Erythromycin fermentation productivity prediction system based on spatio-temporal feature fusion
Through the erythromycin fermentation capacity prediction system with spatiotemporal characteristics fusion, the multimodal sampling sensor and attention mechanism network is used to solve the problem of artificial experience dependence and insufficient prediction accuracy in traditional erythromycin production, and improve the accuracy of production capacity prediction Sex and efficiency.
Patent Information
- Application Number
- CN202410670147.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2024-11-15
AI Technical Summary
The traditional erythromycin production process relies on human experience and cannot accurately respond to changes in complex biological environments, resulting in limited production efficiency and output. Mixed modeling and deep learning face model design, data processing and prediction accuracy in erythromycin production optimization and other technical obstacles.
The erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion is adopted, including multimodal sampling sensors, data preprocessing modules, feature attribute analysis modules, feature fusion modules and attention mechanism capacity prediction networks, which are carried out through spatiotemporal feature analysis and feature fusion. Capacity forecast.
The feature engineering of prediction technology is simplified, the complexity of algorithm parameters is reduced, the accuracy and efficiency of erythromycin fermentation capacity prediction is improved, and the problem of huge feature engineering and distortion of prediction effect of traditional methods is overcome.
Smart Images

Figure CN118966394A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of bioinformatics, hybrid modeling, deep learning and biopharmaceuticals, and in particular to an erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion. Background Art
[0002] With advances in bioengineering and data science, the biopharmaceutical industry is facing unprecedented opportunities and challenges. Optimizing the production process of erythromycin, a key antibiotic, directly impacts the efficiency and stability of the pharmaceutical supply chain. Traditionally, erythromycin production relies on experienced technicians operating and monitoring bioreactors. However, this method, which relies on human experience, often lacks accuracy when dealing with complex biological processes and cannot accurately respond to various microbial environmental changes, resulting in limited production efficiency and output.
[0003] In the modernization of biopharmaceuticals, the importance of data is becoming increasingly prominent. Vast amounts of biological experimental and production data contain crucial information for optimizing production processes, but the complexity and high-dimensionality of this data make traditional data processing methods difficult to handle. Furthermore, the nonlinear nature of biological reactions and the variability of environmental factors complicate analysis and prediction. Therefore, effectively extracting useful information from complex biological data and applying this information to optimize production processes has become the key to improving erythromycin production efficiency.
[0004] In the face of these challenges, hybrid modeling technology, due to its excellent feature extraction capabilities, strong interpretability, and robust model generalization, is considered a powerful tool for addressing these issues. Despite this, the effective application of hybrid modeling and deep learning to erythromycin production optimization still faces numerous technical hurdles, such as model design and optimization, processing and analysis of large-scale biological data, and the accuracy and interpretability of model predictions. To address these challenges, developing a comprehensive hybrid model system integrating deep learning and bioinformatics is crucial for improving erythromycin production efficiency and ensuring the stability and security of the pharmaceutical supply chain. Summary of the Invention
[0005] The purpose of the present invention is to address the deficiencies of the existing technology and provide an erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion.
[0006] The object of the present invention is achieved through the following technical solutions: an erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion, the system comprising: an erythromycin fermentation process module, a multimodal sampling sensor, a central control platform, an online database, a capacity prediction system with spatiotemporal feature fusion, and a prediction result display; the erythromycin fermentation process module is connected to the multimodal sampling sensor, the multimodal sampling sensor is connected to the database, and the central control platform is connected to the erythromycin fermentation process module and the database respectively; the database is connected to the input end of the capacity prediction system with spatiotemporal feature fusion, and the output end of the capacity prediction system with spatiotemporal feature fusion is connected to the prediction result display;
[0007] The multimodal sampling sensor is used to sample the erythromycin process data of the variables in each erythromycin fermentation process in the erythromycin fermentation process module to obtain an erythromycin process data set.
[0008] The central control platform is used to control the sampling frequency of the multimodal sampling sensor in the erythromycin process and control the operation of the online database;
[0009] The online database is used to write, query, modify and export erythromycin process data sets
[0010] The production capacity prediction system based on the fusion of time and space features is used to collect the erythromycin process data set input into the online database. Perform capacity forecast and obtain the capacity forecast result set Y Pred ;
[0011] The forecast result display device is used to display the production capacity forecast result set Y Pred to display.
[0012] Furthermore, the capacity prediction system of spatiotemporal feature fusion includes a data preprocessing module, a feature attribute analysis module, a feature fusion module and an attention mechanism capacity prediction network.
[0013] Furthermore, the production capacity prediction system based on the fusion of spatiotemporal features is used to collect the erythromycin process data set input from the online database. Perform capacity forecast and obtain the capacity forecast result set Y Pred , specifically including the following sub-steps:
[0014] (a.1) Data preprocessing module processes the input erythromycin process data set Perform standardization to obtain the standardized detection data vector set D i ;
[0015] (a.2) The feature attribute analysis module analyzes the standardized detection data vector set D through the collected samples and mechanism equations i The temporal and spatial attributes of the features are used to establish a spatiotemporal feature network between the features;
[0016] (a.3) The feature fusion module performs dual independent channel feature fusion on the spatiotemporal feature network obtained by the feature attribute analysis module, and the time domain feature is fused to obtain the time domain set T * , spatial features realize spatial feature fusion to obtain the spatial set S * , and input into the attention mechanism capacity prediction network;
[0017] (a.4) Attention Mechanism Capacity Prediction Network uses the attention mechanism in the field of natural language processing to design the encoding layer, decoding layer and prediction layer output, and input the time domain set T * and the airspace set S * Vectorize and finally predict the capacity indicators and output the capacity prediction result set Y Pred .
