Method and system for constructing stem cell bank

Abstract

The application provides a stem cell bank construction method and system, relates to the technical field of data processing, and comprises the following steps: performing space-time alignment and multi-modal feature fusion on structured numerical sequence and unstructured image features in real-time collected stem cell multi-source data, and generating an ice crystal nucleation risk space distribution field in the fusion process to obtain a multi-source fusion data set carrying the ice crystal nucleation risk space distribution field; determining a space mapping anchor point according to the gradient of the ice crystal nucleation risk space distribution field in the multi-source fusion data set, and adaptively determining a radial cutoff threshold according to the local variance of the ice crystal nucleation risk space distribution field to define a local influence field. Through space-time alignment of multi-source data, ice crystal nucleation risk field construction, space topology reconstruction, semantic standardization and tensor optimization, and in combination with deep reinforcement learning and distributed storage, stem cell cryopreservation whole-process data fusion, risk identification, intelligent regulation and safety traceability are realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for constructing a stem cell bank. Background Technology

[0002] In the existing construction of stem cell banks, the setting and regulation of cryopreservation parameters usually rely on single-type sensor data or offline manual experience rules, making it difficult to fully utilize real-time acquired multi-source heterogeneous information. Especially during the cooling and rewarming stages, characteristics such as temperature gradients, ice crystal nucleation initiation points, and cell morphological changes at different spatial locations within the stem cell cryopreservation container often change simultaneously. However, current technologies generally lack effective means for spatiotemporal alignment and joint modeling of structured numerical sequences (such as multi-point thermocouple temperature and pressure values) and unstructured image features (such as ice crystal texture and membrane fold morphology in cell micrographs). This may lead to an assessment of ice crystal nucleation risk that is biased towards the global average state, ignoring its spatially uneven distribution, thus making it difficult for subsequent regulatory strategies to make differentiated responses to high-risk areas.

[0003] For example, when constructing an umbilical cord mesenchymal stem cell bank, operators place cryopreservation tubes or bags containing stem cell suspensions into a programmed cooling system. This system typically contains several thermocouple probes to collect temperature data at various points. Simultaneously, during cooling, researchers sometimes use the system's built-in microscopic imaging module or an additional inverted microscope to capture images of cells in different fields of view every few seconds or tens of seconds to observe morphological changes such as cell shrinkage and intracellular ice crystal germination. However, existing systems generally process temperature data and microscopic images separately. Temperature data is recorded in real-time and used to control the cooling rate; while image data is mostly saved as separate files and only manually interpreted offline after cryopreservation, or assessed through simple visual comparison. This creates a temporal and spatial disconnect. For instance, if the thermocouples show all temperatures between -20°C and -22°C after the cooling program has run for 5 minutes, the system considers this a safe state and maintains the planned cooling rate of 1°C / min. However, simultaneously, early signs of intracellular ice crystal nucleation (such as tiny bright spots and localized cell membrane indentations) were observed in microscopic images near the bottom of the cryovials, and the rate of cell morphological change in the peripheral field of this area was significantly faster than in the upper region of the container. Because current technology cannot spatiotemporally align and fuse this early nucleation signal from unstructured images with temperature numerical sequences, it is also impossible to construct a spatial distribution field of ice crystal nucleation risk within the cryovial. Therefore, the control system does not adjust any parameters for this high-risk area at the bottom. As cooling continues, ice crystals in the bottom region may develop into destructive large intracellular ice crystals, but the system only triggers rewarming intervention based on a single temperature value when the overall temperature drops below -40°C, at which point most local damage is already irreversible. Post-resuscitation stem cell counting and viability testing results showed that the cell viability of the bottom cryovials was significantly lower than that of the upper region, and batch-to-batch stability was also affected. Summary of the Invention

[0004] This invention provides a method and system for constructing a stem cell bank, enabling intelligent and differentiated regulation of the cryopreservation process and reducing cell damage caused by local ice crystal abnormalities.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for constructing a stem cell bank, the method comprising: Step 1: Perform spatiotemporal alignment and multimodal feature fusion on the structured numerical sequences and unstructured image features in the real-time acquired stem cell multi-source data, and generate a spatial distribution field of ice crystal nucleation risk during the fusion process to obtain a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. Step 2: Determine the spatial mapping anchor point based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and adaptively determine the radial truncation threshold based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field; perform neighborhood feature density aggregation and spatial vectorization reconstruction processing on the local influence field to obtain the spatiotemporal topology reconstruction dataset. Step 3: Perform ontology semantic mapping and cleaning on the spatiotemporal topology reconstruction dataset to construct entity topology and obtain a standardized dataset; Step 4: Discretize the rate of change features of the cooling and reheating stages in the standardized dataset into parameter update rate vectors, quantize the cumulative trend of historical control trajectories into iterative momentum coefficients, and convert the response delay features of state switching into convergence damping coefficients to construct a high-dimensional state transition matrix; perform tensor dimensionality reduction and boundary solution processing on the high-dimensional state transition matrix to obtain the dynamic optimization feature tensor. Step 5: Input the dynamically optimized feature tensor into the preset deep reinforcement learning network for processing to obtain the target recommendation parameter set. Bind the target recommendation parameter set with the spatiotemporal index, and write and persistently store the bound index data stream through a distributed database protocol to complete the construction of the stem cell bank.

[0006] Secondly, the system for constructing stem cell banks includes: The fusion module is used to perform spatiotemporal alignment and multimodal feature fusion on the structured numerical sequences and unstructured image features in the real-time acquired multi-source stem cell data, and generate a spatial distribution field of ice crystal nucleation risk during the fusion process, thus obtaining a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. The processing module is used to determine the spatial mapping anchor point based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and adaptively determine the radial truncation threshold based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field; and perform neighborhood feature density aggregation and spatial vectorization reconstruction processing on the local influence field to obtain the spatiotemporal topology reconstruction dataset. The mapping module is used to perform ontological semantic mapping and cleaning on the spatiotemporal topology reconstruction dataset, construct entity topology, and obtain a standardized dataset. The solution module is used to discretize the rate of change characteristics of the cooling and reheating stages in the standardized dataset into parameter update rate vectors, quantize the cumulative trend of historical control trajectories into iterative momentum coefficients, and convert the response delay characteristics of state switching into convergence damping coefficients to construct a high-dimensional state transition matrix; tensor dimensionality reduction and boundary solution processing are performed on the high-dimensional state transition matrix to obtain a dynamic optimization feature tensor; The storage module is used to input the dynamically optimized feature tensor into a preset deep reinforcement learning network for processing to obtain the target recommendation parameter set. The target recommendation parameter set is then bound to the spatiotemporal index, and the bound index data stream is written and persistently stored through a distributed database protocol to complete the construction of the stem cell library.

[0007] The above-described solution of the present invention has at least the following beneficial effects: This invention performs spatiotemporal alignment and multimodal fusion of multi-source heterogeneous stem cell data, simultaneously constructing a spatial distribution field of ice crystal nucleation risk. This overcomes the shortcomings of traditional discrete data processing, accurately reflecting the spatial differences in nucleation risk within the cryopreservation environment and avoiding judgment errors caused by single global assessments. Based on the gradient and local variance of the ice crystal nucleation risk spatial distribution field, anchor points and radial truncation thresholds are adaptively determined. Local influence fields are autonomously defined, and feature density aggregation and spatial vectorization reconstruction are performed. This effectively mines local feature correlations and spatiotemporal topological correlations within the cryopreservation space, achieving structured topological reconstruction of multi-source fused data. Through ontological semantic mapping and anomaly cleaning filtering, semantic rules and entity correlations of multi-source data are unified, eliminating temporal anomalies and feature noise, improving dataset standardization and consistency. A high-dimensional state transition matrix is ​​constructed from multiple dimensions, including temperature change rate, regulation history, and state delay. Then, through tensor dimensionality reduction and constraint boundary solving, a dynamically optimized feature tensor is obtained, fully characterizing the dynamic evolution law, state switching characteristics, and environmental constraint correlation of the stem cell cryopreservation regulation process. By leveraging deep reinforcement learning networks to intelligently generate adaptive control parameters, combined with spatiotemporal index binding and distributed persistent storage, intelligent and differentiated control of the cryopreservation process can be achieved, reducing cell damage caused by local ice crystal abnormalities, improving stem cell resuscitation activity and batch stability, while ensuring the traceability of stem cell bank data. Attached Figure Description

[0008] Figure 1 This is a schematic flowchart of a method for constructing a stem cell bank provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of a stem cell bank construction system provided in an embodiment of the present invention. Detailed Implementation

[0010] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0011] like Figure 1 As shown, embodiments of the present invention propose a method for constructing a stem cell bank, the method comprising the following steps: Step 1: Perform spatiotemporal alignment and multimodal feature fusion on the structured numerical sequences and unstructured image features in the real-time acquired stem cell multi-source data, and generate a spatial distribution field of ice crystal nucleation risk during the fusion process to obtain a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. Step 2: Determine the spatial mapping anchor point based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and adaptively determine the radial truncation threshold based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field; perform neighborhood feature density aggregation and spatial vectorization reconstruction processing on the local influence field to obtain the spatiotemporal topology reconstruction dataset. Step 3: Perform ontology semantic mapping and cleaning on the spatiotemporal topology reconstruction dataset to construct entity topology and obtain a standardized dataset; Step 4: Discretize the rate of change features of the cooling and reheating stages in the standardized dataset into parameter update rate vectors, quantize the cumulative trend of historical control trajectories into iterative momentum coefficients, and convert the response delay features of state switching into convergence damping coefficients to construct a high-dimensional state transition matrix; perform tensor dimensionality reduction and boundary solution processing on the high-dimensional state transition matrix to obtain the dynamic optimization feature tensor. Step 5: Input the dynamically optimized feature tensor into the preset deep reinforcement learning network for processing to obtain the target recommendation parameter set. Bind the target recommendation parameter set with the spatiotemporal index, and write and persistently store the bound index data stream through a distributed database protocol to complete the construction of the stem cell bank.

[0012] In this embodiment of the invention, spatiotemporal alignment and multimodal fusion are performed on multi-source heterogeneous stem cell data to simultaneously construct a spatial distribution field of ice crystal nucleation risk. This overcomes the shortcomings of traditional discrete data processing, accurately reflecting the spatial differences in nucleation risk within the cryopreservation environment and avoiding judgment errors caused by a single global assessment. Anchor points and radial truncation thresholds are adaptively determined based on the gradient and local variance of the ice crystal nucleation risk spatial distribution field. Local influence fields are autonomously defined, and feature density aggregation and spatial vectorization reconstruction are performed. This effectively mines local feature associations and spatiotemporal topological associations within the cryopreservation space, achieving structured topological reconstruction of multi-source fused data. Through ontological semantic mapping and anomaly cleaning filtering, semantic rules and entity associations of multi-source data are unified, eliminating temporal anomalies and feature noise, and improving the consistency of dataset specifications. A high-dimensional state transition matrix is ​​constructed from multiple dimensions, including temperature change rate, regulation history, and state delay. Then, through tensor dimensionality reduction and constraint boundary solving, a dynamically optimized feature tensor is obtained, fully characterizing the dynamic evolution law, state switching characteristics, and environmental constraint associations of the stem cell cryopreservation regulation process. By leveraging deep reinforcement learning networks to intelligently generate adaptive control parameters, combined with spatiotemporal index binding and distributed persistent storage, intelligent and differentiated control of the cryopreservation process can be achieved, reducing cell damage caused by local ice crystal abnormalities, improving stem cell resuscitation activity and batch stability, while ensuring the traceability of stem cell bank data.

