CN104715282A - Data prediction method based on improved PSO-BP neural network - Google Patents

Abstract

The invention provides a data prediction method based on an improved PSO-BP neural network. A next datum to be acquired is predicted through a plurality of continuous data acquired by a sensor, the predicted datum and the acquired data are compared, and whether the acquired data are data defective pixels or not is judged; firstly, the input node number, the output node number and the implicit strata node number of the BP neural network are determined according to the features of the data acquired by the sensor; then, the connection weight and the threshold of the BP neural network are optimized through an improved particle swarm algorithm, and a final BP neural network prediction model is obtained; next, a DLL file is generated by the prediction model through the MATLAB7.1 and the VC6; finally, the DLL file is called through programming software, the predicted datum is compared with the acquired data, and whether the acquired data are datum defective indexes or not is judged. The method can effectively process datum defective indexes, and reliability and the hardware cost are considered at the same time.

Description

A kind of data predication method based on improving PSO-BP neural network

Technical field

The present invention relates to data prediction field, especially a kind of data predication method based on improving PSO-BP neural network.

Background technology

In automation control area, particularly at the control field such as petrochemical complex, power house, it is vital for carrying out monitoring in real time to operational outfit.Need to gather dissimilar data when monitoring in real time, comprise temperature, water level, pressure, flow etc.In actual motion monitoring of tools, sensor there will be the situation of data " bad point ", and these data " bad point " directly will affect the judgement of equipment monitoring system.Occur that the situation of data " bad point " mainly contains following two kinds: 1, due to the instability of data source, make data occur beating suddenly.2, because sensor does not in use overhaul or changes for a long time, can lose efficacy gradually, cause the data " bad point " gathered to get more and more.In above two kinds of situations, the second situation is comparatively general.

In order to improve the reliability of the data collected, reduce data " bad point " to the impact of image data, the method of usual employing has following several method: 1, adopt Redundant Control, multiple sensor is adopted to carry out data acquisition in same data area, improve data reliability, but add hardware cost simultaneously.2, before secure threshold, alerting signal is added to ahead of time caution device operations staff, but this must be equipped with equipment operations staff and operations staff will be on high alert state.3, increase rational delay time for eliminating the impact of data " bad point ", but the degree of reliability improves, and also may affect the safe operation of equipment limited.All there are some self shortcomings in above method, and cannot overcome.

Summary of the invention

In order to overcome the deficiency cannot taking into account reliability and hardware cost of existing data " bad point " disposal route, the present invention proposes a kind of data predication method based on improving PSO-BP neural network, it effectively can process data " bad point ", takes into account reliability and hardware cost.

The technical solution adopted for the present invention to solve the technical problems is as following content:

Based on the data predication method improving PSO-BP neural network, the method comprises the steps:

1), BP neural network model build, process is as follows:

(1.1) input layer and output layer nodes are determined

Neural network model formula is described as:

T _out＝f(T _in1,T _in2,…,T _inn) (1)

In formula: T _outthe data value that neural network needs prediction, T _in1~ T _innn data value of neural network input layer input respectively;

(1.2) the number of hidden nodes is determined

BP neural network hidden layers numbers adopts 1 layer, the selection of hidden node uses formula (2) determine:

$p=\sqrt{n+q}+a---\left(2\right)$

In formula: n is input layer number, q is output layer nodes, and a is the constant between 1 ~ 10;

The basic structure formula of BP neural network model:

Y＝Sigmoid[W ₂*Sigmoid(W ₁*X-O ₁)-O ₂] (3)

In formula: X is BP neural network input matrix; Y is BP neural network output matrix; W ₁, W ₂be respectively input layer in BP neural network to hidden layer, hidden layer to the connection weight value matrix of output layer; O ₁, O ₂for input layer in BP neural network is to hidden layer, hidden layer to the threshold matrix of output layer, excitation function Sigmoid is

2), adopt improve PSO algorithm Optimized BP Neural Network model construction, process is as follows:

(2.1) particle cluster algorithm particle rapidity is improved

Particle rapidity adjustment formula is:

$v_{\mathrm{id}}^{k+1}={\mathrm{wv}}_{\mathrm{id}}^k+c_1{r}_1(p_{\mathrm{id}}-z_{\mathrm{id}}^k)+c_2{r}_2(p_{\mathrm{gd}}-z_{\mathrm{id}}^k)+c_3{r}_3{z}_{\mathrm{id}}^k---\left(4\right)$

