Preparation method of plateau senolytic and food dual-purpose fermented food targeting intestinal-lung axis

Abstract

The application discloses a preparation method of a plateau old-age medicine-food dual-purpose fermented food targeting an intestinal-lung axis, and belongs to the technical field of food preparation. The specific steps of the preparation method are as follows: I. selecting and pretreating medicine-food dual-purpose raw materials suitable for plateau old-age people, compounding the pretreated raw materials according to a preset ratio to form a fermentation substrate; II. screening and activating fermentation strains, preparing a composite fermentation liquid according to a preset strain group ratio, and then inoculating the activated strains into the fermentation substrate to form an initial fermentation system. The application significantly improves control precision and process stability, avoids fluctuations caused by human intervention, thereby improving the consistency and repeatability of the fermentation product, improving the adaptability and functional pertinence of the fermentation product to the plateau old-age people from the source, significantly enhancing the target orientation and functional verifiability, improving the search efficiency, and enhancing the robustness and convergence ability in the complex nonlinear fermentation process.

Description

Technical Field

[0001] This invention relates to the field of food preparation technology, and in particular to a method for preparing a fermented food for elderly people in high-altitude areas that targets the gut-lung axis and is both medicinal and edible. Background Technology

[0002] With the accelerating aging of the population, health problems among the elderly in high-altitude areas are becoming increasingly prominent. Due to the long-term low air pressure and low oxygen partial pressure in the high-altitude environment, the human body is prone to chronic hypoxia, decreased immune function, and gut microbiota imbalance, which in turn affects the coordinated stability of the respiratory and digestive systems. Studies have shown a close physiological connection between the gut and lungs, known as the "gut-lung axis," which plays a crucial role in regulating the body's immune response, inflammatory balance, and barrier function. However, existing nutritional interventions for the elderly in high-altitude areas often focus on regulating single systems, lacking a systematic solution for the overall coordinated regulation of the gut-lung axis.

[0003] On the other hand, while traditional fermented foods have some effect on improving gut microbiota, their production processes largely rely on empirical control and lack the ability to precisely regulate specific populations (especially elderly people in high-altitude areas). Furthermore, existing fermentation processes are typically conducted under normal pressure and oxygen conditions, making it difficult to simulate the induction of microbial metabolic pathways by the high-altitude environment, resulting in insufficient adaptability of the generated active ingredients to high altitudes. In addition, current technologies rely heavily on physicochemical indicators or empirical judgments to determine the fermentation endpoint, lacking an evaluation mechanism directly related to targeted physiological functions in the human body (such as the immune status of the gut-lung axis). Therefore, inventing a method for preparing fermented foods that target the gut-lung axis and are suitable for both medicinal and edible use in high-altitude areas is particularly important.

[0004] Existing methods for preparing fermented foods that are both medicinal and edible for the elderly in high-altitude areas suffer from fluctuations due to human intervention, resulting in low consistency and reproducibility of fermentation products. This reduces the suitability and functional specificity of the fermentation products for the elderly population in high-altitude areas, leading to poor target orientation and functional verifiability. Furthermore, these methods exhibit low robustness and convergence in complex nonlinear fermentation processes. To address these issues, we propose a method for preparing fermented foods that target the gut-lung axis for the elderly in high-altitude areas. Summary of the Invention

[0005] The purpose of this invention is to address the deficiencies in the existing technology by proposing a method for preparing a fermented food product that targets the gut-lung axis for elderly people in high-altitude areas and is both medicinal and edible.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for preparing a fermented food product targeting the gut-lung axis for elderly people in high-altitude areas, which can be used as both food and medicine, includes the following specific steps: I. Select and pre-treat medicinal and edible raw materials suitable for elderly people in high-altitude areas, and then mix the pre-treated raw materials according to a preset ratio to form a fermentation substrate; II. Screen fermentation strains and activate them for culture. Prepare a compound fermentation broth according to the preset microbial ratio. Then inoculate the activated strains into the fermentation substrate to form the initial fermentation system. III. Place the inoculated fermentation system in a controllable fermenter, set the basic parameters, and bring the fermentation system into a stable start-up state, and continuously collect data on the state of each organism. IV. Based on the changing trends of various biological state data in the fermentation system, dynamically output adjustment instructions for each basic parameter, and periodically correct the fermentation parameters; V. During the fermentation process, the low-pressure and low-oxygen conditions of the plateau environment are simulated, and the metabolic activity of microorganisms and their environmental response are monitored in real time, while the control mode is dynamically adjusted. VI. Pre-establish a digital twin of the fermentation process, input the real-time parameters in the fermenter into the twin model, simulate the future fermentation trajectory online, and perform counterfactual control optimization; VIII. In the later stage of fermentation, a fingerprint database of the gut-lung axis health status of elderly people in high-altitude areas was established. At the same time, multimodal characteristic signals of the fermentation broth were collected, and the corresponding target active ingredient spectra were selected. IX. After fermentation, the fermentation products are post-processed to obtain the final fermented food or fermented powder, which is then made into fermented food that can be used for both medicinal and food purposes, and packaged and stored.

[0007] As a further aspect of the present invention, the medicinal and edible raw materials mentioned in step I include one or more of the medicinal and edible plant raw materials such as grains, tubers, and fungi.

[0008] As a further aspect of the present invention, the specific steps of selecting and pre-treating medicinal and edible raw materials suitable for elderly people in high-altitude areas in step I, and then compounding the pre-treated raw materials according to a preset ratio to form a fermentation substrate are as follows: P1.1: First, spread the medicinal and edible raw materials suitable for the elderly in high-altitude areas on a clean operating table. Remove impurities such as mud, sand, stones, straw, metal fragments and moldy particles by manual picking, sieving, air separation or magnetic separation, so that the raw materials entering the next step have a high degree of purity and a relatively stable initial state. After the impurity removal is completed, record the net weight of the raw materials and the weight of impurities by batch. P1.2: Place the cleaned raw materials in running water. First, briefly soak them to remove dust and loose dirt from the surface of the raw materials. Then, gently rinse them to remove any remaining impurities. After washing, place the raw materials in a draining sieve, draining platform, or dewatering device for initial draining. Cut the drained raw materials according to their structure and control the size of each batch during the cutting process. Then, send the cut raw materials into low-temperature hot air drying, vacuum drying, or cold air pre-drying equipment for drying. P1.3: The dried raw materials are pulverized using a low-temperature pulverizing device, and the friction temperature rise is controlled during the pulverization process. After pulverization, the particles are screened through a sieve. Particles that are too large are sent back to the low-temperature pulverizing device for further pulverization until powder that meets the set particle size range is obtained. P1.4: Determine the proportion of each raw material according to its nutritional composition, fiber structure, fermentable sugar content and target flavor requirements. Weigh each component in batches and then use mechanical mixing, stirring or batch feeding to uniformly compound and form a fermentation substrate. According to the nutritional requirements of the target fermentation microorganisms, add appropriate amounts of carbon source, nitrogen source, minerals and buffer components to the compounded fermentation substrate. P1.5: Measure the actual moisture content of the current substrate, calculate the amount of water to be added based on the target moisture content, and then spray or slowly drip purified water in batches while continuously stirring to ensure that the water is fully diffused in the system. After the moisture content is adjusted to the appropriate level, the fermentation substrate should reach a state where it can be formed into a ball when squeezed in the hand but crumbles when lightly touched.

