WO2022121923A1 - Smart modelling method and apparatus of complex industrial process digital twin system, device, and storage medium - Google Patents

Abstract

A smart modelling method and apparatus of a complex industrial process digital twin system, a device, and a storage medium. The smart modelling method of the complex industrial process digital twin system comprises: establishing a mechanism model of a complex industrial process, the mechanism model comprising the two parts of an identifiable model and unmodelled dynamics; estimating parameters of the identifiable model to obtain an identification model; using the error of the identification model and the unmodelled dynamics to form an unknown nonlinear dynamic system; establishing a smart model of the unknown nonlinear dynamic system; and using the identification model and the smart model of the unknown nonlinear dynamic system to establish a smart model of the complex industrial process digital twin system; the error of the identification model is a model output error caused when the parameters in the identifiable model are replaced with identification values thereof. Aimed at the problem of the difficulty of ensuring the precision of a complex industrial process digital twin system, a mechanism model-based system identification method is combined with a big data-based deep learning method, and an end-side cloud collaboration mode is used to establish a smart model of the complex industrial process digital twin system, solving the modelling problem of complex industrial process digital twin systems and increasing the modelling precision.

Description

Intelligent modeling method, device, equipment and storage medium for complex industrial process digital twin system technical field

The invention belongs to the technical field of industrial artificial intelligence, and in particular relates to an intelligent modeling method, device, equipment and storage medium of a complex industrial process digital twin system.

Background technique

There are a large number of complex industrial processes in steel, metallurgy, mineral processing, petrochemical, electric power and other process industries. The mechanism models of these industrial processes have the following complexities: the models contain nonlinear terms between input and output variables, multivariable strong coupling , unknown frequently changing disturbances, unknown structure of some models, unknown order of input and output variables, and unknown dynamic characteristics of changes. Therefore, the existing system identification methods based on mechanism models cannot establish dynamic models of these industrial processes. The interaction of material flow, information flow, and energy flow in the production process makes the dynamic characteristics of these industrial processes change unknown with the production time, resulting in the input and output data of these processes being changed, open, and uncertain Therefore, the existing deep learning technology in the complete information space cannot establish the dynamic model of these industrial processes. At present, the identification of operating conditions of these industrial processes and the decision-making of the set value of the process control system still rely on manual identification and decision-making by operators and engineering technicians based on experience and knowledge. Due to the difficulty of human beings to perceive working condition information in a timely and accurate manner, process multi-source heterogeneous information, coupled with human subjectivity and uncertainty, it is difficult to achieve high-performance control and operation optimization of industrial processes, resulting in unstable product quality, energy and material consumption. high. Establishing the digital twin system of industrial process is the key to realize the integration of optimized operation and control of industrial process. However, due to the above-mentioned characteristics of complex industrial processes, it is difficult to use existing mechanism modeling methods or deep learning methods to establish a digital twin system that meets the accuracy requirements.

SUMMARY OF THE INVENTION

The present invention aims to solve one of the technical problems in the related art at least to a certain extent. The technical scheme of the present invention is as follows:

An intelligent modeling method for a complex industrial process digital twin system, comprising the following steps:

Establish a mechanism model of a complex industrial process, the mechanism model includes an identifiable model and an unmodeled dynamic part;

Estimating the parameters of the identifiable model to obtain an identifiable model;

Using the identified model error and the unmodeled dynamic to form an unknown nonlinear dynamic system;

establishing an intelligent model of the unknown nonlinear dynamic system;

Using the identification model and the intelligent model of the unknown nonlinear dynamic system to establish an intelligent model of the complex industrial process digital twin system;

Wherein, the identification model error is the model output error caused when the parameters in the identifiable model are replaced by their identification values.

Further, preferably, the intelligent model of the unknown nonlinear dynamic system includes an offline deep learning model, an online deep learning model, a deep learning correction model and a self-correction mechanism; LSTM is used to establish the offline deep learning model; The online deep learning model is established with the same structure as the deep learning model; the deep learning correction model is established with the same structure as the offline deep learning model; when the output of the intelligent model of the complex industrial process digital twin system is the same as the When the error between the actual outputs of the complex industrial process is greater than the set threshold, a self-correction mechanism is adopted, and the connection weight parameters and bias parameters of the deep learning correction model are used to correct the connection weight parameters and bias of the online deep learning model parameter; wherein, the historical data used by the deep learning correction model is more than the historical data used by the online deep learning model.

Further, preferably, both the online deep learning model and the deep learning correction model include an input layer, a hidden layer, a fully connected layer and an output layer; the connection weight parameters of the hidden layer in the online deep learning model are fixed. and bias parameters, and online correction of the connection weight parameters and bias parameters of the fully connected layer in the online deep learning model; online training of the hidden layer and the fully connected layer in the deep learning correction model When the error between the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process is greater than the set threshold, the The connection weight parameters and bias parameters of the hidden layer of the deep learning correction model replace the connection weight parameters and bias parameters of the hidden layer in the online deep learning model, and the full value of the deep learning correction model is used. The connection weight parameters and bias parameters of the connection layer replace the connection weight parameters and bias parameters of the fully connected layer in the online deep learning model.

