Early warning method and system for information abnormity in stem cell culture process

Abstract

The invention provides a stem cell culture process information abnormity early warning method, which comprises the following steps: preprocessing original data in a stem cell culture process, extracting layered features to obtain basic statistical features, time sequence trend features and environmental factor interaction features, and updating the features in real time through a dynamic sliding window, after the verified feature set is obtained, fusion calculation is carried out to obtain a multi-dimensional fusion feature set; calculating a trend deviation degree, a feature correlation degree and a fluctuation dispersion degree of the multi-dimensional fusion feature set, and carrying out dynamic weight distribution and fusion to obtain a fusion anomaly degree value; and matching the fusion anomaly degree value with a multi-level threshold value to obtain a matching result, judging an abnormal state type in combination with an association relationship between the feature anomaly type and the cultivation stage, and generating an anomaly early warning result. According to the method, through data preprocessing, layered feature extraction and fusion and dynamic anomaly calculation, the accuracy, real-time performance and practicability of stem cell culture anomaly early warning are improved.

Description

Technical Field

[0001] This invention relates to the field of stem cell culture technology, and in particular to a method and system for early warning of abnormal information in the stem cell culture process. Background Technology

[0002] As a core resource for regenerative medicine, cell therapy, and drug development, the quality of stem cell culture directly determines the safety and efficacy of clinical applications. Dynamic changes in environmental parameters such as temperature, pH, and dissolved oxygen concentration, morphological indicators such as cell density and adhesion rate, and various factors including culture medium composition and metabolites can all trigger abnormalities and lead to culture failure. With the development of stem cell culture towards large-scale and refined methods, traditional methods relying on manual sampling and testing can no longer meet the needs of real-time quality control.

[0003] Currently, existing technologies related to early warning of abnormalities in stem cell culture are developing in multiple dimensions: some technologies deploy multiple sets of sensors to collect single types of data, such as metabolite concentration or environmental nutrient information, and combine them with preset thresholds or simple fitting models to achieve basic early warning; other technologies establish predictive models based on cell growth kinetic analysis, or use machine learning algorithms to integrate multi-sensor data and image information to attempt to capture abnormal features in the cell growth process; at the same time, there are also methods that rely on traditional statistical analysis (such as mean and variance calculation) or preliminary multimodal data fusion to monitor key quality attributes in the cell proliferation and differentiation process. These technologies have, to some extent, achieved a shift from "offline detection" to "real-time monitoring".

[0004] However, significant shortcomings remain, making it difficult to meet the refined early warning needs of the stem cell culture process: First, the ability to process multi-source heterogeneous data is weak. Faced with various types of data such as environment, morphology, operation, and components during the culture process, there is a lack of systematic hierarchical cleaning and standardized processes, which can easily lead to deviations in the accuracy of early warnings due to inconsistent data formats and rough handling of outliers. Second, feature extraction and anomaly judgment lack dynamic adaptability. Existing technologies mostly adopt static feature extraction methods, which do not fully consider the characteristic differences of different culture stages such as inoculation, proliferation, and differentiation. Moreover, the threshold settings and weight allocations are fixed, which cannot respond to the differences in anomaly sensitivity between different batches and different stages. Third, the practicality and depth of early warning are insufficient. Existing technologies can only achieve a qualitative judgment of "whether it is abnormal" and lack precise location of the root cause of the anomaly, prediction of the scope of influence, and targeted treatment suggestions. In addition, the generalization ability of some models is limited, and they rely on a large amount of labeled data, making it difficult to adapt to complex and ever-changing actual culture scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for early warning of abnormal information in the stem cell culture process. By standardizing the raw data, extracting and fusing hierarchical features, and calculating the dynamic anomaly degree, the method can accurately identify the types of abnormal states and generate early warning results, thereby improving the accuracy, real-time performance, and practicality of early warning for abnormal stem cell culture.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for early warning of abnormal information in the stem cell culture process includes the following steps: We acquire multi-source heterogeneous raw data during stem cell culture, and perform outlier identification, repair processing, timestamp alignment, and format standardization on the multi-source heterogeneous raw data to obtain a unified dataset. Hierarchical feature extraction is performed on a unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. The features are then updated in real time through a dynamic sliding window to obtain a validated feature set. A multi-dimensional fused feature set is obtained by fusing the verified feature set using a weighted fusion strategy. Calculate the trend deviation, feature correlation, and volatility dispersion of the multi-dimensional fusion feature set, and dynamically assign and fuse the trend deviation, feature correlation, and volatility dispersion through feature synergistic contribution to obtain the fusion anomaly value; The fusion anomaly value is matched with multi-level thresholds to obtain the matching results. Based on the matching results and the correlation between the characteristic anomaly type and the culture stage, the abnormal state type of the stem cell culture process is determined. Anomaly warning results are generated based on the abnormal state type and the fusion anomaly degree value.

[0007] Optionally, the multi-source heterogeneous raw data includes: culture environment monitoring data, cell morphology observation data, culture operation process data, culture medium composition change data, and cell metabolite detection data; culture environment monitoring data includes: temperature, pH, dissolved oxygen concentration, CO2 concentration, and osmotic pressure; cell morphology observation data includes: cell density, cell adhesion rate, cell diameter, and nucleocytoplasmic ratio; culture operation process data includes: medium change time, additive dosage, stirring rate, and culture dish number; culture medium composition change data includes: glucose concentration, glutamine concentration, growth factor concentration, and metabolic waste concentration; cell metabolite detection data includes: lactate concentration, ammonia concentration, reactive oxygen species content, and ATP concentration.

