Real-time monitoring method and system for stem cell storage environment

Abstract

The present invention relates to the field of stem cell storage technology, and discloses a real-time monitoring method and system for a stem cell storage environment. The method includes: collecting multiple environmental parameter data stored in stem cells and performing preprocessing to obtain a preprocessed data set; performing deep compressed sensing and generative model processing to obtain an interpolated and enhanced data set; performing convolutional neural network processing and feature attention mechanism calculation to obtain a weighted feature representation; performing bidirectional long short-term memory network processing and time step attention mechanism calculation on the weighted feature representation to obtain a context vector; performing multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain a multi-parameter prediction result and an environmental status evaluation index; performing multi-level threshold judgment and adaptive feedback processing on the multi-parameter prediction result and the environmental status evaluation index to obtain real-time monitoring warning information. The present invention realizes the intelligent real-time monitoring of the stem cell storage environment.

Description

Technical Field

[0001] The present invention relates to the technical field of stem cell storage, and in particular to a real-time monitoring method and system for a stem cell storage environment. Background Art

[0002] Stem cell storage plays a key role in biomedical research and clinical applications. The precise control of the storage environment directly affects the activity and quality of stem cells. Traditional methods for monitoring the storage environment of stem cells often have problems such as insufficient real-time performance, limited data processing capabilities, and insufficient early warning mechanisms. With the rapid development of the Internet of Things and artificial intelligence technologies, new technical means and solutions have been provided for real-time monitoring of the stem cell storage environment.

[0003] However, existing monitoring systems still face many challenges. The stem cell storage environment involves multiple key parameters, such as temperature, humidity, gas concentration, etc. There are complex interactions between these parameters, and traditional single-parameter monitoring is difficult to fully grasp the environmental status. Secondly, environmental data often contain noise, missing values, and outliers, which affect the accuracy and reliability of monitoring. In addition, because stem cells are extremely sensitive to environmental changes, a more accurate and forward-looking early warning mechanism is needed to detect potential risks in a timely manner and take corresponding measures. Summary of the invention

[0004] The present invention provides a real-time monitoring method and system for a stem cell storage environment, which are used to realize intelligent real-time monitoring of a stem cell storage environment.

[0005] In a first aspect, the present invention provides a method for real-time monitoring of a stem cell storage environment, the method comprising:

[0006] Collecting a variety of environmental parameter data stored in the stem cells, and preprocessing the multiple environmental parameter data to obtain a preprocessed data set;

[0007] Performing deep compressed sensing and generative model processing on the preprocessed data set to obtain an interpolated and enhanced data set;

[0008] Performing convolutional neural network processing and feature attention mechanism calculation on the interpolated and enhanced data set to obtain a weighted feature representation;

[0009] Performing bidirectional long short-term memory network processing and time-step attention mechanism calculation on the weighted feature representation to obtain a context vector;

[0010] Performing multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain multi-parameter prediction results and environmental state evaluation indicators;

[0011] Multi-level threshold judgment and adaptive feedback processing are performed on the multi-parameter prediction results and the environmental status assessment indicators to obtain real-time monitoring and early warning information.

[0012] In a second aspect, the present invention provides a real-time monitoring system for a stem cell storage environment, the real-time monitoring system for a stem cell storage environment comprising:

[0013] A collection module, used for collecting a variety of environmental parameter data stored in the stem cells, and preprocessing the multiple environmental parameter data to obtain a preprocessed data set;

[0014] An interpolation module, used for performing deep compressed sensing and generative model processing on the preprocessed data set to obtain an interpolated and enhanced data set;

[0015] A weighting module, used for performing convolutional neural network processing and feature attention mechanism calculation on the interpolated and enhanced data set to obtain a weighted feature representation;

[0016] A computing module, used for performing bidirectional long short-term memory network processing and time-step attention mechanism calculation on the weighted feature representation to obtain a context vector;

[0017] A prediction module, used to perform multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain multi-parameter prediction results and environmental state evaluation indicators;

[0018] The early warning module is used to perform multi-level threshold judgment and adaptive feedback processing on the multi-parameter prediction results and the environmental status assessment indicators to obtain real-time monitoring early warning information.

[0019] The third aspect of the present invention provides a real-time monitoring device for a stem cell storage environment, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the real-time monitoring device for the stem cell storage environment executes the above-mentioned real-time monitoring method for the stem cell storage environment.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned method for real-time monitoring of a stem cell storage environment.

[0021] In the technical solution provided by the present invention, through deep compressed sensing and generative model processing, a comprehensive analysis of multiple environmental parameters is realized, which can more comprehensively reflect the overall state of the stem cell storage environment and improve the comprehensiveness and accuracy of monitoring. Advanced data preprocessing and interpolation technology is adopted to effectively solve the noise, missing and outlier problems in environmental data, significantly improve the data quality, and combine convolutional neural network and feature attention mechanism to adaptively extract and strengthen key features, improve the sensitivity and generalization ability of the model to environmental changes. Through the bidirectional long short-term memory network and time step attention mechanism, the long-term and short-term temporal dependencies of environmental parameters are effectively captured, and the prediction ability and stability of the model are enhanced. By adopting multi-task parallel processing and comprehensive loss function calculation, the coordinated optimization of environmental parameter prediction and state evaluation is realized, and the overall performance and efficiency of the model are improved. By introducing multi-level threshold judgment and adaptive feedback processing, a flexible and accurate early warning system is constructed, which can timely detect potential risks and give reasonable suggestions. By dynamically adjusting the early warning threshold and weight coefficient, the adaptive optimization of the system is realized, so that the monitoring method can continuously adapt to environmental changes and new monitoring needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0023] Figure 1 A schematic flow chart of a method for real-time monitoring of a stem cell storage environment provided in an embodiment of the present application;

[0024] Figure 2 This is a schematic block diagram of the structure of the real-time monitoring system for the stem cell storage environment provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change based on actual conditions.

[0027] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0028] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0029] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0030] See also Figure 1 , Figure 1 A flow chart of a real-time monitoring method for a stem cell storage environment provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the real-time monitoring method of the stem cell storage environment provided in the embodiment of the present application includes steps S100 to S600.

[0031] Step S100, collecting various environmental parameter data stored in stem cells, and preprocessing the various environmental parameter data to obtain a preprocessed data set;

[0032] It is understandable that the execution subject of the present invention may be a real-time monitoring system for the stem cell storage environment, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject.

