Thermal power plant safety monitoring method based on federated learning

Abstract

The invention provides a thermal power plant safety monitoring method based on federated learning based on a federated learning algorithm, cavity convolution and an LSTM network, and is used for solving the problems that safety monitoring parameters of different types of thermal power plants are not shared and are not structured. The method comprises the following steps: S1, preprocessing safety monitoring data of a thermal power plant; s2, performing feature extraction on the data preprocessed in the step S1, and establishing a federated learning local model based on cavity convolution and a bidirectional long-short-term memory network; and S3, the server side adopts the federal learning local model established in the step S2 to realize joint training, and thermal power plant safety monitoring is carried out through the trained federal learning local model.

Description

Technical Field

[0001] The present invention relates to the fields of data recognition technology, power system and its automation, and particularly relates to a safety monitoring method for thermal power plants based on federated learning. Background Art

[0002] A thermal power plant, abbreviated as a coal-fired power plant, is a factory that uses combustibles (such as coal) as fuel to produce electric energy. Its basic production process is as follows: when the fuel burns, it heats water to generate steam, converting the chemical energy of the fuel into heat energy. The steam pressure drives the steam turbine to rotate, converting heat energy into mechanical energy. Then the steam turbine drives the generator to rotate, converting mechanical energy into electric energy. Thermal power plants are divided into a combustion system, a steam-water system, and a power generation system. Thermal power plants have the following advantages: flexible layout, and the installed capacity can be determined according to the actual usage size; the construction period is relatively short compared to that of hydroelectric power generation, wind power generation, and photovoltaic power generation; the one-time construction investment cost is small and the return is high. However, there are also some problems with thermal power plants, such as: the production cost of coal is relatively high; there are many power equipment; the investment in pollution control is relatively large; there are many potential safety hazards, etc. China is a large consumer of electric power energy, and currently the main source of electric power energy comes from thermal power generation. Therefore, the safety issues of thermal power generation deserve attention. In addition, in some developing countries, there are only private individual thermal power plants, and the existing potential safety hazards are even more worthy of attention.

[0003] The potential safety hazards of thermal power plants mainly come from three aspects: natural disasters, potential safety hazards caused by human factors or equipment failures, and compatibility problems of equipment updates. Except for natural disasters which are uncontrollable, the other two potential safety hazards can be solved through the training of relevant personnel and big data monitoring means. Among them, using big data technology can not only reduce economic losses, but also effectively store and record historical records, which plays a significant role in subsequent potential hazard prediction. According to the theory of Professor Frank in Germany, the methods of fault prediction are divided into three types: model-based methods, signal processing-based methods, and knowledge-based methods. Using the technical means of big data can effectively establish a data model. The present invention has a large amount of power plant data, so the model-based method is adopted for the safety monitoring of thermal power plants.

[0004] Existing safety monitoring methods for thermal power plants all require obtaining original data, which does not conform to the privacy protection requirements of today's society for individuals, enterprises, and social groups. Since some private thermal power plants are equipped with their own databases and R & D departments, the power generation consumables data and conversion efficiency of private enterprises are privacy contents and should not be publicly disclosed casually at present. In the era of big data, data is a precious wealth. At the same time, with the clear legislation of various countries, no organization or individual is allowed to privately collect users' privacy data. The Chinese patent with the application number 201910098522.2 provides a fault monitoring method, a fault monitoring circuit, and a controller. However, this method only conducts fault monitoring on hardware devices; the Chinese patent with the application number 202111549240.3 proposes a method for training a fault monitoring model of manufacturing equipment based on federated learning. However, in the aggregation of the global model, only the method of weighted average is considered; the Chinese patent with the application number 202011612201.9 proposes an intelligent substation fault prediction method based on the LSTM neural network. However, this method only has a good effect on power data in time series and does not consider problems such as problems caused by the aging of other hardware devices over time. Therefore, the deficiencies of the existing safety monitoring methods for thermal power plants are summarized as follows:

[0005] Problem 1: The traditional safety monitoring of thermal power plants is based on the anomalies of hardware devices, which does not conform to the development of the times. The number and scale of thermal power plants are increasing, and simple hardware monitoring is likely to cause large economic losses. At the same time, regarding the security issues of personal data, the state does not allow random data exchange, and data security issues need to be solved;

[0006] Problem 2: The existing data-based safety monitoring only considers information related to time series and does not involve the safety monitoring of relevant devices, and the problems considered are not comprehensive enough. In the issue of safety monitoring of thermal power plants, not only the faults caused by data anomalies need to be considered, but also the faults caused by device anomalies need to be considered. Summary of the Invention

[0007] Based on the federated learning algorithm, dilated convolution, and LSTM network, the present invention proposes a safety monitoring method for thermal power plants based on federated learning to solve problems such as non-sharing of safety monitoring parameters and unstructured parameters for different types of thermal power plants.

[0008] The technical solution adopted by the present invention to solve the above problems is: A safety monitoring method for thermal power plants based on federated learning, characterized by comprising the following three steps:

[0009] S1: Preprocess the safety monitoring data of thermal power plants;

[0010] S2: Extract features from the data after preprocessing in step S1, and establish a local federated learning model based on dilated convolution and bidirectional long short-term memory network;

[0011] S3: The server side uses the local federated learning model established in step S2 to implement joint training, and conducts safety monitoring of thermal power plants through the trained local federated learning model.

