A storage method and system of an intelligent storage cabinet

Abstract

The application discloses a kind of storage methods and systems of intelligent storage cabinet, method includes: receiving the permission parameter and user biological characteristics of online approval request, generate encrypted digital token;Based on encrypted digital token, using multi-modal biological feature fusion algorithm carries out identity authentication, through space-time attention mechanism association face feature and voiceprint feature, output dynamic identity verification code;According to dynamic identity verification code, activate target storage cabinet group, generate article unique identification set;Based on the difference analysis of article unique identification set and preset inventory list, output optimal cabinet compartment allocation scheme;According to optimal cabinet compartment allocation scheme drive modular cabinet group actuating mechanism adjusts physical structure, synchronously encrypts article state data and writes into distributed account book, triggers edge computing node to update global inventory topology graph.Utilize the embodiment of the application, can realize the intelligentization, self-adaptation and distributed collaborative management of storage cabinet under the premise of guaranteeing security.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent storage, and particularly relates to a storage method and system of an intelligent storage cabinet. BACKGROUND

[0002] Traditional storage cabinet systems usually use mechanical locks or single biometric identification technology for identity verification, and have problems such as extensive permission management, low object identification accuracy, and insufficient space utilization. Existing intelligent storage cabinets mostly rely on fixed RFID or two-dimensional code identification schemes, which are difficult to cope with signal interference in complex electromagnetic environments, and lack the ability to dynamically optimize the space layout in the cabinet. In terms of permission management, the traditional approval process usually uses a centralized database for storage, which has the risk of data tampering and cannot meet the needs of high-security scenarios. SUMMARY

[0003] The purpose of the present application is to provide a storage method and system of an intelligent storage cabinet to solve the problems in the prior art and realize the intelligentization, self-adaptation and distributed collaborative management of the storage cabinet under the premise of ensuring security.

[0004] One embodiment of the present application provides a storage method of an intelligent storage cabinet, which comprises:

[0005] Receiving the permission parameters and user biometric features of the online approval request, verifying the compliance of the approval process through a blockchain consensus algorithm, and generating an encrypted digital token, wherein the blockchain consensus algorithm integrates the approval level topology relationship and the timestamp constraint;

[0006] Based on the encrypted digital token, identity authentication is performed using a multi-modal biometric feature fusion algorithm, the face features and voiceprint features are associated through a spatio-temporal attention mechanism, and a dynamic identity verification code is output;

[0007] According to the dynamic identity verification code, a target storage cabinet group is activated, the items in the cabinet are identified through an anti-interference RFID signal optimization model, the item RFID signals and electromagnetic noise are separated using a spatio-temporal convolution network, and a set of unique item identifiers is generated;

[0008] Based on the difference analysis of the set of unique item identifiers and the preset inventory list, a reinforcement learning algorithm is used to dynamically optimize the space layout of the cabinet group, a three-dimensional space value function is constructed according to the item access frequency and volume characteristics, and an optimal cabinet compartment allocation scheme is output;

[0009] According to the optimal cabinet compartment allocation scheme, the physical structure of the modular cabinet group actuator is adjusted, the item state data is encrypted and written into the distributed ledger at the same time, and the edge computing node updates the global inventory topology graph.

[0010] Optionally, the permission parameters for receiving online approval requests and the user's biometric features are used to verify the compliance of the approval process through a blockchain consensus algorithm to generate an encrypted digital token. The blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints, including:

[0011] Based on the permission parameters of the online approval request, the node permission thresholds in the approval hierarchy topology are parsed to generate a directed graph of the approval process with timestamps.

[0012] The user's biometric hash value is spatiotemporally correlated with the directed graph of the approval process, and the legality of the signature sequence of each level of approval node is verified by the Byzantine fault-tolerant algorithm.

[0013] An approval chain hash tree is constructed based on the verification results. A timestamp-weighted voting mechanism is used to prune conflicting approval branches to generate a compliant approval path hash chain.

[0014] By integrating the compliance approval path hash chain with the user's biometric entropy value, a dynamic digital token containing the validity period of permissions is generated through elliptic curve cryptography.

[0015] Optionally, the step of using a multimodal biometric fusion algorithm for identity authentication based on the encrypted digital token, and linking facial features and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code, includes:

[0016] Based on the encrypted digital token, the pre-stored face image sequence and voiceprint temporal spectrum are extracted, and the spatiotemporal motion features of the face are extracted through a three-dimensional convolutional network to generate a dynamic face feature tensor.

[0017] Mel-cepstral coefficient transform is performed on the time spectrum of the voiceprint, and the time sequence features of the voiceprint are extracted by combining a gated recurrent unit, and the voiceprint feature vector sequence is output.

[0018] Construct a spatiotemporal attention alignment matrix, calculate the cross-modal correlation between dynamic face feature tensors and voiceprint feature vector sequences, and generate a feature alignment weight map;

[0019] Based on the feature alignment weight map, an adaptive gated fusion algorithm is used to dynamically weight and fuse face and voiceprint features, and output a spatiotemporally consistent fused biometric feature vector.

