Working condition transfer learning-based yield interval prediction method
By employing working condition transfer learning and information granularization, a network structure with a composite loss function is constructed. This addresses the issues of accuracy and stability in yield prediction under abnormal working conditions, enabling efficient yield interval prediction under such conditions and enhancing the model's adaptability and predictive capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2025-12-15
- Publication Date
- 2026-05-08
AI Technical Summary
Existing yield prediction methods are unable to meet production needs in terms of adaptability and prediction accuracy under abnormal operating conditions, especially when the number of samples is limited, the model training is insufficient and the generalization ability is weak.
A working condition-based transfer learning approach is adopted. By acquiring historical production time-series data with working condition labels, a source domain base model is constructed and transferred to the target domain. Combined with information granular processing, composite loss function and improved network structure, a target domain yield interval prediction model is trained, and the working condition category is perceived by sensors on the production site for prediction.
It improves the accuracy and stability of yield range prediction under abnormal operating conditions, overcomes the insufficient generalization ability of traditional methods under conditions with few samples, and provides intelligent monitoring and optimization support for complex industrial processes.
Smart Images
Figure CN121997218A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial data processing, specifically relating to a method for predicting yield intervals based on working condition transfer learning. Background Technology
[0002] Yield is one of the core indicators for measuring the stability, economy, and product quality of the iron ore sintering process. Predicting yield trends in advance not only guides operators to adjust process parameters promptly but also provides crucial information for equipment operation and batching scheduling.
[0003] The sintering process is characterized by strong coupling, strong nonlinearity, and multiple operating conditions, with significant differences in data distribution between different conditions. Normal operating conditions typically create a relatively stable production environment, resulting in a high yield of sintered ore. However, the yield under abnormal operating conditions is highly volatile and low. Furthermore, the limited sample size under abnormal conditions makes independent modeling for these conditions prone to insufficient training and weak generalization ability, leading to significant prediction bias. Transfer learning provides an effective technical approach to address this issue. By transferring model knowledge trained under normal operating conditions to abnormal operating conditions, the predictive ability of the model under abnormal conditions can be improved while reducing the need for training data.
[0004] Therefore, in the complex sintering environment with variable operating conditions, existing yield prediction methods are insufficient to meet production needs in terms of adaptability and prediction accuracy under abnormal conditions. To address this, a yield interval prediction method based on operating condition transfer learning is invented, which is of great significance for improving the quality and yield of sintered ore. Summary of the Invention
[0005] This invention aims to address the technical problem that existing yield prediction methods are insufficient in terms of adaptability and prediction accuracy under abnormal operating conditions to meet production needs. To solve this problem, this invention provides a yield interval prediction method based on operating condition transfer learning.
[0006] This invention provides a method for predicting yield intervals based on working condition transfer learning, specifically including the following steps: S1: Obtain historical production time series data with working condition labels, use normal working condition data as source domain data and abnormal working condition data as target domain data, and train the source domain base model for yield interval prediction based on the source domain data. S2: Adaptively transfer the knowledge of the source domain base model to the target domain data to obtain the target domain yield interval prediction model; S3: In the prediction phase, use time-series data from sensors on the production site to perceive the operating condition category of the sample to be predicted; S4: Based on the working condition perception results of the sample to be predicted, if it is an abnormal working condition, the target domain interval prediction model obtained through transfer learning is called to predict the yield interval.
[0007] A computer device includes at least: one or more processors; and a memory storing one or more computer programs; wherein the processors invoke the computer programs to implement the steps of the yield interval prediction method based on working condition transfer learning.
[0008] A computer storage device stores a computer program that is invoked by a processor to implement the steps of the yield interval prediction method based on working condition transfer learning.
[0009] The technical solution provided by this invention has the following beneficial effects: By granularizing time-series data, designing a composite loss function, and improving the network structure, this invention effectively improves the accuracy and stability of yield interval prediction under abnormal operating conditions, overcomes the problem of weak generalization ability of traditional methods under conditions with few samples, and provides reliable support for intelligent monitoring and optimization of complex industrial processes. Attached Figure Description
[0010] The present invention will be further described below with reference to the accompanying drawings and examples. In the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of a yield interval prediction method based on working condition transfer learning according to the present invention. Figure 2 This is the predicted yield range under abnormal operating conditions. Detailed Implementation
[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.
