A data-driven programmable logic controller dynamic trust level calculation method

Abstract

The application provides a data-driven programmable logic controller dynamic trust level calculation method, solves the problem that a programmable logic controller is difficult to realize accurate evaluation of a trust level in a complex operation environment, and realizes real-time and accurate dynamic evaluation of the trust level of the programmable logic controller. The innovation mainly includes the following: firstly, a radial basis function neural network is used as a carrier of a data-driven model to construct a dynamic trust calculation model, the structure and parameters of the model are adjusted online according to variable data such as system access time, access failure time and task execution time, and dynamic calculation of the trust level of the programmable logic controller is realized; secondly, a dynamic threshold adjustment mechanism is designed, the trust level division threshold is adaptively determined according to historical data and real-time data, and the accuracy of the trust level evaluation is ensured; finally, the model and the threshold adjustment mechanism are combined to realize accurate division and evaluation of the dynamic trust level of the programmable logic controller.

Description

TECHNICAL FIELD

[0001] The application designs a data-driven programmable logic controller (PLC) dynamic trust level calculation method, which realizes comprehensive evaluation of the state and behavior data of the PLC and real-time division of the trust level thereof. In the field of industrial control, timely and accurate grasp of the trust state of the PLC system can provide key support for stable operation of the system, and can also avoid production interruption, product quality problems and safety accidents caused by system failure or abnormality. Therefore, dynamic analysis and adjustment of the trust level of the system are of great significance. BACKGROUND

[0002] In the field of modern industrial control, PLC systems are widely used in many critical infrastructures and production processes, and the reliability and safety of their operation are crucial. With the increasing complexity of industrial environments and the continuous development of network technology, PLC systems are facing more and more potential security threats, such as network attacks, device failures, human errors, etc. However, traditional trust evaluation methods are mostly static and cannot accurately reflect the trust state changes of PLC systems in dynamic operation processes in real time, which leads to the inability to take effective measures in the face of system state changes and security threats, increasing the risk of system failure and safety accidents. At the same time, the requirements for system safety and trustworthiness in the field of industrial control are becoming higher and higher, and a method is needed to comprehensively, accurately and in real time evaluate the trust level of PLC systems to ensure the continuity, stability and security of industrial production, and to adapt to the trend of industrial intelligentization. Therefore, dynamic trust level calculation of the system is of great significance, which can ensure the safety and trustworthiness of PLC systems in complex and variable environments.

[0003] The application designs a data-driven PLC dynamic trust level calculation method, which mainly analyzes the state and behavior data of the system comprehensively, designs a dynamic trust level calculation model based on radial basis function, and realizes accurate dynamic analysis and trust level adjustment of the system. SUMMARY

[0004] The application obtains a data-driven PLC dynamic trust level calculation method, uses a radial basis function neural network as a carrier of a data-driven model, constructs a dynamic trust calculation model, adjusts the structure and parameters of the model according to variable data such as system access time, access failure time and task execution time, realizes dynamic calculation of the trust level of the PLC, designs an adjustment mechanism of a dynamic threshold, adaptively determines the trust level division threshold according to historical data and real-time data, ensures the accuracy of trust level evaluation, and finally, combined with the model and the threshold adjustment mechanism, the accurate division and evaluation of the PLC dynamic trust level can be realized.

[0005] The application adopts the technical solutions and implementation steps as follows:

[0006] A data-driven PLC dynamic trust level calculation method, comprising the following steps:

[0007] (1) determining the input variables required for calculation and the output variables to be calculated: for dynamic trust level calculation of the state and behavior data of a PLC (programmable logic controller) system, determining 7 variables including access time, access failure time, task execution time, task stop time, access stay time, access frequency and system warning frequency as the relevant variables for dynamic trust level calculation, and taking the calculated trust level as the output variable;

[0008] (2) designing a method for processing abnormal samples in a PLC data set, specifically