In formula: for the current kth needing speed to adjust particle is for position vector; c ₁, c ₂, c ₃for Studying factors; r ₁, r ₂, r ₃for the random number between [0,1];

(2.2) particle cluster algorithm particle inertia weight is improved

Inertia weight w is used to the previous speed of adjustment particle to the influence degree of present speed, is derived, propose following weighed value adjusting formula by quadratic function matching:

$w=(w_{\max }-w_{\min }){\left(\frac{k}{k_{\max }}\right)}^2-\frac{k}{k_{\max }}+w_{\max }---\left(5\right)$

In formula: w _maxfor initial inertia weight, w _minfor final inertia weight, k _maxfor maximum iteration time, k is current iteration number of times;

(2.3) particle dimension in particle cluster algorithm is determined

According to fixed BP neural network prediction model, determine to connect weights and threshold sum in BP neural network, connect weights and threshold sum formula:

d＝np+pq+p+q (6)

In formula: d is for connecting weights and threshold sum, and n is input layer number, and p is the number of hidden nodes, and q is output layer nodes;

(2.4) improve PSO algorithm Optimized BP Neural Network realizes

Regenerate by making the connection weights and threshold of BP neural network constantly update to particle rapidity and position adjustment, make BP neural network total error be less than setting value or reach iterations, process is as follows:

(2.4.1) determine particle dimension according to the threshold value of BP neural network, weights, and produce primary group;

(2.4.2) fitness function is passed through calculate adaptive value f, being less than setting value according to adaptive value f or reaching iterations is principle, evaluates all individualities in every generation population, and therefrom finds current optimal location p _i, then with acquired optimal location p _gcompare and generate new p _g, through successive ignition, until find global optimum position p _g;

(2.4.3) according to p _gdetermine the initial connection weights and threshold of BP neural network;

(2.4.4) train BP neural network, obtain final BP neural network prediction model;

(2.4.5) training terminates;

3), data prediction CMOS macro cell and application

Use MATLAB7.1 to realize the design cycle of (2.4), obtain final BP neural network prediction model, select MATLAB7.1 and VC6 to convert forecast model to dll file as developing instrument; In MATLAB, carry out mex – setup respectively by step, mbuild – setup operates, and sets compiler; After setting up parameter, input mcc – Wcpplib:shujuyuce – T link:lib shujuyuce.m in order line and generate dll file; The dll file generated is called by programming software, and predicted data and image data are contrasted, be as the criterion with actual acquired data, whether the difference of computational prediction data and image data is greater than setting threshold value, if be greater than setting threshold value, is considered as data " bad point ".

Principle of work of the present invention: the present invention predicts by several data continuous collected sensor the data that the next one will gather, is undertaken contrasting by predicted data and image data and judges whether image data is data " bad points ".The feature of the data first collected according to sensor determines the input number of nodes of BP neural network, output node number and the number of hidden nodes; Then use modified particle swarm optiziation to optimize the connection weights and threshold of this BP neural network, and obtain final BP neural network prediction model; Then by MATLAB7.1 and VC6, forecast model is generated dll file; Call dll file finally by programming software predicted data and image data to be contrasted, be as the criterion with actual acquired data, whether the difference of computational prediction data and image data is greater than a certain setting threshold value, if be greater than setting threshold value, is considered as data " bad point ".

Beneficial effect of the present invention shows as: 1, utilize modified particle swarm optiziation to carry out Optimized BP Neural Network and effectively improve BP neural network prediction model precision; 2, facilitate all kinds of programming software to call data prediction CMOS macro cell dll file, be easy to computer system and realize; 3, result is obtained scientific and reasonable.

Accompanying drawing explanation

Fig. 1 is particle group optimizing BP neural network algorithm process flow diagram.

Embodiment

Below in conjunction with example, the present invention is described further.

With reference to Fig. 1, a kind of data predication method based on improving PSO-BP neural network, the method comprises the steps:

1), BP neural network model build, process is as follows:

(1.1) input layer and output layer nodes are determined

BP neural network input layer nodes and output layer nodes are depending on actual conditions.In this patent, predicted data and image data, to predict the data that the next one will gather, carry out contrasting and judge whether image data is data " bad points " by several data continuous that use sensor collects.Therefore determine that the input layer number of neural network model is depending on actual conditions, and output layer nodes 1.Neural network model can be described as with formula:

T _out＝f(T _in1,T _in2,…,T _inn) (1)

In formula: T _outthe data value that neural network needs prediction, T _in1~ T _innn data value of neural network input layer input respectively;

(1.2) the number of hidden nodes is determined

In this patent, BP neural network hidden layers numbers adopts 1 layer.The selection of hidden node use formula (2) determine.

