Deep belief network temperature prediction method based on selective update strategy

Abstract

The invention provides a deep belief network temperature prediction method based on a selective update strategy. The method comprises the following steps: collecting and standardizing historical and real-time sample data; calculating mutual information of the input variable and the target temperature, and weighting the sample distance according to the mutual information to select a local sample; calculating a variable weighting matrix based on the local sample, and constructing a local weighting data set; and calculating the sample center distance of the query sample, multiplexing the existing model when the sample center distance is smaller than a set threshold value, otherwise, retraining the deep belief network model for prediction. The system comprises a data acquisition module, a mutual information calculation module, a sample selection module, a distance determination module and a modeling prediction module. According to the method, through mutual information weighting and selective updating strategies, the real-time performance and prediction precision of the model are improved, the online modeling time is remarkably shortened, and the method has high anti-interference capacity and engineering applicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of online prediction and intelligent control of industrial processes, and specifically discloses a deep belief network temperature prediction method based on a selective update strategy, which is suitable for online prediction and control of the temperature of industrial furnaces such as regenerative smelting furnaces and heating furnaces. BACKGROUND

[0002] In high-temperature industrial processes (such as aluminum smelting and billet heating), accurate prediction of furnace temperature is of great significance to product quality, energy consumption control, and safety management. Existing temperature prediction methods based on neural networks have made some progress in improving nonlinear modeling capabilities, but there are still some problems in online deployment: frequent model updates result in high online computation overhead, the model's generalization ability under different working conditions is insufficient, and the influence of input features is not dynamically weighted, making it difficult for the prediction system to meet the requirements of high precision and high real-time performance at the same time.

[0003] Therefore, there is a need for a temperature prediction method and system that can maintain the nonlinear representation ability of deep models, reduce the number of retraining times in online scenarios through selective updating, improve real-time performance, and enhance the effectiveness of local modeling through feature weighting. SUMMARY

[0004] The purpose of the present application is to propose a deep belief network (DBN) temperature prediction method based on a selective update strategy, which combines mutual information weighting, local sample selection, and a sample center distance-driven selective update mechanism to significantly reduce the number of online retraining times and modeling time while ensuring prediction accuracy, thereby improving the real-time performance and engineering usability of online prediction.

[0005] To achieve the above purpose, the present application provides the following technical solution - a deep belief network temperature prediction method based on a selective update strategy, which comprises the following steps:

[0006] Data acquisition and preprocessing: Collect historical samples and real-time query samples of the smelting furnace / heating furnace during operation. The original data includes process parameters such as furnace temperature readings, heating power, furnace gas flow rate, furnace pressure, exhaust gas temperature, intake air temperature, and ambient temperature at several time points. Perform missing value processing, outlier removal, and standardization (e.g., Z-score standardization) on the collected data to obtain a feature vector set for modeling.

[0007] Considering the nonlinearity of the aluminum smelting process, mutual information is used to calculate the correlation between variables. Mutual information is an indicator that measures the degree of mutual dependence between two variables and has been widely used to assess nonlinear relationships. For two variables and , the mutual information between the two variables is is represented as:

[0008] ,

[0009] where , represent the probability density functions of and respectively, is the joint probability density function.

[0010] Global mutual information weight calculation: The training dataset includes an input data matrix X and y. In the input data matrix, each column represents a sample sequence of an input variable. The mutual information between X and y is represented as , which can be calculated according to the above formula. Then, the weight coefficient of the kth input variable is calculated as the normalized mutual information:

[0011] ,

[0012] A diagonal weight matrix W is constructed , where m is the dimension of the input variable, for the subsequent calculation of the weighted distance.

[0013] Mutual information weighted sample distance and local sample selection: For the current query sample , the mutual information weighted distance between it and the historical sample is calculated:

[0014] ,

[0015] Sort by distance from small to large, select the first samples to form the local modeling sample set (in experiments and implementations, the optional range is 100-300, preferably , to achieve a balance between precision and computational cost).