[0033] Specifically, the data collection frequency of various sensors in the stem cell storage device is set to obtain a preset collection time interval to ensure the efficiency and consistency of data collection. According to the preset collection time interval, the temperature, humidity, CO2 concentration, O2 concentration and pressure sensors are collected to obtain various environmental parameter data to reflect the specific state of the stem cell storage environment. The various environmental parameter data are divided into time windows to obtain data subsets, and the continuous data are divided into block data units that are easy to process. The same sensor data in the data subset is weighted averaged to obtain aggregated environmental parameter data. The weighted average calculation can reduce data noise and improve the reliability and representativeness of the data. The aggregated environmental parameter data is detected for outliers, the abnormal points in the data are identified, and the abnormal data point marks are obtained. According to the abnormal data point marks, the aggregated environmental parameter data is eliminated for outliers, and those abnormal data that may be caused by sensor failure or other reasons are eliminated to obtain a cleaned data set. The cleaned data set is normalized so that the data has the same dimension and scale, which is convenient for subsequent model training and analysis, and standardized environmental parameter data is obtained. The standardized environmental parameter data is reshaped in terms of time step and feature dimension, and the original time series data is converted into a format suitable for model input to obtain a preliminarily constructed data frame. The preliminarily constructed data frame is subjected to a sliding window operation to obtain a data sample set. Through the sliding window operation, multiple overlapping data samples are generated, which helps the model capture the dynamic change characteristics in the time series. The data sample set is quality assessed, including checking the integrity and consistency of the data and its conformity to the expected pattern, to obtain a preprocessed data set, to ensure the quality and representativeness of the data sample, and to ensure the accuracy of subsequent model training and prediction.

[0034] Step S200, performing deep compressed sensing and generative model processing on the preprocessed data set to obtain an interpolated and enhanced data set;

[0035] Specifically, the preprocessed data set is subjected to random matrix projection to obtain a low-dimensional compressed representation. Random matrix projection is an effective dimensionality reduction technique that projects high-dimensional data into a low-dimensional space through a randomly generated matrix, reducing the dimension of the data and retaining the main information of the data. The low-dimensional compressed representation is sparsely encoded to obtain a sparse coefficient vector. Sparse coding is a representation learning method that captures the intrinsic structure of the data by representing the data as a sparse linear combination to improve the effectiveness of the data representation. According to the sparse coefficient vector, the preprocessed data set is reconstructed to obtain a preliminary interpolation result to make up for the missing parts of the data and make the data more complete. The preliminary interpolation result is processed by a variational autoencoder to obtain a latent space representation. A variational autoencoder is a generative model that can capture the distribution characteristics of data by encoding data into a continuous latent space. Random sampling is performed based on the latent space representation to obtain reparameterized samples. The reparameterization technique allows sampling in the latent space and passing these samples to the decoder to obtain reconstructed data. The reconstruction error between the reconstructed data and the preprocessed data set and the KL divergence of the variational autoencoder are calculated to obtain the variational lower bound. The reconstruction error measures the difference between the generated data and the real data, while the KL divergence measures the difference between the latent space distribution and the prior distribution. By maximizing the variational lower bound and optimizing the parameters of the generative model, the model can better generate high-quality data. The optimized parameters of the generative model are used to conditionally generate the missing data to obtain a complete dataset after interpolation, and the dataset is data augmented to obtain an interpolated and augmented dataset.

[0036] Step S300, performing convolutional neural network processing and feature attention mechanism calculation on the interpolated and enhanced data set to obtain a weighted feature representation;

[0037] Specifically, multi-scale convolution processing is performed on the interpolated and enhanced data sets, and convolution kernels of different sizes are used, namely 3x3, 5x5 and 7x7, to obtain the initial feature map. The multi-scale convolution processing method can capture feature information of different scales, making the feature representation richer and more diverse. The initial feature map is batch normalized to obtain a standardized feature map. Batch normalization helps to stabilize the training process and improve the convergence speed and effect of the model. The standardized feature map is processed by the ReLU activation function to obtain a nonlinear feature map. The ReLU activation function can introduce nonlinear transformations, making the model more expressive. The nonlinear feature map is subjected to the maximum pooling operation to obtain a downsampled feature map. The maximum pooling operation can reduce the size of the feature map while retaining important features and improving computational efficiency. According to the downsampled feature map, a residual connection is constructed to obtain a fused feature map. The introduction of the residual connection can effectively prevent the gradient vanishing and gradient exploding problems, and help the model learn more complex feature representations. The fused feature map is subjected to global average pooling to obtain a channel descriptor. Global average pooling can compress the feature map into a vector of channel dimension and retain important information of each channel. The channel descriptor is processed by two layers of fully connected networks to obtain attention weights. The fully connected network plays a role in feature screening. By learning the weight relationship between channels, important feature channels are highlighted. According to the obtained attention weights, the fused feature map is weighted to obtain a weighted feature map, which strengthens important features and suppresses irrelevant or unimportant features. Multi-scale feature fusion is performed on the weighted feature map to obtain a multi-scale feature representation, which integrates feature information of different scales. The multi-scale feature representation is processed by 1x1 convolution to obtain a weighted feature representation. 1x1 convolution helps to recombine channel information without changing the size of the feature map, resulting in a more compact and efficient feature representation.

[0038] Step S400, performing bidirectional long short-term memory network processing and time-step attention mechanism calculation on the weighted feature representation to obtain a context vector;

[0039] Specifically, the weighted feature representation is rearranged in time series to obtain a time series feature sequence. The features are arranged in time order so that the subsequent time series model can effectively capture the time series dynamic change characteristics of the data. The time series feature sequence is processed by forward LSTM to obtain a forward hidden state sequence. LSTM (Long Short-Term Memory Network) is a deep learning model that can capture long-term dependencies and is used to process time series data. At the same time, the time series feature sequence is processed by reverse LSTM to obtain a reverse hidden state sequence, so that the model can obtain information from past and future time points, which helps to improve the comprehensiveness and accuracy of feature extraction. The forward hidden state sequence and the reverse hidden state sequence are concatenated to obtain a bidirectional hidden state sequence. The bidirectional hidden state sequence is processed by a fully connected layer to generate an attention score sequence. The attention score sequence is normalized by softmax to obtain an attention weight sequence. The softmax function converts the original score into a probability distribution so that the sum of the weights is 1, highlighting important features while suppressing minor features. According to the attention weight sequence, the bidirectional hidden state sequence is weighted and summed to obtain the first vector. The first vector and the hidden state of the last time step are concatenated to obtain the second vector. The information of the current time step and the overall timing information are integrated to construct a more complete feature representation. The second vector is processed nonlinearly to obtain a nonlinear feature representation. Nonlinear processing usually uses activation functions such as ReLU or tanh to introduce nonlinear features and enhance the expressiveness of the model. Dropout regularization is performed on the nonlinear feature representation to obtain a context vector. Dropout is an effective regularization technique that prevents overfitting of the model and improves the generalization ability of the model by randomly discarding a portion of neurons.

