Industrial key parameter long-time prediction method based on double-offset fragment division and adaptive normalization

Abstract

The invention provides an industrial key parameter long-time prediction method based on double-offset fragment division and self-adaptive normalization aiming at optimization control requirements of an industrial process with a time delay characteristic. During prediction, preprocessed industrial data are input into the trained model, and the model firstly performs multi-band decoupling and noise reduction. Secondly, modeling branch networks for the low-frequency components and the intermediate-frequency components respectively; normalization processing based on self-adaptive working condition division is carried out on the low-frequency component and the intermediate-frequency component respectively. The low-frequency component is subjected to feature extraction through a multi-layer perceptron, and the intermediate-frequency component is subjected to feature extraction through a segment-level modeling method based on double-stage offset. And then, inputting the two parts of features into a multi-layer perceptron to obtain a prediction result of each branch, adding the prediction results after reverse normalization, and finally generating an overall prediction result. Besides, on the basis of adopting an MAE loss function, a direction alignment loss function based on relative position weighting is introduced as a new item, so that the attention of the model to a specific layer in the training process is enhanced. According to the method, the long-time prediction precision of the industrial key parameters is remarkably improved, and technical support is provided for optimization control of the industrial process with the time delay characteristic.

Description

Technical Field

[0001] The invention belongs to the field of industrial time series prediction, and in particular is a method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization. Background Art

[0002] Large time lags are common in modern industry. Specifically, due to the time lag in the transmission of materials, energy or information in industrial systems, the changes in key parameters detected usually lag behind the actual production dynamics. This time lag makes it difficult for us to track changes in industrial processes in a timely manner. By predicting the change trend of key parameters in the future in advance, the optimal adjustment plan can be formulated in advance to compensate for this time lag, so as to adjust the control parameters in time to ensure that the industrial process is always in the best state and avoid waste of resources and reduced production efficiency.

[0003] Long-term prediction tasks require the use of a large amount of historical data to capture potential long-term patterns, which requires the model to have complex feature extraction capabilities, which is a challenge for traditional statistical learning and machine learning methods. Deep learning methods have become the mainstream method due to their good fitting performance for nonlinear features. At present, most deep learning methods use a single sample value as the basic input unit of the model, which easily distracts attention from key information when faced with a large amount of input. Relatively speaking, using sequence fragments as basic input units has two advantages: one is that the number of input units can be greatly reduced while ensuring that the length of the historical data used remains unchanged, and the other is to alleviate the interference of random noise. However, most of the existing methods that use sequence fragments as basic input units use fixed partition windows and window moving steps. This rigid partitioning method may destroy the complete semantic information in industrial process data, resulting in the capture of incorrect patterns, which is not conducive to obtaining accurate prediction results. In addition, the existing long-term prediction methods for key parameters of industrial processes face many challenges in dealing with the compound nonlinearity, information sparsity, imprecision and non-stationarity of industrial process data. Summary of the invention

[0004] In order to solve the above problems in the prior art, the present invention provides a long-term prediction method for industrial key parameters based on double offset segment partitioning and adaptive normalization. During prediction, the preprocessed industrial data is input into a trained model. The model first decomposes the data into low-frequency, medium-frequency and high-frequency components in the frequency domain, and removes the high-frequency components to reduce the interference of industrial noise. Then, the low-frequency and medium-frequency components are converted back to the time domain and input into two independent branch networks for modeling to reduce the complexity of the model. In order to solve the non-stationarity problem caused by operating condition fluctuations, the model performs normalization processing based on adaptive operating condition partitioning on the low-frequency and medium-frequency components respectively. The low-frequency component is feature extracted by a multi-layer perceptron, and the medium-frequency component is feature extracted by a segment-level modeling method based on a dual-stage offset. Subsequently, the two parts of the features are respectively input into the multi-layer perceptron to obtain the prediction results of each branch, and are added after denormalization to finally generate the overall prediction result. In addition, the present invention optimizes the loss function used for model training. On the basis of adopting the MAE loss function, a direction alignment loss function based on relative position weighting is introduced as a new item to enhance the model's attention to specific layers during training. The present invention significantly improves the long-term prediction accuracy of key industrial parameters and provides technical support for the optimal control of industrial processes with time-lag characteristics.

