A neural network modeling method and device for multidimensional time series data based on time point labels

Abstract

The present invention discloses a method and device for modeling multidimensional time series data neural network based on time point labels. The present invention implements labeling of time series data through a label decoding module (F), which can realize learning of time series data with missing time point data or missing indicator data, allowing the use of fewer samples to obtain better training effects; by embedding a network structure of a series generation mechanism model in the label decoding module and the feature decoding module, feature extraction that allows for missing observations and training of a system state generation mechanism model are achieved in a unified neural network, and the training error is calculated at the overall network level, completing the collaborative training of the network weights of the feature extraction function and the generation mechanism model function, achieving adaptation of the feature extraction network and the generation mechanism model, and improving the model fitting accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of computing methods and is a method and device for neural network modeling of multidimensional time series data based on time point labels. Background Art

[0002] Establishing a method to build a neural network model for multidimensional data time series, learning the driving mechanism of the system's unit time state changes, and using the neural network model to simulate the temporal changes of the system's internal mechanism is an important goal of time series analysis in the field of artificial intelligence.

[0003] For example, modeling and predicting time-series data for biological fermentation processes is a key technology for achieving efficient biomanufacturing. Time-series data for fermentation processes refers to time-tagged sequence data collected at various time points during the fermentation process. Each time point includes a fixed set of metrics, including metabolome data, fermentation process data, and transcriptome data. The goal of modeling and predicting time-series data for fermentation processes is to explore fermentation patterns through simulation and optimize the fermentation process to achieve greater economic and social benefits.

[0004] When learning model parameters, these methods typically use a batch of multi-dimensional system state change curves as a learning case, and require that all state indicators in all learning cases are present. In practical applications, many system state tracking observations cannot achieve zero missing indicators or time points due to sample detection issues. This incomplete time series data cannot be learned, resulting in a waste of learnable data. The typical solution to missing data is interpolation. Common interpolation methods generally analyze and interpolate single-dimensional data, ignoring the interdependencies between multi-dimensional data. Furthermore, when interpolation methods are used independently of the model learning method to fill in missing data, the mechanism for filling in the data is difficult to reconcile with the mechanism driving time series changes. This results in systematic errors in the filled data and cannot effectively support high-precision modeling of time series mechanisms.

[0005] In the application scenarios of multidimensional time series data modeling, time series data is often high-dimensional and small-sample data. This means that the data has high dimensionality, but the absolute number of samples is small or far smaller than the data dimensionality. Furthermore, the data in all dimensions do not vary independently, but rather exhibit strong correlations between them. The data dimensionality is far greater than the number of independent factors that determine data variance. Models that directly use the data dimensionality as input are complex and prone to the "curse of dimensionality," where computational complexity increases significantly with increasing dimensionality. This increase in dimensionality also increases the sample size required for model training. When the amount of actual observed data is insufficient, overfitting can occur, where the model only memorizes the characteristics of the training data itself rather than fitting the inherent connections between them, reducing the model's generalization ability. To eliminate or mitigate the curse of dimensionality and improve the generalization ability of learning models, extracting features from the observed data and reducing its dimensionality are key tasks in multidimensional time series data modeling.

[0006] Current dimensionality reduction methods are typically performed independently as a pre-processing step in the modeling of multidimensional time series data. These methods primarily include principal component analysis, independent component analysis, filtering analysis, and factor analysis. As a pre-processing step, the dimensionality reduction process is independent of the model learning process. The goal of dimensionality reduction methods is to extract characteristic factors that contain key information about the data under specific statistical significance, such as those that determine indicator variation or independence. However, these characteristic factors do not necessarily correspond to the key factors driving the mechanisms of time series variation and therefore cannot effectively support the modeling of time series mechanisms within the low-dimensional characteristic factor space. Furthermore, these pre-processing dimensionality reduction methods require the use of post-processing methods to convert the characteristic factors into observations. This conversion process can result in error amplification, whereby small errors in the characteristic factors can lead to large errors in the predicted observed data.

[0007] In summary, current high-dimensional small sample time series modeling with missing data has the following shortcomings:

[0008] 1) Time series data with missing time point data or missing indicator data cannot be learned;

[0009] 2) The large dimension of the observed data leads to high model complexity, which can easily lead to the "curse of dimensionality", that is, as the dimension increases, the computational complexity increases significantly;

[0010] 3) Missing data filled in by interpolation methods independent of model learning methods may not be consistent with the mechanism driving time series changes and cannot well support the modeling of time series mechanisms;

[0011] 4) The feature factors extracted by feature extraction methods independent of model learning methods do not necessarily correspond to the key factors driving the mechanism of time series changes, and cannot well support the modeling of temporal mechanisms in the low-dimensional feature factor space;

[0012] 5) The post-processing method of converting characteristic factors into observation values may have the problem of error amplification during the conversion process, that is, a small error in the characteristic factor may lead to a large error in the predicted observation data.

[0013] 6) Most time series modeling methods cannot directly correspond to the generation mechanism of time series and cannot provide clues on how to optimize the generation mechanism of time series.

[0014] To address these issues, we propose a method for modeling time series data of multidimensional system state changes within a feature space that allows for missing data. This method uses a label decoding module to convert observation data into feature data, enabling the generative mechanism network to learn and model the time series generative mechanism within the feature space. By integrating the label decoding module with the tandem generative mechanism module, we achieve the goal of both feature extraction and time series generative mechanism learning within the same framework for observation data with missing values. This approach addresses the overfitting issues associated with high model complexity through dimensionality reduction, while also addressing the need for interpolation and feature extraction to align with the learning objectives of the mechanism and the error amplification associated with the conversion from feature data to observation data.

