Clinical medicine warehouse management method, device and equipment and storage medium

Abstract

The invention relates to the technical field of warehouse management, and discloses a clinical drug warehouse management method, device and equipment and a storage medium, and the method comprises the steps: carrying out the data collection and standardization processing of clinical drugs, obtaining an n-dimensional drug feature vector set and a feature weight matrix, and carrying out the adjacency perception calculation, obtaining a drug membership decision sequence of k storage categories; performing selective masking and sequence prediction on a target storage category membership value in the drug membership decision sequence to obtain a drug storage category prediction result; decomposing a drug storage category prediction result according to a storage condition dimension, an inventory level dimension and a timeliness dimension to obtain a multi-dimensional feature representation set; and performing single-dimensional reconstruction and cross-dimensional fusion on the multi-dimensional feature representation set to obtain comprehensive drug storage representation, and performing storage position allocation and inventory parameter calculation to obtain a clinical drug storage management strategy. The intelligent level of medicine storage management is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of warehouse management, and in particular to a clinical drug warehouse management method, device, equipment and storage medium. Background Art

[0002] Clinical drug storage management is an important part of the operation of medical institutions, and its management quality directly affects the safety of patients' medication and the quality of medical care. With the expansion of medical scale and the increase in drug categories, traditional manual management methods have been unable to meet the needs of refined management, especially in drug storage classification, inventory allocation and environmental monitoring.

[0003] At present, the main challenges facing clinical drug warehouse management are the lack of unified storage classification standards and vague management rules. Due to the differences in storage conditions, inventory levels, and timeliness of different drugs, and the complex interactions between these characteristics, it is difficult to establish an accurate storage classification system. At the same time, traditional warehouse management methods often adopt a single-dimensional classification standard, which cannot effectively deal with the fusion of multi-dimensional features, making the allocation of drug storage locations and inventory parameter settings lack scientific basis. In addition, the existing clinical drug warehouse management system generally has problems such as low data utilization efficiency, insufficient feature extraction, and low prediction accuracy. In particular, when dealing with the classification of drugs with similar storage conditions, it is difficult to accurately identify subtle differences, resulting in low storage space utilization, and unable to respond to inventory changes and environmental fluctuations in a timely manner, increasing drug storage risks and management costs. Summary of the invention

[0004] The present invention provides a clinical drug storage management method, device, equipment and storage medium, and the present invention improves the intelligence level of drug storage management.

[0005] In a first aspect, the present invention provides a clinical drug storage management method, the clinical drug storage management method comprising:

[0006] Data collection and standardization of clinical drugs are performed to obtain an n-dimensional drug feature vector set and feature weight matrix, and adjacency perception calculation is performed to obtain a drug membership decision sequence of k storage categories;

[0007] Selectively masking and sequence predicting the target storage category membership values ​​in the drug membership decision sequence to obtain a drug storage category prediction result;

[0008] Decomposing the drug storage category prediction result according to the storage condition dimension, inventory level dimension and timeliness dimension to obtain a multi-dimensional feature representation set;

[0009] The multi-dimensional feature representation set is reconstructed in a single dimension and fused across dimensions to obtain a comprehensive drug storage representation, and storage location allocation and inventory parameter calculation are performed to obtain a clinical drug storage management strategy.

[0010] In a second aspect, the present invention provides a clinical drug storage management device, the clinical drug storage management device comprising:

[0011] The data collection module is used to collect and standardize the data of clinical drugs, obtain an n-dimensional drug feature vector set and a feature weight matrix, and perform adjacency perception calculation to obtain a drug membership decision sequence of k storage categories;

[0012] A sequence prediction module, used for selectively masking and sequence prediction of the target storage category membership values ​​in the drug membership decision sequence to obtain a drug storage category prediction result;

[0013] A decomposition module, used for decomposing the drug storage category prediction result according to the storage condition dimension, the inventory level dimension and the timeliness dimension to obtain a multi-dimensional feature representation set;

[0014] The fusion module is used to perform single-dimensional reconstruction and cross-dimensional fusion on the multi-dimensional feature representation set to obtain a comprehensive drug storage representation, and perform storage location allocation and inventory parameter calculation to obtain a clinical drug storage management strategy.

[0015] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned clinical drug warehouse management method.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on a computer, the computer executes the above-mentioned clinical drug warehouse management method.

[0017] In the technical solution provided by the present invention, by introducing an adjacency-aware fuzzy label learning framework, multi-dimensional analysis and storage category prediction of drug features are performed, thereby improving the accuracy of drug classification and solving the problem of inconsistent storage classification standards; a selective masking and sequence prediction mechanism is adopted to construct a prediction model with adaptive characteristics, thereby reducing error propagation in the prediction process and improving the reliability of the prediction results; based on the three-dimensional feature decomposition space and the dimensional utility matrix, refined decomposition and feature extraction of storage conditions, inventory levels and timeliness dimensions are realized, thereby enhancing the expression ability of multi-dimensional features; through single-dimensional reconstruction and cross-dimensional fusion technology, a comprehensive representation method that takes into account both local features and global associations is established, thereby optimizing the integrity and accuracy of feature representation; by integrating spatial clustering analysis and environmental parameter monitoring, a dynamically adjusted storage management strategy is constructed, thereby improving storage space utilization and environmental control accuracy; based on time series prediction and scheduling optimization algorithms, a drug inventory early warning and replenishment mechanism is established, thereby reducing inventory management costs and improving the intelligence level of drug storage management. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 A schematic diagram of the steps of the clinical drug storage management method in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of a clinical drug storage management device in an embodiment of the present invention;

[0021] Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] Embodiments of the present invention provide a clinical drug storage management method, device, equipment and storage medium. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than the content illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the clinical drug storage management method in the embodiment of the present invention includes:

[0024] Step S1, collect and standardize data on clinical drugs to obtain an n-dimensional drug feature vector set and a feature weight matrix, and perform adjacency perception calculation to obtain a drug membership decision sequence of k storage categories;

[0025] It is understandable that the execution subject of the present invention may be a clinical drug storage management device, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0026] Specifically, basic information of clinical drugs is collected by systematically collecting key attributes of drugs, including drug name, specification, dosage form, storage conditions, expiration date and production batch number. This information is summarized by electronic input to form the original data set. Natural language preprocessing is performed on the text fields in the original data set, and drug-related information is made more standardized and consistent by removing noise data, word segmentation and standardization operations. The text fields are converted into character encoding, such as converting character information into a digital format suitable for calculation to ensure that the text features can be quantified later. The processed text fields are aggregated into a standardized drug text feature set. The standardized drug text feature set is quantized and vectorized. The quantization process converts non-numerical information into a numerical form that can be used for mathematical operations, and converts these features into a numerical feature matrix through vectorization methods. The numerical feature matrix can characterize the multidimensional characteristics of drugs in a unified data format. The numerical feature matrix is ​​normalized to the maximum and minimum values, and each eigenvalue is scaled to the same magnitude so that it is distributed within a fixed range (such as between 0 and 1) to eliminate the influence of dimensional differences between features. In normalization, the distribution of eigenvalues ​​is adjusted through standardized calculation to conform to the normal distribution with a mean of 0 and a variance of 1, thereby improving the robustness of the model in dealing with differences between different features. The standardized feature vectors are subjected to correlation analysis to explore the dependencies between different features, and the core feature dimensions that have the greatest impact on storage management are screened out. Through correlation analysis and feature screening, n core feature dimensions are extracted from the original feature set, and the standardized feature vectors are reorganized based on these dimensions to obtain an n-dimensional drug feature vector set. The feature importance weight value of each feature dimension in the n-dimensional drug feature vector set is calculated based on the information gain algorithm. The information gain algorithm can quantify the contribution of each feature by measuring the degree of uncertainty reduction of the feature to the decision result, and obtain the initial weight vector, which contains the original importance weight of each feature. The initial weight vector is normalized to standardize all weight values ​​to a fixed range (such as a total weight of 1). On the basis of normalization, nonlinear mapping transformation is applied to adjust the distribution characteristics of the weight value, so that the weight matrix is ​​more in line with the actual needs of the complex system, and a feature weight matrix is ​​formed. Combine the n-dimensional drug feature vector set with the feature weight matrix, and classify them based on the similarity and weight between drug features through adjacency-aware calculation. The core of adjacency-aware calculation is to build a distance measurement model, use weighted Euclidean distance or cosine similarity method to calculate the similarity between drug features, and use clustering algorithm (such as K-Means or hierarchical clustering) to divide drugs into k storage categories based on the similarity results. In the clustering process, a membership decision sequence is generated for each drug, in which each element represents the degree to which the drug belongs to a storage category, completing the preliminary classification of drugs.

