Long-term prediction method of industrial key parameters based on double-offset segment division and adaptive normalization

Abstract

The present application is directed to the optimization control requirements of industrial processes with time delay characteristics, and proposes an industrial key parameter long-time prediction method based on double offset segment division and adaptive normalization. In the prediction, the pretreated industrial data is input into the trained model, and the model first decouples and denoises in multiple frequency bands. Then, the low-frequency and medium-frequency components are modeled in branch networks respectively. The low-frequency and medium-frequency components are normalized based on adaptive working condition division respectively. The low-frequency component is feature extracted by multilayer perceptron, and the medium-frequency component is feature extracted by a segment-level modeling method based on two-stage offset. Subsequently, the two features are input into multilayer perceptron to obtain the prediction results of each branch, and are added after inverse normalization to generate the final overall prediction result. In addition, on the basis of using the MAE loss function, the present application introduces a direction alignment loss function based on relative position weighting as a new item to enhance the model's attention to specific layers during the training process. The present application significantly improves the long-time prediction accuracy of industrial key parameters, and provides technical support for the optimization control of industrial processes with time delay characteristics.

Description

Technical Field

[0001] This invention belongs to the field of industrial time series forecasting, specifically a long-term forecasting method for key industrial parameters based on dual-offset segmentation and adaptive normalization. Background Technology

[0002] Significant time lags are prevalent in modern industry. Specifically, due to time delays in the transmission of materials, energy, or information within industrial systems, changes in detected key parameters typically lag behind actual production dynamics. This time lag makes it difficult to track changes in industrial processes in a timely manner. By predicting the changing trends of key parameters over a longer future time step, optimal adjustment plans can be developed in advance to compensate for this time lag. This allows for timely adjustments to control parameters, ensuring that the industrial process remains in optimal condition and preventing resource waste and decreased production efficiency.

[0003] Long-term prediction tasks require capturing potential long-term patterns using vast amounts of historical data, necessitating sophisticated feature extraction capabilities from the model. This poses a challenge to traditional statistical learning and machine learning methods. Deep learning methods, due to their excellent fitting performance to nonlinear features, have become the mainstream approach. Currently, most deep learning methods use single sample values ​​as the basic input unit, which can easily distract from key information when faced with a large amount of input. In contrast, using sequence fragments as the basic input unit has two advantages: firstly, it can significantly reduce the number of input units while maintaining the same length of historical data; secondly, it mitigates the interference of random noise. However, existing methods using sequence fragments as the basic input unit mostly employ fixed window partitioning and window movement steps. This rigid partitioning method may destroy the complete semantic information in industrial process data, leading to the capture of incorrect patterns and hindering accurate prediction results. Furthermore, existing long-term prediction methods for key parameters of industrial processes face numerous challenges in dealing with the complex nonlinearity, information sparsity, inaccuracy, and non-stationarity of industrial process data. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a long-term prediction method for key industrial parameters based on dual-offset segment partitioning and adaptive normalization. During prediction, preprocessed industrial data is input into a trained model. The model first decomposes the data in the frequency domain into low-frequency, mid-frequency, and high-frequency components, removing the high-frequency components to reduce industrial noise interference. Then, the low-frequency and mid-frequency components are converted back to the time domain and input into two independent branch networks for modeling, reducing model complexity. To address the non-stationarity problem caused by fluctuating operating conditions, the model performs normalization processing on the low-frequency and mid-frequency components based on adaptive operating condition partitioning. Low-frequency components are feature-extracted using a multilayer perceptron, while mid-frequency components are feature-extracted using a segment-level modeling method based on dual-stage offset. Subsequently, the two feature sets are input into the multilayer perceptron to obtain the prediction results for each branch, and after inverse normalization, they are summed to generate the final overall prediction result. Furthermore, this invention optimizes the loss function used for model training. Based on the MAE loss function, a direction alignment loss function weighted by relative position is introduced as a new term to enhance the model's focus on specific layers during training. This invention significantly improves the long-term prediction accuracy of key industrial parameters, providing technical support for the optimized control of industrial processes with time-delay characteristics.

