Medical refrigerator warehousing system based on internet of things

Abstract

A medical refrigerator storage system based on the Internet of Things (IoT) includes a refrigerator body, an authentication unit, an inventory unit, a display input unit, and a control unit mounted on the refrigerator body, as well as a remote management platform communicatively connected to the control unit. The refrigerator body stores items; the authentication unit authenticates the items based on their initial identification information; the inventory unit scans and verifies the quantity and information of the stored items after storage / retrieval; the display input unit is used to input or display relevant information about the stored items; the control unit is connected to the authentication unit, inventory unit, and display input unit, receiving information from each unit or sending control commands to each unit; the remote management platform is used to remotely assign operating permissions, monitor the status of the refrigerator body, and query historical data. This invention employs multiple measures to comprehensively ensure the safety of stored items, reduce medical risks, and provides a safer, more efficient, and intelligent storage management system for the medical field.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things (IoT) technology and medical device management technology, and is used for the storage, management and full-process traceability of medical blood. More specifically, it relates to an IoT-based medical refrigerator storage system. Background Technology

[0002] Medical warehousing systems are core equipment in the medical field for storing specialized items such as medicines, reagents, and biological samples. The stability of the storage environment, the security of management, and the standardization of operation directly affect the quality of medical care and patient safety. With the development of medical technology, the requirements for intelligent and precise medical warehousing are increasing, but existing technologies have the following significant shortcomings.

[0003] Low identity verification security: Traditional password and card verification methods are easily misused, and biometric identification is affected by makeup, lighting and facial expressions, resulting in poor verification accuracy and inability to effectively prevent unauthorized operations.

[0004] Inventory counting is inefficient and prone to errors: manual inventory counting is time-consuming and error-prone; RFID inventory systems suffer from crosstalk due to signal interference, making it difficult to accurately identify stored items in different areas.

[0005] Inaccurate storage management: Most medical refrigerators lack independent temperature control and safety locks for different layers, which cannot meet the temperature requirements of multiple types of items and can easily lead to cross-contamination or theft of stored items.

[0006] Lack of remote management capabilities: The existing system is mainly based on local management and cannot remotely assign permissions, monitor status, or query historical data, which is not conducive to the unified scheduling of medical resources across regions.

[0007] Poor ease of use: Information entry and command input are cumbersome, and key statuses such as temperature and inventory cannot be displayed intuitively, affecting operational efficiency.

[0008] Therefore, the existing technology has problems and needs further improvement and development. Summary of the Invention

[0009] (I) Purpose of the invention: In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a medical warehousing system specifically for the Internet of Things, thereby improving the efficiency, safety and accuracy of medical warehousing in the process of blood storage and management, and ensuring the quality and safety of medical blood use.

[0010] (II) Technical Solution: To solve the above-mentioned technical problems, this technical solution provides a medical refrigerator storage system based on the Internet of Things, including a refrigerator body, an identity verification unit, an inventory unit, a display input unit, a control unit, and a remote management platform communicatively connected to the control unit. The refrigerator body is used to store stored items. The identity verification unit uses the identity core feature suppression recognition method to verify the identity based on the first identity information of the stored / retrieval items. The inventory unit scans and verifies the quantity and information of the stored items after the storage / retrieval operation. The display input unit is used to input or display relevant information of the stored items. The control unit is connected to the identity verification unit, inventory unit, and display input unit respectively, and receives information from each unit or sends control commands to each unit. The remote management platform is used to remotely assign operating permissions, monitor the main status of the refrigerator, and query historical data.

[0011] The aforementioned IoT-based medical refrigerator storage system includes a refrigerator body comprising a layered storage unit, a temperature monitoring device, and a door lock device. The layered storage unit is divided into a refrigeration layer, a freezing layer, or an ultra-low temperature layer according to temperature requirements, and each layer is equipped with an independent electromagnetic lock and a radio frequency identification ultra-thin antenna. The temperature monitoring device collects the internal temperature in real time through a built-in sensor and feeds it back to the control unit. The door lock device is used to automatically lock the refrigerator body and includes an external electronic lock and an internal layered electromagnetic lock.

[0012] The aforementioned Internet of Things-based medical refrigerator storage system includes an inventory unit that is a radio frequency identification (RFID) intelligent inventory device. This device consists of an RFID reader / writer and ultra-thin ultra-high frequency antennas distributed across each layer of the refrigerator body. After storing / retrieving items, the device triggers a scan to verify the quantity and information of the stored items.

[0013] The aforementioned IoT-based medical refrigerator storage system includes a touchscreen display input unit for recording stored item information, scanning documents, inputting commands, and displaying the internal temperature of the refrigerator, inventory results, and inventory status.

[0014] The aforementioned IoT-based medical refrigerator storage system further includes an anti-cross-scan component, which comprises an outer glass door shielding film and a sealing strip with shielding function, used to prevent cross-scanning problems caused by radio frequency identification signal crosstalk.

[0015] The aforementioned Internet of Things-based medical refrigerator storage system includes an identity verification unit comprising an identity information collection device, which is used to collect the identity information of personnel storing / retrieving stored items to obtain first identity information. The identity information collection device includes at least two of the following: a facial recognition device, an UHF identity card that supports identity recognition, an authorization code that supports authorization, and an account password.

[0016] The aforementioned IoT-based medical refrigerator storage system includes the following steps: The identity verification unit verifies the first identity information based on the first identity information collected by the identity information collection device using a core identity feature suppression recognition method, and outputs the identity verification result of the first identity information. The biometric data in the collected first identity information is decomposed to extract the core features for identity recognition and suppress non-core interference features; the identity verification result is output through a pre-trained feature suppression model.

[0017] The aforementioned IoT-based medical refrigerator storage system involves performing feature decomposition on the biometric data in the collected first identity information, using multi-dimensional feature decomposition to break down the biometric data into multiple feature components, and distinguishing between core features and non-core interference features.

[0018] The aforementioned IoT-based medical refrigerator storage system includes a remote management platform that constructs a feature suppression model and trains the feature suppression model using a dual-task training strategy of accurate identification of identity features and detection of abnormal identity features to obtain a target feature suppression model.

[0019] The aforementioned IoT-based medical refrigerator storage system includes a feature suppression model comprising an initial model, a gradient inversion layer, and an interference feature recognition subnetwork. The initial model includes a core feature extraction network subnetwork and a core task recognition subnetwork. The core feature extraction network subnetwork extracts core features from the first identity information. The core task recognition subnetwork is connected to the output of the core feature extraction network subnetwork and completes the recognition of the core task based on the extracted core features. The gradient reversal layer connects the output of the core feature extraction network subnetwork with the input of the interference feature recognition subnetwork, reducing the extraction of non-core interference features by the core feature extraction network subnetwork; the interference feature recognition subnetwork extracts non-core interference features from the first identity information and provides a supervision signal to the gradient reversal layer, assisting the core feature extraction network subnetwork in suppressing non-core interference features.

[0020] The aforementioned Internet of Things-based medical refrigerator storage system, wherein the information includes radio frequency identification information and physical sensing information; The inventory unit verifies the accuracy of the stored information by forcibly aligning it. The time series of RFID and physical sensor information of the stored items are extracted; a hidden Markov model is used to force the time series of RFID and physical sensor information to be aligned and matched; if the matching degree is lower than the preset matching threshold, an abnormal alarm is triggered.