[0005] The specific technical solution adopted by the present invention to achieve the above-mentioned purpose is: a long-term prediction method for industrial key parameters based on double offset segment division and adaptive normalization, performing the following steps to predict target parameters through historical data, so as to provide support for the optimization control of industrial processes with time-lag characteristics; the method comprises the following steps:

[0006] Initialize the long-term prediction model of industrial key parameters based on double-offset segment partitioning and adaptive normalization, collect historical key parameter data, and make a training set after data preprocessing; use the training set data to iteratively train the initial model, improve the loss function and calculate the error between the prediction result and the label, back-propagate the gradient, and then update the model parameters to minimize the loss to obtain the ideal model;

[0007] After the model training is completed, the historical key parameter data corresponding to the industrial process target to be predicted is collected and the data is preprocessed; the trained ideal model is input for long-term prediction.

[0008] The key parameter data of industrial processes include material supply or inventory data related to the production process, control parameter data of electrical equipment that performs actions, and feedback data of sensors; the training set uses the future value of a target parameter in the process production as the key parameter data label to be predicted, and observes the target controlled quantity through other key parameter data and its own historical data to achieve learning, thereby predicting its future long-term time series data. The training set data of this application is a training set made for the key parameters in the loose rehumidification process of food processing {historical values ​​of 24 key parameters, and the future value of the outlet moisture parameter as the target parameter label}.

[0009] The establishment of a long-term prediction model for key industrial parameters based on dual-offset segment partitioning and adaptive normalization includes:

[0010] 1) Multi-band decoupling to obtain mid-frequency and low-frequency components;

[0011] 2) Based on the adaptive working condition division, the low-frequency component and the medium-frequency component are normalized respectively;

[0012] 3) For the normalized intermediate frequency components, the semantically complete sequence segments are obtained based on the two-stage offset partitioning method and are sequentially input into the multi-layer perceptron, convolutional layer and multi-layer perceptron for feature extraction between segments, within segments and between variables;

[0013] 4) For the normalized low-frequency components, a multi-layer perceptron is used to extract features;

[0014] 5) Input the features extracted from the intermediate frequency and low frequency branches into two multi-layer perceptrons respectively to obtain the prediction results and perform denormalization;

[0015] 6) Add the output results of the two branches to obtain the long-term prediction results of the target parameters.

[0016] Multi-band decoupling involves converting the preprocessed data into the frequency domain using fast Fourier transform and decomposing it into high-frequency, medium-frequency and low-frequency components; deleting the high-frequency component and converting the low-frequency and medium-frequency components back to the time domain through inverse Fourier transform.

[0017] The normalization of low-frequency components based on adaptive working condition division includes the following steps:

[0018] Split the process data into multiple segments using windows of fixed size and moving step size;

[0019] The mean of each segment is calculated for the low-frequency component; the variance of each segment is calculated for the mid-frequency component;

[0020] Perform first-order difference processing on the obtained mean sequence or variance sequence and calculate the absolute value, and then identify the indexes corresponding to the largest k values ​​in the sequence;

[0021] Keep the largest value among the k indexes and delete the fragment before that index;

[0022] Expand the retained segments and calculate the mean and standard deviation of the retained ensemble;

[0023] The low frequency component sequence or the intermediate frequency component sequence is subtracted from the mean and divided by the standard deviation to obtain normalized results of the low frequency component or the intermediate frequency component, respectively.

[0024] When the normalization method based on adaptive working condition division is applied to low-frequency components and medium-frequency components, the size of the segment division window and the moving step length are kept equal.

