Bone soup vacuum concentration flavor stability regulation method and system

Abstract

The present application relates to the technical field of food intelligent processing, and discloses a bone soup vacuum concentration flavor stability regulation method and system, which comprises the following steps: synchronously collecting macroscopic thermodynamic parameters, material state parameters, gas phase response signals and electrical load parameters; generating an apparent viscosity index from the electrical load parameters, the pressure difference between the inlet and outlet of the heat exchanger and the no-load torque model; inputting the above parameters into a neural network model to calculate and output a predicted volatile partial pressure; generating a volatile risk index according to the ratio of the predicted volatile partial pressure to the vacuum pressure in the vacuum concentration device and the time differential data thereof; and issuing an intervention instruction matrix to an execution component when the index is greater than or equal to a threshold value. The present application can reflect the change of the fluid state of the system without a rheometer, convert the gas-liquid phase equilibrium into a quantitative index, realize the early prediction of the volatilization tendency of the target volatile flavor components, reduce the flavor fluctuation between batches by using a control intervention mechanism, and improve the quality consistency.

Description

Technical Field

[0001] This invention relates to the field of intelligent food processing technology, specifically to a method and system for controlling the flavor stability of bone broth through vacuum concentration. Background Technology

[0002] Vacuum concentration is a crucial step in the industrial production of liquid foods such as bone broth. Its main purpose is to remove excess water while preserving as much of the original flavor compounds as possible. In actual production, when bone broth is heated under vacuum, its volatile flavor components are easily lost along with the secondary steam generated by water evaporation. Maintaining the stability of the concentrated liquid's flavor during this process remains a significant technical challenge in the industry.

[0003] To improve the control precision of the concentration process, the industry typically employs automated control strategies to maintain production conditions. For example, document CN119847024A discloses an intelligent control method for the concentration process of traditional Chinese medicine. This method, targeting the vacuum concentration process, constructs a temperature control model and a soft measurement model for the density of the concentrate, and uses fuzzy self-tuning and expert control algorithms to collect and regulate the vacuum degree and heating temperature of the evaporation chamber in real time. This system, by detecting the density of the concentrate online and automatically determining the concentration endpoint by a PLC program, solves to some extent the problems of loss of heat-sensitive components and poor batch quality consistency. However, the control objectives of this type of method mainly focus on vacuum degree, heating temperature, density, or concentration endpoint determination, and do not incorporate the online response signals of volatile flavor components, material rheological state characterization quantities, and gas-liquid phase volatilization driving characteristics into the intelligent decision-making control of the concentration process.

[0004] While the aforementioned existing technologies have achieved a certain degree of automation by controlling temperature, vacuum level, and material density, their application to the vacuum concentration process of bone broth still has significant limitations. Bone broth contains a large amount of protein, fat, and colloidal components. As the concentration process progresses, the rheological properties of the material undergo nonlinear changes, especially with a significant increase in apparent viscosity in the later stages of concentration. This directly alters the mass and heat transfer resistance within the system. Existing control methods mainly rely on conventional thermodynamic parameters for regulation, lacking online sensing capabilities for such complex fluid state changes. This results in an inability to accurately assess the dynamic shifts in the gas-liquid phase equilibrium within the material. Because mass exchange at the gas-liquid interface is drastically affected by viscosity, temperature and pressure data alone cannot accurately reflect the current volatility tendency of flavor substances, making it difficult to predict the loss of target volatile flavor components in real time. Furthermore, existing conventional closed-loop control systems primarily aim to maintain set constant operating parameters, lacking real-time dynamic intervention mechanisms based on flavor volatility risks. When minor fluctuations in operating conditions or sudden changes in physical properties lead to a significant loss of flavor, the control system is unable to provide coordinated compensation and intervention. This results in differences in flavor characteristics between batches of concentrated bone broth produced under the same process, making it difficult to meet the product's requirements for flavor stability control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for controlling the flavor stability of bone broth during vacuum concentration. This method solves the problem that the risk of flavor substance loss during bone broth concentration is difficult to predict in real time, and the lack of a dynamic intervention mechanism leads to flavor instability between batches of concentrated products.

[0006] To achieve the above objectives, the first aspect of the present invention provides a method for controlling the flavor stability of bone broth through vacuum concentration, comprising: Simultaneously collect macroscopic thermodynamic parameters, material state parameters, gas phase response signals of target volatile flavor components, and bottom electrical load parameters of the circulating pump motor within the vacuum concentrator. The macroscopic thermodynamic parameters include the pressure difference between the inlet and outlet of the heat exchanger and the vacuum pressure within the vacuum concentrator. The apparent viscosity index is generated from the underlying electrical load parameters, the inlet and outlet pressure difference of the heat exchanger, and the reference no-load torque model. The macroscopic thermodynamic parameters, the material state parameters, the gas phase response signal, and the apparent viscosity index are input into a preset neural network model, which outputs the liquid phase equivalent flavor state quantity and the equivalent activity correction coefficient, and calculates the predicted volatile partial pressure of each target volatile flavor component. A corresponding volatility risk index is generated based on the ratio of the predicted volatile partial pressure to the vacuum pressure in the vacuum concentration device and the time derivative data of the ratio; When the volatility risk index is greater than or equal to the volatility risk control threshold, an intervention instruction matrix is ​​issued to the underlying execution component.

[0007] Furthermore, the synchronous acquisition includes: During the main polling cycle of the distributed control system, multi-source acquisition values ​​are acquired synchronously. These multi-source acquisition values ​​include: material temperature, vacuum pressure in the vacuum concentrator, and pressure difference between the inlet and outlet of the heat exchanger as macroscopic thermodynamic parameters; material flow rate and soluble solids content as material state parameters; gas phase response signal of the target volatile flavor component as a gas phase detection parameter; and circulating pump motor stator current, output torque, and circulating pump operating frequency as underlying electrical load parameters. Each collected value is timestamped, and a zero-order hold algorithm is used for individual sensor signal channels with low reporting frequency to form a multi-dimensional initial input data matrix.

[0008] Furthermore, the zero-order hold algorithm logic is used to maintain the valid sampled value of the previous cycle when individual sensor signal channels with low reporting frequencies have not been updated.

[0009] Furthermore, the gas phase response signal is acquired by a metal oxide semiconductor gas sensor array or an industrial-grade online ion mobility spectrometer located in the secondary steam emission path or the gas phase space region at the top of the separation chamber. The raw electrical signal output by the online flavor detection unit is preprocessed, features are extracted, and normalized to obtain a dimensionless gas phase response signal in the range of 0 to 1.

[0010] Furthermore, the generation of the apparent viscosity index includes: During the stage of benchmark calibration operation with atmospheric pressure clean water, the benchmark output torque of the circulating pump under different operating frequencies and different material flow rates is collected, and the benchmark no-load torque model is established by multivariate polynomial surface fitting. During operation, the net fluid load torque is obtained by subtracting the reference torque given by the reference no-load torque model from the output torque. Then, by combining the net fluid load torque, the operating frequency, the material flow rate, the pressure difference between the inlet and outlet of the heat exchanger, the soluble solids content, and the electromechanical conversion efficiency correction coefficient, a dimensionless apparent viscosity index is generated.

[0011] Furthermore, when the output torque is less than the reference torque, the net fluid load torque is set to zero; When the material flow rate is lower than the flow protection lower limit threshold, the apparent viscosity index is assigned a preset standard constant of 1.0 in the first sampling period, and the effective calculated value of the previous sampling period is retained in subsequent sampling periods.

[0012] Furthermore, the neural network model is a feedforward multilayer perceptron. The input layer receives a 10-dimensional feature input vector formed by concatenating a 9-dimensional initial input column vector consisting of the material temperature, the vacuum pressure in the vacuum concentration device, the pressure difference between the inlet and outlet of the heat exchanger, the material flow rate, the soluble solids content, the gas phase response signal, the stator current of the circulating pump motor, the output torque, and the operating frequency of the circulating pump with the apparent viscosity index. The output layer outputs the liquid phase equivalent flavor state quantity and the equivalent activity correction coefficient. The neural network model is trained under supervision using the true values ​​of the liquid phase equivalent flavor state quantities and the true values ​​of the equivalent activity correction coefficients derived from the inversion and derivation of offline gas-liquid phase equilibrium closed distillation experimental data.