[0009] As a further aspect of the present invention, the specific requirements for the cutting process described in P1.2 are as follows: block-shaped raw materials are cut into thin slices, small cubes, or strips; stem and leaf raw materials are cut into small segments; mushroom raw materials are cut into uniform thin slices; and root and rhizome raw materials are cut into slices or granules of uniform thickness. The drying process described in P1.2 adopts a staged temperature control method, first removing surface moisture and then completing the internal moisture migration to reduce the loss of effective ingredients; The carbon source described in P1.4 is used to provide basic energy and promote cell proliferation, the nitrogen source is used to meet the needs of protein synthesis, the minerals are used to maintain enzyme activity, and the buffer components are used to stabilize the acid-base environment of the system.

[0010] As a further aspect of the present invention, the specific steps of screening and activating fermentation strains in step II, preparing a composite fermentation broth according to a preset microbial community ratio, and then inoculating the activated strains into the fermentation substrate to form the initial fermentation system are as follows: P2.1: Select strains to be used from food-grade microbial strain resource banks or known safe microbial strain banks, inoculate each strain into standard culture medium, and culture at 35 ℃~37 ℃ for 12 h~24 h. After they enter the stable proliferation stage, take samples to conduct acid resistance, bile salt resistance and co-culture inhibition tests. P2.2: Take out the selected strains and let the freeze-dried bacterial powder stand at 25 ℃~30 ℃ for 5 min~10 min to warm up. Then transfer it into sterile resuscitation solution or pre-sterilized seed culture medium for activation culture until the bacteria enter the end of the logarithmic growth phase or the early stable growth phase, and then stop the activation. P2.3: Determine the effective concentration of the activation solution for each strain, calculate the volume of each strain to be taken according to the target ratio, and add them sequentially to the composite container under aseptic conditions while gently stirring to ensure that each strain is evenly dispersed in the same liquid phase system to form a composite fermentation broth. Adjust the temperature of the fermentation substrate to 35 ℃~37 ℃ and the pH of the system to 5.8~6.4. Then, inoculate the prepared composite fermentation broth into the aforementioned fermentation substrate under aseptic conditions to form the initial fermentation system. At the same time, after inoculation, continue to maintain short-term uniform mixing to ensure that the bacteria and the substrate are in full contact to form a continuous and stable initial fermentation state.

[0011] It should be further noted that, in the acid resistance test described in P2.1, the strain survival rate should be no less than 85%, preferably 88%–95%, under conditions of pH 3.5–4.0 for 2 hours; the bile salt resistance test should be conducted no less than 80%, preferably 82%–92%, under conditions of bile salt mass fraction of 0.30%–0.50% for 3 hours; the strain survival rate should be no less than 80%, preferably 82%–92%; in the compatibility test, the diameter of the mutual inhibition zone should be no greater than 6 mm, preferably 0–3 mm; if the strain exhibits obvious hemolysis, abnormal color production, or other conditions that do not meet food-grade requirements, it should be directly rejected. If the selected strains in P2.2 are preserved in liquid culture medium, they should be directly inoculated into the seed culture medium under aseptic conditions. The activation culture temperature should be controlled at 35 ℃~37 ℃, pH at 6.0~6.5, shaking speed at 120 r / min~160 r / min, and activation time at 8 h~14 h. During the culture, the system should be kept slightly aerated to promote the recovery of metabolic activity of the bacteria, but strong shear damage should not be caused. The composition of the compound fermentation broth described in P2.3 can be set to a multi-strain synergistic mode according to fermentation requirements.

[0012] As a further aspect of the present invention, the specific steps of dynamically outputting adjustment commands for each basic parameter and periodically correcting fermentation parameters in step IV are as follows: S1.1: Based on the control object of the fermentation process, a fermentation control model is established, which includes a state input layer, a feature encoding layer, a strategy decision layer, and an action output layer. Then, the online monitoring data of each biological state in the fermentation system is defined as the state space input, the adjustment of each basic parameter is defined as the action space output, and the target response of the fermentation process is set as the reward signal. S1.2: After the fermentation control model is constructed, its action boundaries, state normalization method and reward calculation rules are set. Then, the historical fermentation batch data is input into the fermentation control model for initial parameter calibration. After that, the state sequence, control sequence and result sequence in the historical fermentation process are used as training samples and input into the fermentation control model for offline pre-training. S1.3: Linear interpolation is used to resample each training sample onto the same time grid. After time alignment is completed, Z-score standardization is performed on each training sample at all time steps. Then, each time step is traversed in chronological order to extract the standardized state, control, and the state at the next moment, forming a state-action-next state triplet. S1.4: Perform a weighted average of multiple expert actions corresponding to each state to obtain the optimal control action corresponding to each standardized state, and establish an expert demonstration dataset containing standardized states and corresponding optimal control actions. Then, after standardizing the state sequences in each training sample in the current training batch, input them into the fermentation control model, perform layer-by-layer processing based on forward propagation, and output the corresponding action vector. S1.5: Calculate the average loss between the action vector output by the fermentation control model and the action in the optimal control trajectory in the expert demonstration dataset. Then calculate the gradient value of this average loss with respect to the current model parameters. Update the parameters according to the gradient descent rule. After each training cycle, calculate the average loss of all training samples in this training cycle and compare it with the average loss of the previous training cycle. If the absolute value of the change in loss is less than 1×10 for 10 consecutive cycles, the model is considered successful. −5 If the historical batches that did not participate in training are used as the validation set, the average loss on the validation set is calculated. If the average loss on the validation set is less than 0.05, training is stopped and the final parameters are retained. Otherwise, the model parameter iteration is repeated. S1.6: After the fermentation process begins, the collected real-time biological state data is denoised, truncated and standardized, and multi-source data are merged into the input vector of the same state time according to the set time interval. This vector is then input into the pre-trained fermentation control model for online inference and outputs the current optimal action set, which includes temperature correction, pH correction, dissolved oxygen correction and stirring speed correction. S1.7: The fermentation control model outputs a set of adjustment instructions that meet the constraints based on the deviation between the current state and the target state, and transmits these adjustment instructions to the execution unit to correct the fermentation process parameters. After the fermentation control model outputs and executes the adjustment instructions, a new round of online monitoring data is input into the fermentation control model again. After each control cycle is completed, the corresponding reward value is calculated through the reward function based on the degree of closeness between the current fermentation state and the target state, and the control strategy for the next cycle is adjusted accordingly. At the same time, the state, action and reward data of each control cycle are backfilled into the model memory unit for the next round of updates.

