Asynchronous time series data processing and predictive modeling methods, systems, devices and media

Abstract

The present invention discloses a method, system, device and medium for processing and predictive modeling of asynchronous time series data, which are mutually corresponding schemes. In the scheme: by combining the powerful spatial modeling capability of the graph attention network (GAT) and the advantages of the gated recurrent unit (GRU) in processing time series data, the spatial and temporal correlations between the complex and changeable asynchronous time series data in the multi-stage continuous production process can be fully captured, which not only significantly improves the prediction accuracy of production data, enables enterprises to more accurately predict key parameters and state changes in the production process, but also provides strong support for optimizing production control. In addition, the interpretability of the hybrid model also provides an important basis for the transparency and intelligent transformation of the production process, which helps enterprises build a more efficient, green and sustainable production system.

Description

Technical Field

[0001] The present invention relates to the technical field of predictive modeling based on data-driven multi-stage continuous production, and in particular to an asynchronous time series data processing and predictive modeling method, system, device and medium. Background Art

[0002] In the modern complex industrial production field, especially those involving multi-stage continuous production processes such as chemical, metallurgical, and pharmaceutical industries, the use of production data to build intelligent optimization based on data-driven modeling is emerging in an endless stream, which is crucial to improving production efficiency, optimizing product quality, and reducing energy consumption.

[0003] The actual production process usually includes multiple interrelated and sequential stages. Each stage generates a large amount of time series data. These data are not only continuous in time, but also correlated in space.

[0004] How to effectively use these asynchronous time series data to make accurate predictions and optimize control is an important challenge facing the current intelligent optimization of industrial production. However, no effective solution has been found yet.

[0005] In view of this, the present invention is proposed. Summary of the invention

[0006] The purpose of the present invention is to provide an asynchronous time series data processing and predictive modeling method, system, device and medium, which can effectively utilize asynchronous time series data to make accurate predictions and thus achieve optimal control of the production process.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] An asynchronous time series data processing and prediction modeling method, comprising:

[0009] Collect asynchronous time series data generated during multi-stage continuous production;

[0010] Construct a graph structure based on the production stages involved in the multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage;

[0011] Dividing the asynchronous time series data into a plurality of subsequences in combination with a graph structure, and preprocessing each subsequence;

[0012] A hybrid model is constructed by combining the graph attention network and the gated recurrent unit, each preprocessed subsequence is input into the hybrid model, the corresponding spatial features are extracted by the graph attention network, and then the gated recurrent unit combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, and the prediction index is output through the fully connected layer at the end of the hybrid model. The loss function is constructed by combining the prediction index and the data in the asynchronous time series data to train the hybrid model; wherein the data type included in the prediction index is the data type involved in the production stage;

[0013] The multi-stage continuous production data collected in real time is input into the trained hybrid model to predict the prediction indicators corresponding to the production stage at future moments.

[0014] An asynchronous time series data processing and prediction modeling system, comprising:

[0015] A data acquisition unit, used to collect asynchronous time series data generated during a multi-stage continuous production process;

[0016] A graph structure building unit, used to build a graph structure based on the production stages involved in the multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage;

[0017] A data division and preprocessing unit, used to divide the asynchronous time series data into a plurality of subsequences in combination with a graph structure, and preprocess each subsequence;

[0018] A model building and training unit is used to build a hybrid model by combining a graph attention network and a gated recurrent unit, input each preprocessed subsequence into the hybrid model, extract the corresponding spatial features through the graph attention network, and then the gated recurrent unit combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, output the prediction index through the fully connected layer at the end of the hybrid model, build a loss function by combining the prediction index and the data in the asynchronous time series data, and train the hybrid model; wherein the data type included in the prediction index is the data type involved in the production stage;

[0019] The prediction unit is used to input the multi-stage continuous production data collected in real time into the trained hybrid model to predict the prediction indicators corresponding to the production stage at the future moment.

[0020] A processing device, comprising: one or more processors; a memory for storing one or more programs;

[0021] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0022] A readable storage medium stores a computer program, which implements the above method when the computer program is executed by a processor.

