Intelligent constant temperature and humidity storage system for medical supplies and control method thereof

Abstract

The present invention relates to an intelligent constant temperature and humidity storage system for medical supplies and a control method thereof. The method: performs multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage scheme; establishes a temperature and humidity dynamic model for each storage area and performs error prediction and parameter adjustment to obtain an adaptive temperature and humidity control instruction; obtains a first temperature and humidity signal data set, performs multi-region collaborative processing through a multi-task learning model of a long short-term memory network, and obtains a first collaborative control scheme and a multi-region collaborative adjustment instruction; obtains a second temperature and humidity signal data set, calculates the storage environment distribution characteristics; performs optimization, obtains a second collaborative control scheme and performs filtering reconstruction to obtain a reconstructed temperature and humidity signal data set; performs state transition probability modeling to obtain a storage environment prediction result and an abnormal warning index. The implementation of the present invention realizes intelligent constant temperature and humidity storage of medical supplies.

Description

Technical Field

[0001] The present invention relates to the technical field of storage control, and in particular to an intelligent constant temperature and humidity storage system for medical supplies and a control method thereof. Background Art

[0002] As important materials to protect human health and life safety, the storage conditions of medical supplies have a crucial impact on their quality and effectiveness. Traditional storage methods for medical supplies often use fixed temperature and humidity control, which is difficult to adapt to the diverse storage needs of different types of medical supplies. With the continuous advancement of medical technology, various new types of medical supplies have higher and higher requirements for storage environment, which makes traditional storage methods face huge challenges.

[0003] The main problem facing the current storage of medical supplies is the inability to achieve accurate and dynamic temperature and humidity control. Different types of medical supplies have different requirements for temperature and humidity, and these requirements may change over time and with changes in the environment. Traditional fixed parameter control methods are difficult to meet such complex and changing needs, which can easily lead to a decline in the quality of medical supplies or even failure. In addition, existing storage systems lack intelligence and adaptive capabilities, and are unable to adjust control strategies in a timely manner according to real-time environmental changes, and are also difficult to predict and prevent potential abnormal situations. Summary of the invention

[0004] The main purpose of the present invention is to provide an intelligent constant temperature and humidity storage system for medical supplies and a control method thereof, so as to realize intelligent constant temperature and humidity storage of medical supplies.

[0005] To achieve the above object, the present invention provides an intelligent constant temperature and humidity storage method for medical supplies, comprising the following steps:

[0006] Performing multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage scheme, wherein the partition storage scheme includes multiple storage areas;

[0007] Establish a dynamic temperature and humidity model for each storage area and perform error prediction and parameter adjustment to obtain adaptive temperature and humidity control instructions for each storage area;

[0008] Execute the adaptive temperature and humidity control instruction and obtain a first temperature and humidity signal data set of each storage area, input the first temperature and humidity signal data set into a multi-task learning model of a long short-term memory network for multi-area collaborative processing, and obtain a first collaborative control scheme and a multi-area collaborative adjustment instruction;

[0009] Execute the multi-area coordinated adjustment instruction and obtain a second temperature and humidity signal data set of each storage area, and calculate the storage environment distribution characteristics of the second temperature and humidity signal data set;

[0010] The first collaborative control scheme is optimized according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and the second temperature and humidity signal data set is filtered and reconstructed to obtain a reconstructed temperature and humidity signal data set;

[0011] The state transition probability modeling is performed on the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain a storage environment prediction result and an abnormal warning index.

[0012] The present invention also provides an intelligent constant temperature and humidity storage system for medical supplies, comprising:

[0013] An analysis module, used for performing multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage scheme, wherein the partition storage scheme includes multiple storage areas;

[0014] Establishing a module for establishing a dynamic temperature and humidity model for each storage area and performing error prediction and parameter adjustment to obtain adaptive temperature and humidity control instructions for each storage area;

[0015] a processing module, configured to execute the adaptive temperature and humidity control instruction and obtain a first temperature and humidity signal data set of each storage area, input the first temperature and humidity signal data set into a multi-task learning model of a long short-term memory network for multi-area collaborative processing, and obtain a first collaborative control scheme and a multi-area collaborative adjustment instruction;

[0016] a calculation module, configured to execute the multi-region coordinated adjustment instruction and obtain a second temperature and humidity signal data set of each storage region, and calculate a storage environment distribution feature of the second temperature and humidity signal data set;

[0017] an optimization module, configured to optimize the first collaborative control scheme according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and filter and reconstruct the second temperature and humidity signal data set to obtain a reconstructed temperature and humidity signal data set;

[0018] The early warning module is used to perform state transition probability modeling on the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain a storage environment prediction result and an abnormal early warning index.

[0019] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0020] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0021] In summary, the technical solution provided by the present invention can automatically generate the optimal partition storage solution according to the storage requirement parameters of different medical supplies through multi-dimensional similarity calculation and cluster analysis, thereby improving the utilization efficiency of storage space. A dynamic temperature and humidity model is established for each storage area, and error prediction and parameter adjustment are performed to achieve accurate adaptive temperature and humidity control to meet the specific storage requirements of different medical supplies. The multi-task learning model of the long short-term memory network is used to achieve the coordinated processing and adjustment of multiple storage areas, thereby improving the stability and efficiency of the overall storage system. By calculating the distribution characteristics of the storage environment, the control scheme is optimized, and the system's adaptability to complex environmental changes is improved. The temperature and humidity signal data is filtered and reconstructed, which effectively reduces noise interference and improves data quality and control accuracy. Through state transition probability modeling, the prediction of the storage environment is realized, and an abnormal warning index is established, which can timely discover and warn potential storage environment abnormalities. Through decision trees and policy networks, intelligent control strategy generation and execution are realized, thereby improving the autonomy and adaptability of the system. The introduction of environmental complexity indicators can better describe and respond to the dynamic changes of the storage environment, thereby realizing intelligent constant temperature and humidity storage of medical supplies. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a schematic diagram of the steps of a method for intelligent constant temperature and humidity storage of medical supplies in one embodiment of the present invention;

[0023] Figure 2 This is a structural block diagram of an intelligent constant temperature and humidity storage system for medical supplies in one embodiment of the present invention;

[0024] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0025] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] Reference Figure 1 This embodiment provides a method for intelligent constant temperature and humidity storage of medical supplies, including the following steps:

[0028] S1, performing multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage solution, which includes multiple storage areas;

[0029] Among them, the storage requirement parameters of medical supplies are standardized to obtain a standardized parameter set, eliminating the differences between different parameter dimensions and making them comparable on the same scale. A parameter vector set is generated based on the standardized parameter set, and the storage requirement of each medical supply is converted into a high-dimensional vector. The cosine similarity calculation is performed on the parameter vector set to obtain a similarity matrix. The cosine similarity calculation can quantify the similarity between two vectors and construct a similarity matrix that reflects the similarity of the storage requirements of medical supplies. Based on the similarity matrix, a medical supply similarity graph is constructed. The nodes of the graph represent medical supplies, and the weights of the edges represent the similarity between medical supplies. Spectral clustering is performed on the medical supply similarity graph. Spectral clustering is a method for clustering by eigenvalue decomposition of a graph. By performing eigenvalue decomposition on the Laplace matrix of the similarity graph, the medical supplies are divided into multiple clusters to obtain the initial clustering results. The silhouette coefficient is calculated based on the initial clustering results. The silhouette coefficient is an indicator used to evaluate the clustering effect. The larger the value, the better the clustering effect. By calculating the silhouette coefficient, the optimal number of clusters is determined, so that the initial clustering results are clustered and optimized. Density peak detection is performed on the optimized clustering results. Density peak detection is a density-based cluster center detection method. By identifying high-density areas in the data distribution, the center of each cluster is determined to obtain the cluster center set. According to the cluster center set, the nearest neighbor allocation is performed for medical supplies. The nearest neighbor allocation method can ensure that each medical supply is allocated to the most similar storage area to obtain a preliminary partitioning scheme. The boundaries of the preliminary partitioning scheme are adjusted. By adjusting the boundaries of each storage area, the medical supplies in each area are more closely clustered together, reducing the similarity between different areas, and obtaining the final partition storage scheme. The partition storage scheme includes multiple storage areas, and the temperature and humidity control requirements of each storage area are relatively consistent, thereby realizing intelligent constant temperature and humidity storage of medical supplies.