[0012] Step S1 of the present invention includes the following steps:

[0013] S1-1: Make a data set for each thermal power plant as a client based on the obtained data, and divide it into a training set, a test set and a validation set;

[0014] S1-2: According to the relevant characteristic parameters of the thermal power plant obtained, map them into the form of vectors: X = {x1, x2, …, x 20}, where x1 is the output voltage monitoring characteristic parameter, x2 is the output current monitoring characteristic parameter, x3 is the boiler temperature monitoring characteristic parameter, x4 is the ambient humidity monitoring characteristic parameter, x5 is the smoke monitoring characteristic parameter, x6 is the communication monitoring characteristic parameter, x7 is the circuit breaker monitoring characteristic parameter, x8 is the connection joint monitoring characteristic parameter, x9 is the fuse monitoring characteristic parameter, x 10 is the monitoring characteristic parameter; x 11 is the forced draft fan detection, x 12 is the induced draft fan monitoring characteristic parameter, x 13 is the coal mill monitoring characteristic parameter, x 14 is the high and low pressure heater monitoring characteristic parameter, x 15 is the deaerator monitoring characteristic parameter, x 16 is the condenser monitoring characteristic parameter, x 17 is the condensate pump monitoring characteristic parameter, x 18 is the capacitance monitoring characteristic parameter, x 19 is the monitoring characteristic parameter of the screw positions of each device, x 20 is the monitoring characteristic parameter of the output knife switch position;

[0015] S1-3: Process the sample imbalance of the data set;

[0016] S1-4: Normalize the collected data using the maximum and minimum values to obtain the standardized results of the relevant characteristic parameters of the thermal power plant.

[0017] Step S1-3 of the present invention includes the following steps:

[0018] First, use the training set defined in step S1-1, which includes m samples {x i , y i}, i = 1, 2, …, m, where x iis a sample in the n-dimensional feature space X, y i ∈Y = {1, -1}, is the class label associated with x i Define m s as the number of minority class samples, that is, the hardware parameter information of the long-term change of the safety monitoring data of thermal power plants; m l is the number of majority class samples, that is, the real-time change parameter information of the safety monitoring data of thermal power plants; thus, m s ≤ m l , and m s + m l = m; Perform the following calculations:

[0019] ① Calculate the data class imbalance rate d:

[0020]

[0021] ② If d < d th , d th is the set maximum imbalance rate, then calculate the number of minority class samples G that need to be synthesized:

[0022] G = (m l - m s ) × β,

[0023] The parameter β ∈ [0, 1], representing the balance degree required for the processed data set;

[0024] ③ For any minority class sample x i , find K nearest neighbor points according to the Euclidean distance in the n-dimensional space,

[0025] and calculate the ratio r i :

[0026]

[0027] where Δ i is the number of samples belonging to the majority class among the K nearest neighbor points, so r i ∈ [0, 1];

[0028] ④ After normalizing r i get

[0029]

[0030] ⑤ Calculate the number of new samples g i to be generated for each minority class sample x i :

[0031]

[0032] ⑥ Among the K nearest neighbor points of each minority class sample x i select a minority class sample x zi , and generate a new sample s i according to the following rules:

[0033] s i = x i + (x zi - x i ) × λ,

[0034] where λ is a coefficient in [0, 1].

[0035] In steps S1 - 4 of the present invention, the standardization is carried out using the following formula:

[0036]

[0037] where x t is the relevant characteristic parameter of the safety monitoring data of the thermal power plant collected at time t, x max is the maximum value in the collected sample parameters, x min is the minimum value in the collected sample parameters, is the standardization result of the relevant characteristic parameters of the thermal power plant collected at time t.

[0038] Step S2 of the present invention includes the following steps:

[0039] S2 - 1. First, adjust the size of the dilated convolution, and then use the dilated convolution to perform feature selection on the data pre - processed in step S1;

[0040] S2 - 2. The bidirectional long short - term memory network Bi - LSTM is composed of two LSTMs stacked on top of each other, and the two LSTMs are connected to an output layer;

[0041] The forward LSTM takes the left side as the starting input of the series. In text processing, it can be understood as starting from the beginning of the sentence for input, that is, the output of the hidden layer is related to the current input x t and the historical information h tB1 of the hidden layer at the previous moment, and its calculation is carried out using the following formula:

[0042]

[0043] The reverse LSTM takes the right side as the starting input of the series. In text processing, it can be understood as starting from the last word of the sentence for input, that is, the output of the hidden layer is related to the current input x t and the historical information h of the hidden layer at the previous momenttI1 The calculation is carried out using the following formula:

[0044]

[0045] Finally, the two results are spliced ​​and calculated using the following formula:

[0046]

[0047] The output of the hidden layer h t Usually not the final output of the network, but also needs to be Projected to output state o t , which is expressed as follows:

[0048] o t =φ(W h0 h t +b0)

[0049] , where: φ represents the activation function, represents the weight matrix input at the current time t in the forward calculation, b A represents the bias term in the forward calculation process, represents the weight matrix input at the current time t in the reverse calculation, b H represents the bias term in the reverse calculation process, W h0 represents the weight matrix of the output layer input, and b0 represents the output layer bias term.

[0050] S2-3. After the above steps S2-1 to S2-2 are completed, at the output of the Bi-LSTM branch, the information output by each Bi-LSTM branch is fused and output. Specifically, this is achieved using a multilayer perceptron MLP. The multilayer perceptron MLP is a simple three-layer structure of input layer, hidden layer and output layer, where the input is the 1*1 output of 3 Bi-LSTMs, the hidden layer is 10 neurons, and the output is 1 1*1; finally, the output layer of the multilayer perceptron MLP is the local model output, and the local model is now established.

[0051] In step S2-1 of the present invention, the dilated convolution uses the parameter of the dilated rate to adjust the size of the dilated convolution; assuming that the convolution kernel size is 3*1 and the step length is 1, when the dilated rate is 1, the number of filled 0 weights is 0; it can be seen from the following that the mapping range length of the l+1th layer convolution kernel in the lth layer is the same:

[0052]

[0053] The present invention in step S3:

[0054] After updating the local model parameters through the above process, upload them to the terminal server for aggregation to generate a global model;

[0055] The respondent-driven sampling (RDS) method was adjusted. Random walk was used to obtain successor samples. Instead of taking 3-4 successor samples each time, only 1 was taken. The harmonic mean was used for sample estimation, and then the structural characteristics of the network were analyzed using the network average degree as an indicator.