[0009] ① obtaining access time as a vector x1=[x 11 ,…,x 1n ], access failure time as a vector x2=[x 21 ,…,x 2n ], task execution time as a vector x3=[x 31 ,…,x 3n ], task stop time as a vector x4=[x 41 ,…,x 4n ], access stay time as a vector x5=[x 51 ,…,x 5n ], access frequency as a vector x6=[x 61 ,…,x 6n ], and system warning frequency as a vector x7=[x 71 ,…,x 7n ], constructing an input data set matrix X=[x1,…,x j ,…,x n ], x j =[x 1j ,…,x ij ,…,x mj ] T , i=1,…,m, j=1,…,n, where n and m are the number of relevant variables and the number of samples, respectively, and T represents the transpose of the matrix;

[0010] ② determining whether each data in the input data set is abnormal, and obtaining Z value as

[0011]

[0012] where Z ij is the standard for determining whether the jth data in the ith variable is an abnormal value, μj is the average of each input data vector, σ j is the standard deviation of each input data vector, respectively denoted as

[0013]

[0014] wherein ∑ is the summation symbol, when -3≤Z ij ≤3, the jth data in the ith variable is considered abnormal;

[0015] ③ Directly remove the abnormal value, and use the data μ j * to supplement the position where the data is located, denoted as

[0016]

[0017] wherein A is the abnormal value;

[0018] (3) Design a constraint neural network model based on radial basis function neural network (Radial Basis Function Neural Network). The neural network topology structure is divided into three layers, input layer, hidden layer and output layer. The connection mode of neural network n-R-1 is determined, that is, the number of input layer neurons is the number of related variables n, and the number of hidden layer neurons is R determined by grid search method, and the number of output layer neurons is 1. The network center value, width and weight are respectively C(t)=[c1(t),c2(t),…,c j (t),…,c n (t)] T , B(t)=[b1(t),b2(t),…,b j (t),…,b n (t)] T and ω(t)=[ω1(t),ω2(t),…,ω q (t),…,ω R (t)] T , j=1,2,…,n, q=1,2,…,R, c j (t)=[c j1 (t),c j2 (t),…,c jq (t),…,c jR (t)], b j (t)=[b j1 (t),b j2 (t),…,b jq (t),…,b jR (t)], c jq (t) and b jq(t) represents the center value and width between the jth output of the input layer and the qth input of the hidden layer at t time, and the weight value ω q (t) represents the weight value between the qth neuron of the hidden layer and the output layer neuron at t time, and C(t), B(t) and ω(t) are initialized as 1 in the process; the output corresponding to the neural network training is -1, 0 and 1, wherein -1 is defined as a low-level trust level, 0 is a medium-level trust level, and 1 is a high-level trust level, and the topological structure and calculation method of the network are as follows:

[0019] ① Input layer: the connection between the input layer and the hidden layer is realized through the weight, assuming that the input of the hidden layer is z(t), the calculation of a single neuron of the input layer to the input of the hidden layer is

[0020]

[0021] wherein z q (t) is the input of the qth neuron of the hidden layer at t time, w jq (t) is the weight value connecting the jth neuron of the input layer to the qth neuron of the hidden layer at t time, b q is a bias term;

[0022] ② Hidden layer: the activation function (radial basis function) of the neuron in this layer performs a nonlinear transformation on the input of the layer, and a Gaussian function is selected as the activation function of the layer, assuming that the output of the hidden layer is h(t), which is expressed as

[0023]

[0024] wherein v = 1, 2, …, R, R is the number of neurons in the hidden layer, h v (t) is the output of the vth node of the neuron of the hidden layer at t time, exp is an exponential function with the natural constant e as the base, and d(t) is the Euclidean distance between the input vector z(t) of the hidden layer at t time and the vth center vector c v

[0025] ③ Output layer: assuming that the input of the layer at t time is g(t), the output of the neuron of the layer is expressed as

[0026]

[0027] wherein θ1(t) represents a low-medium level threshold value at t time, θ2(t) represents a medium-high level threshold value at t time, -1 represents that the trust level of the current system is low, 0 represents that the trust level of the current system is medium, and 1 represents that the trust level of the current system is high;

