A soft sensing method for spatiotemporal feature fusion in digital twin platform of process industry

Abstract

The present application discloses a soft measurement method for the fusion of spatiotemporal features of a digital twin platform for process industries. This method is based on the real-time, online data reading, transmission and processing capabilities of the digital twin platform for process industries. It divides the process variable data acquired by the platform based on the production line process, generates time series, and realizes the spatiotemporal feature extraction and fusion of various time series through a distributed time series feature extraction network and a feature fusion network, thereby fitting and estimating performance indicators. The method provided by the present application can quickly and accurately perform soft measurement on some performance indicators in the process industry that are difficult to measure directly or are costly to measure, and improve the effect of soft measurement of performance indicators in the process industry, thereby providing reference and guidance for downstream tasks such as parameter optimization and production control in the process industry.

Description

Technical Field

[0001] The present application relates to the fields of computer technology and artificial intelligence, and in particular to a soft measurement method for the fusion of spatiotemporal features of a digital twin platform for process industries. The method is mainly applicable to the real-time and accurate identification and estimation tasks of performance indicators that are difficult to obtain directly in process industry control systems. Background Art

[0002] The following is an introduction to some terms and background technologies involved in this application.

[0003] Process Industry: Process Industry refers to the continuous manufacturing industry involving material transformation and energy conversion, such as chemical, steel, cement, petroleum, papermaking, etc., and its production scale is usually large. In the process industry, from raw materials to final products, it often involves a large number of physical and chemical reactions, including complex energy, material and information flows. Therefore, real-time monitoring, optimization and control of process industries are crucial.

[0004] Digital Twin Platform: A digital twin platform is a type of platform that connects physical entities (such as factories, equipment or process systems) with their virtual models to achieve real-time data synchronization and interaction. This platform usually has monitoring, visualization, decision-making and other functions. In the process industry, the integration and modeling of data generated by industrial processes helps to achieve real-time monitoring, optimization and control of the production process, thereby improving production efficiency, product quality and reliability.

[0005] Process variables: Process variables refer to various physical quantities and parameters that are measured, monitored and / or controlled in industrial processes. These variables can be obtained through direct measurement, such as temperature, pressure, flow, concentration, etc., or calculated through other parameters, such as energy consumption, thermal efficiency, reaction rate, etc., and can also be obtained through manual testing and detection, such as the chemical composition and properties of products or raw materials.

[0006] Time Series: A time series refers to a collection of continuous data points that have a time sequence and are arranged in time order. Compared with a single data point, a time series is a data sequence obtained by observing or measuring a certain variable or a group of variables within a continuous time range, and therefore has a time scale and corresponding time attributes.

[0007] Performance Index: A performance index is a metric used to evaluate, measure or describe the quality, efficiency, reliability and performance of a system, product, service or process. In process industries, performance indexes usually refer to the performance of process lines, systems, equipment or product quality, such as production line energy consumption, equipment remaining life, product dimensional accuracy, appearance quality, chemical composition, etc. In a specific context, process variables and performance indicators can be converted into each other.

[0008] Soft Sensing: Soft Sensing refers to the process of using mathematical models or data analysis techniques to infer and calculate variables that cannot be directly measured or are difficult to measure through easily accessible variables. In the process industry, using some real-time, easily accessible process variable data points or data sequences to perform soft measurements on performance indicators that are difficult to measure directly and in real time is one of the key links in the monitoring, optimization and control of process industry control systems.

[0009] Autoencoder: Autoencoder is a model in the field of machine learning, which consists of two parts: encoder and decoder. The encoder maps the input data from a high-dimensional space to a low-dimensional latent space, while the decoder reconstructs the input data through the latent space. The goal of the autoencoder is to learn the compressed representation of the data and reconstruct the input data as accurately as possible. Therefore, this model has been widely used in data dimensionality reduction, feature extraction, noise reduction and other fields.

[0010] Attention Mechanism: Attention Mechanism is a commonly used technique in machine learning. It enables the model to automatically focus on information related to the current task and ignore irrelevant parts when processing sequence data. This technique helps to extract key features from time series data, improve model performance, and enhance the model's ability to process long sequences.

[0011] Figure 1The figure is a schematic diagram of the structure of an exemplary process industry digital twin platform. The process industry digital twin platform is built based on various physical entities such as production lines, systems, and equipment, and is mainly divided into three layers: data perception layer, data integration layer, and data service layer. The data perception layer is the bottom layer of the process industry digital twin platform, which is used to perceive and collect data related to the production process. The means of data collection include, for example, obtaining through sensors deployed in various devices, obtaining through soft measurement methods, and obtaining through artificial testing, detection, and uploading methods. The data perceived by various methods will be transmitted to the data integration layer. The data integration layer is located in the middle layer of the process industry digital twin platform, and is responsible for integrating, processing, and storing the data obtained from the data perception layer. Specifically, the data integration layer first cleans, integrates, and converts the data collected by the data perception layer through certain rules and protocols, and on this basis stores it in the corresponding database, such as a database built into the DCS system, an edge database that is partially independent of the DCS system, or a database located in the cloud. The data service layer is the top layer of the process industry digital twin platform. Its function is to carry out process monitoring, parameter optimization, production control and other tasks based on the data provided by the data integration layer, to achieve comprehensive monitoring and optimization of the process industry production process, and to improve production efficiency, product quality and resource utilization efficiency.

