A Univariate Time Series Prediction Method and System Based on Number Transfer Process Parameters

Abstract

This invention provides a method and system for univariate time series prediction of process parameters based on knowledge transfer, belonging to the field of process industry technology. The method includes: collecting basic univariate time series information and preprocessing the basic information to obtain training basic information; pre-training a knowledge-embedded DKN network based on the training basic information to obtain denoising parameters; performing secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain a trained network; and performing time series prediction under unstable operating conditions based on the trained network. This invention solves the problems of generalization performance degradation in traditional data-driven univariate time series prediction models under uncertain input scenarios, and the slow convergence and false saturation of traditional DKN networks due to the time-varying characteristics of process industry operating conditions.

Description

Technical Field

[0001] This invention relates to the field of process industry technology, specifically to a method for predicting univariate time series of process parameters based on knowledge transfer and a system for predicting univariate time series of process parameters based on knowledge transfer. Background Technology

[0002] Time-series predictive modeling of process parameters is of significant practical importance for production and operational optimization decisions in the process industry. It enables accurate and efficient operational estimation of certain key variables in the production process, helping decision-makers achieve predictive process monitoring, such as issuing alerts before process parameters become abnormal. Univariate time-series predictive modeling is one of the most common requirements in the process industry. It typically estimates and infers the potential changes of a univariate time series over time based on its historical values. Univariate time series mainly contain key mathematical features such as periodic, trend, and stochastic terms. Therefore, how to separate and determine the corresponding feature terms from the historical raw time-series data directly determines the prediction results. Using signal decomposition algorithms to decompose the sequence into subsequences with relatively simple frequency components has proven to be effective in separating feature terms and facilitates the construction of predictive models. Wavelet transform, as a time-frequency domain signal processing method, inherits and develops the idea of ​​localization from short-time Fourier transform. At the same time, it overcomes the shortcomings such as the window size not changing with frequency, and provides a "time-frequency" window that adapts to frequency changes. It is an ideal tool for time-frequency analysis and processing of signals. Based on this, wavelet decomposition can be based on the principle of multi-scale analysis to analyze and model time series data decomposed into low-frequency and high-frequency parts at different scales, thereby providing a modeling perspective for univariate time series prediction modeling research under different scales and different dimensional variables.

[0003] In actual production activities, the complex characteristics of industrial production sites cause data to be affected by noise from multiple sources during the generation, acquisition, and communication stages. The uncertainty of the input to time series prediction models caused by high noise has long plagued the modeling process of time series prediction. A common approach is to assume that the high noise in industrial data consists of multiple independent random variables with different probability distributions. Therefore, according to the central limit theorem, the normalized sum of these noises will approach a Gaussian distribution as the noise scale increases. Based on this assumption, prediction models considering uncertain inputs have been proposed to obtain highly robust time series prediction results. Deep Kernel Network (DKN) is a representative model for uncertain input proposed in this context. It overcomes the problems of overfitting, getting trapped in local extrema, and difficulty in determining model structure parameters that occur when using traditional models such as ANN, RBF neural networks, and multilayer knowledge networks for uncertain input time series prediction. At the same time, based on probabilistic analysis methods, it uses denoising algorithms to extend the shallow kernel function into a deep structure on the basis of SVM to improve the ability to process high-noise data in complex nonlinear processes. However, in the process industry, the time-varying nature of operating conditions often leads to slow network convergence, false weight saturation, and low modeling efficiency when dealing with high-noise problems under unstable operating conditions. This is mainly because unstable operating conditions are accompanied by numerous uncertainties, such as changes in raw materials and equipment operating states, which interfere with and negate the effective learning process of DKN for Gaussian noise assumptions. In fact, the analysis and research of unstable operating conditions are of greater practical significance for production optimization in the process industry than those of stable operating conditions. Unstable operating conditions not only directly increase energy and material consumption losses but also pose significant challenges to product quality and production stability. Under stable operating conditions in the process industry, DKN still possesses strong capabilities for time-series prediction problems with uncertain inputs. Therefore, transferring the noise knowledge of DKN under stable operating conditions to model under unstable operating conditions would be a worthwhile attempt. However, the form of transfer learning and the depth of knowledge transfer will directly affect the prediction modeling process and the final prediction results, thus requiring further research.

