Intelligent control system for biopharmaceutical process parameters based on big data

Abstract

The present application relates to the technical field of biopharmaceutical process control, and discloses a biopharmaceutical process parameter intelligent control system based on big data, which comprises an offline modeling module, a state sensing module, a feature reconstruction module and a collaborative control module. The offline modeling module processes historical data to generate a benchmark feedforward compensation matrix; the state sensing module calculates an online oxygen uptake rate to generate a real-time sliding data window buffer sequence; the feature reconstruction module combines the benchmark feedforward compensation matrix and the buffer sequence to evaluate a composite time sequence matching degree, and generates reconstructed feedforward parameters; and the collaborative control module performs amplitude limiting processing on the reconstructed feedforward parameters, superimposes them on a bottom layer closed loop to generate control instructions, and executes a hierarchical degradation response when a hardware exception is detected. The present application solves the problem of control instruction lag or over-regulation caused by metabolic index deviation, and prevents abnormal signals from participating in closed loop calculation to cause bottom layer mechanism misoperation.

Description

Technical Field

[0001] This invention relates to the field of biopharmaceutical process control technology, specifically to an intelligent control system for biopharmaceutical process parameters based on big data. Background Technology

[0002] Existing biopharmaceutical fermentation or cell culture processes typically employ a single closed-loop feedback control or static feedforward control strategy to regulate feed or gas supply. Due to the nonlinear characteristics and batch-to-batch variations in bioreaction processes, traditional static control models cannot dynamically adjust feedforward compensation in real time when metabolic indicators such as oxygen uptake exhibit temporal phase shifts or amplitude fluctuations. This can lead to control command lag or overshoot, affecting process stability.

[0003] At the system status acquisition and data processing level, existing control methods, when calculating metabolic parameters, are susceptible to deviations in exhaust flow rate measurement due to exhaust moisture evaporation and reactor pressure changes, resulting in inaccurate online oxygen uptake rates. Furthermore, during batch data comparison and parameter reconstruction, when ratio calculations are involved and the baseline amplitude data approaches zero, the underlying control algorithm experiences overflow errors, triggering system operation interruptions.

[0004] Furthermore, existing process control systems lack degradation protection and interlocking mechanisms for anomalies in underlying sensor hardware. When auxiliary components or main control input channel sensors experience open circuits or short circuits, the system cannot proactively identify and isolate the fault source. Incorrect measurement signals directly participate in closed-loop control loop calculations, generating erroneous compensation biases, which in turn lead to malfunctions in the underlying actuators, causing the process environment to become uncontrollable. Summary of the Invention

[0005] The technical problem addressed by this invention is that in existing biopharmaceutical fermentation or cell culture process control, process parameters often employ a single closed-loop feedback or static feedforward control strategy. Due to the nonlinearity and batch-to-batch variability of bioreaction processes, traditional control models cannot correct feedforward compensation in real time when key metabolic indicators experience time phase shifts or amplitude fluctuations, leading to lag or over-adjustment of feed or gas supply control commands. Furthermore, the lack of system degradation protection logic when underlying sensors malfunction means that erroneous signals directly participate in control loop calculations, potentially causing actuator malfunctions and affecting the process control.

[0006] To address the above problems, the present invention provides the following technical solution: This invention provides an intelligent control system for biopharmaceutical process parameters based on big data, comprising an offline modeling module, a state perception module, a feature reconstruction module, and a collaborative control module. The offline modeling module acquires multiple batches of historical time-series data and performs zero-phase digital filtering to extract baseline delay time features and generate a baseline feedforward compensation matrix. The state perception module calculates the current online oxygen uptake rate and continuously writes it to a memory queue according to the sampling timestamp, generating a real-time sliding data window buffer sequence for the current batch. The feature reconstruction module combines the baseline feedforward compensation matrix and the real-time sliding data window buffer sequence to perform a composite time-series matching degree evaluation, and calculates and generates reconstructed feedforward parameters when the deviation calculation conditions are met. The collaborative control module performs safety boundary limiting processing on the reconstructed feedforward parameters and generates a limiting bias, which is then superimposed on the underlying closed-loop feedback loop to generate a general control command to drive the underlying hardware to perform actions.

[0007] Furthermore, the offline modeling module acquires multiple batches of historical time series data and performs zero-phase digital filtering to extract baseline delay time features and generate a baseline feedforward compensation matrix. Specifically, it performs zero-phase digital filtering on multiple batches of historical time series data using a forward and backward superposition filtering method to obtain the final filtered sequence, and extracts the offline feeding command sequence and the baseline oxygen uptake rate sequence. It calculates the correlation coefficient between the offline feeding command sequence and the baseline oxygen uptake rate sequence using the zero-mean normalized cross-correlation formula, and extracts the baseline delay time features using the baseline delay time feature extraction formula. Based on the baseline delay time features, it establishes a parameter mapping relationship between the input and output dimensions and generates the baseline feedforward compensation matrix.

[0008] Furthermore, the specific method by which the state awareness module calculates the current online oxygen uptake rate and continuously writes it into the memory queue according to the sampling timestamp, and generates the real-time sliding data window buffer sequence for the current batch is as follows: the actual exhaust flow rate correction value is calculated using the actual exhaust flow rate correction formula, and the current online oxygen uptake rate per unit volume is calculated using the online oxygen uptake rate calculation formula; the latest current online oxygen uptake rate and its timestamp are written to the tail of the memory queue, and the earliest data at the head of the queue is deleted, so as to continuously write it into the memory queue according to the sampling timestamp; in the continuous update operation of the memory queue, the real-time sliding data window buffer sequence is generated.

[0009] Furthermore, the feature reconstruction module combines the benchmark feedforward compensation matrix and the real-time sliding data window buffer sequence to perform a composite time series matching degree evaluation, and determines whether the deviation calculation conditions are met in the following specific way: extract the variance of the real-time sliding data window buffer sequence and the corresponding called benchmark oxygen uptake sequence; when the variance is not lower than the set detection lower limit, use the composite time series matching degree evaluation formula, combine the benchmark feedforward compensation matrix and the real-time sliding data window buffer sequence to calculate, and obtain the composite time series matching degree; compare the composite time series matching degree with the preset matching threshold, and determine whether the deviation calculation conditions are met when the composite time series matching degree is lower than the preset matching threshold.

[0010] Furthermore, the specific method by which the feature reconstruction module calculates and generates reconstruction feedforward parameters when it determines that the deviation calculation conditions are met is as follows: After determining that the deviation calculation conditions are met, the phase drift is extracted using a cross-correlation sliding search algorithm, and the amplitude residual coefficient is extracted using an anti-singularity amplitude residual coefficient calculation formula; the phase drift and amplitude residual coefficient are combined for calculation, and the reconstruction feedforward parameters for the current batch are generated using a spatiotemporal reconstruction feedforward parameter calculation formula; a timing consistency constraint not exceeding the preset maximum ramp rate is applied to the reconstruction feedforward parameters, and reconstruction feedforward parameters that meet the execution boundary of the output component are generated.

[0011] Furthermore, the collaborative control module performs safety boundary limiting processing on the reconstructed feedforward parameters and generates limiting biases. The limiting biases are then superimposed onto the underlying closed-loop feedback loop to generate the overall control command that drives the underlying hardware to perform actions. Specifically, this is done by: combining the lower and upper safety thresholds of the feedforward bias, using a feedforward bias limiting formula to perform safety boundary limiting processing on the reconstructed feedforward parameters and generate limiting biases; performing first-order smoothing on the limiting biases to generate smooth feedforward biases, then superimposing the smooth feedforward biases onto the underlying basic closed-loop feedback command, and using the overall control command calculation formula to generate the overall control command; combining the lower and upper execution thresholds of the output components, using a total output limiting formula to finally constrain the overall control command, and generating execution-level control commands to be sent to the corresponding output components.

