Edge detection method
When security detection is carried out at the edge end, we propose an attention-based recurrent convolution net- work(ARCNN), which is shown in Figure 2. It is made up of multiple recurrent convolution network layers (MRCN) and multiple attention layers (ATL). The key module of ARCNN is the MRCN. The states of MRCN units evolve over discrete time steps. For a unit located at (i,j) on the kth feature map in an MRCN, its net input zjk(t) at time step t is given by:
Zijk(t) = (wk)Tu(l’j) (t) + (wk)Tu(i’j')(t — 1) + bk (1)
In the equation (t) and u(i,j)(t — 1) denote the feedforward and recurrent input, respectively, which are the vectorized patches centered at (i,j) of the feature maps in the previous and current layer, wf and wk denote the vectorized feedforward weights and recurrent weights, respectively, and bk is the bias. The first term in (1) is used in standard CNN and the second term is induced by the recurrent connections. The activity or state of this unit is a function of its net input Xjk(t) = g(f (zjk(t))), where f is the rectified linear activation function f (zjk(t)) = max(zijk(t), 0). Then we get the output MRCN Omrcn = f (z%jk(t)) and put it into the attention mechanism. They can be described as follows:
Q = K = V = Omrcn
headi = Attention(QW Q, KWf, VW,y) (2)
Oatl = Concat(head1, ••• , headh)Wo
where wQ E R2dhxdk, WK E R2dh xdk , WV E R2dhxdk and Wo E R2dhx2dh are trainable projection parameters and 2k = 2dh/h. We used three layers of MRCN and ATL, and then we get the final output Oarcnn .
Comprehensive evaluation model in the cloud
Based on the output of the edge end, we propose a comprehensive evaluation model of cloud nodes, which is shown in Figure 3, mainly using graph convolution network. After training, the value of each node is related to the nodes related to it, and then regression analysis is made on these nodes. GCN extracts the relationship matrix information to obtain the graph relation information among the 0ARCNN. GCN is a multi-layer neural network that operates directly on a graph and induces embedding vectors of nodes based on properties of their neighborhoods. We define triples (u, r, f) with respect to a graph G = (V,E), where V(\V| = n) and E are sets of nodes and edges respectively. Here u E V, f E V and r E E. Every node is assumed to be connected to itself, i.e., (v, v) E E for any v. Let X E Rnxm be a matrix containing all n nodes with their features, where m is the dimension of the feature vectors, each row xv E is the feature vector for v. We introduce an adjacency matrix A of G and its degree matrix D, where Dii = j Aij. The diagonal elements of A
are set to l because of self-loops. GCN can capture information only about immediate neighbors with one layer of convolution. When multiple GCN layers are stacked, information about larger neighborhoods are integrated. For a one-layer GCN, the new k-dimensional node feature matrix Ll E Rnxk is computed as Ll = p(AXWo), where A = D-2AD-2 is the normalized symmetric adjacency matrix and W0 E Rmxk is a weight matrix. p is an activation function, e.g. a ReLU p(x) = max(0,x). As mentioned before, one can incorporate higher order neighborhoods information by stacking multiple GCN layers: Lj+1 = p(AL(j) Wj), where j denotes the layer number andL(0) = X. We use two layers in this paper, and its output is Ogcn = p(AL(1) Wi). We define the triples (u, r, f) is the relationship matrix among the output of the edge end, where (u, r, f) represents the output of the edge
261
Power plant edge computing node
Substation integrated automation system ,
Online identification of
conventional power parameters
Multi-type power supply
coordination control of
wide area power generation ,
On-site edge
A
Substation parallel simulation
computing platform Д- Service platform for load modeling
and regulation capability evaluation
Safety and stability control of wide-area operation of power grid
Basic Data Layer Edge Computing Layer Plant Application Layer Scheduling Application Layer
Fig. 4. Application framework of edge computing platform in power system.
TABLE I
Image recognition result of edge segment security detection.
Model
|
p
|
R
|
F1
|
CNN
|
82.31
|
83.78
|
83.04
|
CNN+Pooling
|
82.97
|
84.03
|
83.50
|
CNN+Attention
|
83.13
|
84.56
|
83.84
|
Recurrent-CNN
|
83.48
|
84.71
|
84.09
|
Recurrent-CNN+Pooling
|
83.69
|
84.87
|
84.28
|
our method
|
84.14
|
85.06
|
84.60
|
TABLE II
Experimental results of comprehensive evaluation on cloud
Model
|
Stability
|
Failure Rate
|
GNN+Relu
|
75.31
|
22.78
|
GNN+Tanh
|
76.87
|
21.56
|
GNN+Sigmoid
|
77.39
|
21.03
|
GCN+Relu
|
78.21
|
20.56
|
GCN+Tanh
|
79.82
|
19.87
|
our method
|
80.52
|
18.13
|
(3)
У
262
end, the relation, and the output of the edge end, respectively. Then we use the regression function to predict the nodes, and the formula is as follows:
1
1 + e°GCN
IV. Experiment
We perform extensive experiments to evaluate the performance of our method, and compare its performance with the other models. We present the experimental settings and results in rest of this section.
Do'stlaringiz bilan baham: |