as adopting multiple knowledge representation methods and solution strategies to improve system flexibility; and using machine learning methods to solve the bottleneck problem of knowledge acquisition. The fundamental purpose of the above solutions is to effectively combine knowledge-based symbolic processing methods with fuzzy neural networks, establish a powerful real-time database, and perform in-depth processing of various fault symptoms on a unified data information platform. For power plant boiler fault diagnosis, structural and functional knowledge of the boiler equipment, historical operating data, fault data, and boiler operating procedures are all crucial information resources. A comprehensive understanding of this data is essential for maintaining the normal operation of the power plant and for making production decisions and management. However, due to historical and technical reasons, a significant number of thermal power plants in China currently experience information loss and stagnation on their production lines, hindering rapid data sharing between various application systems. This is highly detrimental to establishing a real-time boiler monitoring and diagnostic system. Therefore, establishing a real-time boiler database using enterprise information resource management standards can standardize and regulate fault diagnosis information, effectively improving the efficiency and accuracy of the diagnostic process. The boiler fault diagnosis knowledge base is essentially a rule base composed of phenomena of abnormal accidents, parameters of abnormal signs, and corresponding handling measures. Using object-oriented knowledge representation methods, these can be treated as different objects. The main difference between these objects lies in their attribute values. Therefore, as long as the database stores all fault symptoms, fault types, fault handling, and other attribute values ​​of the entire diagnostic system and defines them as various abnormal event classes, when dealing with the "bottleneck" of "knowledge acquisition" in the diagnostic system, different instances of the abnormal event classes can be created to inherit the attribute values ​​of the corresponding event classes, dynamically generating various different objects. These objects can then be used as experience samples, trained through a neural network, and stored as knowledge in the form of network weights. Because these objects have a close association mechanism, a mature relational database is well-suited for storing and managing their attributes. This paper focuses on fuzzy neural networks and real-time relational databases, and briefly mentions the system implementation and experimental results. 2. Fault Pattern Recognition Based on Fuzzy Neural Networks Fault diagnosis is essentially a state recognition and classification problem. It identifies the operating state based on various state parameters measured by sensors, and diagnoses the fault immediately upon an anomaly. Currently, various fault diagnosis methods have been proposed and studied, such as fuzzy comprehensive evaluation, fuzzy cluster analysis, pattern recognition, artificial neural networks, and expert systems. Each of these methods has its own advantages and disadvantages. Artificial neural networks have attracted widespread attention in the field of equipment fault diagnosis due to their unique associative, memory, and learning functions. Neural networks process information through the interaction of numerous simple processing units called nodes. By learning from experience samples, expert knowledge is stored in the network in the form of weights, and the information preservation property of the network is used to complete imprecise diagnostic reasoning, which can better simulate expert experience and intuition rather than complex reasoning processes. Therefore, employing the self-learning function, associative memory function, and distributed concurrent information processing function based on neural networks has certain advantages in knowledge representation, acquisition, and parallel reasoning in diagnostic systems. Fuzzy neural network diagnosis is a technology that integrates neural networks and fuzzy logic. A fuzzy neural network (fnn) is a network formed by two or more fuzzy neurons interconnected. Its construction method is to fuzzify the traditional neural network. This type of fnn retains the original neural network structure while fuzzifying the neurons, giving them the ability to analyze fuzzy information. 