[0012] Furthermore, the specific method for the collaborative control module to perform hardware abnormal interlock detection is as follows: read the original analog current signal transmitted by the underlying input channel; convert the original analog current signal into the corresponding engineering physical quantity; when the original analog current signal is detected to be lower than the preset low current diagnostic threshold or higher than the preset high current diagnostic threshold, or the engineering physical quantity is lower than the preset lower limit of the process range or higher than the preset upper limit of the process range, set the interlock flag in the system status register to the hardware abnormal interlock state.

[0013] Furthermore, the specific method for the collaborative control module to perform graded and degraded responses under hardware abnormal interlocking conditions is as follows: after the interlocking flag in the system status register is in a hardware abnormal interlocking state, the auxiliary status sensing component abnormality and the main control loop input component abnormality are distinguished according to the source of the fault; when the source of the fault is the auxiliary status sensing component abnormality, the attenuation feedforward bias is calculated using the linear attenuation blocking factor calculation formula and the attenuation feedforward bias calculation formula, and the feedforward channel is smoothly blocked; when the source of the fault is the main control loop input component abnormality, the calculation and update of the corresponding basic closed-loop feedback instruction are frozen and the preset safety baseline maintenance value is called for output.

[0014] Furthermore, the status sensing module calculates the actual exhaust flow correction value using the actual exhaust flow correction formula, and calculates the current online oxygen uptake rate using the online oxygen uptake rate calculation formula. The specific method is as follows: obtain the molar fraction of nitrogen in the intake air and the molar fraction of nitrogen in the exhaust air, and calculate the intake and exhaust flow ratio correction term based on the inert gas mass conservation; multiply the obtained total intake flow rate by the intake and exhaust flow ratio correction term and the comprehensive correction coefficient of temperature, pressure and humidity, and generate the actual exhaust flow correction value; combine the actual exhaust flow correction value, the intake oxygen flow rate and the exhaust oxygen flow rate to derive the total oxygen consumption rate, and calculate the current online oxygen uptake rate based on the current culture medium volume.

[0015] Furthermore, the feature reconstruction module extracts the amplitude residual coefficient using the anti-singularity amplitude residual coefficient calculation formula as follows: After obtaining the phase drift amount through cross-correlation sliding search algorithm matching, the real-time sequence average amplitude of the real-time sliding data window buffer sequence and the reference sequence average amplitude of the corresponding reference sequence are calculated; an amplitude anti-singularity protection constant is superimposed on the reference sequence average amplitude to obtain the protection correction reference amplitude, and the real-time sequence average amplitude is divided by the protection correction reference amplitude to generate the initial amplitude residual ratio; the lower limit threshold and upper limit threshold of the amplitude residual coefficient are applied to the initial amplitude residual ratio, and the final amplitude residual coefficient is extracted and generated.

[0016] This invention provides an intelligent control system for biopharmaceutical process parameters based on big data. It has the following beneficial effects: 1. This invention uses a feature reconstruction module combined with a baseline feedforward compensation matrix and a real-time sliding data window buffer sequence to perform composite time-series matching degree evaluation, and generates reconstructed feedforward parameters when the deviation calculation conditions are met. This method extracts the drift and amplitude change rate of the actual time-series phase, and superimposes the reconstructed feedforward parameters onto the underlying closed-loop feedback loop, thereby correcting the timing and output magnitude of the feedforward compensation action. This solves the problem of delayed or over-adjusted feed or gas supply control commands caused by traditional static feedforward control when metabolic indicators experience phase shifts or amplitude fluctuations during biological reactions.

[0017] 2. This invention introduces an inert gas mass conservation principle to derive a correction term for the intake and exhaust flow rate ratio in the state sensing stage, thereby calculating the online oxygen uptake rate and eliminating measurement deviations caused by exhaust moisture and pressure changes. When extracting the amplitude residual coefficient, the system superimposes an amplitude anti-singularity protection constant on the average amplitude of the reference sequence to obtain a protected correction reference amplitude. This calculation logic avoids computational overflow errors that occur when the divisor reference amplitude approaches zero, ensuring the underlying computational stability of the control system during complex batch data iteration processes.

[0018] 3. This invention establishes a hardware anomaly interlocking detection mechanism based on the original analog current signal and engineering physical quantities in the collaborative control module, and executes a graded degradation response according to the source of the fault. When an auxiliary component malfunctions, the system calculates the attenuation feedforward bias and smoothly blocks the feedforward channel; when the main control input component malfunctions, the system freezes the closed-loop feedback instruction update and outputs a preset safety baseline maintenance value. This mechanism isolates erroneous measurement signals from the underlying sensors, preventing them from continuously participating in the control loop calculation and avoiding malfunctions of the underlying actuators under hardware failure conditions. Attached Figure Description

[0019] Figure 1 This is the overall architecture diagram of the intelligent control system for biopharmaceutical process parameters based on big data according to the present invention; Figure 2 This is a flowchart illustrating the overall workflow of the intelligent control method for biopharmaceutical process parameters based on big data, as described in this invention. Figure 3 This is a data processing logic diagram of an offline dimensionality reduction modeling module according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the parameter exchange and online status sensing data flow according to an embodiment of the present invention; Figure 5 This is a logic block for state evaluation and spatiotemporal feature reconstruction according to an embodiment of the present invention; Figure 6 This is a block diagram of the collaborative execution and limiting closed-loop control logic of one embodiment of the present invention; Figure 7 This is a hardware interlocking and hierarchical degradation backoff control logic diagram according to an embodiment of the present invention; Figure 8 This is a probability density distribution diagram of dissolved oxygen control deviation in this invention; Figure 9 This is a phase trajectory diagram of the underlying control command response under a hardware disconnection fault according to the present invention. Detailed Implementation

[0020] 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.

[0021] See attached document Figure 1 The intelligent control system for biopharmaceutical process parameters based on big data provided by this invention includes an offline modeling module, a state perception module, a feature reconstruction module, and a collaborative control module.

[0022] The offline modeling module is used to acquire multiple batches of historical time series data, perform zero-phase digital filtering on the multiple batches of historical time series data, and extract the benchmark delay time features between physical operations and biological metabolic responses to generate a benchmark feedforward compensation matrix.

[0023] The state perception module is used to calculate the actual exhaust flow correction value based on the inert gas mass conservation law, and to deduce the current online oxygen uptake rate based on the difference between the intake oxygen flow rate and the exhaust oxygen flow rate. The state perception module is also used to continuously write the current online oxygen uptake rate into the memory queue of the industrial control computer according to the sampling timestamp, so as to form the real-time sliding data window buffer sequence of the current batch.

[0024] The feature reconstruction module is used to evaluate the composite time-series matching degree of the real-time sliding data window buffer sequence, and calculate the phase drift and amplitude residual coefficient when the deviation calculation conditions are met. The feature reconstruction module is also used to shift the time axis in the reference feedforward compensation matrix according to the phase drift, and scale the bias amplitude of the reference control quantity according to the amplitude residual coefficient to generate the reconstruction feedforward parameters for the current batch.

[0025] The collaborative control module is used to limit the reconstructed feedforward parameters and superimpose the limiting bias onto the underlying closed-loop feedback loop to generate the general control command that drives the underlying hardware to perform actions.

[0026] In this embodiment, each of the above modules is deployed in its corresponding hardware device. The system includes an industrial control computer, a programmable logic controller (PLC), and a network control device. The industrial control computer serves as the upper-level centralized processing unit, and it houses an offline modeling module, a state awareness module, and a feature reconstruction module; the PLC serves as the lower-level execution unit, and it houses a collaborative control module.

[0027] Input components include an exhaust gas analyzer, dissolved oxygen sensor, pH sensor, temperature sensor, pressure sensor, and a level sensor or weighing sensor for obtaining the volume of the culture medium. Output components include a gas mass flow controller, a stirrer motor frequency converter, a feed pump, and a pH regulating pump.