[0025] The steps of obtaining semantically complete sequence segments using the dual-stage offset-based segmentation method are as follows:

[0026] Use windows of fixed size and moving step length to obtain sequence segments of equal length;

[0027] Get the three key parameters: the index of the center position of each segment division window in the entire sequence, the distance from the left boundary to the center position, and the distance from the right boundary to the center position;

[0028] The multi-layer perceptron network is used to learn the offsets of the above three key parameters, and the original window endpoint parameters are adjusted accordingly to obtain the position of the new segment division window;

[0029] Perform equidistant sampling within the above window to generate new fragments;

[0030] The index of each point in the new segment relative to the segment itself is input into the multi-layer perceptron for adaptive offset to obtain the final sequence segment with complete semantics and aligned patterns between segments.

[0031] In the step of obtaining semantically complete sequence segments using the dual-stage offset-based segmentation method, the size of the segment segmentation window is maintained at twice the moving step length; and new segments equal in length to the initial segments are obtained by performing equidistant sampling in the new window.

[0032] The loss function used for model training consists of two parts: MAE loss function and direction alignment loss function weighted by relative position;

[0033] The design of the relative position weighted direction alignment loss function is as follows:

[0034]

[0035] in represents the directional similarity between the fragments generated by the first offset, Represents the directional similarity between the finally obtained fragments;

[0036] Sim rel The calculation formula is as follows:

[0037]

[0038] where p i and p j Represent two different sequence fragments, It represents the cosine similarity between two segments, that is, the alignment in direction.

[0039] The statistical information used in the denormalization process is the same as that used in the normalization process, and the denormalization is completed by multiplying the sequence by the standard deviation and adding the mean.

[0040] The present invention has the following beneficial effects and advantages:

[0041] 1. The present invention is a method for long-term prediction of industrial key parameters based on dual-offset segment division and adaptive normalization. In order to meet the needs of timely optimization and control of industrial processes with time-lag characteristics, a multi-band decoupling network structure, a reversible normalization method based on adaptive working condition division, a segment-level modeling method based on dual-stage offset and a direction alignment loss function based on relative position weighting are adopted. The method has the ability to make long-term predictions of industrial key parameters.

[0042] 2. The present invention proposes a long-term prediction method for key parameters in response to the needs of industrial process optimization control with time-lag characteristics, which effectively addresses the complex nonlinearity, information sparsity, inaccuracy and non-stationarity of industrial process data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 For the loose rehydration process of the food processing industry;

[0044] Figure 2 It is a schematic diagram of model training of the present invention;

[0045] Figure 3 is a flow chart of the method of the present invention;

[0046] Figure 4 Schematic diagram of segment division. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0048] like Figure 1 As shown, in the loosening and rehydration process of the food processing industry, the target parameter is the moisture content of the material at the drum outlet. This target parameter is used as an indicator of production control to characterize whether the current industrial control process meets the set ideal state. The present invention uses historical data (including material supply or inventory data related to the loosening and rehydration process, control parameter data of electrical equipment for executing actions, and feedback data of sensors) to train the model. During actual prediction, the historical data corresponding to the target parameter to be predicted collected in advance is input into the model, and the target parameter to be predicted is automatically output: the moisture content of the material at the drum outlet. The purpose of simulating and predicting future target data based on historical data is achieved.

[0049] like Figure 2 As shown, it is a schematic diagram of the model training of the present invention.

[0050] like Figure 3 The flowchart of the prediction process of the method of the present invention is shown. A method for long-term prediction of industrial key parameters based on double offset segment division and adaptive normalization, the programming language used in the program execution steps of the present invention is not limited to Python and the like.

[0051] The present invention provides a long-term prediction method for industrial key parameters based on dual-offset segment division and adaptive normalization, which uses historical data to predict target parameters by performing the following steps to provide support for optimal control of industrial processes with time-lag characteristics; the specific steps are as follows:

[0052] Step 1: Collect industrial key parameter data. The target parameter is the moisture content of the material at the drum outlet. Preprocess the collected data to create a training set.