[0008] Optionally, acquire multi-source heterogeneous raw data during stem cell culture, and perform outlier identification, repair processing, timestamp alignment, and format standardization on the multi-source heterogeneous raw data to obtain a unified dataset, including: Potential outliers in numerical data from multi-source heterogeneous raw data were identified using the quartile method, and then filtered according to rules in the field of stem cell culture to obtain true outliers. True outliers include: deletion outliers and mutation outliers. Missing outliers were repaired using a weighted average of adjacent timestamp data, and abrupt outliers were repaired using a multinomial fitting interpolation method based on the data trend type, resulting in repaired data. Using the cultivation start time as a benchmark, the missing timestamp data in the data is supplemented and repaired by linear interpolation to obtain aligned data; The numerical data in the aligned data is subjected to max-min normalization, the categorical data in the aligned data is subjected to one-hot encoding, and the textual operational data in the aligned data is subjected to keyword extraction and integer encoding to obtain a unified dataset.

[0009] Optionally, hierarchical feature extraction is performed on the unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. These features are then updated in real-time using a dynamic sliding window to obtain a validated feature set, including: Extract the mean, variance, maximum value, minimum value, coefficient of variation, and skewness from the unified dataset to obtain the basic statistical characteristics; The adaptive inflection point detection algorithm identifies the trend inflection points of the data sequence in the unified dataset. The time series in the unified dataset is divided into several monotonic trend segments and subjected to quadratic polynomial fitting to obtain the trend equation. The inflection point time and trend segment duration of the trend equation are extracted by the trend inflection point. The slope, intercept, inflection point time and trend segment duration are integrated into the time series trend features. The product of the squares of two different environmental factors is used as a nonlinear interaction term. When any environmental factor is in a preset critical interval, the correlation increment between another environmental factor and cell growth indicators is used as a conditional interaction term. The nonlinear interaction term and the conditional interaction term are integrated into the environmental factor interaction feature. The sliding window is determined based on the different stages of stem cell culture. Based on the sliding window, the time weight vectors of basic statistical features, time series trend features and environmental factor interaction features are determined by the exponential weighting method. The features of basic statistical features, time series trend features and environmental factor interaction features are then updated according to the time weight vectors to obtain the verified feature set.

[0010] Optionally, the trend deviation, feature correlation, and volatility dispersion of the multi-dimensional fused feature set are calculated, and the trend deviation, feature correlation, and volatility dispersion are dynamically weighted and fused using feature synergistic contribution to obtain a fused anomaly value, including: Through formula The trend deviation is calculated; where, The length of the sliding window. For the current eigenvalue, As the baseline eigenvalue, As a penalty factor; Through formula The feature correlation degree is calculated; where, The weights of the Pearson correlation coefficient, The Pearson correlation coefficient between the two features. The mutual information value between two features. The number of features in the multi-dimensional fusion feature set; Through formula The fluctuation dispersion is calculated; where, The standard deviation within the sliding window. The mean within the sliding window. The maximum value of the feature within the sliding window. The minimum value of the feature within the sliding window. The degree of fluctuation exceeding the limit; The feature synergy consistency coefficient among trend deviation, feature correlation, and volatility dispersion is used as the feature synergy contribution for dynamic weight adjustment; the expression for the dynamic weight adjustment process is: ;in, For the original weights, The coefficient of feature coherence; By using modified dynamic weights, the trend deviation, feature correlation, and volatility dispersion are weighted and fused to obtain a fused anomaly value.

[0011] Optionally, the fusion anomaly value is matched with a multi-level threshold to obtain a matching result. Based on the matching result and the correlation between the characteristic anomaly type and the culture stage, the abnormal state type of the stem cell culture process is determined, including: The first threshold, the second threshold, and the third threshold were determined based on the 90%, 95%, and 99% quantiles of historical normal culture data, respectively. Based on different cultivation stages, stage-specific threshold correction terms are determined, and the first, second, and third thresholds are corrected using these stage-specific threshold correction terms to obtain multi-level thresholds. The anomaly type is determined by the proportion of anomaly indicators in the fusion anomaly value, and the degree of anomaly is determined by the comparison between the fusion anomaly value and the multi-level threshold. Based on the cultivation stage, the characteristic anomaly types are classified into association types, which include: primary association anomaly types and secondary association anomaly types. The triggering conditions for the association type are determined based on the proportion of abnormal indicators and the comparison results, and the abnormal state type is determined based on the abnormality type, abnormality degree and triggering conditions.

[0012] Optionally, anomaly warning results are generated based on the anomaly state type and the fused anomaly degree value, including: The basic weights for abnormal state types are determined based on domain knowledge, and sensitivity coefficients are determined based on different cultivation stages. The anomaly correction value is obtained by adjusting the sensitivity of the fusion anomaly value based on the basic weights and sensitivity coefficients. The warning level boundary of the abnormal state type is refined based on the anomaly correction value to obtain the anomaly warning result; the anomaly warning result includes: warning level, abnormal time and abnormal cause indication.

