A variable working condition industrial temperature prediction method based on a KWNet model

Abstract

The application discloses a variable working condition industrial temperature prediction method based on a KWNet model and belongs to the technical field of industrial process parameter prediction.The method adds a novel double-flow wavelet transform module, fuses the approximate and detail components in the time sequence after wavelet decomposition through a double-flow decomposition and reconstruction mode, and can reconstruct all components without difference compared with a traditional wavelet transform module, which enables independent processing of the approximate and detail components in different branches, prevents cross-stage interference, and ensures that the unique features of each frequency band are retained.Then, through a learnable dynamic gate mechanism, the feature components of the seasonal and trend sub-sequences in the reconstructed time domain are effectively extracted, and the learnable dynamic gate mechanism is innovatively used in the pyramid multi-scale dilated convolution and Fourier modulation attention, so that the sequence information in the time domain and the frequency domain is effectively balanced.

Description

Technical Field

[0001] This invention relates to a method for predicting industrial temperature under varying operating conditions based on the KWNet model, belonging to the field of industrial process parameter prediction technology. Background Technology

[0002] Accurate prediction of industrial process parameters is crucial for the monitoring, control, and optimization of industrial processes, especially for predicting parameter changes over multiple time steps (i.e., multi-step prediction). However, in practical applications, due to the complex coupling relationships between multiple variables under different operating conditions, prediction models struggle to accurately capture the resulting dynamic transition characteristics and transient fluctuations, thus affecting prediction accuracy. For example, multi-step prediction of the temperature inside the extraction tank during industrial extraction suffers from this problem. Industrial extraction is an important separation technology widely used in pharmaceutical manufacturing, food processing, and the chemical industry to separate active ingredients from natural or synthetic sources. In this process, the temperature inside the extraction tank is a key parameter controlling extraction kinetics and reliability; however, due to the high viscosity of the solution and the complex sealed environment, its temperature is difficult to measure. Therefore, current industrial temperature prediction often employs soft sensing techniques, and deep learning methods are crucial for soft sensing techniques. Therefore, finding a suitable deep learning model is of great significance.

[0003] Traditional deep learning methods for temperature parameter prediction include Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), and Transformers. CNNs can use one-dimensional convolutional kernels to slide across a time series to obtain detailed information about each part of the time series, but they cannot capture temporal relationships and lack temporal information. LSTMs and GRUs can store temporal information and reflect the order of data, but due to their sequential nature, they have significant complexity when capturing long-distance time series, which limits their computation and scalability. Transformers, as a parallel processing framework, achieve more efficient long-distance dependency modeling and scalability for large datasets, but Transformers still face challenges in capturing complex temporal patterns. Their self-attention mechanism may suffer information loss when dealing with intricate temporal dependencies, and their pointwise structure often cannot maintain consistent distribution characteristics over long sequences; all of these factors can lead to inaccurate prediction results.

[0004] Recent research has yielded many novel predictive models. For example, Autoformer integrates a sequence decomposition module to separate time series into seasonal and trend-periodic components. However, when constructing these components, it uses a fixed sliding window to average the trend term, failing to distinguish multi-frequency information. Simple subtraction results in high-frequency, mid-frequency, and noise components being mixed together and compressed into the seasonal term. It cannot determine which frequency band or period is more important for the future based on time length, thus degrading model performance. Furthermore, the size of the sliding average window is a hyperparameter that is difficult to determine. DLinear demonstrates that simple linear decomposition combined with explicit trend-residue separation can mitigate distribution bias through normalization and sliding window averaging, thus outperforming complex architectures. However, its fixed sliding window size and static decomposition method cannot adapt to different types of data. These methods validate the feasibility of decomposition approaches in industrial forecasting problems. Physical Information Neural Networks (PINNs) encode physical information as soft constraints during training, demonstrating particular effectiveness under conditions of sparse data or measurement uncertainty. However, existing PINNs rely on precise characterization of physical principles or governing equations, which inherently limits their applicability to strongly nonlinear, coupled, and multivariable real-world industrial processes. Therefore, there is still room for improvement in the prediction accuracy of existing prediction models. Summary of the Invention

[0005] To further improve the prediction accuracy of industrial process parameters under varying operating conditions, this invention provides a method for predicting industrial temperature under varying operating conditions based on the KWNet model, including:

[0006] Step 1: Collect historical parameters from specific industrial processes to construct an industrial time series dataset;

[0007] Step 2: Develop expert knowledge for specific industrial processes;

[0008] Step 3: Construct a prediction model based on KWNet by combining the established expert knowledge;

[0009] Step 4: Train the KWNet-based prediction model using the industrial time series dataset;

[0010] Step 5: Use the trained KWNet-based prediction model to complete the multi-step prediction of industrial temperature parameters under varying operating conditions.