[0056] In step S3 of the present invention, the training steps include the following processes:

[0057] (1) In the t-th round of communication, after the local training of the k-th device is completed, its local model is input into the global server. On the global server, let In each class of the auxiliary dataset aux inference is performed on the data, and the square of the Euclidean norm of the model parameter gradient is calculated as shown in the following formula:

[0058]

[0059] (2) Use the following formula to give an estimation scheme for the class information on the device:

[0060]

[0061] (3) After normalization, the proportion of the c-th class data in the local data of device k is R Z V , calculate the following formula:

[0062]

[0063] Give an intuitive explanation of the class estimation scheme.

[0064] In step S3 of the present invention, in the scenario of class imbalance in federated learning, due to the combinatorial nature of the selection of the device subset in each round of communication, the device selection algorithm is based on the class estimation scheme. Using the aggregated global model in each round of communication and the statistical information of the data class distribution on the current device, it selects the device combination that is most complementary to the deviation degree of the test performance of each class of the global model;

[0065] Based on the device selection, after the next round of communication, the class proportion in the historical training data of the global model will tend to be balanced, thereby improving the performance on the global balanced test set;

[0066] Under the training of the federated learning model, in the t-th round of communication, after the training on the k-th device is completed, the model Return the auxiliary dataset aux on the global server and the class ratio of the training data in the model and the global model class distribution before the previous round of aggregation Using cosine similarity as the evaluation criterion for the balance degree, define the reward for selecting device k as the following formula:

[0067]

[0068] After the selection is completed, perform the steps in the standard FedAvg and update after the training ends and the aggregation is completed.

[0069] In step S3 of the present invention, the mean square error MSE is used as the loss function as an index, as shown in the following formula:

[0070]

[0071] Substituting it in, we get the cost function to be optimized, as follows:

[0072]

[0073] Finally, set an error value to determine the monitoring accuracy of the joint modeling.

[0074] Compared with the prior art, the present invention has the following advantages and effects: On the basis of protecting the privacy of each participating party, the algorithm model of the present invention adopts multi-scale fusion data features, uses dilated convolution for feature extraction on non-temporal feature data, and uses a bidirectional LSTM network for feature extraction on temporal feature data, and finally establishes a local model.

[0075] For problem 1 of the prior art, using federated learning technology can well protect the privacy and security of users. Federated learning takes each thermal power plant as a client and builds a model locally on each client to ensure that the user's data is not leaked; the present invention adopts an improved federated learning algorithm to jointly build a model; when processing local data, a data integration method is used to solve the problem of diverse and inconsistent local data types; to solve the problem of diverse data types among different clients, the present invention improves the federated average aggregation algorithm.

[0076] For problem 2 of the prior art: On the basis of solving problem 1, dilated convolution plus a bidirectional LSTM network is adopted to extract features from the integrated data dataset, build a better local model, and use the model method for the safety monitoring of thermal power plants. On the basis of solving problem 1, it can take into account non-temporal data information for better power plant safety monitoring. Finally, an MLP is used to output the bidirectional LSTM network, and thus the local network model of federated learning is established. Description of the Drawings

[0077] Figure 1 This is the federated learning model diagram of the embodiment of the present invention.

[0078] Figure 2 This is the local model diagram of the client of the embodiment of the present invention.

[0079] Figure 3 This is the schematic diagram of dilated convolution of the embodiment of the present invention.

[0080] Figure 4 This is the Bi-LSTM diagram of the embodiment of the present invention.

[0081] Figure 5 This is the improved federated learning model diagram of the embodiment of the present invention.

[0082] Figure 6 This is the algorithm flow chart of the embodiment of the present invention. Specific implementation manners

[0083] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0084] The method for safety monitoring of thermal power plants based on federated learning according to the embodiment of the present invention includes the following steps:

[0085] S1: Preprocess the safety monitoring data of the thermal power plant. Use historical data to perform local client modeling, and then use the established model to monitor possible fault problems within the next 24 hours. The safety monitoring data of the thermal power plant is now divided into 4 categories: ① Very short-term fault monitoring data, used for monitoring abnormal data within 1 hour; ② Short-term fault monitoring data, used for monitoring abnormal data from one day to one week; ③ Medium-term fault monitoring data, used for monitoring abnormal data from one week to several months; ④ Long-term fault monitoring data, used for monitoring abnormal data from several months to several years. Among them, the data of the third and fourth categories not only includes real-time changing data, but also includes the usage status of hardware devices. Most of the data types of power plants of different enterprises are the same, but the power is different, and there are also some differences in data types. Therefore, it is necessary to preprocess the data, first integrate the collected data to facilitate subsequent processing. This step includes the following steps:

[0086] S1-1. The present invention is directed to the safety monitoring of thermal power plants. Thermal power plants have different production efficiencies in different regions. In China, thermal power plants generally operate 24 hours a day. In countries such as Africa, the operating efficiency of power plants is extremely low. In the actual power generation process, there are significant differences in the coal burned and emission standards in different countries. The safety monitoring of the thermal power plants in the present invention is carried out with the unified 24-hour full-load operation. For the situation where there is insufficient data in some clients, it is uniformly counted as data 0 for supplementation.

