Complex multi-working condition alcohol distillation soft measurement method and system based on neural network

Abstract

This invention provides a complex multi-condition alcohol distillation soft measurement method and system based on neural networks, belonging to the field of industrial process soft measurement technology. The method includes the following steps: acquiring time-series variable data of the multi-condition alcohol distillation process and preprocessing it; dividing the preprocessed time-series data into source condition datasets and target condition datasets according to the operating conditions; inputting the source condition dataset into a preset time-series structured echo network for pre-training to obtain a source condition pre-trained model; inputting the target condition dataset into the source condition pre-trained model and obtaining a target condition soft measurement model through transfer learning; acquiring real-time variable data under the operating condition to be predicted and inputting the real-time variable data into the target condition soft measurement model to obtain the target measurement variable prediction result; outputting and storing the target measurement variable prediction result to complete the alcohol distillation soft measurement. This invention effectively improves the adaptability and prediction accuracy of alcohol distillation soft measurement under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of soft measurement technology in industrial processes, and in particular to a soft measurement method and system for complex multi-condition alcohol distillation based on neural networks. Background Technology

[0002] With the continuous development of industrial automation, process control, and artificial intelligence technologies, soft sensing technology has been widely applied to continuous production processes such as chemical engineering, fermentation, biopharmaceuticals, and food processing. Among these, alcohol distillation, as a typical complex industrial process, typically uses key quality variables such as alcohol concentration as important evaluation indicators for product quality. However, due to the long detection cycle, high maintenance costs, and measurement lag of online alcohol concentration analysis equipment, establishing soft sensing models using easily measurable process variables such as temperature, pressure, and flow rate to achieve real-time prediction of key quality variables has become an important research direction for intelligent control of alcohol distillation processes. In recent years, with the development of deep learning technology, neural network models such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Transformers have been gradually applied to the field of industrial soft sensing. By mining the nonlinear mapping relationships and temporal dynamic characteristics in process data, the prediction accuracy of key quality variables in complex industrial processes has been improved, providing new technical means for online monitoring and optimized control of alcohol distillation processes.

[0003] However, existing neural network-based soft measurement methods for alcohol distillation still have significant limitations in complex operating conditions. First, the alcohol distillation process is characterized by multiple operating conditions, long time-delay coupling, periodic fluctuations, and non-stationary dynamics. The mapping relationship between process variables and alcohol concentration often changes significantly under different operating conditions. Existing models such as CNN, LSTM, and Transformer are typically trained based on a single operating condition or a fixed data distribution, making it difficult to effectively adapt to data distribution drift caused by changes in operating conditions, resulting in poor generalization ability across operating conditions. Second, traditional time series modeling methods mostly rely on fixed network structures or fixed time-series memory mechanisms, making it difficult to simultaneously account for short-term dynamic changes and long-term time-delay dependencies. They lack effective representation capabilities for complex periodic behaviors, phase shift phenomena, and critical state abrupt events in the distillation process, thus affecting prediction accuracy and stability. Furthermore, in actual industrial settings, some operating conditions are run infrequently, and the available historical sample data is limited. Existing methods typically require a large amount of labeled data to support model training. When the target operating condition sample is insufficient, problems such as overfitting, model transfer difficulties, and significant degradation in prediction performance can easily occur.

[0004] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a neural network-based method and system for complex multi-condition alcohol distillation soft measurement, which effectively improves the adaptability and prediction accuracy of soft measurement in complex operating conditions.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: This invention provides a complex multi-condition alcohol distillation soft measurement method based on neural networks, comprising the following steps: S1. Obtain variable time series data of alcohol distillation process under several working conditions, preprocess the variable time series data to obtain preprocessed time series data, and divide the preprocessed time series data into source working condition dataset and target working condition dataset according to the working conditions. S2. Input the source working condition dataset into a preset temporal structure echo network for pre-training to obtain the source working condition pre-trained model; the preset temporal structure echo network includes an input encoder unit, a phase encoder unit, an echo unit, a structure regulator unit, an event-driven multi-scale pooling unit, and a decoding unit. S3. Input the target working condition dataset into the source working condition pre-trained model, and obtain the target working condition soft measurement model through transfer learning; S4. Obtain real-time variable data under the working condition to be predicted, and input the real-time variable data into the target working condition soft measurement model to obtain the prediction result of the target measurement variable; S5. Output and store the prediction results of the target measurement variable to complete the soft measurement of alcohol distillation; The specific data flow direction of inputting the source operating condition dataset into the preset time-series echo network is as follows: The time-series data of variables in the source working condition dataset are encoded by the input encoder unit and the phase encoder unit respectively to obtain low-dimensional hidden state features and phase encoding information; The hidden state features are input into the echo unit for state update, resulting in the state feature matrix; The state feature matrix is ​​input into the structure regulator unit, which generates a sparse adjacency matrix. The state feature matrix is ​​then weighted and fused based on the sparse adjacency matrix to output the weighted fused state features. The weighted fusion state features are input into the event-driven multi-scale pooling unit. The event-driven multi-scale pooling unit identifies the event window and performs multi-scale pooling on the weighted fusion state features within the event window to obtain multi-scale pooled features. The multi-scale pooling features and phase-encoded information are input into the decoding unit. The decoding unit then fuses and maps the multi-scale pooling features and phase-encoded information to output the prediction results of the target measurement variables at future time moments.

