CN106845144A - A kind of trend prediction method excavated based on industrial big data - Google Patents

Abstract

The invention discloses a state prediction method based on industrial big data mining, including steps: Step 1, data collection: taking samples reflecting the historical operating state of the system as a training set Where _xi is the system state variable, that is, the input of the model, and t _i is the predictor of concern, that is, the output of the model; Step 2, OS‑ELM model: use the training samples from Step 1 Establish several OS-ELM models and calculate several predicted values; Step 3, EOS-ELM model: average the prediction results of the OS-ELM model to obtain the prediction results of the EOS-ELM model. The invention solves the problem that the system state is difficult to predict in the current industrial system, and improves the stability and reliability of the prediction.

Description

A State Prediction Method Based on Industrial Big Data Mining

technical field

The invention belongs to the field of industrial big data mining, in particular to a state prediction method based on industrial big data mining

Background technique

With the increasing size and complexity of industrial systems, people have higher and higher requirements for the safety and reliability of system operation. The connection between the systems is closer, and the failure of one component will lead to the failure of the subsystem or even the paralysis of the whole system. These problems will bring huge economic losses to the enterprise, and even cause environmental pollution and even casualties. Estimating the operating status of the system in advance and predicting the time and location of abnormalities are effective means to eliminate potential dangers in time, maintain the normal operation of the system, and improve safety and economic benefits.

Neural network, Markov model, Bayesian estimation and ELM are common prediction methods at present. Among them, the ELM (Extreme Learning Machine, which is a generalized single hidden layer feedforward neural network) method is more suitable for processing massive data, and has the characteristics of fast training speed, less manual intervention, and strong generalization ability. The OS-ELM (Online Sequential Extreme Learning Machine) method can predict the system state online and in real time, but the OS-ELM network is randomly generated based on sequence input data, and the prediction effect may encounter the worst situation. In order to avoid this situation, the EOS-ELM method averages the prediction results of several OS-ELMs, so as to adapt to different data adaptability and improve the stability and reliability of the method. When the input data continuously enters the EOS-ELM system, part of the OS-ELM network model can adapt to the new data faster and better, so that better prediction results can be obtained. The present invention uses the EOS-ELM method to solve the state prediction problem of industrial big data.

Aiming at the appealing problem, the inventor proposed a state prediction method based on industrial big data mining.

Contents of the invention

The purpose of the present invention is to provide a state prediction method based on industrial big data mining, so as to solve the problem that the system state is difficult to predict in the current industrial system, and improve the stability and reliability of the prediction.

In order to achieve the above object, the present invention adopts the following technical solutions:

A state prediction method based on industrial big data mining, comprising the following steps:

Step 1, data collection: use the samples reflecting the historical operation status of the system as the training set where _xi is the system state variable, i.e. the input of the model, and t _i is the predictor of interest, i.e. the output of the model;

Step 2, OS-ELM model: use the training samples from step 1 Establish several OS-ELM models and calculate several predicted values;

Step 3, EOS-ELM model: average the prediction results of the OS-ELM model to obtain the prediction result of the EOS-ELM model.

The OS-ELM model building process includes:

Before initialization, first determine the initial parameters of the network: the network has L hidden nodes, first determine the hidden node type, and the hidden node type includes RBF or additive hidden nodes;

In the initialization phase, from the training samples Select some samples from Initialize, the initialization phase includes the following steps:

Step 1: Randomly assign values to the input parameters

Among them, for the RBF hidden node, the parameters are the center point a _i and the influence factor b _i ; for the additive hidden node, the parameters are the input weight a _i and the deviation b _i ;

Step 2: Calculate the initial hidden layer output matrix H ₀

Among them, when using RBF to hide nodes, G(a _i , b _i , x _j )=g(b _i ||x _j -a _i ||), b _i ∈ R ⁺ , when using additive hidden nodes, G(a _i , b _i , x _j )=g(a _i x _j +b _i ), b _i ∈ R;

Step 3: Estimate initial output weights β ⁽⁰⁾

make The weight solving problem is transformed into minimizing ||H ₀ β-t ₀ ||;

From the solution of the ELM algorithm, it can be seen that The solution result of minimizing ||H ₀ β-t ₀ || is in

Step 4: let k=0, k represents the number of data blocks added to the network;

Continuous learning phase, including the following steps:

The observation value of the k+1 data block is:

Among them, N _k+1 is the number of observations of k+1 data blocks;

Step 5: Calculate the local hidden layer output matrix H _k+1

Step 6: Setting Parameters

Step 7: Calculate the output weight β ^(k+1)

Step 8: set k=k+1, return to step 5;

When all the training data participate in the training, the loop ends and the prediction output is calculated, which is the OS-ELM prediction value;

Steps 1-8 are repeated J times, where J is the number of OS-ELM models.