[0013] A stem cell culture process information anomaly early warning system, comprising: The data acquisition module is used to acquire multi-source heterogeneous raw data during the stem cell culture process, and to perform outlier identification, repair processing, timestamp alignment and format standardization processing on the multi-source heterogeneous raw data to obtain a unified dataset; The feature extraction module is used to perform hierarchical feature extraction on a unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. The features are updated in real time through a dynamic sliding window to obtain a validated feature set. The feature fusion module is used to perform fusion calculations on the verified feature set through a weighted fusion strategy to obtain a multi-dimensional fused feature set. The anomaly detection module is used to calculate the trend deviation, feature correlation, and fluctuation dispersion of the multi-dimensional fused feature set, and to dynamically assign and fuse the trend deviation, feature correlation, and fluctuation dispersion through feature collaborative contribution to obtain the fused anomaly value. The state discrimination module is used to match the fusion anomaly value with multi-level thresholds to obtain the matching result, and to determine the abnormal state type of the stem cell culture process based on the matching result and the correlation between the characteristic anomaly type and the culture stage. The anomaly warning module is used to generate anomaly warning results based on the anomaly state type and the fused anomaly degree value.

[0014] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The stem cell culture process information anomaly early warning method and system provided by the present invention includes: acquiring multi-source heterogeneous raw data during the stem cell culture process; performing outlier identification, repair processing, timestamp alignment, and format standardization processing on the multi-source heterogeneous raw data to obtain a unified dataset; performing hierarchical feature extraction on the unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features, and updating the features in real time through a dynamic sliding window to obtain a verified feature set; performing fusion calculation on the verified feature set through a weighted fusion strategy to obtain a multi-dimensional fused feature set; calculating the trend deviation, feature correlation, and fluctuation dispersion of the multi-dimensional fused feature set, and dynamically assigning and fusing the trend deviation, feature correlation, and fluctuation dispersion through feature synergistic contribution to obtain a fusion anomaly value; matching the fusion anomaly value with multi-level thresholds to obtain a matching result, and judging the abnormal state type of the stem cell culture process based on the matching result and the correlation between the feature anomaly type and the culture stage; and generating an anomaly early warning result based on the abnormal state type and the fusion anomaly value. This method accurately identifies abnormal state types and generates early warning results by standardizing the raw data, extracting and fusing hierarchical features, and calculating dynamic anomalies, thereby improving the accuracy, real-time performance, and practicality of early warning for abnormal stem cell culture. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the stem cell culture process information anomaly early warning method of the present invention; Figure 2 This is a schematic diagram of the stem cell culture process information anomaly early warning system of the present invention. Detailed Implementation

[0017] 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.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, the present invention provides a method for early warning of abnormal information in the stem cell culture process, comprising the following steps: Step 100: Obtain multi-source heterogeneous raw data during stem cell culture, and perform outlier identification, repair processing, timestamp alignment, and format standardization on the multi-source heterogeneous raw data to obtain a unified dataset; Step 200: Perform hierarchical feature extraction on the unified dataset to obtain basic statistical features, time series trend features and environmental factor interaction features, and update the features in real time through a dynamic sliding window to obtain the validated feature set; Step 300: Perform fusion calculation on the verified feature set using a weighted fusion strategy to obtain a multi-dimensional fused feature set; Step 400: Calculate the trend deviation, feature correlation, and volatility dispersion of the multi-dimensional fusion feature set, and dynamically assign and fuse the trend deviation, feature correlation, and volatility dispersion through feature collaborative contribution to obtain the fusion anomaly value; Step 500: Match the fusion anomaly value with the multi-level threshold to obtain the matching result, and determine the abnormal state type of the stem cell culture process based on the matching result and the correlation between the characteristic anomaly type and the culture stage; Step 600: Generate anomaly warning results based on the abnormal state type and the fusion anomaly degree value.

[0020] Preferably, the multi-source heterogeneous raw data includes: culture environment monitoring data, cell morphology observation data, culture operation process data, culture medium composition change data, and cell metabolite detection data; culture environment monitoring data includes: temperature, pH value, dissolved oxygen concentration, CO2 concentration, and osmotic pressure; cell morphology observation data includes: cell density, cell adhesion rate, cell diameter, and nucleocytoplasmic ratio; culture operation process data includes: medium change time, additive dosage, stirring rate, and culture dish number; culture medium composition change data includes: glucose concentration, glutamine concentration, growth factor concentration, and metabolic waste concentration; cell metabolite detection data includes: lactate concentration, ammonia concentration, reactive oxygen species content, and ATP concentration.

[0021] Preferably, the process involves acquiring multi-source heterogeneous raw data during stem cell culture, and then performing outlier identification, repair processing, timestamp alignment, and format standardization on the multi-source heterogeneous raw data to obtain a unified dataset, including: Potential outliers in numerical data from multi-source heterogeneous raw data were identified using the quartile method, and then filtered according to rules in the field of stem cell culture to obtain true outliers. True outliers include: deletion outliers and mutation outliers. Missing outliers were repaired using a weighted average of adjacent timestamp data, and abrupt outliers were repaired using a multinomial fitting interpolation method based on the data trend type, resulting in repaired data. Using the cultivation start time as a benchmark, the missing timestamp data in the data is supplemented and repaired by linear interpolation to obtain aligned data; The numerical data in the aligned data is subjected to max-min normalization, the categorical data in the aligned data is subjected to one-hot encoding, and the textual operational data in the aligned data is subjected to keyword extraction and integer encoding to obtain a unified dataset.