[0007] Preferably, the preprocessing of the variable time series data includes normalization, outlier removal, and time series alignment. The step of dividing the preprocessed time-series data into source operating condition datasets and target operating condition datasets according to operating conditions includes: The preprocessed time-series data are classified according to the amount of historical sample data corresponding to each working condition. The working condition data with a sample data amount higher than the preset sample data amount threshold is classified as the source working condition dataset, and the working condition data with a sample data amount lower than the preset sample data amount threshold is classified as the target working condition dataset.

[0008] Preferably, the input encoder unit uses a linear transform layer combined with a nonlinear activation function to map the time-series variable data into low-dimensional hidden state features; The phase encoder unit introduces a learnable phase vector at each time step, and encodes the periodic information and phase change information of the time series through complex mapping or sine and cosine transform, and outputs phase encoded information. The echo unit consists of several parallel echo subunits, each containing a learnable delay parameter and a linear mapping; the state update mechanism of the echo subunit is defined as follows:

[0009] in, Indicates the first The mapping matrix corresponding to each echo unit This represents the learnable delay parameter. Represents local context information. Indicates the bias term. This represents a non-linear activation function.

[0010] Preferably, the step of generating a sparse adjacency matrix through the structure regulator unit includes: extracting statistical features from the time-series data of the variables, and generating a sparse adjacency matrix based on the statistical features using the Softmax function; the statistical features include variance, error energy, and prediction error; The event-driven multi-scale pooling unit identifies event windows in the time-series data of variables through an event detection mechanism, and performs multi-scale pooling on the weighted fusion state features based on the identified event windows; the event windows include mutation point windows, disturbance point windows, and state transition point windows; The decoding unit fuses the multi-scale pooling features corresponding to each time scale with the phase coding information, and establishes a mapping relationship between process variables and target measurement variables through a multilayer perceptron or residual network, and outputs the target measurement variable prediction results for future time moments.

[0011] Preferably, the pre-training step S2 employs a composite loss function, which includes the main prediction loss, the internal prediction error loss, and the structural sparsity regularization loss, and its expression is:

[0012] in, Forecast the loss. For internal prediction error loss, For structurally sparse regularized loss, and These represent the weighting coefficients of the loss term.

[0013] Preferably, the step of inputting the target working condition dataset into the source working condition pre-trained model and obtaining the target working condition soft measurement model through transfer learning includes: Load the network parameters of the pre-trained model under the source operating conditions; The parameters of the input encoder unit, phase encoder unit, echo unit, and event-driven multi-scale pooling unit are frozen. Using the target operating condition dataset, the structure regulator unit and the decoding unit are fine-tuned based on the master prediction loss to obtain the target operating condition soft measurement model.

[0014] This invention also provides a complex multi-condition alcohol distillation soft measurement system based on neural networks, comprising: The dataset partitioning module is used to acquire variable time-series data of the alcohol distillation process under several working conditions, preprocess the variable time-series data to obtain preprocessed time-series data, and partition the preprocessed time-series data into source working condition dataset and target working condition dataset according to the working conditions. The network pre-training module is used to input the source working condition dataset into a preset time-series structured echo network for pre-training to obtain the source working condition pre-trained model. The transfer learning module is used to input the target working condition dataset into the source working condition pre-trained model and obtain the target working condition soft measurement model through transfer learning. The prediction execution module is used to acquire real-time variable data under the working condition to be predicted, and input the real-time variable data into the target working condition soft measurement model to obtain the prediction result of the target measurement variable. The results output module is used to output and store the prediction results of the target measurement variable to complete the soft measurement of alcohol distillation.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described neural network-based complex multi-condition alcohol distillation soft measurement method.

[0016] The present invention also provides a computer storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described neural network-based complex multi-condition alcohol distillation soft measurement method.

[0017] The beneficial effects of this invention are: effectively improving the adaptability and prediction accuracy of soft measurement in alcohol distillation under complex operating conditions. The source operating condition dataset and the target operating condition dataset are divided according to the abundance of samples. A pre-trained model for the source operating condition is constructed using data from operating conditions with sufficient samples. Then, knowledge transfer to operating conditions with scarce samples is achieved through transfer learning, improving the model's generalization ability and modeling efficiency in multi-operating-condition scenarios. A temporal structured echo network is constructed, consisting of an input encoder unit, a phase encoder unit, an echo unit, a structure regulator unit, an event-driven multi-scale pooling unit, and a decoding unit. The phase encoder unit encodes periodic and phase change information in the time series by introducing learnable phase vectors, enabling the model to simultaneously learn the numerical and temporal evolution patterns of process variables, thereby enhancing the model's ability to represent complex periodic behaviors and phase shift phenomena in alcohol distillation. A multi-scale temporal memory structure is constructed through the echo unit. Multiple echo units with learnable delay parameters are used to model the correlation between latent state features at different times, enabling the model to capture the dynamic coupling relationships and long-term time-delay dependencies between process variables at different time scales, improving the prediction accuracy of complex dynamic processes. The structural regulator unit dynamically generates a sparse adjacency matrix based on the statistical characteristics of the input data. This matrix is ​​then used to weight and fuse the outputs of each echo unit, enabling adaptive adjustment of the network's internal connection structure. This allows the model to dynamically optimize information transmission paths based on changes in data distribution under different operating conditions, improving its robustness in complex and non-stationary environments. Event-driven multi-scale pooling units identify abrupt changes, disturbances, and state transitions in the time series and perform multi-scale feature aggregation within the corresponding event windows. This allows the model to simultaneously retain both short-term dynamic information and long-term trend information corresponding to key events, enhancing its responsiveness to changes in critical operating conditions. A composite loss function consisting of the master prediction loss, internal prediction error loss, and structural sparsity regularization loss is employed. By simultaneously constraining prediction accuracy, state evolution consistency, and network structural sparsity, this improves the model's stability and generalization ability during training, reduces the impact of redundant connections and invalid parameters on prediction results, and further enhances soft measurement accuracy under complex operating conditions. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the method steps of the present invention.