[0013] Furthermore, the predicted volatile partial pressure is calculated from the liquid-phase equivalent flavor state quantity, the equivalent activity correction coefficient, the material temperature, and the material property constants of the target volatile flavor components in the Antoine equation; The generation of the volatility risk index includes calculating the volatility drive ratio between the predicted volatile partial pressure and the vacuum pressure in the vacuum concentration device, and fusing the current value of the volatility drive ratio with the time differential features extracted by the first-order backward differential discretization calculation to obtain the volatility risk index of the target volatile flavor component.

[0014] Furthermore, the intervention command matrix includes vacuum pressure bias, heat exchanger steam valve opening bias, and circulating pump operating frequency bias, and the underlying execution components include the proportional exhaust valve, steam regulating valve, and frequency converter corresponding to the circulating pump control unit of the vacuum system. The vacuum pressure bias is sent to the proportional exhaust valve of the vacuum system in the form of a step signal. The opening bias of the heat exchanger steam valve is fed into a first-order inertial filter with a preset time constant when it is output to the steam regulating valve. The operating frequency bias of the circulating pump is sent to the frequency converter corresponding to the circulating pump control unit.

[0015] Furthermore, during the execution of the intervention command matrix, the intelligent decision-making module corrects the target temperature setpoint of the heating system in the same direction based on the boiling point rise increment corresponding to the vacuum pressure bias. The intelligent decision-making module extracts the maximum value among the volatility risk indices corresponding to multiple target volatile flavor components as the global volatility risk representative quantity. When the global volatility risk representative quantity falls back and stabilizes below the volatility risk control threshold, the execution control module decays and cancels the vacuum pressure bias, heat exchanger steam valve opening bias, and circulating pump operating frequency bias at a set rate through the slope generator.

[0016] After each batch of concentrated bone broth is completed, the system obtains the true flavor retention rate of the current batch obtained from offline measurement, and calculates the cumulative prediction deviation by combining the predicted retention rate of the current batch based on the predicted volatile partial pressure and batch operation data. The intelligent decision module updates the node reference bias corresponding to the equivalent activity correction coefficient in the output layer of the neural network model according to the cumulative prediction deviation, so as to achieve cross-batch adaptive correction.

[0017] A second aspect of the present invention provides a flavor stability control system for vacuum concentration of bone broth, comprising a vacuum concentration device for evaporating and concentrating bone broth materials in a vacuum environment, and a distributed control system communicatively connected to the vacuum concentration device, wherein the distributed control system comprises an online detection module, a data acquisition module, a data processing module, a flavor evaluation module, an intelligent decision-making module, and an execution control module. The online detection module is used to acquire the gas-phase response signal of the target volatile flavor component; The data acquisition module is used to acquire macroscopic thermodynamic parameters, material state parameters, and the underlying electrical load parameters of the circulating pump motor. The macroscopic thermodynamic parameters include the vacuum pressure in the vacuum concentration unit and the pressure difference between the inlet and outlet of the heat exchanger. The data processing module is used to generate the apparent viscosity index; The flavor evaluation module is used to output liquid-phase equivalent flavor state quantities and equivalent activity correction coefficients; The intelligent decision-making module is used to calculate the predicted volatile partial pressure of the target volatile flavor components and generate a volatility risk index and an intervention instruction matrix; The execution control module includes a vacuum pump control unit, a steam valve control unit, and a circulation pump control unit. The vacuum pump control unit is connected to the vacuum pump, the steam valve control unit is connected to the steam regulating valve, and the circulation pump control unit is connected to the circulation pump. The execution control module is used to send the vacuum pressure bias to the proportional exhaust valve of the vacuum system in the form of a step signal.

[0018] Furthermore, the execution control module is also used to parse the intervention command matrix and send the vacuum pressure bias, heat exchanger steam valve opening bias, and circulating pump operating frequency bias in the intervention command matrix to the corresponding underlying execution components.

[0019] This invention provides a method and system for controlling the flavor stability of bone broth through vacuum concentration, which has the following beneficial effects: This invention collects the underlying electrical load parameters and the pressure difference between the inlet and outlet of the heat exchanger, and generates an apparent viscosity index by combining it with a benchmark no-load torque model. This allows for real-time reflection of the fluid state changes of bone broth during vacuum concentration without altering the original pipeline structure or relying on an online rheometer. It overcomes the technical limitation that the viscosity of high-concentration materials is difficult to measure online in the later stages of concentration, and provides reliable physical property data support for subsequent flavor state evaluation.

[0020] This invention inputs macroscopic thermodynamic parameters, gas phase response signals, and apparent viscosity index into a pre-set neural network model to calculate and predict volatile partial pressures and generate a volatile risk index by combining time differential data. This can transform the complex nonlinear gas-liquid phase equilibrium relationship in the concentration system into a specific quantitative indicator, enabling early prediction of the volatility tendency of target volatile flavor components. It solves the problem of the lag in traditional processes where flavor loss is difficult to detect in a timely manner due to relying solely on experience to control temperature and pressure.

[0021] This invention establishes a low-level control intervention mechanism based on the volatility risk index. When the risk exceeds the limit, the system suppresses the risk by issuing an intervention command matrix to coordinate the adjustment of vacuum pressure, valve opening and pump operating frequency. After each concentration batch, the system updates the node baseline bias of the neural network model using the measured true retention rate, enabling the system to have the ability to adaptively correct across batches. This effectively reduces flavor fluctuations between multiple batches of production and improves the consistency of concentrated product quality. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the architecture of a bone broth vacuum concentration flavor stability control system according to an embodiment of the present invention. Figure 2 This is a flowchart of the method for controlling the flavor stability of bone broth through vacuum concentration according to an embodiment of the present invention; Figure 3 This is a flowchart of the data processing and flavor state prediction algorithm according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating the multi-dimensional intervention and coordinated action control logic of an embodiment of the present invention. Figure 5 This is a diagram of the physical field asymmetric decoupling control strategy according to an embodiment of the present invention; Figure 6 This is a dynamic evolution diagram of risk prediction and control intervention in this invention; Figure 7 This is a comparison chart showing the retention effects of the target volatile flavor components in this invention. Figure 8 This is a schematic diagram of the neural network model structure according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, the present invention provides a flavor stability control system for vacuum concentration of bone broth, including a vacuum concentration device and a distributed control system.

[0025] A vacuum concentrator is used to heat, evaporate, and concentrate bone broth materials under negative pressure. It may include a heat exchanger, a separation chamber, circulation pipelines, a circulation pump, a steam regulating valve, a secondary steam discharge path, and a vacuum interface. The separation chamber is the cavity in the vacuum concentrator used to achieve gas-liquid separation and allow secondary steam to escape.

[0026] The vacuum system is used to maintain a negative pressure environment within the vacuum concentrator. It may include a vacuum pump, vacuum pipelines, pressure detection components, and a proportional exhaust valve. The vacuum pump is the actuator in the vacuum system used to pump out gas and establish negative pressure. The proportional exhaust valve is used to adjust the intake or exhaust state of the vacuum system to change the vacuum pressure within the vacuum concentrator.

[0027] The distributed control system is connected to the vacuum concentration unit and its underlying actuators at the communication level. The distributed control system consists of multiple logic control modules, specifically including an online detection module, a data acquisition module, a data processing module, a flavor evaluation module, an intelligent decision-making module, and an execution control module.

[0028] The system comprises the following modules: an online detection module for acquiring the gas-phase response signal of the target volatile flavor components; a data acquisition module for acquiring macroscopic thermodynamic parameters, material state parameters, and underlying electrical load parameters of the circulating pump motor within the vacuum concentration unit; a data processing module for performing data alignment, preprocessing, and generating the apparent viscosity index; a flavor evaluation module for calling a neural network model to output liquid-phase equivalent flavor state quantities and equivalent activity correction coefficients; an intelligent decision-making module for calculating and predicting volatile partial pressure, volatile risk index, and intervention command matrix; and an execution control module for parsing the intervention command matrix and driving the corresponding underlying execution components.