[0087] Thermal power plants are divided into 5 categories: small-capacity thermal power plants (total installed capacity less than 100 MW), medium-capacity thermal power plants (total installed capacity 100 MW - 250 MW), medium-large-capacity thermal power plants (installed capacity 250 - 600 MW), large-capacity thermal power plants (total installed capacity 600 - 1000 MW), and extra-large-capacity thermal power plants (total installed capacity greater than 1000 MW). The present invention takes each thermal power plant as a client and collects data once every 10 minutes. 144 data messages can be obtained throughout the day. In order to unify the time resolution of the data of each thermal power plant, the present invention uniformly interpolates the data collection time and collects data once every 5 minutes. 288 data messages can be obtained throughout the day. The obtained data is made into a data set with the thermal power plant as the client. According to different types, there are differences in the data set sizes of each client. For the time being, the data set is classified and made according to the original data.

[0088] After collecting the data, the present invention divides it into three components: a training set, a test set, and a validation set for better training and more accurate evaluation. The D1 data set is used as the training set, the D2 data set is used as the test set, and the D3 data set is used as the validation set. Among them, the validation set D3 is on the server side and is used to verify the accuracy of the global model.

[0089] S1-2. The present invention uses deep learning algorithms to monitor the safety of thermal power plants, so a large amount of data is required. The safety monitoring data of thermal power plants is mainly divided into two parts. One is the real-time changing energy consumption and production data; the other is the data of hardware device changes. Different monitoring parts contain different parameters. The present invention needs to obtain the characteristic parameters of these two parts separately. According to the obtained relevant characteristic parameters of the thermal power plant, the present invention maps them in the form of vectors: X = {x1, x2, …, x 20}, where x1 is the monitoring characteristic parameter of the output voltage, x2 is the monitoring characteristic parameter of the output current, x3 is the monitoring characteristic parameter of the boiler temperature, x4 is the monitoring characteristic parameter of the ambient humidity, x5 is the monitoring characteristic parameter of the smoke, x6 is the monitoring characteristic parameter of the communication, x7 is the monitoring characteristic parameter of the circuit breaker, x8 is the monitoring characteristic parameter of the joint, x9 is the monitoring characteristic parameter of the fuse, x 10 is the monitoring characteristic parameter; x 11For the detection of the forced draft fan, x 12 For the monitoring characteristic parameters of the induced draft fan, x 13 For the monitoring characteristic parameters of the coal mill, x 14 For the monitoring characteristic parameters of the high and low pressure heaters, x 15 For the monitoring characteristic parameters of the deaerator, x 16 For the monitoring characteristic parameters of the condenser, x 17 For the monitoring characteristic parameters of the condensate pump, x 18 For the monitoring characteristic parameters of the capacitor, x 19 For the monitoring characteristic parameters of the screw positions of each device, x 20 For the monitoring characteristic parameters of the output knife switch position, etc., each sample parameter corresponds to a characteristic label y i , i = 1, 2, … 20. These characteristic data are theoretically collected according to step S1-1, but x 19 and x 20 For these 2 hardware characteristic parameters, there will be no change in short-term collection. Changing the collection time to once every 1 hour will cause sample imbalance. Even if the collection time is not changed, the data samples still have imbalance.

[0090] S1-3. Process the imbalance of the dataset samples. Sample imbalance refers to the situation where the number of samples of different classes in the dataset varies greatly and the data class distribution is uneven. In terms of the safety monitoring of thermal power plants, there are many types of relevant data that change in real time and few types of hardware data, so there is sample imbalance. The current processing of dataset sample imbalance mainly proceeds from two aspects: the data level and the algorithm level. There are four methods at the data level: expanding the dataset, data resampling, artificial data samples, and monitoring based on abnormal data; there are three methods at the algorithm level: different classification algorithms, weighted penalty for misclassification of small classes, and reconstructing the classifier. Among them, the most commonly used is data resampling at the data level. Data resampling is divided into three types: undersampling, oversampling, and hybrid sampling. The present invention adopts hybrid sampling, which comprehensively uses oversampling and undersampling methods to balance the number of samples of each class, weakens the disadvantages of using a single sampling method to a certain extent, and is applicable to datasets with different sample sizes to a certain extent. The present invention adopts the adaptive synthetic oversampling ADASYN, which is an improvement based on the minority synthesis oversampling technique SMOTE (Synthetic Minority Oversampling TEchnique) algorithm. This algorithm adaptively generates new samples of the minority class according to the data sample distribution and has good applicability to various data with different distribution types. Its main process includes the following steps:

[0091] First, use the training set D1 defined in step S1-1, which includes m samples {x i , y i}, where \(i = 1, 2, \ldots, m\), and \(x\) i is a sample in the \(n\)-dimensional feature space \(X\), and \(y\) i \(\in Y=\{1, - 1\}\) is the class label associated with \(x\) i . Define \(m\) s as the number of minority class samples, that is, the hardware parameter information of the long-term change of the safety monitoring data of thermal power plants. For example, it can be understood as \(x\) 19 , \(x\) 20 ; \(m\) l is the number of majority class samples, that is, the real-time change parameter information of the safety monitoring data of thermal power plants, that is, the remaining parameters, but these data can be variable. For example, when a certain sensor is damaged, in this way, \(m\) s \(\leq m\) l , and \(m\) s +\(m\) l \(=m\). Perform the following calculations:

[0092] ⑦ Calculate the data class imbalance rate \(d\):

[0093]

[0094] ⑧ If \(d \lt d\) th , \(d\) th is the set maximum imbalance rate, then calculate the number \(G\) of minority class samples to be synthesized:

[0095] \(G=(m\) l -\(m\) s )\(\times\beta\) (2),

[0096] The parameter \(\beta\in[0,1]\), which is manually set and represents the degree of balance required for the processed data set. When \(\beta = 1\) (the default value of the present invention), it means that a completely balanced data set is generated.