[0011] Optionally, the KWNet-based prediction model includes a two-stream wavelet transform module, a learnable dynamic gating mechanism module, a domain knowledge analysis and formulation module, an adaptive time-knowledge fusion loss function module, and a time-frequency fusion decoder; wherein, the domain knowledge analysis and formulation module is used to formulate the established expert knowledge and construct the expert knowledge loss function. ;

[0012] The dual-stream wavelet transform module is used to reconstruct seasonal and trend components from the original input data after wavelet transform. The learnable dynamic gate mechanism module is used to determine the weights of the seasonal and trend components to obtain the overall reconstructed data. This overall reconstructed data is then input into a time-frequency fusion decoder to obtain preliminary prediction results. Based on these preliminary prediction results, a time-domain loss function is derived. The adaptive time-knowledge fusion loss function module is used to integrate the expert knowledge loss function. and the time-domain loss function The total loss function of the KWNet-based prediction model is obtained. .

[0013] Optionally, the dual-stream wavelet transform module includes a wavelet decomposition unit, a frequency processing unit, and a time-frequency joint feature extractor;

[0014] First, the original input data is decomposed into m-level discrete wavelet decomposition using wavelet decomposition units to generate a set of approximate coefficients. and detail coefficient , j =2,3,…,m;

[0015] The detail coefficients generated by the wavelet decomposition unit The input frequency processing unit obtains , and approximation coefficient The seasonal components were calculated together. ;

[0016] The approximation coefficients generated by the wavelet decomposition unit The input time-frequency joint feature extractor processes the data to obtain the desired feature. , and detail coefficient The trend components are calculated together. .

[0017] Optionally, the learnable dynamic gate mechanism module will convert the seasonal components obtained by the dual-stream wavelet transform module. and trend components The reconstructed signal is obtained by splicing the last d-dimensional phases. The reconstructed signal is obtained by splicing together linear layers with an added Sigmoid function. Projection, generating gated tensors Then, the output data is obtained by weighted summation through gates. .

[0018] Optionally, the time-frequency fusion decoder is designed with a fixed core size in the time branch. and three different expansion rates The pyramid multi-scale dilated convolution is used to capture local temporal abrupt changes and long-range contextual dependencies in a hierarchical manner; at the same time, multi-head Fourier modulated attention is used in the frequency branch to extract discriminative frequency features; the pyramid multi-scale dilated convolution and multi-head Fourier modulated attention are adaptively fused through a learnable dynamic gating mechanism to finally obtain a unified time-frequency representation.

[0019] Optionally, the total loss function of the KWNet-based prediction model obtained by the adaptive time-knowledge fusion loss function module is... for:

[0020]

[0021] in and These are adaptive weight coefficients, where 'e' represents the number of training rounds. .

[0022] Optionally, the specific industrial process in step 1 is an extraction process, and the historical parameters of the extraction process include the temperature inside the extraction tank, the pressure inside the extraction tank, the flow rate of the extraction solvent, the pressure of the secondary heat preservation steam, the secondary heating temperature, the temperature of the storage tank, and the valve opening status.

[0023] Optionally, the expert knowledge defined for the extraction process in step 2 includes:

[0024] Expert Knowledge Rule 1: If the second steam valve changes from closed to open, the predicted temperature of the extraction tank is... With minimum increment Monotonically increasing; otherwise, if the prediction violates expert knowledge rule one, it will be penalized with additional knowledge loss.

[0025] Expert Knowledge Rule 2: If the work procedure changes from the end of the first drainage to the second water addition, then the predicted temperature of the extraction tank... Monotonically decreasing, and the minimum decrease is Otherwise, if the prediction violates Rule 2 of Expert Knowledge, it will be penalized with additional knowledge loss.

[0026] Optionally, the expert knowledge loss function obtained based on expert knowledge rule one and expert knowledge rule two. for:

[0027]

[0028] in,

[0029]

[0030]

[0031] in This represents the predicted tank temperature. and The two terms represent the tank temperature difference predicted by two consecutive predictions based on expert knowledge rule one and expert knowledge rule two, respectively.

[0032] and This is an indicator function for the triggering conditions of Expert Knowledge Rule 1 and Expert Knowledge Rule 2. The total number of violations of Expert Knowledge Rule 1 and Expert Knowledge Rule 2 within the batch; To calculate the number of batches, The predicted time length of the sequence.

[0033] Optionally, the time-domain loss function for:

[0034]

[0035] in , N and N represent the actual tank temperature, the predicted tank temperature, and the sample size, respectively.