[0019] Figure 2 This is a system module diagram of the present invention.

[0020] Figure 3 Box plots showing the prediction error distribution of different models in Embodiment 3 of the present invention.

[0021] Figure 4 This is a schematic diagram comparing the prediction performance curves of different models in Embodiment 3 of the present invention.

[0022] Figure 5 This is a diagram of the internal structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation

[0023] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0024] Example 1: See Figure 1 As shown, this embodiment is a complex multi-condition alcohol distillation soft measurement method based on neural networks, including the following steps: S1. Obtain variable time series data of alcohol distillation process under several working conditions, preprocess the variable time series data to obtain preprocessed time series data, and divide the preprocessed time series data into source working condition dataset and target working condition dataset according to the working conditions. For step S1, the core is to obtain the time series data of process variables under multiple working conditions during the alcohol distillation process and perform multi-dimensional preprocessing and working condition classification, so as to build a high-quality, dimensionless and aligned data foundation for the subsequent training of the network architecture.

[0025] It should be noted that in this embodiment, time-series data of variables in the alcohol distillation process under several operating conditions are acquired, specifically by extracting historical and real-time data using a distributed control system (DCS) database. Due to the highly nonlinear and multi-condition fluctuation characteristics of the alcohol distillation industrial process, the acquired time-series data covers various typical operating conditions with significant process differences, including electric heating, steam heating, and mixed heating. Specifically, the time-series data includes key process monitoring variables reflecting the internal state of the distillation column and system, such as top temperature, bottom temperature, internal pressure, feed flow rate, reflux flow rate, steam flow rate, condensation temperature, and heating power.

[0026] It should also be noted that after obtaining the time series data of the variables, these raw time series data need to be preprocessed to eliminate the interference of industrial field noise, inconsistencies in dimensions, and asynchronous sampling by multiple sensors on the modeling accuracy.

[0027] Furthermore, the preprocessing specifically includes normalization, outlier removal, and time series alignment. More specifically, during normalization, this embodiment employs the MinMaxScaler algorithm to linearly map process variable data across all dimensions to a unified [0,1] interval, thereby eliminating the influence of different physical dimensions on neural network weight optimization. During outlier removal, the Z-score filtering algorithm is used to detect statistical anomalies in the time series data, removing industrial noise or abnormal system jumps exceeding a set standard deviation threshold to ensure the authenticity and reliability of the data samples. During time series alignment, to address the asynchronous issues caused by distributed sampling of process variables such as temperature, pressure, and flow rate, sequence alignment and synchronization calibration are performed on the time axis, ultimately constructing a multi-condition time series dataset that is fully synchronized in time steps and categorized by operating condition.

[0028] Furthermore, after completing the above data cleaning and standardization preprocessing, this embodiment divides the preprocessed time series data into source and target domains according to the working condition category and the richness of sample data under each working condition, in order to support the subsequent transfer learning logic.

[0029] It should also be noted that in this embodiment, the electric heating operating condition data with abundant data and complete historical accumulated samples is divided into the source operating condition dataset, which is used for the basic feature learning of the subsequent network architecture and the dynamic capture of general distillation. Correspondingly, since the steam heating operating condition and the mixed heating operating condition have a low frequency of operation in actual production or there is a lack of data, their sample data is relatively scarce. Therefore, the steam heating operating condition data and the mixed heating operating condition data are divided into the target operating condition dataset, which serves as the target object for subsequent migration fine-tuning.

[0030] By precisely dividing the two datasets according to the richness of the data under the heating conditions, a multi-condition training dataset composed of the source condition dataset and the target condition dataset is finally obtained. This realizes the transformation of the original complex industrial field data into a standardized input matrix that can deeply integrate transfer learning strategies, laying a solid data foundation for the deep pre-training of the temporal structure echo network in the subsequent step S2.

[0031] S2. Input the source working condition dataset into the preset temporal structure echo network for pre-training to obtain the source working condition pre-trained model; the preset temporal structure echo network includes an input encoder unit, a phase encoder unit, an echo unit, a structure regulator unit, an event-driven multi-scale pooling unit, and a decoding unit. For step S2, the core is to use the source working condition dataset constructed in step S1 to pre-train the temporal structure echo network, so that the network can fully learn the dynamic evolution law, variable coupling relationship and multi-scale temporal features in the alcohol distillation process, thereby obtaining a source working condition pre-trained model with cross-working condition knowledge expression ability, and providing basic model parameters for subsequent transfer learning of target working conditions.