[0004] In the process industry, another problem hindering the completion of prediction tasks in time series modeling is that data models, including DKN, are based on purely data-driven modeling methods. However, for most engineering problems, it is difficult to obtain sufficient and high-quality data for constructing network models, especially deep network models. This difficulty manifests in two aspects in the practical application of models: First, the extensive production processes in some process industries result in low-quality collected data, increasing the difficulty and cost of data preprocessing. Therefore, many studies suggest that the performance of time series prediction models ultimately depends to some extent on the knowledge content of the data itself rather than the sophistication of the prediction model. Second, the process parameters in some process industries change at low frequencies, while the experimental costs for adjusting process parameters are very high. This means that deep network models that rely excessively on data scale are difficult to apply directly to these industrial scenarios. To address the degradation of generalization performance of traditional data-driven univariate time series prediction models under uncertain input scenarios, and the slow convergence and false saturation of traditional DKN networks due to the time-varying characteristics of process industry conditions, a new univariate time series prediction scheme is needed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for univariate time series prediction of process parameters based on knowledge transfer, so as to at least solve the problems of generalization performance degradation of traditional data-driven univariate time series prediction models under uncertain input scenarios and the slow convergence and false saturation of traditional DKN networks caused by the time-varying characteristics of process industry conditions.

[0006] To achieve the above objectives, the first aspect of the present invention provides a method for univariate time series prediction of process parameters based on knowledge transfer. The method includes: collecting basic information of univariate time series and preprocessing the basic information to obtain training basic information; pre-training a knowledge-embedded DKN network based on the training basic information to obtain denoising parameters; performing secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain a trained network; and performing time series prediction under unstable operating conditions based on the trained network.

[0007] Optionally, the univariate time series basic information includes: the original univariate time series signal and knowledge information; wherein, the knowledge information is one or more of the following: experience data, experimental data, and historical summary data of the process industry.

[0008] Optionally, the preprocessing of the basic information includes: decomposing the original univariate time series signal, including: performing time-frequency deconstruction of the original univariate time series signal based on wavelet decomposition to obtain multiple decomposed signals; the decomposed signals include: multiple high-frequency signals and one low-frequency signal; performing process knowledge discovery and knowledge instantiation on the knowledge information, including: filtering process knowledge of the process industry from the knowledge information; and converting the process knowledge into binary boundary knowledge instances and relational equations.

[0009] Optionally, the step of pre-training the knowledge-embedded DKN network based on the training foundation information to obtain denoising parameters includes: introducing Gaussian distributed noise into the decomposed signal, and embedding the binary boundary knowledge instances and relational equations into the loss function of iterative training to train the knowledge-embedded DKN network; during the DKN network training process, updating the network parameters through error backpropagation to obtain the denoised network parameters and the original input without added noise as denoising parameters.

[0010] Optionally, the step of performing secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain the trained network includes: adding the denoising network parameters to the initial parameters of the secondary training; based on the initial parameters, using the decomposed signal of the original input without added noise as the original input, using the actual value of the predicted variable as the predicted label, adding the binary boundary knowledge instance to the termination condition, and iteratively adjusting the initial parameters using backpropagation and gradient descent algorithms with the knowledge-embedded loss function as the update criterion; and completing the secondary training of the knowledge-embedded DKN network after the iteration termination condition is met to obtain the trained network.

[0011] Optionally, the step of performing time series prediction under unstable operating conditions based on the trained network includes: using stable operating conditions as the source domain and unstable operating conditions as the target domain, performing online prediction based on the trained network includes: performing time-frequency deconstruction on the collected real-time univariate time series data under unstable operating conditions and then performing time series prediction of the corresponding model according to the operating condition.

