Online intelligent evaluation modeling method for activity parameters of fermentation strain

Abstract

The invention discloses a fermentation strain activity parameter online intelligent evaluation modeling method, particularly relates to the field of fermentation process data processing, and is used for solving the problem that quantity characterization and a real activity state are easy to misplace and are difficult to identify online in a fermentation process. The method comprises the following steps: constructing a quantity characterization sequence and a process response sequence by acquiring viable bacteria concentration data or a thallus concentration estimated value, acid-base data, feeding data, temperature data and ventilation data of a current fermentation batch, and identifying a dislocation window in combination with a reference interval of a historical normal batch; calculating a reference conversion stable quantity time, a maintenance load conversion time and a surface activity mismatch coefficient, generating a state deviation fragment, a stage activity track and a strain activity parameter evaluation result, and outputting an activity state label and a production disposal prompt according to the result, so as to realize continuous judgment and management expression of internal and external activity deviation of the fermentation stage. And an online evaluation result and a fermentation batch treatment link form a direct corresponding relation.

Description

Technical Field

[0001] This invention relates to the field of fermentation process data processing, and more specifically, to an online intelligent evaluation and modeling method for the activity parameters of fermentation strains. Background Technology

[0002] In biofermentation production, microbial activity is a key factor affecting substrate utilization, metabolic transformation, and process stability. Existing technologies typically determine the fermentation state indirectly by online detection of viable cell concentration or by establishing soft-sensor models based on process parameters. For example, the existing technology "Real-time Online Detection Device for Viable Cell Concentration in Fermentation Broth" (application number: CN200420005114) focuses on real-time online detection of viable cell concentration in the fermentation broth; the existing technology "Online Soft-sensor Method for Cell Concentration in Glutamic Acid Fermentation Process" (application number: CN201310460718.4) estimates cell concentration online by collecting real-time process data during fermentation and establishing a soft-sensor model. These approaches demonstrate that existing technologies can obtain continuous information and achieve online characterization of the fermentation process, but they primarily reflect cell quantity, growth status, or external process changes, and have not yet formed an effective and direct online evaluation system for microbial activity, an intrinsic parameter that better reflects true metabolic capacity.

[0003] However, in actual fermentation processes, strain activity does not necessarily change synchronously with viable cell concentration or cell volume concentration. This is because when cells are affected by changes in culture conditions, fluctuations in mass transfer states, switching of nutrient supply, and changes in metabolic load, they often first exhibit changes in metabolic capacity, environmental responsiveness, and continuous conversion capacity. These changes do not immediately manifest as significant changes in cell count and are difficult to accurately reveal using only conventional process variables. Consequently, while existing technologies can characterize cell growth trends or local process states online, they are prone to assessment lag, judgment bias, and distortion of control criteria when strain activity undergoes latent decline, stage transitions, or state mismatches. Therefore, establishing a method that can distinguish between cell count characterization and strain activity characterization and is applicable to online intelligent assessment and modeling of fermentation processes remains a key technical problem to be solved in existing technologies.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an online intelligent evaluation and modeling method for fermentation strain activity parameters. This method acquires live bacteria concentration data or estimated cell concentration, pH data, feeding data, temperature data, and aeration data for the current fermentation batch. It constructs a quantity characterization sequence and a process response sequence, identifies misalignment windows by combining historical normal batch reference intervals, calculates reference-converted stable quantity time, maintenance load conversion time, and surface activity mismatch coefficient, and generates state deviation segments, stage activity trajectories, and strain activity parameter evaluation results. Based on these, it outputs activity status labels and production handling prompts, enabling continuous judgment and management of activity deviations within and outside the fermentation stage. This establishes a direct correspondence between online evaluation results and fermentation batch handling procedures, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1: Obtain all data of the current fermentation batch, rearrange them according to a unified time index to form a quantitative characterization sequence and a process response sequence, and divide them into continuous time windows corresponding to the quantitative characterization sequence and the process response sequence respectively; S2: Read the continuous time windows one by one, map the historical normal batch reference interval of the fermentation stage to the quantity characterization sequence and process response sequence, identify the misalignment window where the process response sequence falls outside the reference interval and the quantity characterization sequence falls within the reference interval, generate the reference conversion steady quantity, the maintenance load conversion time and the liveness mismatch coefficient, and connect them to obtain the state deviation segment. S3: Sort the state deviation segments according to a unified time index, call the reference segments of the corresponding fermentation stage, and align the quantity characterization sequence, process response sequence and surfactant mismatch coefficient in the state deviation segments hourly to determine the deviation start point, deviation duration range and deviation end point, form the stage activity trajectory and determine the strain activity parameter evaluation results. S4: Read the strain activity parameter evaluation results and stage activity trajectory, extract the trajectory connection positions between adjacent fermentation stages and the deviation distribution results within each fermentation stage, generate activity state change results, and map the strain activity parameter evaluation results and activity state change results into activity state tags and production disposal prompts.

[0007] Furthermore, a unique source of quantitative data is first determined between the viable bacteria concentration data and the estimated cell concentration. Then, missing timestamps and missing measurements are removed from the viable bacteria concentration data or estimated cell concentration data, acid-base data, feeding data, temperature data, and ventilation data. When the same timestamp is repeated, the last written item is retained. Subsequently, a unified time index is generated based on the common valid observation interval and fixed sampling interval.

[0008] Furthermore, piecewise linear interpolation is performed on the viable bacteria concentration data or estimated cell concentration, acid-base data, and temperature data along a unified time index. Zero-order preservation is performed on the feeding data and aeration data. After removing the observation breakpoints, continuous index segments are formed. Then, continuous time windows are generated from within each continuous index segment according to the preset window length and the rules for the intersection of adjacent windows.

[0009] Furthermore, each continuous time window is compared with the time range of each fermentation stage one by one. The stage mapping window is determined according to the fermentation stage with the longest overlapping duration. Then, the historical normal batch processed in step S1 is called up, and the quantity reference interval and the process reference intervals of acid-base, feeding, temperature and ventilation are generated at the unified time point covered by the stage mapping window. The reference intervals are then mapped to the quantity characterization sequence and the process response sequence.

[0010] Furthermore, within the stage mapping window, first identify quantity bound segments whose duration reaches the quantity bound duration threshold, then identify process deviation segments whose duration reaches the process deviation duration threshold, and determine the portion of the overlapping time period between quantity bound segments and process deviation segments whose duration reaches the misalignment overlap duration threshold as misalignment windows, and determine the quantity bound time period within the misalignment window as stable quantity segments.

