A method and device for modeling time series data of multi-dimensional state changes of a system

Abstract

The application discloses a kind of to the method and device of multidimensional state change time series data of system, the design of the generator (G unit) of unit time positive change is made in the application, so that model can learn the change of two time states of system, the way of the overall learning of traditional method to time series trajectory is converted to the way of the change learning of specific time point of time series trajectory, the problem that time series data with incomplete points cannot be learned is solved, since the change of training mode, multiple learning examples can be organized by a batch of time series data, so that the number of examples is increased, the information efficiency of training data is improved;The law of system state change on unit time is learned by model, and its essence is a system working mechanism model, new system working mechanism can be superimposed on the model to predict what change of system behavior may be caused by the change of existing system working mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of computing methods and is a method and device for modeling time series data of multi-dimensional state changes of a system. 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] Currently, the primary method for analyzing fermentation processes is metabolic flux modeling. This approach analyzes the potential equilibrium states of a fermentation process, rather than its time-series trajectory. This analytical approach relies on a system of partial differential equations to describe chemical reactions, and therefore has certain limitations. Primarily, it cannot perform cross-class joint analysis of metabolomic, fermentation process, and transcriptomic data, limiting its ability to guide fermentation process optimization.

[0005] On the other hand, regression algorithms based on mathematical statistics and machine learning theory are used in many fields to model various types of time series data. Mathematical statistics methods first design a reasonable functional framework based on the modeler's understanding of the state evolution patterns of the modeled system. The functional parameters are then learned to perform curve fitting for multidimensional time series. Commonly used methods in this category include the periodic factor method, the moving average method, and the ARIMA model. Because the functional framework designed by the modeler differs from the actual mechanism driving the state evolution of the modeled system, the models constructed by these methods have a system-level upper limit on accuracy that cannot be surpassed by increasing experimental observations. Another strategy is to use machine learning methods, based on general-purpose frameworks such as neural networks, to perform curve fitting for multidimensional time series. Machine learning methods address the requirement for mathematical statistics to precisely design a fitting functional framework by designing a functional framework that effectively fits all evolution patterns. Commonly used methods in this category include the K-nearest neighbor algorithm, the SVM algorithm, the LSTM model, and the Seq2seq model. Because general-purpose functionals are more difficult to learn parameters for than specifically designed functionals, these methods typically require a large number of training examples to achieve good fitting results.

[0006] Whether using mathematical statistics or machine learning methods, the state change curves of a batch of the modeled system are typically used as a learning example, and all state indicators in the learning examples must be present. In practical applications, many tracking observations of the modeled system's state cannot be completely complete with missing indicators or time points due to sample detection issues. For example, in modeling time series data from a fermentation process, sampling and observation often occurs at fewer time points during periods of minimal change, while sampling and observation occur at more time points during periods of significant change. Using the sampling frequency during periods of significant change as a benchmark results in missing data for some time points during these periods of minimal change. Omics technologies, such as metabolomics and transcriptomics, are often used to characterize the system state at a single time point during a fermentation process. While these technologies can generate numerous indicators simultaneously, due to technical limitations, they cannot guarantee that each indicator measured at each time will be sufficiently accurate for subsequent modeling, resulting in missing indicators in the state data of the modeled system at that point in time. Such incomplete time series data cannot be directly used for learning.

[0007] If the state data of the modeled system at a certain point in time is missing, interpolation can lead to inaccurate training data and reduced model accuracy. Using an entire batch of time series as a learning example increases the size of the individual example data, complicates the model, increases the cost of acquiring examples for testing, and reduces the number of examples, making model learning more difficult.

[0008] On the other hand, the aforementioned time series modeling methods rely on curve fitting, rather than modeling the mechanisms that generate time series. Therefore, they cannot provide clues for optimizing these mechanisms. For example, the time series data of a fermentation process is driven by the biological mechanisms of the engineered bacteria. The fermentation process model obtained through curve fitting time series modeling does not directly correspond to the biological mechanisms of the engineered bacteria, making it difficult to guide the modification of the engineered bacteria genome.

[0009] The present invention discloses a neural network model that can be used to model the working mechanism of a fermentation system in a bio-fermentation 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. Compared to conventional time series numerical fitting models, this driving mechanism model of time series changes in the bio-fermentation system can better reflect the impact of various regulatory mechanisms of engineered bacteria on the fermentation process, better guide the genome modification of engineered bacteria, and achieve better production performance.