[0015] The neural network method described above can be used to model the working mechanism of a biofermentation system. This neural network model of the fermentation system's working mechanism can calculate the system state after a unit of time based on the current system state, targeting the multi-omics time series of the fermentation process (a type of high-dimensional, small-sample data that may contain missing data). Compared to general numerical fitting models of time series state changes, this model of the driving mechanism of time series state changes in biofermentation systems can better reflect the impact of various regulatory mechanisms of engineered bacteria on the fermentation process, better guide the genomic modification of engineered bacteria, and achieve better production performance. Summary of the Invention

[0016] The present invention discloses a method and apparatus for neural network modeling of multidimensional time series data based on time point labels. The method and apparatus are for modeling multidimensional state change time series data of a system in a feature space that allows for missing data. The present invention is implemented using the following technical solutions:

[0017] A neural network modeling method for multidimensional time series data based on time point labels, comprising the following steps:

[0018] 1) Standardize and organize the original observation data to obtain standardized observation data in a unified form;

[0019] 2) Organizing training examples for artificial neural network training based on canonical observation data;

[0020] 3) Design the structure of artificial neural network and establish artificial neural network model;

[0021] 4) using the training examples in step 2) to train the artificial neural network model established in step 3) to obtain a parameter matrix of the artificial neural network;

[0022] 5) using the training examples in step 2), evaluating the effects of the parameters used in the process of designing the artificial neural network structure and the parameters used in the process of training the established artificial neural network model on the accuracy of the obtained artificial neural network model, and selecting the optimal artificial neural network model under different parameter combinations as the final result model;

[0023] 6) Based on the final result model, an encoder module is established to convert the observation data into feature data;

[0024] 7) Use the encoder module to construct a chimeric prediction structure for predicting system state changes at different time intervals; use the chimeric prediction structure to evaluate model accuracy based on test examples and make predictions about system state changes in real application scenarios;

[0025] The neural network structure designed in step 3) has the following characteristics:

[0026] Its basic structure is a chimeric training structure (F-(G) n -D) neural network; the chimeric training structure is composed of a label decoding module (F), a tandem generation mechanism module (G) n ), feature decoding module (D); the label decoding module and feature decoding module respectively realize the conversion of label data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space;

[0027] A neural network unit in the serial generation mechanism module can model the working mechanism of the system's transformation from the current state to the state after a unit time. Such a network unit is called a unit time positive change generator (G unit). By connecting G units in series, it is possible to model the changes that occur in the system after multiple unit times. The serial generation mechanism module ((G)) is composed of multiple G units in series. n );

[0028] The chimeric predicted structure in step 7) has the following characteristics:

[0029] The chimeric prediction structure is composed of the encoder module (E), the tandem generation mechanism module (G) n), and feature decoding module (D) are sequentially embedded in the system; the encoder module (E) and feature decoding module (D) respectively realize the conversion from observation data to feature data and from feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space.

[0030] As a further improvement, the chimeric training structure of the present invention is established as follows: the chimeric training structure (F-(G) n -D) is composed of the label decoding module (F), the series generation mechanism module ((G) n ), feature decoding module (D), the number of neuron nodes in the input layer of the label decoding module is the total dimension of the training data for all time periods, the number of neuron nodes in the output layer is the dimension of the feature data, and the serial generation mechanism module ((G) n ) is obtained by connecting multiple G units in series. The number of neuron nodes in each layer of the G unit is the dimension of the feature data. The number of neuron nodes in the input layer of the feature decoding module is the dimension of the feature data, and the number of neuron nodes in the output layer is the dimension of the observation data.

[0031] As a further improvement, the chimeric training structure with n time intervals in the present invention is called F-(G) n -D: When n=0, FD does not contain G units, and its training data is the training example with the starting time point data and the ending time point data at the same time point; when n=1, F-(G) 1 -D contains only one G unit, and its training data is a training example with a time interval of 1; when n>1, F-(G) n -D in series generation mechanism module ((G) n ) is obtained by connecting n G units in series, and its training data is training examples with n time intervals.

[0032] As a further improvement, the training method of step 4) of the present invention is specifically as follows:

[0033] 4.1) Calculate the loss value of each chimeric training structure: For a specific chimeric training structure F-(G) n -D, only uses examples with time intervals of n to calculate the model loss, and inputs the label data of the training example starting time point into the chimeric training structure corresponding to the number of time intervals to calculate the loss value (that is, the label data of the training example starting time point with time intervals of n is input into the chimeric training structure (F-(G) n -D)), after the training data is input into the chimeric training structure, the network output data is obtained through forward propagation, and the loss value of the chimeric training structure is calculated by the network output data and the real data, that is, T i Label data at a time point Input to the chimeric training structure F-(G)n -D Get network output data , data output by the network With T i+n Observation data at time points The loss value Loss can be obtained, and the formula is

[0034] 4.2) Based on the model loss value, the gradient calculation of the number of back propagation layers with limited loss is performed: for the chimeric training structure with a time interval of 0, which only contains the label decoding module and the feature decoding module, the starting time point data and the ending time point data of the training example are at the same time point, the weight parameters of the feature decoding module and the label decoding module are updated, and the updated weight parameters are shared with the label decoding module and the feature decoding module in other chimeric training structures; for the chimeric training structure with a time interval of 1, which only contains 1 G unit, the network uses the error back propagation training mechanism to directly obtain the updated gradient of the G unit and the feature decoding module; in the chimeric training structure with a time interval greater than 1, the series generation mechanism module consists of multiple G units in series, and the error back propagation mechanism is used to calculate the gradient, and the gradient on the last G unit is intercepted as the updated gradient of the G unit and the feature decoding module;

[0035] 4.3) Based on the gradient calculation method that limits the number of loss backpropagation layers, multiple chimeric training structures are updated sequentially, and network weight parameters are shared during the sequential weight update process. Specifically, for chimeric training structures at different time intervals, the gradient of the chimeric training structure is calculated in a certain order, and the weight parameters are updated using the gradient descent method. The updated weight parameters of one chimeric training structure are immediately shared with all other chimeric training structures.

[0036] 4.4) For K+1 chimeric training structures FD, F-(G) 1 -D, F-(G) 2 -D, F-(G) 3 -D…F-(G) K -D Input the training examples of the corresponding time interval according to the method of 4.1) to obtain the loss value of each chimeric training structure, and then determine whether the loss value of each chimeric training structure has converged. If they have converged, proceed to the next step. Otherwise, continue with steps 4.1), 4.2), and 4.3) until the loss value of each chimeric training structure has converged to obtain the result model.

[0037] As a further improvement, step 5) described in the present invention specifically refers to optimizing the modeling process from the level of model hyperparameters, and attempting to adjust the hyperparameters used in the remodeling process during the optimization process, specifically including: feature dimension; structural parameters of the network module (label decoding module, cascade generation mechanism module, feature decoding module network, number of nodes per layer, number of hidden layers); learning rate of the gradient descent method; number of training data examples input for each training; number of cycles of cyclic training; using different combinations of the above hyperparameters to complete the work of establishing the neural network structure and neural network training to obtain a new result model; evaluating the fitting accuracy of the result model to the observed data under each hyperparameter combination, and selecting the optimal model.