[0023] It can be seen from the technical solution provided by the present invention that by combining the powerful spatial modeling capabilities of the graph attention network (GAT) and the advantages of the gated recurrent unit (GRU) in processing time series data, it is possible to fully capture the spatial and temporal correlations between the complex and variable asynchronous time series data in the multi-stage continuous production process, which not only significantly improves the prediction accuracy of production data, enabling enterprises to more accurately predict key parameters and state changes in the production process, but also provides strong support for optimizing production control. In addition, the interpretability of the hybrid model also provides an important basis for the transparency and intelligent transformation of the production process, which helps enterprises build a more efficient, green and sustainable production system. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0025] Figure 1 A flowchart of an asynchronous time series data processing and prediction modeling method provided by an embodiment of the present invention;

[0026] Figure 2 A schematic diagram of a coking production process provided by an embodiment of the present invention;

[0027] Figure 3 A schematic diagram of the coking production process provided by an embodiment of the present invention;

[0028] Figure 4 A schematic diagram of the spatial feature extraction principle provided by an embodiment of the present invention;

[0029] Figure 5 A schematic diagram of the workflow of the hybrid model provided by an embodiment of the present invention;

[0030] Figure 6 A schematic diagram of an asynchronous time series data processing and prediction modeling system provided by an embodiment of the present invention;

[0031] Figure 7 A schematic diagram of a processing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0033] First, the terms that may be used in this article are explained as follows:

[0034] The terms "include", "comprises", "contains", "has" or other descriptions with similar semantics should be interpreted as non-exclusive inclusion. For example, including certain technical feature elements (such as raw materials, components, ingredients, carriers, dosage forms, materials, dimensions, parts, components, mechanisms, devices, steps, procedures, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products or products, etc.) should be interpreted as including not only certain technical feature elements explicitly listed, but also other technical feature elements known in the art that are not explicitly listed.

[0035] The term "consisting of..." means excluding any technical feature elements not explicitly listed. If this term is used in a claim, it will make the claim closed, so that it does not contain technical feature elements other than the technical feature elements explicitly listed, except for the conventional impurities related to them. If this term only appears in a clause of a claim, it only limits the elements explicitly listed in the clause, and the elements recorded in other clauses are not excluded from the overall claim.

[0036] The asynchronous time series data processing and prediction modeling method, system, device and medium provided by the present invention are described in detail below. The contents not described in detail in the embodiments of the present invention belong to the prior art known to professional and technical personnel in the field. If no specific conditions are specified in the embodiments of the present invention, they shall be carried out according to the conventional conditions in the field or the conditions recommended by the manufacturer.

[0037] Embodiment 1

[0038] The embodiment of the present invention provides an asynchronous time series data processing and prediction modeling method, such as Figure 1 As shown, it mainly includes the following steps:

[0039] Step 1: Collect asynchronous timing data.

[0040] In the embodiment of the present invention, asynchronous time series data generated in a multi-stage continuous production process is collected, mainly including: the chemical composition of raw materials and outputs in each production stage, process parameters in each production stage, and production environment parameters.

[0041] Step 2: Build the graph structure.

[0042] In the embodiment of the present invention, a graph structure is constructed based on the production stages involved in a multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage.

[0043] The preferred implementation is as follows: each production stage and the related output are taken as nodes, and the output represents the product obtained in the related production stage; according to the correlation and dependency between the nodes, edges between different nodes are established to form a graph structure.

[0044] Step 3: Data division and processing.

[0045] In the embodiment of the present invention, the asynchronous time series data is divided in combination with a graph structure to obtain a plurality of subsequences, and each subsequence is preprocessed.

[0046] (1) Data division.

[0047] In an embodiment of the present invention, nodes in a graph structure are identified, and data corresponding to the nodes are organized into a series of ordered or related subsequences based on the positions, connections, and possible characteristic attributes of these nodes in the graph, and different subsequences contain data of all nodes at different time sequences.