[0028] Industrial control computers communicate with programmable logic controllers (PLCs) via network control devices, and the PLCs centrally control the input and output components of one or more bioreactor production units.

[0029] See attached document Figure 2 The control method includes the following steps: S1, Offline Dimensionality Reduction Modeling: Extract multiple batches of historical time series data, perform zero-phase digital filtering on the multiple batches of historical time series data, and calculate the benchmark delay time characteristics to establish a benchmark feedforward compensation matrix; S2, Parameter handover and online sensing: The reference feedforward compensation matrix is ​​sent to the underlying execution unit, and after the data block verification is completed, the actual exhaust flow correction value is calculated based on the inert gas mass conservation; the current online oxygen uptake rate is derived based on the difference between the intake oxygen flow rate and the exhaust oxygen flow rate, and the current online oxygen uptake rate is continuously written into the memory queue of the industrial control computer according to the sampling timestamp to form a real-time sliding data window buffer sequence; S3, State Assessment and Spatiotemporal Reconstruction: Perform composite time-series matching degree assessment on the real-time sliding data window buffer sequence, and calculate the phase drift and amplitude residual coefficient when the deviation calculation conditions are met; shift the time axis in the reference feedforward compensation matrix according to the phase drift, and scale the bias amplitude of the reference control quantity according to the amplitude residual coefficient to generate the reconstruction feedforward parameters for the current batch. S4, Cooperative Execution and Limiting Closed Loop: Receives the reconstructed feedforward parameters, performs safety boundary limiting processing on the reconstructed feedforward parameters, and superimposes the limiting bias onto the underlying closed loop feedback loop to generate the general control command that drives the hardware to perform actions; S5, Hardware Interlocking and Abnormal Backoff: When a disconnection or over-limit abnormality is detected in the feedback signal of an input component, the collaborative control module performs graded backoff according to the source of the fault. When the source of the fault is an auxiliary status sensing component such as an exhaust gas analyzer, the feedforward channel is blocked and backoff is performed to a single closed-loop feedback control mode. When the source of the fault is a main control loop input component such as a dissolved oxygen sensor or a pH sensor, the corresponding closed-loop feedback calculation is frozen and the preset safety baseline maintenance value is called for output.

[0030] The above steps constitute the basic control flow of this embodiment. The following is in conjunction with the appendix... Figure 3 To be continued Figure 7 The specific implementation methods for each step are explained.

[0031] See attached document Figure 3 In this embodiment, step S1 is executed by an offline modeling module deployed within the industrial control computer. The offline modeling module extracts baseline time delay features between physical operations and biological metabolic responses from multiple batches of historical time-series data, and establishes a baseline feedforward compensation matrix accordingly. Step S1 specifically includes the following sub-steps.

[0032] S101, the offline modeling module extracts multiple batches of historical time series data and performs zero-phase digital filtering on the multiple batches of historical time series data.

[0033] Specifically, the offline modeling module retrieves continuous-time process data from multiple qualified batches from the historical database of the Manufacturing Execution System (MES) according to a preset hardware sampling frequency. The multiple batches of historical time-series data include at least the offline feed instruction sequence and the baseline oxygen uptake sequence calculated based on reactor exhaust composition.

[0034] In the biopharmaceutical culture process, the raw signals collected by input components such as dissolved oxygen sensors typically contain high-frequency electrical noise, bubble disturbance noise, and sampling jitter. If conventional low-pass filtering is used directly to process these signals, additional group delays may be introduced, thereby affecting the accuracy of subsequent hysteresis time identification.

[0035] Therefore, as a preferred approach, the offline modeling module employs a forward and backward superposition filtering method to perform zero-phase digital filtering. Specifically, the offline modeling module first inputs the original data sequence along the forward time axis into a digital low-pass filter to obtain a forward filtered sequence; then, the forward filtered sequence is reversed along the time axis and input into the same digital low-pass filter again to obtain a backward filtered sequence; subsequently, the backward filtered sequence is reversed again along the time axis to obtain the final filtered sequence. Through this process, the influence of high-frequency noise can be reduced while minimizing the impact of the filtering process on the phase of the original physical action.

[0036] S102, the offline modeling module calculates the baseline delay time characteristics between feeding action and oxygen consumption feedback based on filtered historical time series data.

[0037] In high-density cell culture, after the feed pump and other output components execute the feed command, the substrate needs to undergo mixing, mass transfer, and cell uptake before the oxygen uptake rate changes. Therefore, there is a statistically extractable time lag between the feed action and the oxygen uptake rate response.

[0038] Meanwhile, differences may exist in basal metabolic levels, sensor baselines, and culture stages among different batches of cells. To reduce the impact of absolute amplitude differences on hysteresis time recognition, the offline modeling module uses a zero-mean normalized cross-correlation formula to calculate the correlation coefficient between the offline feeding command sequence and the baseline oxygen uptake sequence. The zero-mean normalized cross-correlation formula is as follows:

[0039] in, This indicates that the feed instruction sequence and the baseline oxygen uptake sequence have a time offset of [value missing]. Normalized correlation coefficient at time; Indicates the first One candidate time offset; Indicates the sequence number of the candidate time offset; This indicates the total number of sampling points within the evaluation time window; Indicates the sampling point number within the evaluation time window; This indicates the offline replenishment command sequence is in the first... The values ​​of each sampling point; This represents the average value of the offline replenishment command sequence within the evaluation time window; Indicates the baseline oxygen uptake sequence at the 1st The values ​​of each sampling point; This indicates that the baseline oxygen uptake series has a time offset of 1. The mean value within the corresponding evaluation time window; This is an operator that sums up the samples within the evaluation time window.

[0040] The offline modeling module iterates through each candidate time offset within a preset range and determines the baseline delay time feature based on the normalized correlation coefficient corresponding to each candidate time offset. The offline modeling module uses a baseline delay time feature extraction formula to determine the baseline delay time feature. The baseline delay time feature extraction formula is as follows:

[0041] in, Indicates the characteristics of the reference delay time; The operator for determining the value of the independent variable when the objective function reaches its maximum value; This indicates that the feed instruction sequence and the baseline oxygen uptake sequence have a time offset of [value missing]. Normalized correlation coefficient at time; Indicates the first One candidate time offset.

[0042] S103, the offline modeling module constructs a benchmark feedforward compensation matrix based on the benchmark delay time characteristics.

[0043] Specifically, the offline modeling module associates the baseline delay time characteristics with the corresponding production condition parameters and establishes a parameter mapping relationship between the input and output dimensions. In this embodiment, the input dimensions of the baseline feedforward compensation matrix include the feeding type, the single addition volume of the feeding liquid, the feeding rate, the fermentation culture stage time, and the cell density index obtained from offline testing; the output dimensions of the baseline feedforward compensation matrix include the baseline delay time characteristics, the baseline feedforward start time, the baseline control variable bias amplitude, and the baseline compensation duration.

[0044] Based on the mapping relationship between the input and output dimensions, the offline modeling module stores the corresponding parameters as a reference feedforward compensation matrix, and saves the reference feedforward compensation matrix in the storage area of ​​the industrial control computer. This reference feedforward compensation matrix serves as the basis for subsequent online state perception, feature reconstruction, and feedforward control parameter updates.

[0045] For the construction of the data structure of the reference feedforward compensation matrix, those skilled in the art can use multidimensional lookup tables, grid interpolation or index tables. The specific storage structure and indexing method are conventional techniques in this field and will not be elaborated here.