[0012] A second aspect of the present invention provides a single-variable time series prediction system for process parameters based on knowledge transfer. The system includes: a data acquisition unit for acquiring basic single-variable time series information and preprocessing the basic information to obtain training basic information; a pre-training unit for pre-training a knowledge-embedded DKN network based on the training basic information to obtain denoising parameters; a secondary training unit for performing secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain a trained network; and a prediction unit for performing time series prediction under unstable operating conditions based on the trained network.

[0013] Optionally, the univariate time series basic information includes: the original univariate time series signal and knowledge information; wherein, the knowledge information is one or more of the following: experience data, experimental data, and historical summary data of the process industry.

[0014] Optionally, the acquisition unit is specifically used to decompose the original univariate time series signal, including: decomposing the original univariate time series signal by performing time-frequency decomposition on the original univariate time series signal based on wavelet decomposition to obtain multiple decomposed signals; the decomposed signals include: multiple high-frequency signals and one low-frequency signal; the acquisition unit is also specifically used to perform process knowledge discovery and knowledge instantiation on the knowledge information, including: filtering process knowledge of the process industry from the knowledge information; and converting the process knowledge into binary boundary knowledge instances and relational equations.

[0015] On the other hand, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for univariate time series prediction of process parameters based on knowledge transfer.

[0016] Through the above technical solution, this invention constructs a knowledge-embedded DKN network. During the pre-training and secondary training stages, it accelerates network iteration based on boundary instance conditions and relational equations derived from knowledge mining. This allows the network output to continuously approximate existing process knowledge, breaking through the prediction generalization performance bottleneck of traditional models with uncertain inputs. Furthermore, by separating denoising and prediction tasks and utilizing knowledge transfer, the invention effectively improves the model's prediction performance in uncertain input scenarios, optimizing its modeling efficiency under complex conditions, especially unstable conditions.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of the steps of a single-variable time series prediction method for process parameters based on knowledge transfer provided by one embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of a knowledge-embedded DKN network in the pre-training stage provided by one embodiment of the present invention;

[0021] Figure 3This is a schematic diagram of the time series prediction implementation process based on a knowledge-embedded DKN network provided by one embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of a complete knowledge-embedded DKN training process provided by one embodiment of the present invention;

[0023] Figure 5 This is a system architecture diagram of a process parameter univariate time series prediction system based on knowledge transfer, provided by one embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] Figure 1 This is a flowchart of a univariate time series prediction method for process parameters based on knowledge transfer, provided by one embodiment of the present invention. Figure 1 As shown, this invention provides a univariate time series prediction method for process parameters based on knowledge transfer, the method comprising:

[0026] Step S10: Collect univariate time series basic information and preprocess the basic information to obtain training basic information.

[0027] Specifically, the univariate time series basic information includes: the original univariate time series signal and knowledge information; wherein, the knowledge information is one or more of the following: experience data, experimental data, and historical summary data of the process industry.

[0028] Preferably, the preprocessing of the basic information includes: decomposing the original univariate time series signal, including: performing time-frequency deconstruction of the original univariate time series signal based on wavelet decomposition to obtain multiple decomposed signals; the decomposed signals include: multiple high-frequency signals and one low-frequency signal; performing process knowledge discovery and knowledge instantiation on the knowledge information, including: screening process knowledge of the process industry from the knowledge information; and converting the process knowledge into binary boundary knowledge instances and relational equations.

[0029] In one possible implementation, the original signal f(t) is decomposed layer by layer based on wavelet basis functions, and the final decomposition relationship can be expressed as:

[0030]

[0031] A n D represents the low-frequency component (or approximation signal) derived from the original signal. i, i = 1, 2, ... n represent the high-frequency part of the original signal (also known as the detail signal), and n represents the order of decomposition.

[0032] In the process knowledge discovery phase, existing human experience and experimental results are structurally organized and transformed into unified instance data. For example, for personal experience: under stable working conditions, the carbon dioxide emitted by a gas furnace is maintained in the same proportion as the gas entering the furnace. For an industrial gas furnace with a standard furnace temperature of 850℃ and a surface power of 6kW per square meter, the carbon dioxide emitted is 70-90 cubic feet per second.