[0010] Therefore, the current multidimensional time series modeling methods have the following main shortcomings:

[0011] 1) Time series data with incomplete time points cannot be learned;

[0012] 2) Using a batch of time series as a learning case results in a large single-case data size, which leads to high model complexity. At the same time, the number of cases available for learning is small, which makes model learning difficult.

[0013] 3) 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] Currently, there is a lack of a universal method that can fully implement multidimensional time series data with different missing data points, and achieve learning and simulation of the driving mechanisms of system state changes per unit time. This invention discloses a method for modeling the generation mechanism of time series using a neural network model. This method can utilize time series data with incomplete time points without interpolation or using the entire batch of time series as an example, thus overcoming the shortcomings of the above multidimensional time series modeling methods. Summary of the Invention

[0015] The present invention discloses a method and device for modeling time series data of multi-dimensional state changes of a system. Specifically, the present invention is implemented through the following technical solutions:

[0016] A method for modeling time series data of multi-dimensional state changes of a system, comprising:

[0017] 1) Standardize and organize the original observation data to obtain standardized observation data in a unified form;

[0018] 2) Organizing training examples for artificial neural network training based on canonical observation data;

[0019] 3) Design the structure of artificial neural network and establish artificial neural network model;

[0020] 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;

[0021] 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;

[0022] The neural network structure designed in step 3) has the following characteristics:

[0023] Its basic structure is a neural network unit called a unit time forward change generator (G unit), whose input layer and output layer have the same dimension; this neural network unit is used to model the changes that occur in multidimensional observation data after a unit time;

[0024] The serial training structure is obtained by connecting the G units in series, and the serial training structure is used to model the changing rules of multi-dimensional observation data after multiple unit times.

[0025] As a further improvement, step 1) described in the present invention is specifically as follows: obtaining multiple batches of multidimensional time series observation data through observation of multiple batches, each batch of observation data 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 organizational form, namely batch, time, indicator, and value.

[0026] As a further improvement, step 2) described in the present invention is specifically: organizing the state changes of the modeled system at any two time points in the same batch into a training example, obtaining a group of training examples with different time intervals, wherein each training example includes the state of the modeled system at two time points, the data at the earlier time point is called the starting time point data of the training example; the data at the later time point is called the ending time point data of the training example; the starting time point data and the ending time point data of each training example are represented as an organizational form of observation data quadruple.

[0027] As a further improvement, the forward change generator (G unit) described in the present invention adopts a fully connected structure.

[0028] As a further improvement, the serial training structure described in the present invention is established as follows: for training examples with different time intervals, according to the number of time intervals, for each time interval, which is recorded as n unit time, a serial training structure composed of G units is established, and the serial training structure connects n G units end to end in series.

[0029] As a further improvement, the serial training structure with n time intervals in the present invention is called (G) n , when n=1(G) 1 There is only one G unit in the series training structure (G) 1 The corresponding training data is a training example with a time interval of 1; when n>1, (G) n The serial training structure is obtained by connecting n G units in series, and the serial training structure (G) n The corresponding training data is training examples with n time intervals.

[0030] As a further improvement, the training method of step 4) of the present invention is specifically as follows:

[0031] 4.1 Calculate the loss value of each serial training structure: For each serial training structure (G) n , and the corresponding training example is recorded as S i , note S i The starting time point is T i , S i The end time point is T i+n , record the modeled system in T i The status data at a point in time is In T i+n The status data at a point in time is T i Data at a point in time Input to the cascade training structure (G) n , get the output data of the network Output data based on the network With T i+n Real data at a point in time The model loss value Loss can be calculated using the formula:

[0032] 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 series training structure with a time interval of 1, which contains only one G unit, the network uses the error back propagation training mechanism and directly uses the error back propagation method to obtain the updated gradient of the G unit; For the series training structure with a time interval greater than 1, which consists of multiple G units in series, the error back propagation method is used to calculate the gradient, and the gradient on the last G unit is intercepted as the updated gradient of the G unit of the series training structure;

[0033] 4.3 Based on the gradient calculation method of limiting the number of loss backpropagation layers, multiple series training structures are sequentially weighted and network weight parameters are shared during the sequential weight update process. Specifically, for series training structures at different time intervals, the gradient of the series training structure is calculated in a certain order and the weight is updated using the gradient descent method. The updated weight parameters of one series training structure are immediately shared with all other series training structures.