[0038] As a further improvement, step 6) of the present invention specifically refers to establishing an encoder module that can convert observation data into label data based on the final result model, and establishing and training an encoder module (E) that can convert observation data into label data based on the two sets of parameters of the network structure and network weights of the optimal feature decoder module (D) obtained in step 5):

[0039] First, build the encoder module (E): a fully connected neural network with an input layer whose number of nodes is the dimension r of the observation data, a hidden layer whose number of nodes is s, a hidden layer whose number of layers is f1, and an output layer whose number of nodes is the dimension e of the feature data.

[0040] Then, the optimal encoder module (E) is obtained by training: the encoder module (E) and the feature decoder module (D) are chimerized to obtain a DE chimeric training structure, the feature decoder module (D) in the DE chimeric training structure is fixedly set to the network weight parameters of the optimal feature decoder module (D) according to claim 8, M batches of actinomycete fermentation data are used as training data to train and update the network weight parameters of the encoder module (E) in the DE chimeric training structure, the input data and output data of the DE chimeric training structure are four-element organizational form data of the same batch and the same time point, the DE chimeric training structure is trained multiple times using the stochastic gradient descent method of neural network training until the model loss value converges, and the network weight parameters of the optimal encoder module (E) are obtained.

[0041] As a further improvement, step 7) of the present invention specifically refers to constructing a chimeric prediction structure for predicting system state changes at different time intervals using an encoder module, and constructing a chimeric prediction structure (E-(G)) for predicting system state changes at different time intervals using the G unit, feature decoder module (D) and encoder module (E) based on the optimal G unit, feature decoder module (D) and encoder module (E). n -D) to achieve T i The observation data at time point is used as input to predict T iThe system state at a given time point is the system state after n time intervals. The test data is organized into a four-element form. The system state changes at any two time points in the same batch are organized as training examples. The training examples are used to calculate the chimeric prediction structure (E-(G) n -D) The accuracy of predicting system state changes at different time intervals.

[0042] As a further improvement, step 1) of the present invention is specifically as follows: obtaining multiple batches of multidimensional time series observation data through observation of multiple batches, wherein the observation data of each batch includes a set of time points, the observation data of each time point includes a set of indicators, and the observation data of each indicator is a specific value; organizing the multidimensional time series observation data into a four-tuple organization form, namely batch, time, indicator, and value. According to the total number of time points of the existing data, a unit matrix marked by time tags is set, wherein the jth column of the unit matrix marked by time tags represents the label data of the jth time point; and the chimeric training structure adopts a fully connected structure.

[0043] As a further improvement, step 2) described in the present invention is specifically as follows: select any two time points in the same batch (the two time points can be the same time points), the training data of the previous time point is represented by the label data corresponding to the time point in the unit matrix marked with the time label, and the training data of the latter time point is the state observation data of the modeled system, and the data of these two time points are organized into a training example to obtain a set of training examples with different time intervals, wherein each training example includes the state of the modeled system at two time points, when the two time points are not the same time points, the data of the earlier time point is called the starting time point data of the training example; the data of the later time point is called the ending time point data of the training example; when the two time points are the same time point, the starting time point and the ending time point are the same time point; the starting time point data of each training example is represented as label data, and the ending time point data is represented as an organization form of observation data quadruple.

[0044] The present invention also discloses a multi-dimensional time series data neural network modeling device based on time point labels, comprising:

[0045] Acquisition unit: used to normalize and organize the original observation data to obtain standardized observation data in a unified form;

[0046] Organization unit: used to organize the training examples for artificial neural network training based on normative observation data;

[0047] Building unit: used to design the structure of artificial neural network and establish artificial neural network model;

[0048] Training unit: used to train the established artificial neural network model using training examples to obtain the parameter matrix of the artificial neural network;

[0049] Optimization unit: used to evaluate the influence of the parameters used in the process of designing the artificial neural network structure and the parameters used in the process of training the established artificial neural network model on the accuracy of the obtained artificial neural network model by using training examples, and select the optimal artificial neural network model under different parameter combinations as the final result model;

[0050] Prediction unit: Based on the final result model, an encoder module is established to convert observation data into label data. The encoder module is used to construct a chimeric prediction structure for predicting system state changes at different time intervals to achieve prediction of system state.

[0051] The designed neural network structure has the following characteristics:

[0052] Its basic structure is a chimeric training structure (F-(G) n -D) neural network; the chimeric training structure is composed of a label decoding module (F), a tandem generation mechanism module (G) n ), feature decoding module (D); the label decoding module and feature decoding module respectively realize the conversion of label data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space;

[0053] A neural network unit in the serial generation mechanism module can model the working mechanism of the system's transformation from the current state to the state after a unit time. Such a network unit is called a unit time positive change generator (G unit). By connecting G units in series, it is possible to model the changes that occur in the system after multiple unit times. The serial generation mechanism module ((G)) is composed of multiple G units in series. n );

[0054] Chimeric predicted structures have the following characteristics:

[0055] The chimeric prediction structure is composed of the encoder module (E), the tandem generation mechanism module (G) n ), and feature decoding module (D) are sequentially embedded in the system; the encoder module (E) and feature decoding module (D) respectively realize the conversion from observation data to feature data and from feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space.

[0056] The beneficial effects of the present invention are as follows:

[0057] 1) The label decoding module (F) is used to label time series data, which can be used to learn time series data with missing time point data or missing indicator data, allowing for better training results with fewer samples.

[0058] 2) By embedding a network structure of a series-connected generative mechanism model in the label decoding module and the feature decoding module, it is possible to complete feature extraction that allows for missing observations and training of the system state generative mechanism model in a unified neural network. The training error is calculated at the overall network level, and the network weights for feature extraction and generative mechanism model functions are coordinated to achieve adaptation of the feature extraction network and the generative mechanism model, thereby improving model fitting accuracy.