2.1 Fuzzy Neuron The basic structure of a fuzzy neuron is shown in Figure 1. [img=450,216]http://www.ca800.com/maga/images/2003101615595879690.gif[/img][align=left] The inputs of fuzzy neurons are represented by fuzzy sets a1(x1), a2(x2), ..., an(xn) in the domains u1, u2, ..., un. The "weighted" inputs are accumulated not by summation but by fuzzy accumulation operations. Fuzzy operations can also be performed on the output y depending on the specific situation. In fuzzy neural networks, the design of fuzzy neurons should enable them to have roughly the same functions as non-fuzzy neurons, while also requiring them to reflect the fuzzy properties of neurons and have the ability to process fuzzy information. The process from each input to the output is not always the same. 2.2 Fuzzy neurons described by fuzzy rules In knowledge-based systems, a set of conditional statements "if-then" rules are often used to represent knowledge obtained from human experts. This knowledge is often accompanied by uncertainty and fuzzy terminology. Therefore, in "if-then", the premise and conclusion are treated as fuzzy sets. The i-th type of neuron is described by this rule. [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 6] [Image 7] [Image 8] [Image 9] [Image 1] [Image 1] [Image 1] [Image 2] [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 6] [Image 7] [Image 8] [Image 9] [Image 1] [Image 1] [Image 1] [Image 2] [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 1] [Image 1] [Image 2] [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 6] [Image 7] [Image 8] [Image 9] [Image 1] [Image 1] [Image 1] [Image 2] [Image 1] [Image 2] [Image 1] [Image 2] [Image 3] [Image 4] [Image 1] [Image 2] [Image 1] [Image 2] [Image 3] [Image 4] [Image 1] [Image 2] [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 1 ... [img=93,48]http://www.ca800.com/maga/images/200310161634337179.gif[/img]…and [img=93,48]http://www.ca800.com/maga/images/20031016164774352.gif[/img] then yi=bi（y） Here, x1,x2,…xn are the current inputs, and bi is the current output of the i-th neuron. The i-th neuron is described by the i-th rule among all m rules shown in the fuzzy neural network. That is, each neuron represents one of the m “if-then” rules. According to fuzzy logic theory, the i-th fuzzy neuron can be described by a fuzzy relation ri: ri = [img=93,48]http://www.ca800.com/maga/images/20031016165546200.gif[/img]×[img=93,48]http://www.ca800.com/maga/images/200310161653982521.gif[/img]×…×[img=93,48]http://www.ca800.com/maga/images/200310161655537867.gif[/img] ×bi. Given the current input (fuzzy or non-fuzzy) a1, a2, …, an, according to the inferred composite rule, the output given by the i-th rule is yi = a1o (a2o (…(an ri)…)). Here, "o" represents any compound operation. In the proposed fuzzy neuron, the input is associated with its output through a fuzzy conditional statement or an "if-then" rule. The neuron's experience is stored in the fuzzy relation ri, and its output consists of the current input and past experience ri. The learning algorithm for the i-th type of neuron can be specifically determined by the actual problem. The learning of fuzzy neurons can be achieved through "synaptic" adjustment or "body" adjustment. "Synaptic" adjustment means continuously correcting all inputs and then forward transmitting them to the neuron body. "Body" adjustment, on the other hand, corrects past experience. General fuzzy neural networks adopt a multi-layer feedforward neural network structure. Due to the differences in neurons and the fuzzy components incorporated, different types of fuzzy neural networks have emerged. Fuzzy neural networks inherit the learning algorithms of conventional neural networks, but due to the special nature of fuzzy information, some unique algorithms have also been formed. Fuzzy neural networks integrate the advantages of fuzzy logic and neural networks, capable of representing qualitative knowledge and possessing the ability to self-learn and process quantitative data; they are the culmination of the integration of the two. In equipment fault diagnosis, its construction method is to fuzzify the traditional neural network. It retains the original neural network structure while fuzzifying the neurons to enable them to process fuzzy information. For network structure and other details, please refer to references [1][2][3][4]. 3 Fault Diagnosis Database In the fault diagnosis process, a large amount of time series data needs to be collected, and a large amount of real-time data needs to be stored and processed. Therefore, a real-time database is crucial. Based on the boiler operation procedures, a boiler fault diagnosis database is established, including fault types, fault symptom details, experience sample set tables, parameter symptom analysis tables, etc., to achieve effective storage of knowledge, convenient management and maintenance, and to facilitate reasoning within the system. In order to facilitate and integrate the fault diagnosis system into a unified data information platform, the best choice is to use Wonderware's InSQL Server, a real-time relational database for the development of factory and process data. This database can connect to more than 500 kinds of control devices and data acquisition systems through Wonderware I/O Server. Due to the optimized