[0048] (2) Data preprocessing.

[0049] In the embodiment of the present invention, preprocessing is performed on each subsequence, including: extracting corresponding data according to the node, and performing data cleaning on the extracted data to make the data format unified and the length consistent, and then normalizing the cleaned data.

[0050] Step 4: Build and train the model.

[0051] A hybrid model is constructed by combining the graph attention network (GAT) and the gated recurrent unit (GRU). Each preprocessed subsequence is input into the hybrid model, and the corresponding spatial features are extracted by the graph attention network. The gated recurrent unit then combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, and finally outputs the prediction indicators through the fully connected layer.

[0052] In an embodiment of the present invention, each preprocessed subsequence is used as an input of the hybrid model. In the graph attention network, a data mapping relationship is established according to the graph structure, and data at different production stages are associated to obtain spatial features. The spatial features of each preprocessed subsequence are passed to the gated recurrent unit together with the corresponding preprocessed subsequence. The gated recurrent unit is responsible for processing from the last preprocessed subsequence to the first preprocessed subsequence, gradually extracting and integrating the temporal features, and finally outputting the prediction indicators through the fully connected layer.

[0053] In the embodiment of the present invention, the data type included in the prediction index is the data type included in the asynchronous time series data; for example, the prediction index may include process parameters and production environment parameters of each stage.

[0054] In the embodiment of the present invention, the loss function is calculated by combining the predicted prediction index and the process parameters and production environment parameters of each production stage corresponding to the asynchronous time series data, and the parameters of the hybrid model are optimized by the loss function. Considering that the loss function and the specific parameter optimization process involved in this part can be implemented with reference to conventional technology, they will not be described in detail.

[0055] Step 5: Production forecast.

[0056] The trained hybrid model is obtained through the aforementioned step 4. When making predictions, the multi-stage continuous production data collected in real time is input into the trained hybrid model. For each node, the corresponding prediction indicators for the future moment can be predicted, namely, process parameters, production environment parameters, product quality, etc.

[0057] Preferably, after outputting the predicted values ​​of the indicators of the future production stage, they can also be compared and analyzed with the data in the actual generation process. If the difference exceeds the preset threshold, a feedback mechanism is triggered to optimize the control of the actual generation process.

[0058] The above-mentioned scheme provided by the embodiment of the present invention, by combining the powerful spatial modeling ability of the graph attention network and the advantages of the gated recurrent unit in processing time series data, can comprehensively capture the spatial and temporal correlations between the complex and changeable asynchronous time series data in the multi-stage continuous production process, which not only significantly improves the prediction accuracy of production data, enabling enterprises to more accurately predict key parameters and state changes in the production process, but also provides strong support for optimizing production control; on the one hand, through real-time prediction and dynamic adjustment of production parameters such as temperature, pressure, flow, etc., enterprises can achieve refined management of energy consumption and stable improvement of product quality, thereby enhancing market competitiveness; on the other hand, the interpretability of the hybrid model also provides an important basis for the transparency and intelligent transformation of the production process, which helps enterprises build a more efficient, green and sustainable production system.

[0059] In order to more clearly demonstrate the technical solution and technical effects provided by the present invention, the method provided by the embodiment of the present invention is described in detail with reference to specific embodiments below.

[0060] 1. Data collection.

[0061] In the embodiment of the present invention, the asynchronous time series data generated in the multi-stage continuous production process are fully collected, including but not limited to the chemical composition of the raw materials and outputs in each stage (for example, the composition of the blended coal in coking), the process parameters of each production stage (such as flue temperature, pressure, flow, etc., similar to the coke oven gas parameters), and the production environment parameters (such as humidity, gas temperature, gas concentration, etc.). At the same time, the timestamps of each parameter are recorded to ensure the time series and continuity of the data.

[0062] 2. Build the graph structure.

[0063] In the embodiment of the present invention, a complex graph structure is constructed based on the correlation and dependency of data in a multi-stage continuous production process. The nodes in the graph structure include: all production stages and related outputs, the outputs represent the products obtained in the relevant production stages (for example, coke and methanol, etc.), and the edges represent the interaction and dependency between different nodes. This graph structure can intuitively reflect the complex dynamics in a multi-stage production process.