[0030] S2, respectively establish a temperature and humidity dynamic model for each storage area and perform error prediction and parameter adjustment to obtain adaptive temperature and humidity control instructions for each storage area;

[0031] Among them, the historical temperature and humidity data of each storage area are analyzed in time series to obtain the periodic characteristics of the data. The data is decomposed into trend items, seasonal items and random items by seasonal decomposition. The trend items are fitted to obtain the trend prediction model. By analyzing and fitting the long-term change trend in the temperature and humidity data, the future trend changes are predicted. At the same time, the seasonal items are transformed by Fourier transform to obtain the seasonal periodic function. The Fourier transform can convert the time series data into frequency domain data and extract the periodic components in the data. The initial dynamic model of each storage area is constructed according to the trend prediction model and the seasonal periodic function, and the sensitivity analysis of the initial dynamic model is performed to identify the key parameters affecting the model performance. Through sensitivity analysis, the key parameter subsets can be determined, which have a significant impact on the output of the model. The key parameter subsets are tuned to improve the accuracy and reliability of the model and obtain a more accurate temperature and humidity dynamic model. The temperature and humidity dynamic model is simulated, and a state estimator is constructed through simulation. The estimator is used to monitor and estimate the state of the system in real time. The state estimator is used to calculate the filtered state estimate, and the error between the filtered state estimate and the dynamic model prediction value is calculated to obtain the prediction error sequence. The prediction error sequence is subjected to autocorrelation analysis to reveal the potential patterns and laws in the error and obtain the error pattern characteristics. According to the error pattern characteristics, the initial PID parameter set of the adaptive PID controller is calculated. The PID controller is a commonly used control system that achieves precise control of the system by adjusting the proportional, integral and differential parameters. The initial PID parameter set is discretized to adapt it to the requirements of the actual control system, and the adaptive temperature and humidity control instruction is obtained. The instruction can dynamically adjust the temperature and humidity control strategy of each storage area according to the real-time monitoring data and the model prediction results to ensure that the temperature and humidity of each storage area are kept within an appropriate range.

[0032] S3, executing the adaptive temperature and humidity control instruction and obtaining the first temperature and humidity signal data set of each storage area, inputting the first temperature and humidity signal data set into the multi-task learning model of the long short-term memory network for multi-area collaborative processing, and obtaining the first collaborative control scheme and the multi-area collaborative adjustment instruction;

[0033] Among them, the adaptive temperature and humidity control instruction is executed, and the first temperature and humidity signal data set of each storage area is obtained, which represents the temperature and humidity conditions of different storage areas within a specific time. The first temperature and humidity signal data set is divided into time windows and normalized to obtain standardized time series samples. The differences between different data dimensions are eliminated through normalization to ensure the consistency of the input data. The standardized time series samples are mapped to a high-dimensional space through an embedding layer to obtain an input feature vector, which helps to improve the representation ability of the model so that it can capture complex temperature and humidity change patterns. The input feature vector is input into the multi-task learning model of the long short-term memory network. In the first layer of the long short-term memory network, the cell state and hidden state are updated through the calculation of the forget gate, input gate and output gate to obtain the output sequence of the first layer. In order to prevent overfitting, the output sequence of the first layer is dropout-operated to obtain the feature sequence after dropout, and it is input into the second layer of the long short-term memory network. Through similar calculation steps, the cell state and hidden state are updated to obtain the output sequence of the second layer. The output sequence of the second layer is input into the shared representation layer in the multi-task learning model of the long short-term memory network. Through the fully connected layer and the ReLU activation function, the multi-region shared features are obtained, and the attention mechanism is applied to the multi-region shared features. The weight coefficients of different regions are calculated to obtain the weighted shared features. The weighted shared features are respectively input into the temperature prediction branch and the humidity prediction branch in the multi-task learning model of the long short-term memory network to perform temperature prediction and humidity prediction, and the temperature prediction value and humidity prediction value of each storage area are obtained. The temperature prediction value and the humidity prediction value are denormalized to obtain the actual scale prediction result, and the actual scale prediction result is compared with the set threshold to obtain the temperature and humidity anomaly detection result. According to the temperature and humidity anomaly detection results, a decision tree is constructed, and each leaf node corresponds to a control strategy to form a preliminary control strategy set. The preliminary control strategy set is optimized and searched to obtain the optimized control strategy. The optimized control strategy is input into the policy network, and the control parameters of the continuous action space are obtained through a three-layer fully connected network, each layer is followed by BatchNorm and ReLU activation function. According to the control parameters of the continuous action space, a first collaborative control scheme is created, and the execution instruction sequence of the first collaborative control scheme is parsed to obtain multi-region collaborative adjustment instructions.

[0034] The weighted shared features are separated to obtain the temperature feature vector and the humidity feature vector. The temperature feature vector and the humidity feature vector are respectively expanded in dimension. By expanding the dimension, the feature vector is converted into a more representative feature matrix, the representation ability and prediction accuracy of the model are improved, and the temperature feature matrix and the humidity feature matrix are obtained. The temperature feature matrix is ​​input into the first fully connected layer of the temperature prediction branch for weight matrix multiplication to obtain the original temperature output. Through the multiplication of the weight matrix, the input temperature feature matrix is ​​converted into a more abstract representation. The ReLU activation function is performed on the original temperature output. Through the ReLU activation function, nonlinear factors are introduced to improve the model's ability to fit complex relationships and obtain the intermediate temperature features. In order to prevent overfitting, the temperature intermediate features are dropout operated to obtain the temperature features after dropout. The dropout operation can randomly discard some neurons and enhance the generalization ability of the model. The temperature features after dropout are input into the second fully connected layer of the temperature prediction branch, and the weight matrix multiplication operation is performed again to obtain the temperature prediction value of each storage area. Similarly, the humidity feature matrix is ​​input into the first fully connected layer of the humidity prediction branch, and the weight matrix multiplication operation is performed to obtain the original humidity output. The ReLU activation function is performed on the original humidity output to obtain the intermediate humidity feature. In order to enhance the robustness of the model, the humidity intermediate feature is dropout to obtain the humidity feature after dropout. The humidity feature after dropout is input into the second fully connected layer of the humidity prediction branch, and the weight matrix multiplication operation is performed to obtain the humidity prediction value of each storage area.

[0035] The optimized control strategy is encoded and converted into a form suitable for input into the neural network to obtain the strategy vector. The strategy vector is dimensionally expanded and normalized to obtain the control feature matrix. By expanding the dimension, the feature expression ability is enhanced, and the normalization process helps to improve the training effect of the model. The control feature matrix is ​​input into the first fully connected layer of the strategy network for weight matrix multiplication to obtain the first layer original output. The weight matrix multiplication operation can convert the input features into a more abstract representation. The first layer original output is processed by BatchNorm. Through BatchNorm processing, the original output is standardized to accelerate the training of the model and improve the stability, and the standardized intermediate features are obtained. The standardized intermediate features are processed by ReLU activation function. Through the nonlinear transformation of ReLU activation function, the representation ability of the model is enhanced to obtain the features after the activation of the first layer. In order to prevent overfitting, the features after the activation of the first layer are dropout, and the final output of the first layer is obtained. The final output of the first layer is input into the second fully connected layer of the strategy network, and the weight matrix multiplication operation is performed to obtain the second layer original output. The original output of the second layer is processed by BatchNorm to obtain the standardized features of the second layer. The ReLU activation function is performed on the standardized features of the second layer to obtain the features after the activation of the second layer, and the residual connection is performed on the features after the activation of the second layer. Through the residual connection, the gradient vanishing problem in the deep neural network is effectively alleviated, the feature expression ability is enhanced, and the enhanced features of the second layer are obtained. The second layer enhanced features are subjected to the dropout operation to obtain the final output of the second layer, and are input into the third fully connected layer of the policy network for weight matrix multiplication to obtain the original output of the third layer. The original output of the third layer is processed by BatchNorm to obtain the standardized features of the third layer. The ReLU activation function is performed on the standardized features of the third layer to obtain the features after the activation of the third layer. In order to extract features and reduce the number of parameters, the features after the activation of the third layer are subjected to global average pooling to obtain feature vectors. Global average pooling can reduce each feature map to a single value, reducing computational complexity and overfitting risks. The feature vector is input into the output layer for full connection operation to obtain the control parameters of the continuous action space.

[0036] S4, executing the multi-region coordinated adjustment instruction and obtaining the second temperature and humidity signal data set of each storage area, and calculating the storage environment distribution characteristics of the second temperature and humidity signal data set;

[0037] Among them, the multi-region coordinated adjustment instruction is executed, and the second temperature and humidity signal data set of each storage area is obtained. These data sets reflect the temperature and humidity changes of each storage area under the multi-region coordinated adjustment instruction. The second temperature and humidity signal data set of each storage area is feature mapped, and the original signal data is converted into a more representative environmental state cluster. The environmental state cluster is formed by a clustering algorithm, and each cluster represents a typical temperature and humidity state. The inter-cluster and intra-cluster distances are calculated according to the environmental state cluster to obtain the environmental state distribution index. The calculation of the inter-cluster distance and the intra-cluster distance can quantify the similarities and differences between each environmental state cluster and better describe the distribution of the storage environment. A fuzzy membership matrix is ​​constructed according to the environmental state distribution index, and the entropy value of the environmental state is calculated through the fuzzy membership matrix to obtain the environmental complexity index. The environmental complexity index reflects the complexity of the storage environment through the entropy value. The environmental complexity index is subjected to time series analysis to obtain the complexity change trend. Through time series analysis, the law of environmental complexity change over time is identified, and eigenvalue decomposition is performed according to the complexity change trend to obtain a steady-state distribution vector. The steady-state distribution vector can reflect the distribution characteristics of the environmental state under long-term stability. The importance weight of each environmental state is calculated according to the steady-state distribution vector to obtain a weighted environmental state representation. The weighted environmental state representation more accurately describes the overall status of the storage environment by assigning different importance weights to different environmental states. A two-dimensional environmental state topology map is generated according to the weighted environmental state representation. The environmental state relationship between different storage areas is intuitively displayed in the form of a topology map. The topological structure features are calculated according to the two-dimensional environmental state topology map to obtain the environmental space distribution features. The topological structure features can reflect the distribution law and connection of the environmental state in space. The environmental space distribution features are decomposed to obtain the core tensor and factor matrix. The core tensor and factor matrix extract the main components and structures of the storage environment distribution features through the tensor decomposition method. The storage environment distribution features are reconstructed according to the core tensor and factor matrix.