[0097] ⑨ For any minority class sample \(x\) i , in the data space, with it as the center, find \(K\) nearest neighbor points according to the Euclidean distance in the \(n\)-dimensional space, and calculate the ratio \(r\) i :

[0098]

[0099] In the formula, \(\Delta\) i is the number of samples belonging to the majority class among the \(K\) nearest neighbor points, so \(r\) i \(\in[0,1]\).

[0100] ⑩ After normalizing \(r\) i , we get

[0101]

[0102] Calculate the number g of new samples to be generated for each minority class sample x i The number g of new samples to be generated i :

[0103]

[0104] Among the K nearest neighbor points of each minority class sample x i Select a minority class sample x zi , and generate a new sample s according to the following rules i :

[0105] s i = x i +(x zi - x i ) × λ (6),

[0106] where λ is a coefficient in [0, 1]. In the present invention, in order to improve the randomness of the model, this coefficient is set to a random value within the range of [0, 1].

[0107] S1-4. In the present invention, when collecting data, the data set is divided in step S1-1. Generally speaking, when the input data is close to the "0" average value, the learning efficiency of the deep learning algorithm is the best. The collected data is normalized by the maximum and minimum values, and the data is mapped to between 0 and 1 to obtain the standardized results of the relevant characteristic parameters of the thermal power plant.

[0108] The normalization is carried out using the following formula:

[0109]

[0110] where x t is the relevant characteristic parameter of the safety monitoring data of the thermal power plant collected at time t; x 3>x is the maximum value in the collected sample parameters, that is, the maximum value of this parameter in all data sets; x 3i＝ is the minimum value in the collected sample parameters, that is, the minimum value of this parameter in all data sets; is the standardized result of the relevant characteristic parameters of the thermal power plant collected at time t.

[0111] Finally, taking these data with time as columns and parameter characteristics as rows, a data matrix of 20 rows * T columns is obtained. In this specific embodiment, the data matrix input each time is 20 * 25920, that is, the data for a total of 90 days in 3 months.

[0112] S2: Perform feature extraction on the data preprocessed in step S1, and establish a local model for federated learning based on dilated convolution and bidirectional long short-term memory network.

[0113] After the data preprocessing in step S1, in step S2, the present invention needs to extract features from the data.

[0114] For the parameter information of thermal power plants, feature extraction requires convolution operations. In ordinary convolutional neural networks, the receptive field is relatively single. Since the larger the receptive field is, the more suitable it is for long-time series data changes, and the smaller the receptive field is, the more suitable it is for short-time series data changes, a single receptive field is not suitable for the multi-scale time series parameters brought about by the multi-parameter changes that the present invention needs to process. Therefore, the present invention selects dilated convolution for feature extraction. At the same time, as Figure 2 shown, the local model structure established by the present invention is a multi-branch structure. The establishment steps are as follows:

[0115] (1) First, use dilated convolution combining multiple dilation rates for multi-branch feature selection. Each branch is a 3*3 convolution feature extraction branch, and the dilation rates of the convolution on each branch are different, being 1, 3, and 5. This essentially pays more attention to different time spans and the influence between different parameters.

[0116] (2) Then, unify the size of the feature data. For the feature data obtained in the previous step, repeatedly perform 1*3

[0117] convolution on the upper and lower parameters until 20 parameters are turned into 1 feature parameter; then, in the time dimension, on the branch data with dilation rates of 3 and

[0118] 5, fill 0 at the farthest and nearest time data, so as to unify the feature data into 1*25920.

[0119] (3) Then, input the features extracted by each dilated convolution into a bidirectional long short-term memory network Bi-LSTM (Bi-Long short-term memory) respectively for safety monitoring of the time series of a single branch of the thermal power plant.

[0120] (4) Then, output the information of the Bi-LSTM to the vector machine MLP to calculate the detection result together.

[0121] (5) Finally, establish a local model for federated learning.

[0122] Specifically, this step includes the following steps:

[0123] S2-1. Atrous convolution uses a parameter called dilation factor (DF) to adjust the size of the atrous convolution. The role of atrous convolution can replace the pooling operation, multiply the receptive field, enabling each convolution output to capture feature information in a larger range. For large thermal power plants, more feature information can be recognized, enhancing the role of feature extraction without increasing the computational complexity of the network. The atrous convolution kernels in the atrous residual convolution aggregation network are as shown in Figure 3 the appendix. Assuming the convolution kernel size is 3*1 and the stride is 1, when the dilation rate d is 1, that is, the number of "0" weights filled is 0. It can be seen from the following that the mapping range length of the (l + 1)-th layer convolution kernel in the l-th layer is the same:

[0124]

[0125] If the dilation rate d is 2, the receptive field is expanded to 5*1. Therefore, the advantage of atrous convolution lies in its ability to increase the local receptive field during the convolution operation without introducing additional parameters, capturing more thermal power generation monitoring information.

[0126] S2-2. The recurrent neural network consists of three network layers: the input layer, the hidden layer, and the output layer. Whether it is an RNN (Recurrent Neural Networks) or an LSTM (Long Short-Term Memory) network, it can only predict the monitoring data at the next moment based on the monitoring information at the previous moment. However, in the safety monitoring of thermal power plants, the monitoring at the current moment is not only related to the previous monitoring data but may also be related to the future monitoring data. The so-called "memory function" in the recurrent neural network is reflected in the self-connection W of the hidden layer. A connection means receiving its information. There is an order relationship between each input, and each input is associated with its previous input. By unfolding in time series, the sequence correlation between them can be found. The input to the hidden layer usually includes not only the output of the current input layer but also the output of the hidden layer at the previous moment, that is, the network memorizes the information at the previous moment and applies it to the calculation of the current moment output.