[0032] It should be noted that the Temporal State Enhanced Network (TSE-Net) in this embodiment is a deep neural network architecture designed for complex industrial time-series modeling tasks. It achieves joint modeling of nonlinear, multi-time-delay, and multi-condition dynamic characteristics during alcohol distillation through a multi-module cascade collaborative mechanism. The overall structure of the Temporal State Enhanced Network can be represented as follows:

[0033] in, Represents time series data of variables. Indicates the input encoder unit, Indicates echo unit, This indicates the structural regulator unit. This represents an event-driven multi-scale pooling unit. Indicates the decoding unit. This represents the predicted result of the target measurement variable at a future time.

[0034] Furthermore, the input encoder unit receives the time-series variable data from the source operating condition dataset and maps the time-series signals to a low-dimensional latent space to extract the latent state features of the process variables. The encoder can be implemented using a linear layer or a small feedforward network, and its encoding process is represented as follows:

[0035] in, This represents the hidden state features of the input encoder output.

[0036] Furthermore, to enhance the model's ability to perceive the periodic variations and phase shift characteristics during alcohol distillation, a phase encoder unit is introduced in the encoding stage. The phase encoder unit introduces a learnable phase vector and encodes the periodic and phase change information in the time series through complex mapping or sine / cosine transform to obtain the phase-encoded information. This allows the network to simultaneously learn the numerical changes and temporal evolution of process variables. Unlike the fixed-position encoding of the Transformer, the phase parameters of this module are learnable and can be dynamically adjusted with the data, thereby improving the model's ability to express the cyclic behavior of complex operating conditions.

[0037] Furthermore, the latent state features are input into the echo unit. It should be noted that the echo unit consists of multiple echo units to construct a multi-scale temporal memory structure. Each unit contains a linear mapping and a learnable delay parameter, where the delay can be an integer (implemented via a sliding window) or a continuous value (approximated by interpolation). By introducing a learnable delay parameter and a historical state feedback mechanism, the latent state features at different times are correlated and modeled to obtain a state feature matrix. The state update mechanism of the echo unit is defined as follows:

[0038] in, Indicates the first The mapping matrix corresponding to each echo unit This represents the learnable delay parameter. Represents local context information. Indicates the bias term. This represents a non-linear activation function.

[0039] Through state propagation mechanisms at different delay scales, each echo unit can capture the dynamic changes in the alcohol distillation process at different time scales, thereby establishing long-term dependencies between process variables.

[0040] Furthermore, the state feature matrix output by the echo unit is input to the structure regulator unit. It should be noted that the structure regulator unit is connected via a small learnable network. A sparse adjacency matrix is ​​generated based on the statistical characteristics of the time series data of the variables (such as variance, error energy, prediction error, etc.). Used to dynamically adjust the information interaction weights generated between echo units:

[0041] in, It represents the statistical characteristics of time series data of variables.

[0042] Subsequently, the generated adjacency matrix is ​​used to perform weighted fusion of the state feature matrix output by the echo unit to obtain the weighted fused state features:

[0043] in:

[0044] This represents the state characteristic matrix formed by the outputs of each echo unit.

[0045] Through the aforementioned dynamic topology adjustment mechanism, the network structure can be adjusted in real time according to changes in operating conditions, thereby improving the model's adaptability to complex operating environments and non-stationary time-series data.

[0046] Furthermore, the weighted fused state features are input into the event-driven multi-scale pooling unit. It should be noted that, to enhance the model's responsiveness to key state change events, this unit identifies abrupt changes, disturbances, or state transitions in the time-series data of the variables through an event detection mechanism. Within these event windows, it adaptively pools the weighted fused echo unit output to obtain multi-scale pooling features. The calculation process is as follows:

[0047] in, Indicates the pooling scalar index. Indicates the length of the time window corresponding to the scale. This represents the learnable weights.

[0048] By using pooling operations at different scales, a multi-scale time representation that takes into account both short-term dynamic changes and long-term trend characteristics can be obtained.

[0049] Furthermore, the multi-scale pooling features output by the event-driven multi-scale pooling unit, together with the phase-encoded information, are input into the decoding unit. The decoding unit fuses the multi-scale pooling features and then generates predictions for future moments through a multilayer perceptron or lightweight residual blocks. The prediction process is represented as follows:

[0050] in, This represents the target measurement variable obtained from the prediction.

[0051] It should also be noted that a composite loss function is used to optimize the parameters of the temporal structured echo network throughout the pre-training process. The composite loss function includes the master prediction loss, the internal prediction error loss, and the structural sparsity regularization loss.

[0052] The primary prediction loss is expressed as:

[0053] Used to measure the error between the predicted result and the actual measured value.

[0054] The internal prediction error loss is expressed as:

[0055] Used to constrain the consistency of future state predictions for echo cells.

[0056] The structurally sparse regularized loss is expressed as:

[0057] This is used to reduce redundant connections and improve the model's generalization ability.

[0058] The final loss function is expressed as:

[0059] in, and These represent the weighting coefficients of the loss term.