[0018] Furthermore, the step (a.1) specifically includes the following sub-steps:
[0019] (a.1.1) The input of the production capacity prediction system fused with spatiotemporal features is the erythromycin process data set The erythromycin process data set for: Among them, there are 19 variables in the erythromycin fermentation process: x1, ..., x j ,…,x 19 ; Represents the variable x j In the i-th erythromycin process, the data vector is detected, and each detection data vector There are Q detection data in
[0020] (a.1.2) The data preprocessing module processes the input erythromycin process data set Perform standardization to obtain the standardized detection data vector set D i :D i ={x i1 ,x i2 ,…,x ij ,…,x i19}, the standardization formula is as follows:
[0021]
[0022] in, Represents a detection data set The arithmetic mean of each test data; Represents a detection data set The standard deviation of each test data; ij is the detection data vector after standard post-processing,
[0023] Furthermore, the step (a.2) specifically includes the following sub-steps:
[0024] (a.2.1) The feature attribute analysis module first performs time attribute analysis through the differential equation in the time domain: Assume that for any variable x j and variable x k There is a characteristic relationship of time attributes between them, j≠k, k=1,…,k,…,19, variable x j and variable x k The rate of change between can be expressed by a time-domain discrete difference equation or a continuous differential equation; thus assuming that for the variable x j and variable x k There is a characteristic relationship of time attributes between them, then:
[0025] x j (t) = f(x k (t),t) (2);
[0026] Among them, x j (t) is the mapping relationship of variable xj with respect to time t; x k (t) is the variable x k The mapping relationship of time t changes; f(·) is the mapping relationship x j (t) About time t and mapping relationship x k (t) binary function relationship;
[0027] Under the condition of time domain differentiation, we get:
[0028]
[0029] in, is f(·) with respect to x k The partial derivative of x describes the j How does a small change in x affect k The value of Direct dependence on time;
[0030] By solving the above differential equation: If the equation has a solution, then it proves that the variable x j and variable x kThere is indeed a characteristic relationship of time attributes between them, and then regression analysis, system identification and significance test are performed on the solution parameters to determine the specific form of the function f(·) and the accuracy and reliability of the solution results, and then the variable x is obtained. j and variable x k The characteristic relationship T(x k (t),x j (t)); If the equation has no solution, then prove that the variables xj and x k There is indeed no characteristic relationship of time attributes between them;
[0031] (a.2.2) The feature attribute analysis module repeats step (a.2.1) for every two variables among the 19 variables to obtain a feature relationship network in the time domain;
[0032] (a.2.3) The feature attribute analysis module then performs spatial attribute analysis through differential equations in the spatial domain: Assume that for the variable x j and variable x k There is a characteristic relationship of spatial attributes between variables x j and variable x k The changes and correlations between different spatial dimensions can be expressed by spatial discrete difference equations or continuous differential equations; thus, it is assumed that for the variable x j and variable x k There is a characteristic relationship of spatial attributes between them, then:
[0033] x j (s) = g(x k (s),s) (4);
[0034] Among them, x j (s) is the mapping relationship between the variable xj and the spatial position s; k (s) is the variable x k The mapping relationship of the spatial position s; g(·) is the mapping relationship x j (s) About the spatial position s and the mapping relationship x k (s) binary functional relationship;
[0035] Under the condition of spatial differentiation, we get:
[0036]
[0037] in, is g(·) with respect to x k The partial derivative of x describes the situation when the space is fixed. j How does a small change in x affect k The value of Direct dependence on space;
[0038] In particular, in the erythromycin fermentation process, the spatial dimension involved is two-dimensional space, from which we can deduce the specific characteristic relationship form in the two-dimensional space during the erythromycin fermentation process:
[0039]
[0040] By solving the above differential equation: If the equation has a solution, then it proves that the variable x j and variable x k There is indeed a characteristic relationship of time attributes between them; then regression analysis, system identification and significance test are performed on the solution parameters to determine the specific form of the function g(·) and the accuracy and reliability of the solution results; then the variable x is obtained j and variable x k The feature relationship S(x k (s),x j (s)); If the equation has no solution, then prove that the variables xj and x k There is indeed no characteristic relationship of time attributes between them;
[0041] (a.2.4) The feature attribute analysis module repeats step (a.2.3) for every two variables among the 19 variables to obtain a feature relationship network in the spatial domain;
[0042] (a.2.5) The feature relationship network in the time domain and the feature relationship network in the spatial domain are added together to obtain the spatiotemporal feature network.