[0028] ④ Evaluation function: the cross-entropy loss function of the model at t time is expressed as ​

[0029]

[0030] where C is the number of classification categories, y d (t) is the dth element in the one-hot encoding of the true label at time t, y d * (t) is the predicted output of the dth category at time t;

[0031] (5) training the neural network, updating the parameters of the model under the constraint:

[0032] a. applying a constraint to the output h(t) of the hidden layer at time t, h(t) = [h1(t), …, h q (t), …, h R (t)] T where q = 1, 2, …, R, calculating the proportion of each neuron at time t to the sum of all neurons as

[0033]

[0034] where R is the number of hidden layer neurons, and a fixed value a is set, when p q < a, it is considered that the proportion of neurons does not satisfy the constraint, and the overall impact is small, and the weight value corresponding to the neuron is set to 0, that is, the calculation of the qth neuron is directly ignored when calculating the input of the next layer of neurons;

[0035] b. applying a constraint to the weight ω(t) between the hidden layer and the output layer at time t, ω(t) = [ω1(t), …, ω q (t), …, ω R (t)] T where q = 1, 2, …, R, calculating the proportion of each weight to the sum of all weights as

[0036]

[0037] where R is the number of hidden layer neurons, and a fixed value b is set, when p q ′ < b, it is considered that the weight has a small overall impact, and the calculation of the qth weight can be directly ignored when updating the weight;

[0038] c. updating the parameters of the model by the gradient descent method of the loss function, and the gradient of the loss function to each parameter is expressed as

[0039]

[0040] where The partial derivative symbol is used to update the parameters of the model at time t+1 in the opposite direction of the gradient. The update rules for the connection weights, center value and width of the output layer are

[0041]

[0042] where ω(t+1), c(t+1) and b(t+1) are the weights, center value and width between the hidden layer neurons and output neurons at time t+1, and α is the learning rate.

[0043] d. The initial threshold values θ1(t) and θ2(t) are determined based on the one-third median and two-thirds median of the samples, and the threshold values θ1(t) and θ2(t) are dynamically adjusted. A sliding window is set to dynamically define a data range in the system related data, and the number of samples in the low, medium and high confidence levels within the sliding window is counted, denoted as n -1 , n0 and n1, respectively. The proportions of each confidence level in the window are calculated as

[0044]

[0045] where W is the size of the sliding window. The adaptive threshold values are dynamically adjusted based on the proportions of each confidence level. If the proportions differ greatly, the threshold values are updated by the median method, and the update rule is represented as

[0046]

[0047] where θ1(t+1) and θ2(t+1) are the updated threshold values, M -1 (t), M0(t) and M1(t) are the medians of the feature vectors in the data groups of each confidence level in the sliding window.

[0048] ⑥ If t = 100 or L(t) < L max , L max is the maximum number of learning, then the training is completed; otherwise, go to step ③; otherwise, t = t + 1, and return to step ⑤.

[0049] (4) System dynamic confidence level calculation based on radial basis function (RBF) neural network. The input sample matrix is selected as W(t+a) = [w(t+1), w(t+2), …, w(t+e), …, w(t+a)], w(t+e) = [w1(t+e), w2(t+e), …, w j (t+e), …, w n (t+e)] T , where e = 1, 2, …, a, a is the number of model training, and n is the number of related variables. The confidence level output of the model at time t+n is calculated according to formulas (6) to (22) as y(t+n).

[0050] The creativity of the present application mainly embodies in:

[0051] The present application aims at the problem that the PLC system is difficult to accurately evaluate the trust level in complex and changing operation environment, and proposes a data-driven method to process the trust level calculation of the PLC system, and through the communication module, the system log, user behavior and other multi-source data are widely collected, so that the safety level of the system can be comprehensively evaluated. Moreover, a dynamic threshold adjustment mechanism is designed, which can flexibly change the threshold of trust level division according to historical and real-time data, so that the evaluation is more suitable for the dynamic change of the system. DETAILED DESCRIPTION

[0052] The present application designs a data-driven PLC dynamic trust level calculation method; seven variables including access time, access failure time, task execution time, task stop time, access residence time, access frequency and system warning frequency are selected as input variables for dynamic trust calculation, and the calculated trust level is taken as the output variable.