[0016] The N samples with the strongest correlation to the query sample are selected using the global mutual information weighting method, and the local model is updated online to predict the output of the current query sample . This immediate learning strategy using mutual information weighting is called (MIW-JITL). The immediate learning strategy using mutual information weighting is a dynamic learning method, aiming to quickly respond to changing data environments by analyzing the relationship between the characteristics of the data and the target variable in real time, thereby improving the predictive ability of the model.

[0017] Local mutual information reweighting (variable weighting): In order to make DBN better in modeling, the weight of the feature which has significant influence on the prediction target is enhanced in the sample selection process. In this paper, the N samples which have the strongest correlation with the query sample selected by global mutual information weighting method are further weighted. The mutual information value of N m-dimensional local modeling samples is calculated , and the local mutual information matrix is obtained as:

[0018]

[0019] The local variable weighting matrix used as local modeling samples and query samples is:

[0020]

[0021]

[0022] where is composed of local modeling samples, and are variable-weighted local modeling samples and variable-weighted query samples, respectively.

[0023] Therefore, the new local modeling dataset is used to establish a local model. After obtaining the new local modeling samples, a deep belief network is used as a local modeling algorithm. Thus, the DBN using the new local modeling dataset as input is called a variable-weighted deep belief network (W-DBN). This method fully utilizes the variable characteristics and pre-weights the input data at the input end, which is beneficial to improve the prediction accuracy.

[0024] The modeling data selected by global mutual information weighting is further filtered by local mutual information weighting, and the local modeling dataset with the strongest correlation with the query sample is finally obtained and input into the deep belief network for modeling. Combined with the global mutual information weighting-based immediate learning strategy and the variable-weighted deep belief network, this model is called a double-weighted deep belief network (DW-DBN).

[0025] The sample center distance (SD) determines the selective update strategy as shown in Figure 2 : the local sample center (or the corresponding weighted center) is calculated, and the Euclidean distance of the query sample relative to the center is calculated:

[0026] ,

[0027] where represents the local weighted query sample vector. When , When the threshold is reached, the corresponding trained local model is directly reused for prediction. When the threshold is not reached, the local deep belief network model is retrained using the training set and the prediction result is outputted.

[0028] The threshold is recommended by experiments In the range of 2.3-2.8, preferably This distance determination and selection strategy based on sample center is called a deep belief network (SD-DW-DBN) based on a selection update strategy and double mutual information weighting.

[0029] Construction and training of local model: the local model uses a deep belief network, DBN, which is composed of a stack of several layers of restricted Boltzmann machines (RBM), and is first pre-trained layer by layer in an unsupervised manner, and then fine-tuned based on labeled samples (back propagation to minimize mean square error MSE). In the embodiment, the DBN can use a stack structure of three layers of RBM, and the training uses an early stopping strategy and a validation set to prevent overfitting.

[0030] Prediction output and online feedback: after completing the prediction, the predicted temperature is compared with the real-time observation value, the error is recorded, and the measured sample (together with the label) is added to the historical sample library or sliding window library for subsequent local sample selection and weight reestimation. At the same time, according to the historical error trend, the threshold or the number of local samples can be adaptively adjusted to maintain system stability.

[0031] Preferred embodiment: the mutual information estimation uses common non-parametric methods such as kernel density estimation; the local sample retrieval can first use approximate nearest neighbor pre-screening to reduce retrieval delay; the local model can be saved as a portable model file for fast reuse. The above preferred embodiment is a specific engineering suggestion for the embodiment, which can be selected and adjusted according to the site constraints.

[0032] Beneficial effects: by combining the mutual information weighting and the selective update mechanism driven by the sample center distance, the present application reduces the retraining frequency while ensuring the strong expression ability of DBN, significantly improving the online prediction efficiency. Experimental verification (see the embodiment) shows that the present application method is superior to the traditional method in terms of MAE, RMSE, and other indicators, and the average online processing time is significantly shortened, having good engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flowchart of the deep belief network temperature prediction method based on the selection update strategy in the embodiment of the present application; ​

[0034] Figure 2 is a selective update judgment and local model reconstruction schematic diagram based on sample center distance in the embodiment one of the present application;

[0035] Figure 3 is a deep belief network temperature prediction system structure schematic diagram based on a selective update strategy in the embodiment two of the present application;

[0036] Figure 4 is a deep belief network temperature prediction device structure schematic diagram based on a selective update strategy in the embodiment three of the present application. DETAILED DESCRIPTION

[0037] The present application will be further described below in conjunction with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, in order to facilitate the description, only the parts related to the present application are shown in the drawings, not all structures.