[0043] Furthermore, the step (a.3) specifically includes the following sub-steps:
[0044] (a.3.1) The feature fusion module includes a time domain feature fusion submodule and a spatial domain feature fusion submodule;
[0045] (a.3.2) The time domain feature fusion submodule is used to fusion each feature relationship T(x k (t),x j (t)) Use interlaced matrix transformation in chronological order to facilitate the subsequent attention mechanism capacity prediction module to capture time features
[0046]
[0047] The characteristic relationship T(x k (t),x j The matrix form of (t)) is as follows:
[0048]
[0049] Assume that the variable xj For variable x k With the previous time step, after interleaving transformation, the time feature relationship after matrix transformation fusion is obtained The specific matrix form is as follows:
[0050]
[0051] (a.3.3) and all the time feature relations T(x k (t),x j (t)) * Stored in the form of a set, recorded as the time domain set T * ;
[0052] (a.3.4) The spatial feature fusion submodule is used to fusion multiple spatial feature relationships S(x k (s),x j (s)) is obtained using spatial fusion weights
[0053]
[0054] Where c represents the number of variables with spatial feature relationships in the spatiotemporal feature network; W c The weight matrix of the variables that represent the spatial feature relationship with respect to the spatial feature relationship;
[0055] (a.3.5) and all the spatial feature relations S(x k (s),x j (s)) * Stored in the form of a set, recorded as the spatial set S * ;
[0056] (a.3.6) The feature fusion module then transforms the time domain set T * and the airspace set S * Input to the attention mechanism capacity prediction network.
[0057] Furthermore, the step (a.4) specifically includes the following sub-steps:
[0058] (a.4.1) Initialize model parameters: Set the hidden layer size, number of layers, and number of heads of the encoding and decoding layers of the attention mechanism model; for each head of the encoding and decoding layers, initialize its corresponding weight matrix W Q 、W K 、W V Used to generate query vector, key vector and value vector and output weight matrix W o ;
[0059] (a.4.2) Vectorize the input sequence: Convert the input time domain fusion features and spatial domain features into vector representation through the embedding layer; let the input sequence be The embedding is represented as and
[0060]
[0061] in, Indicates that in the time domain set T * middle Number of occurrences; Indicates that in the airspace set S * middle Number of occurrences; Len(T * ) represents the time domain set T * Total length of Len(S * ) represents the spatial set S * Total length; d model Indicates the dimension of the attention mechanism capacity prediction network; idxT jk Indicates that in the time domain set T * middle The index order of idxS jk Indicates that in the airspace set S * middle The index order of
[0062] (a.4.3) Encoding layer processing: The encoding layer receives the embedded input vector and processes it through the self-attention mechanism. The calculation formula for self-attention is as follows:
[0063]
[0064] in, express The product of the vectorized result and the query vector; express The product of the vectorized result and the key vector; express The product of the vectorized result and the value vector; express The product of the vectorized result and the query vector; express The product of the vectorized result and the key vector; express The product of the vectorized result and the value vector; (·) T represents the transpose operation; d represents the time domain set T * The d-th time feature relationship in h represents the spatial domain set S * The h-th spatial feature relationship
[0065] (a.4.4) Decoding layer processing: The decoding layer uses the output of the encoding layer and the previous output to generate the current output. The attention mechanism of the decoding layer is divided into two parts: masked self-attention and encoder-decoder attention.
[0066] (a.4.4.1) Masked self-attention ensures that only the output before the current moment is used when generating the current moment output;
[0067] (a.4.4.2) Encoding-decoding attention uses the output of the encoding layer as the key K and value V, and the output of the previous layer of the decoding layer as the query Q to achieve decoding:
[0068]
[0069] Among them, F DecodingOutput The final representation function of the coding layer; Represents the representation function presented by the first encoding layer; represents the representation function presented by the second coding layer; Q dec represents the decoding result of the query vector; K enc Represents the encoding result of the key vector; V enc Represents the encoding result of the value vector; K enc Represents the encoding result of the key vector; Q dec Represents the decoding result of the query vector; AttentionOutput represents the decoding result of the query vector. jk The matrix composed of
[0070] This realizes the loop solution of encoding-decoding attention, realizes the update of attention matrix, and then realizes the update of model;
[0071] The application of encoding-decoding attention loop solving technology in the field of biological intelligent manufacturing can improve computing efficiency, optimize computing resources, enhance system flexibility and reduce operational complexity.
[0072] (a.4.5) Prediction layer output: The output of the decoding layer passes through a linear layer and outputs the capacity prediction result set Y Pred :
[0073] Y Pred =(W Pred ×F DecodingOutput +b Pred ) (13);
[0074] Among them, W Predrepresents the weight matrix of the prediction layer; b Pred Represents the bias matrix of the prediction layer.