[0036] The beneficial effects of this invention are:

[0037] To predict multi-step industrial temperatures, this invention proposes a Knowledge-Informed Wavelet Decomposition Neural Network (KWNet) method for industrial temperature prediction under Variable Working Conditions. By combining and fusing expert knowledge and time loss, a novel adaptive time-knowledge fusion module is designed to automatically and rationally allocate time-domain and knowledge losses, dynamically balancing time accuracy with consistency with industrial process expertise. This invention also proposes a novel dual-stream wavelet transform module that simultaneously captures high-frequency and low-frequency details from industrial time-series signals. These details are then reconstructed separately, acting as a fine filter to enhance the corresponding configuration components while suppressing noise. This refinement capability distinguishes it from traditional wavelet transform modules that lack this discriminative ability, enabling the indiscriminate reconstruction of all components. Traditional wavelet decomposition has different time-domain and frequency-domain resolutions at different scales, leading to frequency domain aliasing and phase shifts during reconstruction. The dual-stream wavelet transform proposed in this application overcomes this drawback by processing different branches independently to prevent cross-level interference and ensure the preservation of unique features in each frequency band, resulting in more accurate predictions. By designing a learnable dynamic gate that fuses trend and seasonal representations, an effective balance is maintained between the temporal features of the pyramid-dilated convolution module and the frequency features of the multi-head Fourier attention block, thereby adaptively fusing time-domain and frequency-domain information and effectively extracting key features. This invention also proposes a novel time-frequency fusion decoder that utilizes detailed time recovery from the pyramid multi-scale dilated convolution and spectral correction provided by the frequency-domain attention block to characterize signal components, significantly enhancing the reconstruction of non-stationary industrial signals. This model exhibits higher prediction accuracy and stronger generalization ability than existing time series prediction models, significantly improving the accurate monitoring and soft measurement capabilities of key parameters in the production process. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of a variable-condition industrial temperature prediction method based on the KWNet model provided in one embodiment of the present invention.

[0040] Figure 2This is a KWNet model structure diagram of a variable operating condition industrial temperature prediction method based on the KWNet model provided in an embodiment of the present invention.

[0041] Figure 3 It is a framework diagram of a real industrial extraction process that includes multiple extraction tanks.

[0042] Figure 4 This is a schematic diagram of the expert knowledge rules for a variable-condition industrial temperature prediction method based on the KWNet model provided in an embodiment of the present invention.

[0043] Figure 5 This is a schematic diagram of the expert knowledge rule 2 of the variable operating condition industrial temperature prediction method based on the KWNet model provided in one embodiment of the present invention.

[0044] Figure 6 This is a schematic diagram illustrating the prediction results of the Transformer model over ten minutes in one embodiment of the present invention on the TA3001A dataset.

[0045] Figure 7 This is a schematic diagram of the prediction results of the comparative model Dliner on the TA3001A dataset of the extraction tank in one embodiment of the present invention over ten minutes.

[0046] Figure 8 This is a schematic diagram illustrating the prediction results of the TimeMixer model over ten minutes in one embodiment of the present invention on the TA3001A dataset.

[0047] Figure 9 This is a schematic diagram illustrating the prediction results of the PatchTST model over ten minutes in one embodiment of the present invention on the TA3001A dataset.

[0048] Figure 10 This is a schematic diagram illustrating the prediction results of the comparison model FreTSr over ten minutes in one embodiment of the present invention on the TA3001A dataset.

[0049] Figure 11 This is a schematic diagram of the prediction results for ten minutes using the variable operating condition industrial temperature prediction method based on the KWNet model provided in this invention on the TA3001A dataset of extraction tanks.

[0050] Figure 12 This is a violin plot showing the absolute error distribution of the KWNet model provided by this invention and five other comparative models on the TA3001A dataset in one embodiment of the present invention.

[0051] Figure 13This is a violin plot showing the absolute error distribution of the KWNet model provided by this invention and five other comparative models on the TA3001B dataset in one embodiment of the present invention.

[0052] Figure 14 This is a violin plot showing the absolute error distribution of the KWNet model provided by this invention and five other comparative models on the TA3001C dataset in one embodiment of the present invention.

[0053] Figure 15 This is a violin plot showing the absolute error distribution of the KWNet model provided by this invention and five other comparative models on the TA3001D dataset in one embodiment of the present invention.

[0054] Figure 16 This is a violin plot showing the absolute error distribution of the KWNet model provided by this invention and five other comparative models on the TA3001E dataset in one embodiment of the present invention.