[0038] S5, optimizing the first collaborative control scheme according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and filtering and reconstructing the second temperature and humidity signal data set to obtain a reconstructed temperature and humidity signal data set;

[0039] Among them, the storage environment distribution characteristics and the first collaborative control scheme are feature fused to obtain a fused feature vector. The fused feature vector is nonlinearly mapped to convert the fused feature vector into a high-dimensional representation to capture more feature information. The high-dimensional representation is denoised to obtain denoised features. The denoising process can effectively remove the noise components in the high-dimensional representation and improve the purity and accuracy of the feature representation. According to the denoised features, a probabilistic control model is constructed. A tree search is performed on the probabilistic control model to generate multiple candidate control sequences. The tree search generates a set of candidate control sequences by traversing possible control schemes to ensure that a better solution is found in the search space. Local search is performed on multiple candidate control sequences to obtain a local optimal solution. The local search improves the effect of the control scheme by fine-tuning and optimizing the candidate control sequences. According to the local optimal solution, a mapping relationship between the control parameters and the objective function is constructed. Through the mapping relationship, the optimized control parameters are linked to the objective function of the storage environment to generate a second collaborative control scheme. The second temperature and humidity signal data set is subjected to wavelet transform to obtain multi-scale decomposition coefficients. The wavelet transform extracts various details and features in the signal by decomposing the signal into components of different scales. The multi-scale decomposition coefficients are threshold processed to remove the noise components in the signal and retain important feature information. The processed multi-scale decomposition coefficients are subjected to inverse wavelet transform to obtain the reconstructed temperature and humidity signal data set. The inverse wavelet transform resynthesizes the processed components to form a reconstructed signal data set.

[0040] S6, conduct state transition probability modeling for the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain storage environment prediction results and abnormal warning index.

[0041] That is, the reconstructed temperature and humidity signal data set is segmented into time windows, and the continuous temperature and humidity signal data is divided into multiple manageable small segments to obtain time series data segments. The time series data segments are dynamically time-warped to obtain an aligned data sequence. Dynamic time warping aligns time series of different lengths through nonlinear transformation to make them consistent in the time dimension. The aligned data sequence is state-identified to obtain an initial state sequence. According to the initial state sequence and the second collaborative control scheme, a state-action pair is constructed. The state-action pair refers to the result of executing a specific control action in a specific state. The transition probability matrix and the emission probability matrix are calculated based on the state-action pair. The transition probability matrix describes the probability of transferring to the next state in the current state, while the emission probability matrix describes the probability of observing a certain temperature and humidity signal in a specific state. Through the probability matrix, the dependency relationship between states is generated to describe the mutual influence and conversion law between each state. The dependency relationship between states is sampled by sequence importance to obtain the state prediction distribution. Sequence importance sampling is a Monte Carlo method that estimates the distribution of future states by sampling the dependency relationship between states. According to the state prediction distribution, the storage environment prediction results are calculated to reflect the change trend of the storage environment in the future. Deviation analysis is performed on the storage environment prediction results to calculate the error distribution characteristics. Deviation analysis can identify the error distribution of the model by comparing the difference between the prediction results and the actual observation values. According to the error distribution characteristics, multi-level warning thresholds are set. By setting warning thresholds of different levels, hierarchical warnings for abnormal storage environment conditions are implemented. The trigger probability is calculated based on the multi-level warning thresholds to obtain a preliminary warning index. The trigger probability reflects the possibility of abnormal storage environment conditions under a specific warning threshold. Abnormal pattern recognition is performed on the preliminary warning index to obtain the abnormal warning index. Abnormal pattern recognition can identify potential abnormal conditions and issue warnings in a timely manner by analyzing the change pattern of the warning index.

[0042] In one example, multi-dimensional similarity calculation and cluster analysis are performed on the storage requirement parameters of medical supplies to obtain a partition storage solution, which includes multiple storage areas, including:

[0043] Standardizing the storage requirement parameters of medical supplies to obtain a standardized parameter set, and generating a parameter vector set according to the standardized parameter set;

[0044] Calculate the cosine similarity of the parameter vector set to obtain a similarity matrix, and construct a medical supplies similarity graph based on the similarity matrix;

[0045] Perform spectral clustering on the medical supplies similarity graph to obtain the initial clustering results, and calculate the silhouette coefficient based on the initial clustering results to obtain the optimal number of clusters;

[0046] The initial clustering results are clustered and optimized according to the optimal clustering number to obtain the optimized clustering results, and the density peak detection is performed on the optimized clustering results to obtain the cluster center set;

[0047] The nearest neighbors of the medical supplies are assigned according to the cluster center set to obtain a preliminary partitioning scheme, and the boundaries of the preliminary partitioning scheme are adjusted to obtain a partitioning storage scheme, which includes multiple storage areas.

[0048] In this example, the storage requirement parameters of medical supplies are collected and organized. These parameters may include temperature, humidity, lighting conditions, air quality, etc. By standardizing these parameters, the differences between different parameter dimensions are eliminated, and all parameters are compared on the same scale to obtain a standardized parameter set. A parameter vector set is generated based on the standardized parameter set, and the standardized parameters of each medical product are used as a high-dimensional vector. For example, if the storage parameters of a medical product include temperature, humidity, and lighting conditions, its parameter vector can be expressed as ,in are the standardized values ​​of temperature, humidity and light conditions respectively. The cosine similarity is calculated for the parameter vector set to obtain the similarity matrix. The formula for calculating cosine similarity is:

[0049] ;

[0050] in, and Represent the parameter vectors of two medical supplies, and Respectively represent the modulus of the parameter vector. By calculating the cosine similarity, the similarity matrix between medical supplies is obtained. Each element in the similarity matrix represents the similarity between two medical supplies in terms of storage requirements. According to the similarity matrix, a medical supply similarity graph is constructed. The nodes of the similarity graph represent medical supplies, and the weights of the edges represent the similarity between medical supplies. Spectral clustering is performed on the medical supply similarity graph. Spectral clustering divides the graph into multiple clusters by performing eigenvalue decomposition on the Laplace matrix of the similarity graph, thereby obtaining the initial clustering result. Laplace matrix is defined as:

[0051] ;

[0052] in, is the similarity matrix, It is a diagonal matrix whose diagonal elements are the degrees of the nodes. By calculating the eigenvectors of the Laplacian matrix, the similarity graph is decomposed into different subgraphs, each of which represents a cluster. The silhouette coefficient is calculated based on the initial clustering results to evaluate the clustering effect. The calculation formula of the silhouette coefficient is:

[0053] ;

[0054] in, It is a data point The average distance to other points in the same cluster, It is a data point The average distance to the nearest cluster. By calculating the average value of the silhouette coefficient of all data points, the overall effect of clustering is determined and the optimal number of clusters is obtained. The initial clustering results are clustered and optimized according to the optimal number of clusters. The clustering effect is optimized by adjusting the cluster boundaries to obtain the optimized clustering results. Density peak detection is performed on the optimized clustering results to identify the center point of each cluster. Density peak detection determines the cluster center set by calculating the local density of each data point and the minimum distance to other high-density points. According to the cluster center set, the nearest neighbor assignment is performed for medical supplies, and each medical supply is assigned to the nearest cluster center to obtain a preliminary partitioning scheme. The boundaries of the preliminary partitioning scheme are adjusted to ensure that the medical supplies in each storage area have a high similarity, and the final partition storage scheme is obtained. The partition storage scheme includes multiple storage areas, and the medical supplies in each area have similar storage requirements, thereby optimizing the storage environment and improving the efficiency of storage management.

[0055] In one example, a temperature and humidity dynamic model is established for each storage area, and error prediction and parameter adjustment are performed to obtain an adaptive temperature and humidity control instruction for each storage area, including:

[0056] Perform time series analysis on the historical temperature and humidity data of each storage area to obtain periodic characteristics, and perform seasonal decomposition on the data based on the periodic characteristics to obtain trend items, seasonal items, and random items;

[0057] The trend term is fitted to obtain the trend prediction model, and the seasonal term is Fourier transformed to obtain the seasonal periodic function;

[0058] The initial dynamic model of each storage area is constructed based on the trend prediction model and seasonal cycle function, and the sensitivity analysis of the initial dynamic model is carried out to obtain a subset of key parameters;

[0059] Optimize the key parameters to obtain the temperature and humidity dynamic model, simulate the temperature and humidity dynamic model, and build a state estimator;

[0060] The state estimation value after filtering is calculated by the state estimator, and the error between the state estimation value after filtering and the dynamic model prediction value is calculated to obtain the prediction error sequence, and the prediction error sequence is analyzed by autocorrelation to obtain the error pattern characteristics;

[0061] The initial PID parameter set of the adaptive PID controller is calculated according to the error pattern characteristics, and the initial PID parameter set is discretized to obtain the adaptive temperature and humidity control instructions.