[0053] The experimental data comes from the system state and behavior data records collected by the communication module, and after screening, 1300 groups of data are left;

[0054] The present application adopts the following technical solutions and implementation steps:

[0055] A data-driven PLC dynamic trust level calculation method, comprising the following steps:

[0056] (1) determining the input variables required for calculation and the output variables to be calculated: for the dynamic trust level calculation of the state and behavior data of the PLC (programmable logic controller) system, seven variables including access time, access failure time, task execution time, task stop time, access residence time, access frequency and system warning frequency are determined as the related variables for dynamic trust calculation, and the calculated trust level is taken as the output variable;

[0057] (2) designing a method for processing abnormal samples in the PLC data set, specifically

[0058] ① access time is taken as vector x1=[x 11 ,…,x 1n ], access failure time is taken as vector x2=[x 21 ,…,x 2n ], task execution time is taken as vector x3=[x 31 ,…,x 3n ], task stop time is taken as vector x4=[x 41 ,…,x 4n ], and access residence time is taken as vector x5=[x 51 ,…,x 5nThe number of visits is vector x6 = [x 61 ,…,x 6n The number of system warnings is represented by the vector x7 = [x 71 ,…,x 7n Construct the input dataset matrix as X = [x1, ..., x2]. j ,…,x n ], x j =[x 1j ,…,x ij ,…,x mj ] T , i = 1,…,m,j = 1,…,n, where n = 7 and m = 1300 are the number of relevant variables and the number of samples, respectively, and T represents the transpose of the matrix;

[0059] ② Determine if each data point in the input dataset is abnormal, and calculate the Z value.

[0060]

[0061] Z ij μ is the criterion for determining whether the j-th data point in the i-th type of variable is an outlier. j σ is the average value of each input data vector. j Let the standard deviation of each input data vector be denoted as .

[0062]

[0063] Where Σ is the summation symbol, when -3≤Z ij If the value is ≤3, then the j-th data in the i-th type of variable is considered an anomaly;

[0064] ③ Directly remove the outlier, and use data μ j * Supplement the location of this data, and indicate it as...

[0065]

[0066] Where A represents an outlier;

[0067] (3) Determine the radial basis function neural network model structure for dynamic credibility level calculation: The network topology is three layers: input layer, hidden layer, and output layer; determine the 7-8-1 connection method of the neural network, that is, the number of neurons in the input layer is 7, the number of neurons in the hidden layer is determined to be 8, and the number of neurons in the output layer is 1, determined by grid search method; the network center value, width, and weights are C(t)=[c1(t),c2(t),…,c j (t),…,c n (t)] T, B(t) = [b1(t), b2(t), …, b j (t), …, b n (t)] T and ω(t) = [ω1(t), ω2(t), …, ω q (t), …, ω R (t)] T where j = 1, 2, …, n, q = 1, 2, …, R, c j (t) = [c j1 (t), c j2 (t), …, c jq (t), …, c jR (t)], b j (t) = [b j1 (t), b j2 (t), …, b jq (t), …, b jR (t)], c jq (t) and b jq (t) represent the center value and width between the jth output of the input layer and the qth input of the hidden layer at time t, respectively, and the weight ω q (t) represents the weight between the qth neuron of the hidden layer and the neuron of the output layer at time t, and the internal elements of C(t), B(t) and ω(t) are all 1 in the initialization process; the outputs corresponding to the neural network training are -1, 0 and 1, wherein -1 is defined as a low-level trust level, 0 is a medium-level trust level, and 1 is a high-level trust level, and the topological structure and calculation method of the network are as follows:

[0068] ① Input layer: the connection between this layer and the hidden layer is realized through weights, assuming that the input of the hidden layer is z(t), the calculation of a single neuron of the input layer to the input of the hidden layer is