[0012] At present, in the industrial digital twin platform, soft measurement methods for various performance indicators in industry have been initially applied in practice, thus solving or alleviating the problems of difficult detection, high acquisition cost, poor real-time update, and large fluctuation in accuracy of some performance indicators. The data-driven soft measurement method of performance indicators is a method that uses industrial historical data to establish a performance indicator model through data analysis, statistical learning and other methods. On this basis, real-time industrial data is input into the model to obtain real-time soft measurement results of performance indicators. Data-driven methods are widely used in soft measurement tasks of various industrial performance indicators due to their strong fitting ability and low reliance on expert knowledge.

[0013] However, in the process industry, soft measurement of performance indicators based on data-driven modeling faces two major challenges: First, in many cases, the scale of labeled data is much smaller than that of unlabeled data. For example, some test results for products are often updated only once every few hours, while some data collected by sensors are updated at a frequency of seconds. This results in only a small number of process variable data points having corresponding performance indicator labels. Secondly, due to the uncontrollable fluctuations in raw material composition and equipment operation, as well as the high coupling of physical and chemical reactions in the production process, the correlation between soft measurement performance indicators and process variables is weak, and nonlinear relationships are prominent. The above two problems make it difficult for simple data-driven methods that rely only on labeled data to establish an accurate relationship between performance indicators and process variables, resulting in poor soft measurement results and low precision of performance indicators, which in turn affects the optimization and control effects based on soft measurement results of performance indicators.

[0014] In addition, in process industries, the production of products often needs to be achieved through a series of multiple processes that are connected one after another. Due to the large scale of process industry production and the different production properties of different processes, the time series of process variables collected from different processes often show differences in spatiotemporal information. Time series information can be defined as a combination of static and dynamic levels of process variable time series. Among them, the static level represents the mean of a time series, while the dynamic level covers information such as trends and seasonality. On the other hand, spatial information can be defined as the difference between the information contained in the time series of process variables from different processes. For example, due to the serial characteristics of the process, the time series of process variables collected from different processes have different time ranges of influence on performance indicators. However, the above two differences have not received enough attention in existing research and engineering practice, which has limited the further improvement of the accuracy of process industry performance indicator modeling and soft measurement effect. Combining the advantages of the digital twin platform and making full use of the spatiotemporal information of process industry data will help improve the effect of performance indicator modeling, further improve the accuracy of performance indicator soft measurement, and provide more accurate reference and guidance for downstream tasks such as monitoring, diagnosis, and optimization control of the digital twin platform of the process industry. Summary of the invention

[0015] In view of the above background, this application provides a soft measurement method for spatiotemporal feature fusion of a process industry digital twin platform. The technical solution of the present invention is as follows:

[0016] A soft measurement method for the fusion of spatiotemporal features of a digital twin platform for process industries, comprising: obtaining the operating data of process variables of each production process of the process industry through a data perception module located in the data perception layer of the digital twin platform for process industries, and uploading it to the data integration layer, wherein the data perception module refers to a collection of methods for obtaining the operating status data of each process of the process industry, including various types of physical sensors, soft sensors, and some non-real-time processes of manual testing, detection, and uploading data; process variables refer to various physical quantities and parameters measured, monitored, and / or controlled in the industrial process, such as temperature, pressure, Flow rate, concentration, etc., or energy consumption, thermal efficiency, reaction rate, etc. calculated through other parameters, or obtained through manual testing and detection, such as the chemical composition and properties of products or raw materials; in the data integration layer, the various types of process variable data obtained are scaled and grouped, and a certain time window size is set to generate corresponding time series data. Among them, scale integration refers to giving a sampling interval and resampling the uploaded data through this sampling interval, so that the update frequency of all data is on the same time scale. Data grouping refers to dividing the process variable data into multiple groups based on the data collection process. In particular, the generation of time series data refers to the process of taking the current timestamp as the starting point and a time window size as the span, selecting process variable data within this range, and splicing each process variable data point forward and backward in chronological order to form time series data; in the data service layer, the time series data of each process is read, and the distributed time series feature extraction network is used to extract the spatiotemporal features contained in the time series of the process variables of each process in parallel. Among them, the time series feature extraction network is an autoencoder network composed of an encoder and a decoder. Its function is to extract one-dimensional time series features from two-dimensional time series data from a process. Multiple time series features The feature extraction network extracts features from multiple processes in parallel to obtain the spatiotemporal characteristics of the entire process industry; through the feature fusion network, the spatiotemporal features extracted from each process are fused to obtain the soft measurement results of the performance indicators, that is, the extracted spatiotemporal features are used as the input of the network, and the fusion of spatiotemporal features is realized through the transformation of each level within the network, and the corresponding performance indicator estimation value is obtained by fitting; finally, the soft measurement results of the performance indicators are fed back to the data integration layer for storage, thereby realizing long-term and stable storage of the soft measurement results, and also facilitating the instant calling of some other applications in the data service layer, forming a closed loop of data flow.