[0062] In this example, the historical temperature and humidity data of each storage area is analyzed in time series to identify the periodic features in the data. For example, the data is analyzed by autocorrelation function and partial autocorrelation function to extract the periodic pattern of the data. The data is seasonally decomposed according to the periodic features to obtain trend terms, seasonal terms, and random terms. Seasonal decomposition usually uses classic decomposition methods, such as the moving average method. The formula is:

[0063] ;

[0064] in, It's time The observed value of is a trend item, is the seasonal term, is a random term. Through this step, the original time series data is split into different components. The trend term is fitted to obtain a trend prediction model. The trend term can usually be fitted using linear regression, quadratic curve or other suitable regression models. For example, for a linear trend term, the following formula can be used for fitting:

[0065] ;

[0066] in, and are regression coefficients, estimated by the least squares method. By performing Fourier transform on the seasonal term, its periodic component can be extracted. The formula for Fourier transform is:

[0067] ;

[0068] in, and are the Fourier coefficients, is a period. Through Fourier transform, the seasonal term is expressed as a periodic function to reveal its periodic law. According to the trend prediction model and the seasonal periodic function, the initial dynamic model of each storage area is constructed. The initial dynamic model combines the superpotential term and the seasonal term to describe the dynamic process of temperature and humidity changing over time. A sensitivity analysis is performed on the initial dynamic model to identify the key parameters that have the greatest impact on the model output. Sensitivity analysis is usually performed by changing the model parameters and observing the changes in the model output. Parameter tuning is performed on a subset of key parameters to optimize these parameters to improve the accuracy of the model. Parameter tuning can use methods such as grid search, random search or Bayesian optimization. Through parameter tuning, a more accurate temperature and humidity dynamic model is obtained, and the model is simulated to construct a state estimator. The state estimator is used to estimate the state of the system in real time. Common methods include Kalman filtering and extended Kalman filtering. The filtered state estimate is calculated by the state estimator, and the error between these estimates and the predicted values ​​of the dynamic model is calculated to obtain a prediction error sequence. The error calculation formula is:

[0069] ;

[0070] in, is the error, is the actual observed value, is the predicted value. Autocorrelation analysis is performed on the predicted error sequence to identify the pattern characteristics of the error. Autocorrelation analysis can reveal the periodicity and correlation characteristics in the error sequence, and is usually performed using an autocorrelation function. Based on the error pattern characteristics, the initial PID parameter set of the adaptive PID controller is calculated. The formula for the PID controller is:

[0071] ;

[0072] in, is the control output, is the error, is the proportional gain, is the integral gain, is the differential gain. The initial PID parameter set is obtained through empirical formulas or automatic parameter adjustment algorithms (such as the Ziegler-Nichols method). The initial PID parameter set is discretized to generate adaptive temperature and humidity control instructions. Discretization converts continuous control parameters into discrete control instructions for actual control system execution.

[0073] In one example, an adaptive temperature and humidity control instruction is executed and a first temperature and humidity signal data set of each storage area is obtained, and the first temperature and humidity signal data set is input into a multi-task learning model of a long short-term memory network for multi-area collaborative processing to obtain a first collaborative control scheme and a multi-area collaborative adjustment instruction, including:

[0074] Execute the adaptive temperature and humidity control instruction and obtain the first temperature and humidity signal data set of each storage area;

[0075] Performing time window division and normalization processing on the first temperature and humidity signal data set to obtain standardized time series samples, and mapping the standardized time series samples to a high-dimensional space through an embedding layer to obtain an input feature vector;

[0076] Input the input feature vector into the first layer of the long short-term memory network in the multi-task learning model of the long short-term memory network, update the cell state and hidden state through the calculation of the forget gate, input gate and output gate, and obtain the output sequence of the first layer;

[0077] Perform a dropout operation on the output sequence of the first layer to obtain a feature sequence after dropout, and input the feature sequence after dropout into the second layer of the long short-term memory network. By calculating the forget gate, input gate, and output gate, the cell state and hidden state are updated to obtain the output sequence of the second layer.

[0078] The output sequence of the second layer is input into the shared representation layer in the multi-task learning model of the long short-term memory network. Through the fully connected layer and the ReLU activation function, the multi-region shared features are obtained. The attention mechanism is applied to the multi-region shared features, and the weight coefficients of different regions are calculated to obtain the weighted shared features.

[0079] The weighted shared features are respectively input into the temperature prediction branch and the humidity prediction branch in the multi-task learning model of the long short-term memory network to perform temperature prediction and humidity prediction, and the temperature prediction value and humidity prediction value of each storage area are obtained;

[0080] Denormalize the predicted temperature and humidity values ​​to obtain the actual scale prediction results, and compare the actual scale prediction results with the set thresholds to obtain the temperature and humidity anomaly detection results;

[0081] A decision tree is constructed based on the abnormal temperature and humidity detection results. Each leaf node corresponds to a control strategy, and a preliminary control strategy set is obtained. The preliminary control strategy set is optimized and searched to obtain an optimized control strategy.

[0082] The optimized control strategy is input into the policy network and passed through a three-layer fully connected network, with each layer followed by a BatchNorm and ReLU activation function, to obtain the control parameters of the continuous action space.

[0083] A first collaborative control scheme is created according to the control parameters of the continuous action space, and an execution instruction sequence of the first collaborative control scheme is parsed to obtain a multi-region collaborative adjustment instruction.

[0084] In this example, the adaptive temperature and humidity control instruction is executed, which will be adjusted according to the real-time temperature and humidity data of the storage area to maintain the stability of the environment. After the control instruction is executed, the first temperature and humidity signal data set of each storage area is obtained in real time. The data set is divided into time windows so that the continuous time series data can be divided into segments of fixed length. The divided data segments are normalized to standardize the data so that they are on the same scale to obtain standardized time series samples. The standardized time series samples are mapped to a high-dimensional space through an embedding layer to obtain an input feature vector. The role of the embedding layer is to map the original data to a high-dimensional feature space so that the model can better capture the characteristics and patterns of the data. The input feature vector is input into the multi-task learning model of the long short-term memory network (LSTM). In the first layer of the LSTM, the cell state and hidden state are updated through the calculation of the forget gate, input gate, and output gate to obtain the output sequence of the first layer. The gating mechanism of the LSTM can effectively capture the long-range dependencies in the time series data. Its calculation formula is:

[0085] ;

[0086] in, , , are the activation values ​​of the forget gate, input gate, and output gate, respectively. , , , is the weight matrix, , , , is the bias term, is the cell state, is the hidden state, is the current input, is the hidden state at the previous moment. Dropout operation is performed on the output sequence of the first layer to prevent overfitting, and the feature sequence after dropout is obtained. The feature sequence after dropout is input into the second layer of LSTM, and the cell state and hidden state are updated through similar calculation steps to obtain the output sequence of the second layer. The LSTM layer of the second layer further extracts the deep features of the time series data. The output sequence of the second layer is input into the shared representation layer in the multi-task learning model of LSTM, and the multi-region shared features are obtained through the fully connected layer and ReLU activation function. The fully connected layer of the shared representation layer maps the high-dimensional features to a more compact representation space, and the ReLU activation function introduces nonlinear transformation to improve the expression ability of the model. The attention mechanism is applied to the multi-region shared features, and the weight coefficients of different regions are calculated to obtain the weighted shared features. The attention mechanism can capture the correlation between different storage areas, and the contribution of each area to the shared features is determined by calculating the attention weight. The weighted shared features are respectively input into the temperature prediction branch and the humidity prediction branch in the multi-task learning model of LSTM to perform temperature prediction and humidity prediction, and obtain the temperature prediction value and humidity prediction value of each storage area. The predicted value is denormalized to obtain the actual scale prediction result. The denormalization formula is:

[0087] ;

[0088] in, is the denormalized data value, is the normalized data value, and are the minimum and maximum values ​​of the original data respectively. The predicted results of the actual scale are compared with the set threshold to obtain the temperature and humidity anomaly detection results. Based on the detection results, a decision tree is constructed, and each leaf node corresponds to a control strategy to obtain a preliminary control strategy set. By splitting the input features, the decision tree can generate a series of control strategies for dealing with different temperature and humidity anomalies. The preliminary control strategy set is optimized and searched to obtain the optimized control strategy. The optimization search can use heuristic search algorithms, such as simulated annealing, genetic algorithms, etc., to find the optimal control strategy. The optimized control strategy is input into the strategy network, and the control parameters of the continuous action space are obtained through a three-layer fully connected network, each layer is followed by BatchNorm and ReLU activation functions. The three-layer fully connected network can effectively extract the features of the control strategy, BatchNorm can accelerate training and improve stability, and the ReLU activation function introduces nonlinear transformation. According to the control parameters of the continuous action space, the first collaborative control scheme is created, and the execution instruction sequence of the control scheme is parsed to obtain the multi-region collaborative adjustment instruction. The collaborative adjustment instruction will be used in the actual control system to ensure that the temperature and humidity of each storage area are within the optimal range, thereby realizing intelligent constant temperature and humidity control.

