Biological fermentation measurement method and system fusing physical mechanism and multi-modal data

Abstract

The application provides a biological fermentation measurement method and system fusing physical mechanism and multi-modal data, and the method comprises the following steps: collecting and processing multi-modal operation data of a biological fermentation process; based on a microbial metabolic kinetics core equation and a material / energy balance principle, a set of physical mechanism constraint equations is constructed, and a regularization form is embedded into a loss function of a measurement model based on a neural differential equation; the measurement model based on the neural differential equation is trained and processed by using the processed data and the new loss function; multi-modal operation data of a to-be-measured biological fermentation process is input into the trained model to determine predicted key fermentation parameters. Through the application, high-precision measurement of a biological fermentation process in conformity with biological metabolic rules, strong interference and long-term stability is realized.

Description

Technical Field

[0001] This application relates to the field of bio-fermentation engineering technology, specifically to a bio-fermentation measurement method and system that integrates physical mechanisms and multimodal data. Background Technology

[0002] Bio-fermentation is a core process in biomanufacturing fields such as biopharmaceuticals, food processing, antibiotics, and health products. Precise online measurement of key parameters such as biomass concentration, product concentration, and substrate concentration during fermentation is crucial for achieving intelligent process control and improving production stability and efficiency. Currently, the measurement of key parameters in the fermentation process mainly employs three mainstream technical methods: traditional mechanistic modeling, pure data-driven AI modeling, and conventional multimodal data fusion.

[0003] Traditional mechanistic modeling methods rely on fixed fermentation kinetic equations and manually fitted parameters to build models, which cannot adapt to dynamic conditions such as batch fluctuations and raw material differences during the fermentation process. The prediction accuracy is difficult to meet the needs of industrial production, and there is a lack of online adaptive update mechanisms.

[0004] Purely data-driven AI measurement models lack physical mechanism constraints, making them black-box predictions. This results in weak generalization ability and is prone to producing abnormal results that violate biological metabolic laws. For example, in patent "CN116469478A; A Deep Learning-Based Online Soft Measurement Method and Device for Penicillin Fermentation Process," a one-dimensional convolutional neural network is used to extract high-dimensional spatial features from sensor data, combined with a long short-term memory network to mine the long-term dependence characteristics of the fermentation process. Finally, a fully connected network is used to output predicted values ​​of key parameters such as cell concentration, aiming to solve the problem of traditional methods' difficulty in accurate online measurement. However, this is a "black-box" data-driven model, lacking constraints from the physical mechanism of fermentation, which leads to prediction results that easily violate microbial metabolic laws under extreme operating conditions, resulting in a lack of industrial credibility.

[0005] Conventional multimodal fusion technology simply stitches together multi-source sensing and spectral data, without achieving deep feature fusion and temporal correlation. The measurement results are easily affected by on-site noise and cannot perform adaptive optimization of the model according to the actual production status.

[0006] In summary, existing fermentation measurement methods cannot simultaneously address physical mechanism constraints, in-depth utilization of multimodal data, and dynamic adaptive adjustment, making it difficult to meet the high-precision and robust online detection requirements of industrial-grade bio-fermentation. There is an urgent need to construct a new fermentation measurement method that integrates physical mechanisms and multimodal deep learning to overcome the bottlenecks of existing technologies in terms of generalization ability, prediction accuracy, and long-term reliability. Summary of the Invention In view of the deficiencies in the prior art, the purpose of this application is to provide a bio-fermentation measurement method and system that integrates physical mechanisms and multimodal data.

[0007] The first aspect of this application provides a method for measuring bio-fermentation that integrates physical mechanisms and multimodal data, comprising: Collect multimodal operational data of the bio-fermentation process; The multimodal operating data is standardized and time-aligned to determine the processed data; Based on the core equations of microbial metabolic kinetics and the principle of material / energy balance, a set of physical mechanism constraint equations is constructed. The set of physical mechanism constraint equations is embedded in a regularized form into the loss function of the measurement model based on neural differential equations to determine a new loss function; The processed data and the new loss function are used to train the measurement model based on neural differential equations to determine the trained model. The multimodal operational data of the biological fermentation process to be tested are input into the trained model to determine the predicted key fermentation parameters, including cell concentration, substrate concentration, and product concentration.

[0008] Optionally, the multimodal operating data of the bio-fermentation process includes real-time data from online sensors, spectral analysis data, process operation parameter data, and historical fermentation batch data; The standardization and time-series alignment of the multimodal operational data, to determine the processed data, includes: The multimodal operating data is subjected to outlier removal, noise filtering, missing value interpolation, and zero-mean normalization in sequence to determine the standardized multimodal data. A preset dynamic time warping algorithm is used to perform time-series alignment processing on the standardized multimodal data to determine the processed data.

[0009] Optionally, the construction of a set of physical mechanism constraint equations based on the core equations of microbial metabolic kinetics and the principle of material / energy balance includes: Based on the metabolic laws of microbial fermentation, a cell growth kinetic model was established based on the Monod equation. The cell growth kinetic model is used to characterize the quantitative relationship between substrate concentration and cell specific growth rate. A product synthesis kinetic model was constructed based on the Luedeking-Piret equation. This model was used to distinguish the synthesis patterns of growth-related and non-growth-related products. Based on the material / energy balance principle, a coupling constraint equation is established between substrate consumption, cell growth, and product synthesis. The bacterial cell growth kinetics model, the product synthesis kinetics model, and the coupled constraint equations form the physical mechanism constraint equation set.

[0010] Optionally, the step of embedding the set of physical mechanism constraint equations into the loss function of the measurement model based on neural differential equations in a regularized form to determine a new loss function includes: L_total = L_data + λ × L_physics Where L_total represents the new loss function, L_data represents the data fitting error, λ represents the weight coefficient, and L_physics represents the physical mechanism constraint regularization term.

[0011] Optionally, the step of training the measurement model based on neural differential equations using the processed data and the new loss function to determine the trained model includes: The processed data is input into the measurement model based on neural differential equations to determine unconstrained predicted key fermentation parameters, including cell concentration, substrate concentration, and product concentration. Based on the unconstrained predicted key fermentation parameters and the actual key fermentation parameters, the data fitting error is determined; Based on the data fitting error and the new loss function, the measurement model based on neural differential equations is optimized to determine the trained model.

[0012] Optionally, the method further includes: Based on the prediction error between the predicted key fermentation parameters and the offline laboratory test values, fermentation batch information, and operating condition change information, the trained model is incrementally optimized to determine the optimized model.