[0064] The following is an introduction to the coking production process. Figure 2 The coking production process is shown. Figure 3 Shows an example of a graph structure. Figure 3 The nodes shown include: four production stages (in order: 1-coal blending, 2-coke oven coking, 3-chemical gas refining, 4-methanol synthesis distillation) as well as coke and methanol. The edges include material flow (how the raw coal enters the coke oven after coal blending, how the coking coal gas produced by the coke oven enters the chemical gas refining section, and how the synthesis gas obtained in the chemical gas refining section is used for methanol synthesis), energy flow (heat transfer and energy conversion relationship between stages), information flow (transmission of control signals between equipment) and dependency relationship (the degree of dependence of one production stage on another production stage), etc. Figure 2 The arrows in the processes shown also indicate the flow of material, energy or information.

[0065] 3. Data division and unified processing.

[0066] In the embodiment of the present invention, the collected asynchronous time series data (data set) is first divided. When dividing, the asynchronous time series data is divided into multiple subsequences in combination with the node information, edge information, etc. contained in the graph structure, and each subsequence contains data of all different nodes in the corresponding time series. After that, each subsequence is preprocessed separately, including data extraction (extracting useful information), data cleaning (removing noise, errors and outliers), and normalization (scaling the data to the same scale for easy model processing); the following mainly introduces the preprocessing process in detail.

[0067] 1. Data extraction.

[0068] In the embodiment of the present invention, the data in the subsequence is extracted in combination with the graph structure to obtain relevant data of each node as the node feature.

[0069] Taking the coking production process as an example, the data extracted from each production stage include: the type and proportion of raw coal in the coal blending section, the ash content, sulfur content and other chemical components of the raw coal, and the coal blending process parameters (such as temperature and humidity control); the type of coke oven (such as top-loading coke oven, side-loading coke oven), coking temperature curve, and the composition of coke gas produced in the coke oven coking section; the coke gas treatment process (such as cooling and purification), the type and output of chemical products (such as tar, ammonia, etc.), gas composition analysis (such as hydrogen, carbon monoxide, methane, etc.) in the chemical production gas refining section; the raw gas composition, catalyst type and performance, synthesis reaction conditions (such as temperature, pressure), distillation process parameters (such as the number of plates, reflux ratio) in the methanol synthesis distillation section; the particle size, strength, and chemical composition of coke (such as fixed carbon, ash, volatile matter) in the product; the purity of methanol in the product, the yield and energy consumption of methanol, etc. All the data involved in this part belong to the data in the aforementioned data set, so they are not repeated.

[0070] 2. Data cleaning.

[0071] In an embodiment of the present invention, the extracted data set is cleaned to unify the data formats from different data sources; and missing values ​​in the data are checked and corrected. For data with shorter time series in the data set, the data is supplemented by interpolation, and the average value of the existing data is inserted to ensure that the time length between each segment of data is the same.

[0072] 3. Data normalization.

[0073] In the embodiment of the present invention, the data set is normalized to ensure the consistency and comparability of the same type of data between different data sources. For real-time data streams, fixed usage parameters (such as minimum value, maximum value, mean value, standard deviation) are set through the initial data set. In the dynamic data set, outliers in the data are checked and corrected to avoid instability of normalization results due to data fluctuations.

[0074] After being preprocessed in the above manner, each subsequence can be used as hybrid model input data for model training.

[0075] 4. Build the model and train it.

[0076] In an embodiment of the present invention, a graph attention network (GAT) and a gated recurrent unit (GRU) are combined to construct a GAT-GRU hybrid model. The model uses the GAT layer to capture the complex relationship information between various parameters and equipment in the multi-stage continuous production process, and uses the GRU layer to process time-based multi-stage time series data to achieve comprehensive prediction of asynchronous time series data for the entire production process. The GAT-GRU hybrid model is trained using historical multi-stage continuous production time series data as a training set, and the model parameters are continuously optimized through the back-propagation algorithm to reduce prediction errors and improve prediction accuracy.