[0059] 3) By adding label decoding modules and feature decoding modules on both sides of the series-connected generative mechanism model, the generative model working in the feature space can be trained and predicted using actual observation data. At the same time, it ensures that the feature data output by the generative mechanism model can be accurately restored to the observation data, which is beneficial to the application of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of training steps for chimeric training structure;

[0061] Figure 2 It is a schematic diagram of the serial generation mechanism module structure;

[0062] Figure 3 Schematic diagram of loss-limited backpropagation for chimeric training architecture;

[0063] Figure 4 Schematic diagram of the chimeric predicted structure. DETAILED DESCRIPTION

[0064] The chimeric training structure (F-(G)) designed by the present invention n -D) consists of three modules: label decoding module (F), tandem generation mechanism module (G) n ), a feature decoding module (D), can be used to model the overall mechanism of the fermentation biological system, including the mechanisms of bacterial physiology, catalytic reactions, corresponding processes, and nutritional conditions. By analyzing this mechanism model, possible schemes for mechanism optimization can be obtained to guide the optimization of the fermentation process, including the prediction of target product yield, genome modification of engineered bacteria, and design of fermentation conditions. The method disclosed in the present invention models the overall mechanism of the fermentation biological system through the following steps. Figure 1 Schematic diagram of the training steps for the chimeric training structure.

[0065] 1. Organize the time series data into a quaternary format

[0066] The modeling analysis targets the "biological fermentation system," which includes the bacterial strain and the fermentation environment. Therefore, the model needs to be able to learn and predict indicators reflecting the bacterial strain status and fermentation environment. These indicators can be characterized using various omics techniques. The fermentation process is a time series consisting of multiple time points. Therefore, this fermentation time series data can be organized into a four-dimensional structure: batch, time, indicator, and value. Fermentation time series data is divided into training data and test data. Training data is used to build the model, while test data is used to evaluate model accuracy.

[0067] Assume that both the training data and the test data are time series data obtained during the fermentation of actinomycetes with acarbose as the target product. There are (M+U) batches (for example, M=7, U=3), and each batch has (K+1) time points (for example, K=10, each time point is recorded as T1, T2, T3...T 11 ).

[0068] For samples from each of the (M+U) batches of actinomycete fermentation at (K+1) time points, the abundance of various compounds in each sample was measured by mass spectrometry. Q compounds related to acarbose synthesis were selected (e.g., q = 196). The abundance of these q compounds was extracted from the compound abundance report for each time point and used as the observed value of the fermentation system state at that time point. Compounds whose abundance was not observed by mass spectrometry were recorded as missing. In this batch of samples, data at time points T3 for batch 2 and T2 for batch 3 were missing due to experimental failure in sample measurement.

[0069] The (M+U) batches of actinomycete fermentation data are divided into training data and test data, where M batches are training data batches and U batches are test data batches. Therefore, there are a total of (M+U) batches of data, and each batch has (K+1) time points (T1, T2, T3…T K+1 ), each time point has q-dimensional indicator data values, where the q-dimensional data at the p-th time point in the m-th batch is expressed as

[0070] Set the identity matrix for time labeling based on the number of time points in the existing time series:

[0071] The actinomycete fermentation training data has M batches, each batch has (K+1) time points, so the data has a total of M(K+1) time points, and the identity matrix of the time label is always the Z-dimensional identity matrix I Z , where Z = M(K+1). The jth column in the identity matrix of the time tag represents the tag data at the jth time point. The label data at the p-th time point in the m-th batch is expressed as p≤K+1,m、p∈N * .

[0072] 2. Organize training data examples

[0073] Take M batches of actinomycete fermentation data as training data. According to the quaternary organization form and label data form in step 1, the label data and system status data of any two time points in the same batch are organized as an example. a With T b Represent the starting time point and ending time point of the example respectively. The instance training data is organized as follows:

[0074] Batch 1 example set D1:

[0075] Batch 2 example set D2:

[0076] Batch 3 example set D3:

[0077] Batch m example set D m :

[0078] Where 4≤m≤M-1, m∈N *

[0079] Batch M example set D m :

[0080] Then summarize the examples in each batch, according to T in the example a With T b The interval between two time points is reorganized as:

[0081]

[0082] 3. Establish a chimeric training structure of G units for training examples with different time intervals

[0083] Figure 2 The diagram is a schematic diagram of the serial generation mechanism module structure; the label decoding module (F) and the feature decoding module (D) are constructed: both modules are fully connected neural networks, the number of nodes in the input layer of the label decoding module (F) is the total time point dimension Z of the training data, where Z = M(K+1), the number of nodes in the hidden layer is s, the number of hidden layers is f1, and the number of nodes in the output layer is the feature data dimension e. The number of nodes in the input layer of the feature decoding module (D) is the feature data dimension e, the number of nodes in the hidden layer is s, the number of hidden layers is f1, and the number of nodes in the output layer is the observation data dimension r. Indicates the nth chimeric training structure F-(G) n-The network weight parameters of the label decoding module (F) and the feature decoding module (D) in D (for convenience, the network weight parameters and network bias parameters are collectively referred to as network weight parameters).

[0084] Build a serial generation mechanism module ((G) n ): The network of G units consists of an input layer, a hidden layer, and an output layer. The number of nodes in each layer is e, and the number of hidden layers is f2. For training examples with different time intervals, the G units are connected in series according to the number of time intervals to obtain the corresponding series generation mechanism module: G, (G) 2 、(G) 3 …(G) K ,use Represented as the nth chimeric training structure F-(G) n -D series generation mechanism module (G) i The network weight parameters of the j-th G unit.

[0085] Then the label decoding module (F) and the serial generation mechanism module (G) n ), feature decoding module (D), and the three modules are sequentially combined to form a mosaic training structure (F-(G) n -D).

[0086] The specific parameters of each chimeric training structure are:

[0087] There is no G unit in FD: the weight parameters are

[0088] F-(G) 1 - There is 1 G unit in D: the weight parameter is

[0089] F-(G) 2 - There are 2 G units in D: the weight parameters are

[0090] F-(G) 3 - There are 3 G units in D: the weight parameters are

[0091]

[0092] F-(G) K -D has K G units: the weight parameters are

[0093] The initial weight parameters of the G unit, TI module and D module of each chimeric training structure are set to the same random values: Right now:

[0094] Before training begins, Before training begins, Before training begins,

[0095] 4. Calculate model loss value

[0096] The starting time point T of the training examples of group c a The label data is input into the chimeric training structure of the corresponding number of time intervals to calculate the loss value. c starting time points T in a Label data Enter F-(G) n -D, get output data Where n≤K, n∈N. With T b Multidimensional real data at a time point The loss value Loss can be calculated n , the calculation formula is

[0097] 5. Calculate the gradient through loss-restricted backpropagation and update the network weight parameters through sequential cycle optimization

[0098] Figure 3 Schematic diagram of loss-restricted backpropagation for the chimeric training structure; for the chimeric training structure when the number of time intervals is 0, it only includes the label decoding module and the feature decoding module, and updates the weight parameters of the feature decoding module and the label decoding module.