[0013] As a further aspect of the present invention, the basic parameters mentioned in S1.1 specifically include temperature, pH, dissolved oxygen, and stirring speed; The state space input described in S1.1 contains multiple state vectors, specifically represented as follows: ,in, Representative moment The state vector; Representative moment Characteristic values ​​of short-chain fatty acids; Representative moment Characteristic values ​​of tryptophan metabolites; Representative moment Cytokine characteristic values; Representative moment The fermentation environment characteristics were obtained by standardizing temperature, pH, dissolved oxygen, and stirring conditions. The constraints described in S1.7 are as follows: the single adjustment range of temperature correction is -1.5 ℃ to 1.5 ℃; pH correction is -0.3 to 0.3; dissolved oxygen correction is -8% to 8%; and stirring speed correction is -20 r / min to 20 r / min.

[0014] As a further aspect of the present invention, the model memory unit described in S1.7 is used to perform small-step iterative updates on the fermentation control model parameters, so that the model gradually adapts to the dynamic characteristics of the current batch fermentation system, and prioritizes online fine-tuning during updates instead of retraining the entire batch. Meanwhile, when the magnitude of the action change output by the model is lower than the preset threshold in multiple consecutive cycles and the reward value remains stable, the model can be considered to have entered the stable control stage. At this time, the current strategy can be maintained. If the state deviates significantly in the future, the correction process will be restarted and the closed loop of "online input - action output - feedback correction" will continue to be executed.

[0015] As a further aspect of the present invention, the specific steps for simulating the low-pressure, low-oxygen conditions of a plateau environment during fermentation, as described in step V, are as follows: S2.1: Place the entire fermentation unit, which has been loaded with the fermenter, into the programmable high-altitude environment simulation chamber. Then close the chamber door and lock it in an airtight manner to put the chamber into a sealed working state. After sealing, perform zero-point calibration and linear calibration on each pressure sensor, oxygen partial pressure sensor and pressure regulating actuator in the simulation chamber. During calibration, keep the chamber under normal pressure and normal oxygen reference conditions and read the corresponding baseline values. If the calibration deviation exceeds the preset range, re-zero the sensor or replace the calibration point and recalibrate. S2.2: After completing the cabin calibration, the cabin air pressure and oxygen partial pressure are designed as a dynamic change mode with segmented decrease and platform transition according to the preset plateau simulation target, so as to establish the corresponding gradient transition curve. S2.3: After establishing the gradient transition curve, drive the cabin pressure regulating valve, exhaust and gas extraction components and gas replenishment components to work together according to the predetermined time node to regulate the air pressure, so that the air pressure in the cabin gradually decreases along the set trajectory. While the air pressure regulation is carried out simultaneously, a low-oxygen mixed gas replacement method is adopted to adjust the gas supply ratio and gas replacement rate, so that the oxygen partial pressure in the cabin decreases synchronously according to the preset curve. S2.4: During the synchronous change of air pressure and oxygen partial pressure, the two parameters are judged in a linkage manner at a sampling cycle of 5 minutes. After each sampling cycle, the joint deviation of air pressure and oxygen partial pressure is calculated. If the joint deviation is ≤0.06, the linkage control is considered stable and the next stage transition continues. If 0.06 < joint deviation ≤0.1, it is an allowable correction range and feedback is given to the staff to manually select whether to correct. If 0.1 < joint deviation, it is a forced correction range and the next step action is immediately paused. The correction is performed first and then the descent continues. S2.5: When the air pressure and oxygen partial pressure inside the high-altitude environment simulation chamber drop to the target window, the transition stops and enters a stable maintenance phase, maintaining the chamber environment in a simulated state of low pressure and low oxygen at high altitude. The target window air pressure is 80.0 kPa to 82.5 kPa and oxygen partial pressure is 12.0 kPa to 13.5 kPa. During the stable maintenance phase, the pressure fluctuation range does not exceed ±0.8 kPa and the oxygen partial pressure fluctuation range does not exceed ±0.4 kPa. When the fluctuation of air pressure and oxygen partial pressure inside the chamber is below the constraint range for three consecutive sampling cycles, it is determined that the target simulation state has been stabilized.

[0016] As a further aspect of the present invention, the baseline values ​​in S2.1 include baseline pressure, typically 100.8 kPa to 101.5 kPa, and baseline oxygen partial pressure, typically 20.6 kPa to 21.0 kPa. The calibrated pressure measurement error is preferably no more than ±0.3 kPa, and the oxygen partial pressure measurement error is preferably no more than ±0.15 kPa. The segmented decrease described in S2.2 is as follows: the gas pressure gradually decreases from 101.3 kPa to 78.0 kPa to 84.0 kPa, and the oxygen partial pressure gradually decreases from 21.0 kPa to 11.5 kPa to 14.5 kPa; the duration of each transition segment is preferably 20 min to 45 min, and a buffer transition segment of 3 min to 8 min can be set between adjacent segments.

[0017] In addition, it should be noted that during the air pressure regulation process described in S2.3, each air pressure change is controlled within a single step of 0.5 kPa to 2.5 kPa, preferably 1.0 kPa to 1.8 kPa; if the air pressure in the cabin drops too quickly at a certain stage, the pressure reduction is immediately stopped, and sterile inert gas or clean air is added in the next control cycle for a slight adjustment so that the air pressure returns to the vicinity of the current target trajectory; During the gas supply ratio and gas replacement rate adjustment process described in S2.3, the oxygen partial pressure is controlled to change between 0.2 kPa and 1.0 kPa each time, preferably between 0.4 kPa and 0.8 kPa. If the oxygen partial pressure decreases too slowly, the proportion of low-oxygen gas is increased. If the decrease is too fast, the current ratio is maintained for a short time to bring the oxygen partial pressure back close to the target curve.