[0069]

[0070] where z q (t) is the input of the qth neuron of the hidden layer at time t, w jq (t) is the weight connecting the jth neuron of the input layer to the qth neuron of the hidden layer at time t, and b q is a bias term;

[0071] ② Hidden layer: the activation function (radial basis function) of the neuron in this layer performs a nonlinear transformation on the input of the layer, and a Gaussian function is selected as the activation function of the layer, assuming that the output of the hidden layer is h(t), which is expressed as

[0072]

[0073] where v = 1, 2,..., R, R is the number of hidden layer neurons, h v (t) is the output of the vth node of the hidden layer neuron at time t, exp is the exponential function with base e, d(t) is the Euclidean distance between the input vector z(t) of the hidden layer at time t and the vth center vector c v (t) at time t, and

[0074] ③Output layer: assuming that the input of this layer at time t is g(t), the output of the neurons in this layer is represented as

[0075]

[0076] where θ1(t) represents the low-medium level threshold at time t, θ2(t) represents the medium-high level threshold at time t, -1 indicates that the trust level of the current system is low, 0 indicates that the trust level of the current system is medium, and 1 indicates that the trust level of the current system is high;

[0077] ④Evaluation function: the cross-entropy loss function of the model at time t is represented as

[0078]

[0079] where C is the number of classification categories, y d (t) is the dth element in the one-hot encoding of the real label at time t, y d * (t) is the predicted output of the dth category at time t;

[0080] ⑤Training the neural network and updating the parameters of the model under the constraint condition:

[0081] a. A constraint is imposed on the output h(t) of the hidden layer at time t, h(t) = [h1(t),..., h q (t),..., h R (t)] T where q = 1, 2,..., R, and the proportion of each neuron to the sum of all neurons can be represented as

[0082]

[0083] where R is the number of hidden layer neurons, and a fixed value of 0.1 is set. When p q <0.1, it is considered that the proportion of neurons does not meet the constraint and has little effect on the whole, and the weight corresponding to the neuron is set to 0, i.e. the calculation of the qth neuron is directly ignored when calculating the input of the next layer of neurons;

[0084] b. A constraint is also imposed on the weights ω(t) between the hidden layer and the output layer at time t, ω(t) = [ω1(t), …, ω q (t), …, ω R (t)] T , where q = 1, 2, …, R, the proportion of each weight in the sum of all weights is calculated as

[0085]

[0086] where R is the number of hidden layer neurons, and a fixed value of 0.2 is set. When p q ′ <0.2, it is considered that the weight has a small overall impact, and the calculation of the qth weight can be directly ignored in weight update;

[0087] c. The parameters of the model are updated by the gradient descent method of the loss function, and the gradient of the loss function with respect to each parameter can be expressed as

[0088]

[0089]

[0090] where is the partial derivative symbol, and the parameters of the model at time t+1 are updated in the opposite direction of the gradient. The update rules for the connection weights, center values, and widths of the output layer are

[0091]

[0092] where ω(t+1), c(t+1), and b(t+1) are the weights between the hidden layer neurons and the output neurons, the center values, and the widths at time t+1, respectively, and α is the learning rate;

[0093] d. The initial threshold values θ1(t) and θ2(t) are determined based on the one-third median and the two-thirds median of the samples, and the threshold values θ1(t) and θ2(t) are dynamically adjusted. A sliding window is dynamically defined in a range of system-related data, and the number of samples in the sliding window is counted for low, medium, and high confidence levels, denoted as n -1 , n0, and n1, respectively. The proportions of each confidence level in the window are calculated as

[0094]

[0095] where W is the size of the sliding window, and W = 18 is set. The adaptive threshold values are dynamically adjusted based on the proportions of each confidence level. If the proportions differ greatly, the threshold values are updated by the median method, and the update rule is

[0096]

[0097] where θ1(t+1) and θ2(t+1) are updated thresholds, M -1 (t), M0(t) and M1(t) are the median of the feature vectors in each data group of the sliding window, respectively;