$p=\sqrt{n+q}+a---\left(2\right)$

In formula: n is input layer number, q is output layer nodes, and a is the constant between 1 ~ 10.Neural network the number of hidden nodes is determined finally by test method.

Therefore the basic structure formula of BP neural network model:

Y＝Sigmoid[W ₂*Sigmoid(W ₁*X-O ₁)-O ₂] (3)

In formula: X is BP neural network input matrix; Y is BP neural network output matrix; W ₁, W ₂be respectively input layer in BP neural network to hidden layer, hidden layer to the connection weight value matrix of output layer; O ₁, O ₂for input layer in BP neural network is to hidden layer, hidden layer to the threshold matrix of output layer, excitation function Sigmoid is

2), improve population (PSO) algorithm optimization BP neural network model structure, process is as follows:

(2.1) particle cluster algorithm particle rapidity is improved

In particle cluster algorithm, in order to avoid too small and lose ability of searching optimum in iteration later stage particle search speed, a kind of particle rapidity method of adjustment is proposed in this article.Its particle rapidity adjustment formula is:

$v_{\mathrm{id}}^{k+1}={\mathrm{wv}}_{\mathrm{id}}^k+c_1{r}_1(p_{\mathrm{id}}-z_{\mathrm{id}}^k)+c_2{r}_2(p_{\mathrm{gd}}-z_{\mathrm{id}}^k)+c_3{r}_3{z}_{\mathrm{id}}^k---\left(4\right)$

In formula: for the current kth needing speed to adjust particle is for position vector; c ₁, c ₂, c ₃for Studying factors; r ₁, r ₂, r ₃for the random number between [0,1].Wherein c ₁=c ₂=c ₃=2.

(2.2) particle cluster algorithm particle inertia weight is improved

Inertia weight w is used to the previous speed of adjustment particle to the influence degree of present speed.Derived by quadratic function matching in this patent, propose following weighed value adjusting formula:

$w=(w_{\max }-w_{\min }){\left(\frac{k}{k_{\max }}\right)}^2-\frac{k}{k_{\max }}+w_{\max }---\left(5\right)$

In formula: w _maxfor initial inertia weight, w _minfor final inertia weight, k _maxfor maximum iteration time, k is current iteration number of times.Wherein initial inertia weight w _max=0.9, final inertia weight w _min=0.4, maximum iteration time k _max=1500.

(2.3) particle dimension in particle cluster algorithm is determined

According to fixed BP neural network prediction model, determine to connect weights and threshold sum in BP neural network.Connect weights and threshold sum formula:

d＝np+pq+p+q (6)

In formula: d is for connecting weights and threshold sum, and n is input layer number, and p is the number of hidden nodes, and q is output layer nodes.

(2.4) improve PSO algorithm Optimized BP Neural Network realizes

The connection weights and threshold of modified particle swarm optiziation to BP neural network is used to be optimized, regenerate by making the connection weights and threshold of BP neural network constantly update to particle rapidity and position adjustment, make BP neural network total error be less than setting value or reach iterations.The concrete steps of particle cluster algorithm Optimized BP Neural Network are as shown in Figure 1:

(2.4.1) determine particle dimension according to the threshold value of BP neural network, weights, and produce primary group.

(2.4.2) fitness function is passed through calculate adaptive value f, being less than setting value according to adaptive value f or reaching iterations is principle, evaluates all individualities in every generation population, and therefrom finds current optimal location p _i, then with acquired optimal location p _gcompare and generate new p _g.Through successive ignition, until find global optimum position p _g.

(2.4.3) according to p _gdetermine the initial connection weights and threshold of BP neural network.

(2.4.4) train BP neural network, obtain final BP neural network prediction model.

(2.4.5) training terminates.

3), data prediction CMOS macro cell and application

Use MATLAB7.1 to realize the design cycle of (2.4), obtain final BP neural network prediction model, select MATLAB7.1 and VC6 to convert forecast model to dll file as developing instrument.In MATLAB, carry out mex – setup respectively by step, mbuild – setup operates, and sets compiler.After setting up parameter, input mcc – Wcpplib:shujuyuce – T link:lib shujuyuce.m in order line and generate dll file.The dll file generated can be called by programming software, and predicted data and image data are contrasted, be as the criterion with actual acquired data, whether the difference of computational prediction data and image data is greater than a certain setting threshold value, if be greater than setting threshold value, is considered as data " bad point ".