[0040] Step S500, performing multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain multi-parameter prediction results and environmental state evaluation indicators;

[0041] Specifically, the context vector is processed by the task fully connected layer to generate multiple task-related feature vectors. These feature vectors represent different tasks, such as the prediction of environmental parameters such as temperature, humidity, CO2 concentration, and O2 concentration. The shared context vector is separated into features suitable for different tasks, so that the model can perform specialized processing on each task. The task-related feature vectors are processed by the output layer to obtain the predicted values ​​of each environmental parameter. Each output layer corresponds to a specific environmental parameter. In this way, multiple parameters are predicted at the same time. The task loss is calculated based on the predicted values ​​of the environmental parameters to obtain multiple single-task loss values. These loss values ​​reflect the prediction error of the model on each task, which is usually calculated by loss functions such as mean square error or cross entropy. Multiple task-related feature vectors are concatenated to form a comprehensive feature vector. The comprehensive feature vector is processed by a multi-layer perceptron to generate environmental state features. The multi-layer perceptron extracts complex high-dimensional features through a series of linear and nonlinear transformations to enhance the model's ability to represent environmental states. The environmental state features are classified by Softmax to obtain the probability distribution of the environmental state. The Softmax function converts the output into a probability distribution so that the sum of the probabilities of each environmental state is 1, which helps to determine which state the current environment is in. According to the probability distribution of the environmental state, the classification loss is calculated to obtain the environmental state evaluation loss. By comparing the actual environmental state with the predicted state, the performance of the model in environmental state classification is evaluated. The weighted sum of multiple single-task loss values ​​and environmental state evaluation losses is performed to obtain a comprehensive loss value. The comprehensive loss value represents the overall performance of the model on all tasks. By minimizing this value, the prediction performance of all tasks is optimized at the same time. The uncertainty between tasks is calculated based on the comprehensive loss value to obtain a dynamic weight coefficient. The dynamic weight coefficient measures the uncertainty of the task, and the weight of each task is adjusted according to its uncertainty size, so as to better balance the contribution of each task during the training process. The dynamic weight coefficient is used to update the comprehensive loss value, optimize the parameters of the model, and obtain more accurate multi-parameter prediction results and environmental state evaluation indicators.

[0042] Step S600: Perform multi-level threshold judgment and adaptive feedback processing on the multi-parameter prediction results and environmental status assessment indicators to obtain real-time monitoring and warning information.

[0043] Specifically, the multi-parameter prediction results are fitted with a normal distribution to obtain the probability density function of each parameter. The volatility and uncertainty of the parameters are characterized by the probability model, and then the confidence interval of the predicted value of each parameter is calculated to determine the fluctuation range of the parameter. The confidence interval reflects the possible variation range of the parameter under a certain probability level. The parameter fluctuation range is compared with the preset multi-level warning threshold to determine the preliminary warning level. The multi-level warning threshold is set according to the specific requirements and safety standards of stem cell storage. Each parameter has corresponding multiple thresholds to distinguish different levels of warning conditions such as normal state, attention state, and alert state. At the same time, the state transition probability is calculated according to the environmental state assessment index to construct the state transition matrix. The transition probability describes the possibility of the system transferring from one state to another and reflects the dynamic change trend of the environmental state. The state transition matrix is ​​subjected to Markov chain analysis to obtain the steady-state distribution of the system. The steady-state distribution represents the probability of the system being in each state under long-term operation. Combined with the current state, the probability distribution of the future state is predicted to obtain the state warning index. The state warning index is a prediction of the state that may occur in the future and is used to take preventive measures in advance. The preliminary warning level and the status warning index are weighted and integrated to obtain the comprehensive warning level. According to the comprehensive warning level, real-time monitoring warning information is generated. The real-time monitoring warning information is adaptively fed back to obtain adaptive feedback results. By analyzing the feedback results, the warning threshold and weight coefficient are dynamically adjusted to achieve adaptive optimization. The adaptive feedback mechanism can continuously optimize the parameter settings of the warning system according to the actual situation, and improve the accuracy and response speed of the warning.

[0044] In the embodiment of the present invention, through deep compressed sensing and generative model processing, a comprehensive analysis of multiple environmental parameters is achieved, which can more comprehensively reflect the overall state of the stem cell storage environment and improve the comprehensiveness and accuracy of monitoring. Advanced data preprocessing and interpolation technology is adopted to effectively solve the noise, missing and outlier problems in environmental data, significantly improve the data quality, and combine convolutional neural network and feature attention mechanism to adaptively extract and strengthen key features, improve the sensitivity and generalization ability of the model to environmental changes. Through the bidirectional long short-term memory network and time step attention mechanism, the long-term and short-term temporal dependencies of environmental parameters are effectively captured, and the prediction ability and stability of the model are enhanced. By adopting multi-task parallel processing and comprehensive loss function calculation, the coordinated optimization of environmental parameter prediction and state evaluation is achieved, and the overall performance and efficiency of the model are improved. By introducing multi-level threshold judgment and adaptive feedback processing, a flexible and accurate early warning system is constructed, which can timely detect potential risks and give reasonable suggestions. By dynamically adjusting the early warning threshold and weight coefficient, the adaptive optimization of the system is achieved, so that the monitoring method can continuously adapt to environmental changes and new monitoring needs.

[0045] In a specific embodiment, the process of executing step S100 may specifically include the following steps:

[0046] Setting the data collection frequency for various sensors in the stem cell storage device to obtain a preset collection time interval, and collecting data from temperature, humidity, CO2 concentration, O2 concentration and pressure sensors according to the preset collection time interval to obtain various environmental parameter data;

[0047] Divide the various environmental parameter data into time windows to obtain data subsets, and perform weighted average calculation on similar sensor data in the data subsets to obtain aggregated environmental parameter data;

[0048] Perform outlier detection on the aggregated environmental parameter data to obtain abnormal data point marks, and remove outliers from the aggregated environmental parameter data based on the abnormal data point marks to obtain a cleaned data set;

[0049] The cleaned data set is normalized to obtain standardized environmental parameter data, and the standardized environmental parameter data is reshaped in terms of time step and feature dimension to obtain a preliminarily constructed data frame;

[0050] A sliding window operation is performed on the initially constructed data frame to obtain a data sample set, and a preprocessed data set is obtained based on a quality assessment of the data sample set.