[0033] A more general representation is defined as X(U) t ) indicates that under condition U t The predictor variable X is given below, while Represents the instance value under this condition, where condition U t It consists of several condition variables, U t ={u}. Determined by process knowledge. It is a definite constant value, such as "upper limit 70". Therefore, the general representation of process knowledge instantiation is a tuple of <predictor variable, instance value>, that is, In theory, through a specific mapping method, φ(·), any form of process knowledge can be instantiated into a deterministic binary tuple, φ: <X(U t ), Considering the limited scale of process knowledge available in actual process industry production sites, it is usually possible to obtain corresponding binary pairs by manually instantiating process knowledge, where X(U t )and The relationship can be represented by the following relational equation:

[0034] X+N(X;λ)C=0

[0035] Where N(X; λ) is the operation operator for the predictor variable X, and λ is the set of undetermined parameters determined based on specific process knowledge. Common operation operators in the field of process knowledge discovery in the process industry are summarized in Table 1. The above equation describes the general binary relation for process knowledge discovery. However, for relational equations that are difficult to solve directly, the instance data after process knowledge normalization will be regarded as the solution to the relational equation and will participate in the subsequent knowledge embedding process as boundary instance conditions.

[0036]

[0037] Table 1 Common operators for process knowledge discovery in the process industry

[0038] Step S20: Based on the training foundation information, perform knowledge-embedded DKN network pre-training to obtain denoising parameters.

[0039] Specifically, Gaussian noise is introduced into the decomposed signal, and the binary boundary knowledge instances and relational equations are embedded into the loss function of iterative training to train a knowledge-embedded DKN network. During the network training process, the network parameters are updated through error backpropagation to obtain the denoised network parameters and the original input without added noise, which are used as denoising parameters.

[0040] In one possible implementation, firstly, zero-mean Gaussian distributed noise X... U ~N(0, σ 2 The noise is added to the input X to obtain sample data containing Gaussian noise. Where X consists of detail signals and approximation signals obtained through wavelet decomposition, i.e., X = [D1, D2, ..., D...]. n A n ],and The input samples are fed into the kernel function layer for parameter training. The specific network structure is as follows: Figure 3 As shown, its network structure consists of two parts: a denoising kernel function network and process knowledge of the process industry. The process knowledge is embedded in the network's loss function to participate in the model training.

[0041] During the training phase, the input of Gaussian mixture noise is first propagated forward in the first kernel function layer. The activation value obtained after the kernel function value is applied to the activation function σ(·) is:

[0042]

[0043] σ is the activation function. These represent the forward weight matrix and the bias term, respectively. The inverse activation value is obtained by using it as input to the backpropagation activation function.

[0044]

[0045] calculate The error δ between the input X and the input X without noise (1) The error is then backpropagated, and the weight matrix and bias terms of the first-layer kernel function are updated using the gradient descent algorithm.

[0046]

[0047] The calculated forward activation values ​​are then used as input to the second kernel function layer, and so on until all network parameters are obtained. When updating the parameters of each layer, the boundary instance conditions and relational equations in the process knowledge are added as soft constraints to the loss function, as shown in the following equation:

[0048]

[0049] in This represents the root mean square of the model output and X, while Π represents the mean error of all training samples in the i-th layer; the second term on the right-hand side of the above equation represents the soft constraints of all boundary instances of the predicted value, where Π represents the constraint weight; the third term on the right-hand side represents the relational equation constraint, where Λ represents the equation constraint weight, and this term is optional for process industry applications where the relational equation is difficult to determine; the fourth term is the network weight parameter decay term, used to prevent overfitting, where μ is the corresponding weight coefficient. The updated network parameters... The secondary training process, which passes the data to the network, is used to complete the model's prediction task.

[0050] Step S30: Perform secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain the trained network.