[0027] The n-dimensional drug feature vector set is preliminarily grouped by k-means clustering analysis. The core of k-means clustering is to randomly initialize k storage category core vectors from the drug feature vector set with the goal of minimizing the intra-group square error. According to the Euclidean distance between each drug feature vector and the core vector, the drug is assigned to the category with the closest distance, and the core vector of each category is recalculated based on the assignment result. This iterative process continues until the category core vector is stable and no longer changes, forming an initial storage category core vector set. The adjacency relationship of the initial storage category core vector set is calculated to construct the storage category adjacency weight function. The adjacency relationship is calculated based on the distance between the core vectors, and the weighted cosine similarity or inverse distance function is used to measure the mutual proximity between the categories, and the adjacency weight value between each pair of categories is obtained. The Euclidean distance between each feature vector in the n-dimensional drug feature vector set and the storage category core vector set is calculated to obtain the basic distance matrix, which reflects the preliminary matching degree between the drug and each storage category; in order to improve the classification accuracy, the basic distance matrix is ​​weighted and adjusted according to the feature weight matrix. The feature weight matrix contains the importance weight of each feature. Through weighting operation, the features that have a greater impact on the storage category decision have a higher weight in the distance calculation, and the weighted distance matrix is ​​obtained. The weighted distance matrix is ​​adjusted based on the storage category adjacency weight function. The adjustment process combines the distance between the drug and its target storage category and the relationship between the storage category and the adjacent category to recalculate the membership value of the drug to each storage category. The mutual influence between categories is comprehensively considered, so that the membership calculation is not only based on the characteristics of the individual category, but also reflects the dynamic association between categories, and the initial membership matrix is ​​obtained. The initial membership matrix is ​​optimized based on the sequence joint distribution loss function. By introducing the sequence joint distribution loss function, the drug's membership value is combined with the distribution characteristics of the category, and the membership contribution of each category in the global scope is calculated. The sequence joint distribution loss function aims to minimize the difference in membership between categories. Through multiple iterations of optimization and adjustment of the membership matrix, the final optimized membership matrix can more accurately reflect the belonging relationship of the drug in each storage category. The optimized membership matrix is ​​rearranged according to the drug index order to ensure that the generated membership decision sequence can directly correspond to the actual drug storage demand. In this sequence, each drug corresponds to a membership value set containing k elements, and each value represents the degree to which the drug belongs to a specific storage category.

[0028] Step S2, selectively masking and sequence predicting the target storage category membership values ​​in the drug membership decision sequence to obtain a drug storage category prediction result;

[0029] Specifically, random stratified sampling is performed on the drug membership decision sequence to obtain a training set and a validation set. By performing distribution statistics on the membership values ​​of different storage categories in the membership decision sequence, the distribution ratio of samples of each category in the training set and the validation set is ensured to be consistent with the original data, so as to avoid affecting the performance of the model due to category imbalance. The storage category membership values ​​in the training set are statistically analyzed to reflect the range characteristics of the membership values ​​of different storage categories. By analyzing the distribution of membership values ​​of each storage category, the membership threshold range is determined, which is used to describe the significant characteristics of membership values ​​in different categories. Based on the membership threshold range, all membership values ​​are hierarchically classified, and the membership values ​​are divided into a multi-level masking mark set, each level corresponding to a different degree of importance, reflecting the relative significance of the membership values ​​of each category in different contexts. A selective masking function is constructed based on the generated multi-level masking mark set to define the sequence masking rule. In the process of constructing the selective masking function, the masking strategy is designed by combining the context dependency and the membership distribution characteristics, so that part of the storage category membership values ​​can be effectively masked during the training process, while the rest is retained to provide sufficient context information. According to the generated sequence masking rules, the target storage category membership values ​​in the training set are masked and replaced to obtain the masked training sequence. The masked training sequence is input into the consistency prediction model for parameter optimization training. The consistency prediction model grasps the internal logic and distribution law of the storage category membership values ​​in the sequence by learning the correspondence between the complete sequence and the masked sequence. During the model training process, the parameters of the model are gradually optimized by minimizing the error between the predicted value and the true value while considering the context consistency constraints. After sufficient training, the consistency prediction model can reconstruct the partially masked sequence and accurately predict the membership value of the target storage category, forming a sequence reconstruction model with strong generalization ability. After the model training is completed, the drug membership decision sequence is masked according to the previously generated sequence masking rules to obtain the masked sequence to be predicted. The masked sequence to be predicted contains complete context information, but lacks some target storage category membership values, which are completed and calculated through the sequence reconstruction model. The sequence completion calculation takes the masked sequence to be predicted as input, reconstructs the target storage category membership values ​​through the membership distribution law and context dependency information learned by the model parameters, and generates a completed complete sequence. The reconstructed sequence is the drug storage category prediction result.

[0030] Step S3, decomposing the drug storage category prediction results according to the storage condition dimension, inventory level dimension and timeliness dimension to obtain a multi-dimensional feature representation set;

[0031] Specifically, a three-dimensional feature decomposition space including storage conditions, inventory levels and timeliness is constructed based on the prediction results of drug storage categories. The three-dimensional feature decomposition space is based on the drug prediction results and is mapped to the specific attributes of each feature dimension to obtain the storage condition feature matrix, inventory level feature matrix and timeliness feature matrix respectively. The temperature and humidity correlation analysis is performed on the storage condition feature matrix. By constructing a multivariate regression model or correlation analysis model for storage conditions, the key influencing factors of drugs under different storage conditions are analyzed. Based on this analysis result, a storage condition feature calculation function is generated, which can dynamically calculate the storage condition feature value according to the specific drug prediction results. After the drug storage category prediction results are input into the storage condition feature calculation function, the storage condition feature subvector obtained reflects the adaptability and demand of the drug under storage temperature, humidity and other conditions. At the same time, the inventory level feature matrix is ​​analyzed for inventory fluctuations to reveal the dynamic change law of drug inventory levels. The inventory level feature calculation is mainly realized through fluctuation analysis and demand forecasting models, such as using time series forecasting, seasonal adjustment analysis and other methods to calculate the inventory fluctuation trend and its influencing parameters. Based on these analysis results, an inventory level feature calculation function is constructed, and the drug storage category prediction results are input into it. The obtained inventory level feature subvector can reflect the key information such as the replenishment frequency, maximum inventory and reorder point required for drugs in inventory management. Time series analysis is performed on the timeliness feature matrix. By modeling and predicting the historical data of drug expiration dates and usage cycles, a timeliness feature calculation function is constructed. This function takes the drug storage category prediction results as input and calculates the timeliness feature subvector according to the change law of the time series. This subvector describes the update frequency, optimal use time limit and potential expiration risk of drugs in warehouse management. A dimension utility matrix is ​​constructed to describe the interaction between storage conditions, inventory levels and timeliness. The dimension utility matrix calculates the dimensional interaction coefficients of storage conditions, inventory levels and timeliness by analyzing the degree of influence of each dimension on drug storage and their interaction coefficients in specific applications. These interaction coefficients can quantify the synergistic effect between dimensions. According to the dimensional interaction coefficient, the weight distribution of the feature subvector is calculated to generate a dimension weight matrix. The dimension weight matrix quantifies the relative contribution of each dimension into a weight distribution by normalizing the interaction coefficient. The storage condition feature subvector, inventory level feature subvector, and timeliness feature subvector are weighted combined. The weighted combination process is based on the dimension weight matrix to ensure that the combination result fully considers the relative importance and interaction relationship of each dimension and generates a multi-dimensional feature representation set.