[0011] Furthermore, the product of the quantity value accumulated point by point along the stable quantity segment and the fixed sampling interval is used to obtain the stable quantity load time product. The average result of the median value of the historical normal batch quantity at the same time point corresponding to the aforementioned stable quantity segment is used as the reference quantity benchmark, and the stable quantity load time product is converted into the reference converted stable quantity time. At the same time, the process channels within the stable quantity segment that fall outside the process reference interval are counted point by point and accumulated to obtain the maintenance load converted time. Then, the liveness mismatch coefficient is output by the dual-channel mutual verification learning model.

[0012] Furthermore, the state deviation segments are first sorted according to a unified time index, and the liveness mismatch coefficients of the effective misalignment windows that make up the state deviation segments are spread into a time-spread liveness mismatch sequence. Then, reference segments covering the same unified time point set are called from the historical normal batches. The reference segments include the upper and lower bounds of quantity reference, the upper and lower bounds of process reference, and the upper bound of liveness mismatch reference. They are then aligned with the quantity characterization sequence, the process response sequence, and the time-spread liveness mismatch sequence hourly at the same unified time point.

[0013] Furthermore, based on quantity deviation, process deviation, and coefficient deviation, trajectory deviation identifiers are generated. Continuous trajectory deviation segments whose duration reaches the deviation confirmation duration threshold are identified as initial candidate deviation segments. Adjacent initial candidate deviation segments with an interval not exceeding the deviation allowable gap threshold are then connected together. The candidate deviation segment with the largest cumulative coefficient is used to determine the deviation start point, deviation duration range, and deviation end point.

[0014] Furthermore, the time interval between adjacent deviation durations within the same fermentation stage that does not exceed the stage concatenation interval is concatenated to form a stage activity trajectory fragment, and a complete stage activity trajectory is formed according to the fermentation stage sequence; at the same time, the deviation coverage time ratio, the longest continuous deviation time, and the stage peak coefficient are extracted in each fermentation stage, and the deviation coverage time ratio, the longest continuous deviation time, and the stage peak coefficient are used to form the strain activity parameter evaluation results.

[0015] Furthermore, the results of the change in active state are determined based on the trajectory connection position and connection judgment threshold between adjacent fermentation stages. The active stage code of each fermentation stage is determined based on the coverage judgment threshold, the continuity judgment threshold, and the peak judgment threshold. The first stage number and the main migration type are extracted from the results of the change in active state and combined with the active stage code of the last stage to form a ternary index, which is uniquely mapped to the active state label and production disposal prompt according to the preset rule set.

[0016] The technical effects and advantages of the online intelligent evaluation and modeling method for the activity parameters of fermentation strains of this invention are as follows: This invention uses a unified time index to connect viable cell concentration data or cell concentration estimates with acid-base data, feeding data, temperature data, and aeration data. It incorporates the quantitative characterization and process response during fermentation into the same temporal framework. Then, through staggered window identification, reference-converted steady-state time, maintenance load conversion time, and progressive processing using surfactant mismatch coefficients, it generates state deviation segments, stage activity trajectories, and strain activity parameter evaluation results. Therefore, the online evaluation of fermentation strain activity parameters is no longer limited to a single characterization of viable cell quantity or cell concentration, but can identify and characterize continuous deviations within the fermentation stage, making the evaluation object closer to the actual changes in the strain's activity state.

[0017] Furthermore, this invention maps the stage activity trajectory and strain activity parameter evaluation results into activity status labels and production disposal prompts, ensuring consistency between front-end data processing results and back-end management expressions, forming a complete chain from data input and activity determination to batch disposal output. Compared to processing methods that only perform online detection, single soft measurement, or general status monitoring, this invention emphasizes the misalignment between quantitative representation and process response in its processing logic, and continuously summarizes activity deviations using fermentation stages as organizational units. Therefore, it is more suitable for batch-level online evaluation, stage-level status identification, and disposal decision expression during the fermentation process. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the online intelligent evaluation and modeling method for the activity parameters of fermentation strains according to the present invention. Figure 2 This is a schematic diagram illustrating the unified time index and continuous time window generation for multi-source data in this invention; Figure 3 This is a schematic diagram illustrating the fermentation stage mapping, historical normal batch reference interval, and misalignment window identification of the present invention. Figure 4 This is a schematic diagram illustrating the generation of surfactant mismatch coefficients and the construction of staged activity trajectories in this invention. Figure 5 This is a schematic diagram illustrating the generation of the results of changes in the active state, the active state label, and the production and disposal prompts of the present invention. Figure 6 This is a schematic diagram illustrating the refinement of the state deviation fragment boundary and the determination of the deviation duration range during the fermentation stage of this invention. Detailed Implementation

[0019] The technical solutions of 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.

[0020] Please see Figures 1-6 The present invention provides an online intelligent evaluation and modeling method for the activity parameters of fermentation strains, including steps S1 to S4.

[0021] Step S1: Obtain all data of the current fermentation batch, rearrange them according to a unified time index to form a quantitative characterization sequence and a process response sequence, and divide them into continuous time windows corresponding to the quantitative characterization sequence and the process response sequence respectively.

[0022] Step S101: Determining the source of quantity data and classifying original records.

[0023] When the current fermentation batch enters the online evaluation process for strain activity parameters, it first receives live bacteria concentration data, estimated cell concentration, pH data, feed data, temperature data, and aeration data. The source of quantity data can only be either live bacteria concentration data or estimated cell concentration. If live bacteria concentration data exists for the current fermentation batch, it will be designated as the quantity data stream; if estimated cell concentration data exists but not live bacteria concentration data, it will be designated as the quantity data stream. The source of quantity data remains unchanged within the same current fermentation batch, and switching between live bacteria concentration data and estimated cell concentration is not allowed. If the current fermentation batch has neither live bacteria concentration data nor estimated cell concentration data, it will not enter the online evaluation process for strain activity parameters, and a message indicating missing quantity data will be output. Acid-base data is defined as an acid-base data stream, feed data as a feed data stream, temperature data as a temperature data stream, and ventilation data as a ventilation data stream. Each of the aforementioned data streams consists of a timestamp and a measured value. The timestamp indicates the time when the record occurred, and the measured value indicates the process quantity at the corresponding time.

[0024] Step S102: Cleaning up the original records and determining the common valid observation interval.