[0013] Optionally, the step of incrementally optimizing the trained model based on the prediction error between the predicted key fermentation parameters and the offline laboratory test values, fermentation batch information, and operating condition change information, to determine the optimized model, includes: When the prediction error of the predicted key fermentation parameters is greater than a preset prediction error threshold, the number of batch iterations in the fermentation batch information is equal to a preset batch iteration threshold, or any one or more of the conditions in the operating condition change information exist, an incremental learning strategy is used to incrementally optimize the trained model, and the accuracy of the incrementally optimized model is verified to determine the optimized model. The operating condition change information includes changes in fermentation process, fermentation raw materials, and fermentation strains. The incremental learning strategy includes: loading the trained model, fixing the core network bottom-level parameters of the trained model, and unlocking the output layer and associated fully connected layer weights of the trained model. Collect newly added data that triggers incremental optimization and perform standardization and time-series alignment processing to determine the processed new data; Using the processed new data as training samples, the new loss function and Adam adaptive optimization algorithm are used to iteratively fine-tune the trained model, unlocking layer weights until the loss function converges or reaches the preset maximum number of iterations. The accuracy verification process includes: collecting a verification dataset that did not participate in this incremental optimization and matches the current fermentation conditions; The validation dataset is standardized and time-series aligned to determine the processed validation dataset. The processed validation dataset is input into the incrementally optimized model to determine the predicted key fermentation parameters of the validation dataset. Based on the relative error between the predicted key fermentation parameters of the validation dataset and the actual offline experimental key fermentation parameters of the validation dataset, and a preset accuracy threshold, if the relative error does not exceed the preset accuracy threshold, the accuracy is deemed qualified, and the incrementally optimized model is determined as the optimized model; otherwise, incremental optimization is repeated until the accuracy is qualified.

[0014] A second aspect of this application provides a bio-fermentation measurement system that integrates physical mechanisms and multimodal data, comprising: The multimodal data acquisition module is used to collect multimodal operational data of the bio-fermentation process; The data preprocessing module is used to perform standardization and time-series alignment processing on the multimodal operating data to determine the processed data; The physical mechanism constraint construction module is used to construct a set of physical mechanism constraint equations based on the core equations of microbial metabolic kinetics and the principle of material / energy balance. The physical mechanism constraint fusion module is used to embed the set of physical mechanism constraint equations into the loss function of the measurement model based on neural differential equations in a regularized form to determine a new loss function; The model training module is used to train the measurement model based on neural differential equations using the processed data and the new loss function, and to determine the trained model. The fermentation key parameter prediction module is used to input the multimodal operation data of the biological fermentation process to be tested into the trained model to determine the predicted fermentation key parameters, including cell concentration, substrate concentration and product concentration.

[0015] A third aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of any of the methods provided in the first aspect of this application.

[0016] A fourth aspect of this application provides an electronic device comprising: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of any of the methods provided in the first aspect of this application.

[0017] This application presents a bio-fermentation measurement method that integrates physical mechanisms and multimodal data. It collects multimodal operational data of the bio-fermentation process and performs standardization and time-series alignment processing on the data. The method deeply integrates the multimodal data to achieve efficient utilization of multi-source information and improve noise immunity. Based on the core equations of microbial metabolic kinetics and the principle of material / energy balance, a set of physical mechanism constraint equations is constructed and embedded in the loss function of the measurement model based on neural differential equations in a regularized form, so that the measurement output conforms to the laws of biological metabolism.

[0018] Other technical effects resulting from the additional features will be further illustrated in the corresponding embodiments. Attached Figure Description

[0019] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic flowchart illustrating a bio-fermentation measurement method that integrates physical mechanisms and multimodal data according to an exemplary embodiment.

[0020] Figure 2 This is a schematic diagram illustrating the principle of training a measurement model based on neural differential equations according to an exemplary embodiment.

[0021] Figure 3 This is a schematic diagram of a network structure for a measurement model based on neural differential equations, according to an exemplary embodiment.

[0022] Figure 4 This is a schematic diagram illustrating an incremental optimization process for a measurement model based on neural differential equations, according to an exemplary embodiment.

[0023] Figure 5 This is a schematic diagram illustrating the structure of a bio-fermentation measurement system that integrates physical mechanisms and multimodal data according to an exemplary embodiment. Detailed Implementation

[0024] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0025] The terms "comprising" and "having," and any variations thereof, in the embodiments of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or devices.

[0026] In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0027] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0028] Existing methods for measuring key parameters in fermentation processes include traditional mechanistic modeling methods, purely data-driven AI measurement models, and conventional multimodal fusion techniques. Traditional mechanistic modeling methods cannot adapt to batch fluctuations during fermentation, their prediction accuracy cannot meet industrial needs, and they lack online adaptive update mechanisms. Purely data-driven AI measurement models lack physical mechanism constraints, have weak generalization ability, and are prone to abnormal results that violate biological metabolic laws. Conventional multimodal fusion techniques only perform simple stitching of multi-source sensing and spectral data, and the measurement results are easily affected by on-site noise. All three methods are insufficient to meet the high-precision and robust online monitoring requirements of industrial-grade bio-fermentation. Based on the above problems, this application provides a bio-fermentation measurement method that integrates physical mechanisms and multimodal data to solve the aforementioned problems.

[0029] Figure 1 This is a schematic flowchart illustrating a bio-fermentation measurement method that integrates physical mechanisms and multimodal data according to an exemplary embodiment.

[0030] Reference Figure 1 As shown in one embodiment of this application, a method for measuring bio-fermentation that integrates physical mechanisms and multimodal data is provided, including steps S11 to S16.

[0031] S11 collects multimodal operational data of the bio-fermentation process.

[0032] S12 performs standardization and time-series alignment on the multimodal operating data to determine the processed data.

[0033] S13, based on the core equation of microbial metabolic kinetics and the principle of material / energy balance, constructs a set of physical mechanism constraint equations.

[0034] S14. Embed the set of physical mechanism constraint equations into the loss function of the measurement model based on neural differential equations in a regularized form to determine a new loss function.

[0035] S15, using the processed data and the new loss function, trains the measurement model based on the neural differential equation to determine the trained model.

[0036] S16: Input the multimodal operating data of the bio-fermentation process to be tested into the trained model to determine the predicted key fermentation parameters.

[0037] Specifically, key fermentation parameters include cell concentration, substrate concentration, and product concentration. The embodiments described above in this application collect multimodal operational data of the bio-fermentation process and perform standardization and time-series alignment processing on it. The multimodal data is deeply fused to achieve efficient utilization of multi-source information and improve noise immunity. Based on the core equations of microbial metabolic kinetics and the principle of material / energy balance, a set of physical mechanism constraint equations is constructed and embedded in the loss function of the measurement model based on neural differential equations in a regularized form to ensure that the measurement output conforms to the laws of biological metabolism.

[0038] To make full use of multimodal data, in some specific embodiments of this application, the multimodal operation data collected in the bio-fermentation process includes, but is not limited to, real-time data from online sensors, spectral analysis data, process operation parameter data, and historical fermentation batch data throughout the entire bio-fermentation process.

[0039] For example, the real-time data from the online sensors are time-series data from the online sensors, such as fermentation process temperature, pH value, dissolved oxygen concentration, stirring speed, aeration rate, and feeding rate.

[0040] The spectral analysis data can be near-infrared / Raman spectral data, and the process operation parameters can be feeding strategies and culture medium formulations.