[0098] ⑥If t = 100 or L(t) < L max , L max is the maximum learning number, then the training is completed; otherwise, go to step ③; otherwise, t = t + 1, return to step ⑤;

[0099] (4) System dynamic confidence level calculation based on radial basis function (RBF) neural network, select the calculation input sample matrix, W(t+a) = [w(t+1), w(t+2), …, w(t+e), …, w(t+a)], w(t+e) = [w1(t+e), w2(t+e), …, wn(t+e)], e = 1, 2, …, a, a is the model training number, n is the number of related variables, and the confidence level output of the model at t+n is calculated according to formulas (23)-(44) y(t+n). j n T , w(t+e), …, wn(t+e)], e = 1, 2, …, a, a is the model training number, n is the number of related variables, and the confidence level output of the model at t+n is calculated according to formulas (23)-(44) y(t+n).​​

Claims

1. A data-driven programmable logic controller dynamic trust level calculation method, characterized in that Comprise the following steps: (1) Determine the input variables required for calculation and the output variables to be calculated: dynamic trust level calculation is performed on the state and behavior data of the programmable logic controller (PLC) system, and seven variables including access time, access failure time, task execution time, task stop time, access dwell time, access frequency and system warning frequency are determined as the relevant variables for dynamic trust level calculation, and the calculated trust level is taken as the output variable; (2) Design a method for processing abnormal samples in the PLC data set, specifically 11 1n 21 2n 31 3n 41 4n 51 5n 61 6n 71 7n j n j 1j ij mj T , i = 1, …, m, j = 1, …, n, where n and m are the number of relevant variables and the number of samples, respectively, and T represents the transpose of the matrix.​​​​​​​​​​​​​​​​​​​​​ ②Judge whether each data in the input data set is abnormal, and obtain the Z value as where Z ij is the criterion for determining whether the jth data in the ith variable is an outlier, μ j is the mean of each input data vector, σ j is the standard deviation of each input data vector, respectively represented as where Σ is the summation symbol, when -3 < Z ij ≤ 3, the jth data in the ith class of variables is considered abnormal. ③ Directly remove the outlier and use data μ j * Supplement the position of the data, denoted as Wherein, A is an abnormal value; (3) Design a constraint neural network model based on radial basis function neural network (RBFNN). The neural network topology structure is divided into three layers, input layer, hidden layer and output layer. Determine the connection mode of neural network n-R-1, that is, the number of input layer neurons is the number of related variables n, and the number of hidden layer neurons is R by grid search method, and the number of output layer neurons is 1. The network center value, width and weight are respectively C(t) = [c1(t), c2(t), …, c j (t), …, c n (t)] T , B(t) = [b1(t), b2(t), …, b j (t), …, b n (t)] T and ω(t) = [ω1(t), ω2(t), …, ω q (t), …, ω R (t)] T , where j = 1, 2, …, n, q = 1, 2, …, R, c j (t) = [c j1 (t), c j2 (t), …, c jq (t), …, c jR (t)], b j (t) = [b j1 (t), b j2 (t), …, b jq (t), …, b jR (t)], c jq (t) and b jq (t) represent the center value and width between the jth output of the input layer and the qth input of the hidden layer at time t, and the weight ω q (t) represents the weight between the qth neuron of the hidden layer and the output layer neuron at time t. The internal elements of C(t), B(t) and ω(t) are all 1 during the initialization process. The output corresponding to the neural network training is -1, 0 and 1, where -1 is defined as low confidence level, 0 is defined as medium confidence level, and 1 is defined as high confidence level. The topology structure and calculation method of the network are: ①Input layer: the connection between this layer and the hidden layer is realized through weight, assuming that the input of the hidden layer is z(t), the input calculation of a single neuron of the input layer to the hidden layer is where z q (t) is the input to the qth neuron of the hidden layer at time t, w jq (t) is the weight connecting the jth neuron of the input layer to the qth neuron of the hidden layer at time t, b q is the bias term; ②Hidden layer: the activation function of the neurons in this layer performs nonlinear transformation on the input of this layer, and the Gaussian function is selected as the activation function of this