Further, preferably, the complex industrial process is an fused magnesia smelting process.

An intelligent modeling device for a complex industrial process digital twin system, comprising:

The mechanism model modeling module is used to establish a mechanism model of a complex industrial process, and the mechanism model includes two parts: an identifiable model and an unmodeled dynamic;

A parameter identification module for estimating the parameters of the identifiable model to obtain an identification model;

an unknown nonlinear dynamic acquisition module, used for using the identified model error and the unmodeled dynamic to form an unknown nonlinear dynamic system;

an intelligent model modeling module of an unknown nonlinear dynamic system, used for establishing an intelligent model of the unknown nonlinear dynamic system;

A digital twin system intelligent model modeling module, used for establishing an intelligent model of the complex industrial process digital twin system by using the identification model and the intelligent model of the unknown nonlinear dynamic system;

Wherein, the identification model error is the model output error caused when the parameters in the identifiable model are replaced by their identification values.

Further, preferably, the unknown nonlinear dynamic system intelligent model modeling module includes an offline deep learning model modeling module, an online deep learning model modeling module, a deep learning calibration model modeling module and a self-calibration module; the offline depth The learning model modeling module adopts LSTM to build the offline deep learning model; the online deep learning model modeling module adopts the same structure as the offline deep learning model to build the online deep learning model; the deep learning correction model builds the online deep learning model. The model module adopts the same structure as the offline deep learning model to establish the deep learning correction model; the self-correction module, the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process When the error between them is greater than the set threshold, a self-correction mechanism is used to correct the connection weight parameters and bias parameters of the online deep learning model with the connection weight parameters and bias parameters of the deep learning correction model; wherein, the The deep learning correction model uses more historical data than the online deep learning model.

Further, preferably, the online deep learning model and the deep learning correction model include an input layer, a hidden layer, a fully connected layer and an output layer; the online deep learning model modeling module is fixed in the online deep learning model The connection weight parameters and bias parameters of the hidden layer, and the connection weight parameters and bias parameters of the fully connected layer in the online deep learning model are corrected online; the deep learning correction model modeling module is trained online The connection weight parameters and bias parameters of the hidden layer and the fully connected layer in the deep learning correction model; the self-correction module, in the output of the intelligent model of the complex industrial process digital twin system, is the same as the When the error between the actual outputs of the complex industrial process is greater than the set threshold, a self-correction mechanism is used to replace the connection weight parameters and bias parameters of the hidden layer of the deep learning correction model in the online deep learning model. The connection weight parameters and bias parameters of the hidden layer are replaced by the connection weight parameters and bias parameters of the fully connected layer of the deep learning correction model to replace the connection of the fully connected layer in the online deep learning model weight parameters and bias parameters.

Further, preferably, the complex industrial process is an fused magnesia smelting process.

A device for realizing an intelligent modeling method for a digital twin system of a complex industrial process, including: terminal-side sub-devices, edge-side sub-devices, and cloud-side sub-devices;

The terminal-side sub-device is used to obtain the input data of the online deep learning model and the deep learning correction model;

The edge side sub-device is used to run the identification model and the online deep learning model;

The cloud side sub-device is used to run the deep learning correction model;

The terminal-side sub-device is a process control system of the complex industrial process, the edge-side sub-device is an edge computing device, and the cloud-side sub-device is an industrial cloud.

A computer-readable storage medium storing a computer program, when the program is executed by a processor, realizes the above-mentioned intelligent modeling method for a digital twin system of a complex industrial process.

The invention combines the system identification method based on the mechanism model with the deep learning method based on big data, adopts the terminal-edge-cloud collaboration method, establishes the intelligent model of the digital twin system of the complex industrial process, and solves the problem of the construction of the digital twin system of the complex industrial process. The model problem is improved, and the modeling accuracy is improved.

Description of drawings

Fig. 1 is the realization flow chart of the intelligent modeling method of complex industrial process digital twin system according to the embodiment of the present invention;

Fig. 2 is the realization flow chart of the intelligent modeling method of the digital twin system of the fused magnesia smelting process according to an embodiment of the present invention;

3 is a structural diagram of an LSTM network according to an embodiment of the present invention;

4 is a node diagram of an LSTM unit according to an embodiment of the present invention;

Fig. 5 is the variation curve of the error with the number of neurons of an embodiment of the present invention;

Fig. 6 is the variation curve of the error with the number of element nodes of an embodiment of the present invention;

Fig. 7 is the variation curve of the error with the number of network layers of an embodiment of the present invention;

Fig. 8 is the variation curve of the error with the length of the time series window according to an embodiment of the present invention;

9 is a structural diagram of a digital twin system of an fused magnesia smelting process according to an embodiment of the present invention;

FIG. 10 is a schematic structural diagram of an apparatus for intelligent modeling of a complex industrial process digital twin system according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Fig. 1 is the realization flow chart of the complex industrial process digital twin system intelligent modeling method of the embodiment of the present invention, and the method comprises the following steps:

S1: Establish a mechanism model of a complex industrial process, the mechanism model includes an identifiable model and an unmodeled dynamic part.