[0041] The embodiments described above in this application, by comprehensively collecting online sensor time-series data, spectral analysis data, process operation parameters, and historical fermentation batch data, achieve deep complementarity and full-process coverage of multimodal information. This enables real-time and accurate characterization of the fermentation physicochemical state and metabolic features, effectively overcoming the shortcomings of single data source information being one-sided and susceptible to interference. It significantly improves the generalization ability and decision reliability of the subsequently trained model, providing comprehensive and high-quality data support for the fine monitoring, dynamic optimization, and stable high yield of the fermentation process.

[0042] To improve the utilization rate of multimodal data, in some specific embodiments of this application, for S12, the multimodal operation data is standardized and time-series aligned to determine the processed data, which can be done using S121 to S122. S121, perform outlier removal, noise filtering, missing value interpolation, and zero-mean normalization on the multimodal operating data in sequence to determine the standardized multimodal data.

[0043] Specifically, noise filtering can be performed using moving average or wavelet filtering methods.

[0044] Among them, outlier removal can use the 3σ criterion to identify and remove abnormal data caused by sensor drift and equipment failure; noise filtering can use wavelet filtering algorithm to remove data noise caused by electromagnetic interference and equipment vibration in industrial sites; zero mean normalization can map multi-dimensional data to the [0, 1] interval, thereby eliminating the influence of dimensional differences on model training.

[0045] S122, a preset dynamic time warping algorithm is used to perform time-series alignment processing on the standardized multimodal data to determine the processed data.

[0046] Specifically, batch data from different fermentation sites and different starting points are aligned to a unified timeline to achieve time-series alignment, thereby ensuring the time-series consistency of the model input.

[0047] The embodiments described above in this application standardize multimodal operating data to eliminate on-site noise and data defects. A preset dynamic time warping algorithm is used to perform time-series alignment processing on the standardized multimodal data to unify the time dimension, thereby adapting to batch data with different fermentation cycles and different start nodes, and realizing time-series alignment of fermentation batches with different durations.

[0048] Figure 2 This is a schematic diagram illustrating the principle of training a measurement model based on neural differential equations according to an exemplary embodiment.

[0049] Reference Figure 2As shown, in order to achieve constraints that conform to the laws of biological metabolism, in some specific embodiments of this application, for S13, a set of physical mechanism constraint equations is constructed based on the core equation of microbial metabolic kinetics and the principle of material / energy balance, which can be S131 to S134.

[0050] S131. Based on the metabolic laws of microbial fermentation, a cell growth kinetic model was established using the Monod equation.

[0051] Specifically, the cell growth kinetic model is used to characterize the quantitative relationship between substrate concentration and cell specific growth rate.

[0052] Specifically, the metabolic laws of microbial fermentation indicate that during fermentation, microorganisms, by taking up limiting substrates (such as carbon sources like glucose and sucrose, or nitrogen sources and inorganic salts) in the fermentation system, complete metabolic activities under the catalysis of enzymes. On the one hand, they achieve the growth and reproduction of their own cells, and on the other hand, they synthesize target fermentation products. This is accompanied by energy conversion and the generation of metabolic waste. The entire process strictly follows the laws of conservation of mass and energy, and there is a clear quantitative correlation between cell growth rate, substrate consumption rate, and product synthesis rate. Its core laws can be summarized as follows: substrate consumption provides the material and energy basis for cell growth and product synthesis; cell growth rate shows a specific pattern with changes in substrate concentration; and product synthesis is directly related to the cell growth state (divided into growth-related and non-growth-related types).

[0053] Specifically, the bacterial cell growth kinetics equation (based on the Monod equation): in, This represents the specific growth rate of the bacterial cells (h⁻¹). KS represents the maximum specific growth rate of the bacteria (h⁻¹), S represents the limiting substrate concentration (g / L), and KS represents the half-saturation constant (g / L).

[0054] The half-saturation constant KS is used to characterize the affinity of bacteria for substrates.

[0055] For example, a cell growth kinetic model is established based on the Monod equation as follows: Taking Penicillium chrysogenum, a commonly used cephalosporin antibiotic fermentation agent in industry, as an example, a cell growth kinetic model was established based on the Monod equation, taking into account its metabolic characteristics.

[0056] First, we need to clarify that the core form of the Monod equation is: The specific value range for each parameter is determined by considering the fermentation characteristics of Penicillium chrysogenum: The maximum specific growth rate of the bacteria was taken as 0.15~0.25 h. - ¹, Adjust the fermentation time according to the initial activity of the microbial strain; when the activity is high, take 0.20~0.25 h. - ¹, when the activity is low, use 0.15~0.20 h. - ¹; S is the concentration (g / L) of the limiting substrate (glucose in this example) in the fermentation system, ranging from 0.5 to 30 g / L, covering the substrate concentration changes throughout the fermentation cycle; KS (half-saturation constant) is 0.8 to 1.2 g / L, characterizing the affinity of Penicillium chrysogenum for glucose. This value was determined through fitting multiple batches of preliminary experiments. Specifically, five groups of different initial glucose concentrations (2 g / L, 5 g / L, 10 g / L, 20 g / L, 30 g / L) were set up and cultured under the same fermentation conditions (temperature 28℃, pH 6.5, dissolved oxygen concentration 20%). The cell concentration at different time points was measured, and the cell specific growth rate was calculated. The optimal fitting value of KS was obtained as 1.0 g / L through nonlinear fitting.

[0057] The final bacterial cell growth kinetic model for this embodiment is determined as follows: This cell growth kinetic model can accurately characterize the quantitative relationship between glucose concentration and the specific growth rate of Penicillium chrysogenum. When the glucose concentration S > 10 g / L, the specific growth rate is close to the maximum specific growth rate of 0.22 h. - ¹, When S < 1.0 g / L, the cell growth rate decreased significantly with decreasing glucose concentration, which perfectly matches the actual growth pattern of Penicillium chrysogenum.

[0058] S132, a product synthesis kinetic model was constructed based on the Luedeking-Piret equation.

[0059] Specifically, the product synthesis kinetics model is used to distinguish the synthesis patterns of growth-related and non-growth-related products.

[0060] Specifically, the product synthesis kinetic equations (based on the Luedeking-Piret equations): in, The product synthesis rate is expressed as g / (L·h) X represents the cell growth rate (g / (L·h)) and X represents the cell concentration (g / L). This represents the growth correlation coefficient (g / g). This represents the non-growth correlation coefficient (g / (g·h)).

[0061] For example, the product synthesis kinetic model is constructed based on the Luedeking-Piret equation as follows: Taking the fermentation of Penicillium chrysogenum to produce cephalosporin antibiotics (growth-related products) as an example, the core form of the Luedeking-Piret equation is: Among them, the parameters were determined experimentally in conjunction with the cephalosporin fermentation characteristics: growth correlation coefficient Characterizing the correlation between product synthesis and cell growth, non-growth correlation coefficient. Characterizes the ability of bacteria to synthesize products during the stationary phase.