[0077] 1. Graph Attention Network.

[0078] In an embodiment of the present invention, the graph attention network establishes a data mapping relationship based on the previously constructed graph structure to determine how to associate data from different work sections. For the information with time dimension consistency and continuity that has been mapped, it is input into the gated recurrent unit.

[0079] In an embodiment of the present invention, in a graph attention network, nodes in each preprocessed subsequence learn spatial features through a graph attention layer, capture the complex interaction relationship between nodes, and obtain corresponding spatial features. The graph attention network includes: multiple single-layer graph attention layers arranged in sequence. A single-layer graph attention layer is used as an example for introduction below.

[0080] The input of each single-layer graph attention layer contains N nodes and F features, recorded as:

[0081]

[0082] in, are the characteristics of all nodes, For Node The features of, which contain F feature data, When the graph attention layer is the first layer of the graph attention network, the node features are the data in the preprocessed subsequence. When the graph attention layer is not the first layer of the graph attention network, the node features are the new features of the nodes output by the previous layer.

[0083] In order to convert the input features into high-dimensional features, a shared linear transformation method is used for each point. The weight matrix is ​​used to parameterize the linear transformation.

[0084] In the embodiment of the present invention, the adjacency matrix is ​​constructed based on GAT, and the graph structure is represented as an adjacency matrix. The features of each node are linearly transformed and mapped to the attention score space. The linear transformation of is expressed as:

[0085] ;

[0086] in, For Node The weight matrix of is the adjacency matrix of the graph structure, Representation Node With Node The association between nodes (i.e., whether there is an edge). With Node If connected, ,otherwise, , .

[0087] The weights of the connections between nodes are calculated through a multi-head attention mechanism, and K groups of parallel attention mechanism calculations are performed on the nodes. Each group uses an independent weight matrix and attention mechanism function. The attention weight calculation formula is as follows:

[0088] ;

[0089] Among them, the node is obtained through the above standardization With Node The attention weight between , For Node With Node The graph attention mechanism coefficient between For Node With Node The graph attention mechanism coefficient between For Node The set of neighbor nodes of and All nodes Neighbor nodes of With Node The graph attention mechanism coefficient between , is the attention mechanism function, obtained by inner product, concatenation and full connection layer. For Node The weight matrix of node Neighbor nodes The graph attention mechanism coefficient between , For Node The weight matrix of .

[0090] Based on the feature aggregation of the multi-head attention mechanism, the node New features :

[0091] ;

[0092] in, Calculate the nodes obtained for the kth group With Node The attention weights between is the weight parameter for the kth group calculation, For Node Features, is an activation function, for example, a ReLU (rectified linear unit) function.

[0093] All nodes in each single-layer graph attention layer are processed equally to obtain new features of all nodes, which are used as the input of the next single-layer graph attention layer. The new features of all nodes obtained by the last single-layer graph attention layer are combined to form the spatial features corresponding to the corresponding subsequence.

[0094] like Figure 4 As shown, the above graph attention network is introduced to extract spatial features. Figure 4 The x1~x6 in represents 6 nodes, and the normalized exponential function (Softmax) is the normalization process of the aforementioned attention weight calculation.

[0095] 2. Gated recurrent unit.

[0096] In the embodiment of the present invention, the gated recurrent unit network structure is designed, including the design of the update gate and the reset gate, the size and number of hidden layers, etc. The gated recurrent unit GRU is a variant of the long short-term memory network (LSTM). Based on the LSTM, the GRU is optimized from 3 gates to a 2-gate structure of a reset gate and an update gate. The state vector and output vector inside the LSTM are fused to become the hidden state vector in the GRU. The flow of information is controlled through these two gated units (i.e., the update gate and the reset gate), thereby solving the long-term dependency and gradient disappearance problems.