After the continuous learning phase, a model evaluation phase is also included, which is specifically:

Using the parameters a, b, weight β generated in the continuous learning stage and the input data of the test, the predicted output is obtained, that is, the predicted value of EOS-ELM.

Taking the average of the prediction results of the OS-ELM model is specifically: for the variable x _i , the output of each OS-ELM model is f ^j ( _xi ), j=1,...,J, Then the predicted output of EOS-ELM is

Before the EOS-ELM algorithm, variable selection and feature extraction methods are used to determine output-related variables for several predicted values obtained in step 2, so as to reduce the data dimension.

After adopting the above scheme, the beneficial effects of the present invention are: the present invention averages the prediction results of several OS-ELM models to obtain the prediction results of the EOS-ELM method, which can avoid the possibility that the OS-ELM model may encounter the worst prediction As a result, this method can accurately monitor the system status online in real time, has strong generalization ability, high learning rate, more significant training effect on massive data, low computing cost, and can also lay a foundation for early warning and fault diagnosis. Compared with the traditional OS-ELM method, it can also avoid the worst possible prediction results, enhance the adaptability of the model to new data, and improve the stability and reliability of prediction. In addition, the method can also be used in battery life estimation, load forecasting and other fields, and has a wide range of applications and strong practicability.

The present invention will be further described below in conjunction with the accompanying drawings.

Description of drawings

Fig. 1 is the flow chart of state prediction based on industrial big data mining in the present invention;

Fig. 2 is the prediction output diagram of the EOS-ELM method of nonlinear numerical simulation;

Fig. 3 is a comparison chart of prediction error between EOS-ELM and OS-ELM of nonlinear numerical simulation.

detailed description

As shown in Figure 1, a state prediction method based on industrial big data mining disclosed in this embodiment specifically includes the following steps:

Step 1, data collection: use the samples reflecting the historical operation status of the system as the training set where _xi is the system state variable, i.e. the input of the model, and t _i is the predictor of interest, i.e. the output of the model;

Step 2, OS-ELM model: use the training samples from step 1 Establish several OS-ELM models and calculate several predicted values;

Step 3, EOS-ELM model: average the prediction results of the OS-ELM model to obtain the prediction result of the EOS-ELM model.

The OS-ELM model building process includes:

Before initialization, first determine the initial parameters of the network: the network has L hidden nodes, and determine the type of hidden nodes;

In the initialization phase, from the training samples Select some samples from Initialize, the initialization phase includes the following steps:

Step 1: Randomly assign values to the input parameters

Among them, for the RBF hidden node, the parameters are the center point a _i and the influence factor b _i ; for the additive hidden node, the parameters are the input weight a _i and the deviation b _i ;

Step 2: Calculate the initial hidden layer output matrix H ₀

Among them, when using RBF to hide nodes, G(a _i , b _i , x _j )=g(b _i ||x _j -a _i ||), b _i ∈ R ⁺ , when using additive hidden nodes, G(a _i , b _i , x _j )=g(a _i x _j +b _i ), b _i ∈ R;

Step 3: Estimate initial output weights β ⁽⁰⁾

make The weight solving problem is transformed into minimizing ||H ₀ β-t ₀ ||;

From the solution of the ELM algorithm, it can be seen that The solution result of minimizing ||H ₀ β-t ₀ || is in

Step 4: let k=0, k represents the number of data blocks added to the network;

Continuous learning phase, including the following steps:

The observed value of the k+1 data block is:

Among them, N _k+1 is the number of observations of k+1 data blocks;

Step 5: Calculate the local hidden layer output matrix H _k+1

Step 6: Setting Parameters

Step 7: Calculate the output weight β ^(k+1)

Step 8: set k=k+1, return to step 5;

When all the training data has participated in the training, the loop ends, and the predicted output is calculated, that is, the predicted value;

Steps 1-8 are repeated J times, where J is the number of OS-ELM models.