[0025] A valid timestamp refers to a record time that simultaneously meets the following conditions: the time field is not empty and can be parsed into a uniform time format; the time is between the start and end times of the current fermentation batch; the corresponding measurement value is not empty; the corresponding measurement value passes the range validity check of the channel; if duplicate timestamps exist in the same data stream, only the timestamp corresponding to the last written record is retained as the valid timestamp for that moment.

[0026] Each original record in the quantity, acid-base, feed, temperature, and aeration data streams is checked. Records with missing timestamps, missing measurements, or failed range verification are deleted. If multiple records with the same timestamp appear in the same data stream, only the last record written is retained according to the writing order. The original records of each data stream are then reordered from earliest to latest timestamp. After sorting, the first and last valid timestamps of each data stream are extracted. The latest time among the five first valid timestamps is determined as the common valid observation start point, and the earliest time among the five last valid timestamps is determined as the common valid observation end point. Only when the common valid observation start point is earlier than the common valid observation end point can the quantity, acid-base, feed, temperature, and aeration data streams have a common valid observation interval that can be processed in parallel. If the common valid observation start point is not earlier than the common valid observation end point, the current fermentation batch will not enter the subsequent unified time index generation process.

[0027] Step S103: Candidate unified time index generation and observation fault segment identification.

[0028] Given that the common effective observation interval has been determined, the fixed sampling interval is first determined. The fixed sampling interval is preferentially adopted from the basic sampling period of the fermentation control system; when the original sampling periods of each data stream are inconsistent, a common time interval that can be stably supported by most critical channels and is not higher than the sampling period of the slowest critical control quantity is selected as the fixed sampling interval. In one embodiment, when the control system refreshes the main process quantities every minute, the fixed sampling interval can be one minute.

[0029] Using a fixed sampling interval as the step size, candidate unified time indices are generated by recursively pushing from the common valid observation start point to the common valid observation end point. Each time point in the candidate unified time index is consistent with the fixed sampling interval. Then, a missing span threshold is introduced, representing the maximum time span allowed for interpolation or preservation between two adjacent valid raw records, preferably three to five times the fixed sampling interval; for example, when the fixed sampling interval is one minute, the missing span threshold can be three or five minutes. Adjacent raw records in each data stream are checked. If the time span between two adjacent raw records exceeds the missing span threshold, the entire time between the timestamps of the previous and next raw records is identified as an observation break segment. All observation break segments identified in the quantity data stream, acid-base data stream, feed data stream, temperature data stream, and ventilation data stream are then aggregated into a break segment set. Finally, time points falling into the break segment set are removed from the candidate unified time index to obtain the unified time index.

[0030] Step S104: Classification resampling of multi-source records.

[0031] After the unified time index is generated, piecewise linear interpolation is performed on the quantity data stream, acid-base data stream, and temperature data stream. Since these three types of data are continuous state quantities, at any unified time point between two adjacent valid original records, their resampled values ​​are obtained according to the following rules: First, calculate the difference between the measurement value of the later original record and the measurement value of the earlier original record. Then, divide the difference by the time span between the timestamps of the two original records to obtain the change per unit time. Next, multiply the change per unit time by the time interval between the unified time point and the timestamp of the earlier original record. Finally, add the measurement value of the earlier original record. The result is used as the resampled value at the unified time point. Zero-order hold is performed on the feed data stream and ventilation data stream. Since these two types of data are control hold quantities, at any unified time point between two adjacent valid original records, their resampled values ​​are directly taken from the measurement value of the earlier valid original record until the later valid original record arrives.

[0032] Step S105: Construction of quantitative characterization sequence and process response sequence.

[0033] For each time point in the unified time index, the resampled values ​​of the quantity data stream at that time point are read and arranged in chronological order to form a quantity representation sequence. Simultaneously, at the same unified time point, the resampled values ​​of the acid-base data stream, feed data stream, temperature data stream, and ventilation data stream are read and written into the same process response vector according to the fixed positional order of acid-base, feed, temperature, and ventilation. All process response vectors are then arranged from earliest to latest according to the unified time index to form a process response sequence.

[0034] Step S106: Continuous index segment division and continuous time window generation.

[0035] The unified time index may contain multiple unconnected time blocks due to the removal of observation break segments. Therefore, we continue to check whether the time difference between two adjacent unified time points is equal to the fixed sampling interval, and determine all the longest time point sequences that satisfy the condition that the adjacent time difference is always equal to the fixed sampling interval as continuous index segments.

[0036] Building upon this, window length and handover duration are introduced. Window length represents the duration covered by a single continuous time window, determined based on the shortest duration required for the deviation of bacterial activity from an observable trend on the process side, and is typically not shorter than a complete control response cycle. Handover duration represents the duration shared by two adjacent continuous time windows, and is typically one-third to one-half of the window length. For example, when the fixed sampling interval is one minute and the process response typically forms a stable trend after more than fifteen minutes, the window length can be twenty to thirty minutes, and the handover duration can be five to ten minutes.

[0037] Subsequently, within each consecutive index segment, the first consecutive time window is generated starting from the first unified time point. Thereafter, subsequent consecutive time windows are generated sequentially, using the difference between the number of unified time points contained in a single consecutive time window and the number of shared unified time points as the step size. If the number of unified time points remaining at the end of a consecutive index segment is less than the number required for a single consecutive time window, no new consecutive time window is generated. Each consecutive time window simultaneously corresponds to a quantity representation sequence segment and a process response sequence segment, serving as the direct input for the window-by-window reading in step S2.

[0038] After step S1, the multi-source records related to strain activity in the current fermentation batch have been time-aligned, object-classified, and continuously segmented, forming a unified time index, quantity characterization sequence, process response sequence, and continuous time window, providing a unified input for identifying abnormal disconnections during the fermentation stage.

[0039] Step S2: Read the continuous time windows one by one, map the historical normal batch reference interval of the fermentation stage to the quantity characterization sequence and process response sequence, identify the misalignment window where the process response sequence falls outside the reference interval and the quantity characterization sequence falls within the reference interval, generate the reference conversion stable quantity, the maintenance load conversion quantity and the liveness mismatch coefficient, and connect them to obtain the state deviation segment.

[0040] Step S201: Fermentation stage mapping and historical normal batch reference range generation.

[0041] First, each continuous time window is compared with the time range of each fermentation stage predefined in the process document. The time range of each fermentation stage is pre-set based on the process specification, batch record, or formula control document corresponding to the current batch, and is bound and saved with the strain type, process version, and equipment specifications; the stage boundary is based on the stage switching time point in the production control system. In one embodiment, it can be divided into the inoculation adaptation stage, rapid growth stage, fed-to-stabilize production stage, and post-ripening stage, with the boundary of each stage based on the switching time point recorded in the control system.

