Bayesian classification recognition system based on industrial PaaS platform
Disclosed is a Bayesian classification recognition system based on an industrial PaaS platform, comprising: an IaaS infrastructure service layer, a G-PaaS graph neural network processing layer, an O-PaaS docking service layer and an SaaS system application layer. The G-PaaS graph neural network processing layer is configured for point cloud feature generation, point cloud feature learning, point cloud structure estimation and point cloud model classification; and the recognition accuracy of a workpiece point cloud model is improved through the Bayesian classification recognition system based on the industrial PaaS platform.
1 . A Bayesian classification recognition system based on an industrial PaaS platform, comprising a computer hardware facility, an IaaS infrastructure service layer, a G-PaaS graph neural network processing layer, an O-PaaS docking service layer and an SaaS system application layer, wherein:
the IaaS infrastructure service layer is configured for providing physical resource control service and management for a request submitted by the G-PaaS graph neural network processing layer;
the G-PaaS graph neural network processing layer is configured for providing background underlying computing service support to the SaaS system application layer, and at least comprises a point cloud feature generation module, a point cloud feature learning module, a point cloud structure estimation module and a point cloud model classification module; the point cloud feature generation module pre-processes a point cloud feature through a diffusion-based spatial code, generates an original feature of point cloud, and represents a structural role of each node in the point cloud from a global perspective; the point cloud feature learning module aggregates feature information of a plurality of edges connected with a central node by using a graph neural network with an attention mechanism to represent feature information of the central node; the point cloud structure estimation module calculates a feature similarity of a node with surrounding nodes by using an estimation method self-adaptive to Bayesian network structure learning, and selects an adjacent node as a neighboring node of the node according to the feature similarity to define a spatial local critical region; and the point cloud model classification module completes convergence of a graph neural network model by realizing iterative optimization of the point cloud feature learning module and the point cloud structure estimation module for point cloud model classification;
the O-PaaS docking service layer is configured for providing a data docking service and an authority authorization service between the SaaS system application layer and a private cloud SaaS system and the G-PaaS graph neural network processing layer, and between a point cloud acquisition system of the O-PaaS docking service layer and the G-PaaS graph neural network processing layer; and
the SaaS system application layer is configured for being orientated to a user and connected with the O-PaaS docking service layer, and
at least comprises a user input management module, a platform management module, a collaborative design module, a checking module, a document management module and a workpiece classification result display module.
2 . The Bayesian classification recognition system based on the industrial PaaS platform according to claim 1 , wherein:
the IaaS infrastructure service layer comprises a unified resource management module, a general access authority management module and a monitoring alarm management module;
the O-PaaS docking service layer comprises an interface configuration module, an access authority module, a device networking management module, a device connection management module, a protocol conversion module and a point cloud acquisition system module; and
the SaaS system application layer is configured for providing a capability of supporting the industrial PaaS platform to classify workpieces.
3 . The Bayesian classification recognition system based on the industrial PaaS platform according to claim 1 , wherein:
the point cloud feature generation module is configured for:
representing point cloud data as G={V, E, H}, wherein V is a set of N point cloud nodes, and E is a set of edges and represented by an adjacency matrix A∈R N×N , wherein A i,j represents that one edge exists between a node i and a node j; and H={h 1 , h 2 , . . . , h N }∈R N×D represents a node feature matrix, and each node V i ∈V has a corresponding D-dimensional feature vector h i ; and preprocessing the point cloud feature through a diffusion-based spatial code, generating an original node feature H of the point cloud, and representing a structural role of each node in the point cloud from a global perspective; for the diffusion-based spatial code, a diffusion matrix S between nodes in the point cloud set is calculated by using a graph diffusion technology; then, each node v in the point cloud set is sorted according to a diffusion value of the node, and the sort is used as a data source; and the diffusion-based spatial code is calculated by using a learnable coding function according to a grade, and finally, the nodes are stacked into the original node feature H representing a point cloud attribute; and the diffusion-based spatial code is defined as follows:
H =⊗[linear(rank( S [idenx( v i )]))] T
wherein, idenx ( ) is an index query function, rank ( ) is a sorting function, linear ( ) is a learnable linear mapping, ⊗ represents series operation of matrix, and [ ] T represents transposition of matrix.