[0021] The aforementioned IoT-based medical refrigerator storage system includes an inventory unit that further comprises feature point reference marks attached to the surface of the stored items and an image acquisition device deployed inside the refrigerator body. The inventory unit acquires the orientation data of the stored items by photographing the feature point reference marks on the surface of the stored items. The control unit constructs multiple sets of data sequences based on attitude data, traverses and calculates the attitude relationship between the stored items and the main body of the refrigerator, the correlation between the operator's operating attitude and permissions, selects the attitude with the optimal calibration error as the target attitude, and establishes a linkage rule between the target attitude and operating permissions. The remote management platform performs anomaly monitoring and real-time early warning based on target posture and linkage rules. (III) Beneficial Effects: This invention provides a medical refrigerator storage system based on the Internet of Things (IoT). It employs a core feature suppression and recognition method, combined with a feature suppression model containing a gradient inversion layer and an interference feature recognition subnetwork, effectively filtering non-core interference and improving verification accuracy and robustness. It supports multimodal verification such as facial recognition, UHF identification cards, and authorization codes, providing multiple layers of protection against unauthorized operations. It solves the problem of signal crosstalk, improving inventory efficiency and accuracy, and reducing human error. It achieves precise temperature control and layered safety management to prevent cross-contamination. Automatic locking ensures the safety of stored items. The remote management platform enhances the system's intelligence level, facilitating unified management of cross-regional medical resources. The touchscreen display input unit optimizes user experience and operational efficiency. These multiple measures comprehensively protect the safety of stored items, reduce medical risks, and provide a safer, more efficient, and intelligent storage management system for the medical field. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a medical refrigerator storage system based on the Internet of Things according to the present invention; Figure 2 This is a schematic diagram of the feature suppression model in a medical refrigerator storage system based on the Internet of Things according to the present invention; Figure 3 This is a schematic diagram illustrating the steps of an inventory unit in an IoT-based medical refrigerator storage system of the present invention, which verifies the accuracy of the information of the stored items by forcibly aligning the information. Figure 4 This is a schematic diagram of feature point reference marks on the surface of stored items in a medical refrigerator storage system based on the Internet of Things according to the present invention; 100 - Refrigerator body; 200 - Remote management platform. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to preferred embodiments. More details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention can obviously be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions and derivations based on actual application situations without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0024] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that these drawings are for illustrative purposes only and are not drawn to scale, and should not be construed as limiting the actual scope of protection of the present invention.

[0025] A medical refrigerator storage system based on the Internet of Things, such as Figure 1 As shown, it includes a refrigerator body, an authentication unit, an inventory unit, a display input unit, a control unit, and a remote management platform that is communicatively connected to the control unit.

[0026] The refrigerator body is used for storing items. The authentication unit uses a core feature suppression recognition method to authenticate the items based on their initial identity information. The inventory unit scans and verifies the quantity and information of the stored items after storage / retrieval. The display input unit is used to input or display relevant information about the stored items. The control unit is connected to the authentication unit, inventory unit, and display input unit, respectively, and receives information from each unit or sends control commands to each unit. The remote management platform is used to remotely assign operating permissions, monitor the status of the refrigerator body, and query historical data.

[0027] The authentication unit, inventory unit, and display input unit are each connected to the control unit. The control unit is connected to the remote management platform, which can be wireless or wired; no specific limitation is made here. The control unit sends the information obtained by the authentication unit, inventory unit, and display input unit on the refrigerator body to the remote management platform.

[0028] The refrigerator body includes a layered storage unit, a temperature monitoring device, and a door locking device. The layered storage unit is divided into a refrigerator compartment, a freezer compartment, or an ultra-low temperature compartment according to temperature requirements, and each compartment is equipped with an electromagnetic lock and a radio frequency identification (RFID) ultra-thin antenna. The temperature monitoring device is located inside the refrigerator body and collects the internal temperature of the refrigerator body in real time through built-in sensors and sends the data to the control unit. The door locking device is used to automatically lock the refrigerator body and includes an external electronic lock and internal layered electromagnetic locks.

[0029] The inventory unit is a radio frequency identification (RFID) intelligent inventory device, consisting of an RFID reader and ultra-thin ultra-high frequency antennas distributed on each layer of the refrigerator body. After storing / retrieving stored items, it triggers a scan to verify the quantity and information of the stored items.

[0030] The display input unit is located on the surface of the refrigerator body. The display input unit is a touch screen and is used for entering stored information, scanning documents, inputting commands, and displaying the internal temperature of the refrigerator body, inventory results, inventory status, etc.

[0031] The refrigerator body also includes an anti-cross-scan component, which includes a shielding film on the outer glass door of the refrigerator body, a sealing strip with shielding function, a metal shielding partition, and other components that can prevent RFID radio frequency identification signal crosstalk, thereby preventing cross-scanning problems caused by RFID radio frequency identification signal crosstalk. Adjacent layer antennas can also use different frequencies to prevent cross-scanning problems caused by RFID radio frequency identification signal crosstalk.

[0032] The refrigerator body is also equipped with a network communication device, which connects to an external medical information system and remote management platform via a network cable to achieve real-time transmission and synchronization of blood storage / retrieval information, temperature data, and inventory data. The external medical information system includes the information systems of blood banks and hospitals.

[0033] During the blood collection process, the identity verification unit supports authorization code verification: when the medical staff of the blood-using department enters the authorization code approved by the blood transfusion department, the control unit verifies the validity of the authorization code. If the verification is successful, the corresponding layer electromagnetic lock is opened.

[0034] When the RFID smart inventory device detects a discrepancy between the inventory quantity and the expected information, the control unit triggers an abnormal alarm signal, displays the abnormal details on the interactive display device, and sends an abnormal notification to the remote management platform to remind relevant personnel to investigate.

[0035] The identity verification unit includes an identity information collection device, which is used to collect the identity information of the person storing / retrieving the stored items to obtain first identity information. The identity information collection device includes at least two of the following: a facial recognition device, an UHF identification card supporting identity recognition, an authorization code supporting authorization, and an account password. During the blood collection process, the identity verification unit supports authorization code verification: when medical staff in the blood-receiving department enter an authorization code approved by the blood transfusion department, the control unit verifies the validity of the authorization code; if the verification is successful, it controls the corresponding layered electromagnetic lock to open.

[0036] The first identity information includes facial images, fingerprint patterns, iris textures, etc., which serve as the input basis for feature decomposition and suppression.

[0037] Core features refer to the key factors in identity recognition, such as the proportions of facial features, the core pattern of fingerprints, and the topological structure of the iris. Non-core interference features refer to environmental or other changes that affect the judgment in identity recognition, such as changes in facial makeup, edge wear of fingerprints, and light reflection noise in the iris.

[0038] The identity verification unit verifies the first identity information based on the first identity information collected by the identity information collection device using the identity core feature suppression recognition method, and outputs the identity verification result of the first identity information, including the following steps. The biometric data in the collected first identity information is decomposed to extract the core features for identity recognition and suppress non-core interference features; the identity verification result is output through a pre-trained feature suppression model.

[0039] Biometric data includes facial, fingerprint, and iris data; core features include facial proportions, core fingerprint patterns, and iris topology. Non-core interfering features include makeup variations, fingerprint edge wear, and iris light reflection noise.

[0040] The biometric data in the collected initial identity information is subjected to feature decomposition. Multi-dimensional feature decomposition is used to break down the biometric data into multiple feature components, distinguishing between core features and non-core interfering features. Specifically, this includes: Biometric data is preprocessed to obtain standardized feature data; The standardized feature data is divided into multiple dimensions to obtain the decomposed data. After decomposition, the multi-dimensional data is labeled with core features or non-core interference features. The labeled core features and non-core interference features form the label system of the training dataset.

[0041] The labeling system of the training dataset includes biometric datasets labeled with core features and non-core features, such as image samples labeled with the core facial features and samples labeled with the core fingerprint patterns.

[0042] Biometric data is preprocessed to obtain standardized feature data, specifically including grayscale conversion of face images, binarization of fingerprint images, and denoising of iris images.

[0043] The standardized feature data is divided into multiple dimensions to obtain the decomposed data. When the standardized feature data is facial feature data, it is decomposed into the relative positions of facial features, texture details, and illumination reflection. The relative positions of facial features are core features, such as the distance between the eyes and the ratio of the nasolabial folds; texture details are non-core interfering features, such as makeup and beards; and illumination reflection is also a non-core interfering feature.

[0044] When the standardized feature data is fingerprint feature data, it is decomposed into core ridge orientation, edge wear, and edge smudges. Core ridge orientation is the core feature, such as the spiral center of a whorl pattern, while edge wear and smudges are non-core interference features.

[0045] When the standardized feature data is iris feature data, it is decomposed into topological structure and illumination noise. Topological structure is the core feature, such as the iris crypt distribution, while illumination noise is a non-core interference feature.