Specifically, an unknown constant F is used to represent a nonlinear function F(·) whose model parameters are unknown changes, and a new variable is used.

Represents the nonlinear term of the input _ui (k) and output y _i (k) in the mechanism model, and uses the unknown nonlinear function V( ) to represent the unknown part of the model structure and the unknown order of input and output variables, unknown disturbance, unknown Coupling and dynamic characteristics of unknown changes, so that the mechanism models of complex industrial processes can be represented by identifiable models and unmodeled dynamics.

S2: Estimate the parameters of the identifiable model to obtain an identifiable model.

Specifically, the input and output data u _i (k) and y _i (k) of the process are used to obtain new variable data representing the nonlinear term between the input and output

Use input and output data _ui (k), _yi (k) and new variable data

The identification algorithm is used to estimate the parameters of the identifiable model, so as to obtain the identification model.

S3: Using the identified model error and the unmodeled dynamics to form an unknown nonlinear dynamic system.

Specifically, the identification model error is the model output error caused when the parameters in the identifiable model are replaced by their identification values. Using the identified model errors and the unmodeled dynamics to form an unknown nonlinear dynamic system represented by:

in,

f is a nonlinear function of unknown change, y _i (k) is the ith phase output of the mechanism model at time k, u _i (k) is the ith phase input of the mechanism model at time k,

is a new variable representing the nonlinear term between _ui (k) and y _i (k),

To identify the output of the model,

is the output of the intelligent model of the unknown nonlinear dynamic system at time k-1, and n is the order of the unknown nonlinear dynamic system.

S4: Build an intelligent model of the unknown nonlinear dynamic system.

Specifically, the intelligent model of the unknown nonlinear dynamic system includes an offline deep learning model, an online deep learning model, a deep learning correction model and a self-correction mechanism.

The multi-layer long and short-term memory network LSTM is used, the input variable in formula (1) is selected as the input of a single neuron, the order n is used as the number of neurons, v(k) is used as the label data, and the input of formula (1) is used. , The output data constitutes a large data sample, and a training algorithm is used to make the error Δv (k) between the label data v (k) and the output v (k) of the offline deep learning model as small as possible, and determine the number of neurons in the offline deep learning model. The number n, the number of unit nodes h of the LSTM, the number of layers of the neural network L, the connection weight parameters and bias parameters of each layer, so as to establish an offline deep learning model.

The online deep learning model is established with the same structure as the offline deep learning model, that is, the online deep learning model is established by using LSTM, and the number of neurons, unit nodes and neural network layers of the online deep learning model are the same as those of the offline deep learning model. The same is true for deep learning models. The initial value of the connection weight parameter and bias parameter of each layer of the online deep learning model adopts the connection weight parameter value and bias parameter of the corresponding layer of the offline deep learning model, and the connection weight parameter and the bias parameter of the hidden layer in the online deep learning model are fixed. Bias parameter, using the real-time updated input data and label data of the time series window length N to correct the connection weight parameters and bias parameters of the fully connected layer of the online deep learning model, where the time series window length N is obtained through training Algorithm to make the labeled data v(k) with the output of the online deep learning model

The error Δv(k) is as small as possible.

Using the same structure as the offline deep learning model, a deep learning correction model is established. That is, LSTM is used to establish a deep learning correction model, and the number of neurons, unit nodes and neural network layers of the deep learning correction model are the same as those of the offline deep learning model. The input data and output data of formula (1) at the current moment and all previous moments are used as the input data and label data of the deep learning correction model, and all the connection weight parameters and bias parameters of each layer of the deep learning correction model are trained.

The upper bound of the error interval set by the self-correction mechanism is δ. When the error between the output of the online deep learning model and the output of the identification model superimposed at time k and the real output value of the actual industrial process at time k |Δr(k)|>δ, the deep learning correction model is used for each error. The connection weight parameters and bias parameters of the layers replace the connection weight parameters and bias parameters of the corresponding layers of the online deep learning model.

Among them, the input data of the online deep learning model and the deep learning correction model are obtained by the end-process control system, the online deep learning model runs on the edge-edge computing device, and the deep learning correction model runs on the cloud-industrial cloud.

S5: Use the identification model and the intelligent model of the unknown nonlinear dynamic system to establish the intelligent model of the digital twin system of the complex industrial process.

Further, in one embodiment, the intelligent modeling method of the complex industrial process digital twin system can be used in the fused magnesia smelting process.