[0055] Figure 17 This is a violin plot showing the absolute error distribution of the KWNet model provided by this invention and five other comparative models on the TA3001F dataset in one embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0057] Example 1

[0058] This embodiment provides a method for predicting industrial temperature under varying operating conditions based on the KWNet model. The flowchart of this method is as follows: Figure 1 As shown, firstly, historical parameters from specific industrial processes are collected to construct an industrial time-series dataset. This dataset is then preprocessed and divided into training, validation, and test sets according to a certain ratio. Next, expert knowledge is formulated and formalized for the industrial process. Subsequently, a KWNet-based prediction model is constructed using the obtained expert knowledge. The KWNet model is then trained using the industrial time-series dataset. Finally, the industrial temperature is predicted based on the previously trained model.

[0059] Specifically, the method includes:

[0060] Step 1: Collect historical parameters from specific industrial processes to construct an industrial time series dataset, and preprocess it to divide it into training set, validation set, and test set according to a certain ratio.

[0061] The specific industrial process mainly refers to the extraction process, which is widely used in the pharmaceutical, food and cosmetic industries. The data to be collected varies for different industrial processes, but all include parameters such as pressure and flow rate, including temperature, in the process flow. Specifically, this includes the temperature inside the extraction tank, the pressure inside the extraction tank, the flow rate of the extraction solvent, the pressure of the secondary insulation steam, the secondary heating temperature, the storage tank temperature, and the valve opening status.

[0062] The preprocessing step after data collection is the classic Z-Score normalization operation, which requires calculating the mean and standard deviation of the sequence. The calculation formula is as follows:

[0063]

[0064] in This represents the data after normalization. This represents the data before processing. Represents the mean of the sequence. This represents the standard deviation. After processing, all data will be standardized to the same range, facilitating subsequent calculations.

[0065] Step 2: Develop expert knowledge rules for the industrial process and construct the expert knowledge loss function. .

[0066] Developing expert knowledge requires reflecting the operating principles of the extraction process. Its core is to systematically analyze specific patterns observed during the production process so that the model can cope with different working conditions and clearly reflect the underlying physical principles.

[0067] This embodiment uses Figure 3 The extraction process shown is used as an example to illustrate the specific process of developing expert knowledge and formulating it, such as... Figure 3 As shown, the extraction system comprises six extraction tanks, TA3001A to TA3001F. Each extraction tank is equipped with a heat exchanger, filter, intelligent temperature detector, and program status monitor, which can effectively monitor important parameters during the extraction process, such as the temperature inside the extraction tank, the pressure inside the extraction tank, the flow rate of the extraction solvent, the pressure of the secondary insulation steam, the secondary heating temperature, the storage tank temperature, and the valve opening status. Figure 3 As can be seen, the first steam valve is used to control the steam flow rate entering the heat exchanger, while the second steam valve is used to control the steam flow rate entering the extraction tank.

[0068] for Figure 3 The extraction process shown can be used Figure 4 and Figure 5 This indicates their expert knowledge.

[0069] Figure 4The average dynamic response of the ETT (extraction tank temperature) variable across similar time series segments under different SIV2 (second steam valve) variables shows a clear "decline-rise" phenomenon.

[0070] Figure 5 The average dynamic response of ETT (extraction tank temperature) across similar time series segments under different WS (working steps) shows a clear "decline-rise" phenomenon.

[0071] based on Figure 4 and Figure 5 The expert knowledge displayed, for a specific extraction process, is defined as follows:

[0072] Expert Knowledge Rule 1: If the second steam valve— The temperature of the extraction tank is predicted as it changes from 0.00 (closed) to 1.00 (open). With minimum increment Monotonically increasing. Otherwise, if the prediction result violates expert knowledge rule one, it will be penalized with additional knowledge loss.

[0073] Expert Knowledge Rule Two: If the work steps— From the end of the first drainage to the second water addition, what is the predicted temperature of the extraction tank? Monotonically decreasing, and the minimum decrease is Otherwise, predictions that violate Rule 2 of Expert Knowledge will be penalized with additional knowledge loss.

[0074] in and Let "second steam valve" and "working procedure" represent the control variables at time t, respectively. Represents the prediction time domain The predicted temperature of the storage tank, k, ranges from [value missing]. .

[0075] The expert knowledge loss function is obtained using the aforementioned expert knowledge rules. The expression is:

[0076]

[0077] in, and These represent the temperature difference under expert knowledge rule one and expert knowledge rule two, respectively. and This is an indicator function for the triggering conditions of Expert Knowledge Rule 1 and Expert Knowledge Rule 2. The total number of violations of Expert Knowledge Rule 1 and Expert Knowledge Rule 2 within the batch; To calculate the number of batches, The predicted time length of the sequence.