[0029] The execution control module further includes a vacuum pump control unit, a steam valve control unit, and a circulation pump control unit. The vacuum pump control unit is connected to the vacuum pump, the steam valve control unit is connected to the steam regulating valve, and the circulation pump control unit is connected to the circulation pump. The execution control module is also connected to the proportional exhaust valve of the vacuum system, used to send the vacuum pressure bias to the proportional exhaust valve of the vacuum system in the form of a step signal. Each module interacts with the underlying hardware via an industrial fieldbus, forming a closed-loop control network that links the physical execution layer and the algorithm control layer.

[0030] like Figure 2 As shown, this embodiment of the invention provides a method for controlling the flavor stability of bone broth through vacuum concentration, comprising the following steps: S10, the data acquisition module and the online detection module synchronously acquire multi-source parameters during the main polling cycle of the distributed control system. Specifically, the data acquisition module acquires macroscopic thermodynamic parameters, material state parameters, and the underlying electrical load parameters of the circulating pump motor within the vacuum concentration unit; the online detection module acquires the gas-phase response signals of the target volatile flavor components. The data processing module timestamps and aligns the aforementioned multi-source parameters in the time domain to form a multi-dimensional initial input data matrix, which is then sent to the flavor evaluation module and the intelligent decision-making module. S20, the data processing module extracts the actual output torque and circulating pump motor stator current from the underlying electrical load parameters, and compares them with the heat exchanger inlet and outlet pressure difference under the current fluid operating conditions and the benchmark no-load torque model. The data processing module filters out mechanical loss terms and converts the electrical state data into a dimensionless apparent viscosity index to characterize the mass transfer resistance of polymer colloids. S30, the flavor evaluation module inputs macroscopic thermodynamic parameters, material state parameters, gas phase response signals, and apparent viscosity index into a pre-set neural network model. The neural network model outputs liquid-phase equivalent flavor state quantities and equivalent activity correction coefficients. The intelligent decision module receives the liquid-phase equivalent flavor state quantities and equivalent activity correction coefficients, and calculates the predicted volatile partial pressure of the target volatile flavor component by combining the material temperature and the Antoine constant corresponding to the target volatile flavor component. S40, the intelligent decision-making module calculates the ratio between the predicted volatile partial pressure and the current vacuum pressure in the vacuum concentrator, and extracts the time derivative data of the ratio to generate a dynamic volatile risk index. When the intelligent decision-making module determines that the volatile risk index has reached the set control threshold, it issues an intervention command to the execution control module; S50: After receiving the intervention command, the execution control module outputs an intervention command matrix containing multi-dimensional feedforward bias, instructing the circulating pump control unit to increase the operating frequency to increase the material circulation flow rate. Simultaneously, the execution control module sends the vacuum pressure bias as a step signal to the proportional exhaust valve of the vacuum system to rapidly increase the operating pressure, and instructs the steam valve control unit to incorporate an inertial filter element on top of the superimposed bias to perform a gradual descent action. This constitutes an asymmetric physical field decoupling and collaborative intervention mechanism, avoiding flavor stripping caused by localized stress impacts and intensified boiling. S60: Within the steady-state time window, the data processing module calculates the predicted flavor retention rate of the current batch based on batch operation data and receives the actual flavor retention rate of the current batch obtained from offline detection. The intelligent decision-making module calculates the prediction deviation based on the predicted flavor retention rate and the actual flavor retention rate, and uses the prediction deviation as an error compensation signal. Through the feedback link, it updates the node reference bias corresponding to the equivalent activity correction coefficient in the neural network model, completing the closed-loop convergence of the control parameters.

[0031] During the vacuum concentration of bone broth, the broth matrix is ​​rich in collagen, fat, and flavor compounds. As water evaporates, the system exhibits a complex physical state characterized by high viscosity, non-Newtonian fluid properties, and vigorous boiling in both the gas and liquid phases. The evaporation rate of microscopic flavor components depends not only on macroscopic temperature and pressure boundaries but also on the flow resistance at the mass transfer interface and the binding force of the macromolecular matrix. To achieve precise digital control of this multi-physics coupled evolution process, it is necessary to simultaneously and complementaryly acquire multi-source heterogeneous data on macroscopic thermodynamics, microscopic flavor response, and underlying driving electrical load at the same sampling time segment within the control cycle. This provides time-domain correlated underlying physical properties for subsequent equivalent viscosity analysis and flavor loss risk prediction. In this embodiment, the online synchronous acquisition and preprocessing of multi-source heterogeneous data specifically includes the following steps: S101, the data acquisition module acquires the macroscopic thermodynamic parameters and material state parameters within the vacuum concentration unit.

[0032] In this embodiment, the data acquisition module obtains process parameters through industrial transmitters deployed on the vacuum concentration unit and auxiliary pipelines. Specifically, the data acquisition module obtains the material temperature in the heat exchanger and separation chamber through temperature transmitters. The vacuum pressure inside the vacuum concentrator is obtained through an absolute pressure transmitter. The differential pressure difference between the inlet and outlet of the heat exchanger is obtained through a differential pressure transmitter during the fluid flow stages. The material flow rate output by the circulating pump is obtained through an electromagnetic flow meter. The refractive index of the bone broth fluid is obtained in real time by an online refractometer installed at the detection port of the circulation pipeline, and the corresponding soluble solids content is output by the built-in conversion logic. Regarding the internal circuit structure, measurement principle, and basic wiring installation method of the aforementioned industrial transmitters and flow meters, those skilled in the art can make conventional selections and configurations based on the hygiene requirements of the actual production workshop. These are well-known technologies in the field and will not be elaborated further here.

[0033] S102, the online detection module acquires the gas phase response signal of the target volatile flavor component.

[0034] As an optional implementation, the online detection module includes an online flavor detection unit. This unit is installed in the secondary steam emission path of the vacuum concentrator or in the top gas phase space of the separation chamber, and is connected to the gas phase space of the vacuum concentrator. It is used to capture volatile substances escaping from the bone broth under vacuum boiling conditions. The specific hardware implementation of the online flavor detection unit is characterized by a metal oxide semiconductor gas sensor array or an industrial-grade online ion mobility spectrometer. When the online flavor detection unit comes into contact with the target volatile flavor component, it converts the concentration change of the chemical component in the gas phase into a corresponding raw continuous electrical signal. The data processing module receives the raw continuous electrical signal and performs baseline subtraction, smoothing filtering, and feature extraction to obtain gas phase detection features such as peak area and characteristic response intensity. Subsequently, the gas phase detection features are normalized to obtain the gas phase response signal. Thus, the gas phase response signal is a dimensionless signal formed by preprocessing, feature extraction, and normalization of the raw continuous electrical signal output by the online flavor detection unit. In order to smooth out the range and baseline differences between different sensor channels, the above normalization process adopts the max-min normalization algorithm to map the original electrical signal characteristic intensity proportionally to the dimensionless interval between 0 and 1, so as to ensure the numerical regularity of the input data.

[0035] S103, the data acquisition module reads the underlying electrical load parameters from the inverter communication bus corresponding to the circulating pump control unit.

[0036] Compared to conventional methods that involve adding contact rheometers or external mechanical torque sensors inside high-temperature boiling fluid pipelines, this system employs a non-invasive, low-level data extraction mechanism. The circulating pump control unit uses an industrial frequency converter with a built-in sensorless vector control algorithm to drive the circulating pump motor. During operation, the industrial frequency converter calculates the motor's flux linkage and current components in real time based on its internal control model. The data acquisition module periodically sends read commands to the control register address inside the frequency converter via an industrial fieldbus, directly extracting the circulating pump motor stator current, avoiding the noise of external mechanical sensor measurements. Output torque and the operating frequency of the circulating pump Parameters. Industrial fieldbuses specifically employ the Profinet bus protocol or Modbus communication protocol stack for signal interaction. This non-intrusive acquisition method effectively avoids direct corrosion of mechanical sensing components by high-temperature pipelines and high-viscosity deposits, reducing the risk of erosion and mechanical damage to physical sensors.