[0046] After decomposition, the core features or non-core interference features of the multi-dimensional data can be labeled manually or automatically using an automatic labeling tool. Here, it is preferred to use an automatic labeling tool.

[0047] The remote management platform constructs a feature suppression model and trains the feature suppression model using a dual-task training strategy of accurate identity feature recognition and abnormal identity feature detection to obtain the target feature suppression model.

[0048] like Figure 2 As shown, the feature suppression model includes an initial model, a gradient reversal layer, and an interference feature recognition subnetwork. The initial model includes a core feature extraction network subnetwork and a core task recognition subnetwork. The core feature extraction network subnetwork extracts core features from the first identity information. The core task recognition subnetwork is connected to the output of the core feature extraction network subnetwork and completes the recognition of the core task based on the extracted core features. The gradient inversion layer connects the output of the core feature extraction network subnetwork to the input of the interference feature recognition subnetwork, reducing the extraction of non-core interference features by the core feature extraction network subnetwork. The interference feature recognition subnetwork extracts non-core interference features from the first identity information and provides a supervision signal to the gradient inversion layer, assisting the core feature extraction network subnetwork in suppressing non-core interference features. During the training phase, the interference feature recognition subnetwork receives the output of the core feature extraction network through the gradient inversion layer, and its loss is backpropagated to the core feature extraction network through the gradient inversion layer, guiding the core task recognition subnetwork to suppress interference features.

[0049] The core feature extraction network subnetwork takes the decomposed biometric data as input and outputs core feature vectors, such as low-dimensional vectors of facial feature proportions and encoding of fingerprint core patterns. The interference feature recognition subnetwork is connected to the output of the core feature extraction network subnetwork through a gradient inversion layer. It takes the core feature vector of the core feature extraction network subnetwork as input and outputs the recognition result of interference features, such as the presence of makeup or fingerprint edge wear.

[0050] Gradient reversal layers enable the core feature extraction subnetwork to simultaneously learn to preserve core features and suppress non-core interfering features during training. For example, a gradient reversal layer allows the core feature extraction subnetwork to simultaneously address core task recognition and interference feature suppression during training. Taking phoneme recognition as an example, the interference feature suppression focuses on speaker information. The core task recognition subnetwork minimizes the core task loss, such as the phoneme recognition loss, ensuring the extraction of core features. Through the gradient reversal layer, as the loss of the interference feature recognition subnetwork propagates backpropagation, the core feature extraction subnetwork optimizes in the direction of maximizing the error rate of interference feature recognition, thereby suppressing the extraction of interference features and improving the accuracy of the core task.

[0051] Specifically, the gradient inversion layer achieves feature suppression through gradient backpropagation: The structure connects the output of the core feature extraction subnetwork to the input of the interference feature recognition subnetwork through the gradient inversion layer. Forward propagation: The gradient inversion layer does not modify the input features, that is, it multiplies the input features by 1 to ensure that the core features can be normally transmitted to the interference feature recognition subnetwork; Backpropagation: The gradient inversion layer multiplies the loss gradient of the interference feature recognition subnetwork by -1 and then backpropagates it to the core feature extraction subnetwork, forcing the core feature extraction subnetwork to optimize in the direction of maximizing the error rate of interference recognition, thereby suppressing the extraction of interference features. That is, when the core feature extraction subnetwork is optimizing the interference recognition loss, the parameter update direction will be opposite to the needs of the interference recognition task, thereby suppressing the ability to extract interference features and focusing on the core task features.

[0052] The remote management platform constructs a feature suppression model and trains the feature suppression model using a dual-task training strategy of accurate identification of identity features and detection of abnormal identity features. After obtaining the target feature suppression model, the platform outputs the identity verification result through the target feature suppression model, which specifically includes a training phase and an inference phase.

[0053] The training phase is the phase for learning features that suppress interference. The specific training process is as follows. The input is a biometric dataset labeled with core feature labels and distractor feature labels. For example, the core feature label is identity ID, the distractor feature label is "wearing makeup," and the biometric data is a face image.

[0054] The core feature extraction subnetwork extracts features from the input data and outputs intermediate features containing core features and interference features.

[0055] Intermediate features are fed into the interference feature recognition subnetwork through the gradient inversion layer. The interference feature recognition subnetwork outputs the recognition result of the interference features, such as whether makeup is present.

[0056] The matching loss between the core features output by the core task recognition subnetwork and the identity ID, i.e., the core task loss, such as cross-entropy loss, guides the core feature extraction subnetwork to accurately extract core features. The loss between the output of the interference feature recognition subnetwork and the interference label, i.e., the interference task loss, such as cross-entropy loss, forces the core feature extraction subnetwork to suppress interference features during backpropagation through the gradient reversal layer. Because the gradient direction is reversed, the core feature extraction subnetwork will adjust its parameters in the direction that makes the interference recognition incorrect. The total loss is calculated as: Core Task Loss + λ × Interference Task Loss. The model training is completed by minimizing the total loss using gradient descent. Here, λ is a weight parameter used to balance the importance of the core task and the interference task, controlling the strength of interference feature suppression. It can be modified or input via the input unit of the remote management platform. Increasing λ increases it if stronger interference suppression is needed; decreasing λ decreases it if core task accuracy is prioritized. λ is typically a positive number, such as 0.1 ≤ λ ≤ 10. Initially, λ = 1 is set, and the model's performance on the core and interference tasks is observed. If interference suppression is insufficient (e.g., speaker recognition accuracy is too high), λ is increased (e.g., adjusted to 2 or 5). If core task performance deteriorates (e.g., identity verification accuracy decreases), λ is decreased (e.g., adjusted to 0.5 or 0.1). Finally, the optimal λ value is selected to achieve the best balance between core task accuracy and interference suppression.

[0057] Minimizing the total loss using gradient descent involves iteratively adjusting the model parameters to minimize the total loss. The specific steps are as follows. Forward propagation loss calculation: Input the training data into the core feature extraction sub-network to obtain intermediate features; input the intermediate features into the core task recognition sub-network to calculate the core task loss L. core For example, the ID matching loss; intermediate features are input into the interference feature recognition subnetwork via the gradient inversion layer to calculate the interference task loss L. disturb According to the total loss formula L total =L core +λ×L disturb Calculate the total loss.

[0058] Backpropagation calculates the gradient: for the total loss L total Calculate the partial derivatives of the weights of each model parameter, the core feature extraction subnetwork, the core task subnetwork, and the interference recognition submodule to obtain the gradient ▽L. total During backpropagation, the gradient inversion layer multiplies the gradient of the interference task loss by -1 or another preset negative coefficient, causing the parameter update direction of the core feature extraction subnetwork to be opposite to that of the interference recognition submodule. Specifically, for the core task loss L... core Find the partial derivatives of the parameters of the core task recognition subnetwork and the core feature extraction subnetwork to obtain the gradient ▽L. coreThe direction is to minimize the core task loss, that is, to make the core features more accurately match the identity ID; and to minimize the interference task loss L. disturb Find the partial derivatives of the parameters of the interference recognition sub-module, gradient inversion layer, and core feature extraction sub-network to obtain the gradient ▽L. disturb Because the gradient inversion layer multiplies the gradient of the interfering task loss by -1 or a negative coefficient during backpropagation, the gradient direction of the core feature extraction subnetwork becomes one that maximizes the interfering task loss; the gradient of the total loss, ▽Ltotal, is the weighted sum of the gradients of the core task loss and the interfering task loss, including the negative adjustment of the gradient inversion layer, and is given by the formula: ▽L total =▽L core +λ×(-▽L disturb The negative sign indicates that the gradient direction of the interference task loss is reversed. The parameters of the core feature extraction sub-network affect both the core task loss and the interference task loss; therefore, its gradient needs to be a combination of the backpropagation results of both. The parameters of the core task recognition sub-network only affect the core task loss, and its gradient is determined by ▽L. core This is decided separately. The parameters of the interference identification submodule only affect the interference task loss, and the gradient is determined by ▽L. disturb Independently, the gradient inversion layer does not affect the gradient direction of the interference identification submodule itself.