[0097] Among them, the update gate determines how much information of the hidden state at the previous moment is retained in the hidden state at the current moment; the reset gate determines how much information of the hidden state at the previous moment is written into the candidate hidden state; the hidden layer generates the candidate state by combining the current input and the reset hidden state at the previous moment.

[0098] 3. Model training plan.

[0099] like Figure 5As shown, a schematic diagram of the model workflow is provided. A data set is formed by collecting asynchronous time series data generated in a multi-stage continuous production process. The structural graph structure is divided into L subsequences. Each subsequence is preprocessed separately and used as the input of the hybrid model. On the one hand, each preprocessed subsequence is mapped through the graph attention network to obtain spatial features. On the other hand, each preprocessed subsequence is also input into the gated recurrent unit. The gated recurrent unit receives the preprocessed subsequences and the corresponding spatial features, starting from the last subsequence (i.e., subsequence L) to the first subsequence (i.e., subsequence 1), and gradually extracts and integrates the temporal features. In addition to the graph attention network and the gated recurrent unit, the hybrid model also includes a fully connected layer at the tail. The temporal feature passes through the fully connected layer to obtain the prediction index value x pre .

[0100] In an embodiment of the present invention, the prediction index refers to the parameters of the corresponding production stage, such as process parameters and production environment parameters. The loss function is calculated using the prediction index and the values ​​of the corresponding parameters in the asynchronous time series data, and the network parameters are optimized through the back propagation algorithm and the RMSprop (root mean square propagation) optimization algorithm to improve the prediction accuracy and convergence speed.

[0101] For example, MSE (mean square error) can be selected to calculate the loss function, and the model parameters can be adjusted to improve the generalization ability and prediction accuracy of the model.

[0102] In addition, the asynchronous time series data can be divided into training set and test set according to a set ratio (for example, according to a ratio of 8:2); after the training is completed using the training set, the test set is divided and preprocessed in the same way, and then input into the trained hybrid model, and the performance of the trained hybrid model is evaluated according to the output prediction indicators; for example, MAE (mean absolute error), RMSE (root mean square error), MAPE (mean absolute error) and (coefficient of determination) etc. are used as evaluation indicators to quantify the prediction accuracy of the hybrid model.

[0103] The specific training process involved in this part can be implemented with reference to conventional techniques and will not be elaborated in the present invention.

[0104] 5. Predictive applications.

[0105] The trained hybrid model is applied to real-time multi-stage continuous production data prediction. Specifically, the preprocessed real-time data is input into the trained hybrid model to output the predicted values ​​of indicators (process parameters, production environment parameters, product quality, etc.) for a specified production stage in the future.

[0106] 6. Optimize control.

[0107] In the embodiment of the present invention, key parameters in the production process, such as raw material ratio, temperature, pressure, flow rate, etc., can be adjusted in time according to the prediction results (indicator prediction values) to achieve optimized control of the production process. In the specific implementation process, an intelligent visual monitoring system can be set up to compare and analyze the prediction results with the corresponding types of data in the actual production process. Through the independent analysis of existing information data by engineers, it helps to discover and solve existing problems and assist in timely adjustment of strategies to make better decisions; when the prediction error exceeds the preset threshold, the detection system feedback mechanism is triggered, and the raw material ratio, heating temperature curve, and gas circulation in the furnace are adjusted according to the feedback results to achieve refined control of the coking process and improve the quality and production efficiency of coking production products. This optimization control not only helps to improve product quality, but also effectively reduce energy consumption, improve production efficiency, and provide strong support for the sustainable development of enterprises.