[0051] Specifically, the denoising network parameters are added to the initial parameters of the secondary training; based on the initial parameters, the decomposed signal of the original input without added noise is used as the original input, the actual value of the predicted variable is used as the predicted label, the binary boundary knowledge instance is added to the termination condition, and the initial parameters are iteratively adjusted through backpropagation and gradient descent algorithm using the knowledge embedding loss function as the update criterion; when the iteration termination condition is met, the secondary training of the knowledge embedding DKN network is completed, and the trained network is obtained.

[0052] In one possible implementation, after layer-by-layer pre-training, the process knowledge-embedded DKN network model can extract uncontaminated data from Gaussian mixture noise, enabling it to further model noisy data. Furthermore, to enhance its time-series prediction performance, the network undergoes secondary training, i.e., further adjusting network parameters to make the model's expected output approximate the sample labels. Specifically, the model uses the denoised network parameters obtained in the pre-training phase as initial parameters, and iteratively updates the weight parameters of each layer by backpropagating the error between the network output and the expected output to complete the model training. Since process knowledge embedding has been completed in the pre-training phase, making the output result approximate the true physical value, this overcomes the problem of traditional DKN getting stuck in local optima during the random initial parameter training phase, and also avoids large adjustments to network parameters to a certain extent. The secondary training phase of the knowledge-embedded DKN is as follows: Figure 3 As shown, compared to the pre-training stage, during the secondary training process, the input is adjusted to the decomposed sequence X of the original signal, while the error terms calculated by each layer during error backpropagation are... The error term represents the difference between the actual output and the expected predicted output label, with the subscript R indicating secondary training. Taking the first-layer kernel function as an example, its calculated error term is:

[0053]

[0054] Accordingly, the Loss function for the second training phase is:

[0055]

[0056] The first term on the right-hand side of the equation is the actual output. Mean square error of the expected output Y Actual output After passing through each kernel function layer, we obtain:

[0057]

[0058] in:

[0059]

[0060] In the process of updating network parameters θ, taking the update process of the first-layer kernel function parameters as an example, the first-layer parameters are updated separately. Find the partial derivative, that is,

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] in and Let represent the positive activation value and the negative activation value obtained by the kernel function value after passing through the activation function σ(·) in the secondary training phase, respectively:

[0067]

[0068]

[0069] The procedural knowledge is embedded in the last two terms of the four equations, participating in the calculation of the partial derivative of θ and accelerating its convergence. Therefore, the updated weights and bias terms are obtained as follows:

[0070]

[0071]

[0072]

[0073]

[0074] Where η is the learning update coefficient. After completing the above parameter update, the activation value of the first layer will be used as the input of the second layer to continue the parameter update process. This process is repeated multiple times until the loss function is less than a given threshold ∈ (Loss < ∈) or the predicted output approximately satisfies the instance conditions of the process knowledge. The iteration then terminates, yielding the final shape of the network and its parameters θ. Substituting these parameters into the following formula, we obtain the final predicted output value:

[0075]

[0076] Step S40: Based on the trained network, perform time series prediction under unstable operating conditions.

[0077] Specifically, taking stable operating conditions as the source domain and unstable operating conditions as the target domain, online prediction is performed based on the trained network, including: performing time-frequency deconstruction on the collected real-time univariate time series data under unstable operating conditions and then performing time series prediction of the corresponding model according to the operating conditions.

[0078] In one possible implementation, during the time series prediction modeling process with uncertain inputs, considering that the numerous production uncertainties under unstable operating conditions in the process industry will significantly increase the complexity of the denoising task and thus affect the robustness of the final prediction results, a transfer fusion modeling strategy will be adopted for both stable and unstable production conditions in the process industry. The stable operating condition will be considered as the source domain, D... s In D s After wavelet decomposition, the univariate time-series data of the internal process parameters are used for pre-training and secondary training of a knowledge-embedded distributed knowledge network (DKN) using the decomposed signals. Simultaneously, the pre-trained network parameters θ of the DKN are permanently stored. Unsteady operating conditions are considered as the target domain. T In the target domain D T In this context, the network parameters θ, pre-trained in the source domain for uncertain input noise, are transferred to the target domain; this is a general approach. The prediction task performed within a given learning domain is defined as... in Let D represent the label space, and f(x) represent the prediction function. The source domain D... s They typically exist in the form of instance-label pairs in prediction tasks, i.e. Similarly, target domain Based on the source domain and the corresponding prediction task The learned denoising knowledge is transferred to the target domain and the corresponding prediction task. To improve the prediction function f TThe performance of the knowledge-embedded DKN is improved. After knowledge transfer, the DKN is trained again for time series prediction under unstable operating conditions using θ as the initial parameter. After fine-tuning θ, the time series prediction modeling task under unstable operating conditions is realized. The complete training process of the knowledge-embedded DKN is as follows: Figure 4 As shown.