[0038] Before discussing the example embodiments in more detail, it should be noted that some embodiments of the present application are described in the form of a series of processing steps as a flow chart. Although the flow chart shows the steps in order, many of the steps can be performed in parallel, concurrently or at the same time. In addition, the order of execution of the steps can be adjusted or rearranged according to the specific application scenario. When its operation is completed, the flow can be terminated, or it can contain additional steps not listed in the drawings. The flow can correspond to methods, functions, procedures, subprograms, etc.

[0039] The embodiment of the present application is directed to a deep belief network temperature prediction method based on a selective update strategy, which provides the following embodiments:

[0040] Based on the embodiment 1 of the present application.

[0041] This embodiment is used to illustrate a deep belief network temperature prediction method based on a selective update strategy. Referring to Figure 1 , it is a flow chart of the deep belief network temperature prediction method based on a selective update strategy of the present application, which specifically includes the following steps:

[0042] Engineering background and target: deploy the method of the present application in the temperature prediction task of the regenerative aluminum smelting furnace, the target is to predict the temperature of the aluminum liquid in real time during the operation of the furnace body to assist the control system to adjust the power and optimize the process, so as to improve the product temperature stability and reduce the energy consumption.

[0043] Data collection and sample construction: Several temperature points and relevant working condition quantity collection channels are arranged on the site of the smelting furnace. Several thousand pieces of historical operation data are collected as candidate sample library (about 4000 pieces of historical samples) and part of the samples are divided as reserved test set (285 pieces). The collected data items include but are not limited to: furnace multi-point temperature reading, heating power, combustion air speed, furnace pressure, air inlet temperature, exhaust gas temperature, material level and environmental temperature, etc. All raw data are labeled in time sequence and stored in time sequence database.

[0044] Data preprocessing steps:

[0045] Missing value processing: missing values caused by sensor interruption or communication packet loss are filled by interpolation or forward filling strategy; samples or channels with missing value ratio higher than the preset threshold are removed and recorded for maintenance.

[0046] Abnormal value detection: experience threshold or sliding window based Abnormal readings are removed, and the removal records are fed back to the maintenance system.

[0047] Standardization: the final reserved samples are standardized by Z-score (based on the mean and standard deviation of the training set) to obtain standardized feature vectors, ensuring the consistency of the numerical scale of each variable, facilitating subsequent mutual information estimation and DBN training.

[0048] Global mutual information weight calculation and weighted matrix construction: based on the historical sample set, the mutual information between each input variable and the target temperature is calculated :

[0049] ,

[0050] In the modeling of liquid aluminum temperature prediction, N (N<H) samples are selected from the historical samples , to locally model the query sample. First, the mutual information values between the input variables and the output variables of all historical samples are calculated to obtain the global mutual information matrix:

[0051] ,

[0052] where m is the dimension of the input variable, is the normalized mutual information value. It is used to measure the overall contribution of each input variable and participates in sample distance calculation.

[0053] Local sample selection (mutual information weighted distance): when a query sample is received in real time (the sample is the standardized input vector at the current time), the system calculates the mutual information weighted distance in the historical sample library:

[0054] ,

[0055] And according to distance Sort by size from smallest to largest, select the top... A number of historical samples constitute local modeling samples In engineering practice, to balance accuracy and real-time performance, the preferred method is... When the sample database is extremely large, a KD-tree or approximate nearest neighbor algorithm can be used for coarse screening to reduce retrieval latency.

[0056] Local mutual information reweighting and data construction: in a selected local sample set Above, the mutual information between each variable and the target is re-estimated to obtain local weights. Construct local diagonal moments The local sample features are then weighted twice to strengthen the variables that have a stronger impact on the target under the local distribution, thus obtaining a locally weighted dataset for training.