design for the acquisition and storage of analog and discrete data, InSQL Server outperforms all ordinary relational databases in many aspects under the same hardware environment, making it possible to store high-speed data in relational databases and capture data at high speed. Due to the use of data compression technology, data can be stored in a very small space. Compared with ordinary relational databases, InSQL Server only requires about 2% of the space needed to store the same amount of data. For example, for a factory with 4,000 variables and scanning speeds ranging from seconds to levels, storing two months of historical data would typically require no less than 2GB of disk space, while InSQL Server only requires 40MB. Because of the reduced storage space, the price of storage hardware has become increasingly less important. To ensure users receive high-resolution and high-quality data, the database uses a lossless compression algorithm. General databases do not support time-series data; for example, SQL does not support time-series data, especially in terms of controlling the resolution of returned data and the inability to proactively provide data to users. InSQL Server, however, allows control over resolution and proactive updates, and provides basic time-related functions on the server, such as rate of change and complete process calculations. These features significantly shorten the software development cycle of the fault diagnosis system. 4. Engineering Implementation of Fault Diagnosis Currently, power plants commonly use DCS systems, which are feature-rich, including monitoring and recording operating parameters, real-time trend and historical trend display, etc. By slightly modifying the DCS system, not only can redundant investment be avoided and the potential of existing resources be fully utilized, but more importantly, the development cycle of the fault diagnosis system is greatly shortened, which is in line with my country's national conditions. The core technology of the diagnostic system lies in how to organically combine neural networks and expert systems. Because the two have many differences in knowledge acquisition, knowledge representation, and reasoning mechanisms, if all knowledge is converted into something related to thermal parameters, and the extraction and preprocessing of parameter symptoms meet the input requirements of the neural network, then the problem can be easily solved. Figure 4 is a schematic diagram of the system structure for extracting thermal parameter symptoms. [img=450,290]http://www.ca800.com/maga/images/200310161664242405.gif[/img] The fault diagnosis process of the equipment includes two aspects: abnormal judgment of parameter symptoms and fault diagnosis. The former is the foundation of fault diagnosis. For the parameters collected reflecting the equipment's operating status, it's necessary to extract symptoms and determine whether the operating status is normal or abnormal based on the equipment's normal operating characteristics and established standards (usually set to four states: normal, alarm, danger alarm, and abnormal). If the state is abnormal, further diagnosis is needed to find the fault. Normal and abnormal states are usually judged based on the experience of on-site personnel or a specific standard, but in the field, this standard is not very strict. For example, a unit operating under rated conditions has a main steam temperature of 540℃. Within this rated value, +5℃ to -10℃ is considered normal. In other words, 545℃ and 530℃ can be used as limits to determine the abnormality of the main steam temperature state. However, this standard is not strict. Sometimes the main steam temperature is slightly higher than 545℃ or slightly lower than 530℃, and the unit is still in normal operation. Furthermore, if the parameters are abnormal, it's difficult to clearly distinguish whether it belongs to an alarm state or a danger state. Therefore, it can be said that the boundaries between normal and abnormal, and between the degrees of abnormality, are fuzzy, and fuzzy sets are precisely the mathematical tools for handling these unclear boundaries. For a fuzzy subset 'a' representing the range of values ​​for a certain feature parameter 'x', the membership function 'μa(x)' and its specific value 'membership degree' 'μa(x)' (0≤μ≤1) are used to describe the membership degree to 'a'. When μa(x) = 0.1 or 1, the feature parameter is in a strictly normal or abnormal state; when μa(x) > 0.4, a pre-alarm signal is issued; when μa(x) > 0.5, a danger alarm signal is issued; when μa(x) > 0.8, an abnormal alarm signal is issued. 5. Conclusion Fault diagnosis is very complex. This paper, using a power plant boiler as an example, omits many technical details and briefly discusses a diagnostic method based on fuzzy neural networks. Preliminary experimental results have shown that this method is reasonable and effective.