[0079] Example:

[0080] The effectiveness of the proposed method is validated using a public dataset of CO2 emissions from industrial gasifiers and a dataset of blast furnace gas flow rates from an industrial ironmaking process. The CO2 emission dataset records the input gas rate and emitted CO2 data from an industrial gasifier in a cement production line, with a sampling interval of 9 seconds, containing a total of 296 sets of sample data under normal production conditions. Using CO2 emission data as the primary research object and prediction process parameter, a univariate time-series prediction model is established to complete the denoising and time-series prediction tasks under stable operating conditions.

[0081] The Daubenchies function was used as the wavelet function, with a decomposition order of 3, yielding three high-frequency detail signals (D1, D2, D3) and one low-frequency approximation signal (A3). Additionally, CO2 emissions from the gasifier, Y... t With gas input X t The state transition equation satisfies a multi-order relation, the general relation of which is as follows:

[0082] Y t =δ1Y t-1 +…+δ r Y t-r +ω0X t-b -ω1X t-b-1 -…-ω s X t-b-s +C

[0083] Where C represents the initial value, δ i ω j , i = 1, 2, ..., r; j = 0, 1, ..., s represent weighting coefficients. To determine the value of these coefficients, experimental data on gasifier emissions and inputs were collected, and the instantiated experimental results are shown in Table 2.

[0084]

[0085] Table 2 instantiation of gasifier input and emission experimental results. The multi-order relation obtained by the least squares method is shown in the following formula:

[0086]

[0087] Substituting the above equation into the constraint equation, we obtain:

[0088]

[0089] The boundary instance constraints of the Loss function are:

[0090]

[0091] The aforementioned relationships and boundary instances will be added as constraints to the training process of the knowledge-embedded DKN. As comparative models, wavelet denoising autoregressive (AR) method, the traditional DKN method, and the multilayer backpropagation network (BPNN) method will be used as baseline models for comparative experiments. The input to the neural network method is the decomposed time-series data after wavelet decomposition. The training results and test errors of different models are shown in Table 3.

[0092]

[0093]

[0094] Table 3 Test error and training time for different models

[0095] During the model training phase, the proposed knowledge-embedded DKN achieved the best training performance among all prediction models, with its predictions most closely matching the true values. In contrast, the denoising AR model exhibited poor fitting performance during training, but boasted the highest training efficiency, with a training time of 4.53 seconds, the shortest among all methods. While the traditional DKN method also achieved good prediction fitting results, its training efficiency was the lowest, suggesting that the calculation of kernel functions for each neuron and the wavelet decomposition process consumed a significant portion of its training time. Notably, the introduction of procedural knowledge significantly improved the modeling efficiency of knowledge-embedded DKN. Compared to traditional DKN and NPNN, it not only had a shorter training time but also significantly improved the training fitting effect. This advantage was even more pronounced during the model testing phase. As shown in Table 3, the proposed model's testing accuracy was significantly enhanced in both the error metrics MSE and MAPE, and the goodness-of-fit R-squared was also improved. 2 Since it is closest to 1, knowledge-embedded DKN has improved both modeling efficiency and prediction accuracy to varying degrees after considering input noise and introducing process knowledge.

[0096] In the blast furnace gas (BFG) dataset, the samples are divided into stable and unstable operating conditions, with 80% of the training samples falling under different conditions. This method establishes corresponding time-series prediction models for different operating condition types, and transfers the denoising network parameters pre-trained in the stable operating conditions to the unstable operating conditions for secondary training of the knowledge-embedded DKN.