[0018] As a further aspect of the present invention, the specific steps of pre-establishing a digital twin of the fermentation process in step VI, inputting real-time parameters from the fermenter into the twin model, and simulating the future fermentation trajectory online are as follows: S3.1: Divide the fermentation system into multiple computational units, and then establish mass conservation relationships, temperature transfer relationships and metabolic generation relationships for each unit based on the structural parameters of the fermenter, stirring and mass transfer characteristics, microbial kinetic characteristics and target product generation characteristics, so as to construct a digital twin corresponding to the actual fermenter. S3.2: After the digital twin is established, the real-time parameters collected in the fermenter are preprocessed and mapped into the digital twin. After the real-time parameters are input, the digital twin calculates the future fermentation trajectory by rolling forward at a time step of 90 min to 180 min. At the same time, during the rolling simulation, the digital twin first estimates the state at the next moment based on the current state, and then uses the estimated state as the initial condition for the next round of simulation for iterative updates. S3.3: After obtaining the future fermentation trajectory, extract virtual samples corresponding to the fermentation products from the predicted sequence of the digital twin and map them to indicators. Then, read the predicted product concentration, acid metabolite ratio and low molecular weight active component spectrum every 15 minutes. At the same time, calculate its response to organoid phenotypes, specifically including intestinal organoid response value and lung organoid response value. S3.4: Normalize the response values ​​of intestinal organoids and lung organoids respectively, and then sum them according to preset weights to obtain a comprehensive prediction score of fermentation product suitability. Then compare the comprehensive prediction score with the preset target interval. If the comprehensive prediction scores of multiple prediction points are all within the target window within the next 2 hours, it is considered that the current fermentation trajectory can continue. If the comprehensive prediction score is lower than the target window, it is indicated that the current trajectory of the digital twin deviates from the target and needs to be readjusted.

[0019] In addition, it should be noted that the digital twin described in S3.1 must be consistent with the actual fermentation equipment. Therefore, the geometric boundaries, material boundaries and time step boundaries are set for the digital twin based on the actual fermentation equipment, and the process trajectory that has shown stability in historical batches is used as the initial calibration benchmark. The real-time parameters mentioned in S3.2 include, but are not limited to, local temperature, local pH, local dissolved oxygen, stirring load, foam state, and liquid level changes. The indicators mentioned in S3.3 include, but are not limited to, secretory mucus-related quantities, inflammation balance-related quantities, barrier stability-related quantities, and oxidative stress-related quantities; The comprehensive prediction score described in S3.4 can be used to describe the stability of the epithelial barrier, the level of mucus secretion, the balance of inflammatory factors, and the control of oxidative stress in the co-culture system.

[0020] As a further aspect of the present invention, the specific steps for performing counterfactual control optimization in step VI are as follows: S4.1: Organize the control variables to be optimized in the fermentation process into control sequence vectors, and set the feasible region of each control sequence vector to establish a corresponding search space. At the same time, combine and encode according to the set time window. Then set the target activity benchmark value to 0.94 and the stability window to [0.90, 0.97]. The control variables to be optimized include temperature adjustment, pH adjustment, dissolved oxygen adjustment and stirring speed adjustment. S4.2: After establishing the search space, based on the historical control sequences and their corresponding target activity results, nonlinear fitting is performed on the control sequence-target activity. The correlation between different control parameters is learned through the kernel function. The candidate control sequence is used as input and the corresponding target activity score is used as output. S4.3: After the fitting is completed, calculate the current round of historical best target activity score, then jointly evaluate the predicted mean and predicted variance of the candidate control sequences, and select the sequence with the largest collected value as the optimal control sequence to input into the digital twin to obtain the corresponding comprehensive prediction score, and then compare it with the target activity benchmark to determine whether the current optimization direction deviates from the preset path. S4.4: If the prediction result is within the target window, continue searching around the control sequence; if the prediction result deviates from the target window, perturb the current control sequence, construct a counterfactual control sequence, and after generating the counterfactual control sequence, interpolate and correct it with the current control sequence to form the corrected control sequence. S4.5: The corrected control sequence is sent to the fermentation execution unit. After execution, the new trajectory is evaluated again through the digital twin and its proximity to the target activity is compared. If the target activity of the corrected trajectory is stably within the target window for multiple consecutive iterations, the control sequence is considered to have converged. If there is still a deviation, the iterative update of the control sequence continues.

[0021] As a further aspect of the present invention, the specific steps of establishing a fingerprint database of the gut-lung axis health status of elderly people in high-altitude areas, simultaneously collecting multimodal feature signals from fermentation broth, and selecting the corresponding target active ingredient profiles in step VIII are as follows: S5.1: The gut microbiota metagenomic characteristics, serum metabolomics characteristics, and lung function indicators of each plateau elderly subject sample collected and preprocessed are uniformly encoded and health status fingerprint entries are formed according to the individual dimension to establish a health status fingerprint map library. Each fingerprint entry corresponds to the basic health status of a subject. S5.2: In the later stage of fermentation, the fermentation broth is placed in a fixed flow state, and multimodal characteristic signals, including near-infrared spectral signals, electronic nose response signals and electronic tongue potential signals, are collected simultaneously from the fermentation broth. The near-infrared, electronic nose and electronic tongue signals are then aligned according to a unified timestamp. Baseline drift correction and amplitude normalization are then performed on each channel to unify the data from different sources to the same numerical scale. If any channel shows a sudden change within three consecutive sampling periods, that channel is treated as an abnormal channel and local interpolation is performed for repair. S5.3: After preprocessing, the modal signals are spliced ​​into a one-dimensional input sequence in the order of spectrum-odor-potential, and then fed into a one-dimensional convolutional neural network. The sequence is processed layer by layer through convolutional layer, activation layer, pooling layer and fully connected layer to extract the feature tensor of the fermentation broth in the multimodal state, and then it is matched with each fingerprint in the healthy state fingerprint database for similarity. S5.4: Calculate the similarity value between the feature tensor of each fermentation broth and each healthy fingerprint using cosine similarity, and then correct the matching results by combining individual adaptation weights to generate a comprehensive matching score. Sort each healthy fingerprint from high to low according to the comprehensive matching score, and select the one with the highest comprehensive matching score as the target active ingredient spectrum template corresponding to the current fermentation broth. Then, extract the corresponding ideal active ingredient combination range according to the template and output it as the target spectrum for subsequent updates.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: This method for preparing fermented foods for the elderly in high-altitude regions, targeting the gut-lung axis, constructs a fermentation control model that includes state input, feature encoding, strategy decision-making, and action output. It uses online monitored biological state data as input, process parameter adjustments as output, and the fermentation target response as a reward signal. Parameter calibration and offline pre-training are performed using historical batch data to form a stable intelligent control strategy. After fermentation begins, real-time data is collected, pre-processed, and input into the model to achieve dynamic closed-loop control of parameters such as temperature, pH, dissolved oxygen, and stirring. Simultaneously, the control strategy is continuously optimized by combining reward feedback. The fermentation unit is placed in a programmable high-altitude environment simulation chamber. Through gradient transitions and linkage control of air pressure and oxygen partial pressure, accurate simulation of a low-pressure, low-oxygen environment is achieved. A deviation judgment mechanism ensures environmental stability. A digital twin corresponding to the actual fermentation process is constructed to predict future fermentation trajectories. The system performs rolling predictions and assesses product suitability using organoid response indicators. When predictions deviate from the target, Bayesian optimization and counterfactual intervention strategies are introduced to iteratively correct the control sequence until convergence. A fingerprint database of gut-lung axis health status is established for elderly people in high-altitude areas. Multimodal signals from fermentation broth are collected, and feature tensors are extracted using a one-dimensional convolutional neural network. Similarity matching is performed with the health fingerprints to automatically select the optimal target active ingredient spectrum. This enables personalized determination of the fermentation endpoint and function-oriented optimization, significantly improving control accuracy and process stability, avoiding fluctuations caused by human intervention, and thus improving the consistency and reproducibility of fermentation products. This enhances the suitability and functional specificity of fermentation products for elderly people in high-altitude areas from the source, significantly strengthening target orientation and functional verifiability. It not only improves search efficiency but also enhances robustness and convergence ability in complex nonlinear fermentation processes. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0024] Figure 1 This is a flowchart illustrating a method for preparing a fermented food product for high-altitude elderly people that targets the gut-lung axis, according to the present invention. Detailed Implementation