[0022] In the specific implementation process, step 100, for numerical data in the multi-source heterogeneous original data, calculates the first quartile Q1 and the third quartile Q3, determines the interquartile range (IQR) = Q3 - Q1, and identifies values ​​less than Q1 - 1.5IQR or greater than Q3 + 1.5IQR as potential outliers. Then, combining with the rules of the stem cell culture field, it filters out missing outliers (data records are empty or not collected) and mutation outliers (deviations deviating from the overall data trend and exceeding the reasonable range of the field) as true outliers. For missing outliers, the weight of adjacent timestamp data decreases as the distance from the missing timestamp increases. The weights of adjacent timestamps of missing data are set to 0.6 and 0.4 respectively, and a weighted sum is performed based on the two weights to repair the missing outlier. For mutation outliers, the trend type is first determined by data trend fitting. A polynomial fitting interpolation method is used, substituting the normal data before and after the mutation point to solve for the polynomial coefficients. Then, the interpolated repair value is calculated based on the timestamp of the mutation point to obtain the repaired data. Next, using the stem cell culture start time as a benchmark, and addressing the missing timestamps in the repair data, assuming the missing timestamps lie between the known timestamps t1 and t2, with corresponding data f(t1) and f(t2) respectively, the formula is used... Missing timestamp data was filled in to obtain unified, aligned timestamp data. Finally, numerical data underwent max-min normalization, while categorical data (i.e., classification information without continuous numerical meaning, such as petri dish numbers) was processed using one-hot encoding. This involved mapping each categorical value to a binary vector with a length equal to the total number of categories, where only the corresponding category position was set to 1, and the rest to 0. For textual operational data (i.e., descriptions of operations such as changing the medium or adding growth factors), keywords were first extracted using the TF-IDF algorithm, and then each keyword was mapped to a unique integer code. Finally, these were integrated to obtain a unified dataset.

[0023] Preferably, hierarchical feature extraction is performed on the unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. These features are then updated in real time using a dynamic sliding window to obtain a validated feature set, including: Extract the mean, variance, maximum value, minimum value, coefficient of variation, and skewness from the unified dataset to obtain the basic statistical characteristics; The adaptive inflection point detection algorithm identifies the trend inflection points of the data sequence in the unified dataset. The time series in the unified dataset is divided into several monotonic trend segments and subjected to quadratic polynomial fitting to obtain the trend equation. The inflection point time and trend segment duration of the trend equation are extracted by the trend inflection point. The slope, intercept, inflection point time and trend segment duration are integrated into the time series trend features. The product of the squares of two different environmental factors is used as a nonlinear interaction term. When any environmental factor is in a preset critical interval, the correlation increment between another environmental factor and cell growth indicators is used as a conditional interaction term. The nonlinear interaction term and the conditional interaction term are integrated into the environmental factor interaction feature. The sliding window is determined based on the different stages of stem cell culture. Based on the sliding window, the time weight vectors of basic statistical features, time series trend features and environmental factor interaction features are determined by the exponential weighting method. The features of basic statistical features, time series trend features and environmental factor interaction features are then updated according to the time weight vectors to obtain the verified feature set.

[0024] In the specific implementation process, step 200 calculates the product of the squares of the mean and variance of various data in the unified dataset as a nonlinear interaction term. Then, based on knowledge in the field of stem cell culture, it sets preset critical ranges for each environmental factor. In some embodiments, the critical range for temperature is 36℃-38℃, and the critical range for pH is 7.0-7.6. When any environmental factor is within this critical range, the correlation increment between another environmental factor and cell growth indicators such as cell density and adhesion rate is calculated. As a conditional interaction item The change in environmental factors The changes in cell growth indicators were used, and the interaction characteristics of environmental factors were obtained by integrating nonlinear and conditional interaction terms. Then, the dynamic sliding window length was determined according to different stages of stem cell culture, such as the inoculation phase, proliferation phase, and differentiation phase. During the inoculation phase, when cell state fluctuates greatly, the window length was set to 2 hours; during the proliferation phase, when cell growth is stable, the window length was set to 4 hours; and during the differentiation phase, when cell function matures, the window length was set to 6 hours. Finally, based on this sliding window, the time weight vector was calculated using an exponential weighting method. The calculation formula is as follows: , ; in, This is the attenuation coefficient, which in some embodiments ranges from 0.85 to 0.95, and is used to control the weight attenuation rate. The number of feature data points within the sliding window. The time-series position of the data within the window is used to calculate the time weight vectors corresponding to the basic statistical features, time series trend features, and environmental factor interaction features. Then, the historical data of each feature within the sliding window are weighted and summed with the corresponding weight vectors to complete the real-time update of the three types of features. Finally, the updated features are integrated to obtain the verified feature set.

[0025] In the specific implementation process, step 300 first conducts a pre-analysis based on knowledge and feature contribution in the field of stem cell culture. Among these, time-series trend features are more critical in reflecting dynamic changes in cell growth, and environmental factor interaction features are directly related to the adaptability of the cell's microenvironment for survival. Therefore, the initial weights of the three types of features are determined to be 0.3, 0.4, and 0.3, respectively. Subsequently, the importance index of each type of feature is calculated using the information gain method. The calculation formula is as follows: ; in, The information gain of feature set X. The prior entropy represents the normal or abnormal growth state of a cell. Let X be the posterior entropy given the feature set X. Then, normalize the information gain of each feature to obtain the weight adjustment coefficients, and then... The initial weights are corrected by multiplication. Next, Z-score normalization is used to standardize the specific feature terms within each feature class, eliminating magnitude differences between different features. Then, the standardized feature terms within each feature class are concatenated sequentially to form intra-class feature vectors, which are then weighted and fused according to the corrected weights. Finally, redundancy removal is performed on the preliminary fused feature vectors using Principal Component Analysis (PCA) to calculate the covariance matrix of the feature vectors and solve for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​are sorted from largest to smallest, and principal components with a cumulative variance contribution rate ≥ 95% are selected to form the final multi-dimensional fused feature set. This method retains core information while reducing feature dimensionality, avoiding redundant features from interfering with subsequent anomaly detection accuracy.