[0051] Specifically, determine the appropriate collection frequency for each sensor. Set a reasonable collection time interval that fully considers the changing characteristics of environmental parameters and the response speed of the equipment. For example, if the temperature changes rapidly and has a significant impact on stem cell storage, a shorter time interval can be selected (such as collecting once per second), while for slower changing parameters such as CO2 concentration and pressure, a longer time interval may be selected (such as collecting once per minute). Data from various sensors are collected according to the set collection time interval to obtain environmental parameter data of temperature, humidity, CO2 concentration, O2 concentration and pressure. The collected data of various environmental parameters are divided into time windows. Assume that a time window size of , then the data in each window forms a data subset. Time window division can divide the continuous data stream into smaller segments that are easier to process, thereby improving the efficiency and accuracy of data processing. The data from the same type of sensors in the data subset are weighted averaged to eliminate accidental errors in a single acquisition and improve the reliability of the data. After completing the weighted average, the aggregated environmental parameter data is detected for outliers to identify abnormal points in the data. Outlier detection can use statistical methods such as the three sigma criterion (that is, data values ​​exceeding three times the standard deviation of the mean are considered abnormal) or machine learning methods such as the isolation forest algorithm. The detected abnormal data points will be marked, and then the data will be removed for outliers based on the marks to obtain a cleaned data set. The cleaned data set is normalized to convert data of different dimensions to the same dimension to obtain standardized environmental parameter data. The standardized environmental parameter data is reshaped by time step and feature dimensions to reshape the data into a time series data frame. For example, the data of multiple consecutive time points are spliced ​​into a multidimensional array, and each dimension represents a feature. The data frame constructed initially is subjected to a sliding window operation to generate a data sample set. Sliding window operation is a common data enhancement technique. It generates multiple overlapping samples by applying a window function to the original data, which helps capture the dynamic changes of the data in time series. Assuming that the size of the sliding window is ω and the step length is s, the data in each window constitutes a sample, and the window moves s time steps each time. Perform a quality assessment on the data sample set to determine whether the data meets the requirements of model training and obtain the preprocessed data set. Quality assessment can be performed by calculating statistics such as variance, mean, skewness, etc. of the data to ensure the uniformity and diversity of the data.

[0052] In a specific embodiment, the process of executing step S200 may specifically include the following steps:

[0053] Perform random matrix projection on the preprocessed data set to obtain a low-dimensional compressed representation, and perform sparse coding on the low-dimensional compressed representation to obtain a sparse coefficient vector;

[0054] Reconstruct the preprocessed data set according to the sparse coefficient vector to obtain a preliminary interpolation result, and process the preliminary interpolation result with a variational autoencoder to obtain a latent space representation;

[0055] Random sampling is performed according to the latent space representation to obtain re-parameterized samples, and the re-parameterized samples are processed by a decoder to obtain reconstructed data;

[0056] The reconstruction error and KL divergence are calculated based on the reconstructed data and the preprocessed data set to obtain the variational lower bound, and the variational lower bound is maximized to obtain the optimized generation model parameters;

[0057] The missing data are conditionally generated according to the optimized generation model parameters to obtain a complete data set after interpolation, and data enhancement processing is performed on the complete data set after interpolation to obtain an interpolated and enhanced data set.

[0058] Specifically, the random matrix projection method is used to reduce the dimension of the data. Random matrix projection is an efficient dimensionality reduction method that uses a randomly generated matrix Project high-dimensional data into low-dimensional space. Assume that the preprocessed data set is (Dimensions are ,in is the sample size, is the number of features), then the low-dimensional compressed representation The calculation formula is:

[0059] ;

[0060] in, is a A random matrix, is the dimension of the target low-dimensional space. The projection operation helps to reduce the dimension of the data while preserving the structural information and characteristics of the original data as much as possible. Sparse coding is performed. Sparse coding is a representation learning method that represents the input data as a sparse linear combination. The formula is:

[0061] ;

[0062] in, is a sparse dictionary, is a sparse coefficient vector. By optimizing the dictionary and the sparse coefficient To represent low-dimensional data, improve the representation ability and sparsity of data. According to the sparse coefficient vector For the original data Reconstruction is performed to obtain preliminary interpolation results to fill in the missing data in the data set due to various reasons (such as sensor failure or data loss) to ensure data integrity. The preliminary interpolation results are processed by a variational autoencoder. A variational autoencoder is a generative model that generates new data samples by learning the potential distribution of data. In the variational autoencoder processing, the preliminary interpolation results are mapped to the latent space representation. The latent space representation is a low-dimensional representation of the data that can capture the intrinsic structure of the data. Random sampling is performed based on the latent space representation to obtain reparameterized samples . The reparameterization technique ensures the stability of the model and the diversity of the generated data by introducing noise. The reparameterized samples are then decoded by the decoder to generate reconstructed data. After obtaining the reconstructed data, the reconstruction error between the reconstructed data and the preprocessed data set, as well as the KL divergence between the latent distribution and the standard normal distribution are calculated. The reconstruction error reflects the difference between the generated data and the original data, while the KL divergence measures the difference between the distribution of the latent space and the prior distribution (usually the standard normal distribution). The formula is as follows:

[0063] ;

[0064] in, is the variational lower bound, is the latent distribution of the encoder output, is the reconstructed data distribution output by the decoder, yes Divergence. Maximizing the variational lower bound helps optimize the parameters of the generative model so that the model can better reconstruct and generate data. According to the optimized parameters of the generative model, the missing data is conditionally generated to obtain the complete dataset after interpolation. The interpolated data is generated by the variational autoencoder model, which can supplement the missing parts of the dataset and make the dataset complete and consistent. In order to improve the applicability and diversity of the data, the complete dataset after interpolation is subjected to data enhancement. Data enhancement technology can expand the dataset by adding noise, data transformation (such as rotation, scaling), etc., to prevent overfitting and improve the generalization ability of the model.

[0065] In a specific embodiment, the process of executing step S300 may specifically include the following steps:

[0066] Perform multi-scale convolution processing on the interpolated and enhanced data sets to obtain the initial feature map, where the convolution kernel sizes are 3x3, 5x5, and 7x7 respectively;

[0067] The initial feature map is batch normalized to obtain a standardized feature map, and the standardized feature map is processed by the ReLU activation function to obtain a nonlinear feature map;

[0068] Perform a maximum pooling operation on the nonlinear feature map to obtain a downsampled feature map, and construct a residual connection based on the downsampled feature map to obtain a fused feature map;

[0069] Perform global average pooling on the fused feature map to obtain the channel descriptor, and perform two-layer fully connected network processing on the channel descriptor to obtain the attention weight;

[0070] The fused feature map is weighted according to the attention weight to obtain a weighted feature map, multi-scale feature fusion is performed on the weighted feature map to obtain a multi-scale feature representation, and 1x1 convolution is performed on the multi-scale feature representation to obtain a weighted feature representation.