[0037] S104, the data acquisition module timestamps all acquired data in the time domain to form an initial input data matrix.

[0038] Due to delays caused by physical hardware and communication links in the data reporting cycle between the temperature transmitter, online flavor detection unit, and industrial frequency converter communication bus, the data acquisition module is equipped with a synchronization clock unit. A fixed polling cycle of the distributed control system is used as the reference sampling cycle, with a value range of 50ms to 200ms. At the rising edge of the clock in each reference sampling cycle, the data acquisition module synchronously receives the material temperature... Vacuum pressure Differential pressure between heat exchanger inlet and outlet Material flow soluble solids content Target volatile flavor components gas phase response signal and the stator current of the circulating pump motor output by the frequency converter Output torque Operating frequency of circulating pump Add a timestamp.

[0039] For individual sensor signal channels with low reporting frequencies, the data acquisition module internally retrieves the zero-order hold algorithm logic to maintain the valid sampled value from the previous cycle. To eliminate the algorithm control dead zone caused by the lack of historical values ​​during system startup or initial power-on, if the zero-order hold determines that it cannot retrieve the historical sampled value from the previous cycle in the first sampling cycle, the system will automatically call the pre-stored atmospheric pressure clean water offline calibration parameters as the initial default fill value until the first valid real-time reporting signal is received. All data that are timestamped at the same time are combined and compiled into a one-dimensional digital vector, and are sequentially stored in the memory buffer area of ​​the intelligent decision module as time progresses, aggregating to form a multi-dimensional initial input data matrix for characterizing the transient physical characteristics of the vacuum concentration system. Specifically, corresponding to the project requirements of the subsequent neural network model input layer, at any sampling time... The constructed one-dimensional input vector matrix is ​​as follows: ; In the formula, Indicates the sampling time The initial 9-dimensional input column vector is used, and the meanings of the remaining components are completely consistent with the physical definitions in the aforementioned sub-steps. The structure, dimensions, and physical meaning of this vector remain constant, thus providing a structured data flow guarantee for the subsequent extraction of nonlinear mass transfer features from the hidden layers of the network model.

[0040] After completing the online synchronous acquisition and preprocessing of multi-source heterogeneous data, the acquired data matrix provides a foundation for further analysis of the material state. Because bone broth is rich in collagen and fat during the later stages of concentration, the matrix exhibits shear thinning and viscoelastic characteristics characteristic of non-Newtonian fluids. Under these physical properties, the increase in apparent viscosity increases the resistance to the diffusion of flavor compounds from the liquid phase to the gas phase, directly constrained by the flow resistance at the mass transfer interface and the binding force of the macromolecular matrix, thus slowing down the volatilization rate of flavor components. However, contact rheometers are prone to measurement failures or high maintenance costs under the high-temperature boiling, scaling, and gas-liquid two-phase interweaving conditions inside vacuum concentration devices.

[0041] To overcome this hardware measurement bottleneck, this embodiment introduces an apparent viscosity index analysis stage based on underlying electrical load into the bone broth vacuum concentration flavor stability control method. This stage uses the electrical parameters of the material circulation pump motor to inversely calculate the fluid's apparent viscosity characteristics. As a preferred implementation, the apparent viscosity index analysis based on underlying electrical load specifically includes the following steps: S201, the data processing module constructs and calls the benchmark no-load torque model to establish the benchmark for mechanical and pipeline inherent losses under standard medium.

[0042] In this embodiment, before the vacuum concentration unit is officially put into operation or during the online cleaning cycle, the distributed control system introduces atmospheric pressure clean water for benchmark calibration under preset standard temperature conditions. During this period, the frequency converter driving the material circulation pump operates at different frequencies. With the corresponding material flow rate Under combined operation, the data acquisition module records the reference output torque data fed back by the frequency converter. The data processing module uses a multivariate polynomial surface fitting algorithm to perform mathematical statistical analysis on the calibration point data to construct the output reference torque. The mathematical expression for the reference torque. The specific calculation formula for characterizing the rotor inertia of the circulating pump motor, bearing mechanical friction loss, and the inherent resistance torque generated by the fluid on the pipeline under clean water conditions is as follows: ; In the formula, This represents the reference torque, in N·m. Indicates the operating frequency, in Hz; This indicates the material flow rate, in meters (m). 3 / h; , , These are the regression calibration coefficients. This refers to the calibration constant term. In this embodiment, the regression calibration coefficients and calibration constant term are obtained by fitting more than 20 sets of standard calibration data points using the least squares method, wherein the calibration constant term... The physical meaning corresponds to the pure mechanical static friction torque of the motor, and its value range is set from 2.5 N·m to 4.5 N·m. The establishment of this benchmark model provides data reference for subsequent elimination of inherent mechanical background interference of the system.

[0043] S202, the data processing module calculates the net fluid load torque under the current operating conditions in real time and performs dead zone boundary filtering.

[0044] As the vacuum concentration process of bone broth continues, the apparent viscosity characteristics of the material change due to the evaporation of water. This change occurs at any sampling point during the actual concentration process. The data processing module will extract the output torque from the inverter communication bus. and the corresponding operating frequency Material flow Substituting into the above reference no-load torque model, the current reference torque is calculated. By performing a subtraction operation, the output torque is obtained. Subtracting the reference torque This isolates the net fluid load torque. To prevent data inversion during initial system operation, hardware switching transients, or when sensors are subjected to minor electromagnetic interference, the data processing module is equipped with dynamic boundary filtering logic. If the currently calculated output torque... Less than the reference torque The system forces the result of net fluid load torque to be 0, which effectively avoids the algorithm control dead zone caused by negative torque participating in subsequent volatile partial pressure calculation and volatile risk index calculation, and ensures the integrity of the underlying control flow.

[0045] S203, the data processing module couples multidimensional physical parameters with load torque to dynamically generate a dimensionless apparent viscosity index.

[0046] To comprehensively characterize the matrix binding force and micro-phase mass transfer resistance exerted by the polymer colloidal matrix on microscopic flavor substances, this embodiment employs correlation modeling based on fluid mechanics principles. From a rheological perspective, the apparent viscosity of a non-Newtonian fluid is directly proportional to shear stress and inversely proportional to shear rate. In the pipeline flow field of this system, the net fluid load torque, after deducting inherent losses, macroscopically corresponds to the total energy consumption driving the fluid shear flow. Its ratio to the material flow rate and operating frequency effectively maps the fluid's fundamental shear stress state. The pressure difference between the heat exchanger inlet and outlet reflects the fluid's resistance loss along the flow path, which exhibits a nonlinear increasing trend with increasing material concentration. By introducing soluble solids content to correct for the matrix concentration effect in the pressure difference term, the interference of salting-out effect and solid precipitation on the flow regime can be eliminated.

[0047] Based on the above principles, this embodiment constructs an apparent viscosity index that integrates dynamic transmission load and pipeline shear rheological characteristics, serving as a dimensionless state quantity characterizing the apparent viscosity properties of materials. Material flow soluble solids content The net fluid load torque calculated in step S202 is coupled across physical domains to dynamically output the apparent viscosity index. Apparent viscosity index The specific analytical calculation formula is as follows: ; In the formula, It represents the apparent viscosity index, which is a dimensionless comprehensive characterization of physical properties; This indicates the output torque, expressed in N·m. This represents the reference torque, in N·m. Indicates the operating frequency, in Hz; This indicates the pressure difference between the inlet and outlet of the heat exchanger, expressed in kPa. This indicates the material flow rate, in meters (m). 3 / h; The soluble solids content is expressed as a percentage absolute value in calculations; This represents the electromechanical conversion efficiency correction factor; This represents the feature weighting coefficient.

[0048] In this embodiment, the electromechanical conversion efficiency correction coefficient As a variable related to motor heating and reactive power, its value range is set from 0.82 to 0.94, and it is determined by online efficiency lookup table by reading the real-time output power of the frequency converter; characteristic weighting coefficient The value of the nonlinear contribution of the collagen network structure in bone broth to the flow shear resistance is set to a range of 0.12 to 0.28 and is pre-calibrated by laboratory historical rheological sample test data.

