Induced draft fan intelligent early warning system based on ICS and diagnosis method

Abstract

The invention relates to an intelligent early warning system and method for an induced draft fan of an ICS, and the system comprises a DCS system and a big data platform, the big data platform carries out the information interaction with an intelligent early warning module and an intelligent diagnosis module, and the intelligent early warning module transmits the early warning information to the intelligent diagnosis module; the intelligent early warning module is in signal connection with the big data platform through the data model module, the big data platform transmits data to the data model module, and the data model module is in signal connection with the intelligent early warning module; the intelligent diagnosis module is in signal connection with the big data platform through the mechanism model module, the big data platform transmits data to the mechanism model module, and the mechanism model module is in signal connection with the intelligent diagnosis module; and the mechanism model module is in signal connection with the data model module through the simulation working condition generation and accident deduction module. The method has the advantages that prediction can be carried out in advance, and loss is effectively reduced and controlled.

Description

Technical Field

[0001] The present invention relates to the technical field of early warning and diagnosis of induced draft fans, and in particular to an ICS-based intelligent early warning system and a diagnosis method for induced draft fans. Background Art

[0002] With the in-depth application of advanced concepts and cutting-edge technologies such as industrial Internet, artificial intelligence, big data, and smart hardware in the field of industrial control, the underlying control system of coal-fired thermal power plants has adapted to the development of the times, upgrading from traditional monitoring and control-based DCS software to an intelligent power generation operation control operating system ICS that integrates monitoring, analysis, diagnosis, alarm, control, and optimization. Alarm is the core function for centralized control operators to monitor the operating status of generator sets. In traditional DCS, alarm only implements simple alarm functions based on upper and lower limit violations, resulting in a large number of defects or faults being missed or reported late, which brings great safety risks to the centralized control operation of thermal power plants.

[0003] Statistics show that the shutdown failures of thermal power plant units caused by large rotating equipment such as fans and feed water pumps account for 30% to 50% of the shutdown failures of units, and the unplanned shutdown caused by fan failure is one of the main reasons for unplanned shutdown caused by auxiliary machine failure. Fans and boiler feed water pumps are listed as key equipment that cause failures together with generators and steam turbines. The role of the induced draft fan is to transport flue gas with large impurities and high temperature. The working environment is harsh and the frequency of failure is relatively high. Once a failure occurs, it will affect economic benefits and bring economic losses.

[0004] As the capacity of power plants increases, it is increasingly urgent to improve the reliability of wind turbines and reduce power generation losses. Equipment status monitoring and fault warning systems can detect early signs of equipment failures, allowing relevant personnel of power generation companies to have sufficient time to analyze and diagnose, so that maintenance plans can be formulated before the failure occurs, and corresponding measures can be taken to curb the continued deterioration of equipment operation and eliminate equipment failures in the bud. Fault warning can achieve pre-emptive maintenance, and the corrective measures adopted are proactive and forward-looking. Predictive analysis of faults is carried out, passive maintenance is changed to active maintenance, and unplanned downtime is changed to planned downtime, thereby bringing good economic and social benefits to power generation companies. Summary of the invention

[0005] 1. Technical Problems Solved

[0006] The technical problem to be solved by the present invention is that currently the failure of the induced draft fan cannot be predicted in advance and the control loss cannot be effectively reduced.

[0007] 2. Technical Solution

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: an ICS-based intelligent early warning system for induced draft fans, including a DCS system and a big data platform.

[0009] The big data platform interacts with the intelligent early warning module and the intelligent diagnosis module, and the intelligent early warning module transmits the early warning information to the intelligent diagnosis module;

[0010] The intelligent early warning module is connected to the big data platform signal through the data model module, the big data platform transmits data to the data model module, and the data model module is connected to the intelligent early warning module signal;

[0011] The intelligent diagnosis module is connected to the big data platform signal through the mechanism model module, and the big data platform transmits data to the mechanism model module, and the mechanism model module is connected to the intelligent diagnosis module signal;

[0012] The mechanism model module is connected with the data model module signal through the simulation condition generation and accident deduction module.

[0013] Furthermore, the intelligent early warning module obtains data from the big data platform in real time and performs real-time calculations, and the intelligent diagnosis module is connected to the big data platform signal through a configurable diagnosis algorithm.