[0071] Specifically, a multi-scale convolutional neural network is designed, where the convolution kernel sizes are 3x3, 5x5, and 7x7. Convolution kernels of different sizes can capture features of different scales, which helps to enrich the effect of feature extraction. Assume that the input dataset is , after multi-scale convolution processing, the initial feature map is obtained Among them, The output of a convolution kernel can be expressed as:

[0072] ;

[0073] Among them, * represents the convolution operation, It is The weight of the convolution kernel, is the corresponding bias term. Different convolution kernel sizes enable the system to extract features of different scales. For example, a 3x3 convolution kernel is more suitable for capturing detail information, while a 7x7 convolution kernel is more suitable for capturing global information. Perform batch normalization to obtain a standardized feature map The purpose of batch normalization is to reduce internal covariate shift, speed up network training and stabilize the training process. The calculation formula for the standardized feature map is:

[0074] ;

[0075] in, is the mean of the batch data, is the standard deviation, is a small constant used to avoid the denominator being zero. After batch normalization, the standardized feature map is processed by the ReLU activation function to obtain a nonlinear feature map The ReLU (Rectified Linear Unit) activation function introduces nonlinear features by setting negative values ​​to zero and retaining positive values, thereby improving the model's expressiveness. Perform a maximum pooling operation on the nonlinear feature map to obtain a downsampled feature map. The maximum pooling operation reduces the dimension by selecting the maximum value in the local area, which helps to reduce the amount of calculation and control overfitting while retaining important features. Assume that the pooling window size is , then the calculation formula of the downsampled feature map is:

[0076] ;

[0077] After obtaining the downsampled feature map, construct a residual connection to obtain the fused feature map Residual connection is a technique that can effectively solve the gradient vanishing problem in deep networks by directly connecting the input to the output. The formula for residual connection is:

[0078] ;

[0079] in, is the nonlinear feature map of the input. Through residual connection, the network can more easily learn the identity mapping, which is conducive to lossless transmission of information. Perform global average pooling to obtain channel descriptors . All values ​​in each feature map are averaged to compress the feature map into a single value. The channel descriptor is calculated as:

[0080] ;

[0081] in, and are the height and width of the feature map, It is The feature map of each channel is obtained by global average pooling. Input into a two-layer fully connected network to calculate the attention weight The two-layer fully connected network assigns weights to different channels by learning the importance of each channel. The calculation formula is:

[0082] ;

[0083] in, and is the weight of the fully connected layer, and is the bias term, represents a nonlinear activation function (such as ReLU), and softmax is used to normalize the output to a probability distribution. Attention weight It reflects the relative importance of each channel in the entire feature map. According to the calculated attention weight , for the fusion feature map Weighted to get the weighted feature map , enhance the information of important channels and suppress the interference of unimportant channels. The calculation formula of weighted feature map is:

[0084] ;

[0085] Weighted feature map Perform multi-scale feature fusion to obtain multi-scale feature representation Multi-scale feature fusion improves the richness and robustness of features by combining features of different scales. The multi-scale feature representation is processed by 1x1 convolution to obtain the final weighted feature representation. The role of 1x1 convolution is to linearly combine the input channels, thereby reducing the dimension of the feature map and introducing nonlinear changes. The formula is:

[0086] ;

[0087] in, is the weight matrix of the 1x1 convolution, is the bias term. Through these processing steps, we finally get a comprehensive feature representation that can effectively capture feature information of different scales.

[0088] In a specific embodiment, the process of executing step S400 may specifically include the following steps:

[0089] Rearrange the weighted feature representation in time series to obtain a time series feature sequence;

[0090] Perform forward LSTM processing on the time series feature sequence to obtain a forward hidden state sequence, and perform reverse LSTM processing on the time series feature sequence to obtain a reverse hidden state sequence;

[0091] Concatenate the forward hidden state sequence and the reverse hidden state sequence to obtain a bidirectional hidden state sequence;

[0092] Perform a fully connected layer process on the bidirectional hidden state sequence to obtain an attention score sequence, and perform softmax normalization on the attention score sequence to obtain an attention weight sequence;

[0093] Perform weighted summation on the bidirectional hidden state sequence according to the attention weight sequence to obtain the first vector, and concatenate the first vector and the hidden state of the last time step to obtain the second vector;

[0094] The second vector is processed nonlinearly to obtain a nonlinear feature representation, and the nonlinear feature representation is subjected to Dropout regularization to obtain a context vector.

[0095] Specifically, the original data is organized into a feature matrix arranged in chronological order. Assume that the weighted feature is represented as , which contains data from multiple time steps, each of which has several features. The feature data is organized by time dimension to form a feature sequence ,in , each Represents the time step The feature vector of . By arranging, the time dependency and dynamic change characteristics in the data are captured. Perform forward LSTM (long short-term memory network) processing to obtain the forward hidden state sequence LSTM is a special recurrent neural network that can effectively process long dependencies in time series data. The forward LSTM starts processing data from the starting time step of the sequence, moving forward step by step, and updating the hidden state in sequence. Its calculation formula is:

[0096] ;

[0097] Among them, LSTM is the function of LSTM unit, is the input of the current time step, is the hidden state of the previous time step. At the same time, the time series feature sequence Perform reverse LSTM processing to obtain the reverse hidden state sequence The reverse LSTM starts processing data from the end of the sequence, and moves forward step by step to update the hidden state. This enables the network to capture information in the time series from two directions and enhances the modeling ability of time dependency. The calculation formula of the reverse LSTM is:

[0098] ;

[0099] Concatenate the forward hidden state sequence and the reverse hidden state sequence to obtain a bidirectional hidden state sequence ,in The bidirectional hidden state sequence combines the forward and backward information to more comprehensively represent the characteristics of each time step. Perform full connection layer processing to generate attention score sequence The fully connected layer maps the bidirectional hidden state sequence to a new representation space for calculating the attention score at each time step. The calculation formula for the attention score is:

[0100] ;

[0101] in, and is the weight of the fully connected layer, is the bias term, is the activation function. The attention score sequence is normalized by softmax to obtain the attention weight sequence The Softmax function converts the score into a probability distribution so that the sum of the weights is 1. The calculation formula for the attention weight is:

[0102] ;

[0103] According to the attention weight sequence For the bidirectional hidden state sequence Perform weighted summation to obtain the first vector The formula for weighted summation is:

[0104] ;

[0105] The first vector and the hidden state at the last time step Concatenate and get the second vector The concatenation operation combines the global attention information with the specific state information of the last time step. Perform nonlinear processing to obtain nonlinear feature representation Nonlinear processing usually uses activation functions such as ReLU (Rectified Linear Unit), whose formula is:

[0106] ;

[0107] in, and are the weights and biases of the fully connected layer. Nonlinear processing can increase the expressive power of the model, allowing it to capture more complex patterns and features. Dropout regularization is performed on the nonlinear feature representation to obtain the context vector. Dropout is a regularization technique that prevents overfitting and improves the generalization ability of the model by randomly discarding a portion of neurons.