[0017] The present application also provides a time series feature extraction autoencoder network, which consists of two parts: an encoder for time series feature extraction and a decoder for input sequence reconstruction. Among them, the encoder part inputs process variable time series data, and extracts weighted mean features and weighted dynamic features through a series of network layer transformations and processing, and outputs the feature splicing results of the two. The decoder part uses the feature splicing results to try to reconstruct the input time series and generate a reconstructed time series. In the decoder of this network, not only weighted dynamic features are used, but also process variable weighted mean features are introduced to participate in sequence reconstruction to obtain a reconstructed sequence, thereby guiding the encoder to learn a more comprehensive time series representation and effectively obtain dynamic and static information in the time series.

[0018] The present application also provides a synchronous training method for time series feature extraction and feature fusion network parameters, including: determining the data duration and required process variables required for network parameter training, obtaining process variable data for each process step of network parameter training; scalably integrating and grouping the process variable data, and constructing a network parameter training data set, wherein the construction of the data set includes the steps of time series generation, dividing a training set, a validation set, a test set, etc.; connecting a spatiotemporal feature extraction network including multiple time series feature extraction networks with a feature fusion network front to back, and constructing an end-to-end network structure, that is, establishing a connection between the encoder output layer of each time series feature extraction network and the input layer of the feature fusion network, thereby forming all networks into an end-to-end structure; using a semi-supervised loss function to train the end-to-end network parameters, and uploading the training results, wherein the semi-supervised loss function refers to a composite loss function that simultaneously considers the regression accuracy of the performance indicator label value, the reconstruction accuracy of the input time series, and the distance between the features extracted by each time series feature extraction network, and the model training results refer to the parameters of various networks, the input time series length of each time series feature extraction network, the process variables contained in each time series, and other related information used to configure the performance indicator soft measurement model.

[0019] The present application provides a soft measurement method for the fusion of spatiotemporal features of a digital twin platform for process industries. Based on the real-time and online data reading, transmission and processing capabilities of the digital twin platform for process industries, the process variable data acquired by the platform is divided based on the production line process, and time series are generated. Through the distributed time series feature extraction network and feature fusion network, the spatiotemporal feature extraction and fusion of various time series are realized, and then the performance indicators are fitted and estimated. This method can quickly and accurately perform soft measurement on some performance indicators in the process industry that are difficult to measure directly or are costly to measure, and improve the effect of soft measurement of process industry performance indicators, thereby providing reference and guidance for downstream tasks such as parameter optimization and production control in the process industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural diagram of an exemplary process industry digital twin platform provided in this application.

[0021] Figure 2 This is a flow chart of a soft measurement method for spatiotemporal feature fusion of a process industry digital twin platform provided in this application.

[0022] Figure 3 It is a structural diagram of a temporal feature extraction autoencoder network provided by the present application as an embodiment.

[0023] Figure 4 This is a flow chart of a method for synchronous training of time series feature extraction and feature fusion network parameters provided by the present application as an embodiment. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0025] In the embodiments of the present application, "and / or" describes the access relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B may be singular or plural. In addition, in the embodiments of the present application, "first", "second", "third", etc. are only used to distinguish the contents of different objects and have no other special meanings.

[0026] The technical solution provided by the present application is described in detail below in conjunction with the accompanying drawings.

[0027] Figure 2 A flow chart of a soft measurement method for spatiotemporal feature fusion of a process industry digital twin platform provided in this application.

[0028] See also Figure 2 , the method may include the following steps:

[0029] 101. Through the data perception module located in the data perception layer of the process industry digital twin platform, the operating data of the process variables of each production process of the process industry are obtained and uploaded to the data integration layer.

[0030] In this embodiment, the data perception module refers to a collection of methods that can obtain the operating status data of each process in the process industry, and its purpose is to detect the operating status of each process in real time and convert the operating status into a suitable data representation. The data perception module includes not only various types of physical sensors, such as thermocouples, infrared scanners, pressure gauges, flow meters, etc., but also soft sensors for calculating parameters that cannot be directly obtained, and also includes some non-real-time processes of manual testing, detection, and uploading data.