[0059] Parameter update: Update model parameters θ using gradient descent. new =θ old -α×▽L total Repeat the forward propagation, back propagation, and parameter update steps until the total loss converges, such as when the loss value no longer decreases significantly, or when the preset number of iterations is reached. θ new The updated model parameters; θ old The initial values ​​are the model parameters before the update, randomly initialized (either normally or uniformly distributed, or provided by a pre-trained model), and in iterations, they are the parameters from the previous update; α is the learning rate, controlling the step size of parameter updates and determining the magnitude of parameter adjustment in each iteration, and can be a preset value; ▽L total The gradient, or partial derivative of the total loss with respect to the current model parameters, indicates the direction of parameter updates and is calculated via backpropagation: The gradient is the partial derivative of the total loss with respect to the current model parameters. total Calculate the partial derivatives of each parameter, and combine the gradients of the core task loss and the interference task loss, including the negative adjustment with the gradient reversal layer.

[0060] Model convergence criterion: When the core task loss L core Small enough, meaning the core features match the identity ID accurately, and the interference task loss L disturb The model training is complete when the size is large enough, i.e., when the interfering features are effectively suppressed.

[0061] The inference phase is the phase where the authentication result is output. After training, the interference feature recognition subnetwork no longer participates in inference; only the core feature extraction subnetwork is used for 90% of the work. The specific process is as follows. Input the biometric data to be verified, such as a user's facial image.

[0062] The data to be verified is fed into the core feature extraction sub-network, which outputs core features that suppress interference features, such as facial feature proportions after removing the influence of makeup.

[0063] The similarity between the core features and the identity templates in the database is calculated, specifically using cosine similarity calculation.

[0064] If the similarity is greater than or equal to the preset similarity threshold, the verification passes; otherwise, the verification fails. The preset similarity threshold can be 90%.

[0065] A preset core feature matching threshold, i.e., a similarity threshold of 90%, such as the similarity threshold for core facial features or the matching threshold for core fingerprint patterns, is used to determine whether the core features after suppressing interfering features meet the identity verification requirements. Through feature decomposition, core feature preservation, non-core feature suppression, and identity verification, the system effectively resists the risk of identity impersonation caused by tampering with non-core features, such as makeup or using fake fingerprints.

[0066] After storing / retrieving items, the inventory unit scans and verifies the quantity and information of the stored items. This information includes RFID information and physical sensor information. RFID information refers to the information of the stored items identified by the RFID smart inventory device, while physical sensor information refers to information about the physical characteristics of the stored items, such as color, shape, and weight.

[0067] The inventory unit verifies the accuracy of stored item information by forcibly aligning it: it extracts time series data from the RFID and physical sensor information of the stored items; it uses a Hidden Markov Model to forcibly align and match the time series data of the RFID and physical sensor information; if the matching degree is lower than a preset matching threshold, an abnormal alarm is triggered. Specifically, such as... Figure 3 As shown, time series extraction is performed on the RFID and physical sensor information of the stored items. A Hidden Markov Model (HMM) is constructed based on the time series of RFID and physical sensor information, with RFID information as the observation sequence and physical sensor information as the state sequence. The time series of RFID and physical sensor information are input into the HMM, and the alignment probability between the two is calculated using the Viterbi algorithm to obtain the matching degree. If the matching degree is lower than a preset matching threshold, an abnormal alarm is triggered, indicating that the stored item information may have been tampered with. This solves the problem of inconsistent physical sensor information when stored item information is tampered with, such as RFID tag replacement or expiration date modification.

[0068] The time series extraction includes: Synchronize the timestamp of RFID information with the sampling time of physical sensing information to ensure that the two are consistent in time dimension; Encode the key fields in the RFID information to generate a structured time series. The key fields can be blood identification codes, expiration dates, etc. Physical sensor information, such as temperature and location, is preprocessed to generate continuous time series. Preprocessing includes noise reduction and interpolation.

[0069] The construction of the hidden Markov model includes: Define the observation state of RFID information and the hidden state of physical sensing information; the observation state of RFID information includes ID matching, normal validity period, etc., and the hidden state of physical sensing information includes stable temperature, fixed position, etc. The transition probability matrix and transmission probability matrix of the Hidden Markov Model are trained based on historical data, where the transition probability represents the change pattern of the physical state and the transmission probability represents the probability of the physical state corresponding to the radio frequency identification information. Initialize the initial state probability distribution of the Hidden Markov Model to ensure model convergence.

[0070] This section uses the evaluation of speech data as an example to explain the Hidden Markov Model in detail.

[0071] In the evaluation of speech data, Hidden Markov Models (HMMs) are used to force alignment between the time-series features of the evaluation speech data and the corresponding phonetic unit sequences of the evaluation text. The construction method mainly includes the following steps: Determine the phonetic unit sequence of the evaluation resource: Obtain the evaluation resource, such as the evaluation text, and find the phonetic unit sequence, such as the phoneme sequence, corresponding to each evaluation word by consulting a phonetic dictionary. For example, the phonetic unit sequence corresponding to the evaluation text "apple" is / æ / , / p / , / This sequence of phonetic units is the state sequence of the Hidden Markov Model, i.e., the hidden states.

[0072] Extracting time-series features of the evaluation speech: The evaluation speech data is segmented into frames to obtain multiple evaluation speech frames, such as 25ms per frame with a frame shift of 10ms. Speech features, such as Mel-frequency cepstral coefficients (MFCCs), are extracted from each speech frame to form an observation sequence, i.e., observation values.

[0073] Constructing a Hidden Markov Model: Based on the sequence of articulatory units (i.e., the state sequence) and the sequence of speech features (i.e., the observation sequence), a Hidden Markov Model is constructed to achieve forced alignment between speech frames and articulatory units.

[0074] A Hidden Markov Model consists of a set of states, a set of observations, a transition probability matrix, an emission probability matrix, and an initial state probability distribution.

[0075] A state set refers to the sequence of phonetic units corresponding to the evaluation text. Each state represents a phonetic unit, such as a phoneme. For example, the state set for the evaluation text "apple" is { / æ / , / p / , / ...} The number of states is equal to the length of the vocal unit sequence, for example, 3 states.

[0076] The observation set refers to the time-series features corresponding to the evaluation speech. Each observation is a feature vector of a speech frame, such as Mel-frequency cepstral coefficients (MFCC) features. The number of observations is equal to the number of frames of the evaluation speech; for example, 100 frames of speech correspond to 100 observations.

[0077] The transition probability matrix represents the probability of transitioning from one state to another. For example, the probability of transitioning from state / æ / to state / p / is denoted as a_ij, where i is the current state and j is the next state. For state transitions in a sequence of articulatory units, it is generally assumed that each state can only transition to the next state or loop within itself. For example, state / æ / can only transition to / p / or remain at / æ / , with a low probability of looping. Transition probabilities can be obtained by statistically analyzing the state transition frequencies in a large amount of training data.

[0078] The emission probability matrix represents the probability of generating a specific observation in a given state. For example, the probability of the state / æ / generating the speech frame feature x_t is denoted as b_j(x_t), where j is the state and x_t is the observation. The emission probability is the probability that each speech frame belongs to each phoneme identified by the target acoustic model. For example, if the target acoustic model outputs a probability of 0.8 for the speech in frame t belonging to the phoneme / æ / , then the emission probability b_ / æ / (x_t) = 0.8. The dimension of the emission probability matrix is ​​the number of states × the dimension of the observation features, and each element represents the probability of the state generating the observation.

[0079] The initial state probability distribution represents the probability distribution of the model starting from the initial state. For example, the probability of the initial state being the first state in a phonetic unit sequence is 1.0, and the probability of other states is 0. For a phonetic unit sequence, the initial state is fixed at the first phonetic unit, such as / æ / , so the initial state probability distribution is π=[1.0,0,0], and the number of states is assumed to be 3. If there are multiple possible initial states, such as when the evaluation text has multiple possible pronunciation beginnings, the initial probability is calculated based on the language model or training data.

[0080] The convergence of Hidden Markov Models is guaranteed in the following ways.

[0081] Reasonable state transition constraints: Limit the scope of state transitions, such as only transitioning in the order of the vocal unit sequence, to avoid the model getting stuck in meaningless state loops.