[0005] The specific technical solution adopted by this invention to achieve the above objectives is as follows: a long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization, which performs the following steps to predict target parameters using historical data, providing support for the optimal control of time-delay industrial processes; the method includes the following steps:

[0006] A long-term prediction model for key industrial parameters based on dual-offset segmentation and adaptive normalization is initialized. Historical key parameter data is collected and a training set is created after data preprocessing. The initial model is iteratively trained using the training set data. The loss function is improved and the error between the prediction result and the label is calculated. The gradient is calculated by backpropagation, and the model parameters are updated to minimize the loss, thus obtaining the ideal model.

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

[0008] Key parameter data for industrial processes include material supply or inventory data related to the production process, control parameter data of electrical equipment performing actions, and sensor feedback data. The training set uses the future value of a target parameter in the production process as the label for the key parameter data to be predicted. Learning is achieved by observing the target controlled variable through other key parameter data and its own historical data, thereby predicting its future long-term data. In this application, the training set data is a training set created for key parameters in the loose rehydration process of food processing {historical values ​​of 24 key parameters, with 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 segmentation and adaptive normalization includes:

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

[0011] 2) Based on adaptive operating condition division, the low-frequency components and the mid-frequency components are normalized respectively;

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

[0013] 4) For the normalized low-frequency components, a multilayer perceptron is used for feature extraction;

[0014] 5) Input the features extracted from the mid-frequency and low-frequency branches into two multilayer perceptrons respectively to obtain the prediction results, and perform inverse normalization processing;

[0015] 6) Add the outputs of the two branches to obtain the long-term prediction result of the target parameter.

[0016] Multi-band decoupling involves using Fast Fourier Transform to convert the preprocessed data to the frequency domain and decompose it into high-frequency, mid-frequency, and low-frequency components; deleting the high-frequency components; and converting the low-frequency and mid-frequency components back to the time domain using Inverse Fourier Transform.

[0017] Low-frequency component normalization based on adaptive operating condition partitioning includes the following steps:

[0018] The process data is divided into multiple segments using windows of fixed size and movement step size;

[0019] Calculate the mean of each segment for the low-frequency components; calculate the variance of each segment for the mid-frequency components;

[0020] Perform first-order differencing on the obtained mean or variance sequence and calculate the absolute value. Then identify the indices corresponding to the k largest values ​​in the sequence.

[0021] Keep the largest value among the k indices and delete the fragments preceding that index;

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

[0023] Subtract the mean from the low-frequency component sequence or the mid-frequency component sequence and divide by the standard deviation to obtain the normalized results of the low-frequency component or the mid-frequency component, respectively.

[0024] The normalization method based on adaptive operating condition partitioning maintains the same segment partitioning window size and moving step size when applied to both low-frequency and mid-frequency components.

[0025] The steps for obtaining semantically complete sequence segments using the two-stage offset-based partitioning method are as follows:

[0026] Obtain sequence segments of equal length using a window of fixed size and moving step size;

[0027] Obtain the three key parameters: the index of the center position of each segment partition 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 offsets of the three key parameters mentioned above are learned using a multilayer perceptron network, and the original window endpoint parameters are adjusted accordingly to obtain the position of the new segmentation window.

[0029] Perform isometric sampling within the aforementioned window to generate new fragments;

[0030] The index of each point in the new segment relative to the segment itself is input into a multilayer perceptron for adaptive offset to obtain the final semantically complete sequence segment with pattern alignment between segments.

[0031] In the step of obtaining semantically complete sequence segments using a two-stage offset-based partitioning method, the size of the segment partitioning window is maintained at twice the moving step size; equidistant sampling is performed within the new window to obtain a new segment with the same length as the initial segment.