[0089] In one example, the weighted shared features are respectively input into the temperature prediction branch and the humidity prediction branch in the multi-task learning model of the long short-term memory network to perform temperature prediction and humidity prediction, and the temperature prediction value and humidity prediction value of each storage area are obtained, including:

[0090] The weighted shared features are separated to obtain a temperature feature vector and a humidity feature vector, and the temperature feature vector and the humidity feature vector are respectively dimensionally expanded to obtain a temperature feature matrix and a humidity feature matrix;

[0091] The temperature feature matrix is ​​input into the first fully connected layer of the temperature prediction branch for weight matrix multiplication to obtain the original temperature output, and the ReLU activation function is performed on the original temperature output to obtain the intermediate temperature feature;

[0092] Perform a dropout operation on the intermediate temperature feature to obtain the temperature feature after dropout, and input the temperature feature after dropout into the second fully connected layer of the temperature prediction branch for weight matrix multiplication to obtain the temperature prediction value of each storage area;

[0093] The humidity feature matrix is ​​input into the first fully connected layer of the humidity prediction branch for weight matrix multiplication to obtain the original humidity output, and the ReLU activation function is performed on the original humidity output to obtain the intermediate humidity feature;

[0094] The dropout operation is performed on the intermediate humidity features to obtain the humidity features after dropout, and the humidity features after dropout are input into the second fully connected layer of the humidity prediction branch for weight matrix multiplication to obtain the humidity prediction value of each storage area.

[0095] In this example, the weighted shared features are separated to obtain the temperature feature vector and the humidity feature vector. The shared features come from the multi-task learning model of the long short-term memory network and contain comprehensive information from different storage areas. Through feature separation, the features related to temperature and humidity are extracted separately to ensure that the change patterns of the two can be accurately captured in subsequent predictions. After feature separation, the temperature feature vector and the humidity feature vector are dimensionally expanded respectively. Dimension expansion refers to converting a one-dimensional feature vector into a two-dimensional feature matrix, increasing the expressive power of the feature and improving the prediction accuracy of the model. For example, if the temperature feature vector is , the temperature feature matrix after dimension expansion can be expressed as , similarly, the humidity feature vector The expanded humidity feature matrix is: . The temperature feature matrix Input to the first fully connected layer of the temperature prediction branch for weight matrix multiplication. The calculation formula of the fully connected layer is:

[0096] ;

[0097] in, is the raw output of temperature, is the weight matrix of the fully connected layer, is the bias vector. Through operation, the temperature feature matrix is ​​converted into a new feature representation. Perform ReLU activation function operation. The formula of ReLU activation function is:

[0098] ;

[0099] Through the ReLU activation function, nonlinear factors are introduced to improve the expression ability and prediction accuracy of the model. The output after ReLU activation is the intermediate temperature feature. In order to prevent the model from overfitting, the intermediate temperature feature is dropout. Dropout refers to randomly discarding some neurons during the training process to reduce the complexity of the model and the risk of overfitting. The temperature feature after dropout will be input into the second fully connected layer of the temperature prediction branch, and the weight matrix multiplication operation will be performed again to obtain the temperature prediction value of each storage area. This operation is similar to the first fully connected layer, and the feature is mapped to the output space through linear transformation. Similarly, the humidity feature matrix Input to the first fully connected layer of the humidity prediction branch, perform weight matrix multiplication, and obtain the original humidity output The original humidity output is nonlinearly transformed through the ReLU activation function to obtain the intermediate humidity feature. The humidity intermediate feature is dropout operated to obtain the humidity feature after dropout. The humidity feature after dropout is input into the second fully connected layer of the humidity prediction branch, and the weight matrix multiplication operation is performed to obtain the humidity prediction value of each storage area.

[0100] In one example, the optimized control strategy is input into the policy network, and through a three-layer fully connected network, each layer is followed by a BatchNorm and ReLU activation function to obtain the control parameters of the continuous action space, including:

[0101] The optimized control strategy is encoded to obtain a strategy vector, and the strategy vector is dimensionally expanded and normalized to obtain a control feature matrix;

[0102] The control feature matrix is ​​input into the first fully connected layer of the policy network for weight matrix multiplication to obtain the first layer original output, and the first layer original output is processed by BatchNorm to obtain standardized intermediate features;

[0103] Perform ReLU activation function operation on the standardized intermediate features to obtain the features after activation of the first layer, and perform dropout operation on the features after activation of the first layer to obtain the final output of the first layer;

[0104] The final output of the first layer is input into the second fully connected layer of the strategy network for weight matrix multiplication to obtain the original output of the second layer, and the original output of the second layer is processed by BatchNorm to obtain the standardized features of the second layer;

[0105] Perform ReLU activation function operation on the second layer standardized features to obtain the second layer activated features, and perform residual connection on the second layer activated features to obtain the second layer enhanced features;

[0106] Perform a dropout operation on the enhanced features of the second layer to obtain the final output of the second layer, and input the final output of the second layer into the third fully connected layer of the strategy network for weight matrix multiplication to obtain the original output of the third layer;

[0107] The original output of the third layer is processed by BatchNorm to obtain the standardized features of the third layer, and the ReLU activation function is performed on the standardized features of the third layer to obtain the features after activation of the third layer;

[0108] The features after activation in the third layer are globally averaged pooled to obtain feature vectors, which are then input into the output layer for full connection operation to obtain the continuous action space control parameters.

[0109] In this example, the optimized control policy is converted into a form suitable for model input. The control policy is converted into a policy vector through encoding. The discrete or continuous values ​​of each policy are mapped to a high-dimensional vector space for subsequent neural network processing. The policy vector is dimensionally expanded and normalized to obtain a control feature matrix. Dimensionality expansion refers to converting a one-dimensional policy vector into a higher-dimensional control feature matrix, which helps to increase the expressiveness of the features. Normalization ensures that all features are on the same scale. The processed control feature matrix is ​​input into the first fully connected layer of the policy network for weight matrix multiplication, and the input features are mapped to a new representation space through linear transformation. The formula is:

[0110] ;

[0111] in, is the original output of the first layer, is the weight matrix, is the input feature matrix, is the bias vector. The original output of the first layer is processed by BatchNorm and the original output is standardized to speed up training and improve the stability of the model. The formula for standardization is:

[0112] ;

[0113] in, is the normalized output, It's expectation. is the variance, is a small constant used to prevent zero division errors. The standardized intermediate features will be activated by the ReLU activation function. The formula of the ReLU activation function is:

[0114] ;

[0115] Through the ReLU activation function, nonlinear factors are introduced to improve the expressiveness of the model. After obtaining the features after the activation of the first layer, in order to prevent overfitting, a dropout operation is performed on it. This operation randomly discards some neurons. Usually a certain dropout rate, such as 0.5, is used to enhance the generalization ability of the model. The final output of the first layer will be input into the second fully connected layer of the policy network for weight matrix multiplication. The formula is similar to the first layer:

[0116] ;

[0117] in, is the original output of the second layer, is the weight matrix, is the final output of the first layer, is the bias vector. The second layer original output is processed by BatchNorm to obtain the second layer standardized features. The second layer standardized features are operated by ReLU activation function to obtain the second layer activated features. In order to improve the performance of the model, the second layer activated features are residually connected, and the input features are added to the output features to obtain the second layer enhanced features. The formula of residual connection is:

[0118] ;

[0119] in, is an enhanced feature, is the feature after the second layer is activated, is the input feature of the first layer. Residual connection can effectively alleviate the gradient vanishing problem in deep networks and enhance the expressiveness of features. Dropout operation is performed on the second layer enhanced features to obtain the final output of the second layer. The final output of the second layer is input to the third fully connected layer of the policy network for weight matrix multiplication. The formula is similar to that of the first two layers:

[0120] ;

[0121] in, is the original output of the third layer, is the weight matrix, is the final output of the second layer, is the bias vector. The original output of the third layer is processed by BatchNorm to obtain the standardized features of the third layer, and the features after activation of the third layer are obtained by ReLU activation function operation. In order to further extract features, the features after activation of the third layer are globally averaged pooled, and the formula is:

[0122] ;

[0123] in, is the feature vector after global average pooling, is the feature after activation of the third layer, is the number of features. Global average pooling reduces each feature map to a single value, thereby reducing computational complexity and overfitting risk. The feature vector after global average pooling is input to the output layer and fully connected to obtain the control parameters of the continuous action space. The formula for the fully connected operation is:

[0124] ;

[0125] in, is the control parameter of the continuous action space, is the weight matrix of the output layer, is the feature vector after global average pooling, is the bias vector.

[0126] In one example, executing a multi-region coordinated adjustment instruction and obtaining a second temperature and humidity signal data set of each storage region, and calculating a storage environment distribution feature of the second temperature and humidity signal data set include:

[0127] Execute the multi-region coordinated adjustment instruction and obtain the second temperature and humidity signal data set of each storage area;

[0128] Performing feature mapping on the second temperature and humidity signal data set of each storage area to obtain an environmental state cluster, and calculating the inter-cluster and intra-cluster distances according to the environmental state cluster to obtain an environmental state distribution index;

[0129] According to the environmental state distribution index, a fuzzy membership matrix is ​​constructed, and the entropy value of the environmental state is calculated according to the fuzzy membership matrix to obtain the environmental complexity index;

[0130] Perform time series analysis on environmental complexity indicators to obtain complexity change trends, and perform eigenvalue decomposition based on complexity change trends to obtain steady-state distribution vectors;

[0131] The importance weight of each environmental state is calculated according to the steady-state distribution vector to obtain a weighted environmental state representation, and a two-dimensional environmental state topology map is generated according to the weighted environmental state representation;

[0132] The topological structure characteristics are calculated according to the two-dimensional environmental state topological map to obtain the environmental spatial distribution characteristics, and the environmental spatial distribution characteristics are decomposed to obtain the core tensor and factor matrix;

[0133] The storage environment distribution characteristics are obtained based on the reconstruction of the core tensor and factor matrix.