[0049] To avoid material flow issues caused by pump / valve switching or abnormal blockages in industrial settings. When the denominator term suddenly drops to 0, a mathematical overflow and system crash error occur. The data processing module has internal traffic safety threshold blocking logic. The system sets a lower threshold for traffic protection. 0.2m 3 / h, once real-time material flow is detected Below this traffic protection threshold The data processing module will immediately pause the division operation of the above formula.

[0050] At this point, in order to eliminate the dead zone caused by the lack of historical reference values ​​due to the flow protection being triggered in the first sampling cycle, the data processing module executes the following branch decision: If the current period is the first sampling cycle of the concentration process, then the current apparent viscosity index is directly set. The value is assigned to a preset standard constant of 1.0 to represent the baseline apparent viscosity of water under normal pressure. If the current period is not the first sampling cycle, the data processing module will use the apparent viscosity index of the current cycle. The calculated value remains valid from the previous sampling period. The generated apparent viscosity index... On a macroscopic level, it reflects the degree of obstruction of material flow in the pipeline, while on a microscopic level, it characterizes the mass transfer resistance barrier of the broth matrix to the escape of flavor molecules. Thus, it successfully transforms rheological parameters that are difficult to measure directly into observable controllable variables composed of underlying electrical and hydraulic parameters.

[0051] In the vacuum concentration process of bone broth, the material system contains complex proteins, fats, and carbohydrates, causing the transfer behavior of volatile flavor compounds at the gas-liquid interface to deviate significantly from the ideal solution state. Traditional analytical equations are insufficient to directly describe the component distribution patterns in such complex multiphase fluids. Therefore, this embodiment employs a method combining a gas-liquid phase equilibrium physical model with a neural network to map easily measurable macroscopic engineering parameters into microscopic mass transfer driving forces. (See also...) Figure 3 The flavor state estimation and prediction partial pressure calculation based on neural networks includes the following steps: S301, the flavor evaluation module constructs a multilayer perceptron neural network model and completes the integration of input feature vectors.

[0052] See Figure 8In this embodiment, the flavor evaluation module uses a feedforward multilayer perceptron as the basic neural network model. The neural network model includes a 10-node input layer, a first fully connected hidden layer, a second fully connected hidden layer, and a 2-node output layer. The first fully connected hidden layer contains 64 neurons, and the second fully connected hidden layer contains 32 neurons. Both hidden layers use the ReLU activation function to extract the nonlinear coupling features between multiple physical quantities. The output layer uses linear output, with two output nodes representing the liquid-phase equivalent flavor state quantity and the equivalent activity correction coefficient, respectively. The flavor evaluation module uses the 9-dimensional initial input column vector obtained in the previous steps (… Figure 8 (Simplified expression from x1 to x9) and apparent viscosity index Dimensional concatenation is performed to form a 10-dimensional feature input vector. During the model building and preprocessing stages, the flavor evaluation module uses known-concentration bone broth samples accumulated in the laboratory as the training set. To ensure consistency between the model output and the training objective, the supervision labels for the training set are the true values ​​of the liquid-phase equivalent flavor state quantities and the true values ​​of the equivalent activity correction coefficients, derived from offline gas-liquid phase equilibrium closed distillation experimental data. Specifically, the system uses a gas chromatograph to offline measure the absolute concentrations of flavor components in the gas and liquid phases at different concentration stages, and substitutes these concentrations into the basic equation of Raoult's law for inverse calculation. The basic equation of Raoult's law can be expressed as: ; in, Target volatile flavor components Partial pressure in the gas phase Target volatile flavor components mole fraction in the liquid phase Target volatile flavor components The saturated vapor pressure at the current material temperature.

[0053] In this embodiment, the system uses the concentrations of gas-liquid two-phase flavor components measured by an offline gas-liquid phase equilibrium closed distillation experiment, combined with the above-mentioned basic equations to invert and obtain the true values ​​of the liquid phase equivalent flavor state quantities and the true values ​​of the equivalent activity correction coefficients, and uses them as supervision labels for the neural network model.

[0054] S302, the flavor evaluation module directly outputs equivalent gas-liquid phase parameters through hidden layer feature mapping.

[0055] After the 10-dimensional feature input vector enters the neural network, it undergoes nonlinear activation and weighted calculation of the hidden layer nodes, and is directly output by the output layer along with the target volatile flavor components. The two corresponding characterization parameters are: liquid phase equivalent flavor state quantity. With equivalent activity correction factor Within the realm of complex fluid thermodynamics, the equivalent flavor state quantity in the liquid phase The effective mole fraction used to characterize the actual unbound moles that can freely participate in gas-liquid phase partitioning, while the equivalent activity correction factor... This reflects the physical network constraint and chemical bonding resistance of high-viscosity macromolecular colloids on flavor molecules.

[0056] S303, the intelligent decision-making module predicts volatile partial pressure based on the modified Raoult law.

[0057] Because bone broth concentrate is a typical high-concentration, non-ideal solution, the vapor pressure of flavor components within it cannot directly follow a simple linear distribution according to Raoult's law. To address this issue, the intelligent decision-making module substitutes the state variables output by the model into the built-in gas-liquid equilibrium model. The intelligent decision-making module also considers the current material temperature. The theoretical saturated vapor pressure of the pure component is derived based on the Antoine equation, and the equivalent activity correction factor is directly applied. Deviation correction is performed. The specific calculation formula for predicting volatile partial pressure is as follows: ; In the formula, Indicates the target volatile flavor components The predicted volatile partial pressure is given in kPa. This represents the equivalent activity correction factor, which is a dimensionless empirical correction term. It represents the equivalent flavor state quantity in the liquid phase and is used in the calculation in the form of equivalent mole fraction without dimension; , and All represent the corresponding target volatile flavor components. The specific values ​​of the material property constants in the Antoine equations are determined by looking up tables in the system's built-in database of basic physicochemical parameters. This represents the material temperature synchronously acquired by the data acquisition module, with the unit uniformly defined as °C. The prediction of the volatile partial pressure quantifies the actual driving force for flavor substances to escape from the liquid phase to the gas phase under the current operating conditions, forming a direct digital benchmark for subsequent feedforward intervention control logic.

[0058] After obtaining the predicted volatile partial pressure of the target volatile flavor component, the system needs to quantify the specific risk level of flavor substance loss under the current operating conditions, thus providing a triggering basis for subsequent closed-loop intervention. In a vacuum boiling system, according to the principle of phase equilibrium in chemical thermodynamics, the mass transfer escape rate of a component between the gas and liquid phases depends not only on its own predicted volatile partial pressure but also on the total macroscopic operating pressure of the system. When the local microscopic predicted volatile partial pressure approaches the vacuum pressure within the vacuum concentrator, the mass transfer driving force at the phase interface increases significantly, and the tendency for flavor substances to escape into the gas phase increases accordingly. By introducing a dynamic trend prediction mechanism, this unobservable physical escape tendency can be transformed into a digital judgment benchmark that the control system can execute. (See also...) Figure 4 In this embodiment, the dynamic generation and feedforward intervention determination of the volatility risk index specifically includes the following steps: S401, the intelligent decision-making module calculates and predicts the ratio of the volatile partial pressure to the current vacuum pressure in the vacuum concentrator, and generates the volatile driving ratio.

[0059] Within each baseline sampling period, the intelligent decision-making module extracts the predicted volatile partial pressures of the target volatile flavor components calculated in step S30. and synchronize it with the vacuum pressure acquired by the data acquisition module. Perform a division operation. This ratio, in a physical sense, characterizes the relative potential energy of flavor components overcoming the total system pressure to enter the bulk gas phase. Volatilization drive ratio The specific calculation formula is as follows: ; In the formula, Indicates the target volatile flavor components The volatile driving ratio is a dimensionless transient ratio parameter; This indicates the predicted volatile partial pressure, expressed in kPa. This indicates vacuum pressure, with the unit being kPa.

[0060] S402, the intelligent decision-making module extracts the time differential features of the volatilization driving ratio and dynamically calculates the volatilization risk index.