Evaluating the performance index of the present invention mainly includes prediction error, training time and testing time of the algorithm, compared with other forecasting algorithms, the training time of the EOS-ELM algorithm is relatively short, especially in the field of big data, the effect is more prominent, and the testing time is approximately 0, suitable for online real-time prediction simulation, and the EOS-ELM algorithm has a strong generalization ability, which can avoid problems such as local minimization, inappropriate learning rate and overfitting. With the increase of the number of OS-ELM models J , the prediction effect is also constantly improving, and the stability and reliability of the method are constantly improving, but the training time is also increasing, and a reasonable J needs to be set according to actual needs.

After the continuous learning phase, a model evaluation phase is also included, which is specifically:

Using the parameters a, b, weight β generated in the continuous learning stage and the input data of the test, the predicted output, that is, the predicted value, is obtained.

Taking the average of the prediction results of the OS-ELM model is specifically: for the variable _xi , the output of each OS-ELM model is f ^j ( _xi ), j=1,...,J, then EOS - The predicted output of the ELM is

Before the EOS-ELM algorithm, in order to further reduce the calculation amount of the algorithm and improve the online real-time prediction ability of the algorithm, before using the EOS-ELM algorithm, variable selection and feature extraction methods can be used, for example, variable selection methods based on partial least squares, Determine the variables related to the output, reduce the data dimension, thereby reducing the computational complexity and improving the online real-time prediction performance.

The following is an example of the application of the present invention with actual data, to more clearly illustrate the characteristics of the present invention by adopting a non-linear numerical example, and illustrate its application process in conjunction with the principle of the present invention:

Nonlinear numerical examples are:

Among them, the variable t conforms to the uniform distribution on [-1,1], ε _i (i=1,2,3,4) is the noise uniformly distributed on [-0.1,0.1], and y is the output;

In the simulation, there are 10,000 training data, 1,000 test data, the number of hidden nodes is 25, the number of samples in the initial stage is 100, and the number of samples in each block in the continuous learning stage is 1;

Step 1, use RBF hidden nodes to randomly initialize the input a _i and the impact factor b _i ;

Step 2, calculate the initial hidden layer output matrix H ₀ , H ₀ ∈ R ^100×25 , the calculation formula of each element is G(a _i , b _i , x _j )=g(b _i ||x _j - a _i ||);

Step 3, calculate the initial output weight β ⁽⁰⁾ :

Step 4, let k=0

Step 5, calculate the local hidden layer output matrix H _k+1 , H _k+1 ∈ R ^1×25 ;

Step 6, setting parameter T _k+1 ,

Step 7, calculate the output weight β ^(k+1) ;

Step 8, make k=k+1, return to step 5 until k=9900;

Then, use the parameters a, b and test data obtained from training to calculate the hidden layer output matrix H _te , combined with the output weight β obtained from training, use the formula Y _te = H _te β to calculate the predicted output;

Repeat the above steps 1-8 respectively to obtain the predicted values J=5, J=10, J=15, J=20, J=25 and J=30, and take the average value of the predicted results, which is EOS-ELM;

Referring to Fig. 2 is the EOS-ELM prediction error figure of J=10; Compare OS-ELM method and EOS-ELM method, repeat 50 simulation experiments, the result of the standard deviation of RMSE (standard error) and test is as shown in table 1 Show;

The standard deviation of the test is the standard deviation of 50 RMSE values in the test process after repeated simulations 50 times, which can reflect the stability and reliability of the method; Table 1 below:

The indicators to measure the performance of the method are training time, testing time, training accuracy, testing accuracy and stability, etc., as shown in Table 2;

Table 2

The prediction error comparison of OS-ELM is shown in Figure 3, and the comparison between Figure 2 and Figure 3 further demonstrates the high accuracy of the EOS-ELM method.

The above description shows and describes the preferred embodiments of the present invention, it should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments , and can be modified within the scope of the inventive concept herein through the above teachings or techniques or knowledge in related fields. However, changes and changes made by those skilled in the art do not depart from the spirit and scope of the present invention, and should all be within the protection scope of the appended claims of the present invention.

Claims (5)

1. it is a kind of based on industrial big data excavate trend prediction method, it is characterised in that comprise the following steps：

Step one, data acquisition：The sample of System History running status will be reflected as training set Wherein x_iSystem state variables, i.e. the input of model, t_iIt is the prediction index of concern, the i.e. output of model；

Step 2, OS-ELM models：Using the training sample of step oneSeveral OS-ELM models are set up, and if being calculated Dry predicted value；

Step 3, EOS-ELM models：Predicting the outcome for OS-ELM models is averaged, EOS-ELM model prediction knots are obtained Really.