[0014] Furthermore, the mechanism model module transmits data to the simulation condition generation and accident deduction module, and the simulation condition generation and accident deduction module transmits data to the data model module.

[0015] The diagnostic method of the induced draft fan intelligent early warning system based on ICS includes the following steps:

[0016] S1, data collection, collects historical data from a power plant SIS historical database for the past two years, with an interval of 1 minute;

[0017] S2. Data cleaning, which mainly includes:

[0018] S21, missing value processing,

[0019] S22, outlier processing,

[0020] S23, duplicate value processing,

[0021] S24, data standardization and normalization,

[0022] S25. Ensure format consistency.

[0023] S26. Confirm that the data complies with business logic and does not exceed the reasonable range;

[0024] S3. Determine the model input and output, which mainly includes:

[0025] S31. Establishment of LSTM equipment warning model.

[0026] S32, train the model through the loss function,

[0027] S33, optimization algorithm,

[0028] S34, activation function;

[0029] S4. Fault diagnosis: obtain relevant representations of induced draft fan faults through the shaft system dynamics model, extract fault characteristics of hydropower units, integrate different fault models, summarize the mechanisms of typical faults such as comprehensive misalignment of the hydropower unit shaft system, rotor mass imbalance, oil film vortex, rotor-stator friction, main shaft misalignment, mass imbalance, etc., and obtain unit fault mechanism knowledge including fault number, fault source, fault vibration characteristics and fault mechanism.

[0030] Furthermore, the missing value processing in S21 specifically includes a deletion method, whereby columns with many missing values ​​and no importance are directly deleted;

[0031] Imputation method, using mean, median, mode or other statistical methods to fill missing values, or using machine learning algorithms to predict and fill.

[0032] Furthermore, the abnormal value processing in S22 specifically includes:

[0033] Delete outliers. For some obviously wrong data points, delete them directly.

[0034] Replace outliers, replace outliers with reasonable values, use the median or interpolate;

[0035] Transforming variables, such as logarithmic transformation and square root transformation, can reduce the impact of outliers.

[0036] Furthermore, the S23 duplicate value processing specifically includes deleting duplicate records and removing duplicates based on unique identifiers or key fields.

[0037] Furthermore, the S24 data standardization and normalization specifically include:

[0038] Standardization, transforming the data into a standard normal distribution;

[0039] Normalization scales the data to a fixed interval.

[0040] Furthermore, the S25 ensuring format consistency specifically includes ensuring consistency of data types, unified date formats, and consistent upper and lower case of character strings.

[0041] 3. Beneficial Effects

[0042] The advantages of the present invention compared with the prior art are:

[0043] 1. Based on complex advanced algorithms or multiple mechanism models, develop intelligent alarm and early warning modules to achieve all-round coverage of equipment defects, early warning, and root cause tracing, which not only reduces the monitoring burden of front-line operators, but also helps equipment maintenance personnel to quickly locate defects, isolate and repair equipment in advance, and avoid deterioration of working conditions and equipment status;

[0044] 2. Based on the ICS, an intelligent power generation platform for coal-fired power plants, build an intelligent alarm module, develop intelligent alarm applications based on slope change algorithm, equipment health scoring algorithm, expected value deviation algorithm based on big data learning, neural network algorithm, logical fault tree and expert knowledge base diagnosis, and physical mechanism model, improve the unit's ability to detect defects and warn of faults in advance, and improve the active safety of thermal power generation;

[0045] 3. Development and application of the latest intelligent diagnosis and alarm technology of the ICS platform, the underlying intelligent operating system of coal-fired thermal power plants. By matching different complex operating conditions with multiple models, the active safety capabilities of coal-fired power generation units for rapid intelligent diagnosis, early warning, and accurate alarm can be improved, prompting and guiding operators in operation in a timely manner to avoid further deterioration of operating conditions, thus achieving safe, efficient, and intelligent underlying control;

[0046] 4. The integrated hierarchical display design of DCS realizes the seamless integration of monitoring and early warning information with the DCS system through embedded and alarm linkage. The monitoring screen carousel function reduces the workload of operators who constantly flip through the screen during the monitoring process. The alarm processing realizes closed-loop management, effectively reducing the risk of omissions and misoperation. No additional learning cost is required, and it is easy to view and help make decisions;

[0047] 5. The connection with the DCS system does not require the configuration of complex interfaces. The required measurement point data, hardware points, intermediate points, algorithm points, measurement point quality information, DCS alarm information, etc. can be obtained from the DCS system simply and quickly. The speed and accuracy of data extraction are significantly superior. The platform runs reliably and stably and is easy to maintain.