[0061] Since the concentration process is a typical non-steady-state time-varying process, relying solely on the static ratio at a single moment can easily lead to control lag. To achieve feedforward prediction, the intelligent decision module performs first-order backward differential discretization calculations on the volatile driving ratio, extracts its rate of change over time, and integrates it with the current static ratio. This operation is equivalent to introducing a differential prediction element into the control logic. Volatilization Risk Index The specific calculation formula is as follows: ; In the formula, Indicates the target volatile flavor components At sampling time The volatility risk index below; Indicates the sampling time The evaporation drive ratio is below; This indicates the evaporation drive ratio of the previous sampling period; The reference time interval between two consecutive data samples is expressed in seconds. This represents the prediction time constant, expressed in seconds. As a preferred implementation, the prediction time constant... The value range is set to 2.0s to 5.0s, and the specific value is preset and input by the on-site commissioning personnel based on the overall volume of the concentration unit and the buffer time of the separation chamber.

[0062] In practical engineering control, considering the possibility of missing historical data during the first baseline sampling cycle at the start of the concentration process, the data processing module incorporates dead-zone prevention logic. When it is determined that the current sampling cycle is the first one, the system forcibly adjusts the volatile drive ratio from the previous sampling cycle. The value assigned is the same as the current time. Equal values. This operation makes the differential rate of change term in the first cycle zero, and the volatility risk index is determined only by the current static ratio, thus avoiding the underlying algorithm calculation anomalies caused by the pointer being suspended.

[0063] S403, the intelligent decision-making module determines the intervention conditions based on the control threshold and sends multi-variable collaborative intervention instructions to the execution control module.

[0064] The intelligent decision-making module has pre-set volatile risk control thresholds. The threshold was set based on historical process testing and calibration: engineers recorded volatility driving ratio data under different operating conditions and correlated it with the offline flavor retention rate of the finished bone broth. The critical driving ratio at which flavor loss exceeded the set quality tolerance standard was used as the benchmark, with a value range of 0.85 to 0.92. At the end of each sampling cycle, the intelligent decision module iterates through the volatility risk index of all tracked target volatile flavor components and extracts the maximum value as the global volatility risk representative of the current system.

[0065] When the total volatile risk level does not exceed the set volatile risk control threshold When the intelligent decision-making module determines that the current system is within the flavor safety retention range, the execution control module maintains the original conventional temperature and liquid level control logic unchanged. When the volatility risk index of any target volatile flavor component is detected... Greater than or equal to the volatile risk control threshold Upon detection of a risk of flavor loss, the intelligent decision-making module immediately determines that the system is at risk. At this point, the intelligent decision-making module terminates the conventional single-loop PID control and sends an intervention command matrix to the execution control module via the communication bus. This intervention command matrix includes feedforward setpoint biases requiring synchronized adjustments to the underlying frequency converter, the proportional exhaust valve of the vacuum system, and the steam valve, thereby triggering an asymmetric decoupling collaborative feedforward control mechanism. By proactively intervening by capturing the changing trends of the risk index, the system helps reduce the risk of excessive flavor loss caused by lag in thermodynamic regulation.

[0066] After determining that the system has a risk of flavor stripping and triggering the intervention command matrix, relying solely on a conventional single control loop for adjustment can easily lead to control oscillations due to differences in the dynamic response inertia of different physical fields, affecting the stability of the concentration process. To suppress microscopic flavor escape while maintaining the continuity of the macroscopic process, this embodiment implements cross-domain collaborative control of multiple underlying execution components within the system. (See also...) Figure 5 The asymmetric decoupling collaborative feedforward control of the actuator specifically includes the following steps: S501, the execution control module receives the intervention command matrix and parses the multi-dimensional feedforward bias.

[0067] In this embodiment, the intervention command matrix issued by the intelligent decision-making module to the execution control module includes three dimensions of underlying feedforward adjustment parameters: vacuum pressure bias, heat exchanger steam valve opening bias, and circulating pump operating frequency bias. To ensure the intervention intensity matches the actual risk, these biases are generated by the intelligent decision-making module by multiplying the difference between the previously obtained global volatile risk representative value and the volatile risk control threshold by a pre-tuned feedforward proportional coefficient for each loop. Upon receiving this matrix, the execution control module uses the vacuum pressure bias as the feedforward compensation signal sent to the proportional exhaust valve of the vacuum system, the heat exchanger steam valve opening bias as the feedforward compensation signal sent to the steam regulating valve, and the circulating pump operating frequency bias as the feedforward compensation signal sent to the frequency converter corresponding to the circulating pump control unit, thereby achieving rapid response intervention to the steady-state conventional control logic.

[0068] S502, the execution control module executes asymmetric output response logic based on the difference in physical field inertia.

[0069] Under the operating conditions of industrial concentration units, there are differences in the time constants of the gas phase pressure field, thermodynamic temperature field, and fluid dynamic field. To avoid physical stress shocks caused by synchronous abrupt changes, the execution control module implements asymmetric action regulation based on the characteristics of the object. Specifically, due to the rapid response of the gas phase pressure, the execution control module directly sends the vacuum pressure bias to the proportional exhaust valve of the vacuum system in the form of a step signal. This operation establishes a high-pressure barrier in the gas phase by increasing the absolute pressure in the separation chamber, suppressing the escape of flavor molecules at the gas-liquid interface. At the same time, due to the large thermal inertia of the heat exchanger and connecting pipelines, the execution control module automatically incorporates a first-order inertial filter with a preset time constant when outputting the heat exchanger steam valve opening bias to the steam regulating valve. As a preferred method, the preset time constant is set to a range of 15s to 40s, so that the steam valve opening shows a gradual decreasing trend, preventing severe metal thermal stress in the heat exchange tube bundle. In addition, for the fluid field, the execution control module sends the circulating pump operating frequency bias to the frequency converter, increasing the material flow rate by appropriately increasing the output frequency of the frequency converter. This action utilizes the non-Newtonian fluid dynamics property of shear thinning of the bone broth matrix to reduce the apparent viscosity of the boundary layer fluid, accelerate the peeling rate of bubbles on the heated surface, and weaken local hot spots.

[0070] S503, the execution control module implements decoupling compensation of the temperature control loop through boiling point offset calculation.

[0071] In the operation of increasing vacuum pressure to suppress flavor loss, the increase in absolute system pressure leads to a synchronous rise in the saturated boiling point of the material. Under this coupled state, if the steam supply valve only closes based on its own bias, it can easily cause a reduction in the heat transfer temperature difference within the heat exchanger, thus affecting concentration efficiency. To eliminate this coupling interference, the execution control module internally calls a temperature decoupling compensator. This compensator calculates the boiling point rise increment corresponding to the current vacuum pressure bias in real time, based on the physical correspondence between the saturated vapor pressure and temperature. The execution control module uses the calculated boiling point rise increment as a dynamic compensation value and adds it in the same direction to the target temperature setpoint of the heating system. This decoupling operation ensures that the system can maintain an effective heat transfer temperature difference and guarantee continuous moisture evaporation even when the pressure is increased to suppress flavor loss. The lookup table program or thermodynamic equation calculation for converting saturated vapor pressure to corresponding boiling point temperature can be conventionally configured by those skilled in the art based on fluid property datasheets; this is well-known technology in the field and will not be elaborated upon here.

[0072] S504, the execution control module monitors the dynamic status and implements smooth rollback recovery.

[0073] During the execution of asymmetric decoupling collaborative feedforward control, the intelligent decision-making module continuously tracks and monitors the global volatility risk representative value. When the global volatility risk representative value is detected to have fallen back and stabilized below the set volatility risk control threshold, the system determines that the current operating boundary has moved away from the excessive flavor loss condition. At this time, the execution control module stops intervening and, based on the preset slope generator, gradually reduces the vacuum pressure bias sent to the proportional exhaust valve of the vacuum system, the heat exchanger steam valve opening bias sent to the steam regulating valve, and the circulating pump operating frequency bias sent to the corresponding frequency converter of the circulating pump control unit to zero. In this embodiment, the recovery rate of the slope generator is set to decrease by 1% to 3% of the original set total bias amplitude per second to ensure a smooth system transition. After all feedforward biases are removed, the underlying actuators return to the independent single-loop proportional-integral-differential algorithm for normal steady-state regulation, and the system returns to the basic evaporation and concentration production mode.