Example: be predicted as example with hydraulic turbine top guide bearing temperature data and be described, based on the data predication method improving PSO-BP neural network, comprises following process:

1), BP neural network model builds

In this patent, example is predicted as with hydraulic turbine top guide bearing temperature data, determine that continuous 4 data collected with sensor predict the data that the next one will gather according to data characteristics, therefore determine that the input layer number of neural network model is 4, and output layer nodes 1.According to this patent mentality of designing, determine that BP neural network hidden layers numbers adopts 1 layer.The selection of hidden node use formula (1) determine.

$p=\sqrt{n+q}+a---\left(1\right)$

In formula: n is input layer number, q is output layer nodes, and a is the constant between 1 ~ 10.

Finally according to the span of the number of hidden nodes, determine neural network the number of hidden nodes by test method, determine that in this example, the number of hidden nodes is 9.

Therefore determine the basic structure formula of BP neural network model:

Y＝Sigmoid[W ₂*Sigmoid(W ₁*X-O ₁)-O ₂] (2)

In formula: X is BP neural network input matrix; Y is BP neural network output matrix; W ₁, W ₂be respectively input layer in BP neural network to hidden layer, hidden layer to the connection weight value matrix of output layer; O ₁, O ₂for input layer in BP neural network is to hidden layer, hidden layer to the threshold matrix of output layer, excitation function Sigmoid is

2), improve population (PSO) algorithm optimization BP neural network model to build

(2.1) particle cluster algorithm particle rapidity is improved

Particle rapidity adjustment formula is:

$v_{\mathrm{id}}^{k+1}={\mathrm{wv}}_{\mathrm{id}}^k+c_1{r}_1(p_{\mathrm{id}}-z_{\mathrm{id}}^k)+c_2{r}_2(p_{\mathrm{gd}}-z_{\mathrm{id}}^k)+c_3{r}_3{z}_{\mathrm{id}}^k---\left(3\right)$

In formula: for the current kth needing speed to adjust particle is for position vector; c ₁, c ₂, c ₃for Studying factors; r ₁, r ₂, r ₃for the random number between [0,1].Wherein c ₁=c ₂=c ₃=2.

(2.2) particle cluster algorithm particle inertia weight is improved

Inertia weight w adjustment formula:

$w=(w_{\max }-w_{\min }){\left(\frac{k}{k_{\max }}\right)}^2-\frac{k}{k_{\max }}+w_{\max }---\left(4\right)$

In formula: w _maxfor initial inertia weight, w _minfor final inertia weight, k _maxfor maximum iteration time, k is current iteration number of times.Wherein initial inertia weight w _max=0.9, final inertia weight w _min=0.4, maximum iteration time k _max=1500.

(2.3) particle dimension in particle cluster algorithm is determined

Obtaining population particle dimension according to particle dimension computing formula is 55 dimensions.

(2.4) improve PSO algorithm Optimized BP Neural Network realizes

The connection weights and threshold of modified particle swarm optiziation to BP neural network is used to be optimized, regenerate by making the connection weights and threshold of BP neural network constantly update to particle rapidity and position adjustment, make BP neural network total error be less than setting value or reach iterations.The concrete steps of particle cluster algorithm Optimized BP Neural Network are as shown in Figure 1:

(2.4.1) primary group is produced.

(2.4.2) fitness function is passed through calculate adaptive value f, being less than setting value according to adaptive value f or reaching iterations is principle, evaluates all individualities in every generation population, and therefrom finds current optimal location p _i, then with acquired optimal location p _gcompare and generate new p _g.Through successive ignition, until find global optimum position p _g.

(2.4.3) according to p _gdetermine the initial connection weights and threshold of BP neural network.

(2.4.4) train BP neural network, obtain final BP neural network prediction model.

(2.4.5) training terminates.

3), data prediction CMOS macro cell and application

Use MATLAB7.1 to realize the design cycle of (2.4), obtain final BP neural network prediction model, select MATLAB7.1 and VC6 to convert forecast model to dll file as developing instrument.In MATLAB, carry out mex – setup respectively by step, mbuild – setup operates, and sets compiler.After setting up parameter, input mcc – Wcpplib:shujuyuce – T link:lib shujuyuce.m in order line and generate dll file.The dll file generated can be called by programming software, and predicted data and image data are contrasted, be as the criterion with actual acquired data, whether the difference of computational prediction data and image data is greater than a certain setting threshold value, if be greater than setting threshold value, is considered as data " bad point ".