[0060] It should be noted that, in this embodiment, The value is 0.1. The value is set to 0.001. The optimizer uses Adam, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. During pre-training, 80% of the source work condition dataset is used for parameter updates, and the remaining 20% ​​is used as a validation set. Training is terminated early when the validation loss no longer decreases after 10 consecutive epochs. After training, the parameters of the entire network are saved as the source work condition pre-trained model for transfer learning in step S3.

[0061] Through the above pre-training process, the temporal structure echo network can fully learn the dynamic mechanism characteristics, multi-scale temporal characteristics and coupling relationships between variables of the alcohol distillation process under the source conditions, and finally obtain the source condition pre-trained model with strong knowledge transfer ability, providing basic model parameters for transfer learning in step S3.

[0062] S3. Input the target working condition dataset into the source working condition pre-trained model, and obtain the target working condition soft measurement model through transfer learning; For step S3, the core lies in using the source condition pre-trained model obtained in step S2 to adapt the model to the target condition where historical sampling data is relatively scarce through a transfer learning strategy. While fully preserving the general industrial dynamic mechanisms and underlying spatiotemporal representation characteristics learned from the source condition, the model can quickly and stably adapt to the specific dynamic characteristics and physical mechanism changes of the target condition, thereby obtaining a high-precision soft measurement model for the target condition.

[0063] It should be noted that the target operating condition dataset is the operating condition data segmented in step S1 whose historical sampling data volume is lower than the preset sample size threshold, specifically including at least one of the steam heating operating condition dataset and the mixed heating operating condition dataset. In this embodiment, the process of transfer learning using the steam heating operating condition as the target operating condition is described, and the transfer fine-tuning process for the mixed heating operating condition is exactly the same.

[0064] Furthermore, transfer learning adopts a model parameter transfer-based approach, specifically including the following sub-steps: Sub-step S31: Load the source condition pre-trained model and obtain all network weights and bias parameters of each cascaded unit (specifically including the input encoder unit, phase encoder unit, echo unit, structure modulator unit, event-driven multi-scale pooling unit, and decoding unit).

[0065] Sub-step S32: Freeze all parameters of the input encoder unit and the phase encoder unit so that they remain fixed during subsequent transfer training and do not participate in gradient backpropagation. It should also be noted that the purpose of freezing the low-level encoder units is to fully preserve the ability to extract low-dimensional features of process variables and complex temporal periodic fluctuation patterns learned under the source conditions, and to effectively prevent small sample data from the target conditions from causing destructive interference to the basic feature representation.

[0066] Sub-step S33: Perform partial fine-tuning of the structure regulator unit. Specifically, only the smaller learnable network in the structure regulator unit is fine-tuned. The final layer mapping parameters are set to an updatable state, while the parameters of the remaining intermediate hidden layers remain frozen. It should also be noted that some fine-tuning structure regulator units enable the network's internal topology adjustment mechanism to quickly adapt to changes in the statistical distribution of data under the target operating conditions. This effectively avoids model overfitting under small sample conditions while dynamically adjusting the information interaction weights between echo modules.

[0067] Sub-step S34: Perform full fine-tuning of the decoding unit, that is, set all network parameters of the decoding unit to participate in gradient fine-tuning updates. It should also be noted that the decoding unit, as the final prediction output layer, is directly related to the specific alcohol concentration measurement variable under the target operating condition, and needs to be fully adapted and reconstructed to match and reconstruct the input-output mapping regression relationship under the target operating condition through full parameter updates.

[0068] Sub-step S35: The parameters of the echo unit and the event-driven multi-scale pooling unit remain in the source condition pre-training state during the transfer learning process and do not participate in fine-tuning updates.

[0069] Furthermore, the fine-tuning training process of transfer learning is entirely driven by iterative iteration using the target work condition dataset. To simplify the fine-tuning optimization objective, reduce the risk of non-convergence and divergence under multi-task conflict constraints with small samples, and prevent catastrophic forgetting of the general temporal memory representation of the source work condition, the loss function in this transfer fine-tuning stage has been specifically simplified by dimensionality reduction. Instead of using the composite loss function in the pre-training stage of step S2 (i.e., the joint loss including the internal echo self-supervised prediction error term and the structural sparsity regularization term), it only uses the same main prediction loss term as in the pre-training stage, thereby implementing purely task-oriented end-to-end fine-tuning.

[0070] In this embodiment, the industrial engineering hyperparameters for the transfer learning fine-tuning process are set as follows: the optimizer is the Adam optimizer; since only some specific weights need to be adaptively fine-tuned during the fine-tuning stage, in order to prevent the gradient from oscillating violently and destroying the learned general knowledge, the initial learning rate for fine-tuning is set to one-tenth of the pre-training learning rate, i.e., 0.0001; in order to stabilize the gradient backpropagation under small sample conditions, the batch size for fine-tuning is set to 16; the total number of fine-tuning training rounds is set to 50 rounds.

[0071] Meanwhile, an early stopping mechanism is introduced during the fine-tuning process. When the validation set loss no longer decreases for five consecutive rounds, the fine-tuning training is automatically terminated early. It should also be noted that adopting the aforementioned low learning rate, small batch size, and early stopping mechanism helps to stabilize the fine-tuning parameter space under the harsh conditions of small sample size in the target working condition, preventing drastic fluctuations in parameters.