[0012] Example 1 Please refer to Figure 1 This invention provides a method for predicting yield intervals based on working condition transfer learning, the main steps of which are as follows: S1: Obtain historical production time series data with working condition labels, use normal working condition data as source domain data and abnormal working condition data as target domain data, and train the source domain base model for yield interval prediction based on the source domain data. It should be noted that step S1 is as follows: S11: For historical production time series data with working condition labels, perform information granulation processing according to a fixed sampling window, construct the lower limit and upper limit of the information granules in each window, and form interval information granules; Specifically, historical production time-series data with operating condition tags is acquired, and the time-series data is granulated according to a fixed sampling window, constructing information granules within each window. lower limit and upper limit . and The selection must ensure that the interval it forms To ensure the representativeness of the data within the window, the information granules should cover as many original sampling points as possible; at the same time, the interval width should be as small as possible to reduce redundant data. The interval information granules are used as the input and output objects for training the subsequent interval prediction model.
[0013] S12: Divide the source and target datasets according to the operating condition labels. Normal operating condition data is the source domain, and overheated and underheated operating condition data is the target domain. Specifically, the source domain and target domain datasets are divided based on the working condition labels. The data corresponding to the normal working condition is divided into the source domain dataset, and the data corresponding to the overheated and underheated working conditions are divided into the target domain dataset; and yield prediction samples are constructed for each.
[0014] S13: Normalize the source and target domain data respectively; Specifically, Min-Max normalization is performed on the data in both the source and target domains. The normalized data... It can be represented as:
[0015] in, It is the raw data. and These are the minimum and maximum values in the dataset.
[0016] S14: Construct a base model for predicting the yield interval using source domain data based on a gated temporal attention fusion network. The input is a sequence of historical multi-granular information particles, and the output is the lower or upper limit of the yield of the next particle. Save the model parameters.
[0017] Specifically, a base model for predicting yield intervals is constructed using source domain data and a gated temporal attention fusion network. The input is a sequence of historical multi-granular information particles, and the output is the lower or upper limit of the yield for the next particle. After training, the base model parameters are saved for subsequent transfer learning.
[0018] S2: Adaptively transfer the knowledge of the source domain base model to the target domain data to obtain the target domain yield interval prediction model; It should be noted that step S2 is as follows: S21: Use the parameters of the source domain base model as the initialization parameters of the target domain model; Specifically, the model parameters of the source domain gated temporal attention fusion network model trained in step S1 are used as the initial parameters of the target domain model. Transfer learning is used to preserve the knowledge structure of normal operating conditions, enabling the target domain model to extract features under normal operating conditions from the early stages of training. Let the source domain model parameter set be... Then the initial parameter set of the target domain model is:
[0019] in, Represents the set of parameters for the target domain model, indicated by the superscript. This indicates the initialization phase of the target domain model.
[0020] S22: On the target domain samples, supervised learning is used to fine-tune the model, and a composite loss function is used for optimization. Specifically, supervised learning is used to fine-tune the model on the target domain samples. During fine-tuning, the composite loss function is a weighted average of the interval coverage loss and the interval width loss, expressed as:
[0021] in:
[0022]
[0023] and These are the lower and upper limits of the prediction interval, respectively. and These are adjustable weighting coefficients; N The total number of samples.
[0024] S23: Save the trained target domain model and its corresponding normalizer.
[0025] Specifically, the target domain model and its corresponding normalizer are saved independently for later use.