[0074] During long-term industrial production, scaling on the heat exchanger tube surface, physical aging of sensors, and fluctuations in the micro-composition of bone broth raw materials between batches can cause gradually accumulating data drift in the previously established basic calculation model. To compensate for the aforementioned long-term system errors, this embodiment adds a correction mechanism for batch-to-batch error evolution in addition to the basic high-frequency real-time intervention control loop. The adaptive feedback correction based on slow-loop model parameters specifically includes the following steps: S601, the data processing module extracts historical operating data of the current complete production batch and calculates the cumulative prediction deviation.

[0075] After each batch of bone broth concentration is completed, the data processing module retrieves the historical operating database for the entire cycle of that batch. Based on the retention rate predicted by the model, the data processing module extracts the predicted volatile partial pressures of the target volatile flavor components calculated for each baseline sampling cycle, according to the principles of mass transfer kinetics, and converts them into transient volatile mass rates by combining this with the system's preset total mass transfer coefficient. The specific calculation formula for the transient volatile mass rate is as follows: ; In the formula, This represents the transient evaporation mass rate, expressed in kg / s. This represents the preset total mass transfer coefficient of the equipment, expressed in kg / (s·kPa), obtained through offline equipment calibration. Indicates the predicted volatile partial pressure; Indicates vacuum pressure; This represents the mole fraction of the corresponding component in the gas phase space, and its value is based on the gas phase response signal output by the online detection module. The concentration is obtained by mapping and looking up a table based on the pre-calibrated linear working curve of the gas phase concentration within the system.

[0076] Then, the transient evaporation mass rate is discretely integrated along the entire batch time axis to obtain the total mass of gas phase escape. This total escape mass is then subtracted from the initial material mass base to calculate the predicted retention rate of the target volatile flavor component at the end of the batch. Simultaneously, the system receives the actual retention rate of the batch from offline gas chromatography measurements in the laboratory via an external interface. The data processing module subtracts the predicted retention rate from the actual retention rate to obtain the cumulative prediction deviation for a single batch.

[0077] S602, the data processing module performs anti-jitter dead zone verification on the cumulative prediction deviation.

[0078] Because offline testing instruments inherently possess random measurement errors, responding to minute fluctuations can easily lead to over-adjustment of the parameters within the neural network model. Therefore, the data processing module is configured with an anti-jitter tolerance threshold, ranging from 0.02 to 0.05. The data processing module extracts the absolute value of the calculated cumulative prediction deviation and compares it with the anti-jitter tolerance threshold. When the absolute value of the cumulative prediction deviation is less than or equal to the anti-jitter tolerance threshold, the system determines that the current model's prediction accuracy is within the allowable engineering tolerance range and does not perform parameter correction. When the absolute value of the cumulative prediction deviation is greater than the anti-jitter tolerance threshold, the data processing module determines that the model has experienced a baseline shift and triggers the next step of parameter iteration.

[0079] S603, the data processing module performs adaptive iterative updates of the neural network model parameters based on the prediction deviation.

[0080] Upon receiving the parameter iteration command, the data processing module performs reverse fine-tuning of the output layer parameters of the multilayer perceptron neural network within the aforementioned flavor evaluation module based on the polarity and magnitude of the deviation. From a physical mechanism perspective, the difference between equipment scaling and the colloidal structure of raw materials primarily alters the mass transfer resistance and macromolecular binding capacity of the multiphase interface. This directly maps to changes in the equivalent activity correction coefficient, while having a relatively small impact on the liquid-phase equivalent flavor state quantity, which reflects the conservation of basic materials. Therefore, to avoid overfitting caused by online retraining of the entire network weights and to reduce computational overhead, this embodiment employs a fixed-point bias update algorithm, iterating slowly only on the node baseline bias corresponding to the equivalent activity correction coefficient in the output layer. Since an increase in the node baseline bias ultimately leads to an increase in the predicted volatile partial pressure and a decrease in the predicted retention rate, and these two have a negative physical correlation, the update calculation formula for this node baseline bias is as follows to satisfy the closed-loop negative feedback convergence condition: ; In the formula, Indicates that the updated version is used for the first... Reference bias of output layer nodes of multilayer perceptron neural network for batch concentration control; Indicates the current number Reference bias of output layer nodes of multilayer perceptron neural network used in batch; This represents the preset slow loop learning step size, which is a dimensionless constant. Indicates the target volatile flavor components In the current number The cumulative prediction bias of a batch is calculated by substituting the signed algebraic difference between the actual retention rate and the predicted retention rate. The subtraction mechanism ensures that when the predicted retention rate is underestimated (i.e., the volatile loss is overestimated), the model can reduce the baseline bias and thus reduce the predicted volatile partial pressure, making the error correction direction absolutely physically correct.

[0081] As a preferred method, a preset slow-loop learning step size The value range is set to 0.01 to 0.08. This setting limits the maximum magnitude of a single correction. This slow-cycle update mechanism across batches can gradually compensate for long-term offsets caused by equipment hardware aging and raw material matrix differences without interfering with the underlying high-frequency multivariate collaborative feedforward control, thus helping to maintain the engineering effectiveness of the prediction model in complex production environments.

[0082] A food ingredient factory uses a 5-ton external circulation vacuum concentrator to produce stock concentrate. The goal is to concentrate pork bone broth with an initial soluble solids content of 8% to 35%. This embodiment focuses on monitoring the volatile flavor components. It is 2-methylpyrazine. The system sets a reference time interval between two adjacent data samples. =1.0s, prediction time constant =3.0s, volatility risk control threshold =0.88.

[0083] At the 40-minute mark of the concentration process, the apparent state of the fluid changes, and the parameters acquired online by the system are: operating frequency. =40Hz, material flow rate =10m 3 / h, pressure difference between heat exchanger inlet and outlet =50kPa, soluble solids content =20. Low-level communication extracts output torque. =45 N·m. The system calls the model to calculate the reference torque. =15 N·m. Set the electromechanical conversion efficiency correction factor. =0.9, characteristic weighting coefficient =0.2.

[0084] Substitute into the formula to calculate the apparent viscosity index: ; The increase in this value indicates that the polymer matrix network has a significant inhibitory effect on mass transfer.

[0085] Neural network combination =0.833 Feature extraction, output equivalent activity correction coefficient The system obtains the current material temperature. =65℃, and retrieved the Antoine equation constant for 2-methylpyrazine from the internal database to calculate the predicted volatile partial pressure. =11.147 kPa. Current vacuum pressure =14.2 kPa. Calculate the volatilization drive ratio. .

[0086] Combining the volatility drive ratio of 0.770 from the previous sampling period, the volatility risk index is calculated using the formula. ; See Figure 6 , Figure 6 The horizontal axis represents the concentration time, the vertical axis represents the global volatility risk representative value, and the horizontal dotted line represents the preset volatility risk control threshold of 0.88. Figure 6 In the diagram, solid lines with circular markers represent the experimental group's status. At the 40-minute mark, the system-calculated global volatility risk reached 0.83. At this point, because the differential prediction stage detected an upward trend in the rate of change, the system issued a coordinated intervention command at the 40-minute mark to prevent it from further exceeding the volatility risk control threshold of 0.88. For comparison, Figure 6 The solid line in the graph indicates that if traditional controls are used without early intervention, the risk index will cross the 0.88 threshold line around the 47th minute and continue to rise in the higher risk range.

[0087] The system triggers a coordinated intervention command to change the local mass transfer thermodynamic state: First, a step command is immediately executed to increase the vacuum pressure. The pressure was increased from 14.2 kPa to 16.0 kPa; secondly, the target opening of the steam valve was reduced by 15%, and an inertial filter with a time constant of 20 s was introduced to perform a gradual descent; finally, the operating frequency was adjusted. Increasing the Hz frequency by 5 Hz reduces the apparent viscosity by increasing the flow rate, thus accelerating the rate of bubble peeling off from the heated surface. For example... Figure 6 As shown in the curve of the experimental group, after coordinated intervention, around the 45th minute, The value steadily dropped back to a safe range of approximately 0.77, and then the system maintained a stable operating state.