4 . The Bayesian classification recognition system based on the industrial PaaS platform according to claim 1 , wherein:
the point cloud structure estimation module is configured for:
representing an original undirected graph structure by G obs =(V, Y), wherein Y ab ∈{0,1}, 1≤a≤b≤V; when Y ab is equal to 1, a connected edge exists between a node a and a node b, and when Y ab is equal to 0, no connected edge exists; each node a is related to polynomial distribution π a on a hidden community of the node a; and assuming that a total number a of communities is a K class, probability distribution of a membership degree of a community member of each node is π a =[π a1 , π a2 , . . . , π aK ]; and
a joint posterior probability of the point cloud structure estimation module is defined as:
p
(
π
,
β
❘
G
obs
)
∝
p
(
β
)
p
(
π
)
p
(
G
obs
❘
π
,
β
)
=
{
∏
k
=
1
K
p
(
β
k
)
∏
a
N
p
(
π
a
)
∏
1
≤
a
<
b
≤
N
∑
Z
ab
,
Z
ba
p
(
Y
ab
,
Z
ab
,
Z
ba
❘
π
a
,
π
b
,
β
)
,
if
Z
ab
=
Z
ba
∏
k
=
1
K
p
(
β
k
)
∏
a
N
p
(
π
a
)
∏
∏
1
≤
a
<
b
≤
N
∑
Z
ab
,
Z
ba
p
(
Y
ab
,
Z
ab
,
Z
ba
❘
π
a
,
π
b
,
δ
)
,
otherwise
,
wherein, 0≤β K ≤1 represents a strength of a K th community, δ represents a probability of cross-community linking, Beta distribution is used as a prior of β K , and Dirichlet distribution is used as a prior of π a .
5 . The Bayesian classification recognition system based on the industrial PaaS platform according to claim 1 , wherein:
the point cloud model classification module is configured for:
incorporating the Bayesian method into the graph neural network model to estimate a neighboring node of the point cloud, and regarding an original graph structure as realization of a parametric random graph family by considering the Bayesian method; and a joint posterior coefficient, a node weight coefficient, an attention coefficient and a node label are inferred through the point cloud model classification module to calculate a posterior probability of a label, which is expressed as:
p
(
Z
❘
Y
L
,
X
,
G
obs
)
=
∫
p
(
Z
❘
X
,
α
,
g
,
W
)
p
(
W
❘
Y
L
,
X
,
g
)
p
(
g
❘
λ
)
p
(
λ
❘
G
obs
)
dWdgd
λ
,
wherein, α represents the attention coefficient; W is modeled as a random variable, and represents a weight coefficient on a graph structure g; Y represents a label of a training set; λ represents a parameter of the point cloud structure estimation module, Z represents a low-dimensional vector representation of a node; Y L represents a label of a training set; and G obs represents an original undirected graph structure; and
a posterior of p(W|Y L , X, g) is approximately solved by a Markov Chain Monte Carlo method, and Monte Carlo is approximately:
P
(
Z
❘
Y
L
,
X
,
G
obs
)
≈
1
KV
∑
k
=
1
K
∑
v
=
1
V
1
N
G
S
∑
i
=
1
N
G
∑
s
=
1
S
p
(
Z
❘
α
s
,
i
,
v
,
k
W
s
,
i
,
v
,
g
i
,
v
,
X
)
,
under the approximation, V samples λ v are extracted from p(λ|G obs ); N G graphs g i,v are sampled from p(g|λ v ) through the point cloud structure estimation module; S weight matrices W s,i,v are sampled by p(W|Y L , X, g i,v ) from the graphs g i,v through the point cloud feature learning module; and for an attention coefficient α s,i,v,k , an attention coefficient α ij of a node m and a first-order neighboring node n (n∈N m ) of the node m is learned by the attention coefficient calculation formula in the weight matrices W s,i,v .
6 . The Bayesian classification recognition system based on the industrial PaaS platform according to claim 2 , wherein:
workpiece data is input into the system by a user through the SaaS system application layer, and the O-PaaS docking service layer transmits the workpiece data to the G-PaaS graph neural network processing layer for processing.