[0025] Example 1, referring to Figure 1 A method for preparing a fermented food product for elderly people in high-altitude areas that targets the gut-lung axis, the specific steps of which are as follows: Select and pre-treat medicinal and edible raw materials suitable for elderly people in high-altitude areas, and then combine the pre-treated raw materials in a preset ratio to form a fermentation substrate.

[0026] Fermentation strains are screened and activated, and a compound fermentation broth is prepared according to a preset microbial ratio. The activated strains are then inoculated into the fermentation substrate to form the initial fermentation system.

[0027] The inoculated fermentation system was placed in a controllable fermenter, and the basic parameters were set to bring the fermentation system into a stable start-up state, and data on the state of each organism were continuously collected.

[0028] Based on the changing trends of various biological states in the fermentation system, dynamic output instructions for adjusting each basic parameter are generated, and fermentation parameters are periodically corrected.

[0029] Specifically, a fermentation control model is established based on the controlled object of the fermentation process, comprising a state input layer, a feature encoding layer, a strategy decision layer, and an action output layer. Then, the online monitored biological state data in the fermentation system are defined as the state space input, the adjustment variables of each basic parameter are defined as the action space output, and the target response of the fermentation process is set as the reward signal. After the fermentation control model is constructed, its action boundaries, state normalization method, and reward calculation rules are set. Historical fermentation batch data are then input into the fermentation control model for initial parameter calibration. Subsequently, the state sequence, control sequence, and result sequence of the historical fermentation process are used as training samples and input into the fermentation control model for offline pre-training. Linear interpolation is used to resample each training sample onto the same time grid. After time alignment, Z-score standardization is performed on each training sample at all time steps. Then, each time step is traversed in chronological order to extract the standardized data. The state, control, and next state constitute a state-action-next state triple. A weighted average of multiple expert actions corresponding to each state is used to obtain the optimal control action for each standardized state. An expert demonstration dataset containing standardized states and their corresponding optimal control actions is established. Then, the state sequences of each training sample in the current training batch are standardized and input into the fermentation control model. Processing is performed layer by layer based on forward propagation, and the corresponding action vectors are output. The average loss between the action vector output by the fermentation control model and the action in the optimal control trajectory in the expert demonstration dataset is calculated. The gradient of this average loss with respect to the current model parameters is then calculated, and the parameters are updated according to the gradient descent rule. After each training cycle, the average loss of all training samples in this cycle is calculated and compared with the average loss of the previous training cycle. If the absolute value of the loss change is less than 1 × 10 for 10 consecutive cycles, the model is considered successful. −5If the historical batches not used in training are used as the validation set, the average loss on the validation set is calculated. If the average loss on the validation set is less than 0.05, training is stopped and the final parameters are retained; otherwise, the model parameter iteration is repeated. After the fermentation process begins, the collected real-time biological state data is denoised, truncated, and standardized. Multi-source data are merged into an input vector at the same state time according to a set time interval. This vector is then input into the pre-trained fermentation control model for online inference and outputs the current optimal action set, including temperature correction, pH correction, dissolved oxygen correction, and stirring rotation. The rapid correction parameter is generated by the fermentation control model. Based on the deviation between the current state and the target state, the model outputs a set of adjustment instructions that meet the constraints and transmits these instructions to the execution unit to correct the fermentation process parameters. After the fermentation control model outputs and executes the adjustment instructions, a new round of online monitoring data is input into the fermentation control model again. After each control cycle is completed, the corresponding reward value is calculated through the reward function based on the closeness of the current fermentation state to the target state, and the control strategy for the next cycle is adjusted accordingly. At the same time, the state, action, and reward data of each control cycle are filled back into the model memory unit for the next update.

[0030] During the fermentation process, the low-pressure and low-oxygen conditions of the plateau environment are simulated, and the metabolic activity of microorganisms and their environmental response are monitored in real time, while the control mode is dynamically adjusted.

[0031] Specifically, the entire fermentation unit, already loaded with the fermenter, is placed into the programmable high-altitude environment simulation chamber. The chamber door is then closed and airtightly locked, putting the chamber into a sealed working state. After sealing, zero-point and linear calibrations are performed on all pressure sensors, oxygen partial pressure sensors, and pressure regulating actuators within the simulation chamber. During calibration, baseline values ​​are read under normal pressure and normal oxygen reference conditions within the chamber. If the calibration deviation exceeds the preset range, the sensors are zeroed again or the calibration points are replaced before recalibration. After completing the chamber calibration, the high-altitude simulation objectives are determined according to the preset goals. The design incorporates a segmented, progressively decreasing dynamic pattern for cabin pressure and oxygen partial pressure, with a plateau transition, to establish corresponding gradient transition curves. After establishing these curves, the cabin pressure regulating valve, exhaust / exhaust assembly, and gas supply assembly are activated at predetermined time points to regulate the pressure, causing it to gradually decrease along a set trajectory. Simultaneously, a low-oxygen mixed gas replacement method is employed to adjust the gas supply ratio and replacement rate, ensuring the cabin oxygen partial pressure decreases synchronously along a preset curve. During this synchronous change in pressure and oxygen partial pressure, the pressure is adjusted every 5... A sampling cycle is used to determine the linkage between two parameters. After each sampling cycle, the joint deviation of air pressure and oxygen partial pressure is calculated. If the joint deviation is ≤0.06, the linkage control is considered stable, and the next stage transition continues. If 0.06 < joint deviation ≤0.1, it is within the allowable correction range, and feedback is given to the staff to manually select whether to correct. If 0.1 < joint deviation, it is within the mandatory correction range, and the next step action is immediately paused. Correction is performed before continuing the descent. When the air pressure and oxygen partial pressure in the high-altitude environment simulation chamber drop to the target window, the transition stops, and a stable holding phase begins, maintaining the chamber environment in a simulated high-altitude low-pressure, low-oxygen state. The target window air pressure is 80.0 kPa to 82.5 kPa, and the oxygen partial pressure is 12.0 kPa to 13.5 kPa. During the stable holding phase, the pressure fluctuation does not exceed ±0.8 kPa, and the oxygen partial pressure fluctuation does not exceed ±0.4 kPa. When the cabin air pressure and oxygen partial pressure fluctuations are all below the constraint range for three consecutive sampling cycles, it is determined that the target simulation state has been stabilized.