[0048] 6. The data analysis software supports multiple data interfaces, facilitates online reading of DCS historical data, and has efficient data cleaning and quality inspection capabilities. Use simple visual operations to complete the full process of machine learning modeling from training, verification, and tuning. Model parameter thresholds can be automatically generated, making data analysis and model building simple and efficient;

[0049] 7. Adopt multi-state recognition algorithm to solve the abnormal and fault identification problems of equipment under multi-parameter coupling state, significantly improve fault prediction ability, reduce unexpected equipment downtime, and reduce maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a system block diagram of the present invention;

[0051] Figure 2 This is a data driven model diagram of the present invention;

[0052] Figure 3 This is a schematic diagram of the LSTM structure of the present invention;

[0053] Figure 4 Schematic diagram of the model structure of the multi-hidden layer LSTM of the present invention;

[0054] Figure 5 are the main parameters and dimensions of the LSTM model of the present invention;

[0055] Figure 6 The structure and parameter analysis technology flow chart of the LSTM device early warning model of the present invention;

[0056] Figure 7 It is the state of the LSTM model under different hyperparameter epochs of the present invention;

[0057] Figure 8 This is a schematic diagram of the structure of the human-computer interaction module of the present invention;

[0058] Fig. 9 This is a flow chart of the fault classification / repair suggestion / prediction of the present invention. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] It should be noted that all directional indications such as up, down, left, right, front, back, etc. in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the components in a certain specific posture as shown in the accompanying drawings. If the specific posture changes, the directional indication will also change accordingly.

[0061] In addition, in the present invention, descriptions such as "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0062] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the invention.

[0063] The present invention will now be further described with reference to the accompanying drawings.

[0064] Example

[0065] Generally speaking, the equipment will go through three stages from normal to failure, such as Figure 1 As shown. Only by giving early warning and repairing the equipment when the equipment is in a deteriorating state can the greatest help be provided to the operating personnel and the operating efficiency of the unit be improved. To this end, the software gives the operating personnel a first-level warning when the equipment begins to deteriorate. If the operating personnel have not paid attention, a second-level warning will be issued at the end of the equipment degradation stage to remind the operating personnel again and help them make judgments. In order to achieve degradation warning from the data level, the real-time data input by the model must be real-time, efficient, and stable. The real-time database can meet these requirements well. The historical database accumulates a large amount of thermal process data and connects to the offline modeling module to ensure the integrity of the data.

[0066] Data-driven models

[0067] Data collection

[0068] The historical data of the past two years are collected from a power plant SIS historical database with an interval of 1 minute.

[0069] Data cleaning

[0070] Data cleaning is a very important step in the data analysis process, which mainly includes the following steps:

[0071] Missing value handling:

[0072] Deletion method: For columns with many missing values ​​and unimportant values, they can be deleted directly.

[0073] Imputation method: Use mean, median, mode or other statistical methods to fill missing values, or use machine learning algorithms to predict and fill.

[0074] Outlier handling:

[0075] Delete outliers: Some obviously erroneous data points can be deleted directly.

[0076] Replace outliers: Replace outliers with reasonable values, such as using the median or interpolation.

[0077] Transformation variables: Logarithmic transformation, square root transformation, etc. can reduce the impact of outliers.

[0078] Duplicate value handling:

[0079] De-duplicate records: Remove duplicate records based on unique identifiers or key fields.

[0080] Data standardization and normalization:

[0081] Standardization: Transform the data into a standard normal distribution (mean 0 and standard deviation 1).

[0082] Normalization: Scaling the data to a fixed range (usually 0 to 1).

[0083] Format consistency:

[0084] Ensure that data types are consistent, date formats are unified, string case is consistent, etc.