[0013] In a preferred embodiment of the present invention, step 1 includes: Step 100: The real-time collected cell source data, pre-freezing quality inspection data, dispensing parameter data, cooling curve data, and post-thaw activity data are used as multi-source stem cell data, specifically including: Through a pre-set multi-source data acquisition terminal, cell source data, pre-freezing quality inspection data, dispensing parameter data, cooling curve data, and post-resuscitation activity data are collected in real time. The terminal integrates data acquisition, transmission, preprocessing, and control units, and can establish wired or wireless connections with related equipment such as cell separation, quality inspection, dispensing, cooling, and resuscitation. It has core functions such as equipment linkage, synchronous acquisition, preliminary format standardization, and abnormal warning. Cell source data is collected in real time via a multi-source data acquisition terminal linked to the donor information input terminal and cell separation equipment. This includes basic information such as the donor's age, sex, and health status; process parameters during cell separation such as centrifugation speed, centrifugation time, and separation solution ratio; and data such as absorbance values ​​and viable cell counts for initial cell viability testing after separation. Pre-cryopreservation quality control data is collected in real time via a multi-source data acquisition terminal linked to cell quality control equipment and microbial detection instruments. This includes cell concentration (unit: cells / mL), cell purity (unit: %), colony counts for bacterial and fungal contamination, and cell morphology parameters extracted from microscopic images (such as cell diameter and cell membrane integrity). Dispensing parameter data is collected in real time via a multi-source data acquisition terminal linked to the dispensing equipment control system. This includes cryopreservation tube specifications (e.g., 1.5mL, 2mL), dispensing volume per cryopreservation tube (error controlled within ±0.05mL), and dimethyl sulfoxide content in the cryopreservation solution. The data includes the ratio of the sample to serum (e.g., 1:9, 2:8), the start and end times of the dispensing operation, etc.; cooling curve data is collected in real time through a multi-source data acquisition terminal linked to the temperature and pressure sensors built into the cooling equipment. Specifically, this includes the temperature monitoring value (accuracy ±0.1℃) at each time point during the cooling process, the target value and the actual value of the cooling rate control. The target value is preset to 0.5℃ / min to 2℃ / min according to the stem cell type (e.g., the preset target value is 1℃ / min for umbilical cord mesenchymal stem cells and 0.8℃ / min for hematopoietic stem cells). The actual value is the actual cooling rate, compressor power, liquid nitrogen input, and other equipment operating status parameters monitored during the real-time operation of the cooling equipment; post-resuscitation activity data is collected through a multi-source data acquisition terminal linked to the resuscitation equipment and cell activity detection instruments. Specifically, this includes cell survival rate (unit: %), 24-hour cell proliferation rate, cell surface marker expression levels, and other data on cell proliferation capacity and functional integrity.The system synchronously covers the entire process of stem cell collection, from source documentation, pre-cryopreservation quality testing, sample aliquoting, programmed cooling, to post-thawing viability testing, ensuring comprehensive data acquisition. All quantifiable indicators in cell source data, pre-cryopreservation quality control data, aliquoting parameters, cooling curves, and post-thawing viability data are processed into structured numerical information according to a preset data format by the preprocessing unit of the multi-source data acquisition terminal. Cell microscopic images and ice crystal morphology images captured using an inverted microscope before cryopreservation, during cooling, and after thawing are denoised and grayscaled by the image preprocessing module of the multi-source data acquisition terminal, serving as unstructured image information characterizing cell morphological changes and ice crystal growth status. The multi-source data acquisition terminal then categorizes, integrates, standardizes, and deduplicates the aforementioned structured numerical and unstructured image information, forming a complete and formatted multi-source raw data set covering the entire cryopreservation process and containing both types of heterogeneous information.

[0014] Step 101: Extract hardware clock timestamp labels and device physical space coordinate labels from multi-source stem cell data. Perform linear interpolation resampling alignment on the structured numerical sequence based on the hardware clock timestamp labels. Perform perspective mesh space mapping on the unstructured image features based on the device physical space coordinate labels to obtain a spatiotemporal synchronization data block, specifically including: Two types of tags were extracted from the multi-source raw data set of stem cells: hardware clock timestamp tags (accuracy ±0.1s) marked the acquisition time of structured values, and device physical space coordinate tags (accuracy ±0.5mm) marked the three-dimensional position inside the cryopreservation container and the shooting angle parameters corresponding to the unstructured images. Based on timestamp tags, using the master clock of the acquisition device as a unified reference and 1 second as a fixed sampling interval, a linear interpolation algorithm is used to resample and fill in redundancy in structured numerical sequences with different sampling frequencies (e.g., temperature data sampling frequency of 0.5 seconds / time, pressure data sampling frequency of 2 seconds / time) to eliminate timing misalignment and achieve numerical data timing synchronization. At the same time, based on the extracted physical space coordinate tags of the device, perspective mesh spatial mapping processing is performed on the unstructured image features of cell microscopy. First, a fixed three-dimensional physical space coordinate system is established inside the cryopreservation container (with the center point of the bottom of the cryopreservation container as the origin, the vertical direction as the Z-axis, and the horizontal direction as the X and Y axes). The image pixel coordinates of different shooting angles and different fields of view are mapped to this fixed physical space coordinate system through perspective transformation algorithm to eliminate spatial position deviations caused by differences in shooting angles and fields of view. The spatial mapping error is controlled within ±1 mm, completing the spatial position calibration of image features.

[0015] The calibrated structured numerical features (including quantifiable parameters such as temperature, pressure, and cooling rate) and unstructured image features (including image parameters such as cell morphology and ice crystal texture) are matched one-to-one according to spatiotemporal coordinates. This ensures that under each unified time reference (1-second sampling interval) and under each cryopreservation container's three-dimensional physical space coordinates (accuracy ±0.5mm), there is a unique set of structured numerical data and one unstructured image data. Then, according to the dual dimensions of time sequence nodes and spatial coordinates, the two types of calibrated feature data are dimensionally spliced ​​and integrated, so that each data record simultaneously contains numerical features and image features corresponding to the corresponding time sequence and spatial location. This ensures that each time sequence node and each spatial location simultaneously corresponds to complete numerical features and image features, ultimately resulting in a spatiotemporally synchronized data block that is completely synchronized in spatiotemporal dimensions.

[0016] Step 102: Calculate the Pearson correlation coefficient between the numerical channels and image texture channels in the spatiotemporal synchronization data block, and construct a channel feature weight matrix based on the Pearson correlation coefficient; perform an element-wise weighted summation operation on the channel feature weight matrix and the spatiotemporal synchronization data block to obtain a preliminary fused dataset, specifically including: For the spatiotemporal synchronization data block generated in step 101, two major feature channels are clearly defined according to data type and representational meaning: a numerical channel for representing cryopreservation environmental parameters and an image texture channel for representing cell morphology and ice crystal texture. The numerical channel contains quantifiable environmental parameters such as cooling rate, environmental pressure, and cryopreservation solution concentration in a temporal spatial sequence, while the image texture channel contains image features such as cell morphology contours, ice crystal texture density, and cell membrane wrinkling degree in a temporal spatial sequence. For each of the two channels, the Pearson correlation coefficient r is calculated for the temporal spatial feature sequences within each channel. A r A The value range of r is [-1, 1]. A The closer the absolute value of r is to 1, the higher the linear correlation between the trend of numerical channel changes and the morphological changes of image texture channels. A The closer the correlation coefficient is to 0, the lower the linear correlation between the two. Based on the calculated Pearson correlation coefficients for each channel, corresponding feature weights are assigned, with weight values ​​ranging from 0.1 to 0.9. When |r A When |r ≥ 0.8, the corresponding channel weight ranges from 0.8 to 0.9; when 0.6 ≤ |r A When |<0.8, the corresponding channel weight ranges from 0.6 to 0.7; when 0.4≤|r A When |<0.6, the corresponding channel weight takes values ​​between 0.4 and 0.5; when 0.2≤|r A When |<0.4, the corresponding channel weight ranges from 0.2 to 0.3; when |r AWhen |<0.2, the corresponding channel weights range from 0.1 to 0.15. Channels with higher correlation have larger corresponding feature weight proportions, while channels with lower correlation have smaller corresponding feature weight proportions. This constructs a complete channel feature weight matrix, ensuring that the weight matrix accurately reflects the correlation strength and contribution of each channel feature, and that the sum of all channel weights is 1. The constructed channel feature weight matrix is ​​then used to perform an element-wise weighted summation operation with the spatiotemporal synchronization data block. Let A be the feature matrix of the spatiotemporal synchronization data block. S The channel feature weight matrix is ​​W S Both maintain the same dimensions for matrix A. S The numerical feature elements and image texture feature elements at each temporal spatial location are multiplied by the weight matrix W. S The corresponding weighted elements are then summed and fused together. This operation weakens the interference of weakly correlated channel features and strengthens the contribution of strongly correlated channel features, effectively extracting the effective information from the two types of features. After the operation is completed, a preliminary fused dataset is generated.

[0017] Step 103: Calculate the ratio of the local cooling rate to the standard nucleation rate at each data point in the preliminary fusion dataset. Obtain the spatial distribution field of ice crystal nucleation risk through logarithmic transformation and spatial interpolation. Merge this spatial distribution field as an additional channel with the preliminary fusion dataset to obtain a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. Specifically, this includes: Iterate through each data point in the initial fusion data block generated in step 102, corresponding to all temporal spatial data points. Extract the local cooling rate for each data point through data parsing. This local cooling rate is calculated by dividing the difference in temperature monitoring values ​​between two adjacent temporal nodes by the time interval (1 second between adjacent temporal nodes), ensuring that the cooling rate at each spatial location accurately represents the local temperature change trend. Calculate the ratio between the local cooling rate for each data point and the preset standard nucleation rate to obtain the rate ratio results for each spatial location. The standard nucleation rate is a preset critical cooling rate (1℃ / min) based on the characteristics of stem cell cryopreservation, capable of inhibiting ice crystal nucleation. The rate ratio directly reflects the deviation between the local cooling rate and the critical nucleation rate. A ratio greater than 1 indicates that the local cooling rate is too fast, posing a risk of ice crystal nucleation; a ratio less than 1 indicates that the local cooling rate is within a safe range; and a ratio equal to 1 indicates that the local cooling rate has reached the critical value. Perform a uniform logarithmic transformation on the rate ratio results for all spatial locations. ,in These are the eigenvalues ​​after logarithmic transformation. For the rate ratio result, this logarithmic transformation can compress the numerical dynamic range of the rate ratio, while amplifying the subtle rate differences between different spatial locations, making previously indistinguishable rate differences clearly discernible. Spatial interpolation is then performed on the discretely distributed rate feature points after the logarithmic transformation. A linear interpolation algorithm is used, taking two adjacent discrete feature points as endpoints, and filling the spatial feature gaps between the two points through linear fitting. Let the spatial coordinates of two adjacent discrete feature points be... and The corresponding eigenvalues ​​after logarithmic transformation are respectively and The interpolation step size is set to 0.5mm, meaning an interpolation point is selected every 0.5mm. (The interpolation point lies on the line connecting two discrete feature points), calculate the Euclidean distance D between the two points, for the interpolation point located on the line connecting these two points. Let d be the directed distance along the line connecting point 1 and endpoint 1. S Then the eigenvalues ​​of the interpolation points Calculate according to the following formula, d S This method completes the interpolation between all discrete feature points, thereby continuously fitting the spatial distribution field of ice crystal nucleation risk across the entire cryogenic container. This distribution field can present the ice crystal nucleation risk level and distribution pattern at each spatial location within the cryogenic container, i.e., the degree of deviation from the standard nucleation rate (1℃ / min) as a percentage of the rate. The value is divided into three distinct intervals, with larger deviations corresponding to... The larger the absolute value of the difference from 1, the more... <0.5 or >2.0 indicates a serious deviation, 0.5≤ <0.8 or 1.2< ≤2.0 indicates moderate deviation, 0.8≤ ≤1.2 indicates a slight deviation; correspondingly, the eigenvalues ​​after logarithmic transformation The absolute value also increases with the degree of deviation, and the severely deviated interval corresponds to... ≥0.69 (of which <0.5 ≤-0.69, >2.0 ≥0.69), the moderate deviation interval corresponds to 0.22≤ <0.69 (where 0.5≤z<0.8-0.69<) When z ≤ -0.22, 1.2 < z ≤ 2.0, then 0.22 ≤ <0.69), slightly deviating from the interval corresponding to <0.22 (where 0.8≤ <1 hour - 0.22 < <0, 1 < When ≤1.2, 0 < <0.22); The risk level of ice crystal nucleation is specifically divided into three levels. Severe deviation corresponds to level one (high risk), where localized ice crystal nucleation occurs rapidly, easily forming destructive intracellular ice crystals, leading to cell structure damage and decreased activity. Moderate deviation corresponds to level two (medium risk), where localized areas have a risk of slow ice crystal nucleation; if cooling parameters are not adjusted in time, this can gradually develop into significant ice crystal formation, affecting cell cryopreservation quality. Slight deviation corresponds to level three (low risk), where the probability of localized ice crystal nucleation is extremely low, cell morphology and activity can be effectively protected, and cryopreservation safety is high. The greater the deviation, the lower the risk level. The larger the value, the higher the risk level of ice crystal nucleation. Furthermore, the cooling rate reaches a critical value (rate ratio). The region where =1), =0, indicating a critical state for ice crystal nucleation, requiring close monitoring to prevent risk escalation. The spatial distribution field of ice crystal nucleation risk is used as an independent additional feature channel, merged and stitched with the previously obtained preliminary fusion dataset along the channel dimension. This allows the original numerical and image fusion features to be superimposed with spatial risk distribution features, enabling the fusion dataset to simultaneously carry environmental parameters, cell morphology, and spatial risk distribution information, ultimately resulting in a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk.