[0046] See attached document Figure 4 In this embodiment, step S2 is executed collaboratively by the industrial control computer, the programmable logic controller (PLC), and the state awareness module. The industrial control computer sends the reference feedforward compensation matrix obtained in step S1 to the PLC; the state awareness module calculates the current online oxygen uptake rate based on the feedback data from the input and output components, and forms a real-time sliding data window buffer sequence for the current batch. Step S2 specifically includes the following sub-steps.

[0047] S201, the industrial control computer sends the reference feedforward compensation matrix to the programmable logic controller and completes the data block verification.

[0048] In industrial control systems, industrial control computers and programmable logic controllers (PLCs) typically use asynchronous communication for parameter writing and status reading. If the PLC reads incompletely updated data during matrix parameter writing, it may lead to control parameter misalignment. To mitigate this risk, this embodiment employs a double-buffering and verification mechanism to complete the cross-system handover of the reference feedforward compensation matrix.

[0049] Specifically, the programmable logic controller (PLC) has two storage areas with the same data structure in its internal memory, which are defined as the main data block and the spare data block, respectively; the PLC is also configured with an active pointer to indicate the currently valid read area.

[0050] The industrial control computer appends a 32-bit cyclic redundancy check (CRC) code to the reference feedforward compensation matrix and then writes it into the programmable logic controller's (PLC) spare data block via a network control device. During the writing process, the PLC maintains the main data block as the current read object for the main control loop.

[0051] After the spare data block is written, the programmable logic controller (PLC) calculates the local checksum of the received data in the spare data block and compares it with the cyclic redundancy check (CRC) code issued by the industrial control computer. When the comparison results match, the PLC switches the active pointer from the master data block to the spare data block, allowing the new reference feedforward compensation matrix to be connected to the main control loop. The original master data block then becomes the new spare data block for the next matrix parameter update.

[0052] By using the above methods, the risk of control parameter misalignment caused by incomplete parameter writing during cross-system communication can be reduced.

[0053] S202, the state perception module calculates the correction value of the actual exhaust flow rate based on the conservation of inert gas mass.

[0054] In this embodiment, the exhaust gas analyzer is used to detect the volume fraction or mole fraction of oxygen, carbon dioxide, and nitrogen in the reactor exhaust. The actual exhaust flow rate can be directly measured by the exhaust flow meter, or it can be inferred by the state sensing module based on the conservation of inert gas mass.

[0055] As a preferred approach, the state sensing module uses nitrogen as an inert reference gas. Since nitrogen is not typically a major consumed component in routine biological culture metabolism, the state sensing module can calculate the actual exhaust flow rate correction value based on the correspondence between the nitrogen flow rate on the inlet side and the nitrogen mole fraction on the exhaust side.

[0056] Specifically, the state sensing module acquires the total intake flow rate fed back or set by the gas mass flow controller in the output component, and combines it with the known mole fraction of nitrogen in the intake mixture to obtain the intake-side nitrogen flow rate. The state sensing module also acquires the exhaust nitrogen mole fraction detected by the exhaust gas analyzer.

[0057] Considering the influence of exhaust temperature, system back pressure, and water vapor content on the exhaust flow rate calculation, the state sensing module simultaneously acquires the current exhaust temperature, system back pressure, and humidity-related parameters of the reactor, and performs comprehensive temperature, pressure, and humidity correction on the exhaust flow rate estimate obtained based on the inert gas mass conservation rule, based on the ideal gas law or preset standard condition conversion rules.

[0058] The status sensing module calculates the actual exhaust flow correction value using the actual exhaust flow correction formula, which is:

[0059] in, express Correction value for actual exhaust flow rate at any given time; express Total air intake flow rate at any given time; express Mole fraction of nitrogen in the intake air at any given time; express Mole fraction of exhaust nitrogen at time; This represents the zero protection constant for the mole fraction of nitrogen. express The comprehensive correction factor for temperature, pressure, and humidity at any given time; The operator represents the larger of two input values; This represents the correction term for the ratio of intake and exhaust flow rates obtained based on the conservation of inert gas mass. This represents the estimated exhaust flow rate without temperature, pressure, and humidity corrections.

[0060] S203, the status perception module calculates the current online oxygen uptake rate based on the actual exhaust flow correction value and constructs a real-time sliding data window buffer sequence.

[0061] After obtaining the actual exhaust flow correction value, the status sensing module calculates the total oxygen consumption rate based on the difference between the intake oxygen flow rate and the exhaust oxygen flow rate, and obtains the current online oxygen uptake rate per unit volume by combining it with the current culture medium volume.

[0062] The status sensing module calculates the current online oxygen uptake rate using the online oxygen uptake rate calculation formula, which is: in, express The current online oxygen uptake rate at any given moment; express Total air intake flow rate at any given time; express Mole fraction of intake oxygen at any given time; express Correction value for actual exhaust flow rate at any given time; express The mole fraction of exhaust oxygen at any given time; express The volume of culture medium at any given time; express Absolute flow rate of oxygen intake at any given moment; express Absolute flow rate of exhaust oxygen at any given moment; express Total oxygen consumption at any given time.

[0063] The status sensing module writes the calculated current online oxygen uptake rate into the memory queue of the industrial control computer according to the sampling timestamp. The memory queue is a first-in, first-out queue of preset length, the length of which can be set according to the control sampling cycle and process response time; as an optional method, the memory queue can be set to cover 50 to 200 control sampling cycles.

[0064] Specifically, the status awareness module writes the latest current online oxygen uptake value and its timestamp to the tail of the memory queue and deletes the oldest data at the head of the queue to keep the queue length from exceeding a preset length. The continuously updated memory queue constitutes the real-time sliding data window buffer sequence for the current batch.

[0065] The real-time sliding data window buffer sequence is used to characterize the online trend of the current batch's oxygen uptake rate over time, and serves as the data basis for the subsequent steps S3, including composite time series matching evaluation, phase drift calculation, and amplitude residual coefficient calculation.

[0066] See attached document Figure 5 In this embodiment, step S3 is executed by the feature reconstruction module deployed inside the industrial control computer. The feature reconstruction module receives the real-time sliding data window buffer sequence transmitted by the state awareness module, and calls the reference features corresponding to the cultivation stage in the baseline feedforward compensation matrix to evaluate the metabolic response status of the current batch; when the deviation calculation conditions are met, the feature reconstruction module further generates the reconstruction feedforward parameters for the current batch. Step S3 specifically includes the following sub-steps.

[0067] S301, the feature reconstruction module performs a composite time-series matching degree evaluation on the real-time sliding data window buffer sequence.

[0068] During the cultivation of biopharmaceutical products, the metabolic response of the current batch may exhibit temporal shifts and amplitude variations relative to historical baseline batches. Using correlation coefficients alone is insufficient to reflect the difference between the intensity of real-time oxygen uptake fluctuations and the baseline sequence. Therefore, in this embodiment, the feature reconstruction module extracts the real-time sliding data window buffer sequence of the current batch from the memory of the industrial control computer and simultaneously calls the baseline oxygen uptake sequence corresponding to the cultivation stage from the baseline feedforward compensation matrix.

[0069] Calculating the normalized correlation score Previously, the feature reconstruction module pre-detected the variance of the real-time sliding data window buffer sequence and the baseline oxygen uptake sequence. If the difference between either is lower than the set detection limit, it is determined that the current data does not meet the effective temporal matching conditions, and the normalized correlation score is adjusted accordingly. Set to the default low-match value. As a preferred approach, the default low-match value is set to 0 to reduce the risk of erroneously triggering high-match judgments under constant maintenance conditions.