[0085] Data verification:

[0086] Confirm that the data complies with business logic and does not exceed a reasonable range of values.

[0087] Almost all data points (99.7%) fall within three standard deviations of the mean.

[0088] Model input and output determination:

[0089] The parameters at a certain moment are determined by the relevant parameters of the previous few minutes. The bearing temperature at a certain moment is related to the bearing temperature of the previous few minutes, and also to the temperature of the cooling medium in the previous few minutes, thus establishing a supervised learning regression machine learning model.

[0090] The characteristics of intelligent alarm based on neural network algorithm are: ① There are multiple input variables and multiple output variables, and there is a direct interaction relationship between the input variables and the output variables; ② The functional relationship between multiple independent variables and multiple dependent variables is complex, and the function curve cannot be obtained through simple big data analysis or data statistical regression fitting; ③ Through multi-layer neural networks, the relationship between multiple input variables and multiple output variables can be forcibly matched to obtain the data characteristics of the normal state. Once the deviation between the actual output variable and the ideal output variable exceeds the limit, an alarm of abnormal equipment status is issued.

[0091] Data-driven machine learning models are applied to equipment early warning. Equipment model prediction and early warning can be attributed to time series prediction problems from an algorithmic perspective. The application of machine learning models to equipment early warning converts the equipment parameter time series prediction problem into a supervised learning regression problem, that is, establishing the relationship between the input and output of equipment parameters, thereby realizing the output prediction of equipment parameters.

[0092] The outstanding performance of LSTM in time series analysis tasks is due to its special internal control gate and cell state structure. The cell records the internal state information of the model at different times, while the input gate and forget gate control the preservation and forgetting of the current input information and historical state of the cell, respectively, and the output gate controls the output of the current cell state information. The three control gates and cell state enable the model to better mine the important information and input-output nonlinear mapping relationship in the dynamic change process of the time series. The LSTM cell state is also similar to the equipment parameter state, and LSTM is connected in a recursive form at different times. The state of the model at the current moment depends not only on the input at the current moment, but also on the model state at the previous moment, which is similar to the production and operation of the equipment. The equipment operation process is a highly nonlinear dynamic process with complex relationships between parameters and rapid response. Relying on the recursive connection of LSTM and special control gates and cell states, LSTM is used to deal with equipment warning problems with strong nonlinearity and high complexity, which has great application potential.

[0093] Combined with the actual SIS data, the recursive model can be regarded as a device parameter simulation. After the model state is initialized, the model state is constantly changing with the input of new information at each subsequent moment, and the current state depends not only on the input at the current moment, but also on the model state at the previous moment. This is similar to the operating characteristics of the equipment, that is, the current state of the equipment is not only related to the relevant parameters at the current moment, but also to the state of the equipment at the previous moment.

[0094] The cell state can be regarded as an independent information memory unit in LSTM, which saves important information input at different times of the model, so that LSTM has the ability to store and access historical information for a long time. The three control gates of LSTM use the input information xt at the current moment and the hidden state ht-1 at the previous moment as input data. The outputs of the three control gates further act on the cell state, respectively controlling the forgetting, preservation, updating and output of the cell state for the current input information (xt) and historical information (ht-1, Ct-1), thereby realizing the long and short-term memory function of LSTM.

[0095] The structure of the LSTM device warning model mainly includes the model input and output structure, the number of hidden layers (single or multi-layer) and the number of neuron nodes in the hidden layer. Among them, the input and output structure of the model determines the horizontal depth of the model, while the number of hidden layers and the number of neuron nodes in the hidden layer determine the vertical complexity of the model.

[0096] The equipment early warning model has a multi-input and multi-output structure.

[0097] The number of LSTM hidden layers determines the vertical depth of the model. Generally, the more hidden layers there are, the more complex the model network is, and the stronger the ability to express complex problems and data sets is. The internal connection structure of the multi-hidden layer LSTM model is: the input layer is connected to the first hidden layer, the hidden state ht output by the first hidden layer at each moment is used as the input of the second hidden layer, and the remaining hidden layers are stacked and connected in sequence, and finally the Hth hidden layer is connected to the output layer. The number of hidden layers H is an important parameter of LSTM. The value of H affects the scale of the internal weights and bias parameters of the LSTM device warning model. Adding hidden layers will cause the parameter scale to grow rapidly.