[0018] This embodiment aggregates multi-source data from the entire stem cell cryopreservation process, overcoming the limitations of single-temperature data acquisition and achieving complete data aggregation across the entire chain. It addresses the separation between numerical and image data through timestamp linear interpolation resampling and spatial coordinate perspective grid mapping, achieving two-dimensional spatiotemporal alignment. By leveraging Pearson correlation coefficients to quantize channel associations and construct a weight matrix, it achieves adaptive fusion of multimodal features, avoiding the drawbacks of fixed weights. Based on the ratio of local to standard nucleation rates, it constructs a spatial distribution field of ice crystal nucleation risk using logarithmic transformation and spatial interpolation, accurately characterizing the non-uniform distribution of risk space. This risk field is then incorporated as an additional channel into the fused dataset, improving the comprehensiveness and accuracy of cryopreservation risk assessment. Finally, the calibrated structured numerical features and unstructured image features are matched one-to-one according to spatiotemporal coordinates, and then spliced ​​and integrated along both temporal and spatial dimensions to ensure that each temporal node and each spatial location corresponds to complete numerical and image features, ultimately resulting in a spatiotemporally synchronized data block with complete spatiotemporal synchronization.

[0019] In a preferred embodiment of the present invention, step 2 includes: Step 200a involves performing spatial gradient calculations on the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and determining the coordinates with the largest gradient magnitude as the spatial mapping anchor point. Specifically, this includes extracting the spatial distribution field data of ice crystal nucleation risk from the multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk obtained in step 103. This data is a set of continuously distributed feature values ​​across the entire cryogenic container area, i.e. ,in The three-dimensional physical space coordinates of the cryopreservation container. This represents the logarithmic transformation characteristic value of ice crystal nucleation risk. A spatial gradient calculation is performed on the spatial distribution field of this ice crystal nucleation risk. The spatial gradient calculation uses a three-dimensional gradient calculation method, and the specific calculation formula is as follows: ; in These are the characteristic values ​​of ice crystal nucleation risk. The partial derivatives in the x, y, and z spatial coordinate axes are calculated using the central difference method, specifically as follows: , , ; in The spatial difference step size is set to 0.5 mm, consistent with the spatial interpolation step size in step 103, ensuring consistency between computational accuracy and spatial dimension. After calculating the gradient vector corresponding to each spatial coordinate, the magnitude of this gradient vector is further calculated. This magnitude characterizes the rate of change of ice crystal nucleation risk at the corresponding spatial location. The larger the magnitude, the more drastic the change in ice crystal nucleation risk at that location, indicating it is a core sensitive point for ice crystal nucleation risk. All spatial coordinates across the entire cryogenic container are traversed, and the coordinate point with the largest gradient magnitude is selected. This coordinate point is determined as the spatial mapping anchor point, denoted as... This anchor point is the core location where the risk of ice crystal nucleation changes most drastically and requires the most key monitoring and control.

[0020] Step 201a: Using the spatial mapping anchor point as the center, calculate the local variance of the spatial distribution field of ice crystal nucleation risk within a spherical region centered on the spatial mapping anchor point and with a fixed initial step size as the radius. Combine the ratio of the local variance to the global variance with a preset reference radius to obtain the radial cutoff threshold. Specifically, this includes: using the spatial mapping anchor point determined in step 200a... Centered on the anchor point, a fixed initial step size of 5mm is first set. This value is determined by considering the typical size of cryopreservation containers and the diffusion range of ice crystal nucleation risk, ensuring coverage of the risk-sensitive area around the anchor point while avoiding computational redundancy due to an excessively large initial range. A spherical region is delineated with the spatial mapping anchor point as the center and the fixed initial step size as the radius. The spatial extent of this spherical region satisfies the formula... Extract the ice crystal nucleation risk feature values ​​corresponding to all spatial coordinates within the spherical region. Calculate the local variance within the spherical region, and simultaneously calculate the global variance of the entire spatial distribution field of ice crystal nucleation risk. Calculate the ratio of the local variance to the global variance. This ratio characterizes the relative strength of risk fluctuations in the area surrounding the anchor point compared to the overall risk fluctuations. A larger ratio indicates a more uneven risk distribution and more severe fluctuations in the area surrounding the anchor point. A preset baseline radius of 8 mm is set, based on the spatial scale and risk diffusion characteristics of the cryogenic container, to define a reasonable range for the radial cutoff threshold. The radial cutoff threshold is calculated by combining the ratio with the preset baseline radius. The specific calculation formula is as follows: ,in With a preset baseline radius (8mm), this formula allows for adaptive adjustment of the radial truncation threshold, especially when the risk fluctuations around the anchor point are severe. When the value is large, the radial cutoff threshold increases, expanding the local influence field and ensuring coverage of all high-risk associated areas; when the risk fluctuations around the anchor point are mild ( When the value is small, the radial truncation threshold decreases, reducing the range of the local influence field and reducing redundant data interference.

[0021] Step 202a: Delineate a spherical retrieval boundary with the spatial mapping anchor point as the center and the radial truncation threshold as the radius. Mark the set of feature coordinates falling within the boundary as the local influence field. Specifically, this includes: using the spatial mapping anchor point determined in step 200a... Using the sphere center as the radial cutoff threshold calculated in step 201a A spherical search boundary is defined with radius , and the spatial extent of this spherical search boundary satisfies the formula The entire spatial distribution field of ice crystal nucleation risk is traversed, and all feature coordinates falling within the spherical search boundary are selected. These feature coordinates are then integrated to form a feature coordinate set. This feature coordinate set is marked as the local influence field, which is the region with the highest correlation to the risk of the spatial mapping anchor point and the most significant risk fluctuations.

[0022] Step 200b involves calculating the Euclidean distance between each feature coordinate within the local influence field and the spatial mapping anchor point, and constructing a Gaussian attenuation weight matrix based on the Euclidean distance. Specifically, this includes: For the local influence field marked in step 202a, extract all feature coordinates within the field and calculate the spatial mapping anchor point for each feature coordinate. Euclidean distance This distance characterizes the spatial correlation between the feature coordinates and the anchor point; the smaller the distance, the higher the correlation, and vice versa. Based on all calculated Euclidean distances... Construct a Gaussian decay weight vector, with each weight coefficient... The calculation formula is: ; in The Gaussian kernel bandwidth is set to the radial truncation threshold. One-third of the weight is used to ensure that the weight coefficients decay smoothly with increasing Euclidean distance, giving higher weights to feature coordinates near the anchor point and lower weights to feature coordinates far away, thus highlighting the contribution of core related features. The constructed Gaussian decay weight vector needs to be normalized so that the sum of all weight coefficients is 1, ensuring the rationality of the weight allocation. The normalization formula is: Among them, the first The original weights of the feature coordinates points This represents the total number of characteristic coordinates within the local influence field. These are the normalized Gaussian decay weight coefficients, ultimately resulting in the set of normalized Gaussian decay weight coefficients.

[0023] Step 201b involves performing kernel density estimation on the feature coordinates within the local influence field based on the Gaussian decay weight matrix to obtain a local feature density distribution map. A covariance matrix is ​​constructed from the local feature density distribution map, and the eigenvalues ​​and corresponding orthogonal eigenvectors are solved. The eigenvalues ​​are then sorted in descending order, and several principal component eigenvectors with a cumulative contribution rate reaching a predetermined proportion are extracted. Specifically, this includes: Based on the Gaussian attenuation weight matrix obtained in step 200b, kernel density estimation is performed on all feature coordinates within the local influence field. The kernel density estimation uses a Gaussian kernel function, and the specific calculation formula is as follows: ; in Spatial coordinates The kernel density estimate at that location, For the first The weight coefficients of each feature coordinate (taken from the corresponding row mean of the Gaussian decay weight matrix). The bandwidth for kernel density estimation is set to 0.3 mm, which matches the spatial interpolation and difference step size to ensure the accuracy of density estimation. It is the number of feature coordinates within the local influence field. , , ) is the first Each feature coordinate is used. This operation yields the kernel density estimate for each spatial coordinate within the local influence field. These density values ​​are then integrated according to their spatial coordinate distribution to form a local feature density distribution map. This map clearly shows the spatial density of features within the local influence field; higher density values ​​indicate stronger feature correlations and represent the core area of ​​risk clustering. A covariance matrix is ​​constructed from the local feature density distribution map, with dimensions of 3×3 (corresponding to three-dimensional spatial coordinates x, y, z). To perform eigenvalue decomposition, the characteristic equation is first constructed. ,in Given a 3×3 identity matrix, by expanding the determinant equation and solving it, we obtain three eigenvalues. (in ); then for each eigenvalue Solve the homogeneous linear equations respectively. The non-zero solutions corresponding to each eigenvalue are obtained, which are the eigenvectors that correspond one-to-one with that eigenvalue. Since the covariance matrix is ​​a real symmetric matrix, the eigenvectors corresponding to different eigenvalues ​​are orthogonal to each other, ultimately resulting in three orthogonal eigenvectors. The magnitude of the eigenvalues ​​characterizes the degree of feature dispersion along the corresponding eigenvector direction. Larger eigenvalues ​​indicate more significant feature differences and higher information content along that direction. After arranging the three eigenvalues ​​in descending order, the cumulative contribution rate is calculated using the following formula: ( ); The cumulative contribution rate is set to 90%, meaning that a number of principal component eigenvectors with a cumulative contribution rate of 90% or higher are selected. Typically, the cumulative contribution rate of the first two or three eigenvectors is sufficient to reach 90%. Then only extract ;like Then cut off ;like Then cut off This method preserves the core features while eliminating redundant information.

[0024] Step 202b involves performing a linear projection transformation on the local feature density distribution map along the directions of several principal component eigenvectors to obtain a dimension-reduced coordinate sequence. Based on this dimension-reduced coordinate sequence, an orthogonal feature basis vector matrix is ​​reconstructed. The orthogonal feature basis vector matrix is ​​then used to perform a spatial basis transformation on the feature coordinate set within the local influence field to obtain a spatiotemporal topology reconstruction dataset, specifically including: The local feature density distribution map obtained in step 201b is subjected to a linear projection transformation along the directions of several principal component feature vectors, i.e. ,in The 3×k coordinates formed by all characteristic coordinates within the local influence field A A matrix (each column corresponds to a feature coordinate). The 3×m vectors formed by the extracted principal component eigenvectors A Matrix (m) A (This refers to the number of principal component eigenvectors extracted). The mean vector of the feature coordinates. For the projected m A ×k A A matrix, where each column corresponds to a dimension-reduced coordinate after projection of a feature coordinate. The matrix... Column-wise extraction yields a projected, dimensionality-reduced coordinate sequence. This sequence retains the core information of the local feature density distribution while reducing data dimensionality and subsequent computational complexity. Based on this dimensionality-reduced coordinate sequence, the orthogonal feature basis vector matrix is ​​reconstructed; that is, a covariance matrix is ​​constructed from the dimensionality-reduced coordinate sequence, and eigenvalue decomposition is performed on the covariance matrix to obtain m. A There are m orthogonal eigenvectors, and the eigenvalue decomposition process is consistent with step 201b. These orthogonal eigenvectors are arranged column-wise to construct m. A ×m A orthogonal eigenvector matrix This matrix can characterize the core feature directions of the reduced feature space. Using this orthogonal feature basis vector matrix, a spatial basis transformation is performed on the set of all feature coordinates within the local influence field. The specific calculation formula for the spatial basis transformation is as follows: ,in m after basis transformation A ×k A The matrix, where each column corresponds to an eigenvector obtained by transforming the eigencoordinates of the matrix using a basis transformation. Ice crystal nucleation risk characteristic values ​​corresponding to each characteristic coordinate within the local influence field The original numerical features and image texture features are correlated and integrated to obtain a spatiotemporal topological reconstruction dataset. This dataset not only retains the core spatial topological correlation of the local influence field, but also integrates multimodal feature information, realizing the structured topological reconstruction of multi-source data.

[0025] This embodiment accurately locates high-risk sensitive points by performing spatial gradient calculations on the spatial distribution field of ice crystal nucleation risk, solving the problems of inability to locate core risk points and indiscriminate treatment of risk areas. It adaptively calculates radial truncation thresholds by combining the ratio of local to global variance and a preset baseline radius, accurately delineating the range of the local influence field and avoiding computational redundancy and omission of core risk areas. A Gaussian decay weight matrix is ​​constructed using Euclidean distance to highlight the contribution of core features around anchor points and weaken redundant interference. Based on this matrix, kernel density estimation is performed to mine risk clustering areas. Eigenvalue decomposition of the covariance matrix and principal component truncation are used to eliminate redundancy, reduce dimensionality, and ensure feature integrity. Through linear projection transformation and orthogonal feature basis vector matrix reconstruction, the topological reconstruction of the local influence field is achieved, forming a spatiotemporal topological reconstruction dataset. This dataset mines spatial local and temporal topological correlations, overcoming the limitations of messy and unclear correlations in multi-source data. Finally, the calibrated structured numerical features and unstructured image features are matched one by one according to spatiotemporal coordinates, and then spliced ​​and integrated according to both temporal and spatial dimensions to ensure that each temporal node and each spatial location corresponds to complete numerical and image features, ultimately resulting in a spatiotemporally synchronized data block that is completely synchronized in both spatiotemporal dimensions.