[0078] Indicator Function and The expressions are as follows:

[0079]

[0080]

[0081] in i t represents the batch number, and t represents the time series number.

[0082] Step 3: Combine the expert knowledge obtained in Step 2 to construct a prediction model based on KWNet.

[0083] The KWNet neural network model constructed in this application considers expert knowledge rules and employs wavelet decomposition; therefore, it is named the KWNet model. Figure 2 As shown, it includes a novel dual-stream wavelet transform module, a learnable dynamic gate mechanism module, a domain knowledge analysis and formulation module, an adaptive time-knowledge fusion loss function module, and a time-frequency fusion decoder, which can accurately predict temperature sequences in extraction processes under varying operating conditions.

[0084] like Figure 2 As shown, the dual-stream wavelet transform module contains two wavelet decompositions, involving two signal streams. Specifically, it includes a wavelet decomposition unit, a frequency processing unit, and a time-frequency joint feature extractor. The low-frequency and high-frequency components obtained by the wavelet decomposition unit are processed by the frequency processing unit and the time-frequency joint feature extractor, respectively, to perform corresponding frequency processing and time-frequency joint feature extraction, ultimately yielding the seasonal steering reconstruction signal and the trend steering reconstruction signal.

[0085] The frequency processing unit consists of m frequency feature extractors, each of which comprises a multilayer perceptron (FMLP), a ReLU function, and an addition and normalization module. The time-frequency joint feature extractor includes a frequency feature extractor, a bidirectional long short-term memory network (BiLSTM), a multilayer perceptron (FMLP), a ReLU function, multi-head attention, learnable dynamic gates, and addition and normalization modules.

[0086] Specifically, in the dual-stream wavelet transform module, the discrete wavelet transform (DWT) is used to decompose the signal (i.e., the sequence of acquired industrial production process parameters) into a set of high-frequency and low-frequency basis functions. These basis functions are local in both time and frequency. This multi-resolution analysis capability makes DWT particularly suitable for capturing transient features and hierarchical structures in time series data. DWT can be used to analyze discrete time series signals of length N. Perform multi-level decomposition. At each resolution j ( j Under the condition =2,3,…,m), the transformation generates a set of approximate coefficients. and detail coefficient , where m represents the maximum number of decomposition levels. Decomposition is achieved by cascading low-pass and high-pass filters, followed by downsampling. Approximation coefficients represent the low-frequency components of the signal, while detail coefficients capture the high-frequency components. The calculation formula is:

[0087]

[0088]

[0089] in and These are the low-pass and high-pass filter coefficients derived from the mother wavelet function, respectively, where n is the time index of the input signal. Initial approximation coefficients. Original signal The downsampling operation of 2x is implicit in the filtering process.

[0090] The reconstruction of the original signal from the wavelet coefficients is achieved through the inverse discrete wavelet transform (IDWT), which involves upsampling and filtering operations, expressed as follows:

[0091] in This indicates a reconstructed low-pass filter. This represents a reconstructed high-pass filter. Typically, the initial approximation coefficients are... As a reconstruction quantity for the initial input sequence X, when the wavelet filter satisfies certain orthogonality or bioorthogonality conditions, this reconstruction formula guarantees the perfect reconstruction of the original signal.

[0092] In the time-frequency fusion module, the self-attention mechanism dynamically assigns weights based on the perceived importance of each input element. In the multi-head attention module, the input is defined as a query. ,key ,value ,in w Indicates the number of samples. This indicates the dimensions of the query and the key. The dimension representing the value. The self-attention mechanism is described as follows:

[0093]

[0094]

[0095] in and These correspond to self-attention calculation and the Softmax function, respectively. The Softmax function normalizes the initial alignment score to a probability distribution such that the sum of all weights equals 1.

[0096] like Figure 2 As shown, after preprocessing, the data first enters the dual-stream wavelet transform module, which uses wavelet decomposition units to decompose the input time series into detail coefficients representing high-frequency components. And the approximation coefficients representing low-frequency components The high and low frequency components are processed independently, with the high frequency components processed by the frequency processing unit to obtain... With low-frequency components The seasonal components were calculated together. Low-frequency components are obtained by a time-frequency joint feature extractor. With high-frequency components The trend components are calculated together. The calculation formulas for the two are as follows:

[0097]

[0098]

[0099] in For frequency feature extractor in The output on For time-frequency joint feature extractor in The output on. Represents the initial input sequence Seasonal restructuring volume Represents the initial input sequence Restructuring volume oriented towards trends; and These are the low-pass and high-pass filter coefficients derived from the mother wavelet function, respectively.