[0075] The beneficial effects of the present invention are as follows: the present invention predicts the production capacity of the erythromycin fermentation process, overcomes the problems of large feature engineering, complex algorithm parameter scale, and distorted prediction effect in existing production capacity prediction technologies, introduces a feature attribute analysis module and a feature fusion module to perform feature spatiotemporal attribute analysis and spatiotemporal feature fusion on the collected samples, thereby simplifying the feature engineering of the prediction technology, reducing algorithm parameters, and improving the prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0076] Figure 1 This is a structural diagram of an erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion;
[0077] Figure 2 This is the structural diagram of the capacity forecasting system; Figure 3 This is a simplified production process diagram of erythromycin fermentation process;
[0078] In the figure, 1-erythromycin fermentation process module; 2-multimodal sampling sensor; 3-central control platform; 4-online database; 5-capacity prediction system based on spatiotemporal feature fusion; 6-prediction result display; 7-data preprocessing module; 8-feature attribute analysis module; 9-temporal feature fusion submodule; 10-spatial feature fusion submodule; 11-feature fusion module; 12-attention mechanism capacity prediction network. DETAILED DESCRIPTION
[0079] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to illustrate the present invention, rather than to represent all embodiments. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.
[0080] Example 1
[0081] like Figure 1As shown, the present invention provides an erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion, the system comprising: an erythromycin fermentation process module 1, a multimodal sampling sensor 2, a central control platform 3, an online database 4, a capacity prediction system 5 with spatiotemporal feature fusion, and a prediction result display 6; the erythromycin fermentation process module 1 is connected to the multimodal sampling sensor 2, the multimodal sampling sensor 2 is connected to the database 4, the central control platform 3 is connected to the erythromycin fermentation process module 1 and the database 4, respectively; the database 4 is connected to the input end of the capacity prediction system 5 with spatiotemporal feature fusion, and the output end of the capacity prediction system 5 with spatiotemporal feature fusion is connected to the prediction result display 6.
[0082] The multimodal sampling sensor 2 is used to sample the erythromycin process data of the variables in each erythromycin fermentation process in the erythromycin fermentation process module 1 to obtain the erythromycin process data set.
[0083] The central control platform 3 is used to control the sampling frequency of the multimodal sampling sensor 2 during the erythromycin process and to control the operation of the online database 4 .
[0084] The online database 4 is used to write, query, modify and export erythromycin process data sets
[0085] The production capacity prediction system 5 based on the fusion of time and space features is used to collect the erythromycin process data set input into the online database. Perform capacity forecast and obtain the capacity forecast result set Y Pred .
[0086] The capacity prediction system 5 with spatiotemporal feature fusion includes a data preprocessing module 7, a feature attribute analysis module 8, a feature fusion module 11 and an attention mechanism capacity prediction network 12.
[0087] The production capacity prediction system 5 based on the fusion of time and space features processes the erythromycin process data set input from the online database 4. Perform capacity forecast and obtain the capacity forecast result set Y Pred , specifically including the following sub-steps:
[0088] (a.1) Data preprocessing module 7 processes the input erythromycin process data set Perform standardization to obtain the standardized detection data vector set D i .
[0089] The step (a.1) specifically includes the following sub-steps:
[0090] (a.1.1) The input of the spatiotemporal feature fusion capacity prediction system 5 is the erythromycin process data set The erythromycin process data set for: Among them, there are 19 variables in the erythromycin fermentation process: x1, ..., x j ,…,x 19 ; Represents the variable x j In the i-th erythromycin process, the data vector is detected, and each detection data vector There are Q detection data in
[0091] Reference Figure 3 The variables of the erythromycin fermentation process are shown in Table 1.
[0092] Table 1: Erythromycin fermentation process variables
[0093]
[0094]
[0095] Erythromycin process data is used as input variables for the spatiotemporal feature fusion capacity prediction system 5. It is obtained through manual sampling and analysis, and is collected and analyzed every hour.
[0096] (a.1.2) Since the units and scales of the variables are different, all data are standardized to prevent errors caused by different units. The data preprocessing module 7 processes the input erythromycin process data set. Perform standardization to obtain the standardized detection data vector set D i :D i ={x i1 ,x i2 ,…,x ij ,…,x i19}, the standardization formula is as follows:
[0097]
[0098] in, Represents a detection data set The arithmetic mean of each test data; Represents a detection data set The standard deviation of each test data; ij is the detection data vector after standard post-processing,
[0099] (a.2) The characteristic attribute analysis module 8 analyzes the standardized detection data vector set D through the collected samples and the mechanism equation i The temporal and spatial attributes of features are used to establish a spatiotemporal feature network between features.
[0100] The step (a.2) specifically includes the following sub-steps:
[0101] (a.2.1) The characteristic attribute analysis module 8 first performs time attribute analysis through the differential equation in the time domain: Assume that for any variable x j and variable x k There is a characteristic relationship of time attributes between them, j≠k, k=1,…,k,…,19, variable x j and variable x k The rate of change between can be expressed by a time-domain discrete difference equation or a continuous differential equation; thus assuming that for the variable x j and variable x k There is a characteristic relationship of time attributes between them, then:
[0102] x j (t) = f(x k (t),t) (2);
[0103] Among them, x j (t) is the mapping relationship of variable xj with respect to time t; x k (t) is the variable x k The mapping relationship of time t changes; f(·) is the mapping relationship x j (t) About time t and mapping relationship x k (t) is a binary function relationship.
[0104] Under the condition of time domain differentiation, we get:
[0105]
[0106] in, is f(·) with respect to x k The partial derivative of x describes the j How does a small change in x affect k The value of Direct dependence on time.