[0099] For the chimeric training structure with a time interval of 1, which contains only 1 G unit, the network uses the error back propagation training mechanism to directly obtain the update gradients of the G unit and the feature decoding module.

[0100] The serial generation mechanism module in the chimeric training structure with a time interval greater than 1 is composed of multiple G units connected in series. The error back propagation mechanism is used to calculate the gradient, and the gradient on the last G unit is intercepted as the update gradient of the G unit and the feature decoding module. The specific calculation formula is:

[0101] For the chimeric training structure when the number of time intervals is 0, FD only calculates the update gradients of the label decoding module and the feature decoding module, that is,

[0102]

[0103]

[0104] For the chimeric training structure F-(G) with a time interval of 11 -D, calculate And the updated gradient of the feature decoding module, that is,

[0105]

[0106]

[0107] For the chimeric training structure F-(G) with a time interval of 2 2 -D, calculate And the updated gradient of the feature decoding module, that is,

[0108]

[0109]

[0110] For the chimeric training structure F-(G) with 3 time intervals 3 -D, calculate And the updated gradient of the feature decoding module, that is,

[0111]

[0112]

[0113] For the mosaic training structure F-(G) with K time intervals K -D, calculate And the updated gradient of the feature decoding module, that is,

[0114]

[0115]

[0116] The above gradient calculation method only calculates the last two structural units of the chimeric training structure each time (when the time interval is 0, only the gradients of the feature decoding module and the label decoding module are calculated; when the number of time intervals is greater than 1, only the gradient of the last G unit and the gradient of the label decoding module are calculated), so the present invention calls this method loss-limited backpropagation.

[0117] Based on the gradient calculated by the above method, for the chimeric training structures with different time intervals, the gradient of the chimeric training structure with smaller time intervals is calculated first in the order of the number of time intervals, and the weight is updated using the SGD stochastic gradient descent method. The updated weight is shared with all other chimeric training structures. This training method is called sequential cyclic optimization in the present invention. The number of cycles in the cyclic training is set to H, where the tth cycle training process is:

[0118] First calculate the update gradient of the label decoding module and feature decoding module in FD, that is, Among them, α represents the learning rate of the SGD gradient descent method; Represents the updated weight parameters of the encoder module and decoder module at the end of the t-1th cycle training. Represents the updated weight parameters of the encoder module and decoder module in the tth cycle; the updated gradients of the label decoding module and feature decoding module The weight parameters are shared with other chimeric training structures:

[0119] are equal to are equal to

[0120] Then calculate F-(G) 1 -D The updated weights of the unit and feature decoding module, namely in, Indicates that at the end of the t-1th cycle training, The weight parameters after unit update, Indicates the tth cycle G1 The weight parameters of the unit after update;

[0121] Will Gradients after unit and feature decoding module updates The weight parameters are shared with other chimeric training structures:

[0122] are equal to are equal to

[0123] Calculate F-(G) 2 -D The updated weights of the unit and feature decoding module, namely Then Gradients after unit and feature decoding module updates The weight parameters are shared with all other chimeric training structures.

[0124] Update F-(G) using the same calculation method 3 -D, F-(G) 4 -D, E-(G) 5 -D…F-(G) K -D, when this cycle training calculates F-(G) K -D After the unit and feature decoding modules update their weights and share them with other chimeric training structures, this cycle of training ends and enters the next cycle of training calculation.

[0125] 6. Obtain the result model after all training structure losses converge

[0126] After each cycle of training in step 5, the K+1 chimeric training structures FD, F-(G) 1 -D, F-(G) 2 -D, F-(G) 3 -D…F-(G) K -D Input all training examples of the corresponding time interval according to the method in step 4, and you can get the full data loss value of each series training module. The specific calculation is:

[0127] Calculate the loss value of FD, that is

[0128] Calculate F-(G) 1 -D loss value, that is

[0129] Calculate F-(G) 2 -D loss value, that is

[0130] Calculate F-(G) 3 -D loss value, that is

[0131]

[0132] Calculate F-(G) K -D loss value, that is

[0133]

[0134] Determine Loss0, Loss1, Loss2, Loss3,…, Loss K Are they all converged? If so, proceed to the next step, otherwise continue the calculation of steps 4 and 5 until Loss0, Loss1, Loss2, Loss3, ..., Loss K All converge. Note that the overall loss of the chimeric training structure converges after the training process has gone through the slice cycle

[0135] 7. Adjust model hyperparameters and optimize the resulting model

[0136] After completing training and obtaining the resulting model in step 6, adjust the model's hyperparameters: feature dimension e, network module structural parameters (number of nodes s and number of hidden layers f1 in the hidden layers of the label decoding module and feature decoding module; number of nodes e and number of hidden layers f2 in each layer of the cascade generation mechanism module), learning rate α for the SGD stochastic gradient descent method, number of training examples c in a single training cycle, and number of training cycles H. After adjusting these hyperparameter combinations, repeat steps 4-6 until the network loss difference converges, terminating training.

[0137] 8. Build and train the encoder module

[0138] After the hyperparameter optimization in step 7, the optimal G unit, label decoding module, and feature decoding module are obtained. Save the two sets of parameters of the network structure and network weights of the G unit, label decoding module (F), and feature decoder module (D).

[0139] Based on the two sets of parameters of the optimal feature decoder module (D) obtained in step 7, the network structure and network weights, the encoder module (E) that can convert observation data into label data is established and trained:

[0140] First, build the encoder module (E): It is a fully connected neural network. The number of nodes in the input layer of the encoder module (E) is the dimension r of the observation data, the number of nodes in the hidden layer is s, the number of hidden layers is f1, and the number of nodes in the output layer is the dimension e of the feature data.