[0127] Generally speaking, the bidirectional long short-term memory network (Bi-LSTM) has more powerful expression and learning capabilities, but its LSTM also requires more training data. Bi-LSTM is composed of two LSTMs (forward LSTM and backward LSTM) stacked on top of each other, and the two LSTMs are connected to an output layer. This structure provides the output layer with the complete past and future context information for each point in the input sequence. The forward starts from the left as the starting input of the series. In text processing, it can be understood as starting from the beginning of the sentence for input, that is, the output of the hidden layer is related to the current input x tand the historical information h of the hidden layer at the previous moment tB1 is related, and its calculation formula is as shown in Equation (9):

[0128]

[0129] Backward LSTM starts from the right as the starting input of the series. In text processing, it can be understood as starting from the last word of the sentence as the input (sequence reversal), that is, the output of the hidden layer is related to the current input x t and the historical information h of the hidden layer at the previous moment tI1 is related, and its calculation formula is as shown in Equation (10):

[0130]

[0131] Finally, the two obtained results are concatenated to consider the context information, and its expression is as shown in Equation (11):

[0132]

[0133] The schematic diagram of the expansion of the Bi-LSTM structure in the time series is shown in the appendix Figure 4 as shown. The output h of the hidden layer t is usually not the final output of the network. Generally, it is necessary to project it to the output state o t , and its expression is as shown in Equation (12):

[0134] o t = φ(W h0 h t + b0) (12), where φ represents the activation function. Usually, activation functions such as Tanh, Sigmoid, and Relu are used in the hidden layer, and Softmax is usually used in the output layer. represents the weight matrix of the input at the current t moment in the forward calculation, and b A represents the bias term in the forward calculation process. represents the weight matrix of the input at the current t moment in the backward calculation, and b H represents the bias term in the backward calculation process. W h0 represents the weight matrix of the input of the output layer, and b0 represents the bias term of the output layer.

[0135] S2-3. After the above steps S2-1 to S2-2 are completed, at the output of the Bi-LSTM branch, the present invention fuses the information output by each Bi-LSTM branch and outputs it. Specifically, it is realized by using a multilayer perceptron MLP (MultilayerPerception). The multilayer perceptron MLP is a simple three-layer structure of input layer, hidden layer and output layer, where the input is 1*1 output of 3 Bi-LSTMs, the hidden layer is 10 neurons, and the output is 1 1*1. Finally, the output layer of the multilayer perceptron MLP is the local model output. At this point, the local model is established and is defined as w loZ>ls .

[0136] S3: The server side uses the federated learning local model established in step S2 to implement joint training, and uses the trained federated learning local model to perform safety monitoring of the thermal power plant.

[0137] The purpose of the present invention is to realize the safety monitoring of thermal power plants according to the aggregation model established by each local model without infringing the data of each thermal power plant. Therefore, after the above steps S1 and S2, the present invention establishes a local model for model aggregation, and adopts federated learning to jointly establish the model, which can effectively protect the data privacy security involved in each participating thermal power plant, and at the same time facilitate the use of the jointly trained model for safety monitoring in other thermal power plants. In step S1, the present invention has solved the data category imbalance of the local client thermal power plant. Considering the data category imbalance between different local client thermal power plants, the standard FedAvg algorithm shows poor model convergence effect under the condition of data category imbalance of the local client thermal power plant. The present invention can use an online sampling method to make the test performance of each category of the global model tend to be balanced after selecting the corresponding device combination in each round of communication. Different categories of data show serious distribution imbalance, and for the device level, the extreme value theory is also widely established, that is, due to objective restrictions such as geographical factors, data value, etc., a certain category of majority data in the entire classification task may be distributed in only a few devices. In addition, device selection is often combinatorial. Although there is a large category imbalance in the data on a single device, the degree of category balance of the data after combining multiple devices can be significantly improved. In order to ensure the performance of the model on a balanced global test set, an algorithm for selecting a subset of devices must be designed to fully utilize the data after the device combination to give full play to the value of the client thermal power plant data in federated learning. The design of the client thermal power plant selection algorithm includes two parts: the category estimation scheme and the online learning algorithm framework for device selection. The category estimation algorithm is designed in step S1, and the online learning algorithm is designed in step S3. The improved federated learning model is shown in the attached figure. Figure 5 shown.

[0138] In this step:

[0139] 3-1. After updating the local model parameters through the above process, upload them to the terminal server for aggregation to generate a global model. Aggregate the local model parameters generated in step S2-3 on the trusted server side. Since the present invention is directed to the safety monitoring of thermal power plants, there is an imbalance in the data categories of the local data. In step 1, the problem of class imbalance was addressed at the local end. To solve the different local data category imbalances among different clients, the present invention is based on the classical federated learning framework. In order to ensure the performance of the model on a balanced global test set on the server side, an algorithm for selecting a subset of devices is designed. By making full use of the data after combining the devices, the value of the local device data in federated learning is fully exploited. Subsequently, on this basis, the federated average FedAvg algorithm is used for model aggregation;

[0140] 3-2. Server-side class estimation algorithm. Since, under the federated learning framework, for privacy protection purposes, the global server cannot directly obtain any statistical information other than the sample quantity of the device local data. The present invention adopts the respondent-driven sampling method RDS (Respondent Driven Sampling), which is mainly used for the research of phenomena where both the population and the sample are difficult to determine. This method seeks objects of the same type in the network of the sampling objects to participate in the research, and makes an asymptotically unbiased estimate and inference about the population through the sample characteristics. The RDS method of the respondent-driven sampling method mainly includes three parts:

[0141] (1) Utilize the stationarity of the Markov chain to solve the problem of initial seed sensitivity;

[0142] (2) Use the Hansen-Hurwitz estimation method for stratified sampling control to reduce sampling bias, as shown in formula (13);

[0143]

[0144] where y i is the observation value of the i-th client, and z i is the probability that y i is selected.

[0145] (13) Control the number of subsequent samples included in the sample. This method stipulates that each sample can select at most 3 to 4 subsequent samples to reduce the sample homogeneity bias.