[0032] Step S4: reconstruct the multi-dimensional feature representation set in a single dimension and fuse it across dimensions to obtain a comprehensive drug storage representation, and perform storage location allocation and inventory parameter calculation to obtain a clinical drug storage management strategy.

[0033] Specifically, the storage condition feature subvector, inventory level feature subvector and timeliness feature subvector in the multi-dimensional feature representation set are reconstructed by dimensionality reduction. The principal component analysis or autoencoder model is selected to reduce data redundancy and extract the most representative principal component of each dimensional feature to obtain three groups of single-dimensional reconstruction vectors, which retain the core feature information of storage condition, inventory level and timeliness respectively. After completing the single-dimensional reconstruction, a dimension reconstruction optimization function is constructed to adjust the accuracy and consistency of the single-dimensional reconstruction vector. The dimension reconstruction optimization function quantifies the difference between the original feature and the reconstructed vector by calculating the reconstruction error matrix, where the elements of the reconstruction error matrix represent the amount of information lost in the dimensionality reduction reconstruction process of each dimensional feature. Based on gradient descent or other optimization algorithms, the reconstruction error matrix is ​​iteratively optimized to gradually reduce the reconstruction error and obtain the reconstruction adjustment coefficient used to adjust the reconstruction result. The reconstruction adjustment coefficient improves the representativeness and accuracy of the reconstruction vector by weighting the reconstructed features. The reconstruction adjustment coefficient is fused with each group of single-dimensional reconstruction vectors respectively, and a fusion feature matrix is ​​generated by weighted summing of the feature vectors of each dimension. A feature fusion loss function is constructed to quantify the error value after fusion. The fusion loss function evaluates the matching degree between the fusion matrix and the original multi-dimensional feature representation set, finds missing information or deviations, and adjusts these deviations in further optimization steps. Regularization constraints and dimension complementation calculations are performed on the fusion feature matrix. Regularization constraints constrain the distribution of matrix elements to avoid overfitting caused by too strong features in a certain dimension. Dimension complementation calculations analyze the complementary relationship between the features of each dimension, explore the synergy hidden in different dimensions, and obtain a comprehensive drug storage representation. Storage location allocation and inventory parameter calculations are performed based on the comprehensive drug storage representation. Storage location allocation analyzes the characteristic values ​​of storage conditions in the comprehensive representation, combines the availability of the actual storage environment of the drug, such as temperature and humidity range, storage space restrictions, etc., and uses optimization algorithms such as dynamic programming or genetic algorithms to allocate the best storage location for each drug. At the same time, inventory parameter calculations are based on the inventory level characteristics and timeliness characteristics in the comprehensive representation, and calculate key indicators such as the drug reorder point (ROP), economic order quantity (EOQ) and inventory turnover cycle. The calculation process of these parameters combines the consumption rate of the drug, the supply chain delivery cycle and the expiration date limit to ensure that inventory management can balance efficiency and safety. The clinical drug storage management strategy is obtained by integrating the storage location allocation plan and the inventory parameter calculation results.

[0034] A spatial clustering analysis is performed on the comprehensive drug storage representation. By analyzing the similarity of drugs in different storage condition dimensions and combining the dependence of drugs on storage conditions such as temperature, humidity, and light protection, clustering algorithms (such as K-Means or DBSCAN) are used to divide drugs into several storage categories, and a storage space allocation matrix is ​​constructed based on these categories. The storage space allocation matrix records the distribution ratio and storage requirements of drugs of each storage category in the warehouse. According to the allocation matrix, combined with the actual layout of storage resources (such as shelves, cold storage areas, and normal temperature areas), regional division is performed to generate a storage partitioning scheme to ensure the adaptability of drug storage categories and physical storage areas. Environmental parameters are set for the storage partitioning scheme. By analyzing the storage requirements of each drug category, such as temperature range, humidity threshold, and light conditions, a corresponding set of monitoring indicators is constructed. Based on the set of monitoring indicators, a parameter monitoring matrix is ​​established, which records the environmental monitoring rules in each storage partition, including target parameter values, tolerance deviation ranges, and alarm trigger thresholds. At the same time, a storage condition matching calculation is performed on the comprehensive drug storage representation. By matching the storage demand characteristics of the drug with the environmental parameters of the actual storage partition, a location allocation matrix is ​​generated. The location allocation matrix assigns an optimal storage location to each drug to ensure that the drug can be stored in a suitable environment. To improve the utilization of storage space, the location allocation matrix is ​​optimized for spatial layout. By introducing optimization algorithms (such as genetic algorithms or simulated annealing algorithms), the spatial layout inside the storage area is rearranged to ensure efficient utilization of storage resources and reduce the length of the access path of drugs. The optimized results generate a storage location mapping table to record the correspondence between each drug and the specific storage location. Time series prediction of drug demand. By analyzing the inventory level characteristics in the comprehensive drug storage representation, combining historical demand data and seasonal fluctuation trends, the drug demand forecast matrix is ​​calculated using time series prediction models (such as ARIMA and LSTM). The demand forecast matrix records the expected demand for each drug in the future, and based on this, the inventory control parameters are calculated, including the reorder point (ROP), economic order quantity (EOQ) and the maximum inventory threshold. Scheduling optimization is performed based on the storage location mapping table and inventory control parameters. By comprehensively considering the storage location, demand forecast and inventory level of drugs, a storage scheduling matrix is ​​constructed, which describes the access plan of drugs at different time points and the allocation of storage resources. Combined with environmental monitoring rules, the dynamic operations in the storage scheduling matrix are associated with environmental monitoring requirements to ensure that the storage management strategy meets both scheduling requirements and the stability requirements of the drug storage environment during execution. By combining the storage scheduling matrix with environmental monitoring rules, a complete clinical drug storage management strategy is generated. The strategy includes multiple aspects such as drug storage partition planning, location allocation, environmental monitoring, inventory management, and scheduling optimization.

[0035] In the embodiment of the present invention, by introducing an adjacency-aware fuzzy label learning framework, multi-dimensional analysis and storage category prediction are performed on drug features, thereby improving the accuracy of drug classification and solving the problem of inconsistent storage classification standards; a selective masking and sequence prediction mechanism is adopted to construct a prediction model with adaptive characteristics, thereby reducing error propagation in the prediction process and improving the reliability of the prediction results; based on the three-dimensional feature decomposition space and the dimensional utility matrix, refined decomposition and feature extraction of storage conditions, inventory levels and timeliness dimensions are achieved, thereby enhancing the expression ability of multi-dimensional features; through single-dimensional reconstruction and cross-dimensional fusion technology, a comprehensive representation method that takes into account local features and global associations is established, thereby optimizing the integrity and accuracy of feature representation; by integrating spatial clustering analysis and environmental parameter monitoring, a dynamically adjusted storage management strategy is constructed, thereby improving storage space utilization and environmental control accuracy; based on time series prediction and scheduling optimization algorithms, a drug inventory early warning and replenishment mechanism is established, thereby reducing inventory management costs and improving the intelligence level of drug storage management.

[0036] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0037] Collect and electronically input basic information of clinical drugs to obtain the original data set including drug name, specification, dosage form, storage conditions, expiration date, and production batch number;

[0038] Perform natural language preprocessing and character encoding conversion on the text fields in the original data set to obtain a standardized drug text feature set, and perform feature quantization and vectorization conversion on the standardized drug text feature set to obtain a numerical feature matrix;

[0039] Perform maximum and minimum normalization and standardization calculations on the numerical feature matrix to obtain a standardized feature vector, and perform correlation analysis and feature screening on the standardized feature vector to obtain n core feature dimensions;

[0040] According to n core feature dimensions, the standardized feature vectors are reorganized to obtain an n-dimensional drug feature vector set;

[0041] The feature importance weight value of each dimension feature of the n-dimensional drug feature vector set is calculated by the information gain algorithm to obtain an initial weight vector, and the initial weight vector is normalized and nonlinearly mapped to obtain a feature weight matrix;

[0042] Adjacency perception calculation is performed on the n-dimensional drug feature vector set and feature weight matrix to obtain the drug membership decision sequence of k storage categories.

