Research on Fault Diagnosis of Thermal Power System of Thermal Power Unit under Different Loads

Aiming at the problem of difficulty in acquiring fault data under different loads, a domain-adversarial-based fault diagnosis method for convolutional neural networks is established by combining convolutional neural networks with adversarial domain adaptive methods. The method uses convolutional neural networks to extract the fault features of the source and target domain samples, and simultaneously implements joint adjustment of the distributions of the domain-level and class-level features, aligns the feature distributions of the source and target domains by minimizing the difference in the distributions of the two domains through adversarial learning of the domains and aligns the feature distributions of each class of faults by making the distributions of the same kind of features closer through the loss of the central discriminant. With the help of a 600MW supercritical thermal power unit full-working condition simulation system, experiments are carried out under 100%, 90%, 80% and 70% of the rated load, and the results show that the method has effectiveness and superiority in the fault diagnosis of thermal system under different loads.


Diagnostic method for unit load deviation and cause analysis
With the high-quality development of China's economy, the demand for electrical energy increases year by year.At the same time, people's requirements for electric energy safety and reliability have reached an unprecedented high level.The power industry is the main industry that consumes coal, and thermal power is dominant in the power system [1][2][3].Fault diagnosis methods based on neural networks are more widely used in thermal power system fault diagnosis, but most of these methods are focused on a certain number of steady state load conditions, and there are fewer studies on fault The 2nd International Conference on Smart Energy Journal of Physics: Conference Series 2717 (2024) 012029 IOP Publishing doi:10.1088/1742-6596/2717/1/012029 2 diagnosis under different loads.It is of great significance how to solve the problem of thermal system fault diagnosis under different loads by utilizing the marked data under a certain load [2][3][4][5][6].Shuai Yu [7] proposed a qualitative trend-based approach to identify the severity of faults after extracting the fault characterization information of the thermal system, and completed the fault diagnosis of the thermal system by establishing a rule knowledge base and designing a fuzzy trend matching strategy.Zhao Lijuan [8] completed the prediction and diagnosis of the failure characteristics of thermal system in dynamic process with the theoretical basis of algorithms such as extreme learning machine and particle swarm optimization.Migration learning breaks the limitations of traditional deep learning models, and its main idea is that the migration learning network model first learns knowledge from the labeled source domain, and then migrates the acquired knowledge to the unlabeled target domain to classify the target domain data.By narrowing the differences in the distribution of data features in different domains, the migration effect of fault features of thermal systems under different loads is improved, so as to diagnose the faults of thermal systems under different loads and improve the safety factor and economic efficiency of unit operation.

Description of the problem
A Convolutional Domain Adversarial Network (CDAN) fault diagnosis method based on domain adversarial is proposed for thermal systems under different loads.Specifically, the model extracts domain-invariant features through domain adversarial learning to minimize the difference between the feature distributions in the source and target domains, thus aligning the feature distributions in the two domains.Meanwhile, the center discriminant loss is applied to the network model to align the distribution of each class of fault features.The method implements a joint adjustment of the distributions of domain-level and class-level features to improve the classification performance on unknown target domain data, and a large number of fault diagnosis migration experiments are conducted on a 600MW supercritical thermal unit high-pressure heater system dataset to validate the effectiveness of the method.For the thermal system fault diagnosis problem, a source domain dataset

Feature extraction network
The feature extraction network consists of a one-dimensional convolutional neural network with a two-branch structure, where the two branch networks share weights and biases.Specifically, the network consists of nine layers, of which three are convolutional, three are pooling, two are fully connected, and one is a classification output layer.The first eight layers can be regarded as feature extraction layers, and only the last layer is a feature classification layer.As shown in Table 1, the two branches of the one-dimensional CNN have the same parameter settings, which can halve the training parameters and is more helpful in extracting the fused features of the input data from both domains.For the input sample x the probability of each classification result is estimated.Assuming that the function will output a k-dimensional vector to represent the estimated probability value, the function is represented as follows: ( ; , , , k     is the parameter to be solved for the model, ( ) for normalization of probability distribution, Thus all probabilities sum to 1.The loss values are defined using the cross-entropy loss function, which is expressed as: n denotes the batch size of the sample input and k is the fault category.The optimization function for data classification in the source domain is defined as: min ( ) H  is the cross-entropy loss of the softmax-layer, ( , ) is the distribution of samples and labels in the source domain, C  is the classifier parameter.