The fused magnesia smelting process is a strong nonlinear and strong coupling industrial process with the rotation direction and frequency of the three-phase motor as the input and the three-phase electrode current as the output. The submerged arc method is adopted in the smelting process, and the material is fed while melting. The height of the molten pool changes with the continuous addition and melting of the raw ore, the length of the raw ore particles and the change of the impurity composition. The resistivity of the molten pool changes with the temperature of the melt, the length of the raw ore particles and Changes in impurity composition. Therefore, the model order of the melting current dynamic model is unknown and the dynamic characteristics change is unknown, so the mathematical model cannot be established by the system identification method based on the mechanism model. The input and output data of this process are in a changing, open and uncertain information space, so it is impossible to use the existing deep learning technology of complete information space to establish its dynamic model.

Fig. 2 is the realization flow chart of the intelligent modeling method of the digital twin system of the fused magnesia smelting process of an embodiment of the present invention, and the method comprises the following steps:

S1': Establish the electrode current mechanism model in the smelting process of fused magnesia, and represent the mechanism model with an identifiable model and an unknown nonlinear function.

Specifically, in the smelting process of fused magnesia, the material is added while melting, and the molten pool is continuously increased with the continuous addition and melting of the raw ore.

The relationship between is:

Among them, y _i (t) is the electrode current of the i-th phase at time t, y is the optimal melting current; k _h (y, y _i (t)) is the conversion coefficient between the cumulative feeding amount and the height of the molten pool, the The conversion coefficient changes with the fluctuation of the electrode current y _i (t) around the optimal melting current y; _{m 0} is the initial feeding amount; B ₁ is the length of the raw ore particle; B ₂ is the impurity composition of the raw ore; km (B ₁ , B ₂ ) is the conversion coefficient between the electric vibration frequency and the feeding speed, and the conversion coefficient changes with the changes of B ₁ and B ₂ ;

is the vibration frequency of the electro-vibration feeder at time t; σ(t) is the start-stop sign of the electro-vibration feeder, σ(t)=1 means start, σ(t)=0 means stop;

and σ(t) are determined according to the real-time smelting state.

When h(·) changes according to formula (4) with feeding and melting, the current control system adjusts the distance between the electrode and the molten pool by dragging the motor, so that the electrode current stably tracks the optimal melting current setting value to melt the diamond. Magnesium Ore. In the smelting process of fused magnesia, the electrode current dynamic mechanism model with the rotation direction and frequency u _i (t) of the motor as the input and the electrode current y _i (t) as the output is as follows:

Among them, i=1, 2, 3 represent the three phases of A, B, and C, respectively, Li is the self-inductance of the _{i-th phase circuit, F i (} ) is the unknown nonlinear function of the i-th phase circuit, and s is the motor rotation difference ratio, r _d is the equivalent gear radius of the lifting mechanism, p is the number of pole pairs of the motor, g ₀ is the arc conductivity constant, _ra is the arc column radius, T ₀ is the gas ionization temperature constant, and T ₁ is the arc gap temperature , u _i (t) is the motor rotation direction and frequency of the i-th phase circuit, U _i is the phase voltage of the i-th phase circuit,

is the influence of the three-phase circuit coupling on the i-th phase current.

F _i ( ) is shown in formula (6):

Among them, Δh _ie ( _· ) is the change of electrode height when the electrode is subjected to the electromagnetic force generated by the feeding disturbance and the changing mutual inductance _; changes; f ₀ is the proportional coefficient of the molten pool resistance; S is the cross-sectional area of the molten pool.

As shown in formula (7):

Among them, M ₁₂ ( ) is the mutual inductance between A and B-phase circuits, M ₁₃ ( ) is the mutual inductance between A and C-phase circuits, M ₂₃ ( ) is the mutual inductance between B and C-phase circuits, M ₁₂ (·), M ₁₃ (·) and M ₂₃ (·) vary with the addition and melting of raw ore and the change of melt temperature.

Replace the unknown nonlinear function F _i (·) in equation (5) with an unknown constant F _i , and use the unknown nonlinear function for the model error, the dynamic characteristics of the unknown change, and the coupling of the unknown change.

Representation, that is, equation (5) can be expressed as:

in,

for:

Using Euler's method to discretize equation (8), we can get:

in,

is the unknown nonlinear function shown in (9)

discrete form.

In order to eliminate the integral summation expression of u _i (k ₀ ) in equation (10), divide both sides of equation (10) by y _i (k) and subtract it from the expression at time k-1 to get:

Since y _i (k+1) at time k is unknown, set k in equation (11) equal to k-1, and the discretized current mechanism model can be obtained as:

take a new variable

and

represents the nonlinear term between the input and output in Eq. (12), namely:

The unknown nonlinear function V _i (k) is used to represent the unknown part of the model structure and the unknown order of the input and output variables, the unknown disturbance, the coupling of the unknown change and the dynamic characteristics of the unknown change, let

Substitute _Qi and (13) into _equation (12), so that the electrode current mechanism model (12) in the smelting process of fused magnesia can be expressed as:

Among them, V _i (k) is:

S2': Estimate the identifiable model parameters to obtain the identification model.