[0026] S3: In the prediction phase, use time-series data from sensors on the production site to perceive the operating condition category of the sample to be predicted; It should be noted that step S3 specifically includes: S31: Construct a gated temporal attention fusion network as a working condition perception model. The network includes a temporal feature extraction layer, a temporal attention fusion layer, and a working condition classification layer. The structure of the gated temporal attention fusion network includes: Temporal feature extraction layer: Multiple parallel temporal convolutional modules with different kernel sizes are used to extract multi-scale local temporal features; Temporal attention fusion layer: Receives the multi-scale temporal features, calculates the feature weights of each time step through a gating attention mechanism, and performs dynamic weighted fusion; The gated attention mechanism controls attention weights through learnable gating units, expressed as follows:
[0027] in For the k-th time series feature, For trainable gated weight matrix, For the activation function, the output is... This is the fusion weight vector for each feature; Operating condition classification layer: The fused features are input into the fully connected network, and the probability distribution of the operating condition category is output.
[0028] S32: In the prediction phase, the sensor time series data is input into the working condition perception model, and the category probability distribution of future working conditions is output. S33: Take the category with the highest probability as the result of working condition perception.
[0029] S4: Based on the working condition perception results of the sample to be predicted, if it is an abnormal working condition, the target domain interval prediction model obtained through transfer learning is called to predict the yield interval.
[0030] Specifically, step S4 is as follows: S41: If the working condition perception result is an abnormal working condition, call the target domain interval prediction model and its corresponding normalizer. S42: Normalize the input information particle sequence, input the lower limit prediction model and upper limit prediction model of the target domain respectively, and obtain the two endpoint prediction values of the next particle yield interval; S43: Take the minimum value of the two endpoints as the lower limit and the maximum value as the upper limit to form the yield prediction interval.
[0031] Example 2 A yield interval prediction method based on working condition transfer learning is proposed. First, the time-series data of the sintering process is granulated to obtain interval information granules composed of upper and lower limits. Then, a source domain interval prediction model is constructed using normal working condition data, and transfer learning is used to transfer knowledge from the source domain to abnormal working conditions to form a target domain model. Finally, when the working condition perception result indicates an abnormal working condition, the target domain interval prediction model is invoked to predict the yield interval. The specific steps are as follows: (1) Collect historical data of sintering production to obtain raw sample data Historical sintering production data is stored in the local database of the industrial control computer in the control room in the form of daily reports. Based on the data in the daily reports, half a month's worth of historical data such as sintering endpoint, exhaust gas temperature of the air box, negative pressure of the air box, material layer thickness, trolley speed and yield are collected to form the original sample data.
[0032] (2) Data preprocessing The original time series data was sampled at a 1-minute interval and sliced into 10-minute windows. The information granules of each window were then calculated.
[0033] (3) Construction of the yield interval prediction model The training set data is divided according to working conditions. The normal working condition samples are used as the source domain to construct a yield interval prediction base model based on gated recurrent units. The weights of the source domain model are used as the initial parameters of the target domain model and loaded onto the target domain data for fine-tuning training. After fine-tuning, the target domain interval prediction model is obtained.
[0034] (4) Construction of working condition perception model A condition perception model based on a gated temporal attention fusion network is established using time-series data. During the testing phase, time-series data within the time period to be predicted is input into the condition perception model to obtain the condition perception results.
[0035] (5) Prediction of finished product yield range When the working condition perception result is abnormal, the target domain model is invoked. Historical sequences are input into the lower bound prediction model and the upper bound prediction model respectively to obtain the lower and upper bound predicted values of the next yield. These values are then sorted and combined to form a prediction interval, resulting in the final yield interval output. The yield interval prediction result under abnormal working conditions is as follows: Figure 2 As shown.
[0036] The results of this embodiment show that the present invention can utilize knowledge of normal operating conditions to predict the yield under abnormal operating conditions, and has a good accuracy rate.
[0037] Example 3 A computer device includes at least: one or more processors; and a memory storing one or more computer programs; wherein the processors invoke the computer programs to implement the steps of the yield interval prediction method based on working condition transfer learning.
[0038] A computer storage device stores a computer program that is invoked by a processor to implement the steps of the yield interval prediction method based on working condition transfer learning.
[0039] The preferred embodiments of the present invention disclosed above are only for the purpose of illustrating the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation described herein. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.