[0070] After confirming that the data meets the calculation conditions, the feature reconstruction module calculates the composite time-series matching degree using the composite time-series matching degree evaluation formula. The composite time-series matching degree evaluation formula is as follows:

[0071] in, Indicates the composite time series matching degree; Indicates the weight coefficient of the relevance feature; This represents the normalized correlation score between the real-time sliding data window buffer sequence and the baseline oxygen uptake rate sequence. This represents the variance offset weighting coefficient; The base of the natural logarithm; Indicates the real-time buffer variance; Represents the variance of the baseline series; This represents the variance zero-protection constant; This represents the absolute difference between the real-time buffer variance and the baseline sequence variance; Indicates the zero-correction reference variance; This indicates a variance offset penalty term; This represents the variance matching factor obtained based on the variance offset penalty term; Represents the correlation-weighted components; This represents the variance-matched weighted components.

[0072] in, The data is obtained by the feature reconstruction module performing normalized correlation calculations on the real-time sliding data window buffer sequence and the baseline oxygen uptake sequence. As an optional approach, the feature reconstruction module maps the maximum correlation coefficient obtained from the normalized correlation calculation to a numerical range of 0 to 1, and uses the mapped result as... For the interval mapping of normalized related scores, those skilled in the art can use linear normalization, threshold truncation, or lookup table mapping methods, which are conventional techniques in this field and will not be elaborated here.

[0073] When the calculated composite time series matching degree is lower than the preset matching threshold, the feature reconstruction module determines that there is a time series or amplitude deviation between the real-time sliding data window buffer sequence of the current batch and the benchmark oxygen uptake rate sequence, and enters the candidate deviation calculation process; when the composite time series matching degree is not lower than the preset matching threshold, the feature reconstruction module maintains the current control parameters in the benchmark feedforward compensation matrix and does not perform spatiotemporal reconstruction.

[0074] S302, the feature reconstruction module calculates the phase drift and amplitude residual coefficients.

[0075] After entering the candidate bias calculation process, the feature reconstruction module first performs time-domain matching between the real-time sliding data window buffer sequence and the baseline oxygen uptake sequence to determine the phase shift of the current batch relative to the baseline sequence. Specifically, the feature reconstruction module uses a cross-correlation sliding search algorithm to gradually shift the real-time sliding data window buffer sequence and the baseline oxygen uptake sequence within a preset candidate sliding range, and calculates the normalized cross-correlation value or inner product similarity value at each candidate sliding time step; then, the sliding time step corresponding to when the similarity value reaches its peak is extracted as the current phase shift.

[0076] After obtaining the phase drift, the feature reconstruction module further calculates the amplitude residual coefficient, which characterizes the deviation of the current batch's metabolic intensity from the baseline sequence. To reduce proportional anomalies caused by sensor transient noise or an excessively small average amplitude of the baseline sequence, the feature reconstruction module uses an anti-singularity amplitude residual coefficient calculation formula to extract the amplitude residual coefficient. The anti-singularity amplitude residual coefficient calculation formula is as follows:

[0077] in, Indicates the amplitude residual coefficient; This represents the upper limit threshold of the amplitude residual coefficient, used to limit the maximum amplification ratio of the reconstructed feedforward parameter relative to the reference feedforward parameter, so as to prevent the output component from receiving a feedforward bias that exceeds the process safety boundary. This represents the lower limit threshold of the amplitude residual coefficient, used to limit the minimum reduction ratio of the reconstructed feedforward parameters relative to the reference feedforward parameters, so as to avoid excessive weakening of the feedforward compensation; Indicates the average amplitude of the real-time sequence; Indicates the average magnitude of the reference sequence; This represents the amplitude singularity protection constant; Indicates the protection correction reference amplitude; Indicates the proportion of the initial amplitude residual; This indicates the percentage of amplitude residual after lower limit protection; This represents the amplitude residual coefficient after upper and lower limit protection.

[0078] S303, the feature reconstruction module generates reconstruction feedforward parameters based on the phase drift and amplitude residual coefficients.

[0079] In this embodiment, the feature reconstruction module uses the phase drift as a translation adjustment factor in the time dimension and the amplitude residual coefficient as a scaling factor in the amplitude dimension to perform spatiotemporal decoupling reconstruction of the reference feedforward parameters in the reference feedforward compensation matrix. The feature reconstruction module generates the reconstruction feedforward parameters for the current batch using the spatiotemporal reconstruction feedforward parameter calculation formula, which is as follows:

[0080] in, This indicates the current time step for reconstructing the feedforward parameters; Indicates the current moment; Indicates the amplitude residual coefficient; Indicates the amount of phase shift; when A positive value indicates that the metabolic response of the current batch lags behind the baseline oxygen uptake sequence; when... A negative value indicates that the metabolic response of the current batch occurs earlier than the baseline oxygen uptake sequence; This represents the reference feedforward parameters after time shifting according to the phase drift amount; This represents the reconstructed feedforward parameters at the current moment after amplitude scaling and time shifting.

[0081] when When the effective time index range of the reference feedforward compensation matrix is ​​exceeded, the feature reconstruction module uses the most recent effective boundary value to maintain it, or keeps the reconstruction feedforward parameters of the previous control cycle unchanged.

[0082] To prevent abrupt changes in the reconstructed feedforward parameters between adjacent control cycles, the feature reconstruction module further applies timing consistency constraints to the reconstructed feedforward parameters. Specifically, the feature reconstruction module calculates the rate of change of the difference between the command value of the current control cycle and the command value of the previous control cycle; when the rate of change of this difference exceeds the preset maximum ramp rate of the corresponding output channel, the feature reconstruction module limits the rate of change of this difference to within the preset maximum ramp rate.

[0083] The preset maximum ramp rate can be determined according to the type of output component; when the output component is a feed pump, the flow rate change rate can be used as the ramp rate; when the output component is a gas mass flow controller, the gas flow rate change rate can be used as the ramp rate; when the output component is a stirring motor frequency converter, the speed change rate or frequency change rate can be used as the ramp rate.

[0084] After the timing consistency constraint processing described above, the feature reconstruction module generates reconstruction feedforward parameters that satisfy the execution boundary of the output component, and sends the reconstruction feedforward parameters to the cooperative control module for amplitude limiting closed-loop control in step S4.

[0085] See attached document Figure 6 In this embodiment, step S4 is executed by the collaborative control module deployed inside the programmable logic controller. The collaborative control module receives the reconstruction feedforward parameters issued by the feature reconstruction module and, in conjunction with the underlying closed-loop feedback instructions, performs amplitude limiting, smoothing, and final constraint processing on the control instructions output to the output components. Step S4 specifically includes the following sub-steps.

[0086] S401, the collaborative control module performs safety boundary limiting processing on the reconstructed feedforward parameters.

[0087] In biopharmaceutical cultivation, if reconstructed feedforward parameters are applied directly to the output components without restriction, the output commands may exceed the allowable range of the corresponding actuator. For example, when the output component is a feed pump, an excessively large instantaneous flow rate increment may lead to increased pump load or abnormal pipeline pressure; when the output component is a gas mass flow controller, an excessively large flow rate increment may cause sudden changes in aeration; and when the output component is a stirrer motor frequency converter, an excessively large speed increment may cause mechanical shock.

[0088] Therefore, in this embodiment, the collaborative control module performs feedforward bias limiting processing on the received reconstructed feedforward parameters based on the process safety boundary of the corresponding output channel and the rated parameters of the actuator.

[0089] As a preferred approach, the cooperative control module pre-configures a lower and upper safety threshold for the feedforward bias in the storage area of ​​the programmable logic controller. The upper safety threshold limits the maximum feedforward compensation received by the output component; the lower safety threshold limits the minimum feedforward compensation received by the output component. When reverse compensation is not allowed for the corresponding output channel, the lower safety threshold can be set to zero.