[0043] Specifically, basic information of clinical drugs is collected and electronically input, and the key information of drugs is recorded in the system to form an original data set, including fields such as name, specification, dosage form, storage conditions, expiration date and production batch number. The original data is preprocessed. In the preprocessing, natural language processing is performed on the text fields in the original data set, such as removing redundant symbols, standardizing case, and removing stop words. After processing, the text fields are converted to character encoding, such as converting each text into a fixed-length numerical vector through word embedding technology to generate a standardized text feature set. Through feature quantization and vectorization, all fields are integrated into a numerical feature matrix, denoted as X=[x ij ], where x ij Represents the value of the i-th drug on the j-th feature. The maximum and minimum values ​​of the numerical feature matrix are normalized, and the formula is:

[0044]

[0045] Among them, min(x j ) and max(x j ) represent the minimum and maximum values ​​of feature j, respectively. After normalization, all feature values ​​are scaled to the range of [0,1], eliminating the influence of different dimensions. Then standardization is performed to adjust the mean of each feature to 0 and the standard deviation to 1. The formula is:

[0046]

[0047] Among them, μ j and σ j are the mean and standard deviation of feature j respectively. Through this step, the standardized feature matrix Z is obtained. The standardized feature vector is subjected to correlation analysis, and the correlation coefficient matrix R = [r jk ] represents the correlation between feature j and feature k:

[0048]

[0049] By analyzing the correlation, n core feature dimensions with low correlation but significant impact on the target variable are screened out. These dimensions are used to reorganize the standardized feature vectors to generate an n-dimensional drug feature vector set V = [v ij ], where v ij Represents the value of the i-th drug in the j-th core feature dimension. The importance weight value of each core feature is calculated based on the information gain algorithm. Information gain IG(T,F j )The formula is:

[0050] IG(T,F j )=H(T)-H(T|F j );

[0051] Where T is the target storage category, F j is the jth feature, H(T) and H(T|F j ) represent the entropy of the target and the conditional entropy under given feature conditions. By calculating the information gain of the n-dimensional features, the initial weight vector w = [w1, w2, ..., w n ]. To optimize the weight distribution, the initial weights are normalized and a nonlinear mapping function φ(x) is applied, such as w j ′ =φ(w j ), generate feature weight matrix W = [w jk ]. Perform adjacency perception calculation on the drug feature vector set V and feature weight matrix W. Use the weighted Euclidean distance formula to calculate the similarity between drugs:

[0052]

[0053] Based on the calculated similarity matrix, the drugs are divided into k storage categories through a clustering algorithm (such as K-Means) to generate a membership decision sequence for each drug. The definition of the membership value is based on the distance between the drug and the cluster center:

[0054]

[0055] where u ic represents the membership of drug i to storage category c, d ic is the distance from drug i to category c. Through the above steps, we finally get the drug membership decision sequence of k storage categories.

[0056] In a specific embodiment, the execution step performs adjacency perception calculation on the n-dimensional drug feature vector set and the feature weight matrix to obtain the drug membership decision sequence of k storage categories, which may specifically include the following steps:

[0057] Perform k-means clustering analysis on the n-dimensional drug feature vector set to obtain an initial storage category core vector set, and perform adjacency calculation on the initial storage category core vector set to obtain a storage category adjacency weight function;

[0058] The feature vectors in the n-dimensional drug feature vector set are calculated by Euclidean distance with the storage category core vector set to obtain a basic distance matrix;

[0059] The basic distance matrix is ​​weighted and adjusted according to the feature weight matrix to obtain a weighted distance matrix, and the weighted distance matrix is ​​adjacency-aware adjusted based on the storage category adjacency weight function to obtain an initial membership matrix;

[0060] The initial membership matrix is ​​optimized based on the sequence joint distribution loss function to obtain the optimized membership matrix that takes into account the difference between high and low fuzzy memberships. The optimized membership matrix is ​​rearranged according to the drug index order to obtain the drug membership decision sequence of k storage categories.

[0061] Specifically, for the n-dimensional drug feature vector set X = [x1, x2, ..., x N ] to perform k-means cluster analysis, where x i ∈R n is the feature vector of the ith drug, N is the total number of drugs, and n is the feature dimension. These feature vectors are divided into k storage categories using the k-means algorithm. The clustering process aims to minimize the intra-group square error, and its optimization objective function is:

[0062]

[0063] Among them, c j ∈R n is the core vector (i.e. cluster center) of category j, z ij is an indicator variable. When drug i belongs to category j, z ij =1, otherwise z ij = 0. Update c by iteration j and z ij , and obtain k storage category core vectors {c1,c2,…,c k}, forming the initial storage category core vector set. The adjacency relationship of the initial storage category core vector set is calculated to generate the storage category adjacency weight function. Assume that category j and category j ′ The adjacency weight between is based on the Euclidean distance, and its weight function is defined as:

[0064]

[0065] Among them, ∥c j -c j′ ∥ indicates categories j and j ′ The Euclidean distance between the core vectors of , σ is a scale parameter used to adjust the distance effect. Storage category adjacency weight function w jj′ Describes the proximity between different categories and is used for subsequent adjacency perception calculations. i The Euclidean distance is calculated with the core vector in the storage category core vector set to obtain the basic distance matrix D = [d ij ],in:

[0066] d ij =∥x i -c j ∥;

[0067] Among them, d ij represents the distance between drug i and storage category j. In order to measure the distance more accurately, according to the feature weight matrix W = [w mn ]Weighted adjustment is performed on the basic distance matrix to generate a weighted distance matrix D ′ =[d i ′ j ]. The weighted distance formula is:

[0068]

[0069] Among them, w m represents the weight of feature m, x im and c jm are the values ​​of drug i and category j on the mth feature dimension respectively. By introducing weights, the features that have a greater impact on the storage category decision have a higher weight. After obtaining the weighted distance matrix, the adjacency perception adjustment is performed on it in combination with the storage category adjacency weight function to generate the initial membership matrix U = [u ij ]. The adjusted membership value is expressed as:

[0070]

[0071] in, Indicates that the membership value of drug i is normalized across all storage categories. In order to optimize the membership matrix, the initial membership matrix is ​​optimized based on the sequence joint distribution loss function. The loss function is defined as follows:

[0072]

[0073] in, is the desired membership value, and λ is the regularization coefficient used to balance the entropy loss and the square error loss. Through optimization, the optimized membership matrix is ​​generated This matrix more accurately reflects the relationship between drugs and storage categories. The optimized membership matrix is ​​rearranged according to the drug index order to obtain the final drug membership decision sequence of k storage categories.

[0074] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0075] Perform random stratified sampling on drug membership decision sequences to obtain training sets and validation sets;

[0076] Performing distribution statistics analysis on the storage category membership values ​​in the training set to obtain a membership threshold range, and hierarchically classifying the membership values ​​based on the membership threshold range to obtain a multi-level masking label set;

[0077] Constructing a selective masking function according to a multi-level masking tag set to obtain a sequence masking rule, and replacing the target storage category membership value in the training set according to the sequence masking rule to obtain a masked training sequence;

[0078] Input the masked training sequence into the consistency prediction model for parameter optimization training to obtain a sequence reconstruction model;

[0079] The drug membership decision sequence is masked with the target storage category membership value according to the sequence masking rule to obtain the masked sequence to be predicted, and the sequence reconstruction model is applied to the masked sequence to perform sequence completion calculation to obtain the drug storage category prediction result.