Central discernment loss
The central loss was initially applied in the field of face recognition, where it is a simple algorithm while making the model easier to optimize.The algorithm combines the advantages of clustering loss and softmax loss by constraining intra-class distances with center loss and inter-class distances with softmax loss.
For samples in the source domain with labeling, the following loss will be used to cluster features belonging to the same class: Since the samples in the target domain do not have labels, in this experiment the source domain classifier is utilized to predict the labels for the samples in the target domain and generate pseudo-labels, and then the samples with the same labels in the two domains are treated as the same class with the loss defined as follows: The objective function of the CDAN model consists of three main parts: 1) Categorical loss on the source domai ( S L );2) domain discriminator D of the domain discriminator loss D L ;3) The central discriminant losses CS L and CT L for each class of fault samples in the source and target domains .
The construction yields an overall objective function for the model:  , 1  and 2  are weighting parameters.

Dataset and experimental setup
In this experiment, the high-pressure heater system of the thermal system is taken as the research object, and the experiment is carried out with the help of 600MW supercritical thermal power unit simulation system.The system structure of the high-pressure heater is shown in Fig. 1.The three high-pressure heaters are called No. 1, No. 2 and No. 3 high-pressure heaters according to the order of the pumping pressure at the pumping port from high to low.

Figure 1.
Schematic structure of high pressure heater system The data obtained through the simulation system included normal conditions, leakage faults in high pressure heaters #1, #2 and #3, and short circuit faults in the inlet and outlet chambers of high pressure heaters #1, #2 and #3.For leakage faults and short-circuit faults, there are three different levels of severity: (Low, L), (Moderate, M), and (High, H), corresponding to 2%, 4%, and 6% leakage faults and 10%, 20%, and 40% short-circuit faults in the inlet and outlet chambers, respectively.Each of these severity settings ensures significant changes in its associated process data and invariance of the system topology.Therefore, the dataset has a total of seven health states, denoted by F0 to F6, as shown in Table 2.All samples were obtained at 100% to 70% of the rated load (600-MW, 540-MW, 480-MW, and 420-MW), and the studied operating loads can be reached automatically by a coordinated control system.The data set consists of data from the above four load conditions, and the sampling frequency in the experiments was 1 HZ The experiment collects 500 samples for each of the three severity levels of the seven types of faults at four rated loads, where each sample contains 39 process parameters, including pressure, temperature, flow rate, valve opening, and so on.For migration experiments for each health state, the training set consists of all samples from the source domain and half of the samples from the target domain, and the other half of the samples from the target domain serves as the test set.The ReLU activation function is used in the proposed network model, and the learning rates of the classifier and domain discriminator are set to 0.0001 and 0.0001, with coefficients  of 1.0 and 1  and 2  both of 0.02.The Dropout probability was set to 0.3 to avoid overfitting the training set and to improve the performance of the test set.In addition, in order to minimize the effect of data chance, all experiments were repeated 10 times, and the average diagnostic accuracy was taken as the final result to evaluate the performance of the algorithm.

Performance analysis
Performance analysisof troubleshooting under different loads with very different degrees of severity Four load conditions, i.e., 600-MW, 540-MW, 480-MW, and 420-MW, and three fault severity levels, i.e., Low (L), Moderate (M), and High (H), are considered in the experiment.In this case, the data set 600H represents the fault data at 600-MW load level which is severe.The left side of the arrow of each migration task indicates the source domain and the right side indicates the target domain, e.g., "600H→540H" indicates that the source domain dataset is 600H and the target domain dataset is 540H, and the detailed experimental dataset is shown in Table 3.In order to verify the fault diagnosis performance of the CDAN model under different degrees of load, the following two sets of migration diagnosis experiments were conducted:

( 1 )
TaskA：Migration diagnostic experiments to validate the CDAN model under four loads with the same level of failure;(2)TaskB：Migration diagnostic experiments to validate the CDAN model for three different fault severities under the same load conditions.The average diagnostic results for TaskA are shown in Fig.2, with the horizontal coordinate being the 12 migration tasks established between the four load conditions and the vertical coordinate being the fault diagnostic accuracy.From this figure, it can be seen that the CDAN model obtains good performance in all the migration tasks with a minimum of 93.34% and a maximum accuracy of 96.92%, which fully demonstrates that the CDAN fault diagnosis model is able to identify the high pressure feedwater heater faults effectively.In addition, it can be observed through each set of columns that the accuracy of the model's diagnosis increases with the increase of fault severity, which indicates that the higher the fault severity, the more obvious the fault characteristics are, and the features extracted through the model have a stronger discriminability, which makes the diagnosis more accurate.

Table 1 .
CDAN Network Architecture Parameters The proposed model consists of three main modules: feature extraction network, adversarial domain adaptive network, and central discriminative loss.The detailed network parameters of each constituent module are shown in Table1.

Table 2
Status Names and Labels

Table 3 .
Experimental dataset of fault diagnostics with very different degrees under different loads