[0034] 4.4 Pairs of K Series Training Structures (G) 1 、(G) 2 、(G) 3 ...(G) K, calculate the loss value of each serial training structure, and determine whether the loss value of each serial training structure has converged. If they have converged, the result model is obtained. Otherwise, continue to repeat steps 4.1, 4.2, and 4.3 until the loss value of each serial training structure has converged and the result model is obtained.

[0035] As a further improvement, step 5) of 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: structural parameters of the G unit, including the number of nodes in each layer of the network, the number of hidden layers, the learning rate of the gradient descent method, the number of training data examples input for each training, and the number of cycles of the 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.

[0036] The present invention also discloses a device for modeling a multidimensional data time series using the method for modeling a multidimensional data time series, comprising:

[0037] Acquisition unit: used to normalize and organize the original observation data to obtain standardized observation data in a unified form;

[0038] Organization unit: used to organize the training examples for artificial neural network training based on normative observation data;

[0039] Building unit: used to design the structure of artificial neural network and establish artificial neural network model;

[0040] Training unit: used to train the established artificial neural network model using training examples to obtain the parameter matrix of the artificial neural network;

[0041] 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;

[0042] The neural network structure of the design has the following characteristics:

[0043] Its basic structure is a neural network unit called a unit time forward change generator (G unit), whose input layer and output layer have the same dimension; this neural network unit is used to model the changes that occur in multidimensional observation data after a unit time;

[0044] The serial training structure is obtained by connecting the G units in series, and the serial training structure is used to model the changing rules of multi-dimensional observation data after multiple unit times.

[0045] The beneficial effects of the present invention are as follows:

[0046] 1) By designing a unit-time positive change generator (G unit), the model can learn the changes between two time states of the system. This transforms the traditional method of learning the time series trajectory as a whole into learning the changes at specific time points in the time series trajectory, solving the problem of incomplete time series data being unable to be learned.

[0047] 2) Due to the change in training methods, multiple learning examples can be organized through a batch of time series data, which increases the number of examples and improves the information efficiency of the training data;

[0048] 3) The model learns the law of system state change per unit time. Its essence is a model of the system working mechanism. New system working mechanisms can be superimposed on this model to predict what changes in system behavior may be caused by changes to the existing system working mechanism.

[0049] 4) The structure and training method of the neural network model of the present invention can be applied to the analysis of biological fermentation time series data. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of model training steps;

[0051] Figure 2 This is an example diagram of how training examples are organized;

[0052] Figure 3 Schematic diagram of the series training structure composed of multiple G units;

[0053] Figure 4 Schematic diagram of loss-limited backpropagation for a cascaded training architecture. DETAILED DESCRIPTION

[0054] The forward change generator (G unit) designed in this invention can be used to model the overall mechanisms of fermentation biological systems, including bacterial physiology, catalytic reactions, corresponding processes, and the mechanisms of nutritional regulation. By analyzing this mechanistic model, possible optimization schemes can be obtained, guiding the optimization of the fermentation process, including target product yield prediction, genome modification of engineered bacteria, and fermentation condition design. Figure 1 Schematic diagram of model training steps; the method disclosed in the present invention models the overall mechanism of the fermentation biological system through the following steps.

[0055] 1. Organize the time series data into a quaternary format

[0056] 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.

[0057] Assume that the training data is the 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 ).

[0058] 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.

[0059] 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

[0060]

[0061] 2. Organize training data examples

[0062] Figure 2 This is an example diagram of the organization of training examples. Take M batches of actinomycete fermentation data as training data. 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. a With T bRepresent the starting time point and ending time point of the example respectively. The instance training data is organized as follows:

[0063] Batch 1 example set D1:

[0064] Batch 2 example set D2:

[0065] Batch 3 example set D3:

[0066] Batch m example set D m :

[0067] Where 4≤m≤M-1, m∈N *

[0068] Batch M example set D M :

[0069] 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:

[0070]

[0071] 3. Establish a serial training structure of G units for training examples with different time intervals

[0072] Figure 3 The diagram is a schematic diagram of a serial training structure composed of multiple G units. First, build a network of G units. The G unit 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 f.