[0032] The loss function used for model training consists of two parts: the MAE loss function and the orientation alignment loss function based on relative position weighting;

[0033] The design of the orientation alignment loss function based on relative position weighting is as follows:

[0034]

[0035] in This represents the directional similarity between the fragments generated by the first offset. This represents the directional similarity between the ultimately obtained fragments;

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

[0037]

[0038] Where p i and p j These represent two different sequence segments. The cosine similarity between two segments represents the alignment in direction.

[0039] The statistical information used in the denormalization process is the same as that used in the normalization process. Denormalization is accomplished 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. This invention provides a long-term prediction method for key industrial parameters based on dual-offset segment partitioning and adaptive normalization. To address the timely optimization and control requirements of industrial processes with time-delay characteristics, it employs a multi-band decoupled network structure, a reversible normalization method based on adaptive operating condition partitioning, a segment-level modeling method based on dual-stage offset, and a direction alignment loss function based on relative position weighting. This method possesses the capability for long-term prediction of key industrial parameters.

[0042] 2. This invention proposes a long-term prediction method for key parameters to meet the optimization control requirements of industrial processes with time delay characteristics, effectively addressing the complex nonlinearity, information sparsity, inaccuracy, and nonstationarity of industrial process data. Attached Figure Description

[0043] Figure 1 This refers to the loose rehydration process in the food processing industry.

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

[0045] Figure 3 This is a flowchart of the method of the present invention;

[0046] Figure 4 This is a diagram illustrating the segmentation. Detailed Implementation

[0047] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough 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 modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments 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 serves as an indicator for production control, characterizing whether the current industrial control process meets the set ideal state. This invention uses historical data (including material supply or inventory data related to the loosening and rehydration process, control parameter data of the electrical equipment executing the actions, and sensor feedback data) to train a model. During actual prediction, the historical data corresponding to the target parameter to be predicted is input into the model in advance, and the model automatically outputs the target parameter to be predicted: the moisture content of the material at the drum outlet. This achieves the purpose of simulating and predicting future target data based on historical data.

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

[0050] like Figure 3 The diagram shows a flowchart of the prediction process of the method of the present invention. This invention provides a long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization. The programming language used in the program execution steps of the present invention is not limited to Python, etc.

[0051] This invention discloses a long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization. This method utilizes historical data to predict target parameters through the following steps, providing support for the optimal control of time-delay industrial processes. The specific steps are as follows:

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

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

[0054] Step 1-1: Remove missing values;

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

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

[0057] Steps 1-4: Create a training set. The moisture content of the material at the drum outlet is the target to be predicted. Learning is achieved by observing the controlled variable of the target through other key parameter data and historical data of the target itself, thereby predicting its future long-term data. The training set data in this application includes key parameters in the loosening and rehydration process of food processing {24 key parameters, with the future value of the outlet moisture parameter used as the target parameter label}.

[0058] Step 2: Establish a long-term prediction model for key industrial parameters based on dual-offset segmentation and adaptive normalization. Iteratively train the initial model using training set data, improve the loss function and calculate the error between the prediction result and the label, backpropagate to calculate the gradient, and then update the model parameters to minimize the loss and obtain the ideal model.

[0059] Step 21: Decouple the preprocessed input data using multiple frequencies. Use Fourier transform to convert it to the frequency domain, decomposing it into low-frequency, mid-frequency, and high-frequency components.

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

[0061] Step 23: Perform normalization processing on the low-frequency components and the mid-frequency components based on adaptive operating condition division.

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

[0063] Step 23-1, the low-frequency component normalization based on adaptive operating condition partitioning includes the following steps:

[0064] The process data is divided into multiple segments using windows of fixed size and moving step size, and the mean of each segment is calculated.

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

[0066] Keep the largest value among the k indices and delete the fragments preceding that index;

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

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

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

[0070] The process data is divided into multiple segments using windows of fixed size and moving step size, and the variance of each segment is calculated.

[0071] The obtained variance sequence is subjected to first-order differencing and the absolute value is calculated. Then, the indices corresponding to the k largest values ​​in the sequence are identified.