[0146] The present invention makes a simple adjustment to the standard RDS method. It uses random walk to obtain successor samples, changing the number of successor samples taken each time from 3 - 4 to only 1, and uses the harmonic mean for sample estimation. Subsequently, the present invention analyzes the structural characteristics of the network using the network average degree as an indicator. Equation (14) shows that the uneven distribution of internal relationships is the objective basis for the establishment of the RDS unequal probability estimation method, so this method can obtain a better sampling effect than uniform sampling.

[0147]

[0148] where is a random variable of the random walk sampling degree, B is a normalization constant, and b is an exponent. Therefore, the modulus transmitted by the device in each round of communication can be utilized and the auxiliary dataset aux on the global server. Using the RDS online sampling method, a global model aggregation client - selected class estimation scheme is established.

[0149] 3 - 3. The training steps include the following processes:

[0150] (1) In the t - th round of communication, after the local training of the k - th device ends, its local model is input into the global server. On the global server, let perform inference on the data of each class in aux, and calculate the square of the Euclidean norm of the model parameter gradient as shown in Equation (15):

[0151]

[0152] (2) Therefore, an estimation scheme for the class information on the device can be given by the following Equation (16):

[0153]

[0154] (3) After normalization, the proportion of the c - th class data in the local data of device k is Calculated as Equation (17):

[0155]

[0156] Give an intuitive explanation of the class estimation scheme.

[0157] S3-4. In the scenario of class imbalance in federated learning, due to the combinatorial nature of the selection of the device subset in each round of communication, the device selection algorithm is based on a class estimation scheme. Using the aggregated global model and the statistical information of the data class distribution on the current device in each round of communication, it selects the device combination that is most complementary to the deviation degree of the test performance of each class of the global model. After the device selection, in the next round of communication, the class ratio in the historical training data of the global model will tend to be balanced, thereby improving the performance on the global balanced test set. In the training of the federated learning model, after the training is completed on the k-th device in the t-th round of communication, the model is sent back to the auxiliary dataset aux on the global server. As can be seen from 3-2, the class ratio of the training data in the model and the global model class distribution before the aggregation in the previous round Using the cosine similarity as the evaluation criterion for the balance degree, the reward for selecting device k is defined as formula (18):

[0158]

[0159] After the selection is completed, perform the steps in the standard FedAvg and update after the training ends and the aggregation is completed.

[0160] 3-5. The present invention is a method for safety monitoring of thermal power plants based on federated learning. The goal of learning in the sequence modeling environment is to find a network that can minimize the expected loss between the actual result and the prediction. The present invention uses the mean square error (MSE) as the loss function as an index, that is, as shown in formula (19):

[0161]

[0162] After substituting, the cost function to be optimized is formula (20):

[0163]

[0164] Finally, set an error value to determine the monitoring accuracy of the joint modeling.

[0165] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for safety monitoring of a thermal power plant based on federated learning, characterized in that: It includes the following three steps: S1: Preprocess the safety monitoring data of thermal power plants; S2: Extract features from the data preprocessed in step S1, and build a local federated learning model based on dilated convolution and bidirectional long short-term memory network; S3: The server side uses the federated learning local model established in step S2 to implement joint training, and uses the trained federated learning local model to perform safety monitoring of the thermal power plant.

2. The method for safety monitoring of a thermal power plant based on federated learning according to claim 1 is characterized in that: Step S1 includes the following steps: S1-1. The acquired data is used to create a data set based on the thermal power plant as the client, and is divided into a training set, a test set, and a validation set; S1-2. According to the acquired characteristic parameters of the thermal power plant, map them into a vector form: X = {x1, x2, ..., x 20 }, where x1 is the output voltage monitoring characteristic parameter, x2 is the output current monitoring characteristic parameter, x3 is the boiler temperature monitoring characteristic parameter, x4 is the ambient humidity monitoring characteristic parameter, x5 is the smoke monitoring characteristic parameter, x6 is the communication monitoring characteristic parameter, x7 is the circuit breaker monitoring characteristic parameter, x8 is the lap joint monitoring characteristic parameter, x9 is the fuse monitoring characteristic parameter, x 10 is the monitoring characteristic parameter; x 11 For blower detection, x 12 is the characteristic parameter of the induced draft fan monitoring, x 13 Monitoring characteristic parameters of coal mill, x 14 Monitor characteristic parameters for high and low pressure heaters, x 15 Monitoring characteristic parameters of deaerator, x 16 is the condenser monitoring characteristic parameter, x 17 Condensate pump monitoring characteristic parameters, x 18 is the capacitance monitoring characteristic parameter, x 19 Monitor characteristic parameters and x for each device’s screw position 20 It is the output characteristic parameter of switch position monitoring; S1-3, deal with imbalanced data set samples; S1-4. The collected data are normalized by using the maximum and minimum values ​​to obtain the normalized results of the relevant characteristic parameters of the thermal power plant.

3. The method for safety monitoring of a thermal power plant based on federated learning according to claim 2 is characterized in that: Step S1-3 includes the following steps: First, the training set defined in step S1-1 is used, which includes m samples {x i ,y i }, i = 1, 2, ..., m, where x i is a sample of the n-dimensional feature space X, y i ∈Y={1,-1},is x i The associated class label; define m s is the number of minority class samples, i.e., the hardware parameter information of the thermal power plant safety monitoring data with long-term changes; m l is the number of samples of the majority class, i.e., the real-time changing parameter information of the safety monitoring data of the thermal power plant; thus, m s ≤m l , and m s +m l =m; perform the following calculations: ① Calculate the data category imbalance rate d: ②If d <d th , d th For the maximum imbalance rate set, calculate the number of minority class samples G that need to be synthesized: G=(m l -m s )×β, The parameter β∈[0, 1] represents the degree of balance sought in the processed data set; ③For any minority class sample x i , find the K nearest neighbor points according to the Euclidean distance in n-dimensional space, And calculate the ratio r i : Where Δ i is the number of samples belonging to the majority class among the K nearest neighbors, so r i ∈[0, 1]; ④ For r i After normalization, we get ⑤Calculation needs to be done for each minority class sample x i The number of new samples generated g i : ⑥ In each minority class sample x i Select a minority class sample x from the K nearest neighbors zi , generate new samples s according to the following rules i : s i =x i +(x zi -x i )×λ, Where λ is a coefficient of [0, 1].