[0082] Accurate emission probability estimation: The probability of the sounding unit provided by the target acoustic model ensures a reliable correspondence between the observations and the state.

[0083] The Viterbi Algorithm: In the process of forced alignment, the Viterbi algorithm is used to find the optimal state path, that is, the alignment result between the speech unit sequence and the speech frame, to ensure that a globally optimal solution is found under given model parameters, thereby ensuring model convergence.

[0084] The forced alignment matching includes: The time series of RFID information is used as the observation sequence O, and the time series of physical sensing information is used as the state sequence S. The alignment result is obtained by calculating the maximum probability path of the observation sequence O under the state sequence S using the Viterbi algorithm. The matching degree is calculated based on the alignment results. The formula is: Matching degree = (Number of time points with successful alignment / Total number of time points) × 100%.

[0085] The Viterbi algorithm is used to calculate the maximum probability path of the observation sequence O in the state sequence S, including: initialization: Let the initial state probability distribution be π, where π i This represents the probability of being in state i at the initial moment; Calculate the initial probability δ1(i) of the first observation o1 in the observation sequence O under each state i. i ×b i (o1), where b i (o1) is the emission probability of generating observation value o1 in state i; Initialize the path record array ψ1(i)=0 to record the previous optimal state for each state.

[0086] Recursion: For the t-th observation value o in the observation sequence t (t≥2), iterate through all states j and calculate the maximum probability δ of transitioning from state i to state j. t (j)=max i [δ t-1 (i)×a ij ]×b j (o t ), where a ij Let be the transition probability from state i to state j; Record the previous optimal state ψ for each state j t(j)=argmax i [δ t-1 (i)×a ij ].

[0087] termination: Calculate the last observation o in the observation sequence T The maximum probability P * =max i δ T (i), where T is the length of the observation sequence; Determine the optimal state i of the last state T * =argmax i δ T (i).

[0088] Path backtracking: From the last state i T * Begin by traversing the path record array ψ in reverse order. t Determine the optimal state at each time step in turn. state i t-1 * =ψ t (i t * ); Reverse the state sequence obtained from backtracking to obtain the maximum probability path I of the observation sequence O under the state sequence S. * =[i1 * i2 * ,……,i T * ].

[0089] The state sequence S is the sequence of phonetic units corresponding to the evaluation text, and the observation sequence O is the sequence of speech features of the evaluation speech data.

[0090] The transition probability a ij and the probability of emission b i (o t The target acoustic model is provided by the target acoustic model, which is learned from training samples.

[0091] In the path backtracking step, the state sequence obtained by backtracking is arranged in reverse order to obtain the alignment result between the observation sequence and the state sequence, wherein the alignment result is used to indicate the optimal state corresponding to each observation value.

[0092] The maximum probability path I * This is used to achieve forced alignment between speech frames and speech units, where the forced-aligned speech segments are used for speech evaluation.

[0093] The abnormal alarm triggering includes: The preset matching threshold is dynamically adjusted according to the application scenario. For example, the matching threshold is set to 95% for blood storage scenarios and 85% for ordinary item storage scenarios. Alarm methods include audible and visual alarms, SMS notifications, or system log recordings, while also outputting specific abnormal information, such as a mismatch between the RFID tag identification code and the location trajectory of the physical sensor information.

[0094] The inventory unit includes a data storage module and a model update module. The data storage module stores the time series of RFID information and physical sensor information in real time for Hidden Markov Model (HMM) training and anomaly tracing. The model update module periodically retrains the HMM based on new data, optimizes the transition probability matrix and emission probability matrix, and improves matching accuracy.

[0095] The physical sensing information includes, but is not limited to: real-time temperature values ​​and rate of change from temperature sensors; latitude and longitude coordinates and movement speed from position sensors; and pressure values ​​and trends from pressure sensors.

[0096] After the abnormal alarm is triggered, the refrigerator body automatically performs the following operations: locks the RFID tag of the stored items to prevent further tampering; generates an abnormal report, including a comparative analysis of RFID information and physical sensor information; and notifies the administrator to conduct manual verification to ensure the safety of the stored items.

[0097] The remote management platform adopts a multi-feature scoring strategy for the authenticity of stored items, extracting multi-dimensional features from the RFID information, physical sensor information, and environmental feature information of the stored items; calculating the similarity of each feature with a preset standard feature library, and obtaining the authenticity score by weighted summation; if the score is lower than a preset score threshold, the stored item information is determined to be unauthentic, thus solving the problem of stored items being replaced or information being falsified.

[0098] The remote management platform collects user operation behavior characteristics in real time, such as operation speed, common operation sequence, and click location distribution; extracts core behavior characteristics, such as fixed patterns of common operation sequences, and suppresses non-core behavior characteristics, such as occasional operation errors; when the matching degree between core behavior characteristics and historical patterns is lower than a preset operation threshold, the remote management platform sends a command to the identity verification unit to upgrade the operation permission verification level, which can add secondary authorization code verification to solve the problem of unauthorized operation after identity theft.

[0099] The identity verification unit adopts a multimodal feature fusion suppression strategy, fusing multimodal features from face, fingerprint, and UHF identity card. When training the fusion model, a gradient reversal layer is introduced to suppress interference from a single modality while retaining complementary features of the multimodality. During verification, the model outputs the verification result through a comprehensive judgment of complementary features, thus solving the problem that single-modal identity verification is easily cracked.

[0100] The remote management platform employs a training strategy for a storage anomaly detection model: a dual-task model is trained. The first task is to identify normal information patterns in the stored goods, such as the correspondence between the expiration date and storage time of blood, and the normal range of temperature changes. The second task is to identify abnormal information patterns, such as expiration date tampering and sudden temperature rises. A gradient inversion layer is introduced into the model to suppress the interference of abnormal patterns when identifying normal information in the first task, and to retain the recognizability of abnormal patterns when detecting anomalies in the second task. After training, the platform can accurately identify normal storage status and quickly detect information tampering or abnormal status during monitoring, solving the problem of difficulty in real-time detection of abnormal information in stored goods.

[0101] The control unit uses a hierarchical index to construct hierarchical index data for blood inventory data, which is built by storage area, blood type, expiration date, etc. When performing inventory query or inventory count operations, the target data range is first filtered from the high-level index to remove redundant data from irrelevant levels, thereby improving data processing speed.

[0102] The remote management platform employs a monitoring data area elimination mechanism. Based on the hierarchical inventory index data, it eliminates the areas of interest for refrigerator status data such as temperature, door lock, and inventory changes. It only transmits data from the currently monitored target storage layer or refrigerators with abnormal states, while eliminating normal state data that is not being monitored. This solves the problems of excessive bandwidth consumption and high latency in remote monitoring data transmission.

[0103] The authentication unit improves the authentication speed by using sparse loading of permissions. Specifically, during the authentication process, the user role is first determined, such as whether the user is a blood supply department staff member, a blood use department staff member, or an administrator. Only the operation permission data corresponding to that role is loaded, and the permission data for other roles is not loaded for the time being. This avoids the problem of slow authentication speed caused by loading all data during permission authentication.

[0104] The inventory unit uses a batch counting method to count the number of consecutive inventory requests for the same storage layer. When the number of requests reaches a preset threshold, they are merged into a single batch scan inventory, reducing repetitive scanning operations, avoiding inefficiency caused by frequent and scattered inventory counts, and improving inventory throughput.

[0105] When the control unit detects an abnormal temperature or a discrepancy in inventory quantity, it first triggers an audible and visual alarm on the display input unit. If the problem is not resolved within a preset time, it automatically rolls back to the remote management platform to send an anomaly notification until the anomaly is confirmed and resolved.

[0106] During authentication, the authentication unit first determines whether the user's identity information is stored in the local cache. If it exists, the cached data is used directly to complete the authentication. If it does not exist, the remote management platform's database query is triggered, and the result is updated to the local control unit's cache, thus avoiding the delay problem caused by the same user accessing the database multiple times.

[0107] When executing a blood traceability request, the remote management platform loads only the key traceability node data within the requested blood ID and time range, such as blood storage / removal operation records and temperature abnormality records, while non-key node data is not loaded. If the user needs more detailed data, subsequent dynamic loading is triggered, thus avoiding the slow traceability response problem caused by loading all historical data.