[0026] In a preferred embodiment of the present invention, step 3 includes: Step 300: Based on the cell batch identifier mapping relationship, storage location coordinate mapping relationship, and quality status evaluation mapping relationship defined in the pre-set ontology semantic mapping rule base, perform field alignment and attribute normalization processing on the spatiotemporal topology reconstruction dataset to obtain a topology map with semantic labels, specifically including: The system invokes a pre-built ontology semantic mapping rule library. This library is a standardized set of semantic mapping rules pre-constructed based on the characteristics of the entire stem cell cryopreservation process data, industry standards, and features of previous multi-source data. It explicitly defines three core mapping relationships: cell batch identifier mapping, storage location coordinate mapping, and quality status assessment mapping. Specifically, the cell batch identifier mapping is used to map cell origin-related features (such as donor number, separation time, and cell type) in the spatiotemporal topology reconstruction dataset to a unified format cell batch identifier. The mapping rule uses a combination format of donor number-separation date-cell type abbreviation. The storage location coordinate mapping is used to map the three-dimensional physical space coordinates in the spatiotemporal topology reconstruction dataset... This mapping is used to standardize the storage location identifiers of cryopreservation containers. The mapping rule is to divide the cryopreservation containers into several layers, columns, and rows. Each layer corresponds to the z-axis coordinate interval, each column corresponds to the x-axis coordinate interval, and each row corresponds to the y-axis coordinate interval. For example, if the z-axis (0 to 5 mm) is the first layer (labeled L01), the x-axis (0 to 5 mm) is the first column (labeled C01), and the y-axis (0 to 5 mm) is the first row (labeled R01), then the storage location identifier after mapping the coordinates (3 mm, 3 mm, 3 mm) is L01-C01-R01. The quality status assessment mapping relationship is used to assign ice crystal nucleation risk feature values ​​from the spatiotemporal topology reconstruction dataset. This is mapped to a standardized cell quality status label, and the mapping rule corresponds to the ice crystal nucleation risk level in step 103, i.e. A value ≥0.69 is mapped to high risk (Level 1), and 0.22≤ <0.69 is mapped to medium risk (level 2). <0.22 is mapped to low risk (level 3). =0 is mapped to a critical state.

[0027] Based on the three types of mapping relationships mentioned above, field alignment and attribute normalization are performed on the spatiotemporal topology reconstruction dataset obtained in step 202b. This involves aligning each data record in the dataset with core fields such as cell batch identifier, storage location identifier, quality status label, spatial coordinates, and feature values ​​according to the field order defined in the ontology semantic mapping rule library, ensuring that the field order of all data records is consistent and without missing values. Attribute normalization specifically involves performing a min-max normalization operation on continuous features in the dataset (such as local cooling rate, feature density value, and Euclidean distance). After normalization, the feature values... The range is unified to [0,1] to eliminate the dimensional differences between different feature dimensions. After field alignment and attribute normalization, a topological map is obtained in which each data record carries standardized semantic tags (cell batch identifier, storage location identifier, quality status tag), that is, a topological map with semantic tags. This map realizes the association between data and semantic information, so that messy topological data has a clear semantic meaning.

[0028] Step 301: Perform sliding window temporal difference calculation on the topological map with semantic labels to obtain a sequence of feature difference values ​​for consecutive time steps; calculate the moving average and standard deviation of the sliding window based on the feature difference value sequence to construct a dynamic baseline threshold interval, specifically including: From the semantically labeled topological map obtained in step 300, time-series feature sequences are extracted, sorted by hardware clock timestamps. The time step interval is uniformly 1 second, and each time step contains complete feature data including semantic labels, spatial coordinates, and normalized feature values. A sliding window length of 5 time steps and a sliding step size of 1 time step are set, and the entire time-series feature sequence is traversed segment by segment. Within each sliding window, the feature difference values ​​between adjacent time steps are calculated for each continuous feature, resulting in the feature difference sequence corresponding to each window. Based on the obtained feature difference value sequence, the moving average of the sliding window within each sliding window is calculated. with standard deviation Based on the calculated moving average and standard deviation, a dynamic baseline threshold interval is constructed, i.e. The coefficient 2 is used to control the coverage of the threshold interval, ensuring that it can include more than 95% of the normal feature difference values, while effectively identifying abnormal jump values. The threshold interval is dynamically updated as the sliding window moves, adapting to the dynamic changes of time series features and avoiding the defect that fixed thresholds cannot adapt to data fluctuations.

[0029] Step 302: Filter feature coordinates falling outside the dynamic baseline threshold range and mark the corresponding feature coordinates as anomalous jump coordinates; perform linear temporal interpolation replacement and Gaussian noise smoothing filtering on the anomalous jump coordinates to obtain a filtered topological map with semantic labels; perform serialization encoding on the filtered topological map with semantic labels to obtain a standardized dataset, specifically including: Iterate through all the feature difference values ​​obtained in step 301, comparing each feature difference value with the dynamic baseline threshold range of the corresponding sliding window. Feature difference values ​​falling outside this threshold range are selected; the corresponding time step feature coordinates are the anomalous jump coordinates. These coordinates correspond to abrupt changes in feature values, which may be due to data acquisition errors, equipment operation fluctuations, or abnormal local ice crystal growth. For the selected anomalous jump coordinates, perform linear temporal interpolation replacement processing based on adjacent valid data points. That is, for each anomalous jump coordinate corresponding to a time step... Find the two nearest valid data points before and after it, i.e., the time steps. and time step The corresponding feature coordinates (the feature difference values ​​of the two data points both fall within the dynamic baseline threshold range), set the time step eigenvalues Time step eigenvalues The time step is calculated using a linear interpolation formula. Replacement eigenvalues The calculation formula is: ; The calculated The original feature values ​​of the anomalous jump coordinates are replaced to complete the initial correction of the anomalous data. The replaced topological map with semantic labels is then subjected to Gaussian noise smoothing filtering to eliminate potential local data fluctuations after interpolation. The Gaussian noise smoothing filter uses a 3×3×3 spatial filtering window (corresponding to three-dimensional spatial coordinates), and the weight coefficients of the filtering window are calculated using a Gaussian function. The specific calculation formula is as follows: ,in This represents the offset relative to the center pixel within the filtering window. The standard deviation of the filter is set to 0.5, which matches the previous spatial interpolation and difference step size to ensure filtering accuracy. The filtering window iterates through each feature coordinate of the semantically labeled topological graph, calculating the weighted sum of all feature values ​​and their corresponding weight coefficients within the window. This weighted sum is then used as the filtered feature value at the center feature coordinate of the window. The calculation formula is as follows: ; in It is the filtered output characteristic value obtained after smoothing the current spatial coordinates point through a three-dimensional Gaussian noise filter. Offset within the filter window The corresponding feature values ​​are used to smooth the data and reduce noise interference. After filtering, a topological map with semantic labels is obtained. This map eliminates anomalous jumps and noise interference, significantly improving data stability. Serialization encoding is then performed on the filtered topological map with semantic labels. One-hot encoding is used to encode the semantic labels (cell batch identifier, storage location identifier, and quality status label), converting discrete semantic labels into continuous encoded vectors. For continuous data such as spatial coordinates and normalized feature values, their numerical form is preserved. The encoded semantic label vectors are concatenated with the continuous data in field order to obtain a standardized dataset where each data record is a fixed-length vector. This dataset has a uniform format, no anomalies, and clear semantics.

[0030] This embodiment achieves field alignment and attribute normalization of the spatiotemporal topology reconstruction dataset by pre-setting an ontology semantic mapping rule base, solving problems such as data semantic ambiguity, field confusion, and inconsistent dimensions, and providing a standardized foundation for data interpretation, application, and comparison of data from different batches. Through sliding window temporal difference calculation and the construction of dynamic baseline threshold intervals using moving averages and standard deviations, it accurately identifies anomalous data jump coordinates, improving the accuracy and adaptability of anomaly identification. Linear temporal interpolation replacement of adjacent valid data points and Gaussian noise smoothing filtering are used to correct anomalous data and eliminate noise, ensuring data integrity and reliability. A standardized dataset is obtained through serialization encoding, achieving unified formatting of multi-source heterogeneous data, reducing the complexity of subsequent processing, while preserving data semantics and spatiotemporal topology associations. Finally, the calibrated structured numerical features and unstructured image features are matched one-to-one according to spatiotemporal coordinates, and then spliced ​​and integrated according to both temporal and spatial dimensions, ensuring that each temporal node and each spatial location corresponds to complete numerical and image features, ultimately obtaining a spatiotemporally synchronized data block with complete spatiotemporal synchronization.

[0031] In a preferred embodiment of the present invention, step 4 includes: Step 400a: Extract the temperature control time series curves for the cooling and reheating stages from the standardized dataset to obtain a temperature control time series curve sequence; calculate the first-order difference sequence of the temperature control time series curve sequence within a continuous sampling period; perform fast Fourier transform and bandpass filtering on the first-order difference sequence to extract the dominant frequency component with the highest energy proportion and discretize and encode it to obtain the parameter update rate vector, specifically including: From the standardized dataset obtained in step 302, based on the stage identifier associated with the hardware clock timestamp and quality status label, the temperature control time-series curves corresponding to the cooling and rewarming stages in the entire stem cell cryopreservation process are extracted. The cooling stage corresponds to the time interval where the ice crystal nucleation risk characteristic value gradually changes from the initial state (characteristic value range 0 to 0.22) to the stable state (characteristic value range 0.22 to 0.69). The rewarming stage corresponds to the time interval where the ice crystal nucleation risk characteristic value gradually recovers from the stable state (characteristic value range 0.22 to 0.69) to the activity detection state (characteristic value range 0 to 0.22). The temperature control time-series curves of the two stages are integrated in time-stamp order to obtain a temperature control time-series curve sequence. The sampling interval of this sequence is consistent with steps 101 and 301, both being 1 second, to ensure a unified time-series benchmark. For this temperature control time-series curve sequence, a first-order difference sequence is calculated point-by-point within a continuous sampling period. The first-order difference calculation is used to characterize the rate of change of temperature control parameters at adjacent sampling times, i.e. ,in For the first Temperature control timing values ​​at the sampling time (unit: °C). For the first Temperature control timing values ​​at the sampling time. For the first The first-order difference value corresponding to the sampling period is used to traverse the entire temperature control time series curve to obtain the complete first-order difference sequence.

[0032] Perform a Fast Fourier Transform (FFT) on the obtained first-order difference sequence to convert the difference signal in the time domain into an amplitude-frequency distribution in the frequency domain. The specific FFT operation is as follows: Let the first-order difference sequence be... ( (where the difference sequence length is), the transformation formula is: ,in , The imaginary unit, Pi is a constant, and after transformation, a frequency domain amplitude sequence is obtained. Bandpass filtering is performed on the frequency domain amplitude sequence, with the bandpass frequency range set from 0.01Hz to 1Hz. This range is set in conjunction with the fluctuation characteristics of the temperature control parameters, effectively filtering out low-frequency trend components (caused by steady-state drift of the equipment) and high-frequency random noise components (caused by data acquisition interference), while retaining the effective frequency components corresponding to temperature control fluctuations. The signal energy percentage corresponding to each frequency component after filtering is statistically analyzed, and the signal energy calculation formula is... The formula for calculating the energy percentage is: The frequency component with the largest energy proportion is selected as the dominant frequency component. The three core parameters of the dominant frequency component, namely amplitude, frequency and phase, are subjected to equally spaced discretization and hierarchical encoding to convert the continuous physical quantity into a regular numerical vector form. The encoding interval is uniformly normalized to [0,1], which is consistent with the previous normalization standard. Finally, a parameter update rate vector is generated, which is used to characterize the dynamic change rate of temperature control parameter fluctuations.

[0033] Step 401a involves extracting the time-series evolution trajectory of historical control parameters from the standardized dataset, performing numerical integration and accumulation on the time-series evolution trajectory, and performing extreme value normalization mapping on the accumulation result to obtain the iterative momentum coefficients. Specifically, this includes: From the standardized dataset obtained in step 302, the evolution trajectory of historical control parameters of the stem cell cryopreservation equipment (including cooling rate control values, liquid nitrogen input volume, and resuscitation temperature control values) over time is extracted and arranged in the order of hardware clock timestamps to form a time series evolution trajectory of historical control parameters. This trajectory is consistent with the time reference of the temperature control time series curve sequence. Numerical integration accumulation is performed on this time series evolution trajectory using a trapezoidal numerical integration method, i.e. ;in For the first Historical control parameter values ​​at any given time It is the first Historical control parameter values ​​at any given time For the first The hardware clock timestamp at any given moment. The sampling time interval is fixed (1s). For the first The integral accumulation results at each time point are used to traverse the entire time series evolution trajectory, resulting in a complete set of integral accumulation results. This set of integral accumulation results is then subjected to extreme value normalization mapping to eliminate the dimensional differences between different control parameters. ;in For the cumulative result of a single integral, It is the minimum value among all accumulated integrals. The maximum value among all cumulative integral results is normalized and mapped to a value range between [0,1], consistent with the previous attribute normalization standard. This normalized result is determined as the iteration momentum coefficient, which is used to give the subsequent parameter iteration process historical memory characteristics and avoid iteration oscillation divergence.