[0042] The overlap duration between the continuous time window and the time range of each fermentation stage is calculated. The fermentation stage with the longest overlap duration is then determined as the mapped fermentation stage of the continuous time window, and the overlap portion between the continuous time window and the time range of the mapped fermentation stage is determined as the stage mapping window. A window validity duration threshold is then introduced. This threshold represents the minimum duration required for the stage mapping window to enter subsequent analysis. It is determined based on the minimum analysis length that the continuous time window must cover within a single fermentation stage, preferably at least half the length of a single continuous time window, and not less than the duration corresponding to three consecutive uniform time points. For example, if the window length is 20 minutes and the fixed sampling interval is one minute, the window validity duration threshold can be 10 minutes. When the stage mapping window duration reaches the window validity duration threshold, the stage mapping window enters subsequent analysis; when the stage mapping window duration does not reach the window validity duration threshold, the continuous time window no longer participates in the subsequent processing in step S2.

[0043] After completing the fermentation stage mapping, the historical normal batches corresponding to the mapped fermentation stage are invoked. Within the mapped fermentation stage, for each unified time point, all quantity values ​​of the historical normal batches in the quantity representation sequence, as well as all process values ​​of the historical normal batches in the acid-base channel, feeding channel, temperature channel, and aeration channel, are collected. Then, based on pre-set quantile positions, the following reference bounds are extracted from the historical normal batch value set: lower bound for quantity, upper bound for quantity, lower bound for acid-base, upper bound for acid-base, lower bound for feeding, upper bound for feeding, lower bound for temperature, upper bound for temperature, lower bound for aeration, and upper bound for aeration, thus forming the historical normal batch reference interval. Quantile positions are preferentially obtained through low and high quantile positions; when the number of historical normal batches is sufficient, the 10th and 90th percentiles can be used respectively; when the number of historical normal batches is insufficient, empirical upper and lower bounds confirmed by the process engineer can be used and appropriately relaxed. The aforementioned historical normal batch reference interval is mapped point-by-point to the quantity representation sequence and process response sequence in the stage mapping window according to a unified time index, serving as the comparison boundary for subsequent misalignment window identification.

[0044] Step S202: Identification of quantity-bound segments, process deviation segments, and misaligned windows.

[0045] For each unified time point within the stage mapping window, first read the quantity value corresponding to the quantity representation sequence, then compare the aforementioned quantity value with the lower and upper bounds of the quantity reference at the same unified time point. When the quantity value is between the lower and upper bounds of the quantity reference, the unified time point is recorded as the quantity within the bounds time point; when the quantity value falls outside the quantity reference range, the unified time point is recorded as the quantity outside the bounds time point. Simultaneously, read the acid-base value, feed value, temperature value, and ventilation value of the process response sequence at the same unified time point, and compare them with the acid-base reference range, feed reference range, temperature reference range, and ventilation reference range, respectively. If at least one of the four process values ​​falls outside its corresponding reference range, the unified time point is recorded as the process deviation time point; if all four process values ​​are within their respective reference ranges, the unified time point is recorded as the process within the bounds time point.

[0046] Subsequently, thresholds for the duration of quantity within limits, the duration of process deviation, and the duration of misalignment overlap are introduced. The duration threshold for the duration of quantity within limits is determined based on the duration of short-term jitter in the quantity channel under normal conditions, and is generally not less than the duration corresponding to three consecutive unified time points. The duration threshold for the duration of process deviation is determined based on the boundary duration between instantaneous disturbance and actual deviation in the process channel, prioritizing the duration required for a single actuator action or control response to complete. The duration threshold for misalignment overlap is determined based on the minimum duration required to confirm that there is indeed a time overlap between the quantity steady state and the process deviation, and is generally not higher than the smaller of the aforementioned two duration thresholds, and must cover at least two consecutive unified time points.

[0047] Under the aforementioned rules, the longest consecutive quantity-bounded time period whose duration reaches the quantity-bounded duration threshold is defined as the quantity-bounded segment, and the longest consecutive process deviation time period whose duration reaches the process deviation duration threshold is defined as the effective process deviation segment. Then, the quantity-bounded segments and effective process deviation segments are overlapped segment by segment. Any time period whose overlap duration reaches the misalignment overlap duration threshold is defined as a misalignment window. Since the misalignment window itself is limited to consecutive overlapping time periods where the quantity representation sequence remains within the quantity reference interval while the process response sequence deviates from the process reference interval, the continuous quantity-bounded time range within the misalignment window is directly defined as a stable quantity segment.

[0048] Step S203: Calculation of stable load-bearing time product, reference quantity benchmark and reference converted stable load-bearing time product.

[0049] For each stable quantity segment, the quantity values ​​in the quantity representation sequence are read point-by-point along a unified time index. The quantity value at each unified time point is multiplied by the duration corresponding to a fixed sampling interval. Then, all products within the stable quantity segment are summed to obtain the stable quantity carrying capacity time product. Subsequently, a reference quantity benchmark is generated. Specifically, at each unified time point covered by the stable quantity segment, the quantity values ​​corresponding to all historical normal batches within the mapped fermentation stage are extracted. The median quantity value is taken from the set of quantity values, and then the median quantity values ​​at all unified time points of the stable quantity segment are averaged to obtain the reference quantity benchmark. After completing the above two processes, the quantity-time area corresponding to the stable quantity carrying capacity time product is converted into the equivalent duration corresponding to the reference quantity benchmark level to obtain the reference converted stable quantity time. The reference converted stable quantity time is used to uniformly convert stable quantity segments from different batches and different quantity scales to the same time scale. If the reference quantity benchmark is lower than the preset minimum positive value, the subsequent conversion will not be performed. Instead, the stable quantity segment will be recorded as a segment with insufficient quantity benchmark and will be excluded from the subsequent model discrimination, so as to avoid conversion distortion caused by extremely low benchmark.

[0050] Step S204: Calculate the reference occupancy identifier, dimensionless maintenance load sequence, and maintenance load conversion time.