[0141] Train to obtain the optimal encoder module (E): The encoder module (E) and the feature decoder module (D) are chimerically combined to obtain a DE chimeric training structure. The feature decoder module (D) in the DE chimeric training structure is fixedly set to the network weight parameters of the optimal feature decoder module (D) obtained in step 7. The network weight parameters of the encoder module (E) in the DE chimeric training structure are updated using the M batches of actinomycete fermentation data obtained in step 2 as training data. The input and output data of the DE chimeric training structure are the quaternary tissue form data of the same batch and the same time point. The DE chimeric training structure is repeatedly trained using the stochastic gradient descent method for neural network training until the model loss value converges, thereby obtaining the network weight parameters of the optimal encoder module (E).

[0142] 9. Build a chimeric prediction structure to achieve time series data prediction and verify prediction accuracy

[0143] Figure 4The diagram of the chimeric prediction structure is shown in Figure 7. After steps 7 and 8, the optimal G unit, label decoding module (F), feature decoder module (D), and encoder module (E) are obtained. The G unit, feature decoder module (D), and encoder module (E) are used to construct a chimeric prediction structure for predicting system state changes at different time intervals. The construction of the chimeric prediction structure is similar to the chimeric training structure (F-(G) n -D), whose basic structure is a chimeric predicted structure (E-(G) n -D) neural network; the chimeric prediction structure is composed of an encoder module (E), a tandem generation mechanism module (G) n ), feature decoder module (D); the encoder module and feature decoder module respectively realize the conversion of observation data to feature data and feature data to observation data, so that the tandem generation mechanism module can learn the working mechanism of the system in the feature space;

[0144] A neural network unit in the serial generation mechanism module can model the working mechanism of the system's transformation from the current state to the state after a unit time. Such a network unit is called a unit time positive change generator (G unit). By connecting G units in series, it is possible to model the changes that occur in the system after multiple unit times. The serial generation mechanism module ((G)) is composed of multiple G units in series. n );

[0145] Chimeric predicted structure (E-(G) n -D) are set to the optimal network structure and network weight parameters of the G unit, feature decoder module (D) and encoder module (E) obtained in steps 7 and 8.

[0146] Chimeric predicted structure (E-(G) n -D) by connecting the encoder module (E) and the generation mechanism module (G) in series n ), feature decoding module (D), and the new neural network structure formed by the three modules can realize T i Input the real observation data at the time point, predict T i The system state at the time point is the system state after n time intervals, output T i The predicted value of the system state at a time point based on the actual observation data after n time intervals.

[0147] Using the constructed chimeric prediction structure (E-(G) n -D) Implement time series data prediction:

[0148] The U batches of actinomycete fermentation data described in the first step of the application case are used as test data. Each batch has K+1 time points (T1, T2, T3…TK+1 ), each time point has q indicators. The test data is organized according to the quaternary organization form in step 1. The system state changes at any two time points in the same batch are organized as an example. T a With T b Represent the starting time point and ending time point of the example respectively. The instance training data is organized as follows:

[0149] Batch 1 example collection

[0150] Batch 2 example collection

[0151] Batch 3 example collection

[0152] Batch m example set

[0153] where 4≤m≤U-1, m∈N *

[0154] Batch U example set

[0155] Then summarize the examples in each batch, according to T in the example a With T b The interval between two time points is reorganized as:

[0156]

[0157] in, There are Total N l For example, l≤K, l∈N * .

[0158] The chimeric predicted structure (E-(G) n -D) and evaluate the accuracy of the time series data:

[0159] Calculate the prediction accuracy of ED, that is

[0160] Calculate E-(G) 1 -D prediction accuracy, i.e.

[0161] Calculate E-(G) 2 -D prediction accuracy, i.e.

[0162] Calculate E-(G) 3 -D prediction accuracy, i.e.

[0163]

[0164] Calculate E-(G) K -D prediction accuracy, i.e.

[0165] Thus, the biological fermentation system mechanism model is obtained, which consists of the encoder module (E), the serial generation mechanism module (G) n ), feature decoding module (D), and the three modules are sequentially combined to form a chimeric prediction structure (E-(G) n -D) The expected accuracy for predicting changes in system states in a bio-fermentation system at different time intervals.

[0166] The present invention also discloses a multi-dimensional time series data neural network modeling device based on time point labels, comprising:

[0167] Acquisition unit: used to normalize and organize the original observation data to obtain standardized observation data in a unified form;

[0168] Organization unit: used to organize the training examples for artificial neural network training based on normative observation data;

[0169] Building unit: used to design the structure of artificial neural network and establish artificial neural network model;

[0170] Training unit: used to train the established artificial neural network model using training examples to obtain the parameter matrix of the artificial neural network;

[0171] Optimization unit: used to evaluate the influence of the parameters used in the process of designing the artificial neural network structure and the parameters used in the process of training the established artificial neural network model on the accuracy of the obtained artificial neural network model by using training examples, and select the optimal artificial neural network model under different parameter combinations as the final result model;

[0172] Prediction unit: Based on the final result model, an encoder module is established to convert observation data into label data. The encoder module is used to construct a chimeric prediction structure for predicting system state changes at different time intervals to achieve prediction of system state.

[0173] The designed neural network structure has the following characteristics:

[0174] Its basic structure is a chimeric training structure (F-(G) n -D) neural network; the chimeric training structure is composed of a label decoding module (F), a tandem generation mechanism module (G)n ), feature decoding module (D); the label decoding module and feature decoding module respectively realize the conversion of label data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space;

[0175] A neural network unit in the serial generation mechanism module can model the working mechanism of the system's transformation from the current state to the state after a unit time. Such a network unit is called a unit time positive change generator (G unit). By connecting G units in series, it is possible to model the changes that occur in the system after multiple unit times. The serial generation mechanism module ((G)) is composed of multiple G units in series. n );

[0176] Chimeric predicted structures have the following characteristics:

[0177] The chimeric prediction structure is composed of the encoder module (E), the tandem generation mechanism module (G) n ), and feature decoding module (D) are sequentially embedded in the system; the encoder module (E) and feature decoding module (D) respectively realize the conversion from observation data to feature data and from feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space.