[0034] Step 402a: Perform storage operation mode state recognition processing on the standardized dataset, capture the transition timestamps of mode switching between the programmed cooling storage mode, liquid nitrogen cryogenic preservation mode, and programmed rewarming recovery mode, and obtain the mode switching transition timestamp sequence, specifically including: Based on the standardized data obtained in step 302, the state characteristic parameters (including temperature range, pressure parameters, cooling / reheating rate, equipment operating power, etc.) corresponding to the centralized storage operation mode are stored. The storage operation mode state recognition processing is performed. According to the inherent characteristic thresholds of the three storage operation modes, the programmed cooling storage mode, liquid nitrogen cryogenic preservation mode and programmed reheating recovery mode are automatically distinguished and identified. Among them, the characteristic threshold of the programmed cooling storage mode is -80℃≤-20℃, and the cooling rate is 0.5℃ / min≤2℃ / min. The characteristic threshold of the liquid nitrogen cryogenic preservation mode is the temperature range, that is, the characteristic threshold of the liquid nitrogen cryogenic preservation mode is -201℃≤-191℃, the pressure range is 0.1MPa≤0.15MPa, and the equipment operating power is ≥30% of the rated power and ≤50% of the rated power. The characteristic threshold of the programmed reheating recovery mode is the temperature range, that is, the characteristic threshold of the programmed reheating recovery mode is -80℃≤37℃, and the reheating rate is ≥5℃ / min and ≤10℃ / min. The system captures the hardware clock moment corresponding to the instant when the three working modes switch between each other in real time. At this moment, the state characteristic parameters change abruptly and fully meet all the characteristic threshold requirements of the corresponding target mode. This moment is used as the transition timestamp of the mode switch. All transition moments are integrated in chronological order to form a mode switch transition timestamp sequence. This sequence records the specific moments when the three storage operation modes (programmed cooling storage mode, liquid nitrogen cryogenic preservation mode, and programmed rewarming mode) switch between each other in chronological order. It clearly characterizes the switching order, switching interval, and chronological evolution of storage operation modes, and can intuitively reflect the frequency and time distribution characteristics of different mode switches.

[0035] Step 403a involves calculating the absolute time series deviation between the mode switching transition timestamp sequence and the theoretical response timestamp, inputting the absolute time series deviation into the exponential decay mapping function to obtain the convergence damping coefficient; and then orthogonally concatenating the parameter update rate vector, iteration momentum coefficient, and convergence damping coefficient along the feature dimension to construct a high-dimensional state transition matrix, specifically including: The theoretical response timestamps for the stem cell cryopreservation process are pre-defined. These timestamps represent the theoretical timing of mode switching based on the optimal process parameters for stem cell cryopreservation. The theoretical response timestamp for switching from programmed cooling storage mode to liquid nitrogen cryopreservation mode is 3600 s (1 h) after cryopreservation initiation. The theoretical response timestamp for switching from liquid nitrogen cryopreservation mode to programmed rewarming mode is the full preset cryopreservation duration (e.g., 72000 s, 20 h). The theoretical response timestamp for switching from programmed rewarming mode to subsequent detection mode is 1800 s (0.5 h) after rewarming initiation. All theoretical response timestamps are consistent with the time base of the mode switching transition timestamp sequence (all using the stem cell cryopreservation initiation time as the zero point, with the unit uniformly in seconds). For each actual transition time in the mode switching timestamp sequence, an absolute time series deviation calculation is performed between each actual transition time and the corresponding theoretical response timestamp. ,in This is the actual mode switching transition timestamp (unit: seconds). This is the corresponding theoretical response timestamp (unit: seconds). For the absolute timing bias of a single mode switch, iterate through the entire timestamp sequence to obtain the set of absolute timing biases for all mode switches. Input each absolute timing bias into an exponentially decaying mapping function to map the timing bias to a convergence damping coefficient. The expression for the exponentially decaying mapping function is as follows: ,in The preset time decay constant is set to 30 seconds. This value, combined with the allowable deviation range for mode switching, maps to the output convergence damping coefficient. The value range is (0,1]. The smaller the time series deviation, the closer the convergence damping coefficient is to 1; the larger the time series deviation, the closer the convergence damping coefficient is to 0. This is used to quantify the impact of mode switching hysteresis deviation on system convergence. The parameter update rate vector obtained in step 400a, the iterative momentum coefficient obtained in step 401a, and the convergence damping coefficient obtained in this step are taken as independent feature dimensions. Orthogonal splicing is performed in the order of parameter update rate vector, iterative momentum coefficient, and convergence damping coefficient to ensure that the features of each dimension are orthogonal and uncoupled. After splicing, a high-dimensional state transition matrix is ​​constructed. This matrix fully carries the three types of coupled information: temperature control rate, iterative momentum, and mode switching damping, realizing a high-dimensional unified representation of multi-source control features.

[0036] Step 400b involves performing CP tensor decomposition on the high-dimensional state transition matrix to extract the dominant factor matrix and core tensor blocks, specifically including: The high-dimensional state transition matrix constructed in step 403a is subjected to CP tensor decomposition. This high-dimensional state transition matrix fully carries three types of coupled information: temperature control rate, iteration momentum, and mode switching damping. It serves as a high-dimensional unified representation of multi-source control characteristics. The decomposition process maintains the original dimensional structure of the matrix unchanged, without altering the arrangement order or length of the spatial, temporal, and feature dimensions. First, the high-dimensional state transition matrix is ​​transformed into a multi-dimensional tensor form. Let this multi-dimensional tensor be... Its dimensional scale is ,in Corresponding row dimension (spatial-temporal topological location dimension). The corresponding column dimension (the dimension of the control parameter characteristics) corresponds to the time series dimension and the environmental stress correlation dimension. Let be the total number of dimensions. Using the classic CP tensor decomposition model, this multidimensional tensor is decomposed into a sum of rank components. The specific decomposition expression is as follows: ; In the formula, The rank of the CP tensor decomposition is preset, and the value is set according to the dimension and feature complexity of the high-dimensional state transition matrix, with a value of 5 to 10 (the value is reasonably set in combination with the number of features of the stem cell cryopreservation regulation parameters). For the first The weight coefficients corresponding to the rank are used to quantify the contribution of that rank component to the overall tensor. Their calculation is obtained by minimizing the decomposition error using the least squares method, with the specific error objective function being... ,in This represents the Frobenius inner product. By iteratively solving this objective function, the weight coefficients corresponding to each rank can be obtained. ; Representing the Dimension 1 The dimension factor vector of order, the vector dimension and the corresponding tensor dimension Consistency; Symbol This represents the outer product operation of vectors. After decomposition, the dominant factor matrix and the core tensor block are extracted, where the dominant factor matrix consists of factor vectors of each dimension. The elements are arranged and integrated based on the row dimension (spatial-temporal topological location dimension). The weight proportion of each dimension factor vector is obtained by normalizing the weight coefficients. First, all weight coefficients are calculated. The sum is then calculated, and each weight coefficient is calculated separately. The weight percentage is the proportion of the total, with the weight percentage ranging from [0,1], and the sum of the weight percentages of all factor vectors is 1. The dominant factor matrix first integrates and arranges the factor vectors of each dimension. Taking the spatiotemporal topological position dimension factor vector corresponding to the row dimension as the benchmark, the value of each element inside the row dimension factor vector is multiplied by the normalized weight percentage corresponding to that factor vector in turn to obtain the weighted element value. The magnitude of the weighted element value is used to quantify the strength of the characteristic change of the spatiotemporal topological position corresponding to the row dimension. The larger the weighted element value, the more violent the fluctuation of the temperature control parameters and environmental stress characteristics at that spatiotemporal topological position, and the higher the intensity of the characteristic change. The smaller the weighted element value, the smoother the characteristic change at the corresponding position, and the lower the intensity of the characteristic change. Next, factor vectors corresponding to different row positions in the dominant factor matrix are selected, and Pearson correlation coefficients are calculated for any two pairs of factor vectors. The correlation coefficient ranges from [-1, 1]. The magnitude of the correlation coefficient represents the temporal evolution trend between different spatiotemporal topological positions. When the correlation coefficient approaches positive, the temporal evolution trends of the two positions tend to be synchronous; when the correlation coefficient approaches negative, the temporal evolution trends of the two positions show opposite changes; when the correlation coefficient approaches zero, there is no obvious temporal correlation between the two positions. By comprehensively statistically analyzing the distribution characteristics of the weighted element values ​​of all row positions and the distribution pattern of the Pearson correlation coefficients between factor vectors of all row positions, the local change characteristics and overall evolution correlation of the row dimension can be characterized layer by layer, fully representing the global change pattern of the row dimension. The core tensor block is composed of the rank component tensors of each order obtained from CP tensor decomposition, which are accumulated and integrated step by step. It can completely retain the inherent coupling correlation characteristics and cross-dimensional interaction constraint relationships between the dimensions of multidimensional tensors, and centrally carry the multi-parameter regulation correlation information hidden inside the high-dimensional state transition matrix and the constraint correlation information in the multi-mode switching process.

[0037] Step 401b involves performing mode expansion and singular value decomposition on the core tensor block to obtain a singular value sequence. This sequence is then sorted in descending order and accumulated item by item to construct a singular value cumulative energy distribution sequence. A second-order difference calculation is performed on the singular value cumulative energy distribution sequence to locate abrupt changes in the distribution slope. The cumulative energy percentage corresponding to each abrupt change in the distribution slope is determined as the environmental stress tolerance threshold. Specifically, this includes: The core tensor block obtained from step 400b is subjected to pattern expansion and singular value decomposition. First, the core tensor block is pattern-expanded according to a preset dimensional priority order (spatial dimension - temporal dimension - feature dimension). The core of pattern expansion is to transform the multidimensional tensor mapping into a two-dimensional extended matrix. During the expansion process, a single dimension is fixed as the matrix row vector dimension, and all other dimensions are sequentially concatenated to form the matrix column vector dimension, ensuring that the feature information of the tensor is not lost or distorted before and after expansion. Specifically, the spatial dimension (corresponding to the row dimension, i.e., the spatiotemporal topological position dimension) of the core tensor block is selected as the row dimension of the two-dimensional matrix, and the temporal dimension, feature dimension, and other related dimensions are sequentially concatenated as the column dimensions of the two-dimensional matrix. Let the dimensionality of the core tensor block be... ( For spatial dimensions, For time series dimension, (where the feature dimension is 2), the expanded two-dimensional matrix is ​​obtained after expansion. Dimension size is The value of each element in the matrix is ​​completely consistent with the value of the element at the corresponding dimension position of the core tensor block, realizing the mapping and transformation of multi-dimensional features to two-dimensional space.

[0038] Perform singular value decomposition on the expanded two-dimensional matrix. The specific decomposition expression is as follows: ,in It is a left singular vector matrix with a dimension of . Its column vectors are matrices eigenvectors; It is a singular value diagonal matrix with a dimension of . The elements on the diagonal form a matrix. The singular value of is 0, and all other elements are 0; A right singular vector matrix The transpose of the matrix, with dimensions of Its column vectors are matrices The eigenvectors of the singular value decomposition. The specific calculation process of singular value decomposition is to calculate the matrix... ,get Square matrix of order; construct characteristic equation ,in For the matrix For an identity matrix of the same dimension, solving the characteristic equation yields all eigenvalues. After filtering all eigenvalues ​​by nonnegativity, taking the square root yields the singular values. All singular values ​​form a sequence of singular values, which, along with the matrix... The order is consistent; for each singular value, solve the homogeneous linear equation system. The corresponding feature vectors are obtained. Integrating all eigenvectors yields the right singular vector matrix. Then through The left singular vector matrix is ​​calculated. The entire singular value decomposition process is completed, ultimately yielding a complete sequence of singular values.

[0039] The obtained singular value sequence is sorted in descending order, ensuring that the values ​​decrease sequentially and are not less than 0. The sorted singular value sequence is then subjected to a cumulative energy calculation, successively determining the cumulative energy percentage corresponding to each singular value. This is calculated as: the cumulative energy percentage at a certain order = the sum of all singular values ​​at that order / the total sum of all singular values. After calculating at each order, the cumulative energy percentages of all orders are integrated to construct a complete singular value cumulative energy distribution sequence.

[0040] The second-order difference calculation is performed point-by-point on the singular value cumulative energy distribution sequence. The core function of the second-order difference is to locate the abrupt change points in the slope of the cumulative energy distribution, i.e. ,in For the first The percentage of cumulative energy corresponding to each singular value It is the first The percentage of cumulative energy corresponding to each singular value It is the first The cumulative energy percentage corresponding to each singular value. A threshold of 0.05 is set for the second-order difference value mutation. When the second-order difference value at a certain position is greater than 0.05, and the second-order difference values ​​before and after that position show a trend of increasing from small to large, that position is determined to be a slope mutation point in the singular value cumulative energy distribution. The cumulative energy percentage corresponding to this mutation point is determined as the environmental stress tolerance threshold. This threshold is used to constrain the boundary range of subsequent feature optimization and iterative solution, ensuring that the optimization results meet the environmental stress tolerance limit of stem cell cryopreservation and avoiding invalid feature solutions that exceed the cell cryopreservation tolerance range.