[0097] The Daubenches function was used as the wavelet basis function to perform time-series decomposition on the original BFG data. The decomposed time-series signals were then divided according to their respective operating conditions. In the process knowledge discovery stage, boundary instance condition analysis was performed on both stable and unstable operating conditions. Experts selected several instance data based on historical operating data. In addition, knowledge instances under different operating conditions were summarized by combining on-site experimental test results and operational experience, as shown in Table 4.

[0098]

[0099] Table 4. Results of BFG runtime knowledge instantiation

[0100] Where SC represents the stable operating condition and DSC represents the unstable operating condition. The instantiation results of the running knowledge will be introduced into the loss function of the knowledge-embedded DKN under both the stable and unstable operating conditions, and will participate in the model's pre-training and secondary training processes. The functional expressions for the two different training stages under the stable operating condition are as follows:

[0101]

[0102] Substituting the SC instance results from Table 4 into the above formula, we obtain the last term of the Loss function as:

[0103]

[0104] U i This represents the condition of the i-th instance. Representing condition U respectively i The lower and upper bounds for the following example. Under unstable conditions, there is only a second training process, and the loss function for the training phase is:

[0105]

[0106] Substituting the DSC instance results from Table 4 into the above formula, we get:

[0107]

[0108] Based on this, the instantiation results of process knowledge will be fully involved in the modeling process, enabling the iteration and updating of parameters.

[0109] Test results of different methods under SC and DSC show that all four methods achieved relatively ideal prediction results overall. However, the knowledge-embedded DKN method has the best fitting effect on the true value, followed by the traditional DKN model. The prediction results of the denoising AR and BPNN methods show large fluctuations, while the prediction results of the kernel function network method, especially the knowledge-embedded DKN, are more stable.

[0110] In DSC, the random factors introduced by unstable operating conditions, coupled with uncertain input noise, cause significant data fluctuations, making BFG flow prediction more difficult under unstable operating conditions. Overall, compared to stable operating conditions, the flow prediction results in DSC have larger errors, especially for denoised AR methods, whose prediction results deviate significantly from the true values. This is mainly because these methods rely solely on denoising algorithms to reduce the impact of Gaussian random noise when dealing with high-uncertainty inputs, but their nonlinear processing capabilities are weak. In contrast, knowledge-embedded DKN, which integrates kernel tricks and knowledge transfer, still performs excellently in prediction tests. Both traditional DKN and knowledge-embedded DKN have good prediction performance, and their ability to handle complex DSC is stronger than other methods.

[0111] Figure 5 This is a system architecture diagram of a process parameter univariate time series prediction system based on knowledge transfer, provided by one embodiment of the present invention. (See diagram below.) Figure 5 As shown, this invention provides a univariate time series prediction system for process parameters based on knowledge transfer. The system includes: a data acquisition unit for acquiring basic univariate time series information and preprocessing the basic information to obtain training basic information; a pre-training unit for pre-training a knowledge-embedded DKN network based on the training basic information to obtain denoising parameters; a secondary training unit for performing secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain a trained network; and a prediction unit for performing time series prediction under unstable operating conditions based on the trained network.

[0112] Preferably, the univariate time series basic information includes: original univariate time series signals and knowledge information; wherein, the knowledge information is one or more of the following: experience data, experimental data, and historical summary data of the process industry.

[0113] Preferably, the acquisition unit is specifically used to decompose the original univariate time series signal, including: decomposing the original univariate time series signal by performing time-frequency decomposition on the original univariate time series signal based on wavelet decomposition to obtain multiple decomposed signals; the decomposed signals include: multiple high-frequency signals and one low-frequency signal; the acquisition unit is also specifically used to perform process knowledge discovery and knowledge instantiation on the knowledge information, including: filtering process knowledge of the process industry from the knowledge information; and converting the process knowledge into binary boundary knowledge instances and relational equations.