[0032] Example 2, refer to Figure 1 A method for preparing a fermented food product for elderly people in high-altitude areas that targets the gut-lung axis, the specific steps of which are as follows: A digital twin of the fermentation process is pre-built, and real-time parameters from the fermenter are input into the twin model to simulate the future fermentation trajectory online and perform counterfactual control optimization.

[0033] Specifically, the fermentation system is divided into multiple computational units. Based on the structural parameters of the fermenter, stirring and mass transfer characteristics, microbial kinetics, and target product generation characteristics, mass conservation relationships, temperature transfer relationships, and metabolic generation relationships are established for each unit to construct a digital twin corresponding to the actual fermenter. After the digital twin is established, the real-time parameters collected in the fermenter are preprocessed and mapped into the digital twin. After the real-time parameters are input, the digital twin calculates the future fermentation trajectory by rolling forward at time steps of 90 min to 180 min. During the rolling simulation, the digital twin first estimates the state at the next moment based on the current state, and then uses the estimated state as the initial condition for the next round of simulation for iterative updates. After obtaining the future fermentation trajectory, virtual samples corresponding to the fermentation product generation are extracted from the predicted sequence of the digital twin and their indicators are mapped. Then, every 15 min... At minute intervals, the concentration of predicted products, the proportion of acidic metabolites, and the spectrum of low-molecular-weight active components are read. Simultaneously, their responses to organoid phenotypes are calculated, specifically including intestinal organoid response values ​​and lung organoid response values. The intestinal organoid response values ​​and lung organoid response values ​​are normalized separately, and then weighted and summarized according to preset weights to obtain a comprehensive predicted score for fermentation product suitability. The comprehensive predicted score is then compared with a preset target interval. If the comprehensive predicted scores of multiple prediction points within the next 2 hours are all within the target window, the current fermentation trajectory is considered to be maintained. If the comprehensive predicted score is lower than the target window, it indicates that the current trajectory of the digital twin has deviated from the target and needs to be readjusted.

[0034] Specifically, the control variables to be optimized during the fermentation process are organized into control sequence vectors, and feasible regions are set for each control sequence vector to establish a corresponding search space. Simultaneously, they are combined and encoded according to a set time window. The target activity baseline value is set to 0.94, and the stability window is [0.90, 0.97]. The control variables to be optimized include temperature adjustment, pH adjustment, dissolved oxygen adjustment, and stirring speed adjustment. After establishing the search space, based on historical control sequences and their corresponding target activity results, a nonlinear fitting is performed on the control sequence-target activity. The correlation between different control parameters is learned through a kernel function. Candidate control sequences are used as input, and the corresponding target activity scores are used as output. After fitting, the current round's historical best target activity score is calculated. Then, the predicted mean and predicted variance of the candidate control sequences are jointly evaluated, and the collected values ​​are selected. The largest sequence is used as the optimal control sequence and input into the digital twin to obtain the corresponding comprehensive prediction score. This score is then compared with the target activity benchmark to determine whether the current optimization direction deviates from the preset path. If the prediction result is within the target window, the search continues around the control sequence. If the prediction result deviates from the target window, the current control sequence is perturbed to construct a counterfactual control sequence. After generating the counterfactual control sequence, it is interpolated and corrected with the current control sequence to form a corrected control sequence. The corrected control sequence is then sent to the fermentation execution unit. After execution, the new trajectory is evaluated again through the digital twin, and its proximity to the target activity is compared. If the target activity of the corrected trajectory remains stable within the target window for multiple consecutive iterations, the control sequence is considered to have converged. If there is still a deviation, the iterative update of the control sequence continues.

[0035] In the later stage of fermentation, a fingerprint database of the gut-lung axis health status of elderly people in high-altitude areas was established. At the same time, multimodal characteristic signals of the fermentation broth were collected, and the corresponding target active ingredient spectra were selected.

[0036] Specifically, the metagenomic characteristics of gut microbiota, serum metabolomics, and lung function indicators of pre-collected and pre-processed samples from elderly subjects at high altitudes were uniformly encoded and used to create health status fingerprint entries at the individual level, thus establishing a health status fingerprint database. Each fingerprint entry corresponds to the baseline health status of a subject. During the later stages of fermentation, the fermentation broth was placed in a fixed flow state, and multimodal characteristic signals, including near-infrared spectral signals, electronic nose response signals, and electronic tongue potential signals, were simultaneously collected from the fermentation broth. The near-infrared, electronic nose, and electronic tongue signals were then aligned according to a unified timestamp. Baseline drift correction and amplitude normalization were then performed on each channel to unify data from different sources to the same numerical scale. If any channel exhibits a sudden change within three consecutive sampling periods, that channel is considered an abnormal channel and subjected to local processing. After interpolation and preprocessing, the modal signals are concatenated into a one-dimensional input sequence in the order of spectrum-odor-potential, and then fed into a one-dimensional convolutional neural network. The sequence is processed layer by layer through convolutional layers, activation layers, pooling layers, and fully connected layers to extract the feature tensors of the fermentation broth in multiple modal states. These feature tensors are then matched with fingerprints in the healthy state fingerprint database. Cosine similarity is used to calculate the similarity value between each fermentation broth feature tensor and each healthy fingerprint. The matching results are then corrected by individual adaptation weights to generate a comprehensive matching score. The healthy fingerprints are sorted from high to low according to the comprehensive matching score, and the one with the highest comprehensive matching score is selected as the target active ingredient spectrum template for the current fermentation broth. Subsequently, the ideal combination range of active ingredients is extracted according to this template and output as the target spectrum for subsequent updates.

[0037] After fermentation, the fermentation products are post-processed to obtain the final fermented food or fermented powder, which is then made into fermented food that can be used for both medicinal and edible purposes, and packaged and stored.