layer, assuming that the output of the hidden layer is h(t), which is expressed as where v = 1, 2,..., R, R is the number of hidden layer neurons, h v (t) is the output of the vth node of the hidden layer neuron at time t, exp is the exponential function with base natural number e, d(t) is the Euclidean distance between the input vector z(t) of the hidden layer at time t and the vth center vector c v v at time t; and ③Output layer: assuming that the input of this layer at time t is g(t), the output of the neurons of this layer is expressed as Wherein, θ1(t) represents the low-medium level threshold value at time t, θ2(t) represents the medium-high level threshold value at time t, -1 represents that the trust level of the current system is low, 0 represents that the trust level of the current system is medium, and 1 represents that the trust level of the current system is high; ④Evaluation function: the cross-entropy loss function of the model at time t is expressed as where C is the number of classification classes, y d (t) is the dth element in the one-hot encoding of the true label at time t, y d * (t) is the predicted output for the dth class at time t. ⑤Train the neural network and update the parameters of the model under the constraint condition: a. A constraint is imposed on the output h(t) of the hidden layer at time t, h(t) = [hi(t),...,h q (t),...,h R (t)] T where q = 1, 2,..., R, the proportion of each neuron to the sum of all neurons at time t is calculated as Wherein, R is the number of hidden layer neurons, set a fixed value a, when p q When a, it is considered that the neuron proportion does not satisfy the constraint, and the overall influence is small, so the weight value corresponding to the neuron is set to 0, that is, when calculating the input of the next layer of neurons, the calculation of the qth neuron is directly ignored. b. A constraint is also imposed on the weights ω(t) between the hidden layer and the output layer at time t, ω(t) = [ω1(t),...,ω q (t),...,ω R (t)] T where q = 1, 2,..., R, the proportion of each weight to the sum of all weights is calculated as Wherein, R is the number of hidden layer neurons, set a fixed value b, when p q ′ When b, it is considered that the weight has less influence on the whole, and the calculation of the qth weight can be directly ignored when updating the weight. c. Update the parameters of the model by loss function gradient descent method, and the gradient of the loss function to each parameter is expressed as wherein is the partial derivative symbol, the parameters of the model at time t+1 are updated in the opposite direction of the gradient, and the update rules for the connection weights, center value and width of the output layer are Wherein, ω(t+1), c(t+1) and b(t+1) are respectively the weight between the hidden layer neurons and the output neurons, the center value and the width at time t+1, and α is the learning rate; d. Determine the initial threshold values θ1(t) and θ2(t) according to the one-third median and two-thirds median of the samples, dynamically adjust the threshold values θ1(t) and θ2(t), set a sliding window to dynamically define a data range in the system-related data, and count the number of samples of low, medium, and high confidence levels within the sliding window, denoted as n -1 , n0, n1, respectively, and calculate the proportion of each confidence level within the window as Wherein, W is the size of the sliding window, and the adaptive threshold value is dynamically adjusted according to the proportion of each trust level, if the proportions are quite different, the threshold value is updated by the median method, and the update rule is expressed as where θ1(t+1) and θ2(t+1) are updated thresholds, M -1 M0(t) and M1(t) are the median values of the feature vectors in the data groups of each trust level in the sliding window, respectively. If t = 100 or L(t) < L max , L max is the maximum number of learning times, then the training is completed; otherwise, go to step 3; otherwise, t = t + 1, return to step 5. (4) Based on the system dynamic level of confidence calculation of the radial basis function (RBF) neural network, the input sample matrix W(t+a) = [w(t+1), w(t+2), …, w(t+e), …, w(t+a)] is selected, w(t+e) = [w1(t+e), w2(t+e), …, w j (t+e), …, w n (t+e)] T where e = 1, 2, …, a, a is the number of model training, n is the number of related variables, and the output of the model at time t+n is calculated according to formulas (6)-(22).

Images