[0108] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by means of software plus necessary general hardware platforms. Based on such understanding, the technical solutions of the above embodiments can be embodied in the form of software products, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0109] Embodiment 2

[0110] The present invention also provides an asynchronous time series data processing and prediction modeling system, which is mainly used to implement the method provided in the above embodiment, such as Figure 6 As shown, the system mainly includes:

[0111] A data acquisition unit, used to collect asynchronous time series data generated during a multi-stage continuous production process;

[0112] A graph structure building unit, used to build a graph structure based on the production stages involved in the multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage;

[0113] A data division and preprocessing unit, used to divide the asynchronous time series data into a plurality of subsequences in combination with a graph structure, and preprocess each subsequence;

[0114] A model building and training unit is used to build a hybrid model by combining a graph attention network and a gated recurrent unit, input each preprocessed subsequence into the hybrid model, extract the corresponding spatial features through the graph attention network, and then the gated recurrent unit combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, output the prediction index through the fully connected layer at the end of the hybrid model, build a loss function by combining the prediction index and the data in the asynchronous time series data, and train the hybrid model; wherein the data type included in the prediction index is the data type involved in the production stage;

[0115] The prediction unit is used to input the multi-stage continuous production data collected in real time into the trained hybrid model to predict the prediction indicators corresponding to the production stage at the future moment.

[0116] Considering that the main technical details involved in the system have been introduced in the previous embodiments, they will not be repeated here.

[0117] Technical personnel in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0118] Embodiment 3

[0119] The present invention also provides a processing device, such as Figure 7 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the methods provided in the aforementioned embodiments.

[0120] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0121] In the embodiment of the present invention, the specific types of the memory, input device and output device are not limited; for example:

[0122] The input device may be a touch screen, an image acquisition device, a physical button or a mouse, etc.;

[0123] The output device may be a display terminal;

[0124] The memory may be a random access memory (RAM) or a non-volatile memory such as a disk storage.

[0125] Embodiment 4

[0126] The present invention also provides a readable storage medium storing a computer program, which implements the method provided in the above embodiment when the computer program is executed by a processor.

[0127] In the embodiment of the present invention, the readable storage medium as a computer-readable storage medium can be set in the aforementioned processing device, for example, as a memory in the processing device. In addition, the readable storage medium can also be a U disk, a mobile hard disk, a read-only memory (ROM), a disk or an optical disk, etc., which can store program codes.

[0128] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An asynchronous time series data processing and prediction modeling method, characterized in that: include: Collect asynchronous time series data generated during multi-stage continuous production; Construct a graph structure based on the production stages involved in the multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage; Dividing the asynchronous time series data into a plurality of subsequences in combination with a graph structure, and preprocessing each subsequence; A hybrid model is constructed by combining the graph attention network and the gated recurrent unit, each preprocessed subsequence is input into the hybrid model, the corresponding spatial features are extracted by the graph attention network, and then the gated recurrent unit combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, and the prediction index is output through the fully connected layer at the end of the hybrid model. The loss function is constructed by combining the prediction index and the data in the asynchronous time series data to train the hybrid model; wherein the data type included in the prediction index is the data type involved in the production stage; The multi-stage continuous production data collected in real time is input into the trained hybrid model to predict the prediction indicators corresponding to the production stage at future moments.

2. The asynchronous time series data processing and prediction modeling method according to claim 1, characterized in that: The asynchronous time series data generated in the multi-stage continuous production process include: the chemical composition of raw materials and outputs in each production stage, the process parameters and production environment parameters in each production stage; the data types included in the corresponding prediction indicators include: process parameters and production environment parameters.

3. The asynchronous time series data processing and prediction modeling method according to claim 1, characterized in that: The graph structure constructed based on the production stages involved in the multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage includes: Each production stage and the related output are taken as nodes, and the output represents the products obtained in the related production stage; according to the correlation and dependency between the nodes, the edges between different nodes are established to form a graph structure.

4. The asynchronous time series data processing and prediction modeling method according to claim 1, characterized in that: The combining graph structure divides the asynchronous time series data into a plurality of subsequences, and preprocesses each subsequence, including: According to the graph structure, the data corresponding to the nodes are organized into a series of ordered or related subsequences. Different subsequences contain the data of all nodes at different time series. When preprocessing each subsequence, the data in the subsequence is extracted in combination with the graph structure to obtain the relevant data of each node. The extracted data is cleaned to make the data format unified and the length consistent, and then the cleaned data is normalized.