[0114] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for univariate time series prediction of process parameters based on knowledge transfer.

[0115] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0116] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0117] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A univariate time series forecasting method for process parameters based on knowledge transfer, characterized in that, The method includes: Univariate time series basic information is collected, and the basic information is preprocessed to obtain training basic information; wherein, The univariate time series basic information includes: the original univariate time series signal and knowledge information; wherein, the knowledge information is one or more of the following: empirical data, experimental data, and historical summary data from the process industry; The preprocessing of the basic information includes: decomposing the original univariate time-series signal, including: The original univariate time-series signal is deconstructed using wavelet decomposition to obtain multiple decomposed signals. The decomposed signals include multiple high-frequency signals and one low-frequency signal. Process knowledge discovery and knowledge instantiation are performed on the knowledge information, including: filtering process knowledge of the process industry from the knowledge information; and converting the process knowledge into binary boundary knowledge instances and relational equations. Based on the aforementioned training information, a knowledge-embedded DKN network is pre-trained to obtain denoising parameters, including: Gaussian noise is introduced into the decomposed signal, and the binary boundary knowledge instances and relation equations are embedded into the loss function of iterative training to train the knowledge-embedded DKN network. During the training of the DKN network, the network parameters are updated through backpropagation of the error to obtain the denoised network parameters and the original input without added noise, which are used as the denoising parameters. Based on the denoising parameters and the basic information, a second training of the knowledge-embedded DKN network is performed to obtain the trained network. Based on the trained network, time series prediction under unstable operating conditions is performed.

2. The method according to claim 1, characterized in that, The process of performing secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain the trained network includes: Add the denoising network parameters to the initial parameters of the second training; Based on the initial parameters, the decomposed signal of the original input without noise is used as the original input, the actual value of the predicted variable is used as the predicted label, the binary boundary knowledge instance is added to the termination condition, and the initial parameters are iteratively adjusted through backpropagation and gradient descent algorithm with the knowledge embedding loss function as the update criterion. Once the iteration termination condition is met, the second training of the knowledge-embedded DKN network is completed, and the trained network is obtained.

3. The method according to claim 1, characterized in that, The step of performing time series prediction under unstable operating conditions based on the trained network includes: Using stable operating conditions as the source domain and unstable operating conditions as the target domain, online prediction is performed based on the trained network, including: After time-frequency deconstructing the real-time univariate time series data collected under unstable operating conditions, time series prediction of the corresponding model is performed according to the operating condition.

4. A univariate time series prediction system for process parameters based on knowledge transfer, characterized in that, The system includes: The acquisition unit is used to acquire univariate time-series basic information and preprocess the basic information to obtain training basic information; wherein, The univariate time series basic information includes: the original univariate time series signal and knowledge information; wherein, the knowledge information is one or more of the following: empirical data, experimental data, and historical summary data from the process industry; The preprocessing of the basic information includes: decomposing the original univariate time-series signal, including: The original univariate time-series signal is deconstructed using wavelet decomposition to obtain multiple decomposed signals. The decomposed signals include multiple high-frequency signals and one low-frequency signal. Process knowledge discovery and knowledge instantiation are performed on the knowledge information, including: filtering process knowledge of the process industry from the knowledge information; and converting the process knowledge into binary boundary knowledge instances and relational equations. The pre-training unit is used to pre-train the knowledge-embedded DKN network based on the training foundation information to obtain denoising parameters, including: Gaussian noise is introduced into the decomposed signal, and the binary boundary knowledge instances and relation equations are embedded into the loss function of iterative training to train the knowledge-embedded DKN network. During the training of the DKN network, the network parameters are updated through backpropagation of the error to obtain the denoised network parameters and the original input without added noise, which are used as the denoising parameters. The secondary training unit is used to perform secondary training of the knowledge-embedded DKN network based on the denoising parameters and the basic information to obtain the trained network. The prediction unit is used to perform time series prediction under unstable operating conditions based on the trained network.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the process parameter univariate time series prediction method based on knowledge transfer as described in any one of claims 1-3.

Images