[0051] For each uniform time point within each steady-state segment, the acid / base value is compared with the acid / base reference range, the feed value with the feed reference range, the temperature value with the temperature reference range, and the ventilation value with the ventilation reference range. For any process channel, if the process value falls outside the corresponding reference range, the process channel is recorded as having a valid reference occupancy at that uniform time point; otherwise, it is recorded as having a invalid reference occupancy. The reference occupancy results for the acid / base, feed, temperature, and ventilation channels at the same uniform time point are then directly counted to obtain the dimensionless maintenance load sequence. The dimensionless maintenance load sequence is then read point-by-point along the steady-state segment. The dimensionless maintenance load value at each uniform time point is multiplied by the duration corresponding to the fixed sampling interval. All products within the steady-state segment are then summed to obtain the maintenance load converted to time. The maintenance load converted to time represents the deviation support of how many channel durations the process side has cumulatively occupied within the steady-state segment in order to maintain the apparent stability of the quantitative characterization sequence.

[0052] Step S205: Dual-channel mutual verification learning, generation of expression mismatch coefficients and connection of state deviation segments.

[0053] In one embodiment, the training of the dual-channel mutual verification learning model is completed offline before the current fermentation batch enters online evaluation. In the offline phase, historical normal and abnormal batches are first screened from historical fermentation batches. Then, using stable volume segments as sample units, the reference-converted stable volume time and maintenance load time corresponding to each sample are calculated. Based on historical sample labels, one-dimensional monotonic mappings are established from the reference-converted stable volume time to the stable volume evidence value, and from the maintenance load time to the load evidence value. These one-dimensional monotonic mappings are established using order-preserving regression to ensure that the output risk evidence value is not lower than the risk evidence value corresponding to a lower original indicator when the original indicator is higher. Subsequently, the stable volume evidence value and the load evidence value are arranged in a fixed order to form a two-dimensional evidence vector, which is used to train the liveness mismatch discrimination model.

[0054] In one embodiment, the active mismatch discrimination model employs an extreme random tree classifier. The construction of the extreme random tree classifier can be accomplished as follows: First, all two-dimensional evidence vectors are grouped by batch, ensuring that samples from the same historical fermentation batch remain in the same data subset. Then, the data is divided into training, validation, and test sets according to a preset ratio. When the number of normal samples differs from the number of abnormal samples, a combination of stratified undersampling and repeated sampling is used in the training set to maintain a balance in the batch size of the two types of samples, while the validation and test sets maintain their original distribution. Model parameters can be selected from candidate combinations of candidate tree number, candidate maximum tree depth, minimum number of split samples in internal nodes, minimum number of samples in leaf nodes, and number of random split points. The selection criteria are then based on the ability to identify abnormal samples on the validation set, the ability to pass normal samples, and output stability. Finally, the model structure is determined and stored.

[0055] When the current fermentation batch enters online evaluation, the two-dimensional evidence vector corresponding to the current stable fraction is input into an extreme random tree classifier that has been trained and stored. Each discriminant tree provides the proportion of abnormal samples based on the leaf node where the two-dimensional evidence vector falls. The arithmetic mean of the proportions of abnormal samples given by all discriminant trees is then taken to obtain the surfactant mismatch coefficient. The surfactant mismatch coefficient is between zero and one. The closer the value is to one, the closer the current stable fraction is to the distribution of apparent quantity stability and actual activity mismatch in historical abnormal batches.

[0056] Subsequently, a retention threshold is introduced, determined on the validation samples. The threshold is chosen as the smallest surfactant mismatch coefficient boundary point that simultaneously satisfies both the abnormal sample detection requirement and the normal sample pass requirement. When the surfactant mismatch coefficient corresponding to a misaligned window reaches the retention threshold, the corresponding misaligned window is retained as a valid misaligned window. A connection interval is then introduced, determined based on the typical interval distribution of adjacent valid misaligned windows within the same fermentation stage in historical abnormal batches, preferentially using high quantile interval values ​​or two to five times the fixed sampling interval. As long as the time interval between two consecutive valid misaligned windows does not exceed the connection interval, the two consecutive valid misaligned windows are connected into the same state deviation segment according to a unified time index order. The state deviation segment serves as the direct input object for the sorting, time-by-time alignment, and stage activity trajectory concatenation in step S3.

[0057] After step S2, general process fluctuations in the continuous time window have been filtered out, leaving only the time region where process deviations and quantity steady states coexist and exhibit surface activity mismatch characteristics. The time region is further organized into state deviation segments.

[0058] Step S3: Sort the state deviation fragments according to a unified time index, call the reference fragments of the corresponding fermentation stage, and align the quantity characterization sequence, process response sequence and surfactant mismatch coefficient in the state deviation fragments hourly to determine the deviation start point, deviation duration range and deviation end point, form the stage activity trajectory and determine the strain activity parameter evaluation results.

[0059] Step S301: Sorting of state deviation segments and generation of time-spreading live mismatch sequence.

[0060] First, read all state deviation segments output in step S2 and categorize them according to their respective fermentation stages. Then, within the same fermentation stage, sort them according to the order of their starting unified time points in the unified time index. Next, extract all valid misalignment windows and their corresponding surfactant mismatch coefficients for each state deviation segment. Then, spread the surfactant mismatch coefficients in the valid misalignment windows along the unified time index into the state deviation segment. Specifically, for each unified time point within the state deviation segment, find all valid misalignment windows covering that unified time point. If the unified time point is covered by only one valid misalignment window, determine the surfactant mismatch coefficient of that valid misalignment window as the time-spread surfactant mismatch value for that unified time point. If the unified time point is covered by multiple valid misalignment windows, determine the maximum value among the surfactant mismatch coefficients corresponding to the multiple valid misalignment windows as the time-spread surfactant mismatch value for that unified time point, thus forming a time-spread surfactant mismatch sequence.

[0061] Step S302: Reference segment generation and time-by-time alignment reference establishment.

[0062] For each state deviation segment, records from historical normal batches that are in the same fermentation stage and cover the same set of unified time points are retrieved to generate a reference segment. The reference segment consists of the following references: quantity lower bound trajectory, quantity upper bound trajectory, acid-base lower bound trajectory, acid-base upper bound trajectory, feed lower bound trajectory, feed upper bound trajectory, temperature lower bound trajectory, temperature upper bound trajectory, aeration lower bound trajectory, aeration upper bound trajectory, and surfactant mismatch reference upper bound trajectory. The generation method is as follows: at each unified time point covered by the state deviation segment, collect the quantity, acid-base, feed, temperature, and aeration values ​​from all historical normal batches at the same fermentation stage and unified time point, as well as the time-spread surfactant mismatch values ​​obtained after steps S1 and S2. Then, extract the reference lower bound, reference upper bound, and surfactant mismatch reference upper bound based on pre-set quantile positions. The upper bound trajectory for surfactant mismatch reference is derived from the high quantile boundary of the set of surfactant mismatch values ​​spread over time at the same fermentation stage and at the same unified time point in historical normal batches. Since the reference fragment and the state deviation fragment cover the exact same unified time point set, subsequent hourly alignment does not involve time scaling, but instead directly compares the quantity state, process state, and surfactant mismatch state of the reference fragments of the current fermentation batch with those of historical normal batches at the same unified time point.