[0062] Through multiple batches of fermentation experiments, the cell concentration X (g / L) and product concentration P (g / L) at different time points were measured, and the cell growth rate was calculated. (g / (L·h)) and product synthesis rate (g / (L·h)), determined using a linear fitting method. and The possible values ​​of: =0.08~0.12 g / g, =0.002~0.005 g / (g·h).

[0063] In this embodiment, the optimal fitting value is =0.10 g / g =0.003 g / (g·h), and the final product synthesis kinetic model is as follows: This model can clearly distinguish the synthesis patterns of growth-related products: when the bacterial cells are in the logarithmic growth phase, The product synthesis is relatively large and mainly depends on This refers to the synthesis of certain substances, which occur synchronously with bacterial growth; once the bacteria enter the stationary phase, Approaching 0, the synthesis of the product mainly depends on This refers to the slow synthesis of bacteria during their metabolic processes, which aligns with the actual synthesis patterns of cephalosporin antibiotics.

[0064] For non-growth-related products (such as penicillin G), the results are obtained through experimental fitting. ≈0、 =0.01~0.015g / (g·h), the model simplifies to The characterization product is mainly synthesized during the stationary phase of the cell and is not directly related to cell growth, thus enabling an effective distinction between the synthesis patterns of the two products.

[0065] S133, based on the principle of material / energy balance, establishes a coupling constraint equation between substrate consumption, cell growth and product synthesis.

[0066] Specifically, the coupled constraint equations (based on the material / energy balance principle): in, The substrate consumption rate is expressed as g / (L·h) This represents the yield coefficient (g / g) of the bacterial cells to the substrate. denoted by , which represents the product yield coefficient to substrate (g / g), and m represents the cell maintenance coefficient (g / (g·h)).

[0067] The coupling constraint equations realize the quantitative correlation between substrate consumption, cell growth, and product synthesis, ensuring compliance with the law of conservation of mass.

[0068] For example, based on the principle of material / energy balance, the coupling constraint equations between substrate consumption, cell growth, and product synthesis are established as follows: The coupling constraints between substrate consumption, cell growth, and product synthesis are as follows: The specific values ​​of each parameter were determined based on the fermentation characteristics of Penicillium chrysogenum in this embodiment: the yield coefficient of the cell to the substrate. =0.45~0.55 g / g, representing the mass of bacterial cells produced per 1g of glucose consumed; in this example, 0.50 g / g is used; the product yield coefficient to substrate. =0.04~0.06 g / g, characterizing the mass of cephalosporin antibiotics produced per 1g of glucose consumed; in this example, it is taken as 0.05 g / g; cell maintenance coefficient m=0.02~0.03 g / (g·h), characterizing the mass of substrate consumed per hour to maintain basal metabolism per gram of cells; in this example, it is taken as 0.025 g / (g·h).

[0069] The quantitative correlation between substrate consumption, cell growth, and product synthesis is explained as follows: The coupling constraint equation is essentially the material conservation equation for the substrate (glucose) during fermentation, which clarifies the three main destinations of substrate consumption and realizes the quantitative correlation between substrate consumption, cell growth, and product synthesis.

[0070] Part One: This represents the substrate consumption rate used for bacterial growth, that is, the amount of substrate required to produce 1g of bacterial cells during the bacterial growth process. (In this example, 2g) of glucose is used, and this portion of the substrate is converted into the cell components of the bacteria (such as proteins, nucleic acids, etc.); Part Two: This indicates the substrate consumption rate used in product synthesis, i.e., the amount of substrate required to produce 1g of cephalosporin antibiotic. (In this example, 20g) of glucose is used, and this portion of the substrate is converted into the target product through a metabolic pathway; Part Three: , which represents the substrate consumption rate used by bacteria to maintain metabolism. That is, the substrate consumed by bacteria to maintain their own life activities (such as respiration, material transport, etc.) in addition to growth and product synthesis. This part of the substrate is eventually converted into energy and metabolic waste (such as CO2, water, etc.).

[0071] The sum of the above three factors equals the total consumption rate of the substrate. (The negative sign indicates that the substrate concentration decreases over time), ensuring that the consumption and transformation of substrates strictly follow the law of conservation of mass throughout the fermentation process. At the same time, cell growth, product synthesis and substrate consumption are closely linked, so that the changes of the three are mutually restrictive and mutually influential. For example, an increase in cell growth rate will increase the substrate consumption rate, which in turn will affect the substrate supply for product synthesis. Conversely, insufficient substrate concentration will limit cell growth and product synthesis. This provides a core basis for embedding physical constraints into the subsequent model and avoiding non-physical predictions.

[0072] S134, the bacterial cell growth kinetics model, the product synthesis kinetics model, and the coupled constraint equations form a set of physical mechanism constraint equations.

[0073] The embodiments described above in this application, by integrating Monod growth kinetics, Luedeking-Piret product synthesis kinetics, and the principle of material-energy conservation, construct a set of physical mechanism constraint equations that ensure the model output strictly conforms to the physical laws of biological metabolism. This effectively overcomes the black-box defects of pure data models, prevents abnormal measurement results that do not conform to the laws of biological metabolism, and endows the model with strong interpretability and clear physical meaning of parameters. While strictly ensuring test quality and energy closure, it significantly reduces dependence on experimental data, improves the prediction accuracy and robustness of the model under different fermentation modes and operating conditions, and is widely adaptable to various metabolite synthesis pathways. It provides a highly reliable and computable mechanistic foundation for intelligent optimization control, process scale-up, and digital twins of fermentation processes.

[0074] Reference Figure 2 As shown, to embed a set of physical mechanism constraint equations into a measurement model based on neural differential equations, in some specific embodiments of this application, for S14, the set of physical mechanism constraint equations is embedded in the loss function of the measurement model based on neural differential equations in a regularized form. To determine the new loss function, the following methods can be used: L_total = L_data + λ × L_physics Where L_total represents the new loss function, L_data represents the data fitting error, λ represents the weight coefficient, and L_physics represents the physical mechanism constraint regularization term.

[0075] The value of λ can range from 0.1 to 0.8.

[0076] In the embodiments described above, the set of physical mechanism constraint equations is embedded in the loss function of the measurement model based on neural differential equations in a regularized form. This effectively achieves a deep integration of data-driven flexibility and mechanism rigor. By introducing a mechanism penalty term, the model's dependence on massive experimental data is significantly reduced, avoiding overfitting and non-physical predictions that are prone to occur in pure neural networks. At the same time, the optimization surface is smoothed, improving the training convergence stability and cross-condition generalization ability. Ultimately, the measurement model can accurately fit complex nonlinear features during dynamic fermentation while strictly following metabolic kinetics and conservation laws. This provides a hybrid modeling paradigm with both accuracy and robustness for highly reliable and interpretable online state estimation and process optimization.