[0098] The parameters of the LSTM device warning model refer to the weights and bias parameters (scalars) involved in the interconnection between the input layer, hidden layer, and output layer. They are mainly used for matrix and vector operations between the three control gates of LSTM, the cell unit state, and the model input and output, and each step of the operation is performed according to strict mathematical equations and calculation processes. The parameters of the LSTM device warning model are dimensionless coefficients, which usually only have mathematical significance, and it is generally difficult to analyze the physical meaning of the parameters. The size of the LSTM device warning model parameters mainly depends on the model input feature dimension (m) and the number of output values ​​at each time step, the number of hidden layers, and the number of neuron nodes contained in each hidden layer (u).

[0099] The parameters of the LSTM model are determined by optimization methods. The parameter optimization process of a machine learning model is usually called model training. The model determines the value of the parameters by repeatedly learning the potential relationship between the input data and the output target value in the training sample. The LSTM model uses the Back Propagation Through Time (BPTT) algorithm for supervised learning. Model construction, training, and testing can be implemented based on the open source TensorFlow and Keras deep learning frameworks. According to the call of the LSTM model under the Keras deep learning framework, the factors that affect the determination of model parameters during model training include hyperparameters, loss function, optimization algorithm, activation function, sample set segmentation ratio and its representativeness, etc. These factors are all external configurations of the model.

[0100] 1) Hyperparameters

[0101] Hyperparameters are parameters set before the model starts training. They mainly control the model training process. The value of hyperparameters cannot be estimated from the training sample data. They usually need to be set based on experience or determined through repeated trials in combination with actual problems. The hyperparameters of the LSTM model mainly include epoch and batch size.

[0102] An epoch means that each sample in the training set has participated in one model training process (forward and back propagation calculations). The value of epoch directly affects the fitting state of the model after training, that is, underfitting or overfitting. Underfitting means that the model fails to learn the mapping relationship between input and output data and cannot fit the input and output data well; overfitting means that the model is over-trained, and the simulation effect on the known training set data is good, but the performance in the test set data is poor, and the model generalization ability is weak. Batch size refers to the number of samples used in each training process of the model. Each time, the batch size samples are randomly selected from the training set samples using the method of sampling without replacement to calculate the error and back-propagate the sample error to update the parameters. The number of samples selected indirectly implies the influence of the representativeness of the input data on the model parameter optimization process. In addition, the hyperparameters epoch and batch size of the LSTM model are not independent of each other. They jointly determine the training status and time consumption of the model. Changing the values ​​of epoch and batch size actually affects the number of times and combination of each training sample in the model training phase. Any change in any hyperparameter will change the training process of the model and affect the optimization results of parameters such as the model weight matrix and bias vector.

[0103] 2) Loss Function

[0104] The loss function is a function used to calculate the sample error during the model training phase, that is, the degree of fit or proximity between the model output (calculated) value and the target (true) value of a single sample. The loss function is also a core element of the machine learning model. It is essentially used to measure the simulation performance of the model, evaluate the current fitting effect of the model, and provide direction for further optimization of the model parameters. The training of the machine learning device warning model is a process of optimizing it with the loss function as the objective function, so that the deviation between the model output value of the training set sample and the corresponding true value is minimized as a whole. At present, the commonly used loss function for machine learning models applied to supervised learning regression tasks is the mean absolute error (MAE) or mean square error (MSE), and the calculation formula is as follows:

[0105]

[0106] In the above formula, l represents the number of model output values; y k，out ,y k，obs They represent the kth output value of the model and the corresponding target value respectively.