[0063] Step S303: Determine the deviation start point, the deviation duration range, and the deviation end point.

[0064] For each unified time point within the state deviation segment, first compare the quantity value in the quantity representation sequence with the lower and upper bounds of the quantity reference at the same unified time point. If the quantity value falls outside the lower or upper bound of the quantity reference, the unified time point is recorded as a valid quantity deviation. Then, compare the acid-base value, feed value, temperature value, and ventilation value in the process response sequence with the acid-base reference interval, feed reference interval, temperature reference interval, and ventilation reference interval, respectively. If at least one of the four process values ​​falls outside the corresponding reference interval, the unified time point is recorded as a valid process deviation. Simultaneously, compare the time-spreading surfactant mismatch value at the unified time point with the upper bound of surfactant mismatch reference. If the time-spreading surfactant mismatch value is higher than the upper bound of surfactant mismatch reference, the unified time point is recorded as a valid coefficient deviation.

[0065] Subsequently, trajectory deviation indicators are generated. The generation rule is as follows: only when the coefficient deviation is true, and at least one of the quantity deviation and process deviation is true, is the unified time point recorded as a valid trajectory deviation; otherwise, it is recorded as a invalid trajectory deviation. Next, a deviation confirmation duration threshold and a deviation tolerance gap threshold are introduced. The deviation confirmation duration threshold is determined based on the longest duration of occasional short-term deviations in historical normal batches, and is moderately increased to avoid classifying short-term noise as valid deviations. The deviation tolerance gap threshold is determined based on the typical discontinuity length of brief pullbacks within the same deviation event, preferably one to three times the fixed sampling interval. For example, when similar short-term pullbacks in historical normal batches typically do not exceed two minutes, the deviation tolerance gap threshold can be set to two minutes.

[0066] Under this rule, the time intervals in which trajectory deviations are continuously established and the duration reaches the deviation confirmation duration threshold are first determined as initial candidate deviation segments. Then, the time interval between two adjacent initial candidate deviation segments is compared with the deviation tolerance gap threshold. As long as the time interval does not exceed the deviation tolerance gap threshold, the two initial candidate deviation segments are concatenated end-to-end as the same concatenated candidate deviation segment. If there are multiple concatenated candidate deviation segments within the same deviation segment, the cumulative coefficient of each concatenated candidate deviation segment is calculated. Finally, the concatenated candidate deviation segment with the largest cumulative coefficient is determined as the deviation duration range, and the first unified time point of this deviation duration range is determined as the deviation start point, and the last unified time point of this deviation duration range is determined as the deviation end point.

[0067] Step S304: Construction of stage activity trajectory and complete stage activity trajectory.

[0068] All deviations within the same fermentation stage are sorted in chronological order of their starting points, and a stage concatenation interval is introduced. The stage concatenation interval is determined based on the maximum pause duration allowed between adjacent deviations within the same fermentation stage, typically not exceeding one-tenth of the total stage duration. As long as the time interval between the starting point of the subsequent deviation and the ending point of the preceding deviation does not exceed the stage concatenation interval, the two deviations are concatenated end-to-end to form a single active trajectory segment within the same stage. After concatenation, one or more active trajectory segments are formed within the same fermentation stage. All these active trajectory segments are then arranged chronologically to form the stage active trajectory for the corresponding fermentation stage. Subsequently, the stage active trajectories for each fermentation stage are arranged sequentially according to their chronological order to form a complete stage active trajectory.

[0069] To ensure that subsequent distribution statistics directly revolve around the stage activity trajectory, a stage activity coefficient sequence is established within each fermentation stage. The establishment method is as follows: for each uniform time point covered by the fermentation stage, if the uniform time point falls within a certain deviation range, its value in the time-spreading surfactant mismatch sequence is written into the stage activity coefficient sequence; if the uniform time point does not fall within any deviation range, its value in the stage activity coefficient sequence is recorded as zero. This ensures that the stage activity coefficient sequence corresponds point-by-point to the stage activity trajectory.

[0070] Step S305: Extraction of off-distribution results and generation of strain activity parameter evaluation results.

[0071] For each fermentation stage, the deviation distribution results are extracted from the stage activity trajectory. These deviation distribution results consist of the deviation coverage time ratio, the longest consecutive deviation time, and the stage peak coefficient. Specifically, the deviation coverage time ratio is obtained by comparing the total deviation time within that fermentation stage to the total duration of that fermentation stage; the longest consecutive deviation time is obtained by taking the maximum value among all stage activity trajectory segments within that fermentation stage; and the stage peak coefficient is obtained by taking the maximum value among all stage activity coefficients within the coverage area of ​​that fermentation stage. After completing these three extractions, the deviation coverage time ratio, the longest consecutive deviation time, and the stage peak coefficient corresponding to each fermentation stage are combined into an ordered result set according to the fermentation stage sequence. This ordered result set is then identified as the strain activity parameter evaluation result. The strain activity parameter evaluation result, along with the complete stage activity trajectory, is output to step S4.

[0072] After step S3, the state deviation fragments have been converted into stage activity trajectory fragments with clearly defined deviation start points, deviation durations, and deviation end points, forming complete stage activity trajectories in the order of fermentation stages. Simultaneously, the deviation coverage, continuity, and peak intensity within each fermentation stage are also compiled into the results of strain activity parameter evaluation.

[0073] Step S4: Read the strain activity parameter evaluation results and stage activity trajectory, extract the trajectory connection positions between adjacent fermentation stages and the deviation distribution results within each fermentation stage, generate the activity state change results, and map the strain activity parameter evaluation results and activity state change results into activity state tags and production disposal prompts.

[0074] Step S401: Extract the connection position of adjacent fermentation stage trajectories.

[0075] First, the stage activity trajectory of each fermentation stage is read segment by segment. The total deviation time of all stage activity trajectory segments within each fermentation stage is calculated. If the total deviation time is greater than zero, the fermentation stage is recorded as a stage where the trajectory exists and is established. If the total deviation time is equal to zero, the fermentation stage is recorded as a stage where the trajectory does not exist and is established. For fermentation stages where the trajectory exists and is established, the starting time point of the earliest appearing stage activity trajectory segment is extracted as the first deviation time point of the fermentation stage, and the ending time point of the latest ending stage activity trajectory segment is extracted as the last deviation time point of the fermentation stage.