[0108] In a specific embodiment, the process of executing step S500 may specifically include the following steps:

[0109] The context vector is processed by the task fully connected layer to obtain multiple task-related feature vectors, and the multiple task-related feature vectors are processed by the output layer respectively to obtain the predicted values ​​of each environmental parameter;

[0110] The task loss is calculated based on the predicted values ​​of each environmental parameter to obtain multiple single-task loss values, and multiple task-related feature vectors are concatenated to obtain a comprehensive feature vector.

[0111] The comprehensive feature vector is processed by a multi-layer perceptron to obtain the environmental state features, and the environmental state features are classified by Softmax to obtain the environmental state probability distribution;

[0112] Calculate the classification loss according to the probability distribution of the environment state to obtain the environment state evaluation loss, and perform weighted summation of multiple single task loss values ​​and environment state evaluation losses to obtain a comprehensive loss value;

[0113] The uncertainty between tasks is calculated according to the comprehensive loss value to obtain the dynamic weight coefficient, and the dynamic weight coefficient is used to update the comprehensive loss value to obtain the multi-parameter prediction results and environmental status evaluation indicators.

[0114] Specifically, the context vector The input is sent to different fully connected layers, each of which corresponds to a task. These tasks may include predicting different environmental parameters such as temperature, humidity, CO2 concentration, etc. Task, then The feature vector of the task It can be expressed as:

[0115] ;

[0116] in, It is The weight matrix of each task, is the bias term. Through the processing, we get multiple task-related feature vectors , these feature vectors represent the feature information related to each task. The task-related feature vectors are processed by the output layer to obtain the predicted values ​​of each environmental parameter. The output layer is usually a fully connected layer, and its output directly gives the predicted value. For example, for the temperature prediction task, the result of the output layer is It can be expressed as:

[0117] ;

[0118] in, and are the weight and bias of the temperature task respectively. Similar calculations are also applicable to other environmental parameters, such as humidity, CO2 concentration, etc. After obtaining the predicted values ​​of each environmental parameter, the task loss is calculated based on these predicted values. Assuming the actual value is , then the loss of each task can be calculated by the mean square error (MSE), the formula is:

[0119] ;

[0120] in, It is The loss value of each task, It is The predicted value of samples, is the true value, is the number of samples. Multiple task-related feature vectors are concatenated to obtain a comprehensive feature vector The comprehensive feature vector can be expressed as:

[0121] ;

[0122] The comprehensive feature vector Perform multi-layer perceptron processing to obtain environmental state characteristics The multilayer perceptron consists of multiple fully connected layers and can extract complex nonlinear features. The calculation formula is:

[0123] ;

[0124] in, and are the weights and biases of the multilayer perceptron, is an activation function (such as ReLU). The environmental state features are processed by the Softmax classifier to obtain the probability distribution of the environmental state The function of the Softmax function is to convert the output of the network into a probability distribution so that the sum of the probabilities of each category is 1. Its formula is:

[0125] ;

[0126] in, Indicates The probability of the categories, is the total number of categories. According to the probability distribution of the environment state, the classification loss is calculated, usually using the cross entropy loss. The formula for classification loss is

[0127] ;

[0128] in, It is The actual label of the class. The weighted sum of multiple single-task loss values ​​and environmental state evaluation loss is used to obtain the comprehensive loss value. The calculation formula is:

[0129] ;

[0130] in, and is the corresponding weight coefficient used to balance the loss of each task. In order to dynamically adjust the influence of different tasks, the uncertainty between tasks is calculated according to the comprehensive loss value to obtain the dynamic weight coefficient. These coefficients can be adjusted according to the training situation of each task, so that the model pays more attention to difficult-to-learn tasks. The calculation of dynamic weight coefficients may be based on factors such as the standard deviation of the task and the training progress. The dynamic weight coefficient is used to update the comprehensive loss value and optimize the parameters of the model to obtain more accurate multi-parameter prediction results and environmental status evaluation indicators. The model is trained through the back propagation algorithm, and the weights and bias terms are continuously adjusted to achieve the best performance of the model on each task. Through the multi-task learning method, multiple environmental parameters are predicted and status evaluated at the same time to improve the monitoring capability and accuracy of the system.

[0131] In a specific embodiment, the process of executing step S600 may specifically include the following steps:

[0132] The multi-parameter prediction results are fitted with normal distribution to obtain the probability density function of each parameter, and the confidence interval of the predicted value of each parameter is calculated based on the probability density function to obtain the parameter fluctuation range;

[0133] Compare the parameter fluctuation range with the preset multi-level warning threshold to obtain the preliminary warning level, and calculate the state transition probability based on the environmental state assessment index to obtain the state transition matrix;

[0134] Perform Markov chain analysis on the state transfer matrix to obtain the steady-state distribution, and predict the probability of future states based on the steady-state distribution and the current state to obtain the state warning index;

[0135] The preliminary warning level and the status warning index are weighted and integrated to obtain a comprehensive warning level, and real-time monitoring warning information is generated based on the comprehensive warning level;

[0136] Adaptive feedback is performed on the real-time monitoring warning information to obtain adaptive feedback results, and the warning threshold and weight coefficient are dynamically adjusted according to the adaptive feedback results to achieve adaptive optimization.

[0137] Specifically, the predicted value of each parameter is regarded as sample data, and its distribution characteristics are modeled by fitting the normal distribution. Suppose there is a set of environmental parameter prediction results ,in Indicates The predicted value of samples. The normal distribution parameters obtained by fitting include the mean and standard deviation , its probability density function is expressed as:

[0138] ;

[0139] By maximum likelihood estimation, we can find the mean and standard deviation According to the normal distribution obtained by fitting, the confidence interval of the predicted value of each parameter is calculated. Assuming that the system wants to calculate a 95% confidence interval, the upper and lower bounds of the confidence interval can be expressed as:

[0140] ;

[0141] Among them, 1.96 is the critical value under the standard normal distribution, corresponding to a confidence level of about 95%. The confidence interval provides the parameter fluctuation range. The parameter fluctuation range is compared with the preset multi-level warning threshold. The preset warning threshold usually includes multiple levels of warning standards, such as normal, warning, critical, etc. Assume that the warning threshold of the temperature parameter is set to , the initial warning level of the current parameter is determined based on the upper and lower bounds of the confidence interval. For example, if the upper bound of the confidence interval exceeds But not reached , then the warning level can be set as "warning". After obtaining the preliminary warning level, the system status is analyzed in combination with the environmental status assessment index. By calculating the state transition probability, the state transition matrix is ​​obtained , whose elements Indicates from the state Transfer to state The state transfer matrix is ​​calculated by counting the state transfer frequency in the historical data. Perform Markov chain analysis to obtain the steady-state distribution of the system , that is, the probability distribution of the system in various states under long-term operation. The calculation formula for steady-state distribution is:

[0142] ;

[0143] and ,in Indicates the steady-state distribution The probability of a state. According to the steady-state distribution and the current state, the probability of the future state is predicted to obtain the state warning index. This index reflects the probability of abnormal situations that may occur in the system in the future and provides a quantitative assessment of potential risks. The preliminary warning level and the state warning index are weighted and fused to obtain a comprehensive warning level. Real-time monitoring warning information is generated according to the comprehensive warning level. This information may include the current environmental status, warning level, recommended response measures, etc. Adaptive feedback analysis is performed on the real-time monitoring warning information. The warning system is adjusted and optimized according to actual conditions. Adaptive feedback results are obtained by analyzing actual events and the response of the warning system. These feedback results can be used to dynamically adjust the warning threshold and weight coefficient so that the warning system can more accurately reflect the actual risks.

[0144] See also Figure 2 , Figure 2 A schematic block diagram of the structure of a real-time monitoring system 200 for a stem cell storage environment provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the real-time monitoring system 200 for the stem cell storage environment includes:

[0145] The acquisition module 210 is used to acquire various environmental parameter data stored in the stem cells and pre-process the various environmental parameter data to obtain a pre-processed data set;

[0146] An interpolation module 220 is used to perform deep compressed sensing and generative model processing on the preprocessed data set to obtain an interpolated and enhanced data set;

[0147] A weighting module 230 is used to perform convolutional neural network processing and feature attention mechanism calculation on the interpolated and enhanced data set to obtain a weighted feature representation;

[0148] A computing module 240 is used to perform bidirectional long short-term memory network processing and time-step attention mechanism calculation on the weighted feature representation to obtain a context vector;

[0149] Prediction module 250, used to perform multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain multi-parameter prediction results and environmental status evaluation indicators;

[0150] The early warning module 260 is used to perform multi-level threshold judgment and adaptive feedback processing on the multi-parameter prediction results and environmental status assessment indicators to obtain real-time monitoring early warning information.

[0151] Through the collaboration of the above components, deep compressed sensing and generative model processing, a comprehensive analysis of multiple environmental parameters is achieved, which can more comprehensively reflect the overall state of the stem cell storage environment and improve the comprehensiveness and accuracy of monitoring. Advanced data preprocessing and interpolation techniques are used to effectively solve the problems of noise, missing and outliers in environmental data, significantly improve data quality, and combine convolutional neural networks and feature attention mechanisms to adaptively extract and strengthen key features, thereby improving the sensitivity and generalization ability of the model to environmental changes. Through the bidirectional long short-term memory network and time step attention mechanism, the long-term and short-term temporal dependencies of environmental parameters are effectively captured, and the prediction ability and stability of the model are enhanced. The use of multi-task parallel processing and comprehensive loss function calculation realizes the coordinated optimization of environmental parameter prediction and state evaluation, and improves the overall performance and efficiency of the model. The introduction of multi-level threshold judgment and adaptive feedback processing builds a flexible and accurate early warning system that can detect potential risks in a timely manner and give reasonable suggestions. By dynamically adjusting the warning threshold and weight coefficient, the adaptive optimization of the system is achieved, so that the monitoring method can continuously adapt to environmental changes and new monitoring needs.

[0152] The present application also provides a real-time monitoring device for a stem cell storage environment, the real-time monitoring device for a stem cell storage environment comprising a memory and a processor, the memory storing computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the real-time monitoring method for the stem cell storage environment in the above-mentioned embodiments.

[0153] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the method for real-time monitoring of a stem cell storage environment.

[0154] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0155] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0156] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A real-time monitoring method for stem cell storage environment, characterized in that: include: Collecting a variety of environmental parameter data stored in the stem cells, and preprocessing the multiple environmental parameter data to obtain a preprocessed data set; The preprocessed data set is subjected to deep compressed sensing and generative model processing to obtain an interpolated and enhanced data set; specifically comprising: performing random matrix projection on the preprocessed data set to obtain a low-dimensional compressed representation, and performing sparse coding on the low-dimensional compressed representation to obtain a sparse coefficient vector; reconstructing the preprocessed data set according to the sparse coefficient vector to obtain a preliminary interpolation result, and performing variational autoencoder processing on the preliminary interpolation result to obtain a latent space representation; performing random sampling according to the latent space representation to obtain a reparameterized sample, and performing decoder processing on the reparameterized sample to obtain reconstructed data; calculating the reconstruction error and KL divergence according to the reconstructed data and the preprocessed data set to obtain a variational lower bound, and maximizing the variational lower bound to obtain optimized generative model parameters; conditionally generating missing data according to the optimized generative model parameters to obtain a complete interpolated data set, and performing data enhancement processing on the complete interpolated data set to obtain an interpolated and enhanced data set; Performing convolutional neural network processing and feature attention mechanism calculation on the interpolated and enhanced data set to obtain a weighted feature representation; Performing bidirectional long short-term memory network processing and time-step attention mechanism calculation on the weighted feature representation to obtain a context vector; Performing multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain multi-parameter prediction results and environmental state evaluation indicators; Multi-level threshold judgment and adaptive feedback processing are performed on the multi-parameter prediction results and the environmental status assessment indicators to obtain real-time monitoring and early warning information.

2. The real-time monitoring method of stem cell storage environment according to claim 1, characterized in that: The collecting of the multiple environmental parameter data stored in the stem cells and preprocessing the multiple environmental parameter data to obtain the preprocessed data set includes: Setting the data collection frequency for various sensors in the stem cell storage device to obtain a preset collection time interval, and collecting data for the temperature, humidity, CO2 concentration, O2 concentration and pressure sensors according to the preset collection time interval to obtain various environmental parameter data; Dividing the multiple environmental parameter data into time windows to obtain data subsets, and performing weighted average calculation on similar sensor data in the data subsets to obtain aggregated environmental parameter data; Performing outlier detection on the aggregated environmental parameter data to obtain abnormal data point marks, and removing outliers from the aggregated environmental parameter data according to the abnormal data point marks to obtain a cleaned data set; Normalizing the cleaned data set to obtain standardized environmental parameter data, and reshaping the standardized environmental parameter data in terms of time step and feature dimension to obtain a preliminarily constructed data frame; A sliding window operation is performed on the initially constructed data frame to obtain a data sample set, and a preprocessed data set is obtained based on a quality assessment of the data sample set.