[0026] In the specific implementation process, step 400 first uses the characteristic mean of the historical normal cultivation data at the corresponding timestamp as the benchmark characteristic value, and sets a penalty factor according to the deviation direction to calculate the trend deviation degree. The calculation formula is as follows: ; in, The length of the sliding window. For the current eigenvalue, As the baseline eigenvalue, As a penalty factor, in some embodiments, the penalty factor is set to 1.3 when the culture temperature exceeds 37.5°C and the pH value is below 7.2. Next, the feature correlation is calculated. First, the Pearson correlation coefficient of each pair of features in the multi-dimensional fusion feature set is calculated, and then the mutual information value between the two is calculated. The weight of the Pearson correlation coefficient is set according to the feature association type: features with strong linear correlation, such as glucose concentration and cell density, are assigned a weight of 0.7; features with strong non-linear correlation, such as dissolved oxygen concentration and ATP concentration, are assigned a weight of 0.3. This is achieved using the formula... The feature correlation degree is calculated, where, The weights of the Pearson correlation coefficient, The Pearson correlation coefficient between the two features. The mutual information value between two features. This represents the number of features in the multi-dimensional fusion feature set. Then, the volatility dispersion is calculated, starting with the standard deviation within the sliding window. and mean To determine the maximum and minimum values ​​of the characteristic, the degree of fluctuation exceeding the limit is calculated based on the proportion exceeding the preset normal fluctuation range, where the normal fluctuation range is... The ratio of the excess portion to the range width represents the degree of fluctuation exceeding the limit, which is then expressed using the formula... The fluctuation dispersion is calculated, where, The maximum value of the feature within the sliding window. The minimum value of the feature within the sliding window. The degree of fluctuation exceeding the limit is then determined. Next, the feature synergy consistency coefficient is calculated as the feature synergy contribution. This is achieved by calculating the pairwise Pearson correlation coefficients among trend deviation, feature correlation, and fluctuation dispersion, and taking the average of these three as the feature synergy consistency coefficient. The value range is [-1, 1], with positive values ​​indicating abnormal synergy and negative values ​​indicating weak synergy. Then, dynamic weight correction is performed. Initially, the weights are set to 0.4, 0.3, and 0.3 based on domain knowledge, and then the formula is used to adjust the weights. The original weights are corrected, whereby... For the original weights, Let be the feature consistency coefficient, if If the result is negative, the corresponding weight is reduced. After correction, the weights are normalized to ensure the sum is 1. Finally, the weighted fusion formula is used. The fusion anomaly value is obtained by multiplying the corrected dynamic weights by the corresponding discretenesses and summing the results.

[0027] Preferably, the fusion anomaly value is matched with a multi-level threshold to obtain a matching result. Based on the matching result and the correlation between the characteristic anomaly type and the culture stage, the abnormal state type of the stem cell culture process is determined, including: The first threshold, the second threshold, and the third threshold were determined based on the 90%, 95%, and 99% quantiles of historical normal culture data, respectively. Based on different cultivation stages, stage-specific threshold correction terms are determined, and the first, second, and third thresholds are corrected using these stage-specific threshold correction terms to obtain multi-level thresholds. The anomaly type is determined by the proportion of anomaly indicators in the fusion anomaly value, and the degree of anomaly is determined by the comparison between the fusion anomaly value and the multi-level threshold. Based on the cultivation stage, the characteristic anomaly types are classified into association types, which include: primary association anomaly types and secondary association anomaly types. The triggering conditions for the association type are determined based on the proportion of abnormal indicators and the comparison results, and the abnormal state type is determined based on the abnormality type, abnormality degree and triggering conditions.

[0028] In the specific implementation process, step 500 first collects a large number of historical normal culture data corresponding to fusion anomaly reference values. After sorting them from smallest to largest, the 90th quantile is determined as the first threshold T1, the 95th quantile as the second threshold T2, and the 99th quantile as the third threshold T3 using the quantile calculation method, to ensure that the initial thresholds are based on the fluctuation range of real normal culture scenarios. Then, stage-specific threshold correction terms are determined according to the characteristic differences of the stem cell seeding stage, proliferation stage, and differentiation stage. During the inoculation phase, cells have weak adaptability and are sensitive to abnormalities. The value was 0.05 × T²; during the proliferation phase, cell growth was stable and sensitivity was moderate. The value was -0.03×T2; cells in the differentiation stage have mature functions and slightly higher tolerance to abnormalities. The value is 0.02 × T². The specific threshold correction terms for different cultivation stages are added to the first, second, and third thresholds for correction, resulting in a multi-level threshold. , , .