4. The method for safety monitoring of a thermal power plant based on federated learning according to claim 2 is characterized in that: Step S1-4, standardization is performed using the following formula: where x t is the relevant characteristic parameter of the safety monitoring data of the thermal power plant collected at time t, x max is the maximum value of the collected sample parameters, x min is the minimum value of the collected sample parameters. It is the standardized result of relevant characteristic parameters of thermal power plants collected at time t.

5. The method for safety monitoring of a thermal power plant based on federated learning according to claim 1, characterized in that: The step S2 comprises the following steps: S2-1, first adjust the size of the dilated convolution, and then use the dilated convolution to perform feature selection on the data preprocessed in step S1; S2-2, Bidirectional long short-term memory network Bi-LSTM is composed of two LSTM stacked up and down, and the two LSTM are connected to an output layer; Positive The input from the left is the starting point of the series. In text processing, it can be understood as starting from the beginning of the sentence, that is, the hidden layer The output is the same as the current input x t and the historical information h of the hidden layer at the previous moment tB1 The calculation is carried out using the following formula: Reverse The starting input of the series is from the right. In text processing, it can be understood as taking the last word of the sentence as input, that is, the hidden layer The output is the same as the current input x t and the historical information h of the hidden layer at the previous moment t+1 The calculation is carried out using the following formula: Finally, the two results are spliced ​​and calculated using the following formula: The output of the hidden layer h t Usually not the final output of the network, but also needs to be Projected to output state o t , which is expressed as follows: about t =φ(W h0 h t +b0) Where: φ represents the activation function, represents the weight matrix input at the current time t in the forward calculation, b f represents the bias term in the forward calculation process, represents the weight matrix input at the current time t in the reverse calculation, b H represents the bias term in the reverse calculation process, W h0 represents the weight matrix of the output layer input, and b0 represents the output layer bias term; S2-3. After the above steps S2-1 to S2-2 are completed, at the output of the Bi-LSTM branch, the information output by each Bi-LSTM branch is fused and output. Specifically, this is achieved using a multilayer perceptron MLP. The multilayer perceptron MLP is a simple three-layer structure of input layer, hidden layer and output layer, where the input is the 1*1 output of 3 Bi-LSTMs, the hidden layer is 10 neurons, and the output is 1 1*1; finally, the output layer of the multilayer perceptron MLP is the local model output, and the local model is now established.

6. The method for safety monitoring of a thermal power plant based on federated learning according to claim 5 is characterized in that: In step S2-1, the dilated convolution uses the dilated rate parameter to adjust the size of the dilated convolution. Assuming that the convolution kernel size is 3*1 and the step length is 1, when the dilated rate is 1, the number of weights filled with "0" is 0. It can be seen from the following that the mapping range length of the l+1th layer convolution kernel in the lth layer is the same:

7. The method for safety monitoring of a thermal power plant based on federated learning according to claim 1, characterized in that: In step S3: After updating the local model parameters through the above process, they are uploaded to the terminal server and aggregated to generate a global model; The peer-driven sampling method (RDS) was adopted and adjusted. The subsequent samples were obtained by random walk. Instead of taking 3 to 4 subsequent samples each time, only one was taken. The harmonic mean was used for sample estimation. Then the structural characteristics of the network were analyzed with the network average degree as an indicator.

8. The method for safety monitoring of a thermal power plant based on federated learning according to claim 1, characterized in that: In step S3, the training step includes the following process: (1) In the tth communication, after the local training of the kth device is completed, its local model Input to the global server, on the global server, let In the auxiliary dataset aux, each class Perform inference on data and calculate the square of the Euclidean norm of the model parameter gradient As shown below: (2) An estimation scheme for the category information on the device is given using the following formula: (3) After normalization, the proportion of the cth type of data in the local data of device k is Calculate the following formula: Give an intuitive explanation of the class estimation scheme.

9. The method for safety monitoring of a thermal power plant based on federated learning according to claim 1, characterized in that: In step S3, in the scenario of class imbalance in federated learning, since the selection of device subsets in each round of communication is combinatorial, the device selection algorithm is based on the category estimation scheme, using the global model aggregated in each round of communication and the statistical information of the data category distribution on the current device to select the device combination that is most complementary to the performance deviation of each category of the global model test; After device selection, the category ratio in the historical training data of the global model will tend to be balanced after the next round of communication, thereby improving the performance on the global balanced test set; In the federated learning model training, in the tth round of communication, after the training is completed on the kth device, the model Return the auxiliary dataset aux on the global server, the category ratio of the training data in the model And the global model category distribution before the last round of aggregation Using cosine similarity as the evaluation criterion for the degree of balance, the reward for selecting device k is defined as the following formula: When the selection is complete, the steps in standard FedAvg are performed and updated after training is completed and aggregation is completed.

10. The method for safety monitoring of a thermal power plant based on federated learning according to claim 1, characterized in that: In step S3, the mean square error MSE is used as the loss function as an indicator, which is shown in the following formula: After substituting in, we get the cost function to be optimized, as follows: Finally, an error value is set to determine the monitoring accuracy of the joint modeling.

Images