[0031] In this embodiment, as mentioned above, process variables refer to various physical quantities and parameters measured, monitored and / or controlled in industrial processes. These variables can be obtained by direct measurement, such as temperature, pressure, flow rate, concentration, etc., or calculated by other parameters, such as energy consumption, thermal efficiency, reaction rate, etc., and can also be obtained by manual testing and detection, such as the chemical composition and properties of products or raw materials.

[0032] In actual applications, the data obtained by the data perception module usually needs to be transmitted to the data integration layer of the process industry digital twin platform through the corresponding interface by selecting a suitable data transmission protocol, and saved and encrypted in a certain format.

[0033] 102. In the data integration layer, the acquired process variable data are scaled and grouped, and a certain time window size is set to generate corresponding time series data.

[0034] In this embodiment, scale integration refers to giving a sampling interval and resampling the uploaded data through this sampling interval, so that the update frequency of all data is on the same time scale. In this process, the redundancy of data can be reduced on the basis of retaining most of the data information. At the same time, the scale integration of data is also helpful for the subsequent generation of time series data with unified scale. Data grouping refers to dividing process variable data into multiple groups based on the process of data collection. In practical applications, considering the continuity of process industries, some process variables may affect multiple processes at the same time. In this case, these process variables can be classified into more than one group.

[0035] In this embodiment, for the process variable data after scale integration and data grouping, time series data with a fixed length is generated according to a certain time window size. A time window refers to a specific time period defined on the time axis of a process industry database, which is used to limit the processing scope of the data. In practical applications, for a certain set of data, the current timestamp is usually used as the starting point, and the time window size is used as the span. The process variable data within this range is selected, and each process variable data point is spliced ​​forward and backward in chronological order to form time series data. Considering that the time series data itself has a time dimension and is usually composed of data of multiple variables, the time series data cannot usually be saved in the form of tabular data, but is temporarily stored in the data integration layer in the form of JSON (JavaScript Object Notation) or the like.

[0036] 103. In the data service layer, the time series data of each process is read, and the time and space features contained in the time series of the process variables of each process are extracted in parallel through a distributed time series feature extraction network.

[0037] In this embodiment, considering the real-time requirements of the soft measurement of the performance indicators of the digital twin platform, multiple autoencoder time series feature extraction networks composed of encoders and decoders are pre-trained. Each time series feature extraction network is used to extract one-dimensional time series features from two-dimensional time series data from a process, that is, a comprehensive time series representation containing the mean, trend, seasonality and other information of the corresponding process variable time series. The process of extracting features from different sequences by multiple time series feature extraction networks can not only extract the time information contained in different sequences, but also take into account the differences in spatial information between different sequences, thereby obtaining comprehensive spatiotemporal characteristics of the process industry process from various time series. Considering the distributed characteristics of the time series feature extraction network, the extraction of spatiotemporal features of various time series can be implemented in a parallel manner, thereby reducing the resource consumption of the process industry digital twin platform.

[0038] 104. The spatiotemporal features extracted from each process are fused through the feature fusion network to obtain the soft measurement results of the performance indicators.

[0039] In this embodiment, through the distributed time series feature extraction network, the spatiotemporal features of the process industry with a certain dimension can be extracted from the time series data of each process. There is a certain degree of correlation between the information contained in these features and the performance indicators of the required soft measurement, so they can be used for soft measurement estimation of performance indicators. This process can be completed through a trained feature fusion network, that is, the extracted spatiotemporal features are used as the input of the network, and the spatiotemporal features are fused through the transformation of each level within the network, and the corresponding performance indicator estimation value is obtained by fitting.

[0040] In this embodiment, the feature fusion network can be a simple fully connected network, or it can be various deep networks with more complex structures and more comprehensive functions, or a hybrid network that combines latent variable methods and neural network levels. Therefore, this embodiment does not limit the structure and type of the feature fusion network. A series of networks that have the function of fusing input spatiotemporal features and outputting soft measurement results of performance indicators can be used as the feature fusion network in this embodiment.

[0041] 105. Feedback the soft measurement results of the performance indicators to the data integration layer for storage.

[0042] In this embodiment, through steps 101-104, soft measurement results of performance indicators that are difficult to measure or costly to measure can be obtained. This soft measurement result can not only provide feedback for the operation process of the process industry, but also has a high reference value for downstream tasks such as parameter optimization and production control. By feeding back the soft measurement results of performance indicators to the data integration layer of the process industry digital twin platform, long-term and stable storage of soft measurement results can be achieved, and it is also conducive to the instant call of some other applications in the data service layer, forming a closed loop of data flow.

[0043] In practical applications, the feedback of performance indicator soft measurement results to the data integration layer can be carried out in a variety of ways, such as through the built-in API (application programming interface) of the data service layer, through Web access and database operations, etc., to achieve real-time update of soft measurement results in the data integration layer.