[0041] Step 402b: Extract the column and row dimension parameters of the dominant factor matrix to construct a multidimensional feature optimization space. Based on the environmental stress tolerance threshold, perform convex hull constraint boundary delineation within the multidimensional feature optimization space to obtain the convex hull constraint boundary, specifically including: Extract the column and row dimension parameters of the dominant factor matrix obtained in step 400b. The row dimension parameters correspond to the spatiotemporal topological location dimension, specifically the spatial monitoring point coordinates of the stem cell cryopreservation device (e.g., X-axis coordinate xi, Y-axis coordinate yi). The column dimension parameters correspond to the feature dimensions of the control parameters, specifically the feature values ​​of core control parameters such as temperature control rate, iteration momentum, and convergence damping coefficient. Ensure that the extracted parameters are consistent with the parameters of the previously standardized dataset and the high-dimensional state transition matrix. Construct a closed and continuous multidimensional feature optimization space with the row dimension parameters as the horizontal axis and the column dimension parameters as the vertical axis. The dimension of this space is consistent with the dimension of the dominant factor matrix. If the dominant factor matrix is ​​a two-dimensional matrix (row dimension × column dimension), a two-dimensional feature optimization space is constructed; if the dominant factor matrix is ​​a multidimensional extended matrix, a three-dimensional or higher multidimensional feature optimization space is constructed. The spatial range completely covers the value range of all row and column dimension parameters in the dominant factor matrix.

[0042] Based on the environmental stress tolerance threshold determined in step 401b, the convex hull algorithm is used to perform convex hull constraint boundary delineation within the multidimensional feature optimization space. The core of the convex hull algorithm is to find the smallest convex polygon (two-dimensional space) or convex polyhedron (three-dimensional and above space) that can enclose all valid feature points in the multidimensional space. The specific execution process is as follows: traverse all feature points in the multidimensional feature optimization space. Each feature point is composed of row dimension parameters and column dimension parameters, that is, the feature point coordinates are (xi, yj) (two-dimensional space) or (xi, yj, zk) (three-dimensional space), where xi is the row dimension parameter and yj and zk are the column dimension parameters. Feature points with a cumulative energy percentage greater than or equal to the environmental stress tolerance threshold are selected as valid feature points, while invalid feature points with a cumulative energy percentage less than the environmental stress tolerance threshold are removed to ensure that valid feature points meet the environmental stress constraint requirements. Based on the dimension of the space containing the valid feature points, the corresponding convex hull solving algorithm is selected. If it is a two-dimensional space, the Graham scan method is used to solve the convex hull. First, the point with the smallest Y-axis coordinate among the valid feature points is selected as the starting point. If there are multiple points with the same Y-axis coordinate, the point with the smallest X-axis coordinate is selected. Then, the remaining valid feature points are sorted in ascending order of their polar angle with the starting point. The polar angle is calculated with the starting point as the origin and the positive X-axis as the polar axis. The polar angle calculation formula is: =arctan2(y-ye,x-xe), where (xe, ye) are the coordinates of the starting point and (x, y) are the coordinates of the remaining valid feature points; connect the sorted valid feature points in sequence, determine whether the triangle formed by three adjacent points is in a clockwise direction, if it is in a clockwise direction, remove the middle point, until all valid feature points are traversed, and finally form a closed convex polygon.

[0043] For three-dimensional or higher spaces, the Quickhull algorithm (or a high-dimensional convex hull solution based on linear programming) is used. Taking three-dimensional space as an example, first, all valid feature points are traversed, and the six extreme points corresponding to the maximum and minimum coordinate values ​​in the X, Y, and Z axes (i.e., the points corresponding to Xmin, Xmax, Ymin, Ymax, Zmin, and Zmax) are found respectively. From these six extreme points, four non-coplanar points are selected (for example, Xmin, Ymin, Zmin, and the point farthest from the plane formed by these three points) to construct the initial convex hull tetrahedron. Then, the remaining feature points are traversed. If a point is located outside the current convex hull, the point is added to the convex hull boundary, and the convex hull surface is recursively expanded, deleting occluded edges and faces, until all points are located inside the convex hull or on the boundary, finally obtaining a closed convex polyhedron. For higher dimensions, the algorithm principle remains the same, but the extreme points need to be the points corresponding to the minimum and maximum values ​​along each dimensional axis. The initial convex hull is constructed as a simplex of the corresponding dimension, and the final convex hull is represented as a convex polytope. The final convex hull boundary (the boundary line of a convex polygon in 2D, and the surface of a convex polyhedron in 3D and above) is used as the convex hull constraint boundary. All feature regions outside this boundary are removed, and the effective feature regions within the boundary are retained, thus obtaining the final convex hull constraint boundary.

[0044] Step 403b involves projecting the core tensor block onto the convex hull constraint boundary to obtain the projected core tensor block. Gradient projection iterative solution processing is then performed on the projected core tensor block within the convex hull constraint boundary to obtain the iterative solution result, specifically including: The core tensor block obtained in step 400b is projected onto the convex hull constraint boundary defined in step 402b. The projection mapping process uses an orthogonal projection algorithm to ensure that the feature information of the core tensor block is not distorted after projection and that it all falls within the effective region defined by the convex hull constraint boundary. The specific projection mapping calculation formula is as follows: ,in This refers to the core tensor block obtained from step 400b. Let be the projection matrix corresponding to the convex hull constraint boundary. This is the core tensor block after projection mapping. Projection matrix. The construction process is as follows: First, extract the coordinates of all vertices of the convex hull constraint boundary and organize them into a list. Then, construct a boundary vector matrix with any vertex as the origin. Perform Schmitt orthogonalization on this matrix to obtain an orthogonal matrix. Finally, normalize the column vectors of the orthogonal matrix so that the norm of each column vector is 1, and finally obtain the projection matrix. The dimensions of the projection matrix are kept consistent with those of the core tensor block to ensure the feasibility of the projection operation. During the projection process, invalid feature components that exceed the convex hull constraint boundary in the core tensor block are directly discarded, and only valid feature components that fall within the boundary range are retained. This results in the core tensor block after projection mapping, which retains the core coupling correlation information and meets the environmental stress tolerance threshold requirements.

[0045] Within the optimization space defined by the convex hull constraint boundary, the core tensor block after projection mapping is solved using gradient projection iteration. The core objective of the iteration is to find the tensor solution that satisfies the environmental stress constraint and minimizes the gradient norm, ensuring the smoothness and rationality of the iteration results. The complete objective function is: ,in Core tensor block after projection mapping The gradient; The effective region is defined by the convex hull constraint boundary, and the forced constraint iterative solution always falls within the convex hull constraint boundary. The specific process of iterative solution is as follows: initialize the iteration parameters, and set the initial iterative solution. (i.e., the core tensor block after projection mapping), iteration step size The threshold value is 0.01 (set reasonably based on data accuracy and iteration efficiency), and the iteration termination threshold is... The iteration counter t=0; calculate the current iteration solution. gradient and gradient norm The iterative solution is updated along the negative gradient direction, and the update formula is as follows: Determine the updated iterative solution. Does it fall within the convex hull constraint boundary? If the solution falls within the boundary, the iterative solution is retained; if it exceeds the boundary, it is transformed by orthogonal projection. Projecting back into the convex hull constraint boundary, we obtain the corrected iterative solution. Calculate the gradient norm difference between two consecutive iterations. If the gradient norm difference is less than the iteration termination threshold, the iteration termination condition is met, the iteration stops, and the current iteration solution is taken as the final iteration solution. If the termination condition is not met, let t = t + 1, and repeat the iteration until the iteration termination condition is met, finally obtaining the iteration solution. Step 404b involves truncating and compressing redundant degrees of freedom in the iterative solution results whose gradient norm is less than the tolerance threshold, resulting in a compressed dominant factor matrix and a compressed core tensor block. The compressed dominant factor matrix and the compressed core tensor block are then dimensionally aligned and segmented. Using the indices of each dimension of the compressed core tensor block as a reference, the eigenvectors of the corresponding dimensions in the compressed dominant factor matrix are sequentially expanded by outer product and element-wise multiplicative mapping according to their index order, resulting in a multidimensional block-structure tensor. Dimensional rearrangement and axial concatenation are then performed on the multidimensional block-structure tensor to obtain a dynamically optimized feature tensor carrying multidimensional optimization constraint information, specifically including: The preset gradient norm tolerance threshold is This threshold, set in conjunction with the data accuracy requirements of stem cell cryopreservation regulation parameters, is used to determine whether the degrees of freedom in each dimension of the iterative solution are redundant. Degrees of freedom with gradient norms less than this tolerance threshold have minimal impact on the overall optimization result and contain a large amount of redundant information and noise, thus they can be truncated and compressed. The gradient norms of all dimensions in the iterative solution obtained in step 403b are iterated, and each gradient norm is compared with the tolerance threshold. Positions with gradient norms less than the tolerance threshold are selected as redundant degrees of freedom. Direct truncation and compression are performed on the tensor components corresponding to these redundant degrees of freedom, i.e., the tensor elements corresponding to these positions are deleted, eliminating redundant and invalid dimensional information. Simultaneously, the tensor components corresponding to valid degrees of freedom with gradient norms greater than or equal to the tolerance threshold are retained, resulting in the compressed dominant factor matrix and the compressed core tensor block. The dimensions of the two compressed matrices / tensors are smaller than before compression, and the core effective features are retained, achieving data structure simplification and optimization.

[0046] The compressed dominant factor matrix and the compressed core tensor block are dimensionally aligned and matched to ensure that their dimensional indexes correspond one-to-one, avoiding errors in subsequent calculations due to dimensional misalignment. Specifically, the dimensional indices of the compressed dominant factor matrix and the compressed core tensor block are extracted. The two sets of dimensional indices are compared, retaining those common to both and removing those existing only in a single matrix / tensor to ensure consistency. The column order of the compressed dominant factor matrix and the dimensional order of the compressed core tensor block are adjusted to ensure complete alignment of their corresponding dimensional indices. Based on the aligned dimensional indices, the two matrices / tensors are divided into blocks, with each feature component corresponding to a dimensional index assigned to an independent block, ensuring that each part matches one-to-one after the block division.

[0047] Based on the indices of each dimension of the compressed core tensor block, the eigenvectors of the corresponding dimensions in the compressed dominant factor matrix are sequentially expanded by outer product and mapped by element-wise multiplication according to their index order. The specific outer product expansion formula is as follows: ,in These are the eigenvectors corresponding to each dimension in the compressed dominant factor matrix, each corresponding one-to-one with the dimension index of the compressed core tensor block. The outer product operation is performed on vectors, and the outer product expansion yields a preliminary multidimensional tensor structure. Element-level product mapping is then performed on this preliminary multidimensional tensor structure and the compressed core tensor block. Each element in the preliminary multidimensional tensor structure is multiplied by the corresponding element in the compressed core tensor block. Through this element-level product mapping, the coupling and correlation information of the compressed core tensor block is integrated into the multidimensional tensor structure, ultimately resulting in a multi-level, interconnected multidimensional block-structure tensor. Dimension rearrangement and axial concatenation operations are then performed on the multidimensional block-structure tensor. First, dimension rearrangement is performed, adjusting the dimensional order of the multidimensional block-structure tensor according to a preset dimensional order (spatial dimension - temporal dimension - feature dimension). This ensures that the dimensional order is consistent with the dimensional baseline of the previously standardized dataset and the high-dimensional state transition matrix. Specifically, all dimensions of the multidimensional block-structure tensor are extracted, and the positions of each dimension are rearranged in the order of spatial dimension, temporal dimension, and feature dimension. After adjustment, the dimensional scale of the tensor remains unchanged; only the dimensional order changes. An ordered splicing operation is performed along the axial direction. The core of axial splicing is to integrate the features of each dimension of the rearranged multidimensional block structure tensor along the axial direction. Specifically, the splicing process is as follows: taking the spatial dimension as the reference axis, the feature components corresponding to the temporal dimension and the feature dimension are spliced ​​sequentially along the spatial axis. During splicing, it is ensured that the indices of each dimension feature correspond one-to-one, without feature misalignment or omission. Through axial splicing, multidimensional constraints and temporal evolution information are integrated, and finally a dynamically optimized feature tensor carrying multidimensional optimization constraint information is obtained. This tensor fully integrates spatiotemporal topology, temporal evolution, mode switching and environmental stress multidimensional constraint information, which can be directly used for intelligent parameter regulation, dynamic suppression of ice crystal nucleation risk and smooth switching of multiple modes in the subsequent stem cell cryopreservation process.