[0100] In a learnable dynamic gating mechanism, the reconstructed seasonal components obtained through a dual-stream wavelet transform module and trend components The reconstructed signal is obtained by splicing the last d-dimensional phases. , Represents a real number tensor.

[0101] in b Represents the number of batches.l Represents the length of the time series. d This represents the dimension of the time series.

[0102] After being concatenated through a linear layer pair with the Sigmoid function added Projection, generating gated tensors Then, weighted summation is performed using gates. The process can be represented as follows:

[0103]

[0104]

[0105]

[0106] in and Represents the learnable parameters of a linear layer. This represents the Sigmoid function.

[0107] Then output the data. The input is a time-frequency fusion decoder, which has a fixed core size in the time branch. and three different expansion rates The pyramid multi-scale dilated convolution (PDC) module hierarchically captures local temporal abrupt changes and long-range contextual dependencies; simultaneously, in the frequency branch, multi-head Fourier modulated attention (MFA) is used to extract discriminative frequency features. PDC and MFA are adaptively fused through a learnable dynamic gating mechanism, ultimately yielding a unified time-frequency representation. This step can then be used to obtain temperature predictions. The time-domain loss is obtained after processing with the loss function. :

[0108]

[0109] in , N and N represent the true value, predicted value, and sample size, respectively.

[0110] Next, the total loss will be calculated in the adaptive time-knowledge fusion loss function module. Defined as temporal loss and knowledge loss The weighted sum is expressed as:

[0111]

[0112] in and These are adaptive weighting coefficients. Representing the number of training rounds, they vary with normalized training progress. Linear evolution is defined as:

[0113]

[0114] in The total number of training iterations is represented by the weight update representation:

[0115]

[0116] in , and , Let represent the initial and final values ​​of the time and knowledge loss weights, respectively. To ensure a convex combination throughout the training process, the weights are constrained to sum to 1, i.e. .

[0117] This yields the final total loss of the model. .

[0118] Step 4: Train the KWNet model constructed in Step 3 using the industrial time series dataset from Step 1.

[0119] During training, the parameters of the model are optimized using a specified optimizer to obtain the optimal model.

[0120] Step 5: Predict industrial temperature based on the training model obtained in Step 4.

[0121] Accurate prediction of industrial temperature parameters under varying operating conditions is achieved using a trained KWNet model.

[0122] Example 2

[0123] This embodiment provides a method for predicting industrial temperature under varying operating conditions based on the KWNet model. Based on the prediction model given in Embodiment 1, this embodiment conducts a set of experiments on six real-world datasets (denoted as datasets A, B, C, D, E, and F) collected from industrial extraction tanks in a smart pharmaceutical factory to verify the performance of the KWNet model proposed in this application.

[0124] For each tank, the time series dataset includes records from January 1, 2024 to January 31, 2024, containing 44,470 samples. The collected parameters include the extraction tank temperature, extraction tank pressure, extraction solvent flow rate, secondary insulation steam pressure, secondary heating temperature, and storage tank temperature.

[0125] To meet actual industrial needs, this embodiment performs predictions with lead times of 5 and 10 minutes on all datasets, i.e., predicting the temperature value 5 or 10 minutes later based on the current time t. To highlight the performance of the KWNet model proposed in this application, this embodiment also compares its prediction performance with that of five existing models: Transformer, Dlinear, Timemixer, Patchtst, and Frets.

[0126] Transformer can refer to A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L.Jones, A. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017, pp. 5998–6008.

[0127] For Dlinear, please refer to A. Zeng, M. Chen, L. Zhang, and Q. Xu, "Are transformers effective for time series forecasting?" in Proc. AAAI Conf. Artif. Intell., 2023, pp. 11121–11128.

[0128] For Timemixer, please refer to S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J.Zhang, and J. Zhou, "Timemixer: Decomposable multiscale mixing for timeseries forecasting," in Proc. Int. Conf. Learn. Representations (ICLR), 2024.

[0129] For Patchtst, please refer to Y. Nie, NH Nguyen, P. Sinthong, and J. Kalagnanam, "A time series is worth 64 words: Long-term forecasting with transformers," in Proc. Int. Conf. Learn. Representations (ICLR), 2023.

[0130] For Frets, please refer to K. Yi, Q. Zhang, W. Fan, S. Wang, P. Wang, H. He, N. An, D. Lian, L. Cao, and Z. Niu, "Frequency-domain MLPs are more effective learners in time series forecasting," in Proc. Adv. Neural Inf. Process.Syst. (NeurIPS), 2023, pp.76656-76679.