[0178] The above is not intended to limit the specific embodiments of this patent. It should be noted that those skilled in the art may make changes, modifications, additions, or substitutions without departing from the spirit and scope of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A neural network modeling method for multidimensional time series data based on time point labels, characterized by: The following steps are involved: 1) normalizing and arranging the original observation data to obtain standardized observation data in a unified form; obtaining multiple batches of multidimensional time series observation data through observation of multiple batches, wherein the observation data of each batch includes a set of time points, the observation data of each time point includes a set of indicators, and the observation data of each indicator is a specific value; arranging the multidimensional time series observation data into a four-tuple organizational form, namely, batch, time, indicator, and value; setting a unit matrix marked by time tags based on the total number of time points of the existing data, wherein the jth column of the unit matrix marked by time tags represents the label data of the jth time point; and adopting a fully connected structure as the chimeric training structure; the observation data are specifically as follows: the time series data are measured during the actinomycete fermentation process with acarbose as the target product, the abundance of various compounds in each sample is determined by mass spectrometry, q compounds related to acarbose synthesis are selected, and the abundance of these q compounds is extracted from the compound abundance report of the sample at each time point as the observation value of the state of the biological fermentation system at that time point; 2) Organizing training examples for artificial neural network training based on canonical observation data; 3) Design the structure of artificial neural network and establish artificial neural network model; 4) using the training examples in step 2) to train the artificial neural network model established in step 3) to obtain a parameter matrix of the artificial neural network; 5) using the training examples in step 2), evaluating the effects of the parameters used in the process of designing the artificial neural network structure and the parameters used in the process of training the established artificial neural network model on the accuracy of the obtained artificial neural network model, and selecting the optimal artificial neural network model under different parameter combinations as the final result model; 6) Based on the final result model, an encoder module is established to convert the observation data into feature data; 7) Use the encoder module to construct a chimeric prediction structure for predicting system state changes at different time intervals; use the chimeric prediction structure to evaluate model accuracy based on test examples and make predictions about system state changes in real application scenarios; The neural network structure designed in step 3) has the following characteristics: Its basic structure is a chimeric training structure (F-(G) n -D) neural network; the chimeric training structure is sequentially composed of a label decoding module (F), a tandem generation mechanism module ((G) n ), feature decoding module (D); the label decoding module and feature decoding module respectively realize the conversion of label data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space; A neural network unit in the series generation mechanism module can model the working mechanism of the system changing from the current state to the state after a unit time. Such a network unit is called a unit time positive change generator (G unit); by connecting G units in series, it is possible to model the changes that occur in the system after multiple unit times. The series generation mechanism module ((G)) is composed of multiple G units in series. n ); The chimeric predicted structure described in step 7) has the following characteristics: The chimeric prediction structure is composed of an encoder module (E), a tandem generation mechanism module (G) n ), and feature decoding module (D) are sequentially connected to form the three modules; the encoder module (E) and feature decoding module (D) respectively realize the conversion of observation data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space.

2. The multidimensional time series data neural network modeling method based on time point labels according to claim 1 is characterized in that: The chimeric training structure is established as follows: Chimeric training structure (F -(G) n -D) is composed of the label decoding module (F), the series generation mechanism module ((G) n ), feature decoding module (D), the number of neuron nodes in the input layer of the label decoding module is the total dimension of the training data for all time periods, the number of neuron nodes in the output layer is the dimension of the feature data, and the serial generation mechanism module ((G) n ) is obtained by connecting multiple G units in series. The number of neuron nodes in each layer of the G unit is the dimension of the feature data. The number of neuron nodes in the input layer of the feature decoding module is the dimension of the feature data, and the number of neuron nodes in the output layer is the dimension of the observation data.

3. The multidimensional time series data neural network modeling method based on time point labels according to claim 2 is characterized in that: The chimeric training structure with n time intervals is called F-(G) n -D: When n=0, FD does not contain G units, and its training data is the training example with the starting time point data and the ending time point data at the same time point; when n=1, F-(G) 1 -D contains only one G unit, and its training data is a training example with a time interval of 1; when n>1, F-(G) n -D in series generation mechanism module ((G) n ) is obtained by connecting n G units in series, and its training data is training examples with n time intervals.

4. The multidimensional time series data neural network modeling method based on time point labels according to claim 1, 2 or 3, characterized in that: The training method of step 4) is specifically as follows: 4.1) Calculate the loss value of each chimeric training structure: For a specific chimeric training structure F-(G) n -D, only uses examples with time intervals of n to calculate the model loss, and inputs the label data of the training example starting time point into the chimeric training structure corresponding to the number of time intervals to calculate the loss value (that is, the label data of the training example starting time point with time intervals of n is input into the chimeric training structure (F-(G) n -D)), after the training data is input into the chimeric training structure, the network output data is obtained through forward propagation, and the loss value of the chimeric training structure is calculated by the network output data and the real data, that is, T i Label data at a time point Input to the chimeric training structure F-(G) n -D Get network output data F Output data F from the network With T i+n Observation data at time points The loss value Loss can be obtained, and the formula is 4.2) Based on the model loss value, the gradient calculation of the number of back propagation layers with limited loss is performed: for the chimeric training structure with a time interval of 0, which only contains the label decoding module and the feature decoding module, the starting time point data and the ending time point data of the training example are at the same time point, the weight parameters of the feature decoding module and the label decoding module are updated, and the updated weight parameters are shared with the label decoding module and the feature decoding module in other chimeric training structures; for the chimeric training structure with a time interval of 1, which only contains 1 G unit, the network uses the error back propagation training mechanism to directly obtain the updated gradient of the G unit and the feature decoding module; in the chimeric training structure with a time interval greater than 1, the series generation mechanism module consists of multiple G units in series, and the error back propagation mechanism is used to calculate the gradient, and the gradient on the last G unit is intercepted as the updated gradient of the G unit and the feature decoding module; 4.3) Based on the gradient calculation method of limiting the number of loss backpropagation layers, multiple chimeric training structures are sequentially weighted and network weight parameters are shared during the sequential weight update process. Specifically, for chimeric training structures at different time intervals, the gradients of the chimeric training structures are calculated in a certain order and the weights are updated using the gradient descent method. The updated weight parameters of one chimeric training structure are immediately shared with all other chimeric training structures. 4.4) For K+1 chimeric training structures FD, F-(G) 1 -D, F-(G) 2 -D, F-(G) 3 -D...F-(G) K -D Input the training examples of the corresponding time interval according to the method of 4.1) to obtain the loss value of each chimeric training structure, and then determine whether the loss value of each chimeric training structure has converged. If they have converged, proceed to the next step. Otherwise, continue with steps 4.1), 4.2), and 4.3) until the loss value of each chimeric training structure has converged to obtain the result model.