[0073] For the training examples with different time intervals obtained in step 2, the G units are connected in series according to the number of time intervals to obtain the corresponding series training structure: (G) 1 、(G) 2 、(G) 3 ...(G) K , where each G unit corresponds to a set of network weight parameters (for convenience, network weight parameters and bias parameters are collectively referred to as network weight parameters), and the specific parameters of each series training structure are:

[0074] (G) 1 There is 1 G unit in: The weight parameters are

[0075] (G) 2 There are 2 G units in: The weight parameters are

[0076] (G) 3 There are 3 G units in: The weight parameters are

[0077]

[0078] (G) K There are n G units in : The weight parameters are

[0079] The initial weight parameters of each series training structure are set to the same random value: W0, that is:

[0080] Before training begins, it is equal to W0.

[0081] 4. Calculate model loss value

[0082] The starting time point T of the training examples of group c a The multi-dimensional data is input into the series training structure corresponding to the number of time intervals obtained in step 3 to calculate the loss value.

[0083]

[0084] c starting time points T in a Multidimensional data Input (G) n , get the 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

[0085] 5. Calculate the gradient through loss-restricted backpropagation and update the network weight parameters through sequential cycle optimization

[0086] Figure 4 This is a diagram of the loss-restricted backpropagation of the serial training structure. The loss value of each serial training structure obtained in step 4 is used to calculate the gradient of each serial training structure through restricted backpropagation and update the network weight parameters through sequential loop optimization. The specific calculation process is as follows:

[0087] For a series training structure with a time interval of 1, which contains only one G unit, the network uses the error back propagation training mechanism to directly obtain the updated gradient of the G unit. For a series training structure with a time interval greater than 1, multiple G units are connected in series, and the error back propagation mechanism is used to calculate the gradient. The gradient on the last G unit is intercepted as the updated gradient of the G unit. The specific calculation formula is:

[0088] (G) 1 Only calculate The updated gradient of the unit, i.e.

[0089] (G) 2 Only calculate The updated gradient of the unit, i.e.

[0090] (G) 3 Only calculate The updated gradient of the unit, i.e.

[0091]

[0092] (G) K Only calculate The updated gradient of the unit, i.e.

[0093] The above gradient calculation method only calculates the gradient of the last structure G unit of the series training structure each time, so the present invention calls this method loss-limited backpropagation.

[0094] Based on the gradients calculated by the above method, for serial training structures with different time intervals, the gradients of the serial training structures with smaller time intervals are calculated first in order of the number of time intervals. The weights are then updated using the SGD stochastic gradient descent method. The updated weights are then shared with all other serial training modules. This training method is called sequential cyclic optimization. The number of cycles in the cyclic training is set to H, where the training process for the tth cycle is:

[0095] Calculate (G) first 1 middle The update weight of the unit, that is, Among them, α represents the learning rate of the SGD gradient descent method; Indicates the end of the t-1th cycle training (G) 1 middle The weight parameters after unit update, Indicates the tth cycle (G)1 middle The weight parameters after unit update; Gradient after unit update The weight parameters are shared with other series training structures:

[0096] are equal to

[0097] Calculation (G) 2 middle The update weight of the unit, that is, Then Gradient after unit update The weight parameters are shared with all other serial training structures.

[0098] Update (G) using the same calculation method 3 、(G) 4 ...(G) K , when this cycle training calculates G K middle After the unit's weights are updated and shared with other serial training structures, this cycle training ends and enters the next cycle training calculation. Ultimately, H cycles of training calculations are required.

[0099] 6. Obtain the result model after all training structure losses converge

[0100] After the cycle training in step 5 is completed, the K series training structures (G) are trained 1 、(G) 2 、(G) 3 ...(G) K By inputting all training examples of the corresponding time interval as in step 4, we can get the full data loss value of each series training module. The specific calculation is:

[0101] Calculation (G) 1 The loss value is

[0102] Calculation (G) 2 The loss value is

[0103] Calculation (G) 3 The loss value is

[0104]

[0105] Calculation (G) K The loss value is

[0106] Determine Loss1, Loss2, Loss3,…,Loss K Are they all converged? If so, proceed to the next step, otherwise continue with steps 4 and 5 until Loss1, Loss2, Loss3,…, Loss K All converged.

[0107] 7. Adjust model hyperparameters and optimize the resulting model

[0108] After completing the training and obtaining the resulting model in step 6, adjust the model's hyperparameters: the network structure parameters of the G unit (the number of nodes per layer e and the number of hidden layers f), the learning rate α of the SGD stochastic gradient descent method, the number of training examples in a single training cycle c, and the number of training cycles H. After adjusting the hyperparameter combination, repeat steps 4-6 until the network loss difference converges, and terminate the training.