[0029] Then, the number of abnormal indicators in the multi-dimensional fusion feature set corresponding to the fusion anomaly value is counted, and the proportion of abnormal indicators is calculated. ,like If the proportion of environmental factor-related indicators exceeds 50%, it is judged as an environmental abnormality; if the proportion of culture medium component-related indicators exceeds 50%, it is a component abnormality; if the proportion of culture operation-related indicators exceeds 50%, it is an operation abnormality; if the proportion of cell morphology-related indicators exceeds 50%, it is a morphological abnormality; and if the proportion of metabolite-related indicators exceeds 50%, it is a metabolic abnormality. Simultaneously, the fusion abnormality value is compared with multi-level thresholds. < This indicates no abnormality. ≤ < This indicates a mild abnormality. ≤ < This indicates a moderate abnormality. ≥ This indicates a severe abnormality. Following the association rules between the culture stage and the type of characteristic abnormality, the primary associated abnormality types during the inoculation period are operational and environmental, with secondary association being cytokinesis; during the proliferation period, the primary associated abnormality types are cytokinesis and morphology, with secondary association being environmental; and during the differentiation period, the primary associated abnormality types are metabolic and morphology, with secondary association being operational. The triggering conditions for each association type are then determined, with the primary association type triggered by the percentage of abnormal indicators. ≥50% and the degree of abnormality is moderate or above. Finally, the abnormal state type is determined based on the abnormality type, abnormality degree and triggering conditions. If the primary association triggering condition is met, the abnormal state type is "Primary Association - Type XX - Degree XX Abnormality". If only the secondary association triggering condition is met, it is "Secondary Association - Type XX - Degree XX Abnormality". If both are met, the primary association type takes precedence.

[0030] Preferably, an anomaly warning result is generated based on the anomaly state type and the fused anomaly degree value, including: The basic weights for abnormal state types are determined based on domain knowledge, and sensitivity coefficients are determined based on different cultivation stages. The anomaly correction value is obtained by adjusting the sensitivity of the fusion anomaly value based on the basic weights and sensitivity coefficients. The warning level boundary of the abnormal state type is refined based on the anomaly correction value to obtain the anomaly warning result; the anomaly warning result includes: warning level, abnormal time and abnormal cause indication.

[0031] In the specific implementation process, step 600 first determines the basic weight of the abnormal state type based on knowledge in the field of stem cell culture. The primary correlation anomaly types, due to their more direct impact on culture results, have a base weight set at 0.8-1.0, with environmental and component anomalies assigned 1.0, operational and morphological anomalies assigned 0.9, and metabolic anomalies assigned 0.8. Secondary correlation anomaly types, due to their relatively indirect impact, have a base weight set at 0.5-0.7, with the corresponding secondary correlation terms for the primary correlation types assigned 0.7, 0.6, and 0.5 respectively, to ensure that the base weight matches the priority of the anomaly type's impact. Next, sensitivity coefficients are determined based on different culture stages. During the inoculation stage, cells exhibit weak adaptability and are sensitive to abnormal responses, with a sensitivity coefficient s1 = 1.3. During the proliferation stage, cell growth is stable, and sensitivity is moderate, with a sensitivity coefficient s2 = 1.1. During the differentiation stage, cell function matures, and tolerance is slightly higher, with a sensitivity coefficient s3 = 0.9. Sensitivity adjustments are then made using the formula... Anomaly correction values ​​are calculated to adapt the anomaly values ​​to the importance of anomaly types and the sensitivity of different stages. Based on these correction values, the warning level boundaries are further refined. Combined with the multi-level thresholds corrected in step 500, the warning levels are divided into four levels, with the first-level warning (minor) boundary being... ≤ < The boundary of a Level II warning (mild) is: ≤ < The boundary of a Level 3 (moderate) warning is: ≤ < The boundary of a Level IV (severe) warning is: ≥ Simultaneously, the boundaries are fine-tuned for different cultivation stages: the boundaries at each level are lowered by 10% during the inoculation stage and raised by 5% during the differentiation stage to ensure accurate grading. Finally, anomaly warning results are generated. The anomaly time is directly extracted from the timestamp corresponding to the fusion anomaly value, and the anomaly cause prompt is accurately generated based on the anomaly state type. Finally, the warning level, anomaly time, and anomaly cause prompt are integrated to form a complete anomaly warning result.

[0032] like Figure 2 As shown, the present invention also provides an early warning system for abnormal information in the stem cell culture process, comprising: The data acquisition module is used to acquire multi-source heterogeneous raw data during the stem cell culture process, and to perform outlier identification, repair processing, timestamp alignment and format standardization processing on the multi-source heterogeneous raw data to obtain a unified dataset; The feature extraction module is used to perform hierarchical feature extraction on a unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. The features are updated in real time through a dynamic sliding window to obtain a validated feature set. The feature fusion module is used to perform fusion calculations on the verified feature set through a weighted fusion strategy to obtain a multi-dimensional fused feature set. The anomaly detection module is used to calculate the trend deviation, feature correlation, and fluctuation dispersion of the multi-dimensional fused feature set, and to dynamically assign and fuse the trend deviation, feature correlation, and fluctuation dispersion through feature collaborative contribution to obtain the fused anomaly value. The state discrimination module is used to match the fusion anomaly value with multi-level thresholds to obtain the matching result, and to determine the abnormal state type of the stem cell culture process based on the matching result and the correlation between the characteristic anomaly type and the culture stage. The anomaly warning module is used to generate anomaly warning results based on the anomaly state type and the fused anomaly degree value.