[0108] The inventory unit is also used to acquire the posture data of the stored items. The remote management platform monitors the refrigerator body containing the stored items for anomalies based on the posture data. Specifically, the inventory unit also includes feature point reference markers attached to the surface of the stored items and an image acquisition device deployed inside the refrigerator body. The inventory unit acquires the posture data of the stored items by photographing the feature point reference markers on the surface of the stored items. Image acquisition devices are installed on each layer inside the refrigerator body. The control unit constructs multiple sets of data sequences based on the posture data, iterates and calculates the posture relationship of the stored items relative to the refrigerator body, the correlation between the operator's posture and permissions, selects the posture with the optimal calibration error as the target posture, and establishes a linkage rule between the target posture and operation permissions. The remote management platform performs anomaly monitoring and real-time early warning based on the target posture and linkage rules.

[0109] The feature point reference markers on the surface of the storage object are checkerboard reference markers. The storage object surface is affixed with checkerboard reference markers including corner points. The inventory unit obtains the orientation data of the storage object by identifying the corner point positions of the checkerboard reference markers. The feature point reference markers on the storage object surface can be set separately on different surfaces of the storage object to avoid obstruction of the feature point reference markers.

[0110] To balance attitude calibration accuracy, computational complexity, and adaptability to real-world scenarios, the checkerboard reference markers preferentially adopt a 9×7 corner point pattern. The three-dimensional position information of each corner point is pre-stored in the control unit, and the checker unit calculates the attitude data of the stored object using the PNP (Perspective-n-Point) algorithm.

[0111] A 9x7 grid pattern refers to the reference markings for a checkerboard pattern, composed of alternating black and white squares, such as... Figure 4 As shown. Corner points are the vertices of each cell. Since the number of corner points = the number of cells + 1, the number of cells on the chessboard is 8 × 6.

[0112] The image acquisition device includes a positioning device, which is fixed to the image acquisition device, or the image acquisition device is fixed to the positioning device. There is no specific limitation here. It should be noted that the image acquisition device and the positioning device are mutually fixed, and they must satisfy a rigid body relationship, that is, their relative position and orientation remain unchanged, ensuring the relative orientation of the positioning device and the image acquisition device is stable when capturing images. Specifically, the positioning device and the image acquisition device are installed on each layer inside the refrigerator body.

[0113] The image acquisition device is used to capture images of feature point reference marks inside the refrigerator body, including the surface of the stored items, to obtain the captured data.

[0114] The positioning device diagram is used to acquire the orientation of the positioning device within its own coordinate system during image capture. The positioning device space refers to a three-dimensional local coordinate system established with the positioning device as its origin, or the positioning device's own reference coordinate system, used to describe the spatial state of the positioning device itself. The orientation information refers to the positioning device's pose information within its own coordinate system, typically including: translation parameters, the three-dimensional position of the positioning device in the coordinate system, such as X, Y, and Z coordinates; and rotation parameters, the three-dimensional rotation state of the positioning device, such as rotation matrices, quaternions, or Euler angles. The overall orientation information of the positioning device within its own coordinate system represents the spatial orientation of the positioning device in its own reference system, and is the core foundational data for subsequent calculations of the relative orientation of the stored object.

[0115] When the image acquisition device captures data, the orientation of the positioning device in the positioning device space is obtained through the following steps.

[0116] Real-time data acquisition: The positioning device acquires its own attitude data in the positioning device space in real time through built-in sensors, such as an IMU inertial measurement unit and an optical tracking module; Synchronous recording: At the same time that the image acquisition device captures the data, the positioning device sends the current attitude information to the control unit; Data association: The control unit binds the attitude information with the corresponding shooting data to form a shooting information for subsequent attitude calibration calculation. The shooting information includes shooting data containing the feature point reference mark image of the storage surface, and attitude information of the positioning device.

[0117] The inventory unit acquires N images by using an image acquisition device installed inside the refrigerator body. Each image includes a reference mark image of feature points on the surface of the stored item and the attitude information of the positioning device in the positioning device space during the shooting. N is a positive integer.

[0118] The control unit constructs M shooting data groups based on the N shooting information. The number of shooting information in any shooting data group is H, where H∈[Z,N]. Z is the lower limit threshold of the number of shooting information H in the shooting data group, and the lower limit threshold Z of the number of shooting information H is dynamically adjusted. When the frequency of changes in the attitude of the stored item is higher than a preset frequency threshold, such as when the preset frequency threshold is 10 and the stored item is retrieved or placed ≥10 times per day, Z decreases from 5 to 3 to increase the number of data groups and improve the accuracy of attitude calculation. The initial value of Z is preset based on the minimum data requirements of the attitude estimation algorithm and the robustness requirements in engineering practice.

[0119] The lower limit threshold Z of the number of captured information H is dynamically adjusted. When the frequency of changes in the object's posture exceeds a preset frequency threshold, Z is lowered. Increasing the number of captured data sets M: As Z decreases, the lower limit of H decreases, meaning each captured data set can contain less captured information. Therefore, based on N captured information sets, more M captured data sets can be constructed. For example, when N=20, Z=5 → H≥5 → M is at most C(20,5)=15504; Z=3 → H≥3 → M is at most C(20,3)=1140+...and more. When posture changes frequently, more M captured data sets can traverse more combinations of captured information, capturing feature point changes under different postures of the object. This allows for finding a more accurate target relative posture in subsequent steps by traversing M captured data sets to select the optimal calibration result. After lowering Z, the H of the captured data sets is smaller, shortening the calculation time for a single posture set and enabling faster response to frequent changes in the object's posture. Simultaneously, more captured data sets also reduce the impact of individual abnormal captured information sets on the overall calibration result.

[0120] The control unit traverses the M groups of shooting data, and for the H shooting information in the m-th shooting data group, it calculates the m-th posture of the stored item relative to the current layer of the refrigerator body using the PNP algorithm, as well as the relative relationship between the operator's posture and permissions, where m∈[1,M].

[0121] The m-th pose of the stored item relative to the current layer of the refrigerator body is calculated using the PNP algorithm, including: The three-dimensional position information of each corner point in the feature point reference mark set on the surface of the storage object in the control unit is obtained in the current layer space of the refrigerator body; the three-dimensional position information is calculated by the physical dimensions of the current layer of the refrigerator body and the layout of the feature point reference mark. In the shooting data corresponding to the m-th shooting data group, the two-dimensional position information of each corner point of the feature point reference marker is obtained; the two-dimensional position information can be obtained by corner detection algorithms such as OpenCV's checkerboard corner detection. Based on the mapping relationship between the three-dimensional position information and the two-dimensional position information, the PNP algorithm is used to solve the m-th pose of the stored item relative to the refrigerator layer.

[0122] The PNP algorithm specifically includes the following process.

[0123] Step S10, Input parameters: three-dimensional position information, two-dimensional position information and intrinsic parameters of the image acquisition device. The intrinsic parameters of the image acquisition device include the focal length Fx / Fy, optical center Center, distortion coefficients K1 / K2, etc. of the camera device. Specifically, they can be obtained and stored in advance through Zhang Zhengyou calibration method.

[0124] Step S20, Solve, specifically including the following steps.

[0125] Step S21: Establish corner point correspondence. Extract the two-dimensional corner point coordinates of the feature point reference mark from the shooting data. Associate the three-dimensional coordinates of the feature point reference mark corner point in the current layer space of the refrigerator body to form a set of three-dimensional and two-dimensional point pairs. This process requires at least 4 pairs of non-coplanar points, i.e., H≥Z≥4, to ensure stable convergence of the algorithm.