[0134] In this example, multi-zone collaborative adjustment instructions are executed in the storage environment. These instructions will adjust the temperature and humidity of each storage area according to real-time data and preset strategies. After the instructions are executed, the second temperature and humidity signal data set of each storage area is obtained through the sensor system. These data sets record the changes in temperature and humidity in each area after the adjustment instructions are executed. Feature mapping is performed on the second temperature and humidity signal data set of each storage area to convert the raw data into environmental state clusters. The feature mapping process can use clustering algorithms, such as K-means or DBSCAN, to form clusters representing different environmental states by assigning similar temperature and humidity data to the same cluster. Assuming that the K-means clustering algorithm is used and the number of clusters is K, then each data point Assigned to cluster ,satisfy:

[0135] ;

[0136] in, It is a cluster The centroid of . In this way, the environmental state clusters are obtained. According to the environmental state clusters, the inter-cluster and intra-cluster distances are calculated to obtain the environmental state distribution index. The inter-cluster distance usually uses the Euclidean distance or other measurement methods to indicate the similarity between different clusters, while the intra-cluster distance indicates the closeness between data points in the same cluster. The specific calculation formula is as follows:

[0137] ;

[0138] in, represents the distance between clusters, represents the intra-cluster distance, and are different clusters, is a cluster, and is the data point in the cluster, is the centroid of the cluster. According to the calculated environmental state distribution index, a fuzzy membership matrix is ​​constructed. The fuzzy membership matrix represents the membership of each data point to a cluster, with a value between 0 and 1, and is usually calculated using the fuzzy C-means algorithm. Elements Represents data points Belongs to cluster The membership degree is calculated as follows:

[0139] ;

[0140] in, is the fuzzy coefficient, which usually takes a value between 1.5 and 3. According to the fuzzy membership matrix , calculate the entropy value of the environment state and obtain the environmental complexity index. The entropy value calculation formula is:

[0141] ;

[0142] in, is the number of data points, is the number of clusters. The environmental complexity index reflects the diversity and complexity of the environmental state through the entropy value. The environmental complexity index is analyzed in time series to obtain the complexity change trend. Time series analysis usually uses an autoregressive moving average model or other time series models to identify the change pattern of the complexity index over time. According to the complexity change trend, the eigenvalue decomposition is performed to obtain the steady-state distribution vector. The steady-state distribution vector represents the distribution characteristics of the environmental state under long-term stability. According to the steady-state distribution vector, the importance weight of each environmental state is calculated to obtain the weighted environmental state representation. The weight calculation can be obtained by linear or nonlinear transformation based on the size of the steady-state distribution vector. The weighted environmental state representation integrates the importance of different environmental states to generate a two-dimensional environmental state topology map. The two-dimensional environmental state topology map visualizes the weighted environmental state representation through a graph algorithm, such as force-directed layout, and shows the relationship between different environmental states. According to the two-dimensional environmental state topology map, the topological structure features are calculated to obtain the environmental spatial distribution features. The topological structure features can use indicators such as degree, clustering coefficient, and path length in graph theory to reflect the spatial distribution law of the environmental state. The environmental spatial distribution features are decomposed to obtain the core tensor and factor matrix. The core tensor and factor matrix are used to extract the main components and structures of the environment space distribution characteristics through tensor decomposition methods, such as Tucker decomposition or CP decomposition. According to the core tensor and factor matrix, the storage environment distribution characteristics are reconstructed.

[0143] In one example, the first collaborative control scheme is optimized according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and the second temperature and humidity signal data set is filtered and reconstructed to obtain a reconstructed temperature and humidity signal data set, including:

[0144] Performing feature fusion on the storage environment distribution feature and the first collaborative control scheme to obtain a fused feature vector, and performing nonlinear mapping on the fused feature vector to obtain a high-dimensional representation;

[0145] Perform denoising on the high-dimensional representation to obtain denoised features, and build a probabilistic control model based on the denoised features;

[0146] Perform tree search on the probabilistic control model to generate multiple candidate control sequences, and perform local search on the multiple candidate control sequences to obtain the local optimal solution;

[0147] Constructing a mapping relationship between control parameters and objective functions according to the local optimal solution, and generating a second collaborative control scheme according to the mapping relationship;

[0148] The second temperature and humidity signal data set is subjected to wavelet transformation to obtain multi-scale decomposition coefficients, and the multi-scale decomposition coefficients are subjected to threshold processing and inverse wavelet transformation to obtain a reconstructed temperature and humidity signal data set.

[0149] In this example, the storage environment distribution characteristics and the first collaborative control scheme are fused to obtain a fused feature vector. The storage environment distribution characteristics may include the spatiotemporal distribution of temperature and humidity, the changing trend of environmental conditions, etc., while the first collaborative control scheme involves a specific adjustment strategy for the storage environment. The two types of data are fused together to form a fused feature vector by splicing or adding. The fused feature vector is nonlinearly mapped to obtain a high-dimensional representation. Through a neural network, such as a multilayer perceptron, assuming that the fused feature vector is , then the nonlinear mapping process can be expressed as:

[0150] ;

[0151] in, is a high-dimensional representation, is the weight matrix, is the bias vector, It is a nonlinear activation function, such as ReLU. Through nonlinear mapping, the fused feature vector is converted into a higher-dimensional representation to capture more complex feature relationships. The high-dimensional representation is denoised to obtain denoised features. An autoencoder or variational autoencoder is used to remove the noise components in the high-dimensional representation and retain its main features. The encoder part compresses the high-dimensional representation into a low-dimensional latent space, and the decoder part reconstructs the original data from the latent space to remove the noise. The denoising formula is:

[0152] ;

[0153] in, is a low-dimensional latent representation, is a high-dimensional representation after denoising. Based on the denoising features, a probabilistic control model is constructed. The probabilistic control model can be implemented using a Markov decision process or a partially observable Markov decision process. By modeling the probability distribution of the control strategy, the effects of different strategies under different states are described. The core of the probabilistic control model is the state transition probability and the reward function. The state transition probability represents the probability of transferring to the next state after executing a certain control strategy in a certain state, and the reward function represents the benefit of executing a certain strategy. Perform a tree search on the probabilistic control model to generate multiple candidate control sequences. Tree search usually uses Monte Carlo tree search or heuristic search algorithm to generate a set of candidate control sequences by simulating and evaluating different control strategies. Assuming that the nodes of the search tree represent the state of the storage environment and the edges represent the control strategies, the tree search process can be expressed as:

[0154] ;

[0155] in, and are the current and next states, respectively. is the control strategy, is the state transfer function. Perform local search on the generated multiple candidate control sequences to find the local optimal solution. Local search improves the effect of the control strategy by making subtle adjustments and optimizations based on the candidate sequence. Use local search algorithms, such as simulated annealing or genetic algorithms, to find a better control strategy by gradually adjusting the control parameters. According to the local optimal solution, construct a mapping relationship between the control parameters and the objective function, and generate a second collaborative control scheme based on the mapping relationship. The mapping relationship represents how the control parameters affect the state of the storage environment and the value of the objective function. The objective function can be the stability of the storage environment. The fluctuation range of temperature and humidity, etc. The mapping relationship can be learned through a regression model or a deep learning model, and the formula is:

[0156] ;

[0157] in, is the objective function, is the reward function, s is the state, is the control strategy. By learning the mapping relationship between the control parameters and the objective function, an optimized second collaborative control scheme is generated. The second temperature and humidity signal data set is subjected to wavelet transform to obtain multi-scale decomposition coefficients. Wavelet transform extracts details and trends in the signal by decomposing it into components of different scales. The formula for wavelet transform is:

[0158] ;

[0159] in, are the wavelet transform coefficients, It's a signal. is the wavelet function, and are scale and translation parameters respectively. Threshold processing is performed on the multi-scale decomposition coefficients to remove noise components, and the signal is reconstructed by inverse wavelet transform to obtain a reconstructed temperature and humidity signal data set. Threshold processing can use soft threshold or hard threshold method. By setting a threshold, the coefficients below the threshold are set to zero or reduced. The formula for inverse wavelet transform is:

[0160] ;

[0161] in, is a constant. Through inverse wavelet transform, the decomposed coefficients are recombined to reconstruct the original signal.

[0162] In one example, the state transition probability modeling is performed on the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain the storage environment prediction result and abnormal warning index, including:

[0163] The reconstructed temperature and humidity signal data set is segmented into time windows to obtain time series data segments, and the time series data segments are dynamically time-warped to obtain aligned data sequences;

[0164] Performing state recognition on the aligned data sequence to obtain an initial state sequence, and constructing a state-action pair according to the initial state sequence and the second collaborative control scheme;

[0165] Calculate a transition probability matrix and an emission probability matrix according to the state-action pairs, and generate dependencies between states according to the transition probability matrix and the emission probability matrix;

[0166] Perform sequence importance sampling on the dependencies between states to obtain the state prediction distribution, and calculate the storage environment prediction result based on the state prediction distribution;

[0167] Perform deviation analysis on storage environment prediction results, calculate error distribution characteristics, and set multi-level warning thresholds based on error distribution characteristics;

[0168] The trigger probability is calculated according to the multi-level warning threshold to obtain the preliminary warning index, and the abnormal pattern recognition is performed on the preliminary warning index to obtain the abnormal warning index.