[0107] 3) Optimization algorithm

[0108] The machine learning model optimizes the loss function as the objective function, thereby converting the problem to be solved by the machine learning model into an optimization problem for solution. A good optimization algorithm can significantly improve the training and learning speed of the model, and accelerate the convergence speed and effect of the model. Gradient Descent (GD) is an earlier search-based optimization method used in neural network models. It mainly seeks the minimum value of the loss function or converges to the minimum value through continuous iteration. The most obvious feature of the GD algorithm is that all training samples are used for gradient update calculation in each iteration. Using all training samples can make the model more accurately approach the direction of the global optimal solution, but the model training efficiency is very low under large-scale samples. Stochastic Gradient Descent (SGD) is an improved gradient descent method. It uses one sample in each iteration and can quickly improve the model training speed. Although not every iterative calculation makes the model move towards the global optimal direction, the overall optimization direction is usually still towards the global optimal solution. The Adaptive Moment Estimation (Adam) algorithm is an extended version of the stochastic gradient descent algorithm. The algorithm has high computational efficiency and low memory requirements, and is suitable for calculations of large data sets and high-dimensional feature spaces. The learning rate is also an important parameter in gradient descent optimization algorithms such as SGD and Adam, which affects the speed of gradient descent during model training. The optimization algorithms for neural network models are not limited to the above three. Other methods include Newton's method, conjugate gradient method, and heuristic optimization methods.

[0109] 4) Activation Function

[0110] The activation function is a calculation function of the neuron node in the artificial neural network. It mainly maps the input of the neuron to the output. Selecting different activation functions will affect the convergence speed and effect of the model. The activation function essentially introduces nonlinear factors into the neural network, so that the neural network model can arbitrarily approximate any nonlinear function and improve the nonlinear fitting ability of the model. According to the introduction to the basic principles of LSTM in the previous article, the activation function of the hidden layer neuron node in LSTM can only use sigmoid and tanh functions, while the activation function of the model output layer can select appropriate functions according to different task requirements.

[0111] The number of hidden layers and the number of neuron nodes in the model structure will affect the scale of model parameters, that is, the model structure and parameters are not completely independent. At the same time, there are many external factors (hyperparameters, optimization algorithms, etc.) that affect the determination of model parameters during model training, which adds greater uncertainty to the construction of LSTM equipment warning models. Therefore, building an LSTM equipment warning model is an extremely complex and systematic task, involving model structure, model parameters, and many external factors that affect parameter determination.

[0112] The structure of the LSTM equipment warning model mainly depends on the input and output structure of the model, the number of hidden layers and the number of neuron nodes. Among them, designing the model input and output structure is the key to determining the LSTM model structure, which basically determines the rationality and practicality of the LSTM equipment warning model. First determine the input and output structure of the model: the output is the equipment parameters in different time periods, and the input is the associated variables that affect the equipment parameters.

[0113] The number of hidden layers in a model determines the vertical depth of the model. Increasing the number of hidden layers can expand the scale of model parameters and improve the model's ability to express large-scale and complex data, but it also increases the risk of overfitting and computational complexity of the model. At present, there is no clear method or calculation formula for determining the number of hidden layers. In practical applications, it is usually determined through repeated experiments based on different tasks and specific problems. For simpler tasks or data sets, it is usually sufficient to set up one or two hidden layers. Deeper layers are not only prone to overfitting problems, but also significantly increase the difficulty of training the model and the time required for calculation. However, for complex and massive data sets such as computer vision and expert systems, single-layer or double-layer shallow models usually cannot achieve good results, so more layers are needed, that is, to build a deep model to improve the model's ability to express complex data.

[0114] The number of neuron nodes largely determines the complexity of the internal connection network of the model and the number of model parameters. The input layer, hidden layer and output layer of the model all contain several neuron nodes. Among them, the number of neuron nodes in the input layer and the output layer is relatively simple to determine, which is equal to the number of input features and the number of output values, respectively. However, how to determine the number of neuron nodes in the hidden layer is still an unresolved problem. In practical applications, it is usually determined by repeated adjustment using a trial and error method based on the complexity of the problem and the size of the data set. When dealing with extremely complex problems with large amounts of data and high nonlinear intensity, the number of neuron nodes in the hidden layer of the model should be increased to improve the model's expressiveness. For simpler problems, using a few nodes can also achieve better results.

[0115] On the basis of determining the model structure, the model repeatedly learns the sample data of the training set based on the training process, and finally determines the values ​​of the model weight matrix and bias vector parameters. The LSTM equipment warning model is constructed and trained under the existing Keras deep learning framework. The determination of model parameters mainly involves setting the hyperparameters, loss functions, optimization algorithms and activation functions involved in the model training process.