[0072] After training, the fine-tuned parameters of the entire network are saved, thus obtaining the soft-sensing model for the target operating condition of the alcohol distillation process. It should be noted that this model has been deeply adapted to the unique dynamic characteristics and thermophysical inertia of the target operating condition, and can be directly deployed for real-time high-precision online soft-sensing prediction of alcohol concentration under the target operating condition in step S4.

[0073] It should also be noted that for multiple different target operating conditions (such as steam heating and mixed heating), repeating the migration and adaptation process of sub-steps S31 to S35 above will yield independent soft measurement models for each target operating condition. Each target operating condition model operates independently, but shares the same source operating condition pre-trained model as the underlying common base parameters during offline construction.

[0074] S4. Obtain real-time variable data under the working condition to be predicted, and input the real-time variable data into the target working condition soft measurement model to obtain the prediction result of the target measurement variable; It should be noted that the core of step S4 is to use the target operating condition soft measurement model obtained in step S3 to perform online inference calculations on the real-time operating data in the alcohol distillation process, thereby realizing real-time soft measurement prediction of the target measurement variable.

[0075] Furthermore, during the operation of the alcohol distillation unit, process variable data are continuously collected through a distributed control system, a programmable logic controller, and field sensors, and missing value processing, outlier removal, normalization processing, and time alignment are performed according to the preprocessing method in step S1.

[0076] It should be noted that, in this embodiment, the real-time variable data includes at least one or more of the following: column top temperature, column bottom temperature, column pressure, feed flow rate, reflux flow rate, steam flow rate, condensation temperature, and heating power. The target measurement variable is alcohol concentration, and the prediction step size is... This means outputting the predicted alcohol concentration value for the next time step after the current time window.

[0077] Furthermore, a real-time input sample is constructed using a sliding time window method. The process variables at the current moment and its historical moments are combined into a time series input matrix. The real-time sequence sample is then input into the target working condition soft measurement model obtained in step S3 for forward inference calculation.

[0078] It should be noted that the target operating condition soft measurement model inherits the temporal structure echo network structure constructed in step S2, and completes the target operating condition transfer learning adaptation through step S3. Therefore, it can make full use of the general dynamic features learned in the source operating condition and the unique dynamic laws in the target operating condition to predict the alcohol concentration under the current operating condition.

[0079] Through the above steps, real-time online soft measurement and prediction of key quality indicators in the alcohol distillation process can be achieved without the need for continuous measurement by an online concentration analyzer or when the analyzer has a detection lag.

[0080] S5. Output and store the prediction results of the target measurement variable to complete the soft measurement of alcohol distillation.

[0081] Furthermore, the target operating condition soft measurement model outputs the prediction results for future moments and sends the predicted alcohol concentration results to the host computer monitoring platform, human-machine interface or process control system in real time to display the current concentration change trend and subsequent control decisions.

[0082] It should be noted that the display system refreshes the predicted value every sampling period, enabling operators to observe the dynamic changes in the alcohol concentration at the top of the tower in real time.

[0083] It should also be noted that the stored data includes timestamps, corresponding operating condition identifiers, and predicted alcohol concentrations. The data format conforms to industry standards to facilitate joint analysis and control decisions with other process variables. In this embodiment, the storage strategy employs a cyclic overwrite approach, retaining the predicted data for the most recent 30 days to support historical trend analysis and accident tracing.

[0084] Through step S5 above, the method of the present invention realizes the practical application of model prediction results in the industrial field of alcohol distillation, providing real-time and reliable soft measurement data support for operation guidance, quality control and energy efficiency optimization of the distillation process.

[0085] Example 2: See Figure 2 As shown, this embodiment is a complex multi-condition alcohol distillation soft measurement system based on neural networks, including: The dataset partitioning module is used to acquire variable time-series data of the alcohol distillation process under several working conditions, preprocess the variable time-series data to obtain preprocessed time-series data, and partition the preprocessed time-series data into source working condition dataset and target working condition dataset according to the working conditions. The network pre-training module is used to input the source working condition dataset into a preset time-series structured echo network for pre-training to obtain the source working condition pre-trained model. The transfer learning module is used to input the target working condition dataset into the source working condition pre-trained model and obtain the target working condition soft measurement model through transfer learning. The prediction execution module is used to acquire real-time variable data under the working condition to be predicted, and input the real-time variable data into the target working condition soft measurement model to obtain the prediction result of the target measurement variable. The results output module is used to output and store the prediction results of the target measurement variable to complete the soft measurement of alcohol distillation.

[0086] Example 3: To further verify the effectiveness of the complex-condition alcohol distillation soft measurement method based on Temporal Structured Echo Network (TSE-Net) described in this invention, a comparative experimental example was designed. The experiment selected the distillation process under electrically heated conditions as the test scenario and compared the prediction performance of the method of this invention with traditional deep learning models, including Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Multilayer Perceptrons (MLP), and Transformer models.