[0169] In this example, the continuous temperature and humidity data are divided into time series data segments of fixed length. Assume that the time window length is , then each time series data segment can be expressed as ,in It's time Temperature and humidity data points. Dynamic time warping is performed on the time series data segments to obtain the aligned data series. The purpose of dynamic time warping is to perform nonlinear transformation on the time series so that similar time series are aligned on the time axis and the error caused by time offset in the time series data is eliminated. The formula for dynamic time warping is:

[0170] ;

[0171] in, are two sequences x and y at the point and The cumulative distance of is the Euclidean distance between two points. By minimizing the cumulative distance, dynamic time warping can align two time series to obtain an aligned data sequence. The aligned data sequence is subjected to state recognition to obtain the initial state sequence. State recognition can be achieved by clustering algorithms such as Kmeans or Hidden Markov Model. Assume that K-means clustering is used to divide the aligned data sequence into Each data point is assigned to a certain state to form an initial state sequence ,in It is The state at each time point. According to the initial state sequence and the second coordinated control scheme, a state-action pair is constructed. The state-action pair represents the control action taken in a specific state and its result. Assume that the control action is , then the state-action pair can be expressed as . Based on the state-action pairs, calculate the transition probability matrix and the emission probability matrix. Transition probability matrix Represents the probability of transferring to the next state in a certain state, and the emission probability matrix It represents the probability of observing a certain temperature and humidity data in a certain state. The formula of the transition probability matrix is:

[0172] ;

[0173] in, From the state Transfer to state The probability of the emission probability matrix is:

[0174] ;

[0175] in, is in state Temperature and humidity data were observed According to the probability matrix, the dependencies between states are generated to describe the transition rules between states. Sequential importance sampling is performed on the dependencies between states to obtain the state prediction distribution. Sequential importance sampling is a Monte Carlo method that estimates the distribution of future states by sampling the dependencies between states. Assume that the state prediction distribution is , the state prediction distribution can be obtained by repeated sampling. According to the state prediction distribution, the storage environment prediction result is calculated. The prediction distribution of the current state is combined with the observed data through the Bayesian update formula to update the prediction distribution of the future state. The storage environment prediction result represents the changing trend of the storage environment in the future. Deviation analysis is performed on the storage environment prediction results to calculate the error distribution characteristics. Deviation analysis identifies the error characteristics of the prediction model by comparing the difference between the prediction results and the actual observations. The error distribution characteristics can be calculated by statistical analysis methods, such as mean square error or mean absolute error. The formula for mean square error is:

[0176] ;

[0177] in, is the actual observed value, is the predicted value, is the number of data points. By calculating the error distribution characteristics, the deviation of the prediction model is identified and the model parameters are adjusted. According to the error distribution characteristics, a multi-level warning threshold is set. The multi-level warning threshold is set according to the size and distribution of the error, and can be divided into three levels of warning: low, medium and high. Each level of warning corresponds to a different error range. When the prediction error exceeds a certain threshold, the corresponding warning signal is triggered. According to the multi-level warning threshold, the trigger probability is calculated to obtain the preliminary warning index. The trigger probability indicates the possibility of the warning signal being triggered within a specific error range. The preliminary warning index is obtained by calculating the weighted sum of the trigger probabilities of warnings at all levels. The formula is:

[0178] ;

[0179] in, It is the preliminary warning index. It is The probability of triggering a level 1 warning, It is The weight of the warning level, is the number of warning levels. The abnormal pattern recognition is performed on the preliminary warning index to obtain the abnormal warning index. The abnormal pattern recognition uses machine learning algorithms, such as support vector machines or neural networks, to identify abnormal patterns in the preliminary warning index and generate an abnormal warning index.

[0180] Reference Figure 2 This embodiment provides an intelligent constant temperature and humidity storage system for medical supplies, including:

[0181] Analysis module 1, used for performing multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage scheme, which includes multiple storage areas;

[0182] Establishing module 2, for establishing a temperature and humidity dynamic model for each storage area and performing error prediction and parameter adjustment, and obtaining an adaptive temperature and humidity control instruction for each storage area;

[0183] Processing module 3, used to execute the adaptive temperature and humidity control instruction and obtain the first temperature and humidity signal data set of each storage area, input the first temperature and humidity signal data set into the multi-task learning model of the long short-term memory network for multi-area collaborative processing, and obtain the first collaborative control scheme and the multi-area collaborative adjustment instruction;

[0184] A calculation module 4 is used to execute the multi-region coordinated adjustment instruction and obtain the second temperature and humidity signal data set of each storage area, and calculate the storage environment distribution characteristics of the second temperature and humidity signal data set;

[0185] An optimization module 5 is used to optimize the first collaborative control scheme according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and filter and reconstruct the second temperature and humidity signal data set to obtain a reconstructed temperature and humidity signal data set;

[0186] The early warning module 6 is used to perform state transition probability modeling on the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain the storage environment prediction result and the abnormal early warning index.

[0187] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0188] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0189] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0190] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0191] In summary, an intelligent constant temperature and humidity storage method for medical supplies is proposed in an embodiment of the present invention.

[0192] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0193] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0194] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent constant temperature and humidity storage method for medical supplies, characterized in that: The following steps are involved: Performing multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage scheme, wherein the partition storage scheme includes multiple storage areas; Establish a temperature and humidity dynamic model for each storage area respectively, perform error prediction and parameter adjustment, and obtain an adaptive temperature and humidity control instruction for each storage area; specifically include: perform time series analysis on the historical temperature and humidity data of each storage area to obtain periodic characteristics, and perform seasonal decomposition on the data according to the periodic characteristics to obtain trend items, seasonal items and random items; fit the trend items to obtain a trend prediction model, and perform Fourier transform on the seasonal items to obtain a seasonal periodic function; construct an initial dynamic model for each storage area according to the trend prediction model and the seasonal periodic function, and perform sensitivity analysis on the initial dynamic model to obtain a key parameter subset; perform parameter tuning on the key parameter subset to obtain a temperature and humidity dynamic model, and simulate the temperature and humidity dynamic model to construct a state estimator; calculate the filtered state estimation value through the state estimator, and perform error calculation on the filtered state estimation value and the dynamic model prediction value to obtain a prediction error sequence, and perform autocorrelation analysis on the prediction error sequence to obtain error pattern characteristics; calculate the initial PID parameter set of the adaptive PID controller according to the error pattern characteristics, and discretize the initial PID parameter set to obtain an adaptive temperature and humidity control instruction; Execute the adaptive temperature and humidity control instruction and obtain the first temperature and humidity signal data set of each storage area, input the first temperature and humidity signal data set into the multi-task learning model of the long short-term memory network for multi-region collaborative processing, and obtain a first collaborative control scheme and a multi-region collaborative adjustment instruction; specifically include: execute the adaptive temperature and humidity control instruction and obtain the first temperature and humidity signal data set of each storage area; divide the first temperature and humidity signal data set into time windows and normalize them to obtain standardized time series samples, and map the standardized time series samples to a high-dimensional space through an embedding layer to obtain an input feature vector; input the input feature vector into the first layer of the long short-term memory network in the multi-task learning model of the long short-term memory network, update the cell state and hidden state through the calculation of the forget gate, input gate and output gate, and obtain the output sequence of the first layer; perform a dropout operation on the output sequence of the first layer to obtain a feature sequence after dropout, and input the feature sequence after dropout into the second layer of the long short-term memory network, update the cell state and hidden state through the calculation of the forget gate, input gate and output gate, and obtain the output sequence of the second layer; input the output sequence of the second layer into the multi-task learning model of the long short-term memory network The method comprises the following steps: the first step is to obtain a multi-region shared feature by using a fully connected layer and a ReLU activation function, and the second step is to obtain a weighted shared feature by applying an attention mechanism to the multi-region shared feature, calculating weight coefficients of different regions, and obtaining weighted shared features; the second step is to input the weighted shared features into the temperature prediction branch and the humidity prediction branch in the multi-task learning model of the long short-term memory network respectively to perform temperature prediction and humidity prediction, and obtain temperature prediction values ​​and humidity prediction values ​​of each storage area; the temperature prediction values ​​and humidity prediction values ​​are denormalized to obtain a prediction result of the actual scale, and the prediction result of the actual scale is compared with a set threshold. A comparison is made between the rows to obtain the abnormal temperature and humidity detection result; a decision tree is constructed according to the abnormal temperature and humidity detection result, each leaf node corresponds to a control strategy, a preliminary control strategy set is obtained, and the preliminary control strategy set is optimized and searched to obtain the optimized control strategy; the optimized control strategy is input into the strategy network, and the control parameters of the continuous action space are obtained through a three-layer fully connected network, each layer is followed by a BatchNorm and a ReLU activation function; a first collaborative control scheme is created according to the control parameters of the continuous action space, and the execution instruction sequence of the first collaborative control scheme is parsed to obtain a multi-region collaborative adjustment instruction; Execute the multi-area coordinated adjustment instruction and obtain a second temperature and humidity signal data set of each storage area, and calculate the storage environment distribution characteristics of the second temperature and humidity signal data set; The first collaborative control scheme is optimized according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and the second temperature and humidity signal data set is filtered and reconstructed to obtain a reconstructed temperature and humidity signal data set; The state transition probability modeling is performed on the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain a storage environment prediction result and an abnormal warning index; specifically, the method includes: performing time window segmentation on the reconstructed temperature and humidity signal data set to obtain time series data segments, and dynamically time warping the time series data segments to obtain an aligned data sequence; performing state recognition on the aligned data sequence to obtain an initial state sequence, and constructing a state-action pair based on the initial state sequence and the second collaborative control scheme; calculating a transition probability matrix and a transmission probability matrix based on the state-action pairs, and generating dependencies between states based on the transition probability matrix and the transmission probability matrix; performing sequence importance sampling on the dependencies between the states to obtain a state prediction distribution, and calculating a storage environment prediction result based on the state prediction distribution; performing deviation analysis on the storage environment prediction result, calculating error distribution characteristics, and setting a multi-level warning threshold based on the error distribution characteristics; calculating a trigger probability based on the multi-level warning threshold to obtain a preliminary warning index, and performing abnormal pattern recognition on the preliminary warning index to obtain an abnormal warning index.