[0044] The technical solution provided in the embodiment of the present application is based on the real-time, online data reading, transmission and processing capabilities of the process industry digital twin platform. The process variable data obtained by the platform is divided based on the production line process, and time series are generated. Through the distributed time series feature extraction network and feature fusion network, the spatiotemporal feature extraction and fusion of various time series are realized, and then the performance indicators are fitted and estimated. Therefore, a soft measurement method for spatiotemporal feature fusion of the process industry digital twin platform is proposed. The above method can quickly and accurately perform soft measurement on some performance indicators in the process industry that are difficult to measure directly or have high measurement costs, and improve the effect of soft measurement of process industry performance indicators, thereby providing reference and guidance for downstream tasks such as parameter optimization and production control in the process industry.

[0045] The above is an introduction to a technical solution for a soft measurement method for the fusion of spatiotemporal features in a digital twin platform for process industries. Among them, the extraction of spatiotemporal features is one of the most important links in the soft measurement of performance indicators. The design and application of the timing feature extraction network has a very direct impact on the effect of spatiotemporal feature extraction performed by multiple timing feature extraction networks, and the extracted spatiotemporal features have a decisive role in the effect of soft measurement of performance indicators. Therefore, designing a comprehensive and efficient timing feature extraction network is the key to realizing this technical solution. The following is an introduction to a comprehensive and efficient timing feature extraction network provided by this application as an embodiment.

[0046] Figure 3 is a schematic diagram of the structure of a temporal feature extraction autoencoder network provided by the present application as an embodiment. Figure 3 , this network consists of two parts: an encoder for time series feature extraction and a decoder for input sequence reconstruction. The encoder part inputs the process variable time series data X, and extracts the weighted mean feature M through a series of network layer transformations and processing. att and weighted dynamic features V att , and outputs the feature concatenation result Z of the two. The decoder uses the feature concatenation result Z to try to reconstruct the input time series and generate a reconstructed time series

[0047] In this embodiment, for the construction of the encoder part, a self-attention layer is first introduced to assign attention weights of the time dimension and feature dimension learned by the network to the input time series X of shape L×D. Where L is the length of the time series and D is its feature dimension. The purpose of the self-attention layer is to model the global dependency of the input time series and make the model focus on specific parts of the sequence in a weighted manner. The expression of the self-attention mechanism can be expressed as:

[0048]

[0049] In this expression, d k Represents the dimension of the input sequence X. Using its square root as the denominator helps scale the attention weights to avoid excessive influence of the attention weights on the original sequence. att By averaging along the time dimension, we can get a weighted mean feature M with a shape of 1×D. att Then, a set of long short-term memory (LSTM) networks is used to weight the result X att The dynamic pattern of the network is modeled. In the LSTM network, the working principle of each LSTM unit can be expressed as:

[0050] i t =σ(W i [h t-1 ,x t ]+b i )

[0051] f t =σ(W f [h t-1 ,x t ]+b f )

[0052] o t =σ(W o [h t-1 ,x t ]+b o )

[0053]

[0054]

[0055] h t =o t ⊙tanh(c t )

[0056] The definition of the above expression is as follows: t is the input gate, f t For the forget gate, t is the output gate, is the candidate unit gate, c t is the unit state, h t is the hidden state. In addition, c t and h t are the unit and hidden state of the previous time step respectively. Further, W i , W f , W o , W c are the weights of each gate, b i 、b f 、b o 、b c are the biases of each gate, σ represents the Sigmoid function, which is a nonlinear transformation function, and ⊙ represents the Hadamard product, which is the element-by-element multiplication of two matrix elements. att ={x att,1 ,x att,2 ,…,x att,n}, the LSTM unit captures the dynamic pattern of each time step. The hidden state H = {h1,h2,…,h n} can be regarded as X attOn this basis, another self-attention layer is introduced to weight the dynamic information sequence H and generate a weighted dynamic information sequence H att . Further, a set of fully connected layers is used to att The time dimension is compressed to generate a dynamic information sequence H att A dynamic feature V with generalized properties, whose shape is 1×V, and a set of weights W learned using feature attention a Weight this dynamic feature to obtain the weighted dynamic feature V att :

[0057] V att =softmax(W a )V

[0058] The weighted mean feature M att and dynamic characteristics V att The concatenation along the feature dimension is the output Z of the encoder part:

[0059] Z = concat[M att ,V att ]

[0060] In this embodiment, in order to avoid the autoencoder falling into a local optimal solution during the training process, the weighted dynamic features V extracted by the encoder are att and process variable mean characteristic M att There is a significant information overlap problem between them. In the decoder, not only the weighted dynamic feature V att , and also introduces the process variable weighted mean feature M att Participate in sequence reconstruction to obtain reconstructed sequence This guides the encoder to learn a more comprehensive time series representation and effectively obtain dynamic and static information in the time series. The decoder first expands the encoder output Z with a shape of 1×(V+D) to a reconstructed dynamic information sequence with a shape of L×(V+D) through a set of fully connected layers. Then, through a set of LSTM networks, the reconstructed dynamic information sequence is restored to a reconstructed process variable time series with a shape of L×D Generally speaking, the structure of the fully connected layer and LSTM network used in the decoder should be symmetrical with the network structure in the encoder. It has important reference value for parameter training of time series feature autoencoder network.