[0080] Specifically, for the drug membership decision sequence, assume that its form is matrix U = [u ij ], where u ij represents the membership value of the i-th drug to the j-th storage category. The membership value satisfies Where k is the number of stored categories. In order to train and validate the model, the membership decision sequence is randomly stratified sampled. The purpose of stratified sampling is to keep the category distribution in the training set and validation set consistent with the original data, thereby avoiding the impact of bias on model performance. Assume that in the original data, the distribution ratio of category j is Where N j is the number of samples of category j, and N is the total number of samples. The training set U is extracted from U in proportion train and validation set U val , the formula is:

[0081]

[0082] Among them, N train is the target size of the training set, |U train ,j| represents the number of training samples of category j. After generating the training set, the distribution statistics of the stored category membership values ​​in the training set are analyzed. By calculating the mean μ of each category j and standard deviation σ j , the statistical range of membership value is obtained:

[0083] R j =[μ j -kσ j ,μ j +kσ j ];

[0084] Among them, R jis the membership threshold range for category j, and k is the parameter for the control range (usually 2 to cover 95% of the data). According to the threshold range, the membership values ​​are divided into different levels. For example, below μ j -σ j The value of is marked as low membership, and is located in [μ j -σ j ,μ j +σ j ] is marked as medium membership, higher than μ j +σ j The values ​​of are marked as high membership, generating a multi-level masking tag set. Based on the multi-level masking tag set, a selective masking function is designed to control the masking process of the membership value of the target storage category. The masking function is defined as:

[0085]

[0086] Among them, M(u ij ) represents the membership value after masking, P mask (u ij ) is a masking probability function, which is usually related to the level of membership. For example, low membership values ​​have a higher masking probability. p is a random number in the interval [0,1]. According to the masking rule, the target storage category membership values ​​in the training set are masked and replaced to obtain the masked training sequence U mask The masked training sequence is input into the consistency prediction model for training. The goal of the model is to predict the masked value based on the unmasked membership value. The loss function used in training is the mean square error (MSE):

[0087]

[0088] Where is the total loss, is the membership value predicted by the model, u ij is the original membership value. The model parameters are optimized by back propagation until the loss function converges to obtain the sequence reconstruction model. After the model training is completed, the drug membership decision sequence is masked according to the same masking rule to obtain the masked sequence to be predicted U test-mask . Input it into the sequence reconstruction model and use the model's completion capability to generate a complete prediction sequence. The prediction formula of the model is:

[0089]

[0090] in, is the predicted membership value, f model is the prediction function of the model.

[0091] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0092] Based on the prediction results of drug storage categories, a three-dimensional feature decomposition space is constructed to obtain the storage condition feature matrix, inventory level feature matrix and timeliness feature matrix.

[0093] Perform temperature and humidity correlation analysis on the storage condition feature matrix to obtain a storage condition feature calculation function, and input the drug storage category prediction result into the storage condition feature calculation function to obtain a storage condition feature sub-vector;

[0094] Perform inventory fluctuation analysis on the inventory level feature matrix to obtain the inventory level feature calculation function, and input the drug storage category prediction result into the inventory level feature calculation function to obtain the inventory level feature sub-vector;

[0095] Perform time series analysis on the timeliness feature matrix to obtain a timeliness feature calculation function, and input the drug storage category prediction result into the timeliness feature calculation function to obtain a timeliness feature sub-vector;

[0096] Construct a dimension utility matrix, obtain the dimension interaction coefficients of storage conditions, inventory levels, and timeliness, and calculate the weight distribution of feature subvectors based on the dimension interaction coefficients to obtain the dimension weight matrix;

[0097] The storage condition feature sub-vector, inventory level feature sub-vector and timeliness feature sub-vector are weightedly combined to obtain a multi-dimensional feature representation set.

[0098] Specifically, based on the prediction results of drug storage categories, a three-dimensional feature decomposition space is constructed. Assume that the prediction results of drug storage categories are represented by the matrix P = [p ij ] indicates that p ij is the membership value of the i-th drug to storage category j, satisfying ∑ j p ij =1. The construction of the three-dimensional feature decomposition space decomposes the prediction result matrix into three feature dimensions, corresponding to the storage condition feature matrix C, the inventory level feature matrix L and the timeliness feature matrix T. The matrix decomposition process is expressed by the following formula:

[0099] P≈C·L·T T ;

[0100] Where C = [c im ]、L=[l mn ]、T=[t nj ] represent the relationship matrix between drugs and storage conditions, inventory levels, and timeliness characteristics, and each element p of P ijIt is calculated by combining these three characteristic relationships. For the storage condition characteristic matrix C, the correlation analysis of temperature and humidity is performed to characterize the adaptability of drugs to different storage environments. Assuming that the storage conditions of drugs depend on temperature θ and humidity φ, the correlation analysis is expressed by the correlation function R(θ,φ) as follows:

[0101]

[0102] Among them, θ i and φ i They are the storage temperature and humidity requirements of drug i, is the mean of temperature and humidity, and R(θ,φ) represents the degree of linear correlation between the two. Based on the analysis results, the storage condition characteristic calculation function f is defined c (p ij ), store the prediction result p ij After entering this function, the storage conditional feature vector c is generated i =[c i1 ,c i2 ,…,c im ]. For the inventory level feature matrix L, the inventory fluctuation analysis is performed. Assuming that the change of inventory over time is represented by a time series, the inventory level fluctuation function is defined as:

[0103] S(t)=α+βt+γsin(ωt)+∈;

[0104] Among them, S(t) is the inventory level at time t, α is the initial inventory, β is the trend term of inventory change, γ and ω represent the amplitude and frequency of seasonal fluctuations respectively, and ∈ is the random error term. Based on this model, the inventory level characteristic calculation function f is defined l (p ij ), the prediction result p ij After input, generate the inventory level feature subvector l i =[l i1 ,l i2 ,…,l im ]. For the timeliness characteristic matrix T, a time series analysis is performed to characterize the changing patterns of the drug's validity period and usage cycle. Assuming that the amount of drug usage changes over time, the ARIMA model is used to model the prediction formula:

[0105] X t =φ1X t-1 +φ2X t-2 +…+φ p X t-p +∈ t ;

[0106] Among them, X tis the amount of medicine used at time t, φ1, φ2, …, φ p is the autoregressive coefficient, ∈ t is the error term. Based on the prediction results, define the timeliness feature calculation function f t (p ij ), generate the time-sensitive feature sub-vector t i =[t i1 ,t i2 ,…,t im After obtaining the characteristic subvectors of storage conditions, inventory levels, and timeliness, a dimension utility matrix is ​​constructed to quantify the interaction between dimensions. The element e of the dimension utility matrix E is m n represents the interaction coefficient of dimensions m and n, and the calculation formula is:

[0107]

[0108] Among them, v im and v in are the eigenvalues ​​of drug i in dimensions m and n respectively. The weight distribution of the feature subvector is calculated based on the interaction coefficient, and the dimension weight matrix W is generated. i , inventory level feature vector l i and the timeliness feature vector t i Perform weighted combination to generate a multi-dimensional feature representation set F i :

[0109] F i =W·[c i ,l i ,t i ] T ;

[0110] The multi-dimensional feature representation set integrates the characteristic information of drug storage conditions, inventory levels and timeliness.

[0111] In this embodiment, a dimensional utility matrix is ​​constructed to obtain the dimensional interaction coefficients of storage conditions, inventory levels, and timeliness, and the weight distribution of the feature subvectors is calculated based on the dimensional interaction coefficients to obtain a dimensional weight matrix, including: performing tensor transformation on the storage condition feature subvector, the inventory level feature subvector, and the timeliness feature subvector to obtain a three-dimensional feature tensor, and constructing an input layer matrix of a multi-derivative coupled neural network for the three-dimensional feature tensor; sending the input layer matrix to the hidden layer of the multi-derivative coupled neural network for multi-order derivative decomposition to obtain a first-order derivative matrix and a second-order derivative matrix between dimensions; performing inter-dimensional gradient calculation on the first-order derivative matrix to obtain a dimensional coupling strength vector, and performing feature decomposition on the second-order derivative matrix to obtain dimensional collaborative eigenvalues; based on the dimensional coupling strength vector and the dimensional collaborative The eigenvalues ​​are used to construct the neuron activation function to obtain the dimensional interaction response function, and the dimensional interaction response function is nonlinearly mapped to obtain the dimensional utility matrix; the dimensional utility matrix is ​​subjected to singular value decomposition to obtain the dimensional interaction coefficient matrix, and the correlation scores between dimensions are calculated based on the dimensional interaction coefficient matrix to obtain the dimensional correlation matrix; the dimensional correlation matrix is ​​input into the weight distribution model to obtain the initial weight distribution vector, and the initial weight distribution vector is normalized to obtain the standardized weight vector; matrix multiplication operation is performed on the standardized weight vector and the dimensional interaction coefficient matrix to obtain the weighted feature matrix, and the weighted feature matrix is ​​normalized to obtain the initial dimensional weight matrix; the back propagation algorithm is applied to the initial dimensional weight matrix for optimization iteration to obtain the dimensional weight matrix.