[0072] Keep the largest value among the k indices and delete the fragments preceding that index;

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

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

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

[0076] Step 24: For the normalized mid-frequency components, use a two-stage offset-based segmentation method to obtain semantically complete sequence segments, and then sequentially input them into a multilayer perceptron, a convolutional layer, and another multilayer perceptron for feature extraction between segments, within segments, and between variables. For example... Figure 4 The diagram shown is a schematic representation of the segmentation process from steps 24-1 to 24-5.

[0077] Step 24-1: Obtain sequence segments of equal length using a window of fixed size and moving step size;

[0078] Step 24-2: Obtain the three key parameters: the index of the center position of each segmentation 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: Use a multilayer perceptron to learn the offsets of the above three key parameters, and adjust the original parameters accordingly to obtain the position of the new segmentation window;

[0080] Step 24-4: Perform equidistant sampling within the above window to generate new segments;

[0081] Steps 24-5: Input the index of each point in the new segment relative to the segment itself into a multilayer perceptron for adaptive offset to obtain the final semantically complete sequence segment with pattern alignment between segments;

[0082] Steps 24-6: Input the sequence with segments as the basic input units into the multilayer perceptron, convolutional layer and multilayer perceptron in sequence to extract features between segments, within segments and between variables.

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

[0084] In the step of obtaining semantically complete sequence segments using a two-stage offset-based partitioning method, performing equidistant sampling within a new window yields a new segment with the same length as the initial segment.

[0085] Step 25: Input the normalized low-frequency components into the multilayer perceptron to extract features.

[0086] Step 26: Input the features output from Step 24 and Step 25 into their respective multilayer perceptrons to obtain the prediction results.

[0087] Step 27: Based on the two sets of statistical information obtained in step 23, multiply the two sets of prediction results output in step 26 by the standard deviation and add the mean to complete the inverse normalization. Finally, add the two parts 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. Denormalization is accomplished by multiplying the sequence by the standard deviation and adding the mean.

[0089] Step 3: Iteratively train the initial model using the training set data, improve the loss function, calculate the error between the predicted result and the label, backpropagate to calculate the gradient, and then update the model parameters to minimize the loss, obtaining the ideal model. The loss function used for model training consists of two parts: the MAE loss function and a direction alignment loss function weighted based on relative position.

[0090] The design of the orientation alignment loss function based on relative position weighting is as follows:

[0091]

[0092] in This represents the directional similarity between the fragments generated by the first offset. This represents the directional similarity between the ultimately obtained fragments.

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

[0094]

[0095] Where p i and p j These represent two different sequence segments. Cosine similarity between two segments represents their 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 complete, collect 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] Table 1 shows a performance comparison of the method of this invention with eight other deep learning methods on three metrics: MSE, MAE, and MAPE.

[0098] Table 1

[0099]

[0100] The embodiments described above will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

Claims

1. A long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization, characterized in that: The following steps are performed to predict target parameters using historical data, providing support for the optimal control of time-delay industrial processes; the method includes the following steps: Initialize a long-term prediction model for industrial key parameters based on dual-offset segmentation and adaptive normalization. Collect historical key parameter data and create a training set after data preprocessing. Iteratively train the initial model using the training set data. Improve the loss function and calculate the error between the prediction result and the label. Backpropagate to calculate the gradient and update the model parameters to minimize the loss, obtaining the ideal model. The establishment of the long-term prediction model for industrial key parameters based on dual-offset segmentation and adaptive normalization includes: 1) Multi-band decoupling to obtain mid-frequency and low-frequency components; 2) Based on adaptive operating condition division, the low-frequency components and the mid-frequency components are normalized respectively; 3) For the normalized mid-frequency components, semantically complete sequence segments are obtained based on the two-stage offset partitioning method and then sequentially input into the multilayer perceptron, convolutional layer and multilayer perceptron for feature extraction between segments, within segments and between variables. 4) For the normalized low-frequency components, a multilayer perceptron is used for feature extraction; 5) Input the features extracted from the mid-frequency and low-frequency branches into two multilayer perceptrons respectively to obtain the prediction results, and perform inverse normalization processing; 6) Add the outputs of the two branches to obtain the long-term prediction result of the target parameter; After the model training is completed, historical key parameter data corresponding to the target of the industrial process to be predicted are collected and the data is preprocessed; then the data is input into the trained ideal model for long-term prediction.