[0090] The collaborative control module uses a feedforward bias limiting formula to generate the limiting bias. The feedforward bias limiting formula is as follows:

[0091] in, Indicates the current amplitude limit offset; This represents the saturation limiting operator. This indicates the current time step for reconstructing the feedforward parameters; This indicates the lower safety threshold of the feedforward bias; This indicates the upper limit threshold for the feedforward bias safety. This indicates the condition for determining whether the reconstructed feedforward parameters exceed the upper limit threshold of the feedforward bias at the current moment. This indicates the condition for determining whether the reconstructed feedforward parameters are between the lower and upper safety thresholds of the feedforward bias at the current moment. This indicates the condition for determining if the reconstructed feedforward parameter is lower than the feedforward bias safety lower limit threshold at the current moment.

[0092] Through the aforementioned feedforward bias limiting process, the collaborative control module constrains the reconstructed feedforward parameters that exceed the process safety boundary within a preset range, and obtains... It serves as the feedforward compensation input for subsequent closed-loop superposition.

[0093] S402, the cooperative control module smoothly superimposes the limiting bias onto the basic closed-loop feedback instruction and generates the execution-level control instruction at the current moment.

[0094] In this embodiment, the basic closed-loop feedback command is used to adjust the real-time deviations of dissolved oxygen, pH, or other process parameters; the limiting bias is used to compensate for predictable metabolic disturbances. To reduce the abrupt changes in output commands caused by the direct superposition of feedforward biases, the cooperative control module first performs first-order smoothing on the limiting bias before superimposing it on the basic closed-loop feedback command.

[0095] The collaborative control module uses a smoothed feedforward bias calculation formula to calculate the smoothed feedforward bias for the current cycle. The smoothed feedforward bias calculation formula is as follows:

[0096] in, Indicates the current period's smoothed feedforward bias; Indicates the first-order filter smoothing coefficient; Indicates the current amplitude limit offset; This indicates the smoothed feedforward bias of the previous cycle; Indicates the historical attenuation coefficient; Indicates the current limiting bias weighted component; This represents the weighted component of the smoothed feedforward bias from the previous cycle.

[0097] As an optional method, The settings can be adjusted based on the response speed of the output components and the inertia of the process; under normal operating conditions of the bioreactor, It can be set within the range of 0.1 to 0.3. The collaborative control module calculates... Then, it is written into the register of the programmable logic controller as the basis for the next control cycle. Call.

[0098] After obtaining the current period's smoothed feedforward bias, the collaborative control module generates the total control command using the total control command calculation formula. The total control command calculation formula is as follows:

[0099] in, This indicates the current general control command; This indicates the basic closed-loop feedback instruction at the current moment; Indicates the current period's smoothed feedforward bias; This represents the control command resulting from the superposition of the current basic closed-loop feedback command and the current period smooth feedforward bias.

[0100] To prevent the overall control command from exceeding the execution range of the corresponding output component, the collaborative control module further verifies the output before outputting the command. Final limiting is performed. The cooperative control module uses the total output limiting formula to apply the final constraint to the total control command at the current moment. The total output limiting formula is as follows:

[0101] in, Indicates the execution level control command at the current moment; This represents the saturation limiting operator. This indicates the current general control command; This indicates the lower threshold for output component execution; This indicates the upper limit threshold for output component execution.

[0102] The collaborative control module will execute the control commands at the current moment. It is converted into analog signals, digital signals, or industrial communication commands and sent to the corresponding output components to drive the gas mass flow controller, stirring motor frequency converter, feed pump, or acid-base regulating pump to perform corresponding process actions.

[0103] See attached document Figure 7 In this embodiment, step S5 is executed by a cooperative control module deployed within the programmable logic controller. The cooperative control module monitors the feedback signal status of the input components and, upon detecting a disconnection or limit-crossing anomaly, performs feedforward channel blocking and graded degradation backoff based on the source of the fault. Step S5 specifically includes the following sub-steps.

[0104] S501, the collaborative control module detects disconnection and over-limit anomalies in the feedback signals of the input components.

[0105] During operation in a biopharmaceutical field, input components such as dissolved oxygen sensors, pH sensors, and exhaust gas analyzers may generate abnormal signals due to sterilization, vibration, cable aging, or wear and tear on measuring components. To identify these abnormalities before control commands deviate significantly, the collaborative control module runs a hardware daemon process during the background cyclic scanning cycle of the programmable logic controller.

[0106] As a preferred approach, the hardware daemon reads the raw analog current signal transmitted from the input component to the programmable logic controller's underlying I / O channel. The collaborative control module, combining the NAMUR NE43 industrial electrical diagnostic rules with the corresponding input component's instrument range parameters, sets preset low-current diagnostic thresholds and preset high-current diagnostic thresholds.

[0107] When the current signal detected in the input channel is lower than the preset low current diagnostic threshold, the collaborative control module determines that the corresponding input component has experienced a disconnection abnormality; when the current signal detected in the input channel is higher than the preset high current diagnostic threshold, the collaborative control module determines that the corresponding input component has experienced a short circuit abnormality or a hardware fault abnormality. The preset low current diagnostic threshold and the preset high current diagnostic threshold can be determined based on the instrument manual of the input component, the range setting, and the field verification results.

[0108] In addition, the collaborative control module converts analog current signals within the normal current range into corresponding engineering physical quantities and compares these quantities with the preset process range upper and lower limits. When the engineering physical quantity is lower than the preset process range lower limit or higher than the preset process range upper limit, the collaborative control module determines that the corresponding input component has an out-of-limit anomaly.

[0109] When any of the following anomalies are detected: disconnection, short circuit, hardware failure, or exceeding limits, the collaborative control module sets the interlock flag in the system status register to the hardware anomaly interlock state and triggers the subsequent graded downgrade rollback process.

[0110] S502, the collaborative control module performs feedforward channel blocking and graded degradation backoff.

[0111] After the system enters a hardware fault interlock state, the collaborative control module distinguishes between faults in the auxiliary state sensing components and faults in the main control loop input components based on the source of the fault. When the feedforward state sensing component malfunctions but the main control loop input components remain active, the collaborative control module gradually weakens the feedforward channel while maintaining basic closed-loop feedback control. When the main control loop input components malfunction, the collaborative control module freezes the updates of the corresponding basic closed-loop feedback commands and calls the preset safety baseline maintenance value for output.

[0112] To prevent the feedforward bias from being directly reset to zero during an anomaly, which could cause a step change in the output command, the cooperative control module uses a linear decay method to smoothly block the feedforward channel. Specifically, the cooperative control module records the duration after the system enters the hardware anomaly interlock state and uses this duration as the anomaly duration.

[0113] The collaborative control module uses the linear attenuation blocking factor calculation formula to calculate the linear attenuation blocking factor. The formula for calculating the linear attenuation blocking factor is as follows:

[0114] in, express The linear decay blocking factor at time; express The abnormal duration of time; Indicates the smooth decay buffer period; This represents the ratio of the duration of the anomaly to the smooth decay buffer period. This represents the linear decay blocking factor without amplitude limiting; The operator represents the larger of two input values; This represents the linear decay blocking factor after non-negativity constraint.

[0115] After generating the linear attenuation blocking factor, the cooperative control module calculates the attenuation feedforward bias using the attenuation feedforward bias calculation formula, which is as follows:

[0116] in, express The decaying feedforward bias at each time step; express The linear decay blocking factor at time; This represents the current periodic smooth feedforward bias at the moment when the hardware abnormal interlock state is triggered; This represents the feedforward bias scaled by the linear decay blocking factor; This indicates the moment when the hardware abnormal interlock state is triggered.

[0117] If the fault originates from auxiliary status sensing components such as the exhaust gas analyzer, and the main control loop input components such as the dissolved oxygen sensor or pH sensor are still active, the collaborative control module maintains the calculation logic of the basic closed-loop feedback command and superimposes the attenuation feedforward bias onto the basic closed-loop feedback command to generate a degraded control command. When the linear attenuation blocking factor drops to zero, the feedforward channel is blocked, and the system reverts to a single closed-loop feedback control mode.