Specifically, using the actual process input and output data u _i (k) and y _i (k), the new variable data representing the nonlinear term between the input and output in the formula (14) is obtained from the formula (13).

and

The parameter Q _i in the identifiable model (14) is estimated by the least square identification algorithm, and its estimated value is obtained

for:

Therefore, the output of the identification model can be obtained

for:

S3': construct an unknown nonlinear dynamic system vi ₍ k) by using the identification model error and the unknown nonlinear function V _i (k), and express the mechanism model as the sum of the output of the identification model and the output of the unknown nonlinear dynamic system.

Specifically, using the identification model error

and the unknown nonlinear function V _i (k) form an unknown nonlinear dynamic system v _i (k) expressed by the following formula:

Then v _i (k) can be represented by an unknown variable nonlinear function f _i , namely:

where, the input data vector

for:

From equations (14), (17) and (18), the output vi ( _k ) of the unknown nonlinear dynamic system can be obtained as:

From the formula (21), the current mechanism model of the fused magnesia smelting process can be expressed as the output of the identification model

and the output of the unknown nonlinear dynamic system v _i (k), namely:

S4': Build an intelligent model of the unknown nonlinear dynamic system vi (k) ( _i =1, 2, 3).

Specifically, the intelligent model consists of an offline deep learning model, an online deep learning model, a deep learning correction model and a self-correction mechanism.

S41': Use the multi-layer long and short-term memory network LSTM architecture as shown in Figure 3 to establish an offline deep learning model of v _i (k), specifically:

Select the input data vector in (19)

The order n is used as the number of neurons in each layer; the data vector

As input data, it is input to the _n neurons of the first layer respectively; the label data is v(k)=[v ₁ (k), v ₂ (k), v ₃ (k)] ^T when training the network. Let the number of nodes of a single neuron in LSTM be h, the number of network layers is L, the current moment is time k, and the output of the n _i neuron in the n _l layer is

Figure 4 shows the cell node diagram of the LSTM.

For layer 1 (n _l =1), the input to the n _i neuron is:

The output of the n _i neuron in layer 1

for:

Among them, ⊙ represents the multiplication of elements at the corresponding positions of the two vectors; tanh is the hyperbolic tangent function, that is, tanh(x)=(e ^x -e ^-x )/(e ^x +e ^-x );

is the output gate output,

is the state of the LSTM unit, as shown in the following formula:

Among them, σ is a sigmoid function, that is, σ(x)=(1+e ^-x ) ^-1 ;

output for the forget gate,

is the input gate output,

is the state candidate value, as shown in the following formula:

(25) and (26),

and

is h×(h+15) dimension weight matrix,

and

is an h × 1-dimensional offset column vector.

When n _l ≥ 2, the output of the n _i neuron in the n _l layer is the same as the output of the n _i neuron in the first layer shown in equations (24) to (26), the difference lies in the weight matrix

and

The dimension of is h×(h+h), the input of the n _i neuron is the output of the n _i neuron in the n _l -1 layer, namely:

Among them, when calculating the input and output of the first neuron in each layer,

and

The value of is determined by random initialization when k=0, and its value is equal to the output and state of the last neuron in the layer at the previous moment when k≥1, namely

Put the output of the nth neuron in the Lth layer

Input to the fully connected layer, and the output of the fully connected layer is the output v (k) of the offline deep learning model, as shown in the following formula:

Among them, W _d ∈ R ^3×h and b _d ∈ R ³ are the weights and biases of the fully connected layer.

The actual process data of two heats (10 hours for each heat and 36,000 sets of data) are used as the training set and the test set respectively, and M (M=36000) input and output data of formula (19) are used to form big data Sample, use the following training algorithm to make the model error Δ v (k) between the label data v (k) and the offline deep learning model output v (k) as small as possible to train the network, determine n, h and L, the specific operations are as follows:

Let the number of network layers L=1, the objective function of the training algorithm is:

Where, Δ v (k)=v(k) -v (k), v(k) and v (k) are shown in equations (30) and (31):

in,

It can be obtained from equation (24). Based on error back propagation, gradient descent algorithm is used to train network weights

W _d and Bias

b _d . The training algorithm for all weights and biases is the same, and the output gate weights are used below.

The training is described as an example, and the objective function J is about

The partial derivative of is:

Update the output gate weights according to

where η is the learning rate.

Other weights and biases are determined using the same training algorithm as equations (32) to (33).

Let the number of cell nodes h of the LSTM be equal to the input time series

The dimension of 15, the number of neural network layers L is equal to 1. Let n increase from 1, and calculate the mean absolute error (MAE) of the model error Δv (k), respectively. It can be seen from Figure 5 that when n=3, the MAE of Δ v (k) is the smallest, so the order n of the dynamic system is 3, that is, the number of neurons in each layer is 3.

Fix n=3, L=1, let h increase from 15, and calculate the MAE of Δ v (k) respectively. It can be seen from FIG. 6 that when h=1770, the MAE of Δ v (k) is the smallest, so the number h of unit nodes is 1770.