[0107] By solving the above differential equation, we can understand the variable x j and variable x k Dynamic relationship between: If the equation has a solution, then the variable x j and variable x kThere is indeed a characteristic relationship of time attributes between them, and then regression analysis, system identification and significance test are performed on the solution parameters to determine the specific form of the function f(·) and the accuracy and reliability of the solution results, and then the variables xj and x are obtained. k The characteristic relationship T(x k (t),x j (t)); If the equation has no solution, then prove that the variables xj and x k There is indeed no characteristic relationship of time attributes between them.
[0108] (a.2.2) The feature attribute analysis module 8 repeats step (a.2.1) for every two variables among the 19 variables to obtain a feature relationship network in the time domain.
[0109] (a.2.3) The feature attribute analysis module 8 then performs spatial attribute analysis through differential equations in the spatial domain: Assume that for the variable x j and variable x k There is a characteristic relationship of spatial attributes between variables x j and variable x k The changes and correlations between different spatial dimensions can be expressed by spatial discrete difference equations or continuous differential equations; thus, it is assumed that for the variable x j and variable x k There is a characteristic relationship of spatial attributes between them, then:
[0110] x j (s) = g(x k (s),s) (4);
[0111] Among them, x j (s) is the mapping relationship between the variable xj and the spatial position s; k (s) is the variable x k The mapping relationship of the spatial position s; g(·) is the mapping relationship x j (s) About the spatial position s and the mapping relationship x k (s) is a binary function relationship.
[0112] Under the condition of spatial differentiation, we get:
[0113]
[0114] in, is g(·) with respect to x k The partial derivative of x describes the situation when the space is fixed. j How does a small change in x affect k The value of Direct dependence on space.
[0115] In particular, in the erythromycin fermentation process, the spatial dimension involved is two-dimensional space, from which we can deduce the specific characteristic relationship form in the two-dimensional space during the erythromycin fermentation process:
[0116]
[0117] By solving the above differential equation: If the equation has a solution, then it proves that the variable x j and variable x k There is indeed a characteristic relationship of time attributes between them; then regression analysis, system identification and significance test are performed on the solution parameters to determine the specific form of the function g(·) and the accuracy and reliability of the solution results; then the variable x is obtained j and variable x k The feature relationship S(x k (s),x j (s)); If the equation has no solution, then prove that the variables xj and x k There is indeed no characteristic relationship of time attributes between them.
[0118] (a.2.4) The feature attribute analysis module 8 repeats step (a.2.3) for every two variables among the 19 variables to obtain a feature relationship network in the spatial domain.
[0119] (a.2.5) The feature relationship network in the time domain and the feature relationship network in the spatial domain are added together to obtain the spatiotemporal feature network.
[0120] (a.3) The feature fusion module 11 performs dual independent channel feature fusion on the spatiotemporal feature network obtained by the feature attribute analysis module 8, and the time domain feature is fused to obtain the time domain set T * , spatial features realize spatial feature fusion to obtain the spatial set S * , and input into the attention mechanism capacity prediction network 12.
[0121] The step (a.3) specifically includes the following sub-steps:
[0122] (a.3.1) The feature fusion module 11 includes a time domain feature fusion submodule 9 and a spatial domain feature fusion submodule 10.
[0123] (a.3.2) The time domain feature fusion submodule 9 performs the fusion of each feature relationship T(x k (t),x j (t)) Use interlaced matrix transformation in chronological order to facilitate the subsequent attention mechanism capacity prediction module to capture time features
[0124] The characteristic relationship T(x k (t),x j The matrix form of (t)) is as follows:
[0125]
[0126] Assume that the variable x j For variable x k With the previous time step, after interleaving transformation, the time feature relationship after matrix transformation fusion is obtained The specific matrix form is as follows:
[0127]
[0128] (a.3.3) and all the time feature relations T(x k (t),x j (t)) * Stored in the form of a set, recorded as the time domain set T * .
[0129] The original multi-dimensional time-domain feature relationship is compressed into a univariate time series feature relationship, making the internal characteristics of the time sequence between each time-domain feature relationship more significant, providing high-dimensional time-domain feature fusion input data for the subsequent attention mechanism capacity prediction module, thereby improving the prediction effect.
[0130] (a.3.4) The spatial feature fusion submodule 10 is used to fusion multiple spatial feature relations S(x k (s),x j (s)) is obtained using spatial fusion weights
[0131]
[0132] Where c represents the number of variables with spatial feature relationships in the spatiotemporal feature network; W c The weight matrix of the variables that represent the spatial feature relationship.
[0133] (a.3.5) and all the spatial feature relations S(x k (s),x j (s)) * It is stored in the form of a set, recorded as the spatial set S*.
[0134] (a.3.6) The feature fusion module 11 then transforms the time domain set T * and the airspace set S * Input to the attention mechanism capacity prediction network 12.
[0135] (a.4) Attention Mechanism Capacity Prediction Network 12 uses the attention mechanism in the field of natural language processing to design the encoding layer, decoding layer and prediction layer output, and converts the input time domain set T * and the airspace set S * Vectorize and finally predict the capacity indicators and output the capacity prediction result set Y Pred .