[0076] Subsequently, for any two adjacent fermentation stages, if both the preceding and following fermentation stages are valid fermentation stages with existing trajectories, then the position of the last deviation point of the preceding fermentation stage within the preceding fermentation stage is used to determine the tail connection position of the preceding fermentation stage. Similarly, the position of the first deviation point of the following fermentation stage within the following fermentation stage is used to determine the front connection position of the following fermentation stage. The stage connection interval is determined by the time difference between the deviation boundaries of the two stages. The tail connection position and the front connection position characterize the relative landing point of the deviation within adjacent fermentation stages, and the stage connection interval characterizes whether the deviation is continuous between adjacent fermentation stages.

[0077] Step S402: Fermentation stage intensity classification and activity stage code generation.

[0078] Read the deviation coverage ratio, longest consecutive deviation duration, and stage peak coefficient output from step S3, and introduce coverage determination threshold, persistence determination threshold, and peak determination threshold respectively. Specifically, the coverage determination threshold is preferentially taken as the upper bound of the deviation coverage ratio of historical normal batches within the corresponding fermentation stage; the persistence determination threshold is preferentially taken as the upper bound of the longest consecutive deviation duration of historical normal batches within the corresponding fermentation stage; and the peak determination threshold is preferentially taken as the upper bound of the stage peak coefficient of historical normal batches within the corresponding fermentation stage. When it is necessary to balance detection sensitivity and false alarm rate, the process engineer can make a one-time fine-tuning based on the aforementioned upper bounds, and the fine-tuned thresholds are bound to the process version and saved.

[0079] Subsequently, each fermentation stage is compared item by item. If the deviation from coverage reaches the coverage threshold, an item is recorded as valid; if the longest continuous deviation reaches the persistence threshold, an item is recorded as valid; if the stage peak coefficient reaches the peak threshold, another item is recorded as valid. The sum of these three valid items is determined as the active stage code. If a fermentation stage is a stage where the trajectory does not exist, the active stage code for that fermentation stage is directly recorded as zero. An active stage code of zero indicates that the fermentation stage does not have any stage deviation features that would lead to a mismatch classification; an active stage code of one indicates that the fermentation stage has a single-dimensional deviation; an active stage code of two indicates that the fermentation stage has a two-dimensional deviation; and an active stage code of three indicates that the fermentation stage enters the mismatch classification in all three dimensions: coverage, persistence, and peak intensity.

[0080] Step S403: Stage migration determination and generation of active state change results.

[0081] For each pair of adjacent fermentation stages, the existence of the trajectory of the previous and subsequent fermentation stages, the active stage code, and the stage connection interval are read, and a connection determination threshold is introduced. The connection determination threshold is determined based on the maximum time interval at the boundary of adjacent fermentation stages that can be considered as a continuous migration of the same active mismatch event, and can be taken as the greater of the stage switching buffer time, the control action transmission delay, and the fixed sampling interval. For example, if the subsequent fermentation stage deviates within three minutes after the stage switch, the deviation can be regarded as a continuous migration of the deviation of the previous fermentation stage to the subsequent fermentation stage.

[0082] The stage transition type is then determined according to the following rules: If the previous fermentation stage is a stage where the trajectory does not exist and the next fermentation stage is a stage where the trajectory exists and the transition is initiated, the adjacent fermentation stage is designated as the initial transition; if the previous fermentation stage is a stage where the trajectory exists and the next fermentation stage is a stage where the trajectory does not exist and the transition is initiated, the adjacent fermentation stage is designated as the exit boundary; if both the previous and next fermentation stages are stages where the trajectory exists and the stage connection interval does not exceed the connection determination threshold, and the active stage code of the next fermentation stage is higher than the active stage code of the previous fermentation stage, then the adjacent stage is designated as the exit boundary. Fermentation stages are defined as continuously aggravated. If the active stage code of the subsequent fermentation stage is equal to that of the preceding fermentation stage, the adjacent fermentation stage is defined as continuously maintained. If the active stage code of the subsequent fermentation stage is lower than that of the preceding fermentation stage, the adjacent fermentation stage is defined as continuously alleviated. If both the preceding and following fermentation stages have trajectories and constitute a fermentation stage, but the stage connection interval exceeds the connection determination threshold, the adjacent fermentation stage is defined as discontinuous migration. If both the preceding and following fermentation stages have no trajectories and constitute a fermentation stage, the adjacent fermentation stage is defined as no migration. The stage migration types corresponding to each adjacent fermentation stage are arranged in the order of fermentation stages to form the results of the change in active state.

[0083] Step S404: Generation of active status label and output of production disposal prompt.

[0084] First, extract the initial stage number corresponding to the initial transition from the activity state change results. If there is no initial transition in the activity state change results, record the initial stage number as zero. Next, extract the activity stage code corresponding to the last established fermentation stage with a known trajectory from all fermentation stages, and use it as the final stage activity stage code. If all fermentation stages are without a known trajectory and thus a fermentation stage is established, record the final stage activity stage code as zero. Then, according to a pre-defined migration priority order, extract the main migration type from the activity state change results. The priority order of intermittent migration is higher than continuous aggravation, continuous aggravation is higher than continuous maintenance, continuous maintenance is higher than continuous mitigation, continuous mitigation is higher than initial transition, initial transition is higher than exit boundary, and exit boundary is higher than no migration.

[0085] Then, the initial stage sequence number, the final stage activity stage code, and the main migration type are used as a ternary index and input into the production disposal prompt rule set. The production disposal prompt rule set is pre-established by process engineers, quality personnel, and production supervisors based on historical disposal experience and is bound and saved to the strain type, equipment model, and process version. When the process version changes, the correspondence between the activity status label and the production disposal prompt is updated synchronously. After completing the table lookup, a unique activity status label is obtained, and the corresponding production disposal prompt is output.

[0086] In one embodiment, if the initial stage number is located in a subsequent feeding-related fermentation stage, the final stage activity stage code is three, and the main migration type is continuous aggravation, then the activity status label corresponds to the subsequent continuous mismatch label, and the production handling prompt corresponds to prioritizing the verification of feeding cycle time, aeration settings, and temperature pullback order; if the initial stage number is located in an early culture-related fermentation stage, the final stage activity stage code is one, and the main migration type is continuous mitigation, then the activity status label corresponds to the initial short-term mismatch recovery label, and the production handling prompt corresponds to maintaining the current culture conditions and increasing the frequency of subsequent sampling.