[0033] The beneficial effects of this invention are as follows: 1) Missing and mutation-type real outliers were accurately screened using the quartile method and domain rules. Weighted average method and polynomial fitting interpolation method were used to repair the two types of outliers respectively. At the same time, the timestamps were aligned by linear interpolation. The numerical, categorical and text data were standardized by max-min normalization, one-hot encoding and keyword extraction encoding respectively to form a unified dataset. This ensured data quality from the source, reduced subsequent warning bias and improved the reliability of multi-source heterogeneous data processing. 2) The sliding window length is dynamically adjusted according to the characteristics of the inoculation, proliferation and differentiation stages; the weights of basic statistical features, time series trend features and environmental factor interaction features are updated in real time using the exponential weighting method; the fusion weights of trend deviation, feature correlation and fluctuation dispersion are dynamically corrected based on the feature synergistic contribution; and multi-level thresholds are adjusted through stage-specific threshold correction terms to adapt to the differences in anomaly sensitivity in different stages and batches, thereby improving the targeting of anomaly identification and enhancing the dynamic adaptability of feature extraction and anomaly judgment. 3) By accurately classifying abnormality types into environmental, component, operational, morphological, and metabolic categories based on the proportion of abnormal indicators, and clarifying the primary / secondary correlation of abnormality types in conjunction with the cultivation stage, the severity of mild, moderate, and severe abnormalities is quantified based on multi-level thresholds. The final warning results also include warning level, abnormal time, and abnormal cause indication, realizing the location of the root cause of the abnormality, prediction of the scope of impact, and precise guidance, improving the comprehensiveness and accuracy of abnormality identification, and enhancing the practical application value of the warning. 4) A hierarchical feature extraction strategy is adopted, which not only covers basic statistical features such as mean and variance, but also captures time series trend features through adaptive inflection point detection and quadratic polynomial fitting. Furthermore, the correlation features between environmental factors and cell growth are mined through nonlinear interaction terms and conditional interaction terms. After weighted fusion and PCA redundancy removal, a multi-dimensional core feature set is formed. At the same time, combined with multi-dimensional anomaly quantification of trend deviation, feature correlation, and fluctuation dispersion, the abnormal signals in the cell culture process are comprehensively captured, improving the comprehensiveness and accuracy of anomaly identification. 5) Initial thresholds are determined based on the 90th, 95th, and 99th percentiles of historical normal data, and dynamically adjusted in conjunction with stage-specific correction terms. The abnormality value is then corrected and fused by the basic weight of the abnormal state type and the stage sensitivity coefficient, refining the four-level warning boundary. This makes the warning level both consistent with the objective laws of the data and match the tolerance characteristics of cells at different stages, enhancing the guidance and operability of the warning and optimizing the rationality of the warning level division. 6) Through real-time data acquisition, dynamic feature updates, real-time anomaly calculation and early warning generation throughout the entire process, potential risks such as temperature fluctuations, changes in culture medium composition, and abnormal cell morphology can be captured, thereby providing early warnings and realizing real-time quality control and risk prevention in the culture process. This significantly reduces the probability of failure in large-scale and refined culture, ensuring the safety and effectiveness of stem cell clinical applications.

[0034] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0035] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for early warning of abnormal information in the stem cell culture process, characterized in that, Includes the following steps: Obtain multi-source heterogeneous raw data during stem cell culture, and perform outlier identification, repair processing, timestamp alignment, and format standardization on the multi-source heterogeneous raw data to obtain a unified dataset; Hierarchical feature extraction is performed on the unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. The features are then updated in real time through a dynamic sliding window to obtain a validated feature set. The verified feature set is fused using a weighted fusion strategy to obtain a multi-dimensional fused feature set; Calculate the trend deviation, feature correlation, and fluctuation dispersion of the multi-dimensional fused feature set, and dynamically assign and fuse the trend deviation, feature correlation, and fluctuation dispersion through feature collaborative contribution to obtain the fusion anomaly value; The fusion anomaly value is matched with a multi-level threshold to obtain a matching result. Based on the matching result and the correlation between the characteristic anomaly type and the culture stage, the abnormal state type of the stem cell culture process is determined. An anomaly warning result is generated based on the anomaly state type and the fusion anomaly degree value.

2. The method for early warning of abnormal information in the stem cell culture process according to claim 1, characterized in that, The multi-source heterogeneous raw data includes: culture environment monitoring data, cell morphology observation data, culture operation process data, culture medium composition change data, and cell metabolite detection data; the culture environment monitoring data includes: temperature, pH, dissolved oxygen concentration, CO2 concentration, and osmotic pressure; the cell morphology observation data includes: cell density, cell adhesion rate, cell diameter, and nucleocytoplasmic ratio; the culture operation process data includes: medium change time, additive dosage, stirring rate, and culture dish number; the culture medium composition change data includes: glucose concentration, glutamine concentration, growth factor concentration, and metabolic waste concentration; the cell metabolite detection data includes: lactate concentration, ammonia concentration, reactive oxygen species content, and ATP concentration.

3. The method for early warning of abnormal information in the stem cell culture process according to claim 1, characterized in that, Obtain multi-source heterogeneous raw data during stem cell culture, and perform outlier identification, repair processing, timestamp alignment, and format standardization on the multi-source heterogeneous raw data to obtain a unified dataset, including: Potential outliers in the numerical data of the multi-source heterogeneous raw data were identified by the quartile method, and the potential outliers were screened according to the rules of stem cell culture to obtain the true outliers; the true outliers include: deletion outliers and mutation outliers. The missing outliers are repaired by using a weighted average of adjacent timestamp data, and the abrupt outliers are repaired by using a multinomial fitting interpolation method based on the data trend type, resulting in repaired data. Using the cultivation start time as a benchmark, the missing timestamp data in the repaired data is supplemented by linear interpolation to obtain aligned data; The numerical data in the aligned data is subjected to max-min normalization, the categorical data in the aligned data is subjected to one-hot encoding, and the textual operational data in the aligned data is subjected to keyword extraction and integer encoding to obtain the unified dataset.