2. The intelligent constant temperature and humidity storage method for medical supplies according to claim 1, characterized in that: The multi-dimensional similarity calculation and cluster analysis are performed on the medical supplies storage requirement parameters to obtain a partition storage scheme, which includes multiple storage areas, including: Standardizing the medical supplies storage requirement parameters to obtain a standardized parameter set, and generating a parameter vector set according to the standardized parameter set; Performing cosine similarity calculation on the parameter vector set to obtain a similarity matrix, and constructing a medical supplies similarity graph based on the similarity matrix; Performing spectral clustering on the medical supplies similarity graph to obtain an initial clustering result, and calculating a silhouette coefficient based on the initial clustering result to obtain an optimal number of clusters; Performing cluster optimization on the initial clustering result according to the optimal clustering number to obtain an optimized clustering result, and performing density peak detection on the optimized clustering result to obtain a cluster center set; The medical supplies are assigned to nearest neighbors according to the cluster center set to obtain a preliminary partitioning scheme, and the boundaries of the preliminary partitioning scheme are adjusted to obtain a partitioning storage scheme, which includes multiple storage areas.

3. The intelligent constant temperature and humidity storage method for medical supplies according to claim 1, characterized in that: The weighted shared features are respectively input into the temperature prediction branch and the humidity prediction branch in the multi-task learning model of the long short-term memory network to perform temperature prediction and humidity prediction, and obtain the temperature prediction value and humidity prediction value of each storage area, including: Performing feature separation on the weighted shared features to obtain a temperature feature vector and a humidity feature vector, and performing dimension expansion on the temperature feature vector and the humidity feature vector to obtain a temperature feature matrix and a humidity feature matrix; Input the temperature feature matrix into the first fully connected layer of the temperature prediction branch to perform a weight matrix multiplication operation to obtain an original temperature output, and perform a ReLU activation function operation on the original temperature output to obtain an intermediate temperature feature; Performing a dropout operation on the intermediate temperature feature to obtain a temperature feature after dropout, and inputting the temperature feature after dropout into the second fully connected layer of the temperature prediction branch to perform a weight matrix multiplication operation to obtain a temperature prediction value of each storage area; Input the humidity feature matrix into the first fully connected layer of the humidity prediction branch to perform a weight matrix multiplication operation to obtain a humidity original output, and perform a ReLU activation function operation on the humidity original output to obtain a humidity intermediate feature; A dropout operation is performed on the intermediate humidity feature to obtain a humidity feature after dropout, and the humidity feature after dropout is input into the second fully connected layer of the humidity prediction branch to perform a weight matrix multiplication operation to obtain a humidity prediction value for each storage area.

4. The intelligent constant temperature and humidity storage method for medical supplies according to claim 1, characterized in that: The optimized control strategy is input into the strategy network, and through a three-layer fully connected network, each layer is followed by a BatchNorm and ReLU activation function to obtain the control parameters of the continuous action space, including: Encoding the optimized control strategy to obtain a strategy vector, and dimensionally expanding and normalizing the strategy vector to obtain a control feature matrix; Inputting the control feature matrix into the first fully connected layer of the policy network to perform weight matrix multiplication operation to obtain the first layer original output, and performing BatchNorm processing on the first layer original output to obtain standardized intermediate features; Performing a ReLU activation function operation on the standardized intermediate features to obtain features after activation of the first layer, and performing a dropout operation on the features after activation of the first layer to obtain a final output of the first layer; The final output of the first layer is input into the second fully connected layer of the policy network for weight matrix multiplication to obtain the original output of the second layer, and the original output of the second layer is subjected to BatchNorm processing to obtain the standardized features of the second layer; Performing a ReLU activation function operation on the second layer standardized features to obtain the second layer activated features, and performing a residual connection on the second layer activated features to obtain the second layer enhanced features; Performing a dropout operation on the second layer enhanced features to obtain the second layer final output, and inputting the second layer final output into the third fully connected layer of the strategy network for weight matrix multiplication to obtain the third layer original output; Performing BatchNorm processing on the original output of the third layer to obtain standardized features of the third layer, and performing ReLU activation function operation on the standardized features of the third layer to obtain features after activation of the third layer; The features after activation of the third layer are globally averaged pooled to obtain a feature vector, and the feature vector is input into the output layer for full connection operation to obtain continuous action space control parameters.

5. The intelligent constant temperature and humidity storage method for medical supplies according to claim 1, characterized in that: The executing the multi-area coordinated adjustment instruction and acquiring the second temperature and humidity signal data set of each storage area, and calculating the storage environment distribution characteristics of the second temperature and humidity signal data set, includes: Execute the multi-area coordinated adjustment instruction and obtain a second temperature and humidity signal data set of each storage area; Performing feature mapping on the second temperature and humidity signal data set of each storage area to obtain an environmental state cluster, and calculating the inter-cluster and intra-cluster distances according to the environmental state cluster to obtain an environmental state distribution index; Constructing a fuzzy membership matrix according to the environmental state distribution index, and calculating the entropy value of the environmental state according to the fuzzy membership matrix to obtain an environmental complexity index; Performing time series analysis on the environmental complexity index to obtain a complexity change trend, and performing eigenvalue decomposition according to the complexity change trend to obtain a steady-state distribution vector; Calculating the importance weight of each environmental state according to the steady-state distribution vector to obtain a weighted environmental state representation, and generating a two-dimensional environmental state topology map according to the weighted environmental state representation; Calculating topological structure features according to the two-dimensional environment state topological map to obtain environment space distribution features, and decomposing the environment space distribution features to obtain a core tensor and a factor matrix; The storage environment distribution characteristics are obtained by reconstructing the core tensor and the factor matrix.

6. The intelligent constant temperature and humidity storage method for medical supplies according to claim 1, characterized in that: The first collaborative control scheme is optimized according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and the second temperature and humidity signal data set is filtered and reconstructed to obtain a reconstructed temperature and humidity signal data set, including: Performing feature fusion on the storage environment distribution feature and the first collaborative control scheme to obtain a fused feature vector, and performing nonlinear mapping on the fused feature vector to obtain a high-dimensional representation; Performing denoising processing on the high-dimensional representation to obtain denoised features, and constructing a probabilistic control model based on the denoised features; Performing a tree search on the probability control model to generate multiple candidate control sequences, and performing a local search on the multiple candidate control sequences to obtain a local optimal solution; Constructing a mapping relationship between a control parameter and an objective function according to the local optimal solution, and generating a second collaborative control scheme according to the mapping relationship; The second temperature and humidity signal data set is subjected to wavelet transformation to obtain multi-scale decomposition coefficients, and the multi-scale decomposition coefficients are subjected to threshold processing and inverse wavelet transformation to obtain a reconstructed temperature and humidity signal data set.

7. An intelligent constant temperature and humidity storage system for medical supplies, characterized in that: A system for performing an intelligent constant temperature and humidity storage method for medical supplies according to any one of claims 1 to 6, the system comprising: An analysis module, used for performing multi-dimensional similarity calculation and cluster analysis on the storage requirement parameters of medical supplies to obtain a partition storage scheme, wherein the partition storage scheme includes multiple storage areas; Establishing a module for establishing a dynamic temperature and humidity model for each storage area and performing error prediction and parameter adjustment to obtain adaptive temperature and humidity control instructions for each storage area; a processing module, configured to execute the adaptive temperature and humidity control instruction and obtain a first temperature and humidity signal data set of each storage area, input the first temperature and humidity signal data set into a multi-task learning model of a long short-term memory network for multi-area collaborative processing, and obtain a first collaborative control scheme and a multi-area collaborative adjustment instruction; a calculation module, configured to execute the multi-region coordinated adjustment instruction and obtain a second temperature and humidity signal data set of each storage region, and calculate a storage environment distribution feature of the second temperature and humidity signal data set; an optimization module, configured to optimize the first collaborative control scheme according to the storage environment distribution characteristics to obtain a second collaborative control scheme, and filter and reconstruct the second temperature and humidity signal data set to obtain a reconstructed temperature and humidity signal data set; The early warning module is used to perform state transition probability modeling on the reconstructed temperature and humidity signal data set and the second collaborative control scheme to obtain a storage environment prediction result and an abnormal early warning index.

Images