[0077] Reference Figure 2 As shown, in order to embed a set of physical mechanism constraint equations into a measurement model based on neural differential equations, some specific embodiments of this application construct a time-series prediction network with neural differential equations as the core as a measurement model based on neural differential equations, and adapt to the long-cycle, strong temporal, and slow-dynamic characteristics of the bio-fermentation process. The measurement model based on neural differential equations includes an input layer, a hidden layer, and an output layer. The input layer is used to input multimodal operating data, the hidden layer is a neural differential equation used to predict key fermentation parameters and capture long-cycle metabolic cumulative effects, and the output layer is used to output the predicted key fermentation parameters without the addition of a set of physical mechanism constraint equations.

[0078] Reference Figure 2 As shown in some specific embodiments of this application, for S15, the measurement model based on neural differential equations is trained using processed data and a new loss function to determine the trained model, which can be done using S151 to S153.

[0079] S151, the processed data is input into a measurement model based on neural differential equations to determine unconstrained predicted key fermentation parameters.

[0080] Specifically, the processed data is a training set that has been standardized and time-aligned, and an adaptive optimization algorithm is used to train a measurement model based on neural differential equations.

[0081] Key fermentation parameters include cell concentration, substrate concentration, and product concentration.

[0082] S152, determine the data fitting error based on the unconstrained predicted key fermentation parameters and the actual key fermentation parameters.

[0083] S153, based on the data fitting error and the new loss function, optimize the measurement model based on neural differential equations to determine the trained model.

[0084] Specifically, a new loss function is used to optimize the measurement model based on neural differential equations, and the data fitting error and physical mechanism constraint regularization term are optimized simultaneously to obtain a trained model, namely a soft measurement model that integrates physical constraints, which can generate predicted values ​​of key fermentation parameters that conform to metabolic laws.

[0085] The new loss function is: L_total = L_data + λ × L_physics Where L_total represents the new loss function, L_data represents the data fitting error, λ represents the weight coefficient, and L_physics represents the physical mechanism constraint regularization term.

[0086] The value of λ ranges from 0.1 to 0.8.

[0087] In some specific embodiments of this application, validation data from multiple batches of validation sets are used to validate the trained model, verifying its performance and ensuring that the prediction error meets the requirements of industrial applications.

[0088] Specifically, the validation set includes multimodal operational data of bio-fermentation processes that have been standardized and time-aligned.

[0089] Prediction error represents the error between the predicted key fermentation parameters and the actual key fermentation parameters under the constraints of the physical mechanism constraint equation set. When the prediction error is not greater than the preset error threshold, it indicates that the performance of the trained model meets the requirements; otherwise, training continues until the prediction error of the verification process is not greater than the preset error threshold.

[0090] The embodiments described above in this application effectively improve the prediction accuracy and generalization ability of soft measurement of key fermentation parameters by combining a measurement model based on neural differential equations with a new loss function for model training. The new loss function integrates data fitting error with prior physical mechanism constraints during the optimization process, achieving a deep integration of data-driven and physical mechanism constraints. This avoids the overfitting and physical inconsistency problems of pure data-driven models, while improving training stability and convergence efficiency. Ultimately, it enables accurate and reliable monitoring of the fermentation process, providing high-quality data support for process optimization and intelligent control.

[0091] Figure 3This is a schematic diagram of a network structure for a measurement model based on neural differential equations, according to an exemplary embodiment.

[0092] Reference Figure 3 As shown in some specific embodiments of this application, for S16, inputting the multimodal operational data of the bio-fermentation process to be tested into the trained model to determine the predicted key fermentation parameters can be done by: S161, standardize and time-align the multimodal operation data of the bio-fermentation process to be tested, and determine the real-time multimodal operation data of the bio-fermentation process to be tested after processing.

[0093] S162, input the processed real-time multimodal operation data of the bio-fermentation process to be tested into the trained model to determine the predicted key fermentation parameters constrained by physical mechanisms.

[0094] Specifically, key fermentation parameters may include, but are not limited to, cell concentration, substrate concentration, and product concentration.

[0095] Specifically, the trained model consists of an input layer, a hidden layer, a physical constraint layer, and an output layer connected in sequence.

[0096] Based on the above steps S161 to S162, for example, 100-dimensional processed multimodal operating data, sourced from online sensor real-time data, spectral analysis data, process operation parameter data, and historical fermentation batch data, are input into the input layer of the trained model.

[0097] The input layer maps features to the hidden layer and performs temporal modeling based on neural differential equations to model temporal features.

[0098] Specifically, the hidden layer includes sub-layer 1, sub-layer 2 and sub-layer 3. Sub-layer 1 is a feature decoupling layer, sub-layer 2 is a differential equation calculation layer, and sub-layer 3 is a feature fusion layer. Sub-layer 1, sub-layer 2 and sub-layer 3 are connected in sequence, and sub-layer 3 outputs temporal features.

[0099] Temporal features are input into the physical constraint layer, and after being verified by metabolic patterns, physical consistency correction is performed to correct predicted values ​​that violate metabolic patterns.

[0100] The predicted values ​​of key fermentation parameters, after physical consistency correction, are output through the output layer, which are the predicted key fermentation parameters constrained by physical mechanisms.

[0101] The embodiments described above in this application effectively overcome the problems of scale differences and asynchronous sampling of multi-source data by standardizing and aligning real-time multimodal data, ensuring the temporal consistency and high quality of model input. Combined with the predicted key fermentation parameters constrained by physical mechanisms output by the trained model, the online soft measurement results are ensured to have both high precision driven by data and physical rationality of the bio-fermentation process. This significantly improves the reliability of prediction, real-time response capability, and anti-interference ability, providing a robust and reliable decision-making basis for dynamic monitoring and closed-loop intelligent control of the fermentation process.

[0102] Figure 4 This is a schematic diagram illustrating an incremental optimization process for a measurement model based on neural differential equations, according to an exemplary embodiment.

[0103] Reference Figure 4 As shown, to achieve online adaptive updating of the measurement model based on neural differential equations, in some specific embodiments of this application, for S17, incremental optimization is performed on the trained model based on the prediction error of the predicted key fermentation parameters, fermentation batch information, and operating condition change information to determine the optimized model. This can be achieved by: When any one or more of the following conditions are met: the prediction error of the predicted key fermentation parameters is greater than the preset prediction error threshold, the number of batch iterations in the fermentation batch information is equal to the preset batch iteration threshold, or there is a change in operating conditions, an incremental learning strategy is used to incrementally optimize the trained model, and the accuracy of the incrementally optimized model is verified to determine the optimized model.

[0104] Specifically, the change information on operating conditions may include, but is not limited to, any one or more combinations of changes in fermentation process, fermentation raw materials, and fermentation strains.

[0105] Specifically, the prediction error between the predicted key fermentation parameters and the laboratory offline test values ​​represents the relative error between the predicted key fermentation parameters and the laboratory offline test values, with the preset error threshold set at 5%.

[0106] Specifically, The smaller the prediction error, the higher the prediction accuracy.

[0107] The preset batch iteration threshold can be 3 fermentation batches.