[0126] Step S22: Constructing the perspective projection model. Using the intrinsic parameters of the camera device, establish the perspective projection relationship from 3D points to 2D image points. Where s is the scale factor, a scalar, referring to the depth scaling factor when a 3D point is projected onto a 2D image. It represents the influence of the distance from the 3D point to the optical center of the camera device on the projected size. The greater the distance, the smaller s is, and the smaller the projected pixel is. In the PNP algorithm solution process, it is used together with the rotation matrix R and the translation vector t as unknowns, and is calculated by minimizing the error between the 3D point projection and the 2D pixel; (u,v) are the 2D pixel coordinates, referring to the pixel positions of the feature point reference marker corner points in the captured image, that is, the projection positions of the 3D spatial points on the imaging plane of the camera device. They are extracted from the captured data through corner detection algorithms, such as OpenCV's checkerboard corner detection, that is, the 2D position information of each corner point in the captured image; K is the intrinsic parameter matrix of the image acquisition device, a matrix used to describe the optical characteristics of the camera device itself, in the form of Fx / Fy is the focal length, Cx / Cy is the optical center coordinates, and distortion coefficients K1 / K2 need to be added if distortion is considered. These are pre-calibrated using the Zhang Zhengyou calibration method, and the calibration results are stored in the control unit. The R rotation matrix and t translation vector are the attitude parameters of the stored object relative to the current layer of the refrigerator. The R rotation matrix is ​​a 3×3 orthogonal matrix used to describe the rotation relationship from the coordinate system of the 3D point to the coordinate system of the image acquisition device, representing the attitude rotation of the 3D object relative to the image acquisition device. It is solved by the PNP algorithm, using the correspondence between the 3D point (X,Y,Z) and the 2D point (u,v), combined with the intrinsic parameter K. The t translation vector is a 3×1 vector used to describe the translation relationship from the coordinate system of the 3D point to the coordinate system of the image acquisition device, representing the position offset of the 3D object relative to the camera device. (X,Y,Z) are the 3D spatial coordinates, which are the 3D coordinates of the feature point reference marker corner point in the target space, that is, the real spatial position of the feature point on the 3D object, used to establish the spatial reference of the projection.

[0127] Step S23, solving for the pose parameters: By minimizing the reprojection error, i.e., the difference between the coordinates of the 3D point projected onto the 2D image and the actual detected coordinates, the rotation matrix R and the translation vector t are solved. π is the perspective projection function, and n is the number of corner points. The PNP algorithm is used to solve for the optimal R and t.

[0128] Step S24: Output the attitude result. Combine the obtained rotation matrix R and translation vector t to get the m-th attitude of the stored item relative to the current layer of the refrigerator body.

[0129] The m-th pose of the stored item relative to the current layer of the refrigerator body includes: rotation parameters: rotation matrix (3×3), quaternion or Euler angle, describing the three-dimensional rotation state of the stored item; translation parameters: three-dimensional translation vector (X,Y,Z), describing the position coordinates of the stored item in the current layer space of the refrigerator body; reprojection error: used to evaluate the accuracy of the pose solution, the smaller the error, the more reliable the result.

[0130] The relative relationship between staff operating postures and permissions includes: The refrigerator's built-in camera or inertial sensor collects information about the operator's posture during operation. The operation posture information is compared with a preset permission posture library, which stores legal operation postures corresponding to different permission levels, such as administrator posture and ordinary user posture. If the operation posture information matches the posture corresponding to a certain permission level in the permission posture library, then the corresponding operation permission is granted, such as opening the refrigerator layer or modifying the stored item information.

[0131] By understanding the relationship between staff operating postures and permissions, dynamic verification of operating permissions can be achieved, preventing unauthorized operations and improving the security and compliance of storage management.

[0132] The remote management platform selects the posture with the smallest calibration error as the target posture and uses a dual error threshold judgment: when the calibration error between the storage posture and the standard posture is ≤5 pixels and the matching error between the operation posture and the permission posture is ≤10 degrees, it is judged as a legal state.

[0133] When the operating posture exceeds the legal range, such as when the refrigerator door opening angle is greater than 90 degrees, the control unit intercepts the operating permission and locks the electromagnetic lock of the corresponding layered storage unit.

[0134] The remote management platform monitors the refrigerator door's posture data in real time. When an abnormal posture is detected, such as the refrigerator door not being closed tightly, or the opening angle being less than 5 degrees and lasting for more than 30 seconds, an early warning message is immediately sent to the administrator.

[0135] The control unit periodically acquires K test data, including test images and the posture of the positioning device, selects L data with a test reference error > 15 pixels and adds them to the captured data, and iteratively updates the posture model. K and L are positive integers.

[0136] The posture model is trained by introducing a gradient reversal layer to suppress non-core interfering postures, such as slight hand tremors of the staff, while retaining core operational postures, such as the opening angle of the refrigerator door and the gesture of picking up items, thereby improving the robustness of the model.

[0137] From the perspective of the core requirements of the pose estimation algorithm, the PNP algorithm needs a sufficient number of 3D and 2D feature point correspondences to ensure the stability of pose estimation. A 9×7 corner point pattern provides 63 feature points (9 horizontally × 7 vertically), far exceeding the minimum requirement of the PNP algorithm. Typically, there are ≥4 feature points, effectively reducing the impact of single-point position errors on the overall pose calculation. If the number of corner points is too small, such as 3×3=9 points, the sparse feature points will lead to ambiguity in the pose solution, causing multiple solutions and making it impossible to accurately distinguish similar poses. If the number is too large, such as 15×10=150 points, it will increase the computational burden and false detection rate of corner point detection, especially in low-resolution images.

[0138] From the perspective of balancing accuracy and computational complexity, attitude calibration requires traversing M sets of captured data. Each set of data needs to process the corner information of H captured data. The number of 9×7 corner points can ensure that each captured image provides sufficient attitude constraints without excessively increasing the computational load or the time cost of corner detection and 3D-2D mapping.

[0139] To ensure a high corner detection rate across multiple shooting angles, data is acquired from N different shooting angles. A 9×7 corner pattern layout covers a reasonable display area of ​​the stored object, ensuring that the camera can detect at least 10-20 corners from different shooting angles, meeting the robustness requirements of the PNP algorithm. If the number of corners is too small, insufficient corners may not be detected at some shooting angles, leading to pose calculation failure; if the number is too large, optical distortion will blur the corner edges, reducing detection accuracy.

[0140] From the perspective of optimizing the accuracy of calibration error calculation, the attitude calibration accuracy is measured by the corner mapping error. The 9×7 corner points provide multi-dimensional error samples. The mapping error of each corner point, i.e., the pixel distance, can be averaged or summed to obtain the overall calibration error, effectively suppressing the influence of noise from individual corner points. If the number of corner points is too small, the error calculation is easily affected by individual outliers, leading to misjudgment of the calibration results; if the number is too large, it will increase the redundancy of the error calculation, failing to significantly improve accuracy while wasting computational resources.

[0141] Fewer than 9×7 corner points, such as 5×5 corner points: Insufficient feature points, unstable pose estimation, prone to multiple solutions or excessive errors. More than 9×7 corner points, such as 12×9 corner points: Computational complexity increases dramatically, false detection rate rises, and edge corner points are prone to 3D position mapping errors due to optical distortion. If the corner point array is circular, the corner point distribution becomes uneven when the shooting angle changes, affecting the accuracy of pose estimation. Therefore, using a 9×7 corner point pattern satisfies the requirement for the number of feature points and is suitable for practical scenarios.

[0142] The inventory unit is also used to acquire the posture data of the stored items. The remote management platform monitors the refrigerator body containing the stored items for anomalies based on the posture data, solving the problems of stored item posture recognition and operation permission linkage. Furthermore, it ensures the accuracy of stored item posture recognition, reduces inventory error rate, improves the unauthorized operation interception rate, reduces remote monitoring latency, and improves the anomaly warning response speed.