[0109] 8. Get the optimal model

[0110] After the hyperparameter optimization in step 7, the optimal G unit is obtained, and two sets of parameters, the G unit network structure and network weights, are saved. The prediction accuracy of the model is then verified using test data.

[0111] 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…T K+1 ), each time point has q indicators, and the test data is organized in the manner of steps 1-2 as follows:

[0112]

[0113] in, There are Total N l For example, l≤K, l∈N * .

[0114] Calculation (G) 1 The prediction accuracy of

[0115] Calculation (G) 2 The prediction accuracy of

[0116] Calculation (G) 3 The prediction accuracy of

[0117]

[0118] Calculation (G) K The prediction accuracy of

[0119] The obtained biological fermentation system mechanism model (G unit) is used to predict the accuracy of system state changes in the biological fermentation system at different time intervals.

[0120] The present invention also discloses a device for modeling a multidimensional data time series using the method for modeling a multidimensional data time series, comprising:

[0121] Acquisition unit: used to normalize and organize the original observation data to obtain standardized observation data in a unified form;

[0122] Organization unit: used to organize the training examples for artificial neural network training based on normative observation data;

[0123] Building unit: used to design the structure of artificial neural network and establish artificial neural network model;

[0124] Training unit: used to train the established artificial neural network model using training examples to obtain the parameter matrix of the artificial neural network;

[0125] 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;

[0126] The designed neural network structure has the following characteristics:

[0127] Its basic structure is a neural network unit called a unit time forward change generator (G unit), whose input layer and output layer have the same dimension; this neural network unit is used to model the changes that occur in multidimensional observation data after a unit time;

[0128] The serial training structure is obtained by connecting the G units in series, and the serial training structure is used to model the changing rules of multi-dimensional observation data after multiple unit times.

[0129] 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 method for modeling time series data of multi-dimensional state changes of a system, characterized in that: The model is used to model the multi-dimensional state change time series data of the fermentation system, including the following steps: 1) Normalize and organize the original observation data to obtain standardized observation data in a unified form; the time series data of the fermentation process refers to the sequence data with time point labels collected at various time points during the fermentation process of the microorganism. Each time point includes a set of fixed indicators, including metabolome data, fermentation process data, and transcriptome data; 2) organizing training examples for artificial neural network training based on standardized observation data; the training data is time series data measured during actinomycete fermentation with acarbose as the target product; each time point has a q-dimensional indicator data value; 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; The neural network structure designed in step 3) has the following characteristics: Its basic structure is a neural network unit, called the unit time positive change generator, G unit. The G unit is a fully connected neural network used to simulate the driving mechanism of the system's state change in unit time. Its input layer and output layer have the same dimension. This neural network unit is used to model the changes in multidimensional observation data after a unit time. The serial training structure is obtained by connecting G units in series. The serial training structure is used to model the changing patterns of multidimensional observation data after multiple unit times, and simulate the state changing patterns of the system in multiple unit times; the biological fermentation system mechanism model, G unit, is obtained, which is used to predict the expected accuracy of system state changes in the biological fermentation system at different time intervals.

2. The method for modeling time series data of multi-dimensional state changes of a system according to claim 1, characterized in that: The step 1) is specifically as follows: the multidimensional time series observation data is organized into a four-tuple organization form, namely batch, time, indicator, and value, and multiple batches of multidimensional time series observation data are obtained through observation of multiple batches. 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.

3. The method for modeling time series data of system multi-dimensional state changes according to claim 1, characterized in that: The step 2) is specifically as follows: organizing the system state changes at any two time points in the same batch into a training example, obtaining a set of training examples with different time intervals, wherein each training example includes the states of two time points, the data at the earlier time point is called the starting time point data of the training example; the data at the later time point is called the ending time point data of the training example; the starting time point data and the ending time point data of each training example are expressed as an organization form of observation data quadruple.

4. The method for modeling time series data of multi-dimensional state changes of a system according to claim 1, characterized in that: The forward change generator, G unit, adopts a fully connected structure.