[0116] The values ​​of the hyperparameters epoch and batch size directly affect the model's fitting (simulation) and extrapolation (prediction) performance. A machine learning model with good performance can not only accurately simulate known data (training set), but also perform well on unknown data (test set). When the epoch value is small, the LSTM model is not fully trained, resulting in underfitting of the trained model, and weak simulation and prediction capabilities; when the epoch is large, the model repeatedly learns the training set samples excessively, and is prone to overfitting, resulting in insufficient generalization ability, that is, it simulates the known training set data very well, but performs poorly on the test sample set. The figure shows a schematic diagram of the LSTM model training status under different epoch values. The value of batch size is between 1 and the total number of training set samples. When batch size is equal to 1, one sample is selected as input each time during model training. This selection method can fully consider the particularity of each sample, but it will make the model difficult to converge and the generalization ability of the model is relatively poor. When batch size is equal to the number of training set samples, all training samples are used as input each time during model training. The amount of calculation of the model is large each time, and as the number of training sample sets increases, the computing and storage performance of the hardware is required to be higher, which makes it difficult to apply to large data sets. Therefore, in actual tasks, model training usually uses batch size samples as input. In a large number of existing engineering applications, there is no clear principle and basis for setting the hyperparameters of machine learning models. It is mainly to continuously summarize the debugging experience of the model in solving specific problems and use trial algorithms to determine the optimal hyperparameter combination.

[0117] In supervised learning regression tasks of machine learning models, MAE and MSE are commonly used loss functions in the training stage of machine learning models. Compared with the MAE loss function, using the MSE loss function for model training will give greater weight to outliers. The two loss functions can be used as a benchmark for mutual comparison, and they need to be determined comprehensively based on actual needs and effects in different applications.

[0118] Among the numerous optimization algorithms for machine learning models, the Adam algorithm is currently a very popular algorithm in the field of machine learning, and is even recommended as the default algorithm for model training. This algorithm has high computational efficiency and low memory requirements. Since its proposal, it has been widely used in the training process of machine learning models, especially in practical tasks such as computer vision, image recognition, and natural language processing. It has achieved good application results and has better performance than other random optimization algorithms. In addition, when the learning rate of the Adam algorithm is small, the model converges too slowly; and if the learning rate is too large, the loss function curve will oscillate, and the model will be difficult to converge. In practical applications, the value of the learning rate is also mainly determined by repeated attempts and based on the performance of the model. It can usually be set to the default value of 0.001. The default value is generally determined based on the calculation principle of the optimization algorithm and its performance in different practical tasks.

[0119] The activation function of the LSTM device warning model is mainly determined for the model output layer. Based on the basic principles and internal calculation process of LSTM, the sigmoid function is selected as the activation function.

[0120] Troubleshooting

[0121] Through the shaft system dynamics model, the relevant representations of the induced draft fan fault are obtained, the fault characteristics of the hydropower unit are extracted, and different fault models are integrated to summarize the mechanisms of typical faults such as comprehensive misalignment of the hydropower unit shaft system, rotor mass imbalance, oil film vortex, rotor-stator friction, main shaft misalignment, mass imbalance, etc., and obtain the unit fault mechanism knowledge including fault number, fault source, fault vibration characteristics and fault mechanism.

[0122] The initial knowledge base of typical fault mechanism of induced draft fan, which includes fault type, fault time domain and frequency domain characteristics, fault mechanism and other fault knowledge, is constructed to store the inherent fixed information of typical fault of induced draft fan. The fault vibration characteristics in the typical fault mechanism knowledge base mainly include the description of the time domain waveform, axis trajectory, vibration frequency, spectrum diagram, phase, vibration direction and speed of typical fault; the fault mechanism knowledge should include the fault cause and fault process when different faults occur. The initial knowledge base of typical fault mechanism of induced draft fan adopts the storage method of "framework + rule" to store the fault mechanism knowledge according to the division of vibration fault. The framework is expressed as the top event or intermediate event in the fault tree structure. The rule refers to the rule after attribute simplification. The fault mechanism knowledge base includes the framework table, rule table, framework rule table, fault fact table, vibration feature table, fault mechanism table and fault symptom table and other related contents. The knowledge of typical fault mechanism of induced draft fan is stored in the initial knowledge base of typical fault mechanism of unit in the form of dictionary for query and use, so as to solve the deficiencies in the knowledge base construction of induced draft fan fault mechanism.