[0048] In this embodiment, various signal processing and computation methods are used to obtain the parameter update rate vector, iteration momentum coefficient, and convergence damping coefficient, respectively. These three are orthogonally concatenated to construct a high-dimensional state transition matrix, achieving a high-dimensional unified representation of multi-source regulation features. The matrix is ​​then subjected to CP tensor decomposition for dimensionality reduction, adaptively determining the environmental stress tolerance threshold and defining the convex hull constraint boundary. Through gradient projection iterative optimization, redundant degrees of freedom compression, and multi-step feature fusion, a dynamically optimized feature tensor is generated, providing high-dimensional feature support for intelligent regulation of stem cell cryopreservation and improving cryopreservation effect and intelligence level.

[0049] In a preferred embodiment of the present invention, step 5 includes: Step 500: Input the dynamically optimized feature tensor into a pre-defined deep reinforcement learning network to perform forward inference computation to obtain forward inference features, specifically including: The dynamically optimized feature tensor generated in step 404b is fully input into the pre-trained and converged deep reinforcement learning network for layer-by-layer forward inference computation, ensuring that the forward inference features can accurately carry the multi-dimensional constraints and regulatory information of the entire stem cell cryopreservation process. The construction and training process of this deep reinforcement learning network is as follows: In terms of construction, the network adopts a hierarchical structure of "input layer - feature mapping layer - pooling abstraction layer - policy evaluation layer - output layer." The input layer dimension is consistent with the dynamic optimized feature tensor dimension, used to receive high-dimensional feature data; the feature mapping layer has 3 layers, each using the ReLU activation function to achieve non-linear feature fusion; the pooling abstraction layer has 2 layers, using a 3×3 window and a stride of 2 max pooling algorithm to filter redundant features; the policy evaluation layer is a fully connected layer used to complete the evaluation of regulatory behavior value and feature weight allocation; the output layer dimension matches the forward inference feature dimension, used to output the inference results. All layers are connected using a fully connected method to ensure lossless feature transmission. For training, a multi-condition sample of stem cell cryopreservation was selected as the training dataset, covering different environmental stresses, different cryopreservation periods, and different mode switching scenarios. All sample data underwent prior standardization to ensure consistency with the dynamic optimization feature tensor format. The training objective was to minimize the deviation between the inferred features and the actual cryopreservation parameters. Gradient descent was used to iteratively update the network weights and bias parameters, with an iteration step size of 0.001. The iteration termination condition was when the loss function value converged to a certain value. After each iteration, the network performance is verified using a validation set. Outliers are removed, and parameters are adjusted. This process is repeated until the network converges, thus solidifying the parameters. After network training converges, the dynamically optimized feature tensor undergoes dimensionality normalization preprocessing to ensure that its spatial, temporal, and feature dimensions perfectly match the network input layer ports. During normalization, the internal feature correlations and element values ​​of the tensor remain unchanged. The normalized tensor is then fed into the input layer, where standardized normalization calculations are performed on each dimension element. This involves subtracting the mean of all elements in that dimension from the value of each element and then dividing by the standard deviation of that dimension, uniformly mapping the values ​​to the [0,1] interval to eliminate dimensional differences. After being processed by the input layer, the data flows through three feature mapping layers and two pooling abstraction layers to achieve deep fusion and core condensation of high-dimensional features. It is then sent to the policy evaluation layer for weighted fusion, which is calculated as: Weighted fusion result = weight matrix × input feature + bias term (the weight matrix is ​​a 32×64-dimensional matrix solidified after training convergence, and the bias term is a 32-dimensional vector solidified after training, with all elements in the range [0.01, 0.05]). Finally, the output layer outputs forward inference features with regular dimensions and complete structure. These features fully carry the spatiotemporal evolution, environmental stress constraints, and multi-mode regulatory correlation laws of the entire stem cell cryopreservation process.

[0050] Step 501: Within the convex hull constraint boundary, perform cross-entropy loss function minimization optimization on the forward inference features to obtain the target recommendation parameter set; perform globally unique batch identifier and timestamp sequence generation on the target recommendation parameter set to obtain the batch identifier and timestamp sequence, specifically including: Based on the effective feature optimization region defined by the convex hull constraint boundary in step 402b, the forward inference features output in step 500 are optimized by minimizing the cross-entropy loss function to ensure that the optimization result meets the environmental stress constraints and cell cryopreservation tolerance limits, while also conforming to the requirements of actual cryopreservation conditions. First, a cross-entropy loss function is constructed, using the probability difference between the predicted parameter distribution obtained from the forward inference feature mapping and the standard parameter distribution for actual stem cell cryopreservation as the optimization objective. By quantifying the probability difference between the two distributions, a precise measurement of the fit between the predicted parameters and the standard parameters is achieved. During the optimization process, the forward inference features are substituted into the cross-entropy loss function to calculate the initial loss value. The gradient descent method is used to iteratively adjust the parameter values, with an iteration step size set to 0.001 (reasonably set based on the accuracy of stem cell cryopreservation parameters and optimization efficiency). The loss value is recalculated after each iteration. The specific iterative update formula is as follows: = - × L( ),in For the first The parameter values ​​for the next iteration. The iteration step size, L( ) is the first The gradient of the loss function is calculated in each iteration, and the iteration continues until the cross-entropy loss function value converges to a preset minimum threshold. The parameter set obtained at this point is the target recommended parameter set that satisfies the convex hull constraint boundary range and adapts to the stress tolerance limit of the cell cryopreservation environment. This parameter set includes the optimal values ​​of all control dimensions such as temperature control rate, iteration momentum, and mode switching damping, and all parameter values ​​fall within the convex hull constraint boundary defined in step 402b. After obtaining the target recommended parameter set, the entire batch of frozen storage control data bound to this parameter set is processed to generate a globally unique batch identifier and timestamp sequence. The timestamp sequence is generated according to the data generation time sequence, based on the year, month, day, hour, minute, second, and millisecond time scale of the data collection time. The generated timestamp sequence is a continuously increasing sequence to ensure the continuity and uniqueness of the time sequence. The globally unique batch identifier is generated using the UUID algorithm based on the batch data dimension scale and feature dimension distribution characteristics. The specific generation process is as follows: extract the number of dimensions of the target recommended parameter set, the mean of each dimension element, and the variance of the feature distribution. The above parameters are hashed to obtain a 32-bit hash value, and then encoded to generate a globally unique and conflict-free batch identifier. This ensures that each set of target recommended parameter sets corresponds to a unique batch identifier and a continuously arranged timestamp sequence, realizing the time sequence traceability of the parameter set and the binding of the unique batch identifier.

[0051] Step 502: Perform hash encoding on the batch identifier and timestamp sequence to obtain the spatiotemporal index key; perform key-value mapping encapsulation on the spatiotemporal index key and the target recommendation parameter set to obtain the bound index data stream, specifically including: The batch identifier and timestamp sequence generated in step 501 are subjected to hash encoding. The SHA-256 hash algorithm is used to convert the combined character and numeric batch identifier time sequence into a fixed-length hash feature string, ensuring the uniqueness and irreversibility of the spatiotemporal index key. The specific encoding calculation process is as follows: the batch identifier (character type) and the timestamp sequence (numeric type) are concatenated in sequence to obtain a combined string; the combined string is encoded using UTF-8 to convert it into a binary byte stream; the binary byte stream is substituted into the SHA-256 hash function for calculation, specifically by grouping, padding, and compressing the byte stream, finally outputting a 256-bit (32-byte) hash value. This hash value is converted into a hexadecimal string, which is the spatiotemporal index key. This index key is unique, with different batches of target recommendation parameter sets corresponding to different spatiotemporal index keys, effectively avoiding index key duplication and conflict. After the spatiotemporal index key is generated, the spatiotemporal index key is used as the index field, and the target recommendation parameter set is used as the numeric field. Key-value mapping encapsulation processing is carried out to establish a one-to-one binding relationship between the spatiotemporal index key and the target recommendation parameter set. During the encapsulation process, in accordance with the distributed data storage specification, the spatiotemporal index key is set to a string type (fixed length of 64 bits), and each control parameter in the target recommendation parameter set is converted to floating-point data (precision retained to 6 decimal places). The data type and field format are unified. Then, the structure is encapsulated according to the key-value pair of the index key and parameter set to generate a bound index data stream with standard format, regular structure, and complete association binding. This data stream contains key information such as the spatiotemporal index key, the target recommendation parameter set, and data format identifier.

[0052] Step 503: Perform sharding mapping on the bound index data stream using the consistent hashing routing protocol, routing the sharded data stream to the corresponding storage nodes in the distributed database cluster to obtain the sharding mapping result; perform multi-replica synchronization verification and persistent storage on the sharding mapping result to complete the construction of the stem cell bank, specifically including: A consistent hashing routing protocol is introduced to perform sharding mapping on the bound index data stream, first constructing a value range of [0,2]. 32 The hash ring of [-1] uses the SHA-1 hash algorithm to calculate the IP address and port number of each storage node. This involves concatenating the IP address and port number into a binary byte stream, substituting it into the SHA-1 hash function to obtain a 160-bit hash value, and then converting this hash value into a decimal integer and dividing it by 2. 32The modulo operation yields the hash value for each node. Starting from the hash value of each node, the interval extending clockwise to the next node's hash value constitutes the fixed hash interval for each node: Node 1 hash interval [0, 858993459], Node 2 hash interval [858993460, 1717986919], Node 3 hash interval [1717986920, 2576980379], Node 4 hash interval [2576980380, 3435973839], and Node 5 hash interval [3435973840, 4294967295]. These intervals are continuous, non-overlapping, and completely cover the entire hash ring. The time-space index key (hexadecimal string) is then converted to a decimal integer, and the result is calculated using the modulo operation. 32 The modulo operation yields the corresponding hash value. Based on the principle of matching the nearest hash ring, the data stream is routed to the corresponding storage node that covers the hash value within the hash range. All data streams are traversed to complete the full sharding mapping, ensuring load balancing across storage nodes.

[0053] After sharding is completed, multi-replica synchronization verification and persistent storage processing are performed on the sharded data distributed to each storage node. The preset number of replicas is 3 (1 primary node replica and 2 standby node replicas). Replicas of the corresponding sharded data are generated on different standby nodes within the same cluster. During the synchronization verification process, the data of the primary node and each replica node are compared bit by bit. The specific verification content includes the integrity of the spatiotemporal index key, the consistency of the target recommendation parameter set values, and the standardization of the data format. If the absolute value of the difference between the corresponding field of the primary node data and the corresponding field of the replica node is less than 1, the verification is performed. If the data passes the verification, it is considered valid; otherwise, it is considered a verification failure. In the event of a verification failure, the data is resynchronized and verified again until all replicas are completely consistent with the master node. After successful verification, the master node data and replica node data are synchronously written to disk for persistent storage. The specific storage path is set according to the hierarchical structure of cluster node ID / spatiotemporal index key / data type to ensure orderly data storage and easy retrieval. After all sharded data has completed routing, synchronization verification, and persistent storage, the collection, archiving, and distributed deployment of multi-dimensional control features, time-series indexes, and parameter data are completed, ultimately achieving the complete construction of a structured, time-series, and distributed data system for the stem cell bank.

[0054] This embodiment, through forward inference computation of a deep reinforcement learning network, can fully mine the multidimensional correlation information in the dynamically optimized feature tensor, transforming high-dimensional features into forward inference features that fit the actual cryopreservation conditions. This avoids inference bias caused by redundancy of high-dimensional features and ensures that the recommended parameters meet the core requirements of stem cell cryopreservation. Optimizing the cross-entropy loss function within the convex hull constraint boundary ensures that the target recommended parameter set meets the environmental stress tolerance threshold and the cell cryopreservation tolerance limit. Furthermore, through precise loss quantification and iterative calculation, the recommended parameters closely match the actual cryopreservation standard parameters, effectively improving the temperature control accuracy and mode switching smoothness of stem cell cryopreservation, and reducing the risk of ice crystal nucleation. Through hash encoding, key-value mapping encapsulation, and consistent hash routing distribution, the unique identifier, ordered association, and load-balanced storage of the target recommended parameter set are achieved. The uniqueness of the spatiotemporal index key ensures efficient data retrieval, and the consistent hash routing protocol avoids the problem of excessive load on a single node, improving the stability and scalability of data storage. Multi-copy synchronous verification and persistent storage ensure the integrity, consistency and reliability of stem cell bank data, and prevent data loss or tampering. The constructed structured, time-series, and distributed stem cell bank not only realizes the standardized archiving of multi-source regulatory data.