[0118] If the fault source includes main control loop input components such as dissolved oxygen sensors or pH sensors, the collaborative control module freezes the calculation and update of the corresponding basic closed-loop feedback command and calls the preset safety baseline maintenance value output. As a preferred approach, the preset safety baseline maintenance value may include the preset maintenance speed of the stirring motor, the preset maintenance flow rate of the gas mass flow controller, or the preset low-speed maintenance flow rate of the feed pump.

[0119] Through the above-mentioned graded degradation and fallback processing, the collaborative control module can reduce the risk that erroneous feedforward bias or abnormal closed-loop feedback commands will continue to act on the output components when an abnormality occurs in the input components.

[0120] The following example uses the fed-batch stage of a high-density microbial fermentation process with a nominal working volume of 1000L as a parameter calculation example to illustrate the calculation process of the feature reconstruction module.

[0121] In this example, the state awareness module deployed inside the industrial control computer constructs a real-time sliding data window buffer sequence with a length of 100 sampling periods based on exhaust analysis data. Since uneven substrate mixing or fluctuations in cell metabolic intensity may occur during the feed-in stage, the feature reconstruction module needs to determine whether to reconstruct the baseline feedforward parameters based on the real-time sliding data window buffer sequence.

[0122] Feature reconstruction module extracts real-time buffer variance The value is 1.20, and the reference sequence variance is obtained from the pre-configured reference feedforward compensation matrix. The score was 0.80. Meanwhile, the normalized correlation score between the real-time sliding data window buffer sequence and the baseline oxygen uptake sequence was [missing information]. The calculated value is 0.65. The feature reconstruction module retrieves the system's built-in relevance feature weight coefficients. Variance bias weighting coefficient and the variance zero protection constant .

[0123] The feature reconstruction module uses a composite temporal matching degree evaluation formula to calculate the composite temporal matching degree. The composite temporal matching degree evaluation formula is as follows:

[0124] Substituting the above values ​​into the composite time series matching degree evaluation formula, we get:

[0125] The resulting composite time-series matching degree is 0.637. Since this value is lower than the preset matching threshold of 0.75, the feature reconstruction module determines that there is a time-series or amplitude deviation between the real-time sliding data window buffer sequence of the current batch and the baseline oxygen uptake sequence, and enters the candidate deviation calculation process.

[0126] In the candidate bias calculation process, the feature reconstruction module initiates a cross-correlation sliding search algorithm. The calculation shows that the metabolic response of the current batch lags behind the baseline oxygen uptake sequence by two sampling periods, which is the extracted phase shift. It lasts for 20 seconds.

[0127] To address the bias in the amplitude dimension, the feature reconstruction module extracts the average amplitude of the real-time sequence. The average amplitude of the baseline sequence is 15.5. The value is set to 12.0, and the system's built-in amplitude singularity protection constant is retrieved. Lower limit threshold of amplitude residual coefficient and the upper limit threshold of the amplitude residual coefficient .

[0128] The feature reconstruction module uses the anti-singularity amplitude residual coefficient calculation formula to calculate the amplitude residual coefficient. The anti-singularity amplitude residual coefficient calculation formula is as follows:

[0129] Substituting the above values ​​into the formula for calculating the anti-singularity amplitude residual coefficient, we get:

[0130] The resulting amplitude residual coefficient is 1.2906.

[0131] At the present moment The time-shifted reference control quantity of the corresponding output channel in the reference feedforward compensation matrix The unit is 20.0 engineering units. The engineering unit can be determined according to the type of output component; when the output component is a feed pump, the engineering unit is mL / min; when the output component is a gas mass flow controller, the engineering unit is L / min or standard volumetric flow unit; when the output component is a stirring motor frequency converter, the engineering unit is speed increment or frequency increment.

[0132] In this example, the corresponding output channel is the feed pump. The feature reconstruction module uses the spatiotemporal reconstruction feedforward parameter calculation formula to calculate the reconstruction feedforward parameters at the current moment. The spatiotemporal reconstruction feedforward parameter calculation formula is as follows:

[0133] Substituting the above values ​​into the formula for calculating the feedforward parameters of spatiotemporal reconstruction, we get: ; The reconstructed feedforward parameter obtained at the current moment is 25.81 engineering units. After this parameter is sent to the collaborative control module, it is subjected to feedforward bias limiting processing. Since this parameter does not exceed the set feedforward bias safety upper limit threshold of 40.0 engineering units, it can be used as the current moment limiting bias input, and after smoothing, it is superimposed on the basic closed-loop feedback instruction for feedforward compensation of metabolic deviation in the current batch.

[0134] Furthermore, this embodiment uses a pilot plant comprising three parallel bioreactors as a comparative test example to illustrate the differences in control response under different control architectures. The nominal volume of each of the three reactors is 50L, and they are connected to the same basic process parameters. Under the same basic process parameters, reactor A deploys a conventional PID control algorithm, reactor B deploys a fixed feedforward auxiliary closed-loop control algorithm, and reactor C deploys the intelligent control system and method of this invention. The target cultivation stage is set as a high-density constant-flow yeast feeding period, and the core control objective is to maintain the dissolved oxygen concentration within a preset stable range.

[0135] During the normal operating period from 0 to 20 hours, the system continuously records the dissolved oxygen concentration deviation data of each reactor and converts the deviation data into a probability density distribution for comparing the dissolved oxygen deviation distribution of each control architecture under nonlinear biological metabolic fluctuations.

[0136] At the 21st hour of operation, the current feedback lines of the main control loop sensors (i.e., dissolved oxygen sensors) of the three reactors were disconnected at the bottom control cabinet to simulate a physical disconnection fault in an industrial setting. Under this hardware interlocking condition, the system recorded the control command response phase trajectories of the bottom-level actuators in each reactor; these actuators included the stirring motor and the gas mass flow controller. After data processing, a probability density distribution map of dissolved oxygen control deviation and a phase trajectory map of the bottom-level control command response under the hardware disconnection fault were obtained.

[0137] Figure 8 In the diagram, the horizontal axis represents the percentage error of dissolved oxygen concentration deviating from the set value, and the vertical axis represents the frequency of occurrence of the corresponding probability density.

[0138] Figure 9 In the diagram, the vertical axis represents the rate of change of control commands, and the star-shaped marker represents the system's set safe steady-state point.

[0139] Depend on Figure 8 As shown in the error probability density distribution, under the test conditions, the error distribution corresponding to the control scheme of this embodiment is relatively concentrated. Conventional PID control, which mainly relies on feedback deviation for adjustment, has a relatively flat probability density distribution curve and a wide error distribution range. Fixed feedforward control improves some feedback lag problems by introducing empirical compensation parameters, but fixed compensation parameters are difficult to adapt to the differences in cell activity between batches, and its error distribution center is offset from the zero error baseline. In contrast, the scheme of this invention reconstructs the benchmark feedforward compensation matrix in time and space according to the real-time sliding data window buffer sequence, and superimposes the reconstructed feedforward parameters onto the bottom closed-loop feedback loop, so that the action of the output component can be adjusted according to the current online oxygen uptake rate change trend. Therefore, its error distribution center is closer to the zero error baseline.

[0140] Further integration Figure 9 This section describes the changes in control commands when an input component experiences a disconnection. In the production environment, a physical disconnection of a sensor can cause the current feedback signal to fall below the preset low-current diagnostic threshold. In this situation, the conventional PID control algorithm, due to continuously receiving abnormal deviation signals, may experience integrator saturation, causing the system's phase plane trajectory to deviate from the safe steady-state point, manifesting as a continuous increase in the actuator's amplitude.