[0108] The incremental learning strategy includes: loading the trained model, fixing the core network parameters of the trained model, and unlocking the output layer and associated fully connected layer weights of the trained model. Collect newly added data that triggers incremental optimization and perform standardization and time-series alignment processing to determine the processed new data; Using the newly processed data as training samples, a new loss function and the Adam adaptive optimization algorithm are used to iteratively fine-tune the trained model, unlocking layer weights until the loss function converges or the preset number of iterations is reached.

[0109] Accuracy verification processing includes: collecting verification datasets that did not participate in this incremental optimization and that match the current fermentation conditions; The validation dataset is standardized and time-aligned to determine the processed validation dataset. The processed validation dataset is input into the incrementally optimized model to determine the key fermentation parameters for the validation dataset's predictions. Based on the relative error between the predicted key fermentation parameters of the validation dataset and the actual offline experimental key fermentation parameters of the validation dataset, and a preset accuracy threshold, if the relative error does not exceed the preset accuracy threshold, the accuracy is deemed qualified, and the model that has undergone incremental optimization is identified as the optimized model; otherwise, incremental optimization is repeated until the accuracy is qualified.

[0110] An incremental learning strategy is used to incrementally optimize the trained model, fine-tuning the model weights without requiring full retraining. The model update can be completed within 2 minutes, and the model accuracy is automatically verified after the update to ensure that the optimized model is restored to the preset accuracy range before being put back into operation.

[0111] The embodiments described above in this application incrementally optimize the trained model based on the prediction error of the predicted key fermentation parameters, fermentation batch information, and operating condition change information, thereby achieving online adaptive updating of the model, realizing long-term stable measurement of bio-fermentation engineering, and improving the accuracy and robustness of the measurement.

[0112] Taking the scenario of producing cephalosporin antibiotics by fermentation of Penicillium chrysogenum as an example, this paper combines two typical triggering conditions to illustrate the specific implementation process of the incremental learning strategy, ensuring that the optimization process is feasible and reproducible.

[0113] Example 1: Incremental optimization triggered by operating condition change information.

[0114] When the fermentation feedstock changes, such as replacing the limiting substrate glucose with food-grade glucose, the impurity content decreases and the purity increases, leading to a slight increase in the metabolic rate of Penicillium chrysogenum, triggering incremental optimization processes. The first step is to load the pre-trained measurement model based on neural differential equations that incorporates physical mechanism constraints, fix the parameters of the bottom layer of the core network (feature decoupling layer, differential equation calculation layer), and only unlock the weights of the model output layer and associated fully connected layers to avoid damage to the feature extraction capability of the bottom mechanism. The second step is to collect multimodal operation data of the first two fermentation batches after the change of raw materials. Following the data processing flow of step S12 above, outlier removal (using the 3σ criterion), noise filtering (wavelet filtering), missing value interpolation, and zero mean normalization are performed in sequence. Then, the time sequence is aligned using the preset dynamic time warping algorithm and used as new training samples. The third step is to adopt a new loss function that incorporates physical mechanism constraints: L_total=L_data+λ×L_physics (λ is 0.3, consistent with the original model training). The Adam adaptive optimization algorithm is used, with a learning rate of 0.001 and 50 iterations. The weights of the unlocked output layer and associated fully connected layer are fine-tuned through a small number of iterations. The fourth step involves selecting the multimodal data from the third fermentation batch after the raw material change as an independent validation set. After performing the same data processing, the data is input into the trained model to calculate the relative errors between the predicted key fermentation parameters (cell concentration X, cephalosporin potency P, and glucose concentration S) and the offline laboratory test values. The validation results show that the relative errors of the three parameters are all ≤3.5%, which is lower than the preset 5% error threshold. The accuracy is deemed qualified, and the model is identified as optimized. The entire optimization process takes 1.2 minutes and does not require full retraining, quickly adapting to the new operating conditions after the raw material change.

[0115] Example 2: Incremental optimization based on prediction error triggering.

[0116] When the model ran for two consecutive fermentation batches, the predicted relative errors for cephalosporin potency were 5.8% and 6.2%, respectively, both exceeding the preset error threshold of 5%, triggering the incremental optimization process. The first step is to load the currently running model, fix the underlying parameters of the core network, and unlock the weights of the output layer and associated fully connected layers. The second step is to collect multimodal data from nearly three fermentation batches (including two batches with excessive errors), and use the preprocessed data as new training samples, focusing on supplementing the time node data with excessive errors to enhance the model's adaptability to abnormal working conditions. The third step involves using the new loss function and optimization algorithm, adjusting the learning rate to 0.0008, and iterating 40 times to fine-tune the weights of the unlocking layer, focusing on correcting the prediction bias of cephalosporin potency. The fourth step involved selecting one independent fermentation batch that did not participate in this incremental training as a validation set to verify the accuracy of the trained model. The validation results showed that the relative error of cephalosporin potency prediction was reduced to 3.2%, and the prediction errors of cell concentration and glucose concentration were 2.8% and 2.5%, respectively, all of which met the accuracy requirements. The incremental optimization was completed in 1.5 minutes, and the model was put back into online operation, effectively solving the problem of prediction accuracy decay caused by batch fluctuations.

[0117] In addition, when the number of batch iterations reaches the preset 3 fermentation batches, the implementation process of triggering incremental optimization is similar to the above embodiment. Only the training samples are added as the complete data of the 3 fermentation batches. After fine-tuning, the optimization can be completed through accuracy verification, ensuring that the model continues to adapt to the small fluctuations in the fermentation process with batch iterations and maintains long-term stable prediction accuracy.

[0118] This application provides a biological fermentation measurement method that integrates physical mechanisms and multimodal data. It employs deep multimodal data fusion and time-series alignment techniques, enabling efficient utilization of multi-source information and improved anti-interference capabilities. This addresses the technical problems of conventional fusion methods, which only perform surface-level stitching, resulting in low data utilization and susceptibility to noise. Furthermore, it utilizes physical mechanism-constrained embedded neural differential equation technology, ensuring that measurement outputs conform to biological metabolic laws. This overcomes the issues of unreliable, kineticly incompatible, and purely data-driven AI black-box models. Finally, it employs online adaptive model updating technology, achieving long-term stable and high-precision predictions. This addresses the problem of traditional models remaining static and experiencing rapid accuracy decay due to batch fluctuations and changes in operating conditions, maintaining high accuracy and robustness even in industrial scenarios with small samples, strong fluctuations, and multiple interferences.

[0119] This application provides a biological fermentation measurement method that integrates physical mechanisms and multimodal data, which is particularly suitable for real-time monitoring and process control of fermentation processes in biomanufacturing scenarios such as antibiotics, amino acids, food fermentation, and biomedicine.

[0120] The preferred features in the above embodiments can be used individually in any embodiment, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the embodiments can be implemented using existing technologies.

[0121] The following examples and comparative examples will be used to further illustrate this application in order to better understand the above-mentioned technical solutions. It should be understood that the following are only some examples and are not intended to limit this application.