3. The real-time monitoring method of stem cell storage environment according to claim 1, characterized in that: The interpolated and enhanced data sets are processed by convolutional neural network and feature attention mechanism is calculated to obtain weighted feature representation, including: Performing multi-scale convolution processing on the interpolated and enhanced data set to obtain an initial feature map, wherein the convolution kernel sizes are 3x3, 5x5, and 7x7 respectively; Performing batch normalization processing on the initial feature map to obtain a standardized feature map, and performing ReLU activation function processing on the standardized feature map to obtain a nonlinear feature map; Performing a maximum pooling operation on the nonlinear feature map to obtain a downsampled feature map, and constructing a residual connection based on the downsampled feature map to obtain a fused feature map; Performing global average pooling on the fused feature map to obtain a channel descriptor, and performing two-layer fully connected network processing on the channel descriptor to obtain an attention weight; The fused feature map is weighted according to the attention weight to obtain a weighted feature map, multi-scale feature fusion is performed on the weighted feature map to obtain a multi-scale feature representation, and 1x1 convolution processing is performed on the multi-scale feature representation to obtain a weighted feature representation.

4. The real-time monitoring method of stem cell storage environment according to claim 1, characterized in that: The weighted feature representation is subjected to bidirectional long short-term memory network processing and time-step attention mechanism calculation to obtain a context vector, including: Rearranging the weighted feature representation in time series to obtain a time series feature sequence; Performing forward LSTM processing on the time series feature sequence to obtain a forward hidden state sequence, and performing reverse LSTM processing on the time series feature sequence to obtain a reverse hidden state sequence; Concatenating the forward hidden state sequence and the reverse hidden state sequence to obtain a bidirectional hidden state sequence; Performing a fully connected layer processing on the bidirectional hidden state sequence to obtain an attention score sequence, and performing softmax normalization on the attention score sequence to obtain an attention weight sequence; Performing weighted summation on the bidirectional hidden state sequence according to the attention weight sequence to obtain a first vector, and concatenating the first vector and the hidden state of the last time step to obtain a second vector; Nonlinear processing is performed on the second vector to obtain a nonlinear feature representation, and Dropout regularization is performed on the nonlinear feature representation to obtain a context vector.

5. The real-time monitoring method of stem cell storage environment according to claim 1, characterized in that: The multi-task parallel processing and comprehensive loss function calculation of the context vector to obtain a multi-parameter prediction result and an environmental state evaluation index include: Performing task fully connected layer processing on the context vector to obtain multiple task-related feature vectors, and performing output layer processing on the multiple task-related feature vectors respectively to obtain predicted values ​​of each environmental parameter; Calculating the task loss according to the predicted values ​​of the environmental parameters to obtain a plurality of single-task loss values, and concatenating the plurality of task-related feature vectors to obtain a comprehensive feature vector; Performing multi-layer perceptron processing on the comprehensive feature vector to obtain environmental state features, and performing Softmax classification on the environmental state features to obtain environmental state probability distribution; Calculating the classification loss according to the environmental state probability distribution to obtain the environmental state evaluation loss, and performing weighted summation on the multiple single-task loss values ​​and the environmental state evaluation loss to obtain a comprehensive loss value; The uncertainty between tasks is calculated according to the comprehensive loss value to obtain a dynamic weight coefficient, and the comprehensive loss value is updated using the dynamic weight coefficient to obtain a multi-parameter prediction result and an environmental status evaluation index.

6. The real-time monitoring method of stem cell storage environment according to claim 1, characterized in that: The multi-parameter prediction results and the environmental status assessment indicators are subjected to multi-level threshold judgment and adaptive feedback processing to obtain real-time monitoring and early warning information, including: Performing normal distribution fitting on the multi-parameter prediction results to obtain the probability density function of each parameter, and calculating the confidence interval of the predicted value of each parameter according to the probability density function to obtain the parameter fluctuation range; Compare the parameter fluctuation range with the preset multi-level warning threshold to obtain a preliminary warning level, and calculate the state transition probability according to the environmental state assessment index to obtain a state transition matrix; Performing Markov chain analysis on the state transfer matrix to obtain a steady-state distribution, and predicting the probability of future states based on the steady-state distribution and the current state to obtain a state warning index; Performing weighted fusion on the preliminary warning level and the status warning index to obtain a comprehensive warning level, and generating real-time monitoring warning information according to the comprehensive warning level; Adaptive feedback is performed on the real-time monitoring warning information to obtain an adaptive feedback result, and the warning threshold and weight coefficient are dynamically adjusted according to the adaptive feedback result to achieve adaptive optimization.

7. A real-time monitoring system for stem cell storage environment, characterized in that: A method for performing real-time monitoring of a stem cell storage environment as claimed in any one of claims 1 to 6, comprising: A collection module, used for collecting a variety of environmental parameter data stored in the stem cells, and preprocessing the multiple environmental parameter data to obtain a preprocessed data set; An interpolation module is used to perform deep compressed sensing and generative model processing on the preprocessed data set to obtain an interpolated and enhanced data set; specifically comprising: performing random matrix projection on the preprocessed data set to obtain a low-dimensional compressed representation, and performing sparse coding on the low-dimensional compressed representation to obtain a sparse coefficient vector; reconstructing the preprocessed data set according to the sparse coefficient vector to obtain a preliminary interpolation result, and performing variational autoencoder processing on the preliminary interpolation result to obtain a latent space representation; performing random sampling according to the latent space representation to obtain reparameterized samples, and performing decoder processing on the reparameterized samples to obtain reconstructed data; calculating the reconstruction error and KL divergence according to the reconstructed data and the preprocessed data set to obtain a variational lower bound, and maximizing the variational lower bound to obtain optimized generative model parameters; conditionally generating missing data according to the optimized generative model parameters to obtain a complete interpolated data set, and performing data enhancement processing on the complete interpolated data set to obtain an interpolated and enhanced data set; A weighting module, used for performing convolutional neural network processing and feature attention mechanism calculation on the interpolated and enhanced data set to obtain a weighted feature representation; A computing module, used for performing bidirectional long short-term memory network processing and time-step attention mechanism calculation on the weighted feature representation to obtain a context vector; A prediction module, used to perform multi-task parallel processing and comprehensive loss function calculation on the context vector to obtain multi-parameter prediction results and environmental state evaluation indicators; The early warning module is used to perform multi-level threshold judgment and adaptive feedback processing on the multi-parameter prediction results and the environmental status assessment indicators to obtain real-time monitoring early warning information.

8. A real-time monitoring device for stem cell storage environment, characterized in that: The real-time monitoring device for the stem cell storage environment comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the real-time monitoring device for the stem cell storage environment to execute the real-time monitoring method for the stem cell storage environment according to any one of claims 1-6.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the real-time monitoring method for the stem cell storage environment according to any one of claims 1 to 6 is implemented.

Images