[0112] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0113] The storage condition feature sub-vector, inventory level feature sub-vector and timeliness feature sub-vector in the multi-dimensional feature representation set are reconstructed by dimensionality reduction, and three sets of single-dimensional reconstruction vectors are obtained;

[0114] Construct a dimensional reconstruction optimization function to obtain a reconstruction error matrix, and perform iterative parameter optimization based on the reconstruction error matrix to obtain a reconstruction adjustment coefficient;

[0115] Each group of single-dimensional reconstruction vectors and reconstruction adjustment coefficients are dimensional-fused to obtain a fusion feature matrix, and a feature fusion loss function is constructed to obtain a fusion error value;

[0116] Regularization constraints and dimension complementation calculations are performed on the fused feature matrix to obtain a comprehensive drug storage representation;

[0117] Based on the comprehensive drug storage representation, storage location allocation and inventory parameter calculation are performed to obtain the clinical drug storage management strategy.

[0118] Specifically, extract the storage condition feature sub-vector C = [c1, c2, …, c N , inventory level feature sub-vector L = [l1, l2, …, l N , and timeliness feature sub-vector T = [t1, t2, …, t N from the multi-dimensional feature representation set, where c i , l i , t i ∈ R n is the feature representation of the i-th drug in the corresponding dimension, N is the number of drugs, and n is the dimension size of the original features. To reduce redundant information and dimensional complexity, principal component analysis is used to perform dimensionality reduction and reconstruction on each group of feature sub-vectors, compressing the n-dimensional feature vectors into m-dimensional reconstruction vectors (m < n). By maximizing the variance of the data in the new coordinate system, the most representative m principal components are selected. The dimensionality reduction formula is:

[0119] y i = W T x i ;

[0120] where, y i ∈ R m is the feature vector after dimensionality reduction, W ∈ R n×m is the dimensionality reduction matrix, which is composed of the first m eigenvectors of the original feature covariance matrix, and x i ∈ R n is the original feature vector. Through dimensionality reduction, the reconstructed feature matrices C ′ , L ′ , and T ′ are generated for the storage condition, inventory level, and timeliness respectively. A dimensional reconstruction optimization function is constructed to optimize the dimensionality reduction result. Assuming the reconstructed result after dimensionality reduction is The reconstruction error matrix is defined as:

[0121]

[0122] where, E represents the total reconstruction error of all samples, x i is the original feature vector, is the reconstructed feature vector. The parameter iteration optimization of W is performed by the gradient descent method, and the update formula is:

[0123]

[0124] where, η is the learning rate, and W (t) is the dimensionality reduction matrix at the t-th iteration. The optimized result generates the reconstruction adjustment coefficient W *, further improving the reconstruction accuracy. Each set of single-dimensional reconstruction vectors and the corresponding reconstruction adjustment coefficients are dimensional-fused to generate a fusion feature matrix F. The fusion process uses a weighted average method, and the formula is:

[0126] f i =αc i ′+βl i ′+γt i ′;

[0127] Among them, f i ∈R m is the fusion feature representation of drug i, α, β, γ are the weights of storage conditions, inventory levels and timeliness dimensions, satisfying α+β+γ=1. In order to optimize the fusion result, the feature fusion loss function is constructed It is defined as

[0128]

[0129] in, is the ideal fusion feature vector, is a regularization term used to avoid overfitting. By minimizing, the fusion error value is obtained and the fusion parameters are adjusted. Regularization constraints and dimension complementation calculations are performed on the fusion feature matrix to generate a comprehensive drug storage representation S. Regularization is performed through normalization operations so that all eigenvalues ​​fall within the range of [0,1], while dimension complementation calculations are performed by analyzing the synergistic relationship of different dimensions and using the utility matrix E mn The interaction weights are modified by:

[0130] S = F·E;

[0131] Where E represents the weight matrix of dimension interaction. Storage location allocation and inventory parameter calculation are performed based on the comprehensive drug storage representation. Storage location allocation uses an optimization algorithm (such as dynamic programming) to allocate drugs to the most suitable storage area based on their storage condition characteristic values. Inventory parameter calculation includes economic order quantity (EOQ) and reorder point (ROP), and the formula is:

[0132]

[0133] ROP = d·L;

[0134] Among them, D is the annual demand, S is the single ordering cost, H is the unit holding cost, d is the daily demand, and L is the ordering lead time. Through these calculations, the clinical drug storage management strategy is generated.

[0135] In a specific embodiment, the execution step performs storage location allocation and inventory parameter calculation based on the comprehensive drug storage representation to obtain a clinical drug storage management strategy, which may specifically include the following steps:

[0136] Perform spatial cluster analysis on the comprehensive drug storage representation to obtain a storage space allocation matrix, and perform regional division based on the storage space allocation matrix to obtain a storage partitioning scheme;

[0137] Set environmental parameters for the storage partitioning scheme to obtain a set of monitoring indicators, and establish a parameter monitoring matrix based on the set of monitoring indicators to obtain environmental monitoring rules;

[0138] Perform storage condition matching calculation on the comprehensive drug storage representation to obtain a location allocation matrix, and perform spatial layout optimization based on the location allocation matrix to obtain a storage location mapping table;

[0139] Perform time series forecasting on the comprehensive drug storage representation to obtain a demand forecast matrix, and calculate the inventory threshold based on the demand forecast matrix to obtain inventory control parameters;

[0140] Scheduling optimization is performed based on the storage location mapping table and inventory control parameters to obtain a storage scheduling matrix, and a clinical drug storage management strategy is generated based on the storage scheduling matrix and environmental monitoring rules.

[0141] Specifically, the comprehensive drug storage representation matrix S = [s1, s2, ..., s N ] is used as input to perform spatial cluster analysis, where s i ∈R n represents the storage feature vector of the i-th drug, N is the number of drugs, and n is the storage feature dimension. The goal of cluster analysis is to divide drugs into several storage categories to optimize storage space utilization. The k-means clustering algorithm is used to iteratively calculate the cluster center c j And sample allocation, realize drug grouping, and the optimization objective function is:

[0142]

[0143] Among them, c j ∈R n is the cluster center of category j, z ij ∈{0,1} is an indicator variable. When drug i belongs to category j, z ij = 1. Finally, the storage space allocation matrix C = [c ij ], where c ijrepresents the probability of drug i being assigned to storage category j. Based on C, the storage space is divided into regions to form a storage partitioning scheme, where each partition corresponds to a group of drugs with similar storage condition requirements. The environmental parameters of the storage partitioning scheme are set, and the environmental parameters such as temperature, humidity, and light are determined according to the characteristics of the drugs in each partition. Assume that the environmental parameter set of storage partition j is E j =[θ j ,φ j ,γ j ], where θ j ,φ j , γ j Respectively represent the target temperature, humidity and light level of partition j. Based on these settings, a set of monitoring indicators is generated, and a parameter monitoring matrix M = [m jk ], where m jk represents the target value of the kth environmental parameter of partition j. The environmental monitoring rule is defined based on the monitoring matrix and its form is:

[0144]