[0136] The step (a.4) specifically includes the following sub-steps:
[0137] (a.4.1) Initialize model parameters: Set the hidden layer size, number of layers, and number of heads of the encoding and decoding layers of the attention mechanism model; for each head of the encoding and decoding layers, initialize its corresponding weight matrix W Q 、W K 、W V Used to generate query vector, key vector and value vector and output weight matrix W o .
[0138] (a.4.2) Vectorize the input sequence: Convert the input time domain fusion features and spatial domain features into vector representation through the embedding layer; let the input sequence be The embedding is represented as and
[0139]
[0140] in, Indicates that in the time domain set T * middle Number of occurrences; Indicates that in the airspace set S * middle Number of occurrences; Len(T * ) represents the time domain set T * Total length of Len(S * ) represents the spatial set S * Total length; d model Indicates the dimension of the attention mechanism capacity prediction network; idxT jk Indicates that in the time domain set T * middle The index order of idxS jk Indicates that in the airspace set S * middle The index order of .
[0141] (a.4.3) Encoding layer processing: The encoding layer receives the embedded input vector and processes it through the self-attention mechanism. The calculation formula for self-attention is as follows:
[0142]
[0143] in, express The product of the vectorized result and the query vector; express The product of the vectorized result and the key vector; express The product of the vectorized result and the value vector; express The product of the vectorized result and the query vector; express The product of the vectorized result and the key vector; express The product of the vectorized result and the value vector; (·) T represents the transpose operation; d represents the time domain set T * The d-th time feature relationship in h represents the spatial domain set S * The h-th spatial feature relationship
[0144] (a.4.4) Decoding layer processing: The decoding layer uses the output of the encoding layer and the previous output to generate the output at the current moment; the attention mechanism of the decoding layer is divided into two parts: masked self-attention and encoding-decoding attention.
[0145] (a.4.4.1) Masked self-attention ensures that only the outputs before the current moment are used when generating the output at the current moment.
[0146] (a.4.4.2) Encoding-decoding attention uses the output of the encoding layer as the key K and value V, and the output of the previous layer of the decoding layer as the query Q to achieve decoding:
[0147]
[0148] Among them, F DecodingOutput The final representation function of the coding layer; Represents the representation function presented by the first encoding layer; represents the representation function presented by the second coding layer; Q dec represents the decoding result of the query vector; K enc Represents the encoding result of the key vector; V enc Represents the encoding result of the value vector; K enc Represents the encoding result of the key vector; Q dec Represents the decoding result of the query vector; AttentionOutput represents the decoding result of the query vector.jk The matrix composed of .
[0149] This realizes the cyclic solution of encoding-decoding attention, updates the attention matrix, and then updates the model.
[0150] The application of encoding-decoding attention loop solving technology in the field of biological intelligent manufacturing can improve computing efficiency, optimize computing resources, enhance system flexibility and reduce operational complexity.
[0151] (a.4.5) Prediction layer output: The output of the decoding layer passes through a linear layer and outputs the capacity prediction result set Y Pred :
[0152] Y Pred =(W Pred ×F DecodingOutput +b Pred ) (13);
[0153] Among them, W Pred represents the weight matrix of the prediction layer; b Pred Represents the bias matrix of the prediction layer.
[0154] Through the above steps, the attention mechanism capacity prediction network 12 will independently fuse the input dual-channel spatiotemporal features, and encode and decode them through the attention mechanism, and finally output the prediction result Y of erythromycin fermentation capacity Pred .
[0155] The forecast result display device 6 is used to display the capacity forecast result set Y Pred to display.
[0156] There are three indicators for erythromycin fermentation products, namely bacterial concentration, viscosity and chemical efficiency. The final production capacity prediction result is displayed on the result display device 6.
[0157] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A system for predicting erythromycin fermentation capacity based on spatiotemporal feature fusion, characterized in that: The system comprises: an erythromycin fermentation process module, a multimodal sampling sensor, a central control platform, an online database, a capacity prediction system of spatiotemporal feature fusion, and a prediction result display instrument; the erythromycin fermentation process module is connected to the multimodal sampling sensor, the multimodal sampling sensor is connected to the database, and the central control platform is connected to the erythromycin fermentation process module and the database respectively; the database is connected to the input end of the capacity prediction system of spatiotemporal feature fusion, and the output end of the capacity prediction system of spatiotemporal feature fusion is connected to the prediction result display instrument; The multimodal sampling sensor is used to sample the erythromycin process data of the variables in each erythromycin fermentation process in the erythromycin fermentation process module to obtain an erythromycin process data set. The central control platform is used to control the sampling frequency of the multimodal sampling sensor in the erythromycin process and control the operation of the online database; The online database is used to write, query, modify and export erythromycin process data sets The production capacity prediction system integrating spatiotemporal features is used to collect the erythromycin process data set input into the online database. Carry out capacity forecast and obtain the capacity forecast result set Y Pred ; The prediction result display device is used to display the capacity prediction result set Y Pred to display.
2. The erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion according to claim 1, characterized in that: The capacity prediction system of spatiotemporal feature fusion includes a data preprocessing module, a feature attribute analysis module, a feature fusion module and an attention mechanism capacity prediction network.