5. The method for modeling time series data of system multi-dimensional state changes according to claim 4, characterized in that: The serial training structure is established as follows: for training examples with different time intervals, according to the number of time intervals, each time interval is recorded as n unit time, and a serial training structure composed of G units is established. The serial training structure connects n G units end to end in series.

6. The method for modeling system multi-dimensional state change time series data according to claim 5, characterized in that: The serial training structure with n time intervals is called (G) n , (G) n Represents a series structure of n G units, used to model the state change law within n unit time. Different values ​​of n correspond to training examples with different time spans. When n = 1 (G) 1 There is only one G unit in the series training structure (G) 1 The corresponding training data is a training example with a time interval of 1; when n>1, (G) n The serial training structure is obtained by connecting n G units in series, and the serial training structure (G) n The corresponding training data is training examples with n time intervals.

7. The method for modeling time series data of multi-dimensional state changes of a system according to any one of claims 1 to 6, characterized in that: The training method of step 4) is specifically as follows: 4.1) Calculate the loss value of each serial training structure: For each serial training structure (G) n , and the corresponding training example is recorded as S i , note S i The starting time point is T i , S i The end time point is T i+n , record the modeled system in T i The status data at a point in time is In T i+n The status data at a point in time is T i Data at a point in time Input to the cascade training structure (G) n , get the network output data Output data based on the network With T i+n Real data at a point in time The model loss value Loss can be calculated using the formula: 4.2) Based on the model loss value, the gradient of the backpropagation layer with limited loss is calculated: For a series training structure with a time interval of 1, which contains only one G unit, the network uses the error backpropagation training mechanism and directly uses the error backpropagation method to obtain the updated gradient of the G unit; For a series training structure with a time interval greater than 1, which consists of multiple G units in series, the error backpropagation method is used to calculate the gradient, and the gradient of the last G unit is intercepted as the updated gradient of the G unit of the series training structure; 4.3) Sequential Weight Update of Serial Training Structures: Based on the aforementioned training mechanism with limited loss backpropagation layers, multiple serial training structures are sequentially weighted, and network weight parameters are shared during the sequential weight update process. Specifically, for serial training structures at different time intervals, the gradient of the serial training structure is calculated in a certain order, and the weight is updated using the gradient descent method. The updated weights are then shared with all other serial training structures. 4.4) For K series training structures (G) 1 、(G) 2 、(G) 3 ...(G) K Calculate the loss value of each serial training module and determine whether the loss value of each serial training structure has converged. If so, the result model is obtained. Otherwise, continue to repeat 4.1), 4.2), and 4.3) until the loss value of each serial training structure has converged and the result model is obtained.

8. The method for modeling time series data of system multi-dimensional state changes according to claim 7, characterized in that: The step 5) specifically refers to optimizing the modeling process from the level of model hyperparameters. The hyperparameters adjusted during the modeling process include: structural parameters of the G unit (the number of nodes in each layer of the network, the number of hidden layers); the learning rate of the gradient descent method; the number of training data examples input for each training; the number of cycles of the cyclic training; using different combinations of the above hyperparameters, according to steps 3) and 4), complete the work of establishing the neural network structure and neural network training to obtain a new result model; evaluate the fitting accuracy of the result model to the observed data under each hyperparameter combination, and select the optimal model.

9. A device for modeling time series data of multi-dimensional state changes of a system, characterized by: include: Acquisition unit: used to normalize and organize the original observation data to obtain standardized observation data in a unified form; Time series data of the fermentation process refers to the sequence data with time point labels collected at various time points during the fermentation process of microorganisms. Each time point includes a fixed set of indicators, including metabolome data, fermentation process data, and transcriptome data. Organization unit: used to organize training examples for artificial neural network training based on standardized observation data; the training data is time series data measured during the actinomycete fermentation process with acarbose as the target product; each time point has a q-dimensional indicator data value; 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; The designed neural network structure has the following characteristics: Its basic structure is a neural network unit, called the unit time positive change generator, G unit. The G unit is a fully connected neural network used to simulate the driving mechanism of the system's state change in unit time. Its input layer and output layer have the same dimension. This neural network unit is used to model the changes in multidimensional observation data after a unit time. The serial training structure is obtained by connecting G units in series. The serial training structure is used to model the changing patterns of multidimensional observation data after multiple unit times, and simulate the state changing patterns of the system in multiple unit times; the biological system mechanism model, G unit, is obtained, which is used to predict the accuracy expectation when the system state changes in the biological system at different time intervals.

Images