[0122] To further verify the feasibility and practicality of the bio-fermentation measurement method that integrates physical mechanisms and multimodal data provided in this application, specific embodiments of three typical industrial scenarios—antibiotic fermentation, food fermentation, and health product fermentation—are described below.

[0123] Example 1: Measurement of antibiotic fermentation process, such as cephalosporins: Using industrial-grade cephalosporin antibiotic fermentation as an application scenario, the method provided in this application is adopted to achieve online measurement of key parameters.

[0124] Sensor data such as fermenter temperature, pH, dissolved oxygen (DO), stirring speed, aeration rate, and feeding rate are collected at a frequency of 1 minute / time, and near-infrared spectral data of fermentation broth are collected at a frequency of 5 minutes / time, while historical fermentation data of 160 batches are retrieved simultaneously. After anomaly removal, filtering, and normalization are completed, the time axis of different batches of data is unified by using a preset dynamic time warping algorithm to adapt to the working conditions of long cephalosporin fermentation cycle and large differences in metabolic timing. Based on the Luedeking-Piret and Monod equations, and combined with the metabolic characteristics of cephalosporin strains, a set of exclusive fermentation physical constraint equations was constructed to fit the special kinetic laws of cephalosporin fermentation cell growth and product synthesis. The set of fermentation physical constraint equations was embedded into a measurement model based on neural differential equations and the model was trained. The equilibrium coefficient λ was optimized to adapt to the characteristics of cephalosporin fermentation data. The trained model outputs real-time parameters related to biomass concentration, cephalosporin potency, substrate residual sugar concentration, and mycelial morphology, with a prediction relative error stable within 2.8%, fully meeting the high-precision detection requirements for pharmaceutical-grade cephalosporin fermentation. The trained model is automatically updated every three complete fermentation batches or when the prediction error exceeds 5%. The prediction accuracy does not significantly decrease after running continuously for more than six months and is not affected by fluctuations in raw material batches or strain attenuation.

[0125] This embodiment enables online, contactless detection of key parameters in cephalosporin fermentation throughout the entire process, completely replacing the traditional manual offline sampling and laboratory testing procedures. It improves detection efficiency by more than 10 times, significantly reduces labor costs and detection lag, and provides core data support for closed-loop intelligent control of cephalosporin fermentation process, improving product yield and drug quality stability.

[0126] Example 2: Measurement of food fermentation (glutamic acid) fermentation process The bio-fermentation measurement method that integrates physical mechanisms and multimodal data provided in this application enables online measurement of cell concentration, glutamate concentration, and reducing sugar concentration.

[0127] Data on temperature, pH, dissolved oxygen, defoamer dosage, Raman spectroscopy, and culture medium formulation were collected. Standardize and time-series aligned multimodal operating data to adapt to the characteristics of small sample size and variable operating conditions in food fermentation. A set of physical mechanism constraint equations was constructed based on the kinetics of sugar-acid conversion. The set of physical mechanism constraint equations is embedded in a regularized form into the loss function of the measurement model based on neural differential equations to determine a new loss function; The measurement model based on neural differential equations was trained using 50 batches of data and a new loss function, and the model showed excellent generalization performance. The trained model outputs key parameters every minute, providing real-time data for automatic feeding, dissolved oxygen optimization, and pH control. When the culture medium formula or bacterial strain is changed, the model adaptation is completed within 2 hours, and the trained model is incrementally optimized to quickly match the new working conditions.

[0128] The model prediction error in this embodiment is ≤3.2%, which completely solves the industry pain points of low measurement accuracy and poor adaptability in small sample scenarios of food fermentation.

[0129] Example 3: Measurement of the fermentation process of health supplements, such as coenzyme Q10: This embodiment takes high-value-added coenzyme Q10 fermentation as the application scenario and uses the biological fermentation measurement method provided in this application, which integrates physical mechanisms and multimodal data, to achieve accurate measurement of viable cell count, coenzyme Q10 content, and substrate consumption.

[0130] Collect multimodal operating data such as pH, temperature, dissolved oxygen, oxidation-reduction potential (ORP), stirring speed, and spectral data; Standardize and time-series align the multimodal operational data to determine the processed data; A set of physical mechanism constraint equations was constructed based on the coenzyme Q10 metabolic pathway to ensure that the prediction results conform to the physiological laws of bacteria. The set of physical mechanism constraint equations is embedded in a regularized form into the loss function of the measurement model based on neural differential equations to determine a new loss function; The neural differential equation-based measurement model is trained using processed data and a new loss function to obtain a high-precision measurement model. The trained model outputs key indicators in real time, supporting intelligent judgment of fermentation endpoint and improving the yield of coenzyme Q10 products by about 10%. The trained model is incrementally optimized based on the prediction errors of key fermentation parameters, batch iteration information, and operating condition change information, and runs continuously and stably to meet the high-value-added and high-precision detection requirements of health product fermentation.

[0131] This embodiment has a prediction error of ≤2.5%, breaking through the accuracy bottleneck of high value-added fermented product measurement and helping the health product fermentation industry to upgrade intelligently.

[0132] Figure 5This is a schematic diagram illustrating the structure of a bio-fermentation measurement system that integrates physical mechanisms and multimodal data according to an exemplary embodiment.

[0133] Reference Figure 5 As shown in another embodiment of this application, a bio-fermentation measurement system 100 integrating physical mechanisms and multimodal data is provided, including: a multimodal data acquisition module 110, a data preprocessing module 120, a physical mechanism constraint construction module 130, a physical mechanism constraint fusion module 140, a model training module 150, a fermentation key parameter prediction module 160, and a model adaptive update module 170.

[0134] The multimodal data acquisition module 110 is used to acquire multimodal operational data of the bio-fermentation process; The data preprocessing module 120 is used to perform standardization and time-series alignment on multimodal operating data to determine the processed data. The physical mechanism constraint construction module 130 is used to construct a set of physical mechanism constraint equations based on the core equations of microbial metabolic kinetics and the principle of material / energy balance. The physical mechanism constraint fusion module 140 is used to embed the set of physical mechanism constraint equations into the loss function of the measurement model based on neural differential equations in a regularized form to determine a new loss function. The model training module 150 is used to train the measurement model based on neural differential equations using processed data and a new loss function, and to determine the trained model. The fermentation key parameter prediction module 160 is used to input the multimodal operation data of the biological fermentation process to be tested into the trained model to determine the predicted fermentation key parameters, including cell concentration, substrate concentration and product concentration.

[0135] The embodiments described above in this application collect multimodal operational data of the bio-fermentation process and perform standardization and time-series alignment processing on it. The multimodal data is deeply fused to achieve efficient utilization of multi-source information and improve noise immunity. Based on the core equations of microbial metabolic kinetics and the principle of material / energy balance, a set of physical mechanism constraint equations is constructed and embedded in the loss function of the measurement model based on neural differential equations in a regularized form to ensure that the measurement output conforms to the laws of biological metabolism.