3. A system for predicting erythromycin fermentation capacity based on spatiotemporal feature fusion according to claim 2, characterized in that: The production capacity prediction system based on the fusion of spatiotemporal features processes the erythromycin process data set D input into the online database. i *Perform capacity forecast and obtain capacity forecast result set Y Pred , specifically including the following sub-steps: (a.1) The data preprocessing module processes the input erythromycin process data set Perform standardization to obtain the standardized detection data vector set D i ; (a.2) The feature attribute analysis module analyzes the standardized detection data vector set D through the collected samples and mechanism equations i The temporal and spatial attributes of the features are used to establish a spatiotemporal feature network between the features; (a.3) The feature fusion module performs dual independent channel feature fusion on the spatiotemporal feature network obtained by the feature attribute analysis module, and the time domain feature realizes the time domain feature fusion to obtain the time domain set T * , the spatial features realize the fusion of spatial features to obtain the spatial set S * , and input into the attention mechanism capacity prediction network; (a.4) Attention Mechanism Capacity Prediction Network uses the attention mechanism in the field of natural language processing to design the encoding layer, decoding layer and prediction layer output, and converts the input time domain set T * and the space set S * Vectorize and finally predict the capacity indicators and output the capacity prediction result set Y Pred .
4. The erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion according to claim 3, characterized in that: The step (a.1) specifically includes the following sub-steps: (a.1.1) The input of the capacity prediction system with spatiotemporal feature fusion is the erythromycin process data set The erythromycin process data set for: Among them, there are 19 variables in the erythromycin fermentation process: x1, ..., x j , …, x 19 ; Represents variable x j In the i-th erythromycin process, the data vector is detected, and each detection data vector There are Q detection data in (a.1.2) The data preprocessing module processes the input erythromycin process data set Perform standardization to obtain the standardized detection data vector set D i :D i ={x i1 ,x i2 ,…,x ij ,…,x i19 }, the standardized formula is as follows: in, Represents a collection of detection data The arithmetic mean of each test data; Represents a collection of detection data The standard deviation of each test data; x ij is the detection data vector after standard post-processing, 5. The erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion according to claim 4, characterized in that: The step (a.2) specifically includes the following sub-steps: (a.2.1) The feature attribute analysis module first performs time attribute analysis through differential equations in the time domain: Assume that for any variable x j and variable x k There is a characteristic relationship of time attribute between them, j≠k, k=1,…,k,…,19, variable x j and variable x k The rate of change between can be expressed by a time-domain discrete difference equation or a continuous differential equation; thus assuming that for the variable x j and variable x k There is a characteristic relationship of time attribute between them, then: x j (t)=f(x k (t),t) (2); Among them, x j (t) is the mapping relationship of variable xj with respect to time t; x k (t) is the variable x k The mapping relationship of the change in time t; f(·) is the mapping relationship x j (t) About time t and mapping relationship x k (t) binary function relationship; Under the condition of time domain differentiation, we get: in, is f(·) with respect to x k The partial derivative of x describes the j How does a small change in x affect k The value of Direct dependence on time; By solving the above differential equation: If the equation has a solution, then it proves that the variable x j and variable x k There is indeed a characteristic relationship of time attributes between them. Then, regression analysis, system identification and significance test are performed on the solution parameters to determine the specific form of the function f(·) and the accuracy and reliability of the solution results, and then the variable x is obtained. j and variable x k The characteristic relationship T(x k (t),x j (t)); If the equation has no solution, then prove that the variables xj and x k There is indeed no characteristic relationship of time attributes between them; (a.2.2) The feature attribute analysis module repeats step (a.2.1) for every two variables among the 19 variables to obtain a feature relationship network in the time domain; (a.2.3) The feature attribute analysis module then performs spatial attribute analysis through differential equations in the spatial domain: Assume that for the variable x j and variable x k There is a characteristic relationship of spatial attributes between variables x j and variable x k The changes and correlations between different spatial dimensions can be expressed by spatial discrete difference equations or continuous differential equations; thus, it is assumed that for the variable x j and variable x k There is a characteristic relationship of spatial attributes between them, then: x j (s)=g(x k (s),s) (4); Among them, x j (s) is the mapping relationship between the variable xj and the spatial position s; k (s) is the variable x k The mapping relationship about the change of spatial position s; g(·) is the mapping relationship x j (s) Regarding the spatial position s and the mapping relationship x k (s) binary function relationship; Under the condition of spatial differentiation, we get: in, is g(·) with respect to x k The partial derivative of describes the situation when x is fixed in space. j How does a small change in x affect k The value of Direct dependence on space; In particular, in the erythromycin fermentation process, the space dimension involved is two-dimensional space, from which we can deduce the specific characteristic relationship form in the two-dimensional space during the erythromycin fermentation process: By solving the above differential equation: If the equation has a solution, then it proves that the variable x j and variable x k There is indeed a characteristic relationship of time attributes between them; then regression analysis, system identification and significance test are performed on the solution parameters to determine the specific form of the function g(·) and the accuracy and reliability of the solution results; then the variable x is obtained j and variable x k The feature relationship S(x k (s),x j (s)); If the equation has no solution, then prove that the variable x j and variable x k There is indeed no characteristic relationship of time attributes between them; (a.2.4) The feature attribute analysis module repeats step (a.2.3) for every two variables among the 19 variables to obtain a feature relationship network in the spatial domain; (a.2.5) The feature relationship network in the time domain and the feature relationship network in the spatial domain are added together to obtain the spatiotemporal feature network.