[0061] In practical applications, various time series can be extracted through parallel feature extraction of multiple time series feature extraction autoencoder networks to complete the task of spatiotemporal feature extraction of process industrial processes.

[0062] The present application provides a temporal feature extraction autoencoder network, which can obtain mean features and dynamic features weighted by an attention mechanism from the input time series, thereby maximally and comprehensively extracting dynamic and static information in the time series to improve the accuracy of soft measurement of process industry performance indicators.

[0063] The above is an introduction to a temporal feature extraction autoencoder network provided by the present application. Before the spatiotemporal feature extraction network and the feature fusion network are deployed in the data service layer of the digital twin platform and used for soft measurement of performance indicators, the above networks need to be trained to ensure that the parameters of each network can be adapted to the corresponding performance indicator soft measurement tasks. The design of the network parameter training method has a very important influence on whether the trained network can achieve spatiotemporal feature extraction and fusion. The following introduces a fast and efficient method for synchronous training of temporal feature extraction and feature fusion network parameters provided by the present application as an embodiment.

[0064] Figure 4 This is a flowchart of a synchronous training method for time series feature extraction and feature fusion network parameters provided by the present application as an embodiment. This training method is mainly used for synchronous training of time series feature extraction and feature fusion network parameters, thereby reducing the problems of high training cost, poor integrity, information loss, etc. caused by separate network training and parameter adjustment, and strengthening the network's spatiotemporal feature extraction capability and improving the effect of soft measurement of performance indicators. Figure 4 , the method may include the following steps:

[0065] 201. Determine the data duration and required process variables required for network parameter training, and obtain process variable data for each process step of network parameter training.

[0066] In this embodiment, the parameters of the training network often need to obtain sufficient data, so as to fully consider the diversity and complexity of the process industry operation data and improve the generalization ability of the network. Before obtaining data, it is necessary to first determine the data range for network parameter training, including the length of the training data and the required process variables. On this basis, process variable data with a certain time length that can cover each process is obtained.

[0067] In practical applications, data can be obtained by reading and downloading through API access, database query, file system access, Web service access, or loading through physical data storage media. In addition, the collected data can be pre-processed through data cleaning, missing value filling, outlier processing, and data filtering to improve data quality and optimize the effect of network parameter training.

[0068] 202. Scale and group the process variable data, and construct a network parameter training data set.

[0069] In this embodiment, the scale of the data needs to be integrated first, so that the data used for network parameter training is under the same time scale. Then, based on the process of the process industry, the acquired data is grouped to match the corresponding input for each time series feature extraction network. On this basis, the parameter training data set of each time series feature extraction network is constructed respectively. In this embodiment, the data set refers to the collection of data used for training, verifying and evaluating the network. The method for constructing the training data set is: first, a fixed time window size is set, based on each performance indicator label value, with the label value timestamp as the starting point and the time window size as the range, each group of process variable time series is generated through each group of data. Then, according to a certain ratio, the performance indicator label value data points and their corresponding time series are divided into a training set, a verification set and a test set. In the process of network parameter training, the training set is usually used to train the parameters of the network, the performance of the network in the training process is verified by the verification set, and the performance and generalization ability of the network are evaluated by the test set.

[0070] In some optional embodiments, in order to avoid the problem that the order of magnitude of each variable in the process variable time series is not uniform, thereby affecting the effect of network parameter training, the data in the data set can be normalized, standardized, etc. Among them, standardization refers to the operation of adjusting the mean of each variable data to 0 and the variance to 1, while normalization is the process of scaling the eigenvector of each data point to the unit norm. Both can effectively unify the data sets of each process variable, increase the convergence of the model, and avoid the occurrence of overfitting problems.

[0071] 203. Connect the spatiotemporal feature extraction network including multiple temporal feature extraction networks and the feature fusion network front and back to build an end-to-end network structure.

[0072] In this embodiment, various types of networks are first instantiated through a series of programming techniques, and then a connection is established between the encoder output layer of each time series feature extraction network and the input layer of the feature fusion network, and then all networks are combined into an end-to-end structure to realize the function of network input time series and directly output performance indicator soft measurement.

[0073] 204. Use the semi-supervised loss function to train the end-to-end network parameters and upload the training results.