[0136] Regarding the embodiments of the above system, the specific ways in which each module performs operations have been described in detail in the embodiments of the method, and will not be elaborated here.

[0137] Based on the same technical concept, in some specific embodiments of this application, a terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can be used to execute a bio-fermentation measurement method that integrates physical mechanisms and multimodal data.

[0138] Based on the same technical concept, in some specific embodiments of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, can be used to perform a bio-fermentation measurement method that integrates physical mechanisms and multimodal data.

[0139] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs and functional modules that implement the above methods), computer instructions, etc., and the aforementioned computer programs and computer instructions can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0140] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0141] A processor is used to execute a computer program stored in memory to implement the various steps of the methods involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0142] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0143] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] The foregoing has described some specific embodiments of this application. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A method for measuring bio-fermentation that integrates physical mechanisms and multimodal data, characterized in that, include: Collect multimodal operational data of the bio-fermentation process; The multimodal operating data is standardized and time-aligned to determine the processed data; Based on the core equations of microbial metabolic kinetics and the principle of material / energy balance, a set of physical mechanism constraint equations is constructed. The set of physical mechanism constraint equations is embedded in a regularized form into the loss function of the measurement model based on neural differential equations to determine a new loss function; The processed data and the new loss function are used to train the measurement model based on neural differential equations to determine the trained model. The multimodal operational data of the biological fermentation process to be tested are input into the trained model to determine the predicted key fermentation parameters, including cell concentration, substrate concentration, and product concentration.

2. The bio-fermentation measurement method integrating physical mechanisms and multimodal data according to claim 1, characterized in that, The multimodal operation data of the bio-fermentation process includes real-time data from online sensors, spectral analysis data, process operation parameter data, and historical fermentation batch data. The standardization and time-series alignment of the multimodal operational data, to determine the processed data, includes: The multimodal operating data is subjected to outlier removal, noise filtering, missing value interpolation, and zero-mean normalization in sequence to determine the standardized multimodal data. A preset dynamic time warping algorithm is used to perform time-series alignment processing on the standardized multimodal data to determine the processed data.

3. The bio-fermentation measurement method integrating physical mechanisms and multimodal data according to claim 1, characterized in that, The set of physical mechanism constraint equations, constructed based on the core equations of microbial metabolic kinetics and the principle of material / energy balance, includes: Based on the metabolic laws of microbial fermentation, a cell growth kinetic model was established based on the Monod equation. The cell growth kinetic model is used to characterize the quantitative relationship between substrate concentration and cell specific growth rate. A product synthesis kinetic model was constructed based on the Luedeking-Piret equation. This model was used to distinguish the synthesis patterns of growth-related and non-growth-related products. Based on the material / energy balance principle, a coupling constraint equation is established between substrate consumption, cell growth, and product synthesis. The bacterial cell growth kinetics model, the product synthesis kinetics model, and the coupled constraint equations form the physical mechanism constraint equation set.

4. The bio-fermentation measurement method integrating physical mechanisms and multimodal data according to claim 1, characterized in that, The step of embedding the set of physical mechanism constraint equations into the loss function of the measurement model based on neural differential equations in a regularized form to determine a new loss function includes: L_total = L_data + λ × L_physics Where L_total represents the new loss function, L_data represents the data fitting error, λ represents the weight coefficient, and L_physics represents the physical mechanism constraint regularization term.

5. The bio-fermentation measurement method integrating physical mechanisms and multimodal data according to claim 1, characterized in that, The step of training the measurement model based on neural differential equations using the processed data and the new loss function to determine the trained model includes: The processed data is input into the measurement model based on neural differential equations to determine unconstrained predicted key fermentation parameters, including cell concentration, substrate concentration, and product concentration. Based on the unconstrained predicted key fermentation parameters and the actual key fermentation parameters, the data fitting error is determined; Based on the data fitting error and the new loss function, the measurement model based on neural differential equations is optimized to determine the trained model.

6. The bio-fermentation measurement method integrating physical mechanisms and multimodal data according to claim 1, characterized in that, The method further includes: Based on the prediction error between the predicted key fermentation parameters and the offline laboratory test values, fermentation batch information, and operating condition change information, the trained model is incrementally optimized to determine the optimized model.

7. The bio-fermentation measurement method integrating physical mechanisms and multimodal data according to claim 6, characterized in that, The step of incrementally optimizing the trained model based on the prediction error between the predicted key fermentation parameters and the offline laboratory test values, fermentation batch information, and operating condition change information, to determine the optimized model, includes: When the prediction error between the predicted key fermentation parameters and the offline laboratory test values ​​is greater than a preset prediction error threshold, the batch iteration number in the fermentation batch information is equal to a preset batch iteration threshold, or any one or more of the conditions in the operating condition change information exist, an incremental learning strategy is used to incrementally optimize the trained model, and the accuracy of the incrementally optimized model is verified to determine the optimized model. The operating condition change information includes changes in fermentation process, fermentation raw materials, and fermentation strains. The incremental learning strategy includes: loading the trained model, fixing the core network bottom-level parameters of the trained model, and unlocking the output layer and associated fully connected layer weights of the trained model. Collect newly added data that triggers incremental optimization and perform standardization and time-series alignment processing to determine the processed new data; Using the processed new data as training samples, the new loss function and Adam adaptive optimization algorithm are used to iteratively fine-tune the trained model, unlocking layer weights until the loss function converges or reaches the preset maximum number of iterations. The accuracy verification process includes: collecting a verification dataset that did not participate in this incremental optimization and matches the current fermentation conditions; The validation dataset is standardized and time-series aligned to determine the processed validation dataset. The processed validation dataset is input into the incrementally optimized model to determine the predicted key fermentation parameters of the validation dataset. Based on the relative error between the predicted key fermentation parameters of the validation dataset and the actual offline experimental key fermentation parameters of the validation dataset, and a preset accuracy threshold, if the relative error does not exceed the preset accuracy threshold, the accuracy is deemed qualified, and the incrementally optimized model is determined as the optimized model; otherwise, incremental optimization is repeated until the accuracy is qualified.

8. A bio-fermentation measurement system integrating physical mechanisms and multimodal data, characterized in that, include: The multimodal data acquisition module is used to collect multimodal operational data of the bio-fermentation process; The data preprocessing module is used to perform standardization and time-series alignment processing on the multimodal operating data to determine the processed data; The physical mechanism constraint construction module is used to construct a set of physical mechanism constraint equations based on the core equations of microbial metabolic kinetics and the principle of material / energy balance. The physical mechanism constraint fusion module is used to embed the set of physical mechanism constraint equations into the loss function of the measurement model based on neural differential equations in a regularized form to determine a new loss function; The model training module is used to train the measurement model based on neural differential equations using the processed data and the new loss function, and to determine the trained model. The fermentation key parameter prediction module is used to input the multimodal operation data of the biological fermentation process to be tested into the trained model to determine the predicted fermentation key parameters, including cell concentration, substrate concentration and product concentration.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-7.

10. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1-7.

Images