[0087] During the experiment, historical operating condition data were first preprocessed and the source operating condition model was pre-trained according to steps S1-S2 in Example 1 to construct the TSE-Net source operating condition pre-trained model. Then, for the test operating condition, real-time variable data was obtained using step S4 and input into each model for alcohol concentration prediction. The prediction error and concentration prediction results of each model across the entire test sample sequence were recorded. All comparative experiments were based on the same electrically heated operating condition dataset for unified pre-training and testing evaluation. The input feature dimension, sliding time window length, and prediction forward step size, and other temporal hyperparameters of each model remained completely consistent.

[0088] See Figure 3 As shown, Figure 3 This is a box plot showing the prediction error distribution of five soft sensor models on the test set. The horizontal axis represents the model type, the vertical axis represents the prediction error (%), and the central horizontal solid line represents the zero baseline of the error.

[0089] Box plots precisely quantify the error concentration and generalization robustness of each model by using the upper and lower edges, quartile intervals (box height), and median position. Analysis Figure 3 The statistical characteristics show that: The TSE-Net model (reddish-brown box) of this invention has the flattest box height among all comparison models, and its overall prediction error strictly converges to... Within a very small range, and with its median and mean perfectly aligned with the zero-error baseline, the model exhibits no systematic prediction bias and possesses extremely high measurement accuracy and industrial stability. The Transformer model (green box) and the LSTM model (orange box) have significantly larger box spans, with error fluctuation ranges reaching [missing information]. and Furthermore, the Transformer model exhibits significant concave deformation, indicating its weak generalization ability for complex industrial noise. The MLP model (purple box) has the widest error distribution, with the maximum positive and negative errors approaching... and Furthermore, the median deviates significantly from the zero line. Although the CNN model (blue box) has a narrow box, its overall error distribution falls completely below the zero line, indicating a significant negative systematic error bias, and there are multiple significant outliers outside the upper and lower edges.

[0090] See Figure 4 As shown, Figure 4 This is a comparison curve of the predictive performance of five soft sensor models under electrically heated conditions in this embodiment. The horizontal axis represents the test sample index, and the vertical axis represents the alcohol concentration (%). The solid black line represents the true value of the alcohol concentration at the top of the distillation column measured by an online analyzer or laboratory. Through analysis of... Figure 4 By comparing and analyzing the evolution trends of the various characteristic curves, the following conclusions can be drawn: The fitting performance of the TSE-Net model of this invention: The TSE-Net model prediction curve, represented by the blue dashed line in the figure, shows a high degree of consistency and overlap with the black solid line throughout the entire sampling period. Whether in the linear region where alcohol concentration rises steadily or in the region of abrupt slope changes caused by non-stationary perturbations, the method of this invention can accurately and in real-time track the dynamic thermophysical inertia of alcohol concentration. The CNN model, represented by the red short dashed line, produces significant overestimation in the sample interval [0,50] and severe underfitting in the interval [60,95], failing to effectively establish a long-delay regression mapping. The LSTM model, represented by the green double-dotted line, deviates significantly in the initial stage, and the overall response exhibits obvious phase lag, indicating that the traditional cyclic structure is insufficient in dealing with the nonlinear strongly coupled variables of the distillation column. Although the Transformer model, represented by the purple dotted line, has an overall trend close to the true value, it exhibits relatively violent high-frequency numerical oscillations at non-stationary inflection points (such as around sample indices 40-60). The MLP model represented by the orange double dashes exhibits extremely high volatility instability, with local predictions showing jagged divergence and extremely high fitting noise.

[0091] In summary, the Temporal Structured Echo Network (TSE-Net) constructed in this invention, thanks to the learnable phase encoder introduced internally to capture dynamic offsets and the real-time reconstruction of adaptive topological features by echo unit and structure adjuster unit, significantly outperforms existing conventional deep learning temporal modeling methods in terms of soft measurement accuracy, disturbance resistance and error convergence speed under complex alcohol distillation conditions, and has extremely high value for industrial field application.

[0092] Example 4: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0093] This computer device can be a server, and its internal structure diagram can be as follows: Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores server data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a complex multi-condition soft measurement method for alcohol distillation based on neural networks.

[0094] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0095] Example 5: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0096] If the functions implemented by the method are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art or the current technical solution, can be embodied in the form of a software product. This current computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0098] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0099] The technical features of this invention not described can be implemented by or using existing technology, and will not be repeated here. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.

Claims

1. A complex multi-condition alcohol distillation soft measurement method based on neural networks, characterized in that, Includes the following steps: S1. Obtain variable time series data of alcohol distillation process under several working conditions, preprocess the variable time series data to obtain preprocessed time series data, and divide the preprocessed time series data into source working condition dataset and target working condition dataset according to the working conditions. S2. Input the source working condition dataset into a preset temporal structure echo network for pre-training to obtain the source working condition pre-trained model; the preset temporal structure echo network includes an input encoder unit, a phase encoder unit, an echo unit, a structure regulator unit, an event-driven multi-scale pooling unit, and a decoding unit. S3. Input the target working condition dataset into the source working condition pre-trained model, and obtain the target working condition soft measurement model through transfer learning; S4. Obtain real-time variable data under the working condition to be predicted, and input the real-time variable data into the target working condition soft measurement model to obtain the prediction result of the target measurement variable; S5. Output and store the prediction results of the target measurement variable to complete the soft measurement of alcohol distillation; The specific data flow direction of inputting the source operating condition dataset into the preset time-series echo network is as follows: The time-series data of variables in the source working condition dataset are encoded by the input encoder unit and the phase encoder unit respectively to obtain low-dimensional hidden state features and phase encoding information; The hidden state features are input into the echo unit for state update, resulting in the state feature matrix; The state feature matrix is ​​input into the structure regulator unit, which generates a sparse adjacency matrix. The state feature matrix is ​​then weighted and fused based on the sparse adjacency matrix to output the weighted fused state features. The weighted fusion state features are input into the event-driven multi-scale pooling unit. The event-driven multi-scale pooling unit identifies the event window and performs multi-scale pooling on the weighted fusion state features within the event window to obtain multi-scale pooled features. The multi-scale pooling features and phase-encoded information are input into the decoding unit. The decoding unit then fuses and maps the multi-scale pooling features and phase-encoded information to output the prediction results of the target measurement variables at future time moments.