[0143] A medical refrigerator storage system based on the Internet of Things adopts at least two multimodal verification methods, such as facial recognition, UHF identity cards, and authorization codes, combined with the identity core feature suppression recognition method. It extracts core features such as facial proportions and fingerprint core patterns through feature decomposition, and suppresses non-core interference such as makeup and lighting, effectively resisting the risk of identity theft.For example, even with makeup or worn fingerprint edges, the system can still accurately verify identity through core features, increasing the accuracy rate to over 90%. The remote management platform monitors operational behavior characteristics in real time, and automatically upgrades the verification level when the matching degree between core behavioral characteristics and historical patterns falls below a threshold, preventing unauthorized operations after identity theft. The refrigerator body adopts layered storage units, divided into refrigeration, freezing, or ultra-low temperature layers according to temperature requirements. Each layer is equipped with an independent electromagnetic lock and an ultra-thin RFID antenna, achieving precise temperature control and secure isolation for different types of items, avoiding cross-contamination or theft of stored items. The inventory unit uses RFID, combined with ultra-thin ultra-high frequency antennas distributed in each layer and anti-scanning components, automatically triggering scanning after storage / retrieval operations. The system employs a Hidden Markov Model to force alignment and matching between RFID and physical sensor information. An alarm is triggered when the matching degree falls below a threshold, resolving issues such as signal crosstalk and information tampering, thus improving inventory efficiency by over 80%. The remote management platform supports remote allocation of operating permissions, monitoring of refrigerator status, and querying of historical data. Through hierarchical inventory indexing and monitoring data area exclusion technology, only data from the target storage layer or abnormal states is transmitted, reducing bandwidth consumption and latency, and enabling unified scheduling of cross-regional medical resources. Real-time data synchronization with external medical information systems via network communication devices constructs a multi-feature scoring strategy for the authenticity of stored items, evaluating RFID information, physical sensor information, and environmental characteristics from multiple dimensions. If the score falls below a threshold... The system determines if information is inaccurate and locks the tag, generating an anomaly report to notify the administrator for verification, ensuring the safety of stored items. The display input unit uses a touchscreen, supporting stored item information entry, document scanning, and command input, and displays temperature, inventory results, and inventory status in real time, simplifying the operation process by more than 50%. The control unit employs sparse loading of permissions, batch counting of inventory, and local cache verification, significantly improving verification speed and data processing efficiency, avoiding repetitive operations and delays. Through automatic locking devices, temperature monitoring devices, and anomaly alarm mechanisms, the system comprehensively ensures the safety of stored items and reduces medical risks. For blood storage scenarios, the system supports authorization code verification; the blood-using department enters the authorization code approved by the transfusion department to unlock the corresponding layer of locks. This invention utilizes a storage authenticity scoring and anomaly detection model to quickly identify issues such as expiration date tampering and label replacement, ensuring the quality and safety of medical blood use. Integrating IoT, AI, and biometric technologies, it pioneers a core identity feature suppression recognition method and a hidden Markov model-driven alignment matching, addressing pain points in traditional medical warehousing systems such as low identity verification security, poor inventory efficiency, and lack of remote control. It provides the medical field with a safer, more efficient, and intelligent storage management system, achieving intelligent management of the entire medical warehousing process. This significantly improves security, accuracy, and efficiency, reduces medical risks, and can be widely applied in hospitals, blood banks, disease control centers, and other scenarios, promoting the intelligent upgrade of medical equipment management.

[0144] The above description illustrates preferred embodiments of the present invention and helps those skilled in the art to more fully understand the technical solution of the present invention. However, these embodiments are merely illustrative and should not be construed as limiting the specific implementation of the present invention to these embodiments. For those skilled in the art, several simple deductions and modifications can be made without departing from the inventive concept, and all such modifications should be considered within the protection scope of the present invention.

Claims

1. A medical refrigerator storage system based on the Internet of Things, characterized in that, It includes a refrigerator body, an authentication unit, an inventory unit, a display input unit, a control unit, and a remote management platform communicatively connected to the control unit. The refrigerator body is used to store stored items. The identity verification unit uses the identity core feature suppression recognition method to verify the identity based on the first identity information of the stored / retrieval items. The inventory unit scans and verifies the quantity and information of the stored items after the storage / retrieval operation. The display input unit is used to input or display relevant information of the stored items. The control unit is connected to the identity verification unit, inventory unit, and display input unit respectively, and receives information from each unit or sends control commands to each unit. The remote management platform is used to remotely assign operating permissions, monitor the main status of the refrigerator, and query historical data. The inventory unit is also used to acquire the posture data of the stored items, and the remote management platform monitors the main body of the refrigerator containing the stored items for abnormalities based on the posture data of the stored items.

2. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The refrigerator body includes a layered storage unit, a temperature monitoring device, and a door lock device; the layered storage unit is divided into a refrigerator layer, a freezer layer, or an ultra-low temperature layer according to temperature requirements, and each layer is equipped with an independent electromagnetic lock and a radio frequency identification ultra-thin antenna; the temperature monitoring device collects the internal temperature in real time through a built-in sensor and feeds it back to the control unit; the door lock device is used to realize the automatic locking of the refrigerator body, including an external electronic lock and an internal layered electromagnetic lock.

3. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The inventory unit is a radio frequency identification (RFID) intelligent inventory device, consisting of an RFID reader and ultra-thin ultra-high frequency antennas distributed on each layer of the refrigerator body. After storing / retrieving stored items, it triggers a scan to verify the quantity and information of the stored items.

4. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The display input unit is a touch screen used for entering stored item information, scanning documents, inputting commands, and displaying the internal temperature of the refrigerator, inventory results, and inventory status.

5. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The refrigerator body also includes an anti-cross-scan component, which includes an outer glass door shielding film and a sealing strip with shielding function, used to prevent cross-scanning problems caused by radio frequency identification signal crosstalk.

6. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The identity verification unit includes an identity information collection device, which is used to collect the identity information of the person storing / retrieving the stored items to obtain first identity information. The identity information collection device includes at least two of the following: a facial recognition device, an UHF identity card that supports identity recognition, an authorization code that supports authorization, and an account password.

7. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The identity verification unit verifies the first identity information based on the first identity information collected by the identity information collection device using the identity core feature suppression recognition method, and outputs the identity verification result of the first identity information, including the following steps. The biometric data in the collected first identity information is decomposed to extract the core features for identity recognition and suppress non-core interference features; the identity verification result is output through a pre-trained feature suppression model.

8. The medical refrigerator storage system based on the Internet of Things according to claim 7, characterized in that, The biometric data in the collected first identity information is decomposed into features. Multi-dimensional feature decomposition is used to break down the biometric data into multiple feature components and distinguish between core features and non-core interference features.

9. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The remote management platform constructs a feature suppression model and trains the feature suppression model using a dual-task training strategy of accurate identity feature recognition and abnormal identity feature detection to obtain the target feature suppression model.

10. A medical refrigerator storage system based on the Internet of Things according to claim 9, characterized in that, The feature suppression model includes an initial model, a gradient reversal layer, and an interference feature recognition subnetwork. The initial model includes a core feature extraction network subnetwork and a core task recognition subnetwork. The core feature extraction network subnetwork extracts core features from the first identity information. The core task recognition subnetwork is connected to the output of the core feature extraction network subnetwork and completes the recognition of the core task based on the extracted core features. The gradient reversal layer connects the output of the core feature extraction network subnetwork with the input of the interference feature recognition subnetwork, reducing the extraction of non-core interference features by the core feature extraction network subnetwork; the interference feature recognition subnetwork extracts non-core interference features from the first identity information and provides a supervision signal to the gradient reversal layer, assisting the core feature extraction network subnetwork in suppressing non-core interference features.

11. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The information includes radio frequency identification information and physical sensing information; The inventory unit verifies the accuracy of the stored information by forcibly aligning it. The time series of RFID and physical sensor information of the stored items are extracted; a hidden Markov model is used to force the time series of RFID and physical sensor information to be aligned and matched; if the matching degree is lower than the preset matching threshold, an abnormal alarm is triggered.

12. The medical refrigerator storage system based on the Internet of Things according to claim 1, characterized in that, The inventory unit also includes feature point reference marks attached to the surface of the stored items, and an image acquisition device deployed inside the refrigerator body. The inventory unit obtains the orientation data of the stored items by photographing the feature point reference marks on the surface of the stored items. The control unit constructs multiple sets of data sequences based on attitude data, traverses and calculates the attitude relationship between the stored items and the main body of the refrigerator, the correlation between the operator's operating attitude and permissions, selects the attitude with the optimal calibration error as the target attitude, and establishes a linkage rule between the target attitude and operating permissions. The remote management platform performs anomaly monitoring and real-time early warning based on target posture and linkage rules.

Images