[0055] like Figure 2 As shown, embodiments of the present invention also provide a system for constructing a stem cell bank, comprising: The fusion module is used to perform spatiotemporal alignment and multimodal feature fusion on the structured numerical sequences and unstructured image features in the real-time acquired multi-source stem cell data, and generate a spatial distribution field of ice crystal nucleation risk during the fusion process, thus obtaining a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. The processing module is used to determine the spatial mapping anchor point based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and adaptively determine the radial truncation threshold based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field; and perform neighborhood feature density aggregation and spatial vectorization reconstruction processing on the local influence field to obtain the spatiotemporal topology reconstruction dataset. The mapping module is used to perform ontological semantic mapping and cleaning on the spatiotemporal topology reconstruction dataset, construct entity topology, and obtain a standardized dataset. The solution module is used to discretize the rate of change characteristics of the cooling and reheating stages in the standardized dataset into parameter update rate vectors, quantize the cumulative trend of historical control trajectories into iterative momentum coefficients, and convert the response delay characteristics of state switching into convergence damping coefficients to construct a high-dimensional state transition matrix; tensor dimensionality reduction and boundary solution processing are performed on the high-dimensional state transition matrix to obtain a dynamic optimization feature tensor; The storage module is used to input the dynamically optimized feature tensor into a preset deep reinforcement learning network for processing to obtain the target recommendation parameter set. The target recommendation parameter set is then bound to the spatiotemporal index, and the bound index data stream is written and persistently stored through a distributed database protocol to complete the construction of the stem cell library.

[0056] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0057] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0058] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a stem cell bank, characterized in that, The method includes: Step 1: Perform spatiotemporal alignment and multimodal feature fusion on the structured numerical sequences and unstructured image features in the real-time acquired stem cell multi-source data, and generate a spatial distribution field of ice crystal nucleation risk during the fusion process to obtain a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. Step 2: Determine the spatial mapping anchor point based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and adaptively determine the radial truncation threshold based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field; perform neighborhood feature density aggregation and spatial vectorization reconstruction processing on the local influence field to obtain the spatiotemporal topology reconstruction dataset. Step 3: Perform ontology semantic mapping and cleaning on the spatiotemporal topology reconstruction dataset to construct entity topology and obtain a standardized dataset; Step 4: Discretize the rate of change features of the cooling and reheating stages in the standardized dataset into parameter update rate vectors, quantize the cumulative trend of historical control trajectories into iterative momentum coefficients, and convert the response delay features of state switching into convergence damping coefficients to construct a high-dimensional state transition matrix; perform tensor dimensionality reduction and boundary solution processing on the high-dimensional state transition matrix to obtain the dynamic optimization feature tensor. Step 5: Input the dynamically optimized feature tensor into the preset deep reinforcement learning network for processing to obtain the target recommendation parameter set. Bind the target recommendation parameter set with the spatiotemporal index, and write and persistently store the bound index data stream through a distributed database protocol to complete the construction of the stem cell bank.

2. The method for constructing a stem cell bank according to claim 1, characterized in that, Step 1 includes: Real-time collected cell source data, pre-freezing quality inspection data, dispensing parameter data, cooling curve data, and post-resuscitation activity data are used as multi-source stem cell data. Hardware clock timestamp labels and device physical space coordinate labels are extracted from multi-source stem cell data. Linear interpolation resampling alignment is performed on structured numerical sequences based on hardware clock timestamp labels. Perspective mesh spatial mapping is performed on unstructured image features based on device physical space coordinate labels to obtain spatiotemporal synchronized data blocks. Calculate the Pearson correlation coefficient between the numerical channels and image texture channels in the spatiotemporal synchronization data block, and construct the channel feature weight matrix based on the Pearson correlation coefficient; perform element-wise weighted summation operation on the channel feature weight matrix and the spatiotemporal synchronization data block to obtain the preliminary fused dataset; The ratio of the local cooling rate to the standard nucleation rate at each data point in the preliminary fusion dataset is calculated. The spatial distribution field of ice crystal nucleation risk is obtained by logarithmic transformation and spatial interpolation. The spatial distribution field of ice crystal nucleation risk is used as an additional channel and merged with the preliminary fusion dataset to obtain a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk.

3. The method for constructing a stem cell bank according to claim 2, characterized in that, Spatial mapping anchor points are determined based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and radial truncation thresholds are adaptively determined based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field, including: Spatial gradient calculation is performed on the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and the coordinate point with the largest gradient magnitude is determined as the spatial mapping anchor point; Centered on the spatial mapping anchor point, the local variance of the spatial distribution field of ice crystal nucleation risk is calculated in a spherical region with the spatial mapping anchor point as the center and a fixed initial step size as the radius. The ratio of the local variance to the global variance is combined with the preset reference radius to obtain the radial cutoff threshold. A spherical retrieval boundary is defined with the spatial mapping anchor point as the center and the radial truncation threshold as the radius. The set of feature coordinates falling within the boundary is marked as the local influence field.

4. The method for constructing a stem cell bank according to claim 3, characterized in that, Neighborhood feature density aggregation and spatial vectorization reconstruction are performed on the local influence field to obtain a spatiotemporal topology reconstruction dataset, including: Calculate the Euclidean distance between each feature coordinate in the local influence field and the spatial mapping anchor point, and construct a Gaussian decay weight matrix based on the Euclidean distance; Based on the Gaussian decay weight matrix, kernel density estimation is performed on the feature coordinates in the local influence field to obtain the local feature density distribution map; the covariance matrix is ​​constructed on the local feature density distribution map and the eigenvalues ​​and corresponding orthogonal eigenvectors are solved; the eigenvalues ​​are arranged in descending order and a number of principal component eigenvectors with a cumulative contribution rate reaching a set proportion are extracted. A linear projection transformation is performed on the local feature density distribution map along the directions of several principal component feature vectors to obtain the projected dimension-reduced coordinate sequence. An orthogonal feature basis vector matrix is ​​reconstructed based on the dimension-reduced coordinate sequence. The orthogonal feature basis vector matrix is ​​then used to perform a spatial basis transformation on the feature coordinate set in the local influence field to obtain a spatiotemporal topology reconstruction dataset.

5. The method for constructing a stem cell bank according to claim 4, characterized in that, Step 3 includes: Based on the cell batch identifier mapping relationship, storage location coordinate mapping relationship and quality status evaluation mapping relationship defined in the pre-set ontology semantic mapping rule base, field alignment and attribute normalization processing are performed on the spatiotemporal topology reconstruction dataset to obtain a topology map with semantic labels. Perform sliding window temporal difference calculation on the topological map with semantic labels to obtain the feature difference value sequence of continuous time steps; calculate the moving average and standard deviation of the sliding window based on the feature difference value sequence to construct a dynamic baseline threshold interval; Feature coordinates falling outside the dynamic baseline threshold range are filtered and the corresponding feature coordinates are marked as anomalous jump coordinates; linear temporal interpolation replacement based on adjacent valid data points and Gaussian noise smoothing filtering are performed on the anomalous jump coordinates to obtain the filtered topological map with semantic labels; The filtered topological map with semantic labels is serialized and encoded to obtain a standardized dataset.

6. The method for constructing a stem cell bank according to claim 5, characterized in that, The rate of change characteristics of the cooling and reheating stages in the standardized dataset are discretized into parameter update rate vectors, the cumulative trend of historical control trajectories is quantified into iterative momentum coefficients, and the response delay characteristics of state switching are converted into convergence damping coefficients to construct a high-dimensional state transition matrix, including: Temperature control time series curves for the cooling and reheating stages are extracted from the standardized dataset to obtain a temperature control time series curve sequence. The first-order difference sequence of the temperature control time series curve sequence within a continuous sampling period is calculated. Fast Fourier Transform and bandpass filtering are performed on the first-order difference sequence to extract the dominant frequency component with the highest energy proportion and discretize it to obtain the parameter update rate vector. The time series evolution trajectory of historical control parameters is extracted from the standardized dataset. Numerical integration accumulation is performed on the time series evolution trajectory. The accumulation result is subjected to extreme value normalization mapping to obtain the iterative momentum coefficient. Perform storage operation mode state recognition processing on the standardized dataset, capture the transition timestamps of mode switching between program cooling storage mode, liquid nitrogen cryogenic preservation mode and program rewarming recovery mode, and obtain the mode switching transition timestamp sequence. The absolute time series deviation is calculated between the mode switching transition timestamp sequence and the theoretical response timestamp. The absolute time series deviation is then input into the exponential decay mapping function to obtain the convergence damping coefficient. The parameter update rate vector, iterative momentum coefficient, and convergence damping coefficient are orthogonally concatenated according to the feature dimension to construct a high-dimensional state transition matrix.

7. The method for constructing a stem cell bank according to claim 6, characterized in that, Tensor dimensionality reduction and boundary value solving are performed on the high-dimensional state transition matrix to obtain a dynamically optimized feature tensor, including: Perform CP tensor decomposition on the high-dimensional state transition matrix to extract the dominant factor matrix and core tensor block; The core tensor block is subjected to mode expansion and singular value decomposition to obtain a singular value sequence. The singular value sequence is sorted in descending order and accumulated item by item to construct a singular value cumulative energy distribution sequence. The singular value cumulative energy distribution sequence is subjected to second-order difference calculation to locate the distribution slope abrupt point. The cumulative energy ratio corresponding to the distribution slope abrupt point is determined as the environmental stress tolerance threshold. The column and row dimension parameters of the dominant factor matrix are extracted to construct a multidimensional feature optimization space. Based on the environmental stress tolerance threshold, the convex hull constraint boundary is delineated in the multidimensional feature optimization space to obtain the convex hull constraint boundary. The core tensor block is projected onto the convex hull constraint boundary to obtain the projected core tensor block. Gradient projection iterative solution is then performed on the projected core tensor block within the convex hull constraint boundary to obtain the iterative solution result. Redundant degrees of freedom with gradient norms less than the tolerance threshold in the iterative solution are truncated and compressed to obtain a compressed dominant factor matrix and a compressed core tensor block. The compressed dominant factor matrix and the compressed core tensor block are then dimensionally aligned and block-matched. Based on the indexes of each dimension of the compressed core tensor block, the eigenvectors of the corresponding dimensions in the compressed dominant factor matrix are sequentially expanded by outer product and element-wise multiplicative mapping according to their index order to obtain a multidimensional block-structured tensor. The multidimensional block-structured tensor is then subjected to dimensional rearrangement and axial concatenation operations to obtain a dynamically optimized feature tensor carrying multidimensional optimization constraint information.

8. The method for constructing a stem cell bank according to claim 7, characterized in that, Step 5 includes: The dynamically optimized feature tensor is input into a pre-defined deep reinforcement learning network to perform forward inference computation to obtain forward inference features; Within the convex hull constraint boundary, the cross-entropy loss function is minimized to optimize the forward inference features, resulting in the target recommendation parameter set; the target recommendation parameter set is then processed to generate a globally unique batch identifier and timestamp sequence, resulting in the batch identifier and timestamp sequence. The batch identifier and timestamp sequence are hashed to obtain the spatiotemporal index key; the spatiotemporal index key and the target recommendation parameter set are encapsulated by key-value mapping to obtain the bound index data stream. The bound index data stream is processed by a consistent hashing routing protocol to perform sharding mapping, and the sharded data stream is routed to the corresponding storage node of the distributed database cluster to obtain the sharding mapping result. The sharding mapping result is then subjected to multi-replica synchronization verification and persistent storage to complete the construction of the stem cell library.

9. A system for constructing a stem cell bank, the system implementing the method as described in any one of claims 1 to 8, characterized in that, include: The fusion module is used to perform spatiotemporal alignment and multimodal feature fusion on the structured numerical sequences and unstructured image features in the real-time acquired multi-source stem cell data, and generate a spatial distribution field of ice crystal nucleation risk during the fusion process, thus obtaining a multi-source fusion dataset carrying the spatial distribution field of ice crystal nucleation risk. The processing module is used to determine the spatial mapping anchor point based on the gradient of the spatial distribution field of ice crystal nucleation risk in the multi-source fusion dataset, and adaptively determine the radial truncation threshold based on the local variance of the spatial distribution field of ice crystal nucleation risk to define the local influence field; and perform neighborhood feature density aggregation and spatial vectorization reconstruction processing on the local influence field to obtain the spatiotemporal topology reconstruction dataset. The mapping module is used to perform ontological semantic mapping and cleaning on the spatiotemporal topology reconstruction dataset, construct entity topology, and obtain a standardized dataset. The solution module is used to discretize the rate of change characteristics of the cooling and reheating stages in the standardized dataset into a parameter update rate vector, quantize the cumulative trend of the historical control trajectory into an iterative momentum coefficient, and convert the response delay characteristics of state switching into a convergence damping coefficient, so as to construct a high-dimensional state transition matrix. Tensor dimensionality reduction and boundary value solving are performed on the high-dimensional state transition matrix to obtain the dynamically optimized feature tensor. The storage module is used to input the dynamically optimized feature tensor into a preset deep reinforcement learning network for processing to obtain the target recommendation parameter set. The target recommendation parameter set is then bound to the spatiotemporal index, and the bound index data stream is written and persistently stored through a distributed database protocol to complete the construction of the stem cell library.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images