[0087] After step S4, the complete stage activity trajectory and strain activity parameter evaluation results have been summarized into activity state change results, and further mapped into a unique activity state label and corresponding production disposal prompts. The output no longer focuses solely on activity identification, but unifies the deviation characteristics within the fermentation stage and the migration relationships between stages into a single batch judgment framework, forming a clear expression for fermentation management scenarios. Specifically, the above are merely preferred embodiments of this application and are not intended to limit the scope of this application.

[0088] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0089] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for online intelligent evaluation and modeling of the activity parameters of fermentation strains, characterized in that, Including the following steps: S1: Obtain all data of the current fermentation batch, rearrange them according to a unified time index to form a quantitative characterization sequence and a process response sequence, and divide them into continuous time windows corresponding to the quantitative characterization sequence and the process response sequence respectively; S2: Read the continuous time windows one by one, map the historical normal batch reference interval of the fermentation stage to the quantity characterization sequence and process response sequence, identify the misalignment window where the process response sequence falls outside the reference interval and the quantity characterization sequence falls within the reference interval, generate the reference conversion steady quantity, the maintenance load conversion time and the liveness mismatch coefficient, and connect them to obtain the state deviation segment. S3: Sort the state deviation segments according to a unified time index, call the reference segments of the corresponding fermentation stage, and align the quantity characterization sequence, process response sequence and surfactant mismatch coefficient in the state deviation segments hourly to determine the deviation start point, deviation duration range and deviation end point, form the stage activity trajectory and determine the strain activity parameter evaluation results. S4: Read the strain activity parameter evaluation results and stage activity trajectory, extract the trajectory connection positions between adjacent fermentation stages and the deviation distribution results within each fermentation stage, generate activity state change results, and map the strain activity parameter evaluation results and activity state change results into activity state tags and production disposal prompts.

2. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 1, characterized in that, Step S1 includes: First, a unique source of quantitative data is determined between the viable bacteria concentration data and the estimated bacterial cell concentration. Then, missing timestamps and missing measurements are removed from the viable bacteria concentration data or estimated bacterial cell concentration data, acid-base data, feeding data, temperature data, and ventilation data. When the same timestamp is repeated, the last written item is retained. Subsequently, a unified time index is generated based on the common valid observation interval and fixed sampling interval.

3. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 2, characterized in that, Step S1 also includes: Piecewise linear interpolation is performed on viable bacterial concentration data or estimated bacterial cell concentration, acid-base data, and temperature data along a unified time index. Zero-order preservation is performed on feeding data and aeration data. After removing observation breakpoints, continuous index segments are formed. Then, continuous time windows are generated from each continuous index segment according to the preset window length and the rules for the intersection of adjacent windows.

4. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 3, characterized in that, Step S2 includes: First, each continuous time window is compared with the time range of each fermentation stage. The stage mapping window is determined according to the fermentation stage with the longest overlapping duration. Then, the historical normal batch processed in step S1 is called up. At the unified time point covered by the stage mapping window, the quantity reference interval and the process reference intervals for acid-base, feeding, temperature and aeration are generated respectively. The reference intervals are then mapped to the quantity characterization sequence and the process response sequence.

5. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 4, characterized in that, Step S2 also includes: Within the phase mapping window, first identify quantity bound segments whose duration reaches the quantity bound duration threshold, then identify process deviation segments whose duration reaches the process deviation duration threshold, and determine the portion of the overlapping time period between quantity bound segments and process deviation segments whose duration reaches the misalignment overlap duration threshold as misalignment windows, and determine the quantity bound time period within the misalignment window as stable quantity segments.

6. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 5, characterized in that, Step S2 also includes: The stable load time product is obtained by accumulating the quantity value point by point along the stable quantity segment and multiplying it by the fixed sampling interval. The average result of the median value of the historical normal batch quantity at the same time point corresponding to the aforementioned stable quantity segment is used as the reference quantity benchmark. The stable load time product is converted into the reference converted stable quantity time. At the same time, the process channels within the stable quantity segment that fall outside the process reference interval are counted point by point and accumulated to obtain the maintenance load converted time. Then, the liveness mismatch coefficient is output by the dual-channel mutual verification learning model.

7. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 6, characterized in that, Step S3 includes: First, sort the state deviation segments according to the unified time index, and then spread the liveness mismatch coefficients of the effective misalignment windows that make up the state deviation segments into a time-spread liveness mismatch sequence. Then, call the reference segments covering the same unified time point set from the historical normal batches. The reference segments include the upper and lower bounds of quantity reference, the upper and lower bounds of process reference, and the upper bound of liveness mismatch reference. At the same unified time point, they are aligned with the quantity characterization sequence, the process response sequence, and the time-spread liveness mismatch sequence hourly.

8. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 7, characterized in that, Step S3 also includes: Based on quantity deviation, process deviation, and coefficient deviation, trajectory deviation identifiers are generated. Continuous trajectory deviation segments whose duration reaches the deviation confirmation duration threshold are identified as initial candidate deviation segments. Adjacent initial candidate deviation segments with an interval not exceeding the deviation allowable gap threshold are then connected together. The candidate deviation segment with the largest cumulative coefficient is used to determine the deviation start point, deviation duration range, and deviation end point.

9. The online intelligent evaluation and modeling method for fermentation strain activity parameters according to claim 8, characterized in that, Step S3 also includes: Within the same fermentation stage, the time interval between adjacent deviation durations that does not exceed the stage concatenation interval is concatenated to form a stage activity trajectory fragment, and a complete stage activity trajectory is formed according to the fermentation stage sequence. At the same time, the deviation coverage time ratio, the longest continuous deviation time, and the stage peak coefficient are extracted in each fermentation stage, and the deviation coverage time ratio, the longest continuous deviation time, and the stage peak coefficient are used to form the strain activity parameter evaluation results.

10. The online intelligent evaluation and modeling method for the activity parameters of fermentation strains according to claim 9, characterized in that, Step S4 includes: The results of the change in active state are determined based on the trajectory connection position and connection judgment threshold between adjacent fermentation stages. The active stage code of each fermentation stage is determined based on the coverage judgment threshold, the continuity judgment threshold and the peak judgment threshold. The first stage number and the main migration type are extracted from the results of the change in active state and combined with the active stage code of the last stage to form a ternary index, which is uniquely mapped to the active state label and production disposal prompt according to the preset rule set.

Images