[0131] The mean absolute error (MAE), mean squared error (MSE), and weighted average absolute percentage error (WMAPE) are used as the criteria for evaluating the performance of each model, and are expressed as follows:

[0132]

[0133]

[0134]

[0135] in , N and N represent the true value, predicted value, and sample size, respectively.

[0136] A key difference between Mean Absolute Error (MAE) and Mean Squared Error (MSE) lies in their sensitivity to outliers and error distribution. MAE offers greater robustness and is less affected by outliers, but it fails to capture the variance in the prediction error. In contrast, MSE amplifies larger errors through squaring, making it more sensitive to outliers. While MAE and MSE focus on error magnitude, the Weighted Average Absolute Percentage Error (WMAPE) normalizes the total absolute error to the sum of actual values, providing a scale-invariant measure of prediction accuracy.

[0137] The comparison results of different models on datasets A, B, and C are shown in Table 1 below, and the comparison results of different models on datasets D, E, and F are shown in Table 2 below.

[0138] Table 1: Performance comparison of different models on datasets A, B, and C

[0139]

[0140] Table 2: Performance comparison of different models on multiple datasets D, E, and F

[0141]

[0142] As shown in Tables 1 and 2, on dataset A with a 5-minute lead time, the MAE of the KWNet model proposed in this application reaches 0.5489, which is a significant improvement of 17.20% compared to the second-best model, FreTS (MAE=0.6628). Correspondingly, the MSE decreases to 3.1446 (a decrease of 28.88% compared to FreTS's 4.421), and the WMAPE decreases to 0.67% (a decrease of 17.28% compared to 0.81%). This is attributed to the effective decomposition of the original signal by the novel dual-stream wavelet transform module proposed in this application, which can cleverly capture transient fluctuations and seasonal patterns.

[0143] Similarly, in dataset E, with a lead time of 10 minutes, the MAE of our model is 1.3896, which is 21.30% better than TimeMixer (MAE=1.7659). At the same time, the MSE is reduced by 25.15% (17.4851 vs. 23.3592), and the WMAPE is reduced by 21.95% (0.0160 vs. 0.0205). This is due to the integration of knowledge-driven rules, which keeps the predicted trend consistent with the physical production rules under different operating conditions. Furthermore, on dataset D with a 5-minute lead time, KWNet reduced the MAE to 0.5425, a significant decrease of 27.64% compared to the suboptimal DLinear (MAE=0.7499), while also reducing MSE by 34.19% (3.6697 vs. 5.5768) and improving WMAPE by 27.06% (0.62% vs. 0.85%). This improvement is attributed to the multi-scale dilated convolutional pyramid set in the time-frequency fusion decoder, which optimizes the reconstruction of non-stationary signal components. Additionally, the MAE for 10-minute lead time predictions was reduced by an average of 21.31%, and the MSE by an average of 17.71%. These continuous improvements across multi-tank datasets highlight KWNet's strong generalization capabilities, primarily benefiting from the integration of knowledge-driven rules, ensuring that predicted trends align with production principles under different operating conditions.

[0144] In addition, this application also conducted ablation experiments on the KWNet model to further explore the impact of each module on the model performance. The specific data is shown in Table 3 below:

[0145] Table 3: Ablation Experiment Results for Different Module Configurations

[0146]

[0147] Table 3 shows the progressive performance improvement observed in each variant from the basic module to the complete model, which strongly validates the necessity and effectiveness of each module in the KWNet framework.

[0148] Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 , Figure 11 The figures compare the prediction performance of Tank A on the training set for 10 minutes using Transformer, DLinear, TimeMixer, PatchTST, FreTS, and KWNet. As can be seen from the figures, the prediction curve of the KWNet-based prediction model proposed in this application is highly consistent with the actual temperature curve, and it outperforms the other five comparative models at the points of steady temperature change and transition.

[0149] Figure 12 , Figure 13 , Figure 14 , Figure 15 , Figure 16 , Figure 17 The violin plots showing the absolute error distribution of different models on six datasets reveal that the KWNet-based prediction model proposed in this application exhibits a narrow and compact chord plot, with its median closer to zero and its interquartile range shorter. This indicates that the absolute error of most prediction results is close to zero, the predicted values ​​are close to the true values, and large errors are rare. This demonstrates that the KWNet-based prediction model proposed in this invention can achieve accurate prediction of industrial temperatures under varying operating conditions.