Claims (1)

1., based on the data predication method improving PSO-BP neural network, it is characterized in that: the method comprises the steps:

1), BP neural network model build, process is as follows:

(1.1) input layer and output layer nodes are determined

Neural network model formula is described as:

T _out＝f(T _in1,T _in2,…,T _inn) (1)

In formula: T _outthe data value that neural network needs prediction, T _in1~ T _innn data value of neural network input layer input respectively;

(1.2) the number of hidden nodes is determined

BP neural network hidden layers numbers adopts 1 layer, the selection of hidden node uses formula (2) determine:

$p=\sqrt{n+q}+a---\left(2\right)$

In formula: n is input layer number, q is output layer nodes, and a is the constant between 1 ~ 10;

The basic structure formula of BP neural network model:

Y＝Sigmoid[W ₂*Sigmoid(W ₁*X-O ₁)-O ₂] (3)

In formula: X is BP neural network input matrix; Y is BP neural network output matrix; W ₁, W ₂be respectively input layer in BP neural network to hidden layer, hidden layer to the connection weight value matrix of output layer; O ₁, O ₂for input layer in BP neural network is to hidden layer, hidden layer to the threshold matrix of output layer, excitation function Sigmoid is

2), adopt improve PSO algorithm Optimized BP Neural Network model construction, process is as follows:

(2.1) particle cluster algorithm particle rapidity is improved

Particle rapidity adjustment formula is:

$v_{\mathrm{id}}^{k+1}=w{v}_{\mathrm{id}}^k+c_1{r}_1(p_{\mathrm{id}}-z_{\mathrm{id}}^k)+c_2{r}_2(p_{\mathrm{gd}}-z_{\mathrm{id}}^k)+c_3{r}_3{z}_{\mathrm{id}}^k---\left(4\right)$

In formula: for the current kth needing speed to adjust particle is for position vector; c ₁, c ₂, c ₃for Studying factors; r ₁, r ₂, r ₃for the random number between [0,1];

(2.2) particle cluster algorithm particle inertia weight is improved

Inertia weight w is used to the previous speed of adjustment particle to the influence degree of present speed, is derived, propose following weighed value adjusting formula by quadratic function matching:

$w=(w_{\max }-w_{\min }){\left(\frac{k}{k_{\max }}\right)}^2-\frac{k}{k_{\max }}+w_{\max }---\left(5\right)$

In formula: w _maxfor initial inertia weight, w _minfor final inertia weight, k _maxfor maximum iteration time, k is current iteration number of times;

(2.3) particle dimension in particle cluster algorithm is determined

According to fixed BP neural network prediction model, determine to connect weights and threshold sum in BP neural network, connect weights and threshold sum formula:

d＝np+pq+p+q (6)

In formula: d is for connecting weights and threshold sum, and n is input layer number, and p is the number of hidden nodes, and q is output layer nodes;

(2.4) improve PSO algorithm Optimized BP Neural Network realizes

Regenerate by making the connection weights and threshold of BP neural network constantly update to particle rapidity and position adjustment, make BP neural network total error be less than setting value or reach iterations, process is as follows:

(2.4.1) determine particle dimension according to the threshold value of BP neural network, weights, and produce primary group;

(2.4.2) fitness function is passed through calculate adaptive value f, being less than setting value according to adaptive value f or reaching iterations is principle, evaluates all individualities in every generation population, and therefrom finds current optimal location p _i, then with acquired optimal location p _gcompare and generate new p _g, through successive ignition, until find global optimum position p _g;

(2.4.3) according to p _gdetermine the initial connection weights and threshold of BP neural network;

(2.4.4) train BP neural network, obtain final BP neural network prediction model;

(2.4.5) training terminates;

3), data prediction CMOS macro cell and application

Use MATLAB7.1 to realize the design cycle of (2.4), obtain final BP neural network prediction model, select MATLAB7.1 and VC6 to convert forecast model to dll file as developing instrument; In MATLAB, carry out mex – setup respectively by step, mbuild – setup operates, and sets compiler; After setting up parameter, input mcc – W cpplib:shujuyuce – T link:lib shujuyuce.m in order line and generate dll file; The dll file generated is called by programming software, and predicted data and image data are contrasted, be as the criterion with actual acquired data, whether the difference of computational prediction data and image data is greater than setting threshold value, if be greater than setting threshold value, is considered as data " bad point ".