[0057] Sample center distance determination (selective update strategy): such as Figure 2 The diagram shown illustrates selective update determination and local model reconstruction based on sample center distance, calculating local sample centers (or weighted centers). And calculate the weighted Euclidean distance (sample center distance) from the query sample to the center:

[0058] ,

[0059] in This is a locally weighted query sample. When... When the system determines that the distribution of the query samples matches the cached local model, it directly reuses the cached model for fast inference; when The system determines that a distribution shift has occurred and triggers [a certain action]. Reconstruct the local DBN model. Optimize the threshold in engineering practice. The optimal value is found within the interval [2.3, 2.8]. It can also be combined with field error indicators to achieve adaptive fine-tuning (see adaptive strategy below).

[0060] Local DBN model training and inference process: The local model adopts a deep belief network architecture, which is composed of multiple layers of restricted Boltzmann machine (RBM) stacking. The training process includes: layer-by-layer unsupervised pre-training to initialize the parameters (obtain the initial weight by greedy training of each layer of RBM), and then supervised fine-tuning with labeled local weighted samples (back propagation to reduce MSE). After training, the model is serialized and cached (saved as a model file) to support fast reuse of subsequent queries of the same distribution. In order to balance online and engineering feasibility, online training can be done in small batches on the edge computing node or in the cloud with higher computing power to complete and issue the latest model.

[0061] Predicted output and online feedback: Local DBN output predicts temperature The system calculates the error between the predicted value and the subsequent measured temperature and records it in the performance log. If the error exceeds the preset threshold for several consecutive times, or the error shows an upward trend over time, the system triggers the threshold adaptive module to dynamically adjust or increase / decrease to ensure the balance between accuracy and computational load. In addition, the measured samples together with the labels can enter the sliding window sample library to replace or supplement the historical samples, realizing the timeliness maintenance of the sample library.

[0062] Engineering implementation points and guarantees: In order to ensure on-site operability, it is recommended to deploy inference and quick judgment on edge devices (industrial computers), and to use asynchronous strategy or complete retraining tasks in low peak period by cloud; At the same time, it is recommended to implement version management for model cache, sample library and log, in order to audit and rollback. The system should also design a degradation strategy for sensor abnormalities and communication interruptions to ensure that the control system can fall back to a safe baseline when the prediction module is abnormal.

[0063] Based on the embodiment 2 of the application

[0064] The deep belief network temperature prediction system 300 based on the selective update strategy provided by the embodiment two of the application can execute the deep belief network temperature prediction method based on the selective update strategy provided by the embodiment one of the application, and has the corresponding function modules and beneficial effects of the execution method. The system can be realized by software and / or hardware (integrated circuit), and can be integrated in an industrial control server or a temperature prediction terminal device. Figure 3 The structure diagram of the deep belief network temperature prediction system 300 based on the selective update strategy in the embodiment two of the application. Referring to Figure 3 , the deep belief network temperature prediction system 300 based on the selective update strategy of the embodiment of the application specifically includes:

[0065] The data acquisition module 310 is configured to acquire working condition data of the regenerative aluminum smelting furnace in different operation stages, including parameters such as furnace temperature, heating power, fuel gas flow, regenerator reversing period, furnace pressure, and inlet and outlet gas temperature; and real-time acquisition of temperature distribution information measured by multiple-point temperature sensors (K-type thermocouples) in the furnace chamber to form a historical sample data set and a real-time input data set.

[0066] The feature processing and selection module 320 is configured to perform data cleaning, outlier removal, and normalization processing on the collected original working condition data; and determine a global feature weight matrix through mutual information analysis , and calculate a local mutual information weight based on a local sample set , to form a local weighted matrix for enhancing the feature signal sensitive to the furnace temperature change.

[0067] The deep belief network construction module 330 is configured to construct a deep belief network structure of a multi-layer restricted Boltzmann machine (RBM) stack based on the weighted input sample, perform layer-by-layer unsupervised pre-training on the furnace temperature time series feature, extract high-dimensional nonlinear features, and initialize network parameters.