[0088] To verify the effectiveness of this invention, the same batch of raw materials was compared and verified using conventional control (control group) and the control method of this invention (experimental group) on the same equipment (10 batches of each were run and the average value was taken). (Refer to...) Figure 7The horizontal axis represents concentration time, and the vertical axis represents 2-methylpyrazine retention. Regarding the retention of characteristic flavors... Figure 7 The solid line shows that, after the 40th minute, due to the lack of state intervention, the retention rate of the control group decreased non-linearly, dropping to 60.0% by the end of 100 minutes.

[0089] and Figure 7 The solid line with square markers shows that after the intervention at 40 minutes, the rate of decline in retention rate slowed in the experimental group, and by the end of 100 minutes, the retention rate stabilized at 92.5%, an increase of 32.5% compared to the control group. In terms of the overall sensory evaluation score, the control group scored 76 points, while the experimental group scored 92 points. The objective detection data is consistent with the artificial sensory evaluation, indicating that the present invention effectively retains the target volatile flavor components.

[0090] In terms of production efficiency, the average concentration time per batch in the control group was 125 minutes, while in the experimental group it was 118 minutes, indicating that improving the fluid dynamic field conditions through intervention mechanisms did not affect the overall concentration efficiency. Regarding equipment maintenance, after 10 batches, the average scale thickness on the pipe walls of the control group was 3.2 mm, while that of the experimental group was only 1.5 mm, demonstrating that multivariate synergy effectively prevented abnormal local heating surface temperatures and reduced solid adhesion.

[0091] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for controlling the flavor stability of bone broth through vacuum concentration, characterized in that, include: Simultaneously collect macroscopic thermodynamic parameters, material state parameters, gas phase response signals of target volatile flavor components, and bottom electrical load parameters of the circulating pump motor within the vacuum concentrator. The macroscopic thermodynamic parameters include the pressure difference between the inlet and outlet of the heat exchanger and the vacuum pressure within the vacuum concentrator. The apparent viscosity index is generated from the underlying electrical load parameters, the inlet and outlet pressure difference of the heat exchanger, and the reference no-load torque model. The macroscopic thermodynamic parameters, the material state parameters, the gas phase response signal, and the apparent viscosity index are input into a preset neural network model, which outputs the liquid phase equivalent flavor state quantity and the equivalent activity correction coefficient, and calculates the predicted volatile partial pressure of each target volatile flavor component. Based on the ratio of the predicted volatile partial pressure to the vacuum pressure in the vacuum concentration device and the time derivative data of the ratio, a corresponding volatile risk index is generated. When the volatility risk index is greater than or equal to the volatility risk control threshold, an intervention instruction matrix is ​​issued to the underlying execution component.

2. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 1, characterized in that, Multi-source acquired values ​​are synchronously obtained during the main polling cycle of the distributed control system. These multi-source acquired values ​​include: Material temperature, vacuum pressure in the vacuum concentration unit, and pressure difference between the inlet and outlet of the heat exchanger are macroscopic thermodynamic parameters. Material flow rate and soluble solids content are parameters of material state. The gas phase response signal of the target volatile flavor component as a gas phase detection parameter; And the stator current, output torque and operating frequency of the circulating pump motor, which are the underlying electrical load parameters; Each collected value is timestamped, and a zero-order hold algorithm is used for sensor signal channels with low reporting frequency to form a multi-dimensional initial input data matrix.

3. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 1, characterized in that, The gas phase response signal is acquired by a metal oxide semiconductor gas sensor array or an industrial-grade online ion mobility spectrometer located in the secondary steam emission path or the gas phase space region at the top of the separation chamber. The raw electrical signal output from the metal oxide semiconductor gas sensor array or industrial-grade online ion mobility spectrometer is preprocessed, features are extracted, and normalized to obtain a dimensionless gas phase response signal in the range of 0 to 1.

4. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 2, characterized in that, The generation of the apparent viscosity index includes: During the stage of benchmark calibration operation with atmospheric pressure clean water, the benchmark output torque of the circulating pump under different operating frequencies and different material flow rates is collected, and the benchmark no-load torque model is established by multivariate polynomial surface fitting. During the operation phase, the net fluid load torque is obtained by subtracting the reference torque given by the reference no-load torque model from the output torque. The dimensionless apparent viscosity index is generated based on the net fluid load torque, the operating frequency of the circulating pump, the material flow rate, the pressure difference between the inlet and outlet of the heat exchanger, the soluble solids content, and the electromechanical conversion efficiency correction coefficient.

5. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 4, characterized in that, When the output torque is less than the reference torque, the net fluid load torque is set to zero; When the material flow rate is lower than the flow protection lower limit threshold, the apparent viscosity index is assigned a preset standard constant of 1.0 in the first sampling period, and the effective calculated value of the previous sampling period is retained in subsequent sampling periods.

6. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 2, characterized in that, The neural network model is a feedforward multilayer perceptron. The input layer of the neural network model receives a 9-dimensional initial input column vector consisting of material temperature, vacuum pressure in the vacuum concentration device, pressure difference between the inlet and outlet of the heat exchanger, material flow rate, soluble solids content, gas phase response signal, stator current of the circulating pump motor, output torque and operating frequency of the circulating pump. This column vector is then concatenated with the apparent viscosity index to form a 10-dimensional feature input vector. The output layer of the neural network model outputs liquid-phase equivalent flavor state quantities and equivalent activity correction coefficients.

7. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 2, characterized in that, The predicted volatile partial pressure is calculated based on the liquid-phase equivalent flavor state quantity, the equivalent activity correction coefficient, the material temperature, and the Antoine constant corresponding to the target volatile flavor component.

8. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 7, characterized in that, The generation of the volatility risk index includes: The ratio of the predicted volatile partial pressure to the vacuum pressure inside the vacuum concentrator is calculated to obtain the volatile driving ratio; The volatilization drive ratio is discretized using a first-order backward difference method to obtain the time differential characteristics; The current value of the volatility driving ratio is weighted and fused with the time differential feature to obtain the volatility risk index of the target volatile flavor component.

9. The method for controlling the flavor stability of bone broth through vacuum concentration according to claim 1, characterized in that, The intervention command matrix includes vacuum pressure offset, heat exchanger steam valve opening offset, and circulating pump operating frequency offset. The underlying execution components include the frequency converters corresponding to the proportional exhaust valve, steam regulating valve, and circulating pump control unit in the vacuum system. The vacuum pressure bias is sent to the proportional exhaust valve of the vacuum system in the form of a step signal. The opening bias of the heat exchanger steam valve is fed into a first-order inertial filter with a preset time constant when it is output to the steam regulating valve. The operating frequency bias of the circulating pump is sent to the frequency converter corresponding to the circulating pump control unit.

10. A system for controlling the flavor stability of bone broth through vacuum concentration, characterized in that, The method for controlling the flavor stability of bone broth through vacuum concentration according to any one of claims 1 to 9 includes a vacuum concentration device and a distributed control system, wherein the distributed control system includes an online detection module, a data acquisition module, a data processing module, a flavor evaluation module, an intelligent decision-making module, and an execution control module. The online detection module is used to acquire the gas-phase response signal of the target volatile flavor component; The data acquisition module is used to acquire macroscopic thermodynamic parameters, material state parameters, and the underlying electrical load parameters of the circulating pump motor. The macroscopic thermodynamic parameters include the vacuum pressure in the vacuum concentration unit and the pressure difference between the inlet and outlet of the heat exchanger. The data processing module is used to generate the apparent viscosity index; The flavor evaluation module is used to output liquid-phase equivalent flavor state quantities and equivalent activity correction coefficients; The intelligent decision-making module is used to calculate the predicted volatile partial pressure of the target volatile flavor components and generate a volatility risk index and an intervention instruction matrix; The execution control module includes a vacuum pump control unit, a steam valve control unit, and a circulation pump control unit. The vacuum pump control unit is connected to the vacuum pump, the steam valve control unit is connected to the steam regulating valve, and the circulation pump control unit is connected to the circulation pump. The execution control module is used to send the vacuum pressure bias to the proportional exhaust valve of the vacuum system in the form of a step signal.

Images