Claims

1. A method for preparing a fermented food product for high-altitude elderly people that targets the gut-lung axis, characterized in that, The specific steps of this preparation method are as follows: I. Select and pre-treat medicinal and edible raw materials suitable for elderly people in high-altitude areas, and then mix the pre-treated raw materials according to a preset ratio to form a fermentation substrate; II. Screen fermentation strains and activate them for culture. Prepare a compound fermentation broth according to the preset microbial ratio. Then inoculate the activated strains into the fermentation substrate to form the initial fermentation system. III. Place the inoculated fermentation system in a controllable fermenter, set the basic parameters, and bring the fermentation system into a stable start-up state, and continuously collect data on the state of each organism. IV. Based on the changing trends of various biological state data in the fermentation system, dynamically output adjustment instructions for each basic parameter, and periodically correct the fermentation parameters; V. During the fermentation process, the low-pressure and low-oxygen conditions of the plateau environment are simulated, and the metabolic activity of microorganisms and their environmental response are monitored in real time, while the control mode is dynamically adjusted. VI. Pre-establish a digital twin of the fermentation process, input the real-time parameters in the fermenter into the twin model, simulate the future fermentation trajectory online, and perform counterfactual control optimization; VIII. In the later stage of fermentation, a fingerprint database of the gut-lung axis health status of elderly people in high-altitude areas was established. At the same time, multimodal characteristic signals of the fermentation broth were collected, and the corresponding target active ingredient spectra were selected. IX. After fermentation, the fermentation products are post-processed to obtain the final fermented food or fermented powder, which is then made into fermented food that can be used for both medicinal and food purposes, and packaged and stored.

2. The method for preparing a fermented food for high-altitude elderly people that targets the gut-lung axis, as described in claim 1, is characterized in that... The specific steps for dynamically outputting adjustment commands for each basic parameter and periodically correcting fermentation parameters as described in step IV are as follows: S1.1: Based on the control object of the fermentation process, a fermentation control model is established, which includes a state input layer, a feature encoding layer, a strategy decision layer, and an action output layer. Then, the online monitoring data of each biological state in the fermentation system is defined as the state space input, the adjustment of each basic parameter is defined as the action space output, and the target response of the fermentation process is set as the reward signal. S1.2: After the fermentation control model is constructed, its action boundaries, state normalization method and reward calculation rules are set. Then, the historical fermentation batch data is input into the fermentation control model for initial parameter calibration. After that, the state sequence, control sequence and result sequence in the historical fermentation process are used as training samples and input into the fermentation control model for offline pre-training. S1.3: Linear interpolation is used to resample each training sample onto the same time grid. After time alignment is completed, Z-score standardization is performed on each training sample at all time steps. Then, each time step is traversed in chronological order to extract the standardized state, control, and the state at the next moment, forming a state-action-next state triplet. S1.4: Perform a weighted average of multiple expert actions corresponding to each state to obtain the optimal control action corresponding to each standardized state, and establish an expert demonstration dataset containing standardized states and corresponding optimal control actions. Then, after standardizing the state sequences in each training sample in the current training batch, input them into the fermentation control model, perform layer-by-layer processing based on forward propagation, and output the corresponding action vector. S1.5: Calculate the average loss between the action vector output by the fermentation control model and the action in the optimal control trajectory in the expert demonstration dataset. Then calculate the gradient value of this average loss with respect to the current model parameters. Update the parameters according to the gradient descent rule. After each training cycle, calculate the average loss of all training samples in this training cycle and compare it with the average loss of the previous training cycle. If the absolute value of the change in loss is less than 1×10 for 10 consecutive cycles, the model is considered successful. −5 If the historical batches that did not participate in training are used as the validation set, the average loss on the validation set is calculated. If the average loss on the validation set is less than 0.05, training is stopped and the final parameters are retained. Otherwise, the model parameter iteration is repeated. S1.6: After the fermentation process begins, the collected real-time biological state data is denoised, truncated and standardized, and multi-source data are merged into the input vector of the same state time according to the set time interval. This vector is then input into the pre-trained fermentation control model for online inference and outputs the current optimal action set, which includes temperature correction, pH correction, dissolved oxygen correction and stirring speed correction. S1.7: The fermentation control model outputs a set of adjustment instructions that meet the constraints based on the deviation between the current state and the target state, and transmits these adjustment instructions to the execution unit to correct the fermentation process parameters. After the fermentation control model outputs and executes the adjustment instructions, a new round of online monitoring data is input into the fermentation control model again. After each control cycle is completed, the corresponding reward value is calculated through the reward function based on the degree of closeness between the current fermentation state and the target state, and the control strategy for the next cycle is adjusted accordingly. At the same time, the state, action and reward data of each control cycle are backfilled into the model memory unit for the next round of updates.

3. The method for preparing a fermented food for high-altitude elderly people that targets the gut-lung axis, as described in claim 1, is characterized in that... Step V describes the specific steps for simulating the low-pressure, low-oxygen conditions of a plateau environment during fermentation: S2.1: Place the entire fermentation unit, which has been loaded with the fermenter, into the programmable high-altitude environment simulation chamber. Then close the chamber door and lock it in an airtight manner to put the chamber into a sealed working state. After sealing, perform zero-point calibration and linear calibration on each pressure sensor, oxygen partial pressure sensor and pressure regulating actuator in the simulation chamber. During calibration, keep the chamber under normal pressure and normal oxygen reference conditions and read the corresponding baseline values. If the calibration deviation exceeds the preset range, re-zero the sensor or replace the calibration point and recalibrate. S2.2: After completing the cabin calibration, the cabin air pressure and oxygen partial pressure are designed as a dynamic change mode with segmented decrease and platform transition according to the preset plateau simulation target, so as to establish the corresponding gradient transition curve. S2.3: After establishing the gradient transition curve, drive the cabin pressure regulating valve, exhaust and gas extraction components and gas replenishment components to work together according to the predetermined time node to regulate the air pressure, so that the air pressure in the cabin gradually decreases along the set trajectory. While the air pressure regulation is carried out simultaneously, a low-oxygen mixed gas replacement method is adopted to adjust the gas supply ratio and gas replacement rate, so that the oxygen partial pressure in the cabin decreases synchronously according to the preset curve. S2.4: During the synchronous change of air pressure and oxygen partial pressure, the two parameters are judged in a linkage manner at a sampling cycle of 5 minutes. After each sampling cycle, the joint deviation of air pressure and oxygen partial pressure is calculated. If the joint deviation is ≤0.06, the linkage control is considered stable and the next stage transition continues. If 0.06 < joint deviation ≤0.1, it is an allowable correction range and feedback is given to the staff to manually select whether to correct. If 0.1 < joint deviation, it is a forced correction range and the next step action is immediately paused. The correction is performed first and then the descent continues. S2.5: When the air pressure and oxygen partial pressure inside the high-altitude environment simulation chamber drop to the target window, the transition stops and enters a stable maintenance phase, maintaining the chamber environment in a simulated state of low pressure and low oxygen at high altitude. The target window air pressure is 80.0 kPa to 82.5 kPa and oxygen partial pressure is 12.0 kPa to 13.5 kPa. During the stable maintenance phase, the pressure fluctuation range does not exceed ±0.8 kPa and the oxygen partial pressure fluctuation range does not exceed ±0.4 kPa. When the fluctuation of air pressure and oxygen partial pressure inside the chamber is below the constraint range for three consecutive sampling cycles, it is determined that the target simulation state has been stabilized.