[0145] in, is the real-time monitoring value, Δm jk Represents the monitoring error. When the error exceeds the tolerance range, the alarm mechanism is triggered. The storage condition matching calculation is performed on the comprehensive drug storage representation to generate the location allocation matrix P = [p ij ]. Assume that the storage condition characteristic of drug i is s i , the condition requirement for storage partition j is e j , the matching degree is defined as:

[0146]

[0147] Where σ is a tuning parameter used to control the matching sensitivity. Based on P, the storage space layout is optimized, and a genetic algorithm or simulated annealing algorithm is used to maximize the matching degree, and a storage location mapping table L = [l ij ], where l ij represents the optimal storage location of drug i. After the storage location allocation is completed, the time series prediction of the comprehensive drug storage representation is performed to generate the demand prediction matrix D = [d it ], where d it represents the demand for drug i at time t. Assuming that the demand change satisfies the ARIMA model, the prediction formula is:

[0148] X t =φ1X t-1 +φ2X t-2 +…+φ p X t-p +∈ t ;

[0149] Among them, φ1, φ2, …, φ p are the model coefficients, ∈ t is a random error. Based on the forecast results, the inventory threshold is calculated and the reorder point (ROP) and economic order quantity (EOQ) are defined:

[0150] ROP = d·L;

[0151]

[0152] Where d is the daily demand, L is the ordering lead time, D is the annual demand, S is the single ordering cost, and H is the unit holding cost. The inventory control parameter matrix is ​​generated through D. Combined with the storage location mapping table and inventory control parameters, scheduling optimization is performed to generate the storage scheduling matrix T = [t ij ], where t ij represents the scheduling plan of drug i in partition j. The optimization goal is to minimize the scheduling cost, and the objective function is:

[0153]

[0154] Among them, c ij is the storage and access cost of drug i in partition j, and λ is the weight factor, which balances the relationship between scheduling and environmental monitoring. Combined with environmental monitoring rules, a complete clinical drug storage management strategy is finally generated.

[0155] In this embodiment, generating a clinical drug storage management strategy based on a storage scheduling matrix and environmental monitoring rules also includes: sliding segmenting the environmental parameter sequence in the storage scheduling matrix according to the time window to obtain a set of time series sample pairs, and performing difference quantization calculation on the set of time series sample pairs to obtain a sample difference matrix, using a rank sum test statistic on the sample difference matrix to obtain a significance level sequence, setting a threshold interval based on the significance level sequence to obtain an environmental change intensity index; inputting the environmental change intensity index into a deep perception network to obtain a multi-layer feature mapping matrix, performing inter-layer correlation analysis on the multi-layer feature mapping matrix to obtain a feature correlation graph, and performing feature correlation analysis based on the feature correlation. The environmental change classifier is constructed by connecting graphs to obtain the environmental change type matrix. The change intensity evaluation function is designed using the environmental change type matrix, and the environmental fluctuation parameters are obtained through iterative optimization calculation. The storage scheduling matrix is ​​decomposed into time series using a three-parameter exponential smoothing algorithm to obtain a horizontal component sequence, a trend component sequence, and a periodic component sequence. A linear combination prediction model is constructed based on the horizontal component sequence and the trend component sequence to obtain an initial prediction sequence. The periodic correction is combined with the periodic component sequence to obtain a storage prediction vector. A dynamic reference point generator is constructed to generate an adaptive reference point set based on the historical optimal solution set, and the storage prediction vector is calculated from the adaptive reference point set. The Euclidean distance matrix between the two groups is used to obtain the position deviation matrix, and the kernel function transformation is used to perform nonlinear mapping on the position deviation matrix to obtain the prediction correction coefficient; the prediction correction coefficient is fused with the stored prediction vector by the weighted linear combination method to obtain the correction prediction result, and a multi-objective optimization model is constructed. The correction prediction result is substituted into the objective function to calculate the fitness value, and the evolution parameter set is obtained by non-dominated sorting; the evolution parameter set is clustered based on the environmental fluctuation parameter to obtain the parameter distribution matrix, and the distances between and within the class are calculated to obtain the environmental fitness matrix. The main mutation operator based on Gaussian distribution and the secondary mutation operator based on Cauchy distribution are designed, and the double mutation probability matrix is ​​constructed by combination. The mutation probability vector is obtained by applying the mutation probability vector to the correction prediction result according to the adaptive weight allocation strategy, and the initial population optimization sequence is obtained. The initial population optimization sequence is evaluated for diversity by using the improved crowding degree calculation method, and the population distribution characteristic vector is obtained. The diversity evaluation index is constructed based on the population distribution characteristic vector, and the diversity index set is obtained. The strategy optimization objective function is constructed according to the diversity index set, and the parameters of the clinical drug storage management strategy are optimized and the strategy is reorganized to obtain the candidate strategy set. The non-inferior solution set is obtained by screening through the Pareto dominance relationship, and the solution with the best comprehensive performance is selected from the non-inferior solution set as the optimal storage management strategy.

[0156] The clinical drug storage management method in the embodiment of the present invention is described above. The clinical drug storage management device in the embodiment of the present invention is described below. Figure 2 In one embodiment of the present invention, a clinical drug storage management device includes:

[0157] The data collection module is used to collect and standardize the data of clinical drugs, obtain an n-dimensional drug feature vector set and a feature weight matrix, and perform adjacency perception calculation to obtain a drug membership decision sequence of k storage categories;

[0158] A sequence prediction module is used to selectively mask and sequence-predict the target storage category membership values ​​in the drug membership decision sequence to obtain a drug storage category prediction result;

[0159] A decomposition module is used to decompose the drug storage category prediction results according to the storage condition dimension, inventory level dimension and timeliness dimension to obtain a multi-dimensional feature representation set;

[0160] The fusion module is used to perform single-dimensional reconstruction and cross-dimensional fusion of the multi-dimensional feature representation set to obtain a comprehensive drug storage representation, and to perform storage location allocation and inventory parameter calculation to obtain a clinical drug storage management strategy.

[0161] Through the collaborative cooperation of the above components, by introducing the adjacency-aware fuzzy label learning framework, multi-dimensional analysis of drug features and storage category prediction are carried out, which improves the accuracy of drug classification and solves the problem of inconsistent storage classification standards; by adopting selective masking and sequence prediction mechanisms, a prediction model with adaptive characteristics is constructed, which reduces the error propagation in the prediction process and improves the reliability of the prediction results; based on the three-dimensional feature decomposition space and dimensional utility matrix, the refined decomposition and feature extraction of storage conditions, inventory levels and timeliness dimensions are realized, and the expression ability of multi-dimensional features is enhanced; through single-dimensional reconstruction and cross-dimensional fusion technology, a comprehensive representation method that takes into account local features and global associations is established, and the integrity and accuracy of feature representation are optimized; by integrating spatial clustering analysis and environmental parameter monitoring, a dynamically adjusted storage management strategy is constructed, which improves the storage space utilization and environmental control accuracy; based on time series prediction and scheduling optimization algorithms, a drug inventory early warning and replenishment mechanism is established, which reduces inventory management costs and improves the intelligence level of drug storage management.

[0162] 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 3As 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.

[0163] 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.

[0164] 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.

[0165] 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 can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many 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.

[0166] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0167] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0168] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A clinical drug storage management method, characterized in that: The clinical drug storage management method comprises: Data collection and standardization of clinical drugs are performed to obtain an n-dimensional drug feature vector set and feature weight matrix, and adjacency perception calculation is performed to obtain a drug membership decision sequence of k storage categories; Selectively masking and sequence predicting the target storage category membership values ​​in the drug membership decision sequence to obtain a drug storage category prediction result; Decomposing the drug storage category prediction result according to the storage condition dimension, inventory level dimension and timeliness dimension to obtain a multi-dimensional feature representation set; The multi-dimensional feature representation set is reconstructed in a single dimension and fused across dimensions to obtain a comprehensive drug storage representation, and storage location allocation and inventory parameter calculation are performed to obtain a clinical drug storage management strategy.