Fix n=3, h=1770, let L increase from 1, and calculate the MAE of Δ v (k). It can be seen from Figure 7 that when L=7, the MAE of Δ v (k) is the smallest, so the number of neural network layers L is 7.

S42’: Use the same structure as the offline deep learning model to build an online deep learning model, specifically:

LSTM is used to establish an online deep learning model, and the number of neurons, unit nodes and neural network layers of the online deep learning model are the same as those of the offline deep learning model. The connection weight parameters of each layer of the online deep learning model and The initial value of the bias parameter adopts the connection weight parameter value and bias parameter of the corresponding layer of the offline deep learning model, and the connection weight parameter and bias parameter of the 1st to 7th layers in the online deep learning model are fixed, and the time series window length is used. Correct the connection weight parameter W _d (k) and bias parameter b _d (k) of the fully connected layer of the online deep learning model for the real-time updated input data and label data of N, and obtain the online deep learning of v(k) model output

for:

Among them, W _d (k)∈R ^3×h and b _d (k)∈R ³ are the weights and biases of the fully connected layer,

is the output of the third neuron in the seventh layer.

At time k+1, the updated dataset is

for:

The objective function and correction algorithm for online correction of W _d (k+1) and b _d (k+1) are shown in equations (36) and (37), respectively:

in,

Represents labeled data v(k) and online deep learning model output

model error,

and

As shown in the following formula:

Fix n=3, h=1770, L=7, adopt the training algorithm shown in equations (36) to (38), let N increase from 1, and calculate the root mean square error (RMSE) of Δv(k) respectively. It can be seen from Figure 8 that when N=1330, the RMSE of Δv(k) is the smallest, so the time series window length N is 1330.

S43': Using the exact same structure as the offline deep learning model, that is, n=3, h=1770, L=7, establish the deep learning correction model shown in the following formula.

in,

is the output of the deep learning correction model at time k+1. The k+1 moment adopts all the data of all the past moments, and the updated data set is

Online training of this deep learning corrects all connection weight parameters and bias parameters of each layer of the model.

The upper bound of the set error interval is δ=100A. At time k+1, if the output of the online deep learning model and the output of the identification model are superimposed, the output of the electrode current intelligent model is obtained.

The error between the actual value y _i (k+1) of the electrode current in the smelting process of fused magnesia

When , all connection weight parameters and bias parameters of each layer of the deep learning correction model are used to correct the connection weight parameters and bias parameters of the corresponding layers of the online deep learning model.

Among them, the input data of the online deep learning model and the deep learning correction model are obtained by the end-PLC control system of the fused magnesia smelting process. Industrial cloud runs.

S5’: Establish the electrode current intelligent model of the digital twin system of the fused magnesia smelting process by using the identification model and the intelligent model of the unknown nonlinear dynamic system.

Specifically, the electrode current intelligent model of the digital twin system of the fused magnesia smelting process is:

Among them, the identification model output

Obtained from (17), the intelligent model output of the unknown nonlinear dynamic system

It can be obtained from equation (34). The structure diagram of the digital twin system of the fused magnesia smelting process is shown in Figure 9.

Select 300 sets of test set data, and use the identification model to output

and in the embodiment of the present invention, the identification model is output

Intelligent model output with unknown nonlinear dynamic systems

The two current models established by addition are compared, and the model outputs are respectively

and

and

They are obtained from equations (17) and (34), respectively. The performance evaluation indicators root mean square error (RMSE) and mean absolute percentage error (MAPE) shown in equations (41) and (42) are used to evaluate the model accuracy. The results are shown in Table 1.

Among them, N ₀ =300 is the number of time steps, y _i (k) is the real value,

Output values for the model.

Table 1 Three-phase current model accuracy performance evaluation table

It can be seen from Table 1 that the modeling accuracy of the embodiment of the present invention is obviously better than the method of using the identification model output, and the RMSEs of the three-phase currents y ₁ (k), y ₂ (k), and y ₃ (k) are reduced respectively. 93.14%, 91.82%, 93.72%, MAPE decreased by 96.22%, 95.23%, 96.14%, respectively. The accuracy of the three-phase current model established by the embodiment of the present invention satisfies the requirements for establishing a digital twin system of the fused magnesia smelting process, and creates conditions for realizing the intelligent optimization control of the fused magnesia smelting process.

In one embodiment, as shown in FIG. 10 , an intelligent modeling device for a digital twin system of a complex industrial process is provided, including: a mechanism model modeling module, a parameter identification module, an unknown nonlinear dynamic acquisition module, an unknown nonlinear dynamic The system intelligent model modeling module and the digital twin system intelligent model modeling module, including:

The mechanism model modeling module is used to establish a mechanism model of a complex industrial process, and the mechanism model includes two parts: an identifiable model and an unmodeled dynamic;

The parameter identification module is used for estimating the parameters of the identifiable model to obtain the identification model;

The unknown nonlinear dynamic acquisition module is used for using the identified model error and the unmodeled dynamic to form an unknown nonlinear dynamic system;

The unknown nonlinear dynamic system intelligent model modeling module is used to establish the intelligent model of the unknown nonlinear dynamic system;

The digital twin system intelligent model modeling module is used to establish the intelligent model of the complex industrial process digital twin system by using the identification model and the intelligent model of the unknown nonlinear dynamic system;

Wherein, the identification model error is the model output error caused when the parameters in the identifiable model are replaced by their identification values.