2. a kind of trend prediction method excavated based on industrial big data as claimed in claim 1, it is characterised in that described OS-ELM models set up process to be included：

Before initialization, network initial parameter is first determined：Network has L implicit node, it is first determined implicit node type, implies Node type includes implying node for RBF or additive；

Initial phase, from training sampleMiddle selected part sampleN₀>=L is initialized, including following Step：

Step 1：At random to |input paramete assignment

Wherein, node, point a centered on parameter are implied to RBF_iWith factor of influence b_i；Node is implied to additive, parameter is defeated Enter weight a_iWith deviation b_i；

Step 2：Calculate initial hidden layer output matrix H₀

$H_0={\left[\begin{array}{ccc}G(a_1,b_1,x_1)& \mathrm{...}& G(a_L,b_L,x_1)\\ .& & .\\ .& \mathrm{...}& .\\ .& & .\\ G(a_1,b_1,x_{N_0})& \mathrm{...}& G(a_L,b_L,x_{N_0})\end{array}\right]}_{N_0\×L}$

Wherein, when implying node using RBF, G (a_i,b_i,x_j)=g (b_i||x_j-a_i| |), b_i∈R⁺, implied when using additive During node, G (a_i,b_i,x_j)=g (a_i·x_j+b_i), b_i∈R；

Step 3：Estimate initial output weight beta⁽⁰⁾

OrderWeight Solve problems are converted into minimum | | H₀β-t₀||；

From the solution of ELM algorithms, it is known thatMinimize | | H₀β-t₀| | solving result beWherein

Step 4：K=0, k is made to represent the quantity of the data block for adding network；

In the successive learning stage, comprise the following steps：

The observation of k+1 data blocks is：

Wherein, N_k+1It is the quantity of the observation of k+1 data blocks；

Step 5：Calculate local hidden layer output matrix H_k+1

$H_{k+1}={\left[\begin{array}{ccc}G(a_1,b_1,x_{i=\left({\Σ}_{j=0}^k{N}_j\right)+1})& \mathrm{...}& G(a_L,b_L,x_{i=\left({\Σ}_{j=0}^k{N}_j\right)+1})\\ .& & .\\ .& \mathrm{...}& .\\ .& & .\\ G(a_1,b_1,x_{i=\left({\Σ}_{j=0}^{k+1}{N}_j\right)+1})& \mathrm{...}& G(a_L,b_L,x_{i=\left({\Σ}_{j=0}^{k+1}{N}_j\right)+1})\end{array}\right]}_{N_{k+1}\×L}$

Step 6：Arrange parameter

$T_{k+1}={\[t_{i=\left({\Σ}_{j=0}^k{N}_j\right)+1},\mathrm{...},t_{{\Σ}_{j=0}^{k+1}{N}_j}\]}_{N_{k+1}\×m}^T$

Step 7：Calculate output weight beta^(k+1)

$P_{k+1}=P_k-P_k{H}_{k+1}^T{(I+H_{k+1}{P}_k{H}_{k+1}^T)}^{-1}{H}_{k+1}{P}_k$

${\mathrm{\β}}^{(k+1)}={\mathrm{\β}}^{\left(k\right)}+P_{k+1}{H}_{k+1}^T(T_{k+1}-H_{k+1}{\mathrm{\β}}^{\left(k\right)})$

Step 8：Make k=k+1, return to step 5；

When all training datas are involved in training, circulation terminates, and calculates prediction output, i.e. OS-ELM predicted values；

Step 1-8 is repeated J times, and J is OS-ELM model quantity.

3. a kind of trend prediction method excavated based on industrial big data as claimed in claim 2, it is characterised in that step 2 In, the model evaluation stage is also included after the successive learning stage, the stage is specially：

The input data of parameter a, b, weight beta and test that are produced using the successive learning stage, obtains prediction output, i.e. EOS- ELM predicted values.

4. a kind of trend prediction method excavated based on industrial big data as claimed in claim 2, it is characterised in that to described Predicting the outcome for OS-ELM models is averaged specially：To the variable x_i, each OS-ELM model is output as f^j (x_i), j=1 ..., J, the then prediction of EOS-ELM is output as

5. a kind of trend prediction method excavated based on industrial big data as claimed in claim 1, it is characterised in that：Described Before EOS-ELM algorithms, several predicted values obtained to step 2 using variables choice and feature extracting method are determined and output Related variable, to reduce data dimension.