[0053] The industrial key parameter data include 24 key parameters, such as actual water addition value, heat exchanger damper opening, moisture exhaust damper opening, loose outlet temperature meter display and loose inlet moisture meter display.

[0054] Step 1-1, remove missing values;

[0055] Step 1-2: Remove outliers according to the set error range;

[0056] Steps 1-3: Perform maximum and minimum standardization processing on the industrial key parameter data and normalize the data to between (0,1).

[0057] Step 1-4, create a training set, take the moisture of the material at the drum outlet as the target to be predicted, observe the target controlled quantity through other key parameter data and the historical data of the target to be predicted, and learn to predict its future long-term data. The training set data of this application includes the key parameters in the loosening and rehydration process of food processing {24 key parameters, the future value of the outlet moisture parameter is used as the target parameter label}.

[0058] Step 2: Establish a long-term prediction model for key industrial parameters based on double-offset segment partitioning and adaptive normalization, use the training set data to iteratively train the initial model, improve the loss function and calculate the error between the prediction result and the label, back-propagate the gradient, and then update the model parameters to minimize the loss and obtain the ideal model.

[0059] Step 21: Perform multi-frequency decoupling on the pre-processed input data, convert it into the frequency domain using Fourier transform, and decompose it into low-frequency components, medium-frequency components, and high-frequency components.

[0060] Step 22: Delete the high-frequency components, and convert the low-frequency components and the medium-frequency components back to the time domain and input them into the two branch networks for feature extraction.

[0061] Step 23: normalize the low-frequency component and the medium-frequency component based on the adaptive working condition division.

[0062] The normalization method based on adaptive working condition division includes different processing steps in the application of low-frequency components and medium-frequency components.

[0063] Step 23-1, normalization of low-frequency components based on adaptive working condition division includes the following steps:

[0064] The process data is split into multiple segments using a window of fixed size and moving step, and the mean of each segment is calculated;

[0065] Perform first-order difference processing on the obtained mean sequence and calculate the absolute value, and then identify the indexes corresponding to the largest k values ​​in the sequence;

[0066] Keep the largest value among the k indexes and delete the fragment before that index;

[0067] Expand the retained segments and calculate the mean and standard deviation of the remaining whole;

[0068] The low frequency component sequence is subtracted from the mean and divided by the standard deviation to obtain a normalized result.

[0069] Step 23-2, normalization of intermediate frequency components based on adaptive working condition division includes the following steps:

[0070] The process data is split into multiple segments using a window of fixed size and moving step, and the variance of each segment is calculated;

[0071] Perform first-order difference processing on the obtained variance sequence and calculate the absolute value, and then identify the indexes corresponding to the largest k values ​​in the sequence;

[0072] Keep the largest value among the k indexes and delete the fragment before that index;

[0073] Expand the retained segments and calculate the mean and standard deviation of the remaining whole;

[0074] The mean value is subtracted from the intermediate frequency component sequence and the result is divided by the standard deviation to obtain a normalized result.

[0075] Among them, the normalization method based on adaptive working condition division keeps the size of the segment division window and the moving step equal when applied to low-frequency components and medium-frequency components.

[0076] Step 24: For the normalized intermediate frequency components, a two-stage offset-based partitioning method is used to obtain semantically complete sequence segments and sequentially input them into a multi-layer perceptron, a convolutional layer, and a multi-layer perceptron to extract features between segments, within segments, and between variables. Figure 4 The figure shows a schematic diagram of the division segments from step 24-1 to step 24-5.