2. The complex multi-condition alcohol distillation soft measurement method based on neural networks according to claim 1, characterized in that, The preprocessing of the variable time series data includes normalization, outlier removal, and time series alignment. The step of dividing the preprocessed time-series data into source operating condition datasets and target operating condition datasets according to operating conditions includes: The preprocessed time-series data are classified according to the amount of historical sample data corresponding to each working condition. The working condition data with a sample data amount higher than the preset sample data amount threshold is classified as the source working condition dataset, and the working condition data with a sample data amount lower than the preset sample data amount threshold is classified as the target working condition dataset.

3. The complex multi-condition alcohol distillation soft measurement method based on neural networks according to claim 2, characterized in that, The input encoder unit uses a linear transformation layer combined with a nonlinear activation function to map variable time-series data into low-dimensional hidden state features. The phase encoder unit introduces a learnable phase vector at each time step, and encodes the periodic information and phase change information of the time series through complex mapping or sine and cosine transform, and outputs phase encoded information. The echo unit consists of several parallel echo subunits, each containing a learnable delay parameter and a linear mapping; the state update mechanism of the echo subunit is defined as follows: in, Indicates the first The mapping matrix corresponding to each echo unit This represents the learnable delay parameter. Represents local context information. Indicates the bias term. This represents a non-linear activation function.

4. The complex multi-condition alcohol distillation soft measurement method based on neural networks according to claim 3, characterized in that, The process of generating a sparse adjacency matrix using the structure regulator unit includes: extracting statistical features from the time-series data of the variables, and generating a sparse adjacency matrix based on the statistical features using the Softmax function; the statistical features include variance, error energy, and prediction error. The event-driven multi-scale pooling unit identifies event windows in the time-series data of variables through an event detection mechanism, and performs multi-scale pooling on the weighted fusion state features based on the identified event windows; the event windows include mutation point windows, disturbance point windows, and state transition point windows; The decoding unit fuses the multi-scale pooling features corresponding to each time scale with the phase coding information, and establishes a mapping relationship between process variables and target measurement variables through a multilayer perceptron or residual network, and outputs the target measurement variable prediction results for future time moments.

5. The complex multi-condition alcohol distillation soft measurement method based on neural networks according to claim 4, characterized in that, The pre-training process in step S2 employs a composite loss function, which includes the main prediction loss, the internal prediction error loss, and the structural sparsity regularization loss. Its expression is as follows: in, Forecast the loss. For internal prediction error loss, For structurally sparse regularized loss, and These represent the weighting coefficients of the loss term.

6. The complex multi-condition alcohol distillation soft measurement method based on neural networks according to claim 5, characterized in that, The step of inputting the target working condition dataset into the source working condition pre-trained model and obtaining the target working condition soft measurement model through transfer learning includes: Load the network parameters of the pre-trained model under the source operating conditions; The parameters of the input encoder unit, phase encoder unit, echo unit, and event-driven multi-scale pooling unit are frozen. Using the target operating condition dataset, the structure regulator unit and the decoding unit are fine-tuned based on the master prediction loss to obtain the target operating condition soft measurement model.

7. A complex multi-condition alcohol distillation soft measurement system based on neural networks, characterized in that, The steps for performing the neural network-based complex multi-condition alcohol distillation soft measurement method according to any one of claims 1 to 6 include: The dataset partitioning module is used to acquire variable time-series data of the alcohol distillation process under several working conditions, preprocess the variable time-series data to obtain preprocessed time-series data, and partition the preprocessed time-series data into source working condition dataset and target working condition dataset according to the working conditions. The network pre-training module is used to input the source working condition dataset into a preset time-series structured echo network for pre-training to obtain the source working condition pre-trained model. The transfer learning module is used to input the target working condition dataset into the source working condition pre-trained model and obtain the target working condition soft measurement model through transfer learning. The prediction execution module is used to acquire real-time variable data under the working condition to be predicted, and input the real-time variable data into the target working condition soft measurement model to obtain the prediction result of the target measurement variable. The results output module is used to output and store the prediction results of the target measurement variable to complete the soft measurement of alcohol distillation.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the neural network-based soft measurement method for complex multi-condition alcohol distillation as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the complex multi-condition alcohol distillation soft measurement method based on neural networks as described in any one of claims 1 to 6.

Images