In one embodiment, the unknown nonlinear dynamic system intelligent model modeling module includes an offline deep learning model modeling module, an online deep learning model modeling module, a deep learning calibration model modeling module and a self-calibration module; the offline depth The learning model modeling module adopts LSTM to build the online deep learning model; the online deep learning model modeling module adopts the same structure as the offline deep learning model to build the online deep learning model; the deep learning correction model builds the online deep learning model. The model module adopts the same structure as the offline deep learning model to establish the deep learning correction model; the self-correction module, the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process When the error between them is greater than the set threshold, a self-correction mechanism is used to correct the connection weight parameters and bias parameters of the online deep learning model with the connection weight parameters and bias parameters of the deep learning correction model; wherein, the The deep learning correction model uses more historical data than the online deep learning model.

In one embodiment, both the online deep learning model and the deep learning correction model include an input layer, a hidden layer, a fully connected layer and an output layer; the online deep learning model modeling module fixes the online deep learning connection weight parameters and bias parameters of the hidden layer in the model, and online correction of the connection weight parameters and bias parameters of the fully connected layer in the online deep learning model; the deep learning correction model modeling module The connection weight parameters and bias parameters of the hidden layer and the fully connected layer in the deep learning correction model are trained online; the self-correction module, in the output of the intelligent model of the complex industrial process digital twin system, is the same as the When the error between the actual outputs of the complex industrial process is greater than a set threshold, a self-correction mechanism is used to replace the online deep learning model with the connection weight parameters and bias parameters of the hidden layer of the deep learning correction model The connection weight parameters and bias parameters of the hidden layer in the deep learning correction model are used to replace the fully connected layer in the online deep learning model with the connection weight parameters and bias parameters of the fully connected layer of the deep learning correction model. The connection weight parameters and bias parameters of .

In one of the embodiments, the complex industrial process is an fused magnesia smelting process.

For specific limitations on the intelligent modeling device for a digital twin system of a complex industrial process, reference may be made to the limitations on an intelligent modeling method for a digital twin system of a complex industrial process above, which will not be repeated here. Each module in the above-mentioned intelligent modeling device for a digital twin system of a complex industrial process can be implemented in whole or in part by software, hardware and combinations thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

In one embodiment, a device for implementing the intelligent modeling method for a digital twin system of a complex industrial process in the above embodiments is provided, including: terminal-side sub-devices, edge-side sub-devices, and cloud-side sub-devices; The terminal side sub-device is used to obtain the input data of the online deep learning model and the deep learning correction model; the edge side sub-device is used to run the identification model and the online deep learning model; the cloud side The sub-device is used for running the deep learning correction model; the terminal-side sub-device is a process control system of the complex industrial process, the edge-side sub-device is an edge computing device, and the cloud-side sub-device is an industrial cloud.

In one embodiment, a computer-readable storage medium is provided, which stores a computer program, and when the program is executed by a processor, realizes the intelligent modeling method for a digital twin system of a complex industrial process in each of the foregoing embodiments.

Those skilled in the art may combine and combine the different embodiments described in this specification and the features of the different embodiments without contradicting each other.

To sum up, the intelligent modeling method, device and equipment for a complex industrial process digital twin system proposed in the embodiments of the present invention are aimed at the problem that the accuracy of the complex industrial process digital twin system is difficult to guarantee. Combining the deep learning methods of big data and using the device-edge-cloud collaboration method, an intelligent model of the complex industrial process digital twin system is established, which solves the modeling problem of the complex industrial process digital twin system and improves the modeling accuracy.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited to this. Any person skilled in the art who is familiar with the technical field disclosed in the present invention can easily think of various changes or Replacement, these should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (10)

1. An intelligent modeling method for a complex industrial process digital twin system, characterized in that the method comprises:

   Establish a mechanism model of a complex industrial process, the mechanism model includes an identifiable model and an unmodeled dynamic part;

   Estimating the parameters of the identifiable model to obtain an identifiable model;

   Using the identified model error and the unmodeled dynamic to form an unknown nonlinear dynamic system;

   establishing an intelligent model of the unknown nonlinear dynamic system;

   Using the identification model and the intelligent model of the unknown nonlinear dynamic system to establish an intelligent model of the complex industrial process digital twin system;

   Wherein, the identification model error is the model output error caused when the parameters in the identifiable model are replaced by their identification values.
2. The method according to claim 1, wherein the intelligent model of the unknown nonlinear dynamic system comprises an offline deep learning model, an online deep learning model, a deep learning correction model and a self-correction mechanism;

   Use LSTM to build the offline deep learning model;