[0068] The selective update strategy module 340 is configured to determine whether to trigger a parameter update operation in the model training and inference stage; calculate the change rate of the prediction error at consecutive time points , and compare it with a set threshold value , when , automatically trigger a local parameter update, otherwise keep the model in a frozen state, to reduce the calculation overhead and prevent overfitting.

[0069] The prediction and output module 350 is configured to perform forward propagation of the multi-dimensional working condition input at the current time through the deep belief network, and output the temperature prediction value of the furnace chamber at the future t time; and after the prediction result is subjected to inverse normalization processing, the prediction result is transmitted to the upper computer control system in real time, to realize intelligent prediction and regulation of the smelting furnace.

[0070] The model maintenance and visualization module 360 is configured to record the update state of the model, the threshold change curve, and the prediction error statistical result; and through a human-computer interaction interface, display the comparison trend of the furnace temperature prediction curve and the actual measurement value in real time, to provide visual decision support for the operator.

[0071] In addition to the above-mentioned modules, the deep belief network temperature prediction system 300 based on the selective update strategy can also include other components, such as a database interface, a communication module, and a cloud storage module. However, since these components are irrelevant to the core technology of the present application, their illustrations and descriptions are omitted here.

[0072] The specific working process of the deep belief network temperature prediction system 300 based on the selective updating strategy is described with reference to the deep belief network temperature prediction method based on the selective updating strategy in the above-described embodiment 1, and will not be described again.

[0073] Based on the embodiment 3 of the present application

[0074] The deep belief network temperature prediction device based on the selective updating strategy provided in the embodiment 3 of the present application can also execute the deep belief network temperature prediction method based on the selective updating strategy provided in the embodiment 1 of the present application, and has the corresponding function modules and beneficial effects of the execution method. Refer to Figure 4 The deep belief network temperature prediction device based on the selective updating strategy provided in the embodiment 3 of the present application can also execute the deep belief network temperature prediction method based on the selective updating strategy provided in the embodiment 1 of the present application, and has the corresponding function modules and beneficial effects of the execution method. Refer to

[0075] The regenerative aluminum smelting furnace is used to obtain multi-source working condition data in a smelting process, and collect real-time parameter data including multi-point temperature of a hearth, heating power, gas flow, furnace pressure, inlet and outlet gas temperature, regenerative switching period, furnace door opening and closing state, and environment temperature. High-frequency sampling is realized through the thermocouple and flowmeter sensors arranged, and the temperature distribution data and power signal are transmitted to a data processing terminal.

[0076] The electronic device is provided as a data processing unit and a furnace temperature prediction unit, wherein:

[0077] The data processing unit includes: establishing a furnace temperature feature sample database by using the multi-source working condition data collected by the smelting furnace; performing denoising and standardization preprocessing on the original data, and calculating a global weight matrix based on mutual information between each feature variable and the target furnace temperature ; further calculating a local mutual information weight based on real-time input samples , forming a local weighted matrix , and constructing a local modeling sample set according to the historical sample similarity; when the Euclidean distance between the current input sample and the local sample center is less than a set threshold , the deep belief network model of the last round of training is reused, otherwise the local parameter updating is performed again; the deep belief network realizes hierarchical extraction of furnace temperature time series features through a multi-layer restricted Boltzmann machine stacking structure, and completes model optimization through layer-by-layer pre-training and supervised fine-tuning, and finally outputs a furnace temperature prediction value.

[0078] The furnace temperature prediction unit comprises: inputting the real-time collected furnace temperature and working condition into the trained deep belief network model, performing forward propagation operation, and outputting the furnace temperature prediction result at the future time; after the prediction output is subjected to inverse normalization processing, readable temperature value is generated, and the result is sent to the upper computer control system in real time to assist in adjusting the heating power and gas supply, and realizing intelligent dynamic control of the furnace temperature.

[0079] The specific working process of the deep belief network temperature prediction device based on the selective updating strategy (as shown in Figure 4 The specific working process of the deep belief network temperature prediction device based on the selective updating strategy (as shown in

[0080] As can be seen from the above embodiments of the present application, the deep belief network temperature prediction method and device based on the selective updating strategy provided by the present application can fully combine the nonlinear and multivariate dynamic characteristics of the regenerative aluminum smelting furnace, realize high precision and high real-time performance of furnace temperature prediction, and the like.