2. The clinical drug storage management method according to claim 1, characterized in that: The clinical drugs are subjected to data collection and standardization processing to obtain an n-dimensional drug feature vector set and a feature weight matrix, and adjacency perception calculation is performed to obtain a drug membership decision sequence of k storage categories, including: Collect and electronically input basic information of clinical drugs to obtain the original data set including drug name, specification, dosage form, storage conditions, expiration date, and production batch number; Performing natural language preprocessing and character encoding conversion on the text fields in the original data set to obtain a standardized drug text feature set, and performing feature quantization and vectorization conversion on the standardized drug text feature set to obtain a numerical feature matrix; Performing maximum and minimum value normalization and standardization calculation on the numerical feature matrix to obtain a standardized feature vector, and performing correlation analysis and feature screening on the standardized feature vector to obtain n core feature dimensions; According to the n core feature dimensions, the standardized feature vectors are reorganized to obtain an n-dimensional drug feature vector set; Calculating the feature importance weight value of each dimension feature of the n-dimensional drug feature vector set by an information gain algorithm to obtain an initial weight vector, and normalizing and nonlinearly mapping the initial weight vector to obtain a feature weight matrix; Adjacency perception calculation is performed on the n-dimensional drug feature vector set and the feature weight matrix to obtain drug membership decision sequences of k storage categories.

3. The clinical drug storage management method according to claim 2, characterized in that: The adjacency perception calculation is performed on the n-dimensional drug feature vector set and the feature weight matrix to obtain a drug membership decision sequence of k storage categories, including: Performing k-means clustering analysis on the n-dimensional drug feature vector set to obtain an initial storage category core vector set, and performing adjacency calculation on the initial storage category core vector set to obtain a storage category adjacency weight function; Performing Euclidean distance calculation on the feature vectors in the n-dimensional drug feature vector set and the storage category core vector set to obtain a basic distance matrix; Performing weighted adjustment on the basic distance matrix according to the feature weight matrix to obtain a weighted distance matrix, and performing adjacency-aware adjustment on the weighted distance matrix based on the storage category adjacency weight function to obtain an initial membership matrix; The initial membership matrix is ​​optimized and calculated based on the sequence joint distribution loss function to obtain an optimized membership matrix that takes into account the difference between high and low fuzzy memberships, and the optimized membership matrix is ​​rearranged according to the drug index order to obtain a drug membership decision sequence of k storage categories.

4. The clinical drug storage management method according to claim 3, characterized in that: The selective masking and sequence prediction of the target storage category membership values ​​in the drug membership decision sequence to obtain the drug storage category prediction result includes: Performing random stratified sampling on the drug membership decision sequence to obtain a training set and a validation set; Performing distribution statistics analysis on the storage category membership values ​​in the training set to obtain a membership threshold range, and hierarchically classifying the membership values ​​based on the membership threshold range to obtain a multi-level masking label set; Constructing a selective masking function according to the multi-level masking tag set to obtain a sequence masking rule, and replacing the target storage category membership value in the training set according to the sequence masking rule to obtain a masked training sequence; Inputting the masked training sequence into a consistency prediction model for parameter optimization training to obtain a sequence reconstruction model; The drug membership decision sequence is masked with the target storage category membership value according to the sequence masking rule to obtain a masked sequence to be predicted, and the sequence reconstruction model is applied to the masked sequence to be predicted to perform sequence completion calculation to obtain a drug storage category prediction result.

5. The clinical drug storage management method according to claim 4, characterized in that: The drug storage category prediction result is decomposed according to the storage condition dimension, inventory level dimension and timeliness dimension to obtain a multi-dimensional feature representation set, including: Based on the drug storage category prediction results, a three-dimensional feature decomposition space is constructed to obtain a storage condition feature matrix, an inventory level feature matrix, and a timeliness feature matrix; Performing temperature and humidity correlation analysis on the storage condition feature matrix to obtain a storage condition feature calculation function, and inputting the drug storage category prediction result into the storage condition feature calculation function to obtain a storage condition feature subvector; Performing inventory fluctuation analysis on the inventory level feature matrix to obtain an inventory level feature calculation function, and inputting the drug storage category prediction result into the inventory level feature calculation function to obtain an inventory level feature sub-vector; Performing time series analysis on the timeliness feature matrix to obtain a timeliness feature calculation function, and inputting the drug storage category prediction result into the timeliness feature calculation function to obtain a timeliness feature subvector; Constructing a dimension utility matrix, obtaining dimension interaction coefficients of storage conditions, inventory levels, and timeliness, and calculating weight distribution of feature subvectors based on the dimension interaction coefficients to obtain a dimension weight matrix; The storage condition feature sub-vector, the inventory level feature sub-vector and the timeliness feature sub-vector are weightedly combined to obtain a multi-dimensional feature representation set.

6. The clinical drug storage management method according to claim 5, characterized in that: The single-dimensional reconstruction and cross-dimensional fusion of the multi-dimensional feature representation set are performed to obtain a comprehensive drug storage representation, and storage location allocation and inventory parameter calculation are performed to obtain a clinical drug storage management strategy, including: Respectively performing dimensionality reduction reconstruction on the storage condition feature subvector, the inventory level feature subvector and the timeliness feature subvector in the multi-dimensional feature representation set to obtain three sets of single-dimensional reconstruction vectors; Constructing a dimensional reconstruction optimization function to obtain a reconstruction error matrix, and performing iterative parameter optimization based on the reconstruction error matrix to obtain a reconstruction adjustment coefficient; Each group of single-dimensional reconstruction vectors and the reconstruction adjustment coefficients are dimensional-fused to obtain a fusion feature matrix, and a feature fusion loss function is constructed to obtain a fusion error value; Performing regularization constraints and dimension complementation calculations on the fused feature matrix to obtain a comprehensive drug storage representation; Based on the comprehensive drug storage representation, storage location allocation and inventory parameter calculation are performed to obtain a clinical drug storage management strategy.

7. The clinical drug storage management method according to claim 6, characterized in that: The step of performing storage location allocation and inventory parameter calculation based on the comprehensive drug storage representation to obtain a clinical drug storage management strategy includes: Performing spatial cluster analysis on the comprehensive drug storage representation to obtain a storage space allocation matrix, and performing regional division based on the storage space allocation matrix to obtain a storage partitioning scheme; Setting environmental parameters for the storage partition scheme to obtain a monitoring indicator set, and establishing a parameter monitoring matrix based on the monitoring indicator set to obtain environmental monitoring rules; Performing storage condition matching calculation on the comprehensive drug storage representation to obtain a location allocation matrix, and performing spatial layout optimization based on the location allocation matrix to obtain a storage location mapping table; Performing time series forecasting on the comprehensive drug storage representation to obtain a demand forecast matrix, and calculating an inventory threshold based on the demand forecast matrix to obtain an inventory control parameter; Scheduling optimization is performed based on the storage location mapping table and the inventory control parameters to obtain a storage scheduling matrix, and a clinical drug storage management strategy is generated based on the storage scheduling matrix and the environmental monitoring rules.

8. A clinical drug storage management device, characterized in that: Used to execute the clinical drug storage management method according to any one of claims 1 to 7, the clinical drug storage management device comprises: The data collection module is used to collect and standardize the data of clinical drugs, obtain an n-dimensional drug feature vector set and a feature weight matrix, and perform adjacency perception calculation to obtain a drug membership decision sequence of k storage categories; A sequence prediction module, used for selectively masking and sequence prediction of the target storage category membership values ​​in the drug membership decision sequence to obtain a drug storage category prediction result; A decomposition module, used for decomposing the drug storage category prediction result according to the storage condition dimension, the inventory level dimension and the timeliness dimension to obtain a multi-dimensional feature representation set; The fusion module is used to perform single-dimensional reconstruction and cross-dimensional fusion on the multi-dimensional feature representation set to obtain a comprehensive drug storage representation, and perform storage location allocation and inventory parameter calculation to obtain a clinical drug storage management strategy.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the clinical drug warehouse management method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the clinical drug storage management method according to any one of claims 1 to 7.

Images