[0141] Under the same abnormal hardware input conditions, the collaborative control module, based on a tiered degradation and fallback mechanism, freezes the main loop closed-loop operation instructions after determining that a fault has occurred in the main control loop input component. It then gradually reduces the impact of the feedforward bias on the overall control instructions through a linearly decaying blocking factor, causing the system instructions to transition towards the preset safety baseline maintenance value. Figure 9 It can be seen that, under test conditions, the collaborative control module can reduce the risk of continuously increasing control commands by attenuating the output through the feedforward channel and maintaining the safety baseline.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data-based intelligent control system for biopharmaceutical process parameters, characterized in that, include: The offline modeling module is used to acquire multiple batches of historical time series data and perform zero-phase digital filtering to extract the baseline delay time features and generate the baseline feedforward compensation matrix. The status awareness module is used to calculate the current online oxygen uptake rate and continuously write it into the memory queue according to the sampling timestamp, and generate the real-time sliding data window buffer sequence of the current batch; The feature reconstruction module is used to perform composite time-series matching degree evaluation by combining the benchmark feedforward compensation matrix and the real-time sliding data window buffer sequence, and to calculate and generate reconstruction feedforward parameters when the deviation calculation conditions are met. The collaborative control module is used to perform safety boundary limiting processing on the reconstructed feedforward parameters and generate limiting bias, and to superimpose the limiting bias onto the underlying closed-loop feedback loop and generate a general control command to drive the underlying hardware to perform actions.

2. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The offline modeling module is used to acquire the multiple batches of historical time series data and perform the zero-phase digital filtering process, extract the benchmark delay time features, and generate the benchmark feedforward compensation matrix, specifically including: The zero-phase digital filtering is performed on the multiple batches of historical time series data using a forward and backward superposition filtering method to obtain the final filtered sequence, and the offline feeding instruction sequence and the baseline oxygen uptake rate sequence are extracted. The correlation coefficient between the offline feeding command sequence and the baseline oxygen uptake sequence is calculated using the zero-mean normalized cross-correlation formula, and the baseline delay time feature is extracted using the baseline delay time feature extraction formula. Based on the baseline delay time characteristics, establish the parameter mapping relationship between the input dimension and the output dimension, and generate the baseline feedforward compensation matrix.

3. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The state awareness module is used to calculate the current online oxygen uptake rate and continuously write it into the memory queue according to the sampling timestamp, and generate the real-time sliding data window buffer sequence for the current batch, specifically including: The actual exhaust flow correction value is calculated using the actual exhaust flow correction formula, and the current online oxygen uptake rate per unit volume is calculated using the online oxygen uptake rate calculation formula. The latest current online oxygen uptake rate and its timestamp are written to the tail of the memory queue, and the earliest data at the head of the queue is deleted, so that the data is continuously written to the memory queue according to the sampling timestamp. During the continuous update operation of the memory queue, the real-time sliding data window buffer sequence is generated.

4. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The feature reconstruction module is used to perform a composite time-series matching degree evaluation by combining the benchmark feedforward compensation matrix and the real-time sliding data window buffer sequence, and to determine whether the deviation calculation conditions are met, specifically including: Extract the variance between the real-time sliding data window buffer sequence and the corresponding reference oxygen uptake sequence; When the variance is not lower than the set detection limit, the composite time series matching degree evaluation formula is used, combined with the benchmark feedforward compensation matrix and the real-time sliding data window buffer sequence, to calculate and obtain the composite time series matching degree. The composite timing matching degree is compared with a preset matching threshold, and the deviation calculation condition is determined to be met when the composite timing matching degree is lower than the preset matching threshold.

5. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The feature reconstruction module is used to calculate and generate the reconstruction feedforward parameters when the deviation calculation conditions are met, specifically including: After determining that the deviation calculation conditions are met, the phase drift is extracted using the cross-correlation sliding search algorithm, and the amplitude residual coefficient is extracted using the anti-singularity amplitude residual coefficient calculation formula. The reconstruction feedforward parameters for the current batch are generated by combining the phase drift amount and the amplitude residual coefficient and using the spatiotemporal reconstruction feedforward parameter calculation formula. A timing consistency constraint not exceeding a preset maximum ramp rate is applied to the reconstructed feedforward parameters, and the reconstructed feedforward parameters that satisfy the execution boundary of the output component are generated.

6. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The feature reconstruction module is used to calculate and generate the reconstruction feedforward parameters when the deviation calculation conditions are met, specifically including: After determining that the deviation calculation conditions are met, the phase drift is extracted using the cross-correlation sliding search algorithm, and the amplitude residual coefficient is extracted using the anti-singularity amplitude residual coefficient calculation formula. The reconstruction feedforward parameters for the current batch are generated by combining the phase drift amount and the amplitude residual coefficient and using the spatiotemporal reconstruction feedforward parameter calculation formula. A timing consistency constraint not exceeding a preset maximum ramp rate is applied to the reconstructed feedforward parameters, and the reconstructed feedforward parameters that satisfy the execution boundary of the output component are generated.

7. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The feature reconstruction module is used to calculate and generate the reconstruction feedforward parameters when the deviation calculation conditions are met, specifically including: After determining that the deviation calculation conditions are met, the phase drift is extracted using the cross-correlation sliding search algorithm, and the amplitude residual coefficient is extracted using the anti-singularity amplitude residual coefficient calculation formula. The reconstruction feedforward parameters for the current batch are generated by combining the phase drift amount and the amplitude residual coefficient and using the spatiotemporal reconstruction feedforward parameter calculation formula. A timing consistency constraint not exceeding a preset maximum ramp rate is applied to the reconstructed feedforward parameters, and the reconstructed feedforward parameters that satisfy the execution boundary of the output component are generated.

8. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 7, characterized in that, The collaborative control module is also used to perform graded and degraded responses under the abnormal hardware interlocking state, specifically including: After the interlock flag in the system status register is set to the hardware abnormal interlock state, the abnormality of the auxiliary status sensing component and the abnormality of the main control loop input component are distinguished according to the source of the fault. When the fault source is an abnormality of the auxiliary state sensing component, the attenuation feedforward bias is calculated using the linear attenuation blocking factor calculation formula and the attenuation feedforward bias calculation formula, and the feedforward channel is smoothly blocked. When the fault originates from an abnormality in the main control loop input component, the calculation and update of the corresponding basic closed-loop feedback command are frozen, and the preset safety baseline maintenance value is invoked for output.

9. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 1, characterized in that, The status sensing module is used to calculate the actual exhaust flow correction value using the actual exhaust flow correction formula, and to calculate the current online oxygen uptake rate using the online oxygen uptake rate calculation formula, specifically including: Obtain the molar fraction of nitrogen in the intake air and the molar fraction of nitrogen in the exhaust air, and calculate the intake and exhaust flow rate ratio correction term based on the inert gas mass conservation. The obtained total intake flow rate is multiplied by the intake and exhaust flow rate ratio correction term and the comprehensive correction coefficient for temperature, pressure and humidity, and the actual exhaust flow rate correction value is generated. The total oxygen consumption rate is derived by combining the actual exhaust flow rate correction value, the intake oxygen flow rate, and the exhaust oxygen flow rate, and the current online oxygen uptake rate is calculated by combining the current culture medium volume.

10. The intelligent control system for biopharmaceutical process parameters based on big data according to claim 5, characterized in that, The feature reconstruction module is used to extract the amplitude residual coefficients using the anti-singularity amplitude residual coefficient calculation formula, specifically including: After obtaining the phase drift by performing a cross-correlation sliding search algorithm, the real-time average amplitude of the real-time sliding data window buffer sequence and the average amplitude of the corresponding reference sequence are calculated. An amplitude anti-singularity protection constant is superimposed on the average amplitude of the reference sequence to obtain a protection correction reference amplitude, and the average amplitude of the real-time sequence is divided by the protection correction reference amplitude to generate an initial amplitude residual ratio; The initial amplitude residual ratio is constrained by a lower threshold and an upper threshold of the amplitude residual coefficient, and the final amplitude residual coefficient is extracted and generated.

Images