[0081] By introducing mutual information weighting and local sample selection mechanism, the model redundancy is effectively reduced, and the correlation of input features is improved.

[0082] Through the selective updating strategy, the repeated training of the model under stable working conditions is reduced, and the calculation efficiency is improved.

[0083] Through the multi-layer feature learning structure of the deep belief network, the representation ability of the model to complex coupling relationship is enhanced, so that the prediction result is more smooth and stable.

[0084] The present application can be widely applied to temperature prediction and process control of aluminum smelting furnaces, billet heating furnaces, ceramic sintering kilns and other industrial equipment, and provides reliable support for energy saving and automatic operation of metallurgical and manufacturing processes.

[0085] It should be noted that the above is only the preferred embodiment of the present application and the technical principle applied. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and those skilled in the art can make various obvious modifications, adjustments or replacements without departing from the spirit and essence of the present application. Although the present application has been described in detail through the above embodiments, the protection scope of the present application is not limited thereto, but should be limited by the scope defined in the appended claims and their equivalent forms.

Claims

1. A method for temperature prediction based on deep belief network with selective update strategy, characterized in that, Comprise: Step 1: Collect and standardize historical process data and real-time query sample data; the data includes furnace temperature, furnace pressure, combustion air parameters, exhaust gas temperature, air-fuel ratio and other working condition variables; Step 2: Calculate the correlation between input variables and target temperature based on mutual information, and construct a global mutual information weight matrix; Step 3: Select the first N local samples most similar to the query sample according to the mutual information weighted sample distance, and construct a local modeling sample set; Step 4: Calculate the mutual information weight of the local sample, locally weight the input variable, and get the variable weighted local data set; Step 5: Calculate the sample center distance of the query sample, when the distance is less than a set threshold, call the existing local model for prediction; when the distance is greater than the threshold, retrain the local model ​​ Step 6: The local model is a deep belief network model based on restricted Boltzmann machine, which is trained through layer-by-layer unsupervised pre-training and supervised fine-tuning to realize the prediction output of the target temperature. 2.The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, The mutual information weight calculation formula is: ， wherein is the mutual information value of the kth input variable and the target temperature. 3.The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, The calculation formula of mutual information weighted sample distance is: ， wherein is a global mutual information weight matrix, is a query sample, is a history sample. 4.The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, The sample center distance calculation formula is: ， where N is the number of local samples; when the multiplexing model, when the reconstruction model. 5.The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, Threshold value The range of values of the local sample number N includes: Threshold value is in the range of 2.3 - 2.8, preferably 2.4; The value range of the number of local samples N is 100-300, preferably 200. 6.The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, The deep belief network comprises a plurality of restricted Boltzmann machine (RBM) stacked structures, and the initial value of the weight is obtained through layer-by-layer greedy unsupervised pre-training, and then fine-tuned through supervised learning to optimize the prediction accuracy.

7. The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, The calculation of mutual information uses the empirical method based on kernel density estimation, and the mutual information value of the input variable and the target output is calculated through the probability density of sample distribution.

8. The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, The deep belief network temperature prediction method based on the selective updating strategy further comprises: error evaluation on the predicted output and the actual measured value, and automatic updating of the threshold value when the error exceeds a preset threshold value to realize adaptive optimization of the threshold value. 9.The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, Through the combination of mutual information weighting and sample center distance judgment, the dynamic maintenance of the local model is realized, the repeated modeling under stable working conditions is avoided, and the online prediction time is reduced and the real-time performance is improved.

10. The deep belief network temperature prediction method based on selective update strategy of claim 1, wherein, Comprise: Data acquisition module: for acquiring and standardizing historical and real-time working condition data; Feature processing and selection module: for calculating the mutual information value of the input variable and the target temperature and constructing the weight matrix, and selecting the local modeling sample according to the mutual information weighted sample distance; Deep belief network construction module: for training or calling the local DBN model for temperature prediction; Selective update strategy module: for calculating the sample center distance and determining whether to update the model according to the threshold value; Prediction and output module: for outputting the prediction result and feeding back the updated model parameters.

Images