Claims
1. A method for predicting yield intervals based on working condition transfer learning, characterized in that, Includes the following steps: S1: Obtain historical production time series data with working condition labels, use normal working condition data as source domain data and abnormal working condition data as target domain data, and train the source domain base model for yield interval prediction based on the source domain data. S2: Adaptively transfer the knowledge of the source domain base model to the target domain data to obtain the target domain yield interval prediction model; S3: In the prediction phase, use time-series data from sensors on the production site to perceive the operating condition category of the sample to be predicted; S4: Based on the working condition perception results of the sample to be predicted, if it is an abnormal working condition, the target domain interval prediction model obtained through transfer learning is called to predict the yield interval.
2. The yield interval prediction method based on working condition transfer learning according to claim 1, characterized in that, Step S1 is as follows: S11: For historical production time series data with working condition labels, perform information granulation processing according to a fixed sampling window, construct the lower limit and upper limit of the information granules in each window, and form interval information granules; S12: Divide the source and target datasets according to the operating condition labels. Normal operating condition data is the source domain, and overheated and underheated operating condition data is the target domain. S13: Normalize the source and target domain data respectively; S14: Construct a base model for predicting the yield interval using source domain data based on a gated temporal attention fusion network. The input is a sequence of historical multi-granular information particles, and the output is the lower or upper limit of the yield of the next particle. Save the model parameters.
3. The yield interval prediction method based on working condition transfer learning according to claim 1, characterized in that, Step S2 is as follows: S21: Use the parameters of the source domain base model as the initialization parameters of the target domain model; S22: On the target domain samples, supervised learning is used to fine-tune the model, and a composite loss function is used for optimization. S23: Save the trained target domain model and its corresponding normalizer.
4. The yield interval prediction method based on working condition transfer learning according to claim 1, characterized in that, In step S22, the composite loss function is composed of a weighted average of the interval coverage loss and the interval width loss, and its expression is: in: and These are the lower and upper limits of the prediction interval, respectively. and These are adjustable weighting coefficients; N The total number of samples.
5. The yield interval prediction method based on working condition transfer learning according to claim 1, characterized in that, Step S3 specifically includes: S31: Construct a gated temporal attention fusion network as a working condition perception model. The network includes a temporal feature extraction layer, a temporal attention fusion layer, and a working condition classification layer. S32: In the prediction phase, the sensor time series data is input into the working condition perception model, and the category probability distribution of future working conditions is output. S33: Take the category with the highest probability as the result of working condition perception.
6. The yield interval prediction method based on working condition transfer learning as described in claim 5, characterized in that, The structure of the gated temporal attention fusion network includes: Temporal feature extraction layer: Multiple parallel temporal convolutional modules with different kernel sizes are used to extract multi-scale local temporal features; Temporal attention fusion layer: Receives the multi-scale temporal features, calculates the feature weights of each time step through a gating attention mechanism, and performs dynamic weighted fusion; The gated attention mechanism controls attention weights through learnable gating units, expressed as follows: in For the k-th time series feature, For trainable gated weight matrix, For the activation function, the output is... This is the fusion weight vector for each feature; Operating condition classification layer: The fused features are input into the fully connected network, and the probability distribution of the operating condition category is output.
7. The yield interval prediction method based on working condition transfer learning as described in claim 1, characterized in that, Step S4 is as follows: S41: If the working condition perception result is an abnormal working condition, call the target domain interval prediction model and its corresponding normalizer. S42: Normalize the input information particle sequence, input the lower limit prediction model and upper limit prediction model of the target domain respectively, and obtain the two endpoint prediction values of the next particle yield interval; S43: Take the minimum value of the two endpoints as the lower limit and the maximum value as the upper limit to form the yield prediction interval.
8. A computer device, characterized in that, It includes at least: one or more processors; a memory storing one or more computer programs; wherein the processor calls the computer programs to implement the steps of the yield interval prediction method based on working condition transfer learning as described in any one of claims 1-7.
9. A computer storage device, characterized in that, A computer program is stored, which is invoked by a processor to implement the steps of the yield interval prediction method based on working condition transfer learning as described in any one of claims 1-7.