[0074] In this embodiment, the parameters of the constructed end-to-end network are adjusted by a network training method. For example, the training of network parameters is divided into multiple rounds, and in each round, the data in the training set and the validation set are divided into multiple batches and input into the network, and under the guidance of the semi-supervised loss function, the network parameters are guided to be updated in the direction of minimizing the loss function, so that the network has the ability to extract spatiotemporal features and soft-measure the corresponding performance indicators.

[0075] In this embodiment, the semi-supervised loss function refers to a composite loss function that simultaneously considers the regression accuracy of the performance indicator label value, the reconstruction accuracy of the input time series, and the distance between the features extracted by each time series feature extraction network. By introducing the semi-supervised loss function to train the network parameters, unlabeled data can be effectively utilized, thereby providing additional information for the soft measurement of performance indicators and improving the accuracy and robustness of the soft measurement results. The expression of the semi-supervised loss function is:

[0076]

[0077] In this expression, the semi-supervised loss function is mainly composed of three parts: regression loss, sequence reconstruction loss and feature distance loss. Among them, B represents the size of the batch used in the training process, N represents the feature dimension of each time series input into a single time series feature extraction network, and Respectively represent the values ​​of the input and reconstructed time series at the bth batch, the lth duration, and the nth feature dimension, Y b and represents the value of the real and regressed performance indicator labels at the bth batch, and It represents the feature distance loss, which aims to maximize the distance between the temporal features extracted by different networks, so as to better represent the spatial relationship between these features. In addition, regularization parameters λ1, λ2 and λ3 are introduced to adjust the influence strength of each part of the loss function.

[0078] In some optional embodiments, The calculation of can be performed by the following method: for each group of features extracted from different time series feature extraction networks, first, the kernel principal component analysis method is used to extract the principal component features with consistent dimensions from the features of all groups, and then the Euclidean distance is used to calculate the distance between the two principal component features, sum them up and take the reciprocal:

[0079]

[0080] Among them, A ijand B ij Respectively represent the values ​​of the two sets of time series features on the jth feature dimension in the i-th data point.

[0081] In this embodiment, in each round of network parameter training, an optimizer can be used to update the network parameters. The optimizer can use various algorithms, such as gradient descent, Adam algorithm, Autograd algorithm, etc. The selection of a suitable optimizer depends on the specific task, data, and model structure. The usual practice is to try multiple optimizers and then select the optimizer with the best performance based on the performance of the network on the test set.

[0082] In this embodiment, the training results of the network need to be uploaded and stored for real-time loading, calling and soft measurement of performance indicators of the digital twin platform. The training results are composed of multiple parts, such as the parameters of various networks, the input time series length of each time series feature extraction network, the process variables contained in each time series and other configuration information. In some optional embodiments, it also includes the configuration of the processing method for the data set, such as normalization, standardization and other methods. The training results can be uploaded to the data service layer, or uploaded to other places where the training results can be saved, and loaded and called through certain channels when they need to be called.

[0083] The present application provides a method for synchronously training parameters of time series feature extraction and feature fusion networks. By introducing a semi-supervised loss function to train the parameters of an end-to-end network, the parameters of each time series feature extraction network and feature fusion network can be quickly and efficiently trained synchronously, eliminating the training cost of training networks one by one on the process industry digital twin platform. At the same time, it also realizes the global optimization of network parameters during the training process, avoiding the problems of poor integrity and information loss caused by separately trained networks in performance indicator soft measurement applications.