4. The method for early warning of abnormal information in the stem cell culture process according to claim 1, characterized in that, Hierarchical feature extraction is performed on the unified dataset to obtain basic statistical features, time series trend features, and environmental factor interaction features. These features are then updated in real time using a dynamic sliding window to obtain a validated feature set, including: The mean, variance, maximum value, minimum value, coefficient of variation, and skewness are extracted from the unified dataset to obtain the basic statistical characteristics; The trend inflection points of the data sequence in the unified dataset are identified by an adaptive inflection point detection algorithm. The time series in the unified dataset is divided into several monotonic trend segments and subjected to quadratic polynomial fitting to obtain the trend equation. The inflection point time and trend segment duration of the trend equation are extracted by the trend inflection point. The slope, intercept, inflection point time and trend segment duration are integrated into the trend features of the time series. The square product of two different environmental factors is used as a nonlinear interaction term. When any of the environmental factors is located in a preset critical interval, the correlation increment between the other environmental factor and the cell growth index is used as a conditional interaction term. The nonlinear interaction term and the conditional interaction term are integrated into the environmental factor interaction feature. The sliding window is determined based on the different stages of stem cell culture. Based on the sliding window, the time weight vectors of the basic statistical features, the time series trend features, and the environmental factor interaction features are determined by the exponential weighting method, and the features of the basic statistical features, the time series trend features, and the environmental factor interaction features are updated according to the time weight vectors to obtain the verified feature set.

5. The method for early warning of abnormal information in the stem cell culture process according to claim 1, characterized in that, Calculate the trend deviation, feature correlation, and volatility dispersion of the multi-dimensional fused feature set, and dynamically assign and fuse the trend deviation, feature correlation, and volatility dispersion using feature collaborative contribution to obtain a fused anomaly value, including: Through formula The trend deviation is calculated; where, The length of the sliding window. For the current eigenvalue, As the baseline eigenvalue, As a penalty factor; Through formula The feature correlation degree is calculated; where, The weights of the Pearson correlation coefficient, The Pearson correlation coefficient between the two features. The mutual information value between two features. The number of features in the multi-dimensional fusion feature set; Through formula The fluctuation dispersion is calculated; wherein, The standard deviation within the sliding window. The mean within the sliding window. The maximum value of the feature within the sliding window. The minimum value of the feature within the sliding window. The degree of fluctuation exceeding the limit; The feature synergy consistency coefficient among the trend deviation, feature correlation, and volatility dispersion is used as the feature synergy contribution for dynamic weight adjustment; the expression for the dynamic weight adjustment process is: ;in, For the original weights, The coefficient of feature coherence; The trend deviation, feature correlation, and volatility dispersion are weighted and fused using the modified dynamic weights to obtain the fused anomaly value.

6. The method for early warning of abnormal information in the stem cell culture process according to claim 1, characterized in that, The fusion anomaly value is matched with multi-level thresholds to obtain matching results. Based on the matching results and the correlation between the characteristic anomaly type and the culture stage, the abnormal state type of the stem cell culture process is determined, including: The first threshold, the second threshold, and the third threshold were determined based on the 90%, 95%, and 99% quantiles of historical normal culture data, respectively. Based on different cultivation stages, stage-specific threshold correction terms are determined, and the first threshold, the second threshold, and the third threshold are corrected using the stage-specific threshold correction terms to obtain the multi-level thresholds; The anomaly type is determined based on the proportion of anomaly indicators in the fusion anomaly value, and the degree of anomaly is determined based on the comparison result between the fusion anomaly value and the multi-level threshold. The characteristic anomaly types are classified into associations based on the cultivation stage to obtain association types; the association types include: primary association anomaly types and secondary association anomaly types; The triggering conditions for the association type are determined based on the proportion of the abnormal indicators and the comparison results, and the abnormal state type is determined based on the abnormal type, the degree of abnormality, and the triggering conditions.

7. The method for early warning of abnormal information in the stem cell culture process according to claim 1, characterized in that, An anomaly warning result is generated based on the anomaly state type and the fusion anomaly degree value, including: The basic weights for abnormal state types are determined based on domain knowledge, and sensitivity coefficients are determined based on different cultivation stages. The fusion anomaly value is adjusted based on the basic weights and the sensitivity coefficients to obtain an anomaly correction value. The warning level boundary of the abnormal state type is refined according to the abnormality correction value to obtain the abnormal warning result; the abnormal warning result includes: warning level, abnormal time, and abnormal cause indication.

8. A stem cell culture process information anomaly early warning system, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous raw data during the stem cell culture process, and to perform outlier identification, repair processing, timestamp alignment and format standardization processing on the multi-source heterogeneous raw data to obtain a unified dataset; The feature extraction module is used to perform hierarchical feature extraction on the unified dataset to obtain basic statistical features, time series trend features and environmental factor interaction features, and to update the features in real time through a dynamic sliding window to obtain a verified feature set. The feature fusion module is used to perform fusion calculation on the verified feature set through a weighted fusion strategy to obtain a multi-dimensional fused feature set; An anomaly detection module is used to calculate the trend deviation, feature correlation, and fluctuation dispersion of the multi-dimensional fused feature set, and to dynamically assign and fuse the trend deviation, feature correlation, and fluctuation dispersion through feature collaborative contribution to obtain a fused anomaly value. The state discrimination module is used to match the fusion anomaly value with multi-level thresholds to obtain the matching result, and to determine the abnormal state type of the stem cell culture process based on the matching result and the correlation between the characteristic anomaly type and the culture stage. An anomaly warning module is used to generate anomaly warning results based on the anomaly state type and the fusion anomaly degree value.

Images