5. The multidimensional time series data neural network modeling method based on time point labels according to claim 4 is characterized in that: The step 5) specifically refers to optimizing the modeling process from the level of model hyperparameters, and attempting to adjust the hyperparameters used in the remodeling process during the optimization process, specifically including: feature dimension; structural parameters of the network module (label decoding module, cascade generation mechanism module, number of nodes per layer of the feature decoding module network, number of hidden layers); learning rate of the gradient descent method; number of training data examples input for each training; number of cycles of cyclic training; using different combinations of the above hyperparameters to complete the work of establishing the neural network structure and neural network training to obtain a new result model; evaluating the fitting accuracy of the result model to the observed data under each hyperparameter combination, and selecting the optimal model.

6. The multidimensional time series data neural network modeling method based on time point labels according to claim 1, 2, 3 or 5, characterized in that: Said step 6) specifically refers to establishing an encoder module capable of converting observed data into labeled data based on said final result model, and establishing and training an encoder module (E) capable of converting observed data into labeled data based on the two sets of parameters of the network structure and network weights of the optimal feature decoder module (D) obtained in step 5): First, build the encoder module (E): a fully connected neural network with an input layer whose number of nodes is the dimension r of the observation data, a hidden layer whose number of nodes is s, a hidden layer whose number of layers is f1, and an output layer whose number of nodes is the dimension e of the feature data. Then, the optimal encoder module (E) is obtained by training: the encoder module (E) and the feature decoder module (D) are chimerized to obtain a DE chimeric training structure, the feature decoder module (D) in the DE chimeric training structure is fixedly set to the network weight parameters of the optimal feature decoder module (D) described in claim 8, and M batches of actinomycete fermentation data are used as training data to train and update the network weight parameters of the encoder module (E) in the DE chimeric training structure. The input data and output data of the DE chimeric training structure are four-element organizational form data of the same batch and the same time point. The DE chimeric training structure is trained multiple times using the stochastic gradient descent method of neural network training until the model loss value converges, thereby obtaining the network weight parameters of the optimal encoder module (E) for training.

7. The multidimensional time series data neural network modeling method based on time point labels according to claim 6 is characterized in that: Said step 7) specifically refers to constructing a chimeric prediction structure for predicting system state changes at different time intervals using an encoder module, and constructing a chimeric prediction structure (E-(G)) for predicting system state changes at different time intervals using the G unit, feature decoder module (D) and encoder module (E) based on the optimal G unit, feature decoder module (D) and encoder module (E) n -D) to achieve T i The observation data at time point is used as input to predict T i The system state at a given time point is the system state after n time intervals. The test data is organized into training examples for the system state changes at any two time points in the same batch according to the four-element organization form. The chimeric prediction structure (E-(G)) is calculated using the training examples. n -D) The accuracy of predicting system state changes at different time intervals.

8. The multi-dimensional time series data neural network modeling method based on time point labels according to claim 7 is characterized in that: The step 2) is specifically as follows: select any two time points in the same batch (the two time points can be the same time points), the training data of the previous time point is represented by the label data corresponding to the time point in the unit matrix marked with the time label, and the training data of the latter time point is the state observation data of the modeled system, and the data of these two time points are organized into a training example to obtain a set of training examples with different time intervals, wherein each training example includes the state of the modeled system at two time points. When the two time points are not the same time points, the data of the earlier time point is called the starting time point data of the training example; the data of the later time point is called the ending time point data of the training example; when the two time points are the same time point, the starting time point and the ending time point are the same time point; the starting time point data of each training example is represented as the label data, and the ending time point data is represented as the organization form of the observation data quadruple.

9. A multi-dimensional time series data neural network modeling device based on time point labels, characterized in that: include: Acquisition unit: used to normalize and organize the original observation data to obtain standardized observation data in a unified form; Multiple batches of observations are obtained, where each batch of observation data includes a set of time points, each time point observation data includes a set of indicators, and each indicator observation data is a specific value. The multidimensional time series observation data are organized into a four-tuple organization form, namely, batch, time, indicator, and value. Based on the total number of time points in the existing data, a unit matrix marked by time tags is set, where the jth column of the unit matrix marked by time tags represents the label data of the jth time point. The chimeric training structure adopts a fully connected structure. The observation data is specifically as follows: the time series data is measured during an actinomycete fermentation process with acarbose as the target product, the abundance of various compounds in each sample is determined by mass spectrometry, q compounds related to acarbose synthesis are selected, and the abundance of these q compounds is extracted from the compound abundance report of the sample at each time point as the observation value of the biological fermentation system state at that time point. Organization unit: used to organize the training examples for artificial neural network training based on normative observation data; Building unit: used to design the structure of artificial neural network and establish artificial neural network model; Training unit: used to train the established artificial neural network model using training examples to obtain the parameter matrix of the artificial neural network; Optimization unit: used to evaluate the influence of the parameters used in the process of designing the artificial neural network structure and the parameters used in the process of training the established artificial neural network model on the accuracy of the obtained artificial neural network model by using training examples, and select the optimal artificial neural network model under different parameter combinations as the final result model; Prediction unit: Based on the final result model, an encoder module is established to convert observation data into label data. The encoder module is used to construct a chimeric prediction structure for predicting system state changes at different time intervals to achieve prediction of system state. The designed neural network structure has the following characteristics: Its basic structure is a chimeric training structure (F-(G) n -D) neural network; the chimeric training structure is sequentially composed of a label decoding module (F), a tandem generation mechanism module ((G) n ), feature decoding module (D); the label decoding module and feature decoding module respectively realize the conversion of label data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space; A neural network unit in the series generation mechanism module can model the working mechanism of the system changing from the current state to the state after a unit time. Such a network unit is called a unit time positive change generator (G unit); by connecting G units in series, it is possible to model the changes that occur in the system after multiple unit times. The series generation mechanism module ((G)) is composed of multiple G units in series. n ); The chimeric prediction structure has the following characteristics: The chimeric prediction structure is composed of an encoder module (E), a tandem generation mechanism module (G) n ), and feature decoding module (D) are sequentially connected to form the three modules; the encoder module (E) and feature decoding module (D) respectively realize the conversion of observation data to feature data and feature data to observation data, so that the series generation mechanism module can learn the working mechanism of the system in the feature space.

Images