[0084] It should also be noted that the terms "such as", "including" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, the elements defined by the sentences "such as...", "including..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0085] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A soft sensing method for spatiotemporal feature fusion of a process industry digital twin platform, characterized in that: include: Through the data perception module located in the data perception layer of the process industry digital twin platform, the operation data of the process variables of each production process of the process industry are obtained respectively and uploaded to the data integration layer, wherein the data perception module refers to a collection of methods that can obtain the operation status data of each process of the process industry, including various types of physical sensors, soft sensors, and some non-real-time processes of manual testing, detection, and data uploading. The process variables refer to various physical quantities and parameters measured, monitored and / or controlled in the industrial process, including temperature, pressure, flow, concentration obtained by direct measurement, or energy consumption, thermal efficiency, reaction rate calculated by other parameters, or obtained by manual testing and detection, including the chemical composition and properties of products or raw materials; In the data integration layer, the acquired process variable data are scaled and grouped, and a certain time window size is set to generate corresponding time series data, wherein the scale integration refers to a given sampling interval, and the uploaded data is resampled through this sampling interval, so that the update frequency of all data is on the same time scale, the data grouping refers to dividing the process variable data into multiple groups based on the data collection process, and the generation of time series data refers to the process of taking the current timestamp as the starting point and a time window size as the span, selecting the process variable data within this range, and splicing each process variable data point in chronological order, thereby forming the time series data; In the data service layer, the time series data of each process is read, and the time and space features contained in the time series of the process variables of each process are extracted in parallel through a distributed time series feature extraction network, wherein the time series feature extraction network is an autoencoder network composed of an encoder and a decoder, which is used to extract one-dimensional time series features from two-dimensional time series data from one process, and multiple time series feature extraction networks extract features from multiple processes in parallel to obtain the time and space features of the entire process industry; The spatiotemporal features extracted from each process are fused through a feature fusion network to obtain a soft measurement result of the performance index, wherein the feature fusion network uses the extracted spatiotemporal features as the input of the network, realizes the fusion of spatiotemporal features through the transformation of each level within the network, and obtains the corresponding performance index estimation value by fitting; Feedback the soft measurement results of performance indicators to the data integration layer for storage, thereby achieving long-term and stable storage of soft measurement results and forming a closed loop of data flow; The autoencoder network is composed of two parts: an encoder for extracting time series features and a decoder for reconstructing input sequences. The encoder part inputs the process variable time series data, and extracts the weighted mean feature and the weighted dynamic feature through a series of network layer transformations and processing, and outputs the feature concatenation result of the two. The decoder part uses the feature concatenation result to try to reconstruct the input time series and generate a reconstructed time series. The encoder part introduces a self-attention layer to assign attention weights to the input time series X of shape L×D: Among them, d k represents the dimension of the input sequence X, X att Represents the weighted result of the input sequence, averaging it along the time dimension to obtain a weighted mean feature M with a shape of 1×D att ; Then, a set of long short-term memory (LSTM) networks is used to weight the result X att The dynamic pattern of the network is modeled, and the hidden state H = {h1, h2, ..., h n } can be regarded as X att On this basis, another self-attention layer is introduced to weight the dynamic information sequence H to generate a weighted dynamic information sequence H att ; Further, a set of fully connected layers is used to att The time dimension is compressed to generate dynamic features V, whose shape is 1×V, and the feature attention weight W is used a Weight this dynamic feature to obtain the weighted dynamic feature V att : V att =softmax(W a )V The weighted mean feature M att and dynamic characteristics V att Concatenate along the feature dimension to get the output Z of the encoder part: Z=concat[M att ,In att ]; The decoder part also uses the weighted dynamic feature V att and weighted mean feature M att To reconstruct the sequence, the decoder expands the encoder output Z with a shape of 1×(V+D) into a reconstructed dynamic information sequence with a shape of L×(V+D) Then, through a set of LSTM networks, the reconstructed dynamic information sequence is restored to a reconstructed process variable time series with a shape of L×D 2. The method according to claim 1, characterized in that: A method for synchronously training time series feature extraction and feature fusion network parameters, comprising: Determine the data duration and required process variables required for network parameter training, and obtain process variable data for each process step of network parameter training; The process variable data is scaled and grouped, and a network parameter training data set is constructed, wherein the construction of the data set includes the steps of generating a time series, dividing a training set, a validation set, and a test set; Connecting a spatiotemporal feature extraction network including multiple temporal feature extraction networks to a feature fusion network front and back to construct an end-to-end network structure, wherein the construction of the end-to-end network structure refers to establishing a connection between an encoder output layer of each temporal feature extraction network and an input layer of the feature fusion network, thereby forming all networks into an end-to-end structure; The end-to-end network parameters are trained using a semi-supervised loss function, and the end-to-end network parameter training results are uploaded, wherein the semi-supervised loss function refers to a composite loss function that simultaneously considers the regression accuracy of the performance indicator label value, the reconstruction accuracy of the input time series, and the distance between the features extracted by each time series feature extraction network; the end-to-end network parameter training results refer to the parameters of various networks, the input time series length of each time series feature extraction network, and the process variables contained in each time series for configuring the relevant information of the performance indicator soft measurement model.

3. The method according to claim 2, characterized in that The expression of the semi-supervised loss function of the network parameter training method is: Among them, B represents the batch size used in the training process, N represents the feature dimension of each time series input into a single time series feature extraction network, and Respectively represent the values ​​of the input and reconstructed time series at the bth batch, the lth duration, and the nth feature dimension, Y b and represents the value of the real and regressed performance indicator labels at the bth batch, and It represents the feature distance loss, which is used to maximize the distance between the temporal features extracted by different networks, so as to better represent the spatial relationship between these features. In addition, regularization parameters λ1, λ2 and λ3 are introduced to adjust the influence intensity of each part of the loss function.

4. The method according to claim 3, characterized in that The semi-supervised loss function, feature distance loss One calculation method is: for each group of features extracted from different time series feature extraction networks, first, use the Kernel Principal Component Analysis method to extract principal component features with consistent dimensions from all groups of features, and then use the Euclidean distance to calculate the distance between the two principal component features, sum them up and take the reciprocal: Among them, A ij and B ij Respectively represent the values ​​of the two sets of time series features on the jth feature dimension in the i-th data point.

Images