6. The erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion according to claim 5, characterized in that: The step (a.3) specifically includes the following sub-steps: (a.3.1) The feature fusion module includes a time domain feature fusion submodule and a spatial domain feature fusion submodule; (a.3.2) The time domain feature fusion submodule performs the fusion of each feature relation T(x k (t),x j (t)) uses interlacing to perform matrix transformation in chronological order, which makes it easier for the subsequent attention mechanism capacity prediction module to capture time features and obtain The characteristic relationship T(x k (t),x j The matrix form of (t) is as follows: Assume that the variable x j For variable x k With the previous time step, after interleaving transformation, the time feature relationship after matrix transformation fusion is obtained The specific matrix form is as follows: (a.3.3) and use the time domain feature fusion submodule to transform all time feature relations T(x k (t),x j (t)) * Stored in the form of a set, recorded as the time domain set T * ; (a.3.4) The spatial feature fusion submodule is used to fusion multiple spatial feature relations S(x k (s),x j (s)) is obtained by using spatial fusion weights Where c represents the number of variables with spatial feature relationships in the spatiotemporal feature network; W c The weight matrix of the variables with spatial feature relationships with respect to the spatial feature relationships; (a.3.5) and use the spatial feature fusion submodule to combine all the spatial feature relations S(x k (s),x j (s)) * Stored in the form of a set, recorded as the space domain set S * ; (a.3.6) Then the feature fusion module transforms the time domain set T * and the space set S * Input to the attention mechanism capacity prediction network.
7. The erythromycin fermentation capacity prediction system based on spatiotemporal feature fusion according to claim 6, characterized in that: The step (a.4) specifically includes the following sub-steps: (a.4.1) Initialize model parameters: Set the hidden layer size, number of layers, and number of heads of the encoding and decoding layers of the attention mechanism model; for each head of the encoding and decoding layers, initialize its corresponding weight matrix W Q , W K , W V Used to generate query vectors, key vectors, value vectors, and output weight matrix W o ; (a.4.2) Vectorize the input sequence: Convert the input time domain fusion features and spatial domain features into vector representation through the embedding layer; let the input sequence be The embedding is expressed as and in, Represents the time domain set T * middle Number of occurrences; Indicates that in the airspace set S * middle Number of occurrences; Len(T * ) represents the time domain set T * Total length of Len(S * ) represents the space set S * Total length; d model Indicates the dimension of the attention mechanism capacity prediction network; idxT jk Represents the time domain set T * middle The index order of idxS jk Indicates that in the airspace set S * middle The index order of (a.4.3) Encoding layer processing: The encoding layer receives the embedded input vector and processes it through the self-attention mechanism; the calculation formula of self-attention is as follows: in, express The product of the vectorized result and the query vector; express The product of the vectorized result and the key vector; express The product of the vectorized result and the value vector; express The product of the vectorized result and the query vector; express The product of the vectorized result and the key vector; express The product of the vectorized result and the value vector; (·) T represents the transposition operation; d represents the time domain set T * The d-th time feature relationship in h represents the spatial domain set S * The h-th spatial feature relationship in (a.4.4) Decoding layer processing: The decoding layer uses the output of the encoding layer and the previous output to generate the output at the current moment; the attention mechanism of the decoding layer is divided into two parts: masked self-attention and encoding-decoding attention; (a.4.4.1) Masked self-attention ensures that only the output before the current moment is used when generating the current moment output; (a.4.4.2) Encoding-decoding attention uses the output of the encoding layer as the key K and value V, and the output of the previous layer of the decoding layer as the query Q to achieve decoding: Among them, F DecodingOutput The final representation function of the coding layer is presented to the outside world; Represents the representation function presented by the first encoding layer; represents the representation function presented by the second coding layer; Q dec represents the decoding result of the query vector; K enc Represents the encoding result of the key vector; V enc Represents the encoding result of the value vector; K enc Represents the encoding result of the key vector; Q dec Represents the decoding result of the query vector; AttentionOutput represents the decoding result of the query vector. jk The matrix composed of This realizes the loop solution of encoding-decoding attention, realizes the update of attention matrix, and then realizes the update of model; The application of encoding-decoding attention loop solving technology in the field of biological intelligent manufacturing can improve computing efficiency, optimize computing resources, enhance system flexibility and reduce operational complexity. (a.4.5) Prediction layer output: The output of the decoding layer passes through a linear layer and outputs the capacity prediction result set Y Pred : Y Pred =(W Pred ×F DecodingOutput +b Pred ) (13); Among them, W Pred represents the weight matrix of the prediction layer; b Pred Represents the bias matrix of the prediction layer.
Citation Information
Cited By
Yield prediction method and device for produced product, storage medium and program product
CN121212427A