[0077] Step 24-1, using a window with a fixed size and moving step length to obtain sequence segments of equal length;

[0078] Step 24-2, obtaining three key parameters: the index of the center position of each segment division window in the entire sequence, the distance from the left boundary to the center position, and the distance from the right boundary to the center position;

[0079] Step 24-3, using a multi-layer perceptron to learn the offsets of the above three key parameters, and adjusting the original parameters accordingly to obtain the position of a new segment division window;

[0080] Step 24-4, performing equidistant sampling within the above window to generate a new segment;

[0081] Step 24-5, input the index of each point in the new segment relative to the segment itself into the multi-layer perceptron for adaptive shifting to obtain the final sequence segment with complete semantics and aligned patterns between segments;

[0082] Step 24-6: Input the sequence with the fragment as the basic input unit into the multi-layer perceptron, the convolutional layer and the multi-layer perceptron in turn to extract features between fragments, within fragments and between variables.

[0083] In the step of obtaining semantically complete sequence segments using a segmentation method based on a dual-stage offset, the size of the segment segmentation window is maintained at twice the moving step size.

[0084] In the step of obtaining semantically complete sequence segments by using a partitioning method based on a dual-stage offset, equidistant sampling is performed in a new window to obtain a new segment having the same length as the initial segment.

[0085] Step 25: Input the normalized low-frequency component into a multi-layer perceptron to extract features.

[0086] Step 26: Input the features outputted from step 24 and step 25 into respective multi-layer perceptrons to obtain prediction results.

[0087] Step 27: According to the two sets of statistical information obtained in step 23, the two sets of prediction results output in step 26 are respectively multiplied by the standard deviation and added to the mean to complete the denormalization, and finally the two parts are added to obtain the final prediction result.

[0088] The statistical information used in the denormalization process is the same as that used in the normalization process, and denormalization is completed by multiplying the sequence by the standard deviation and adding the mean.

[0089] Step 3: Use the training set data to iteratively train the initial model, improve the loss function and calculate the error between the prediction result and the label, back-propagate the gradient, and then update the model parameters to minimize the loss and obtain the ideal model. The loss function used for model training consists of two parts: MAE loss function and direction alignment loss function based on relative position weighting.

[0090] The design of the relative position weighted direction alignment loss function is as follows:

[0091]

[0092] in represents the directional similarity between the fragments generated by the first offset, Represents the directional similarity between the final obtained fragments.

[0093] Sim rel The calculation formula is as follows:

[0094]

[0095] where p i and p j Represent two different sequence fragments, It represents the cosine similarity between two segments, that is, the alignment in direction. σ is defined as the standard deviation of the normal distribution, and μ is defined as the mean of the normal distribution.

[0096] Step 4: After the model training is completed, collect the key parameter data of the industrial process and preprocess the data. Then, input the preprocessed data into the trained model for long-term prediction.

[0097] As shown in Table 1, the performance of the method of the present invention is compared with eight deep learning methods in terms of MSE, MAE and MAPE.

[0098] Table 1

[0099]

[0100] The embodiments described in the above description will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

Claims

1. A long-term prediction method for industrial key parameters based on double-offset segment partitioning and adaptive normalization, characterized in that: The following steps are performed to predict target parameters through historical data, so as to provide support for optimal control of industrial processes with time-lag characteristics; the method comprises the following steps: Initialize the long-term prediction model of industrial key parameters based on double-offset segment partitioning and adaptive normalization, collect historical key parameter data, and make a training set after data preprocessing; use the training set data to iteratively train the initial model, improve the loss function and calculate the error between the prediction result and the label, back-propagate the gradient, and then update the model parameters to minimize the loss to obtain the ideal model; After the model training is completed, the historical key parameter data corresponding to the industrial process target to be predicted is collected and the data is preprocessed; the trained ideal model is input for long-term prediction.

2. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 1 is characterized in that: The key parameter data of industrial processes include material supply or inventory data related to the production process, control parameter data of electrical equipment that performs actions, and feedback data of sensors; the training set uses the future value of a target parameter in the process production as the key parameter data label to be predicted, and observes the target controlled quantity through other key parameter data and its own historical data to achieve learning, thereby predicting its future long-term time series data. The training set data of this application is a training set made for the key parameters in the loose rehumidification process of food processing {historical values ​​of 24 key parameters, and the future value of the outlet moisture parameter as the target parameter label}.

3. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 1 is characterized in that: The establishment of a long-term prediction model for key industrial parameters based on dual-offset segment partitioning and adaptive normalization includes: 1) Multi-band decoupling to obtain mid-frequency and low-frequency components; 2) Based on the adaptive working condition division, the low-frequency component and the medium-frequency component are normalized respectively; 3) For the normalized intermediate frequency components, the semantically complete sequence segments are obtained based on the two-stage offset partitioning method and are sequentially input into the multi-layer perceptron, convolutional layer and multi-layer perceptron for feature extraction between segments, within segments and between variables; 4) For the normalized low-frequency components, a multi-layer perceptron is used to extract features; 5) Input the features extracted from the intermediate frequency and low frequency branches into two multi-layer perceptrons respectively to obtain the prediction results and perform denormalization; 6) Add the output results of the two branches to obtain the long-term prediction results of the target parameters.

4. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 3 is characterized in that: Multi-band decoupling involves converting the preprocessed data into the frequency domain using fast Fourier transform and decomposing it into high-frequency, medium-frequency and low-frequency components; deleting the high-frequency component and converting the low-frequency and medium-frequency components back to the time domain through inverse Fourier transform.

5. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 3 is characterized in that: The normalization of low-frequency components based on adaptive working condition division includes the following steps: Split the process data into multiple segments using windows of fixed size and moving step size; The mean of each segment is calculated for the low-frequency component; the variance of each segment is calculated for the mid-frequency component; Perform first-order difference processing on the obtained mean sequence or variance sequence and calculate the absolute value, and then identify the indexes corresponding to the largest k values ​​in the sequence; Keep the largest value among the k indexes and delete the fragment before that index; Expand the retained segments and calculate the mean and standard deviation of the retained ensemble; The low frequency component sequence or the intermediate frequency component sequence is subtracted from the mean and divided by the standard deviation to obtain normalized results of the low frequency component or the intermediate frequency component, respectively.

6. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 5 is characterized in that: When the normalization method based on adaptive working condition division is applied to low-frequency components and medium-frequency components, the size of the segment division window and the moving step length are kept equal.

7. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to any one of claims 3 to 6, characterized in that: The steps of obtaining semantically complete sequence segments using the dual-stage offset-based segmentation method are as follows: Use windows of fixed size and moving step length to obtain sequence segments of equal length; Get the three key parameters: the index of the center position of each segment division window in the entire sequence, the distance from the left boundary to the center position, and the distance from the right boundary to the center position; The multi-layer perceptron network is used to learn the offsets of the above three key parameters, and the original window endpoint parameters are adjusted accordingly to obtain the position of the new segment division window; Perform equidistant sampling within the above window to generate new fragments; The index of each point in the new segment relative to the segment itself is input into the multi-layer perceptron for adaptive offset to obtain the final sequence segment with complete semantics and aligned patterns between segments.

8. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 7 is characterized in that: In the step of obtaining semantically complete sequence segments using the dual-stage offset-based segmentation method, the size of the segment segmentation window is maintained at twice the moving step length; and new segments equal in length to the initial segments are obtained by performing equidistant sampling in the new window.

9. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 1, characterized in that: The loss function used for model training consists of two parts: MAE loss function and direction alignment loss function weighted by relative position; The design of the relative position weighted direction alignment loss function is as follows: in represents the directional similarity between the fragments generated by the first offset, Represents the directional similarity between the finally obtained fragments; Sim rel The calculation formula is as follows: where p i and p j Represent two different sequence fragments, It represents the cosine similarity between two segments, that is, the alignment in direction.

10. The method for long-term prediction of industrial key parameters based on double-offset segment division and adaptive normalization according to claim 1, characterized in that: The statistical information used in the denormalization process is the same as that used in the normalization process, and the denormalization is completed by multiplying the sequence by the standard deviation and adding the mean.

Images