[0123] The invention and its implementation methods are described, and such description is not restrictive. The drawings show only one implementation method of the invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and implementation method similar to the technical solution without creativity, which should fall within the protection scope of the invention.

Claims

1. The ICS-based intelligent early warning system for induced draft fans includes a DCS system and a big data platform, which is characterized by: The big data platform interacts with the intelligent early warning module and the intelligent diagnosis module, and the intelligent early warning module transmits the early warning information to the intelligent diagnosis module; The intelligent early warning module is connected to the big data platform signal through the data model module, the big data platform transmits data to the data model module, and the data model module is connected to the intelligent early warning module signal; The intelligent diagnosis module is connected to the big data platform signal through the mechanism model module, and the big data platform transmits data to the mechanism model module, and the mechanism model module is connected to the intelligent diagnosis module signal; The mechanism model module is connected with the data model module signal through the simulation condition generation and accident deduction module.

2. The ICS-based induced draft fan intelligent early warning system according to claim 1 is characterized in that: The intelligent early warning module obtains data from the big data platform in real time and performs real-time calculations, and the intelligent diagnosis module is connected to the big data platform signal through a configurable diagnosis algorithm.

3. The ICS-based induced draft fan intelligent early warning system according to claim 1 is characterized in that: The mechanism model module transmits data to the simulation condition generation and accident deduction module, and the simulation condition generation and accident deduction module transmits data to the data model module.

4. The diagnostic method of the ICS-based induced draft fan intelligent early warning system is characterized by: The following steps are involved: S1, data collection, collects historical data from a power plant SIS historical database for the past two years, with an interval of 1 minute; S2. Data cleaning, which mainly includes: S21, missing value processing, S22, outlier processing, S23, duplicate value processing, S24, data standardization and normalization, S25. Ensure format consistency. S26. Confirm that the data complies with business logic and does not exceed the reasonable range; S3. Determine the model input and output, which mainly includes: S31. Establishment of LSTM equipment warning model. S32, train the model through the loss function, S33, optimization algorithm, S34, activation function; S4. Fault diagnosis: obtain relevant representations of induced draft fan faults through the shaft system dynamics model, extract fault characteristics of hydropower units, integrate different fault models, summarize the mechanisms of typical faults such as comprehensive misalignment of the hydropower unit shaft system, rotor mass imbalance, oil film vortex, rotor-stator friction, main shaft misalignment, mass imbalance, etc., and obtain unit fault mechanism knowledge including fault number, fault source, fault vibration characteristics and fault mechanism.

5. The diagnostic method of the ICS-based induced draft fan intelligent early warning system according to claim 4 is characterized in that: The missing value processing in S21 specifically includes a deletion method, where columns with many missing values ​​and no importance are directly deleted; Imputation method, using mean, median, mode or other statistical methods to fill missing values, or using machine learning algorithms to predict and fill.

6. The diagnostic method of the ICS-based induced draft fan intelligent early warning system according to claim 4 is characterized in that: The abnormal value processing in S22 specifically includes: Delete outliers. For some obviously wrong data points, delete them directly. Replace outliers, replace outliers with reasonable values, use the median or interpolate; Transforming variables, such as logarithmic transformation and square root transformation, can reduce the impact of outliers.

7. The diagnostic method of the ICS-based induced draft fan intelligent early warning system according to claim 4 is characterized in that: The S23 duplicate value processing specifically includes deleting duplicate records and removing duplicates based on unique identifiers or key fields.

8. The diagnostic method of the ICS-based induced draft fan intelligent early warning system according to claim 4 is characterized in that: The S24 data standardization and normalization specifically include: Standardization, transforming the data into a standard normal distribution; Normalization scales the data to a fixed interval.

9. The diagnostic method of the ICS-based induced draft fan intelligent early warning system according to claim 4 is characterized in that: The S25 ensuring format consistency specifically includes ensuring data type consistency, date format uniformity, and character string case consistency.

Images