2. The long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to claim 1, characterized in that, Key parameter data for industrial processes include material supply or inventory data related to the production process, control parameter data of electrical equipment performing actions, and sensor feedback data. The training set uses the future value of a target parameter in the process production as the label of the key parameter data to be predicted. It learns by observing the target controlled variable through other key parameter data and its own historical data, thereby predicting its future long-term data. The training set data in this application is a training set made for key parameters in the loose rehydration 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 long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to claim 1, characterized in that, Multi-band decoupling involves using Fast Fourier Transform to convert the preprocessed data to the frequency domain and decompose it into high-frequency, mid-frequency, and low-frequency components; deleting the high-frequency components; and converting the low-frequency and mid-frequency components back to the time domain using Inverse Fourier Transform.

4. The long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to claim 1, characterized in that, Low-frequency component normalization based on adaptive operating condition partitioning includes the following steps: The process data is divided into multiple segments using windows of fixed size and movement step size; Calculate the mean of each segment for the low-frequency components; calculate the variance of each segment for the mid-frequency components; Perform first-order differencing on the obtained mean or variance sequence and calculate the absolute value. Then identify the indices corresponding to the k largest values ​​in the sequence. Keep the largest value among the k indices and delete the fragments preceding that index; Expand the retained segments and calculate the mean and standard deviation of the retained whole; Subtract the mean from the low-frequency component sequence or the mid-frequency component sequence and divide by the standard deviation to obtain the normalized results of the low-frequency component or the mid-frequency component, respectively.

5. The long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to claim 4, characterized in that, The normalization method based on adaptive operating condition partitioning maintains the same segment partitioning window size and moving step size when applied to both low-frequency and mid-frequency components.

6. The long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to any one of claims 3-5, characterized in that, The steps for obtaining semantically complete sequence segments using the two-stage offset-based partitioning method are as follows: Obtain sequence segments of equal length using a window of fixed size and moving step size; Obtain the three key parameters: the index of the center position of each segment partition 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 offsets of the three key parameters mentioned above are learned using a multilayer perceptron network, and the original window endpoint parameters are adjusted accordingly to obtain the position of the new segmentation window. Perform isometric sampling within the aforementioned window to generate new fragments; The index of each point in the new segment relative to the segment itself is input into a multilayer perceptron for adaptive offset to obtain the final semantically complete sequence segment with pattern alignment between segments.

7. The long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to claim 6, characterized in that, In the step of obtaining semantically complete sequence segments using a two-stage offset-based partitioning method, the size of the segment partitioning window is maintained at twice the moving step size; equidistant sampling is performed within the new window to obtain a new segment with the same length as the initial segment.

8. The long-term prediction method for key industrial parameters based on dual-offset segmentation and adaptive normalization according to claim 1, characterized in that, The loss function used for model training consists of two parts: the MAE loss function and the orientation alignment loss function based on relative position weighting; The design of the orientation alignment loss function based on relative position weighting is as follows: ； in This represents the directional similarity between the fragments generated by the first offset. This represents the directional similarity between the ultimately obtained fragments; The calculation formula is as follows: ； in and These represent two different sequence segments. The cosine similarity between two segments is also known as directional alignment. Defined as the standard deviation of a normal distribution. It is defined as the mean of a normal distribution.

9. The long-term prediction method for key industrial parameters based on dual-offset segmentation 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. Denormalization is accomplished by multiplying the sequence by the standard deviation and adding the mean.

Images