4. The method for preparing a fermented food for high-altitude elderly people that targets the gut-lung axis, as described in claim 1, is characterized in that... Step VI, which involves pre-establishing a digital twin of the fermentation process, inputting real-time parameters from the fermenter into the twin model, and simulating the future fermentation trajectory online, follows these specific steps: S3.1: Divide the fermentation system into multiple computational units, and then establish mass conservation relationships, temperature transfer relationships and metabolic generation relationships for each unit based on the structural parameters of the fermenter, stirring and mass transfer characteristics, microbial kinetic characteristics and target product generation characteristics, so as to construct a digital twin corresponding to the actual fermenter. S3.2: After the digital twin is established, the real-time parameters collected in the fermenter are preprocessed and mapped into the digital twin. After the real-time parameters are input, the digital twin calculates the future fermentation trajectory by rolling forward at a time step of 90 min to 180 min. At the same time, during the rolling simulation, the digital twin first estimates the state at the next moment based on the current state, and then uses the estimated state as the initial condition for the next round of simulation for iterative updates. S3.3: After obtaining the future fermentation trajectory, extract virtual samples corresponding to the fermentation products from the predicted sequence of the digital twin and map them to indicators. Then, read the predicted product concentration, acid metabolite ratio and low molecular weight active component spectrum every 15 minutes. At the same time, calculate its response to organoid phenotypes, specifically including intestinal organoid response value and lung organoid response value. S3.4: Normalize the response values ​​of intestinal organoids and lung organoids respectively, and then sum them according to preset weights to obtain a comprehensive prediction score of fermentation product suitability. Then compare the comprehensive prediction score with the preset target interval. If the comprehensive prediction scores of multiple prediction points are all within the target window within the next 2 hours, it is considered that the current fermentation trajectory can continue. If the comprehensive prediction score is lower than the target window, it is indicated that the current trajectory of the digital twin deviates from the target and needs to be readjusted.

5. The method for preparing a fermented food for high-altitude elderly people that targets the gut-lung axis, as described in claim 4, is characterized in that... The specific steps for performing counterfactual control optimization as described in step VI are as follows: S4.1: Organize the control variables to be optimized in the fermentation process into control sequence vectors, and set the feasible region of each control sequence vector to establish a corresponding search space. At the same time, combine and encode according to the set time window. Then set the target activity benchmark value to 0.94 and the stability window to [0.90, 0.97]. The control variables to be optimized include temperature adjustment, pH adjustment, dissolved oxygen adjustment and stirring speed adjustment. S4.2: After establishing the search space, based on the historical control sequences and their corresponding target activity results, nonlinear fitting is performed on the control sequence-target activity. The correlation between different control parameters is learned through the kernel function. The candidate control sequence is used as input and the corresponding target activity score is used as output. S4.3: After the fitting is completed, calculate the current round of historical best target activity score, then jointly evaluate the predicted mean and predicted variance of the candidate control sequences, and select the sequence with the largest collected value as the optimal control sequence to input into the digital twin to obtain the corresponding comprehensive prediction score, and then compare it with the target activity benchmark to determine whether the current optimization direction deviates from the preset path. S4.4: If the prediction result is within the target window, continue searching around the control sequence; if the prediction result deviates from the target window, perturb the current control sequence, construct a counterfactual control sequence, and after generating the counterfactual control sequence, interpolate and correct it with the current control sequence to form the corrected control sequence. S4.5: The corrected control sequence is sent to the fermentation execution unit. After execution, the new trajectory is evaluated again through the digital twin and its proximity to the target activity is compared. If the target activity of the corrected trajectory is stably within the target window for multiple consecutive iterations, the control sequence is considered to have converged. If there is still a deviation, the iterative update of the control sequence continues.

6. The method for preparing a fermented food for high-altitude elderly people that targets the gut-lung axis, as described in claim 1, is characterized in that... The specific steps for establishing the gut-lung axis health status fingerprint database of the elderly population in plateau regions, as described in step VIII, and simultaneously collecting multimodal feature signals from fermentation broth and selecting the corresponding target active ingredient profiles are as follows: S5.1: The gut microbiota metagenomic characteristics, serum metabolomics characteristics, and lung function indicators of each plateau elderly subject sample collected and preprocessed are uniformly encoded and health status fingerprint entries are formed according to the individual dimension to establish a health status fingerprint map library. Each fingerprint entry corresponds to the basic health status of a subject. S5.2: In the later stage of fermentation, the fermentation broth is placed in a fixed flow state, and multimodal characteristic signals, including near-infrared spectral signals, electronic nose response signals and electronic tongue potential signals, are collected simultaneously from the fermentation broth. The near-infrared, electronic nose and electronic tongue signals are then aligned according to a unified timestamp. Baseline drift correction and amplitude normalization are then performed on each channel to unify the data from different sources to the same numerical scale. If any channel shows a sudden change within three consecutive sampling periods, that channel is treated as an abnormal channel and local interpolation is performed for repair. S5.3: After preprocessing, the modal signals are spliced ​​into a one-dimensional input sequence in the order of spectrum-odor-potential, and then fed into a one-dimensional convolutional neural network. The sequence is processed layer by layer through convolutional layer, activation layer, pooling layer and fully connected layer to extract the feature tensor of the fermentation broth in the multimodal state, and then it is matched with each fingerprint in the healthy state fingerprint database for similarity. S5.4: Calculate the similarity value between the feature tensor of each fermentation broth and each healthy fingerprint using cosine similarity, and then correct the matching results by combining individual adaptation weights to generate a comprehensive matching score. Sort each healthy fingerprint from high to low according to the comprehensive matching score, and select the one with the highest comprehensive matching score as the target active ingredient spectrum template corresponding to the current fermentation broth. Then, extract the corresponding ideal active ingredient combination range according to the template and output it as the target spectrum for subsequent updates.

Images