5. The asynchronous time series data processing and prediction modeling method according to claim 1, characterized in that: The corresponding spatial features are extracted by the graph attention network, and then the gated recurrent unit combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, including: Each preprocessed subsequence is used as an input of the hybrid model. In the graph attention network, a data mapping relationship is established according to the graph structure, and the data at different production stages are associated to obtain spatial features. The spatial features of each preprocessed subsequence are passed to the gated recurrent unit together with the corresponding preprocessed subsequence. The gated recurrent unit is responsible for processing from the last preprocessed subsequence to the first preprocessed subsequence, and gradually extracting and integrating the temporal features.

6. The asynchronous time series data processing and prediction modeling method according to claim 1 or 5, characterized in that: The graph attention network comprises: a plurality of single-layer graph attention layers arranged in sequence; The input of each single-layer graph attention layer contains the feature data of N nodes, recorded as: ； in, are the characteristics of all nodes, For Node The features of, which contain F feature data, ; When the graph attention layer is the first layer of the graph attention network, the node features are the data in the preprocessed subsequence. When the graph attention layer is not the first layer of the graph attention network, the node features are the new features of the nodes output by the previous layer. The graph structure is represented as an adjacency matrix, and the features of each node are linearly transformed and mapped to the attention score space; for the node The linear transformation of is expressed as: ； in, For Node The weight matrix of is the adjacency matrix of the graph structure, Representation Node With Node The correlation between nodes With Node If connected, ,otherwise, , ; The weights of the connections between nodes are calculated through a multi-head attention mechanism, and K groups of parallel attention mechanism calculations are performed on the nodes. Each group uses an independent weight matrix and attention mechanism function. The attention weight calculation formula is as follows: ； in, To the node With Node The attention weights between For Node With Node The graph attention mechanism coefficient between For Node With Node The graph attention mechanism coefficient between For Node The set of neighbor nodes of and All nodes Neighbor nodes, nodes With Node The graph attention mechanism coefficient between , is the attention mechanism function, For Node The weight matrix of the node Neighbor nodes The graph attention mechanism coefficient between , For Node The weight matrix of Based on the feature aggregation of the multi-head attention mechanism, the node New features : ； in, Calculate the nodes obtained for the kth group With Node The attention weights between is the weight parameter for the kth group calculation, For Node Features, is the activation function; All nodes in each single-layer graph attention layer are processed equally to obtain new features of all nodes, which are used as the input of the next single-layer graph attention layer. The new features of all nodes obtained by the last single-layer graph attention layer are combined to form the spatial features corresponding to the corresponding subsequence.

7. The asynchronous time series data processing and prediction modeling method according to claim 1, characterized in that: The method further includes: After obtaining the predicted indicators of the production stage at the future moment, they are compared and analyzed with the corresponding types of data in the actual generation process. If the gap exceeds the preset threshold, the feedback mechanism is triggered to optimize the actual generation process.

8. An asynchronous time series data processing and prediction modeling system, characterized in that: include: A data acquisition unit, used to collect asynchronous time series data generated during a multi-stage continuous production process; A graph structure building unit, used to build a graph structure based on the production stages involved in the multi-stage continuous production process, the data involved in each production stage, and the correlation and dependency between each production stage; A data division and preprocessing unit, used to divide the asynchronous time series data into a plurality of subsequences in combination with a graph structure, and preprocess each subsequence; A model building and training unit is used to build a hybrid model by combining a graph attention network and a gated recurrent unit, input each preprocessed subsequence into the hybrid model, extract the corresponding spatial features through the graph attention network, and then the gated recurrent unit combines the spatial features and the corresponding preprocessed subsequences to gradually integrate the temporal features, output the prediction index through the fully connected layer at the end of the hybrid model, build a loss function by combining the prediction index and the data in the asynchronous time series data, and train the hybrid model; wherein the data type included in the prediction index is the data type involved in the production stage; The prediction unit is used to input the multi-stage continuous production data collected in real time into the trained hybrid model to predict the prediction indicators corresponding to the production stage at the future moment.

9. A processing device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images