   The online deep learning model is established using the same structure as the offline deep learning model;

   The deep learning correction model is established using the same structure as the offline deep learning model;

   When the error between the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process is greater than a set threshold, a self-correction mechanism is adopted, and the deep learning is used to correct the connection weight parameters of the model and The bias parameter corrects the connection weight parameter and the bias parameter of the online deep learning model;

   Wherein, the historical data used by the deep learning correction model is more than the historical data used by the online deep learning model.
3. The method according to claim 2, wherein the online deep learning model and the deep learning correction model both comprise an input layer, a hidden layer, a fully connected layer and an output layer;

   Fixing the connection weight parameter and bias parameter of the hidden layer in the online deep learning model, and correcting the connection weight parameter and bias parameter of the fully connected layer in the online deep learning model online;

   online training the connection weight parameters and bias parameters of the hidden layer and the fully connected layer in the deep learning correction model;

   When the error between the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process is greater than a set threshold, a self-correction mechanism is adopted to correct the hidden layer of the model with the deep learning The connection weight parameters and bias parameters of the online deep learning model are replaced by the connection weight parameters and bias parameters of the hidden layer in the online deep learning model, and the connection weight parameters and bias parameters of the fully connected layer of the deep learning correction model are used. The parameters replace the connection weight parameters and bias parameters of the fully connected layer in the online deep learning model.
4. The method according to any one of claims 1-3, wherein the complex industrial process is an fused magnesia smelting process.
5. An intelligent modeling device for a complex industrial process digital twin system, characterized in that the device comprises:

   The mechanism model modeling module is used to establish a mechanism model of a complex industrial process, and the mechanism model includes two parts: an identifiable model and an unmodeled dynamic;

   A parameter identification module for estimating the parameters of the identifiable model to obtain an identification model;

   an unknown nonlinear dynamic acquisition module, used for using the identified model error and the unmodeled dynamic to form an unknown nonlinear dynamic system;

   an intelligent model modeling module of an unknown nonlinear dynamic system, used for establishing an intelligent model of the unknown nonlinear dynamic system;

   A digital twin system intelligent model modeling module, used for establishing an intelligent model of the complex industrial process digital twin system by using the identification model and the intelligent model of the unknown nonlinear dynamic system;

   Wherein, the identification model error is the model output error caused when the parameters in the identifiable model are replaced by their identification values.
6. The device according to claim 5, wherein the unknown nonlinear dynamic system intelligent model modeling module comprises an offline deep learning model modeling module, an online deep learning model modeling module, a deep learning correction model modeling module and self-calibration module;

   The offline deep learning model modeling module adopts LSTM to establish the offline deep learning model;

   The online deep learning model modeling module adopts the same structure as the offline deep learning model to establish the online deep learning model;

   The deep learning correction model modeling module adopts the same structure as the offline deep learning model to establish the deep learning correction model;

   The self-correction module, when the error between the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process is greater than a set threshold, adopts a self-correction mechanism, and uses the deep learning to correct The connection weight parameters and bias parameters of the model correct the connection weight parameters and bias parameters of the online deep learning model;

   Wherein, the historical data used by the deep learning correction model is more than the historical data used by the online deep learning model.
7. The device according to claim 6, wherein the online deep learning model and the deep learning correction model both comprise an input layer, a hidden layer, a fully connected layer and an output layer;

   The online deep learning model modeling module fixes the connection weight parameter and bias parameter of the hidden layer in the online deep learning model, and online corrects the connection weight of the fully connected layer in the online deep learning model parameters and bias parameters;

   The deep learning correction model modeling module trains the connection weight parameters and bias parameters of the hidden layer and the fully connected layer in the deep learning correction model online;

   The self-correction module, when the error between the output of the intelligent model of the complex industrial process digital twin system and the actual output of the complex industrial process is greater than a set threshold, adopts a self-correction mechanism, and uses the deep learning to correct The connection weight parameters and bias parameters of the hidden layer of the model replace the connection weight parameters and bias parameters of the hidden layer in the online deep learning model, and the deep learning is used to correct the fully connected layer of the model. The connection weight parameter and the bias parameter replace the connection weight parameter and the bias parameter of the fully connected layer in the online deep learning model.
8. The device according to any one of claims 5-7, wherein the complex industrial process is an fused magnesia smelting process.
9. A device for implementing the method according to claim 2 or 3, characterized in that, the device comprises: terminal-side sub-device, edge-side sub-device, and cloud-side sub-device;

   The terminal-side sub-device is used to obtain the input data of the online deep learning model and the deep learning correction model;

   The edge side sub-device is used to run the identification model and the online deep learning model;

   The cloud side sub-device is used to run the deep learning correction model;

   The terminal-side sub-device is a process control system of the complex industrial process, the edge-side sub-device is an edge computing device, and the cloud-side sub-device is an industrial cloud.
10. A computer-readable storage medium storing a computer program, characterized in that, when the program is executed by a processor, the method according to any one of claims 1-4 is implemented.

Images