[0150] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting industrial temperature under varying operating conditions based on the KWNet model, characterized in that, The method includes: Step 1: Collect historical parameters from specific industrial processes to construct an industrial time series dataset; Step 2: Develop expert knowledge for specific industrial processes; Step 3: Construct a prediction model based on KWNet by combining the established expert knowledge; Step 4: Train the KWNet-based prediction model using the industrial time series dataset; Step 5: Use the trained KWNet-based prediction model to complete the multi-step prediction of industrial temperature parameters under varying operating conditions. The KWNet-based prediction model includes a two-stream wavelet transform module, a learnable dynamic gating mechanism module, a domain knowledge analysis and formulation module, an adaptive time-knowledge fusion loss function module, and a time-frequency fusion decoder. The domain knowledge analysis and formulation module is used to formulate the established expert knowledge and construct the expert knowledge loss function. ; The dual-stream wavelet transform module is used to reconstruct seasonal and trend components from the original input data after wavelet transform. The learnable dynamic gate mechanism module is used to determine the weights of the seasonal and trend components to obtain the overall reconstructed data. This overall reconstructed data is then input into a time-frequency fusion decoder to obtain preliminary prediction results. Based on these preliminary prediction results, a time-domain loss function is derived. The adaptive time-knowledge fusion loss function module is used to integrate the expert knowledge loss function. and the time-domain loss function The total loss function of the KWNet-based prediction model is obtained. .

2. The method according to claim 1, characterized in that, The dual-stream wavelet transform module includes a wavelet decomposition unit, a frequency processing unit, and a time-frequency joint feature extractor. First, the original input data is decomposed into m-level discrete wavelet decomposition using wavelet decomposition units to generate a set of approximate coefficients. and detail coefficient , j =2,3,…,m; The detail coefficients generated by the wavelet decomposition unit The input frequency processing unit obtains , and approximation coefficient The seasonal components were calculated together. ; The approximation coefficients generated by the wavelet decomposition unit The input time-frequency joint feature extractor processes the data to obtain the desired feature. , and detail coefficient The trend components are calculated together. .

3. The method according to claim 2, characterized in that, The learnable dynamic gate mechanism module will convert the seasonal components obtained by the dual-stream wavelet transform module. and trend components The reconstructed signal is obtained by splicing the last d-dimensional phases. The reconstructed signal is obtained by splicing together linear layers with an added Sigmoid function. Projection, generating gated tensors Then, the output data is obtained by weighted summation through gates. .

4. The method according to claim 3, characterized in that, The time-frequency fusion decoder uses a fixed core size in the time branch design. and three different expansion rates The pyramid multi-scale dilated convolution is used to capture local temporal abrupt changes and long-range contextual dependencies in a hierarchical manner; at the same time, multi-head Fourier modulated attention is used in the frequency branch to extract discriminative frequency features; the pyramid multi-scale dilated convolution and multi-head Fourier modulated attention are adaptively fused through a learnable dynamic gating mechanism to finally obtain a unified time-frequency representation.

5. The method according to claim 4, characterized in that, The total loss function of the KWNet-based prediction model obtained by integrating the adaptive time-knowledge fusion loss function module. for: in and These are adaptive weighting coefficients. .

6. The method according to claim 5, characterized in that, The specific industrial process in step 1 is the extraction process, and the historical parameters of the extraction process include the temperature inside the extraction tank, the pressure inside the extraction tank, the flow rate of the extraction solvent, the pressure of the secondary heat preservation steam, the secondary heating temperature, the temperature of the storage tank, and the valve opening status.

7. The method according to claim 6, characterized in that, The expert knowledge defined for the extraction process in step 2 includes: Expert Knowledge Rule 1: If the second steam valve changes from closed to open, the predicted temperature of the extraction tank is... With minimum increment Monotonically increasing; otherwise, if the prediction violates expert knowledge rule one, it will be penalized with additional knowledge loss. Expert Knowledge Rule 2: If the work procedure changes from the end of the first drainage to the second water addition, then the predicted temperature of the extraction tank... Monotonically decreasing, and the minimum decrease is Otherwise, if the prediction violates Rule 2 of Expert Knowledge, it will be penalized with additional knowledge loss.

8. The method according to claim 7, characterized in that, The expert knowledge loss function obtained based on expert knowledge rule one and expert knowledge rule two for: in, in This represents the predicted tank temperature. and The two terms represent the tank temperature difference predicted by two consecutive predictions based on expert knowledge rule one and expert knowledge rule two, respectively. Indicates the prediction time domain; and This is an indicator function for the triggering conditions of Expert Knowledge Rule 1 and Expert Knowledge Rule 2. The total number of violations of Expert Knowledge Rule 1 and Expert Knowledge Rule 2 within the batch; To calculate the number of batches, The predicted time length of the sequence.

9. The method according to claim 8, characterized in that, The time-domain loss function for: in , N and N represent the actual tank temperature, the predicted tank temperature, and the sample size, respectively.

Images