[0020] The fused biometric vectors are input into a two-stream Transformer network, and a one-time dynamic authentication code is generated through a timestamp-bound dynamic hash function.

[0021] Optionally, the step of activating the target storage locker group based on the dynamic authentication code, identifying items inside the locker using an anti-interference RFID signal optimization model, and separating the RFID signal and electromagnetic noise of the items using a spatiotemporal convolutional network to generate a unique set of item identifiers includes:

[0022] The space access tag in the dynamic authentication code is parsed to activate the radio frequency antenna array of the target safe deposit box group and transmit multi-band RFID scanning signals.

[0023] The time-frequency characteristics of the reflected signals inside the cabinet are collected, and the item tag signals and environmental electromagnetic noise are separated by a spatiotemporal convolutional network to construct a denoised signal tensor.

[0024] An adaptive threshold decoding algorithm is used to parse the EPC code in the signal tensor and to reconstruct the overlapping label signals by phase difference separation.

[0025] The EPC code is matched with the pre-stored item fingerprint database using Hamming distance to generate a set of unique item identifiers with confidence scores.

[0026] Optionally, the step of analyzing the difference between the unique identifier set of the items and the preset inventory list, using reinforcement learning algorithms to dynamically optimize the spatial layout of the cabinet groups, constructing a three-dimensional spatial value function based on the item access frequency and volume characteristics, and outputting the optimal cabinet allocation scheme includes:

[0027] Based on the difference between the unique identifier set of items and the inventory list, the volume, weight, and historical access frequency of newly added items are extracted to construct an item feature matrix;

[0028] Define a three-dimensional spatial value function, divide the cabinet into voxel units, and calculate the access efficiency weight, space utilization rate and item correlation weight of each voxel.

[0029] A deep reinforcement learning environment is constructed, and the item feature matrix and three-dimensional spatial value function are input into the policy network. The action sequence of cabinet spatial layout is explored through the proximal policy optimization algorithm.

[0030] The course learning mechanism is designed to first train the optimal positioning strategy for a single item, then gradually increase the number of items to the complexity of the actual scenario, and generate a set of candidate layout schemes.

[0031] Candidate layout schemes are screened based on Pareto front analysis, and the comprehensive score is calculated by the entropy weight-TOPSIS algorithm to output the optimal cabinet allocation layout scheme that balances space utilization and access efficiency.

[0032] Optionally, the step of driving the modular cabinet group actuator to adjust the physical structure according to the optimal cabinet allocation scheme, synchronously encrypting and writing the item status data into the distributed ledger, and triggering the edge computing nodes to update the global inventory topology map includes:

[0033] The optimal cabinet allocation scheme is transformed into a sequence of motion control instructions for the robotic arm, and a collision-free path trajectory is planned using inverse kinematics algorithms.

[0034] Pressure sensor data is collected in real time during the drive process, and the motor torque is dynamically adjusted using a PID closed-loop control algorithm.

[0035] Synchronize the item displacement data with the digital twin model, and generate a state change proof chain using a zero-knowledge proof algorithm;

[0036] Trigger the Byzantine Broadcast Protocol between edge computing nodes to write encrypted change data into the distributed ledger and update the spatiotemporal state of the global inventory topology.

[0037] Another embodiment of this application provides a storage system for an intelligent safe deposit box, the system comprising:

[0038] The receiving module is used to receive the permission parameters and user biometrics of the online approval request, verify the compliance of the approval process through a blockchain consensus algorithm, and generate an encrypted digital token. The blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints.

[0039] The authentication module is used to perform identity authentication based on the encrypted digital token using a multimodal biometric fusion algorithm, and to associate facial features and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code.

[0040] The identification module is used to activate the target storage cabinet group according to the dynamic identity verification code, identify the items in the cabinet through the anti-interference RFID signal optimization model, and use the spatiotemporal convolutional network to separate the RFID signal of the items from the electromagnetic noise, and generate a unique set of item identifiers.

[0041] The optimization module is used to dynamically optimize the cabinet space layout based on the difference analysis between the unique identifier set of the items and the preset inventory list, and to construct a three-dimensional spatial value function based on the item access frequency and volume characteristics, and output the optimal cabinet allocation scheme.

[0042] The execution module is used to drive the modular cabinet group execution mechanism to adjust the physical structure according to the optimal cabinet allocation scheme, synchronously encrypt and write the item status data into the distributed ledger, and trigger the edge computing node to update the global inventory topology map.

[0043] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0044] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0045] Compared with existing technologies, the present invention provides a method for storing intelligent safes, which receives permission parameters and user biometrics from online approval requests to generate an encrypted digital token; based on the encrypted digital token, a multimodal biometric fusion algorithm is used for identity authentication, and facial features and voiceprint features are associated through a spatiotemporal attention mechanism to output a dynamic identity verification code; the target safe group is activated according to the dynamic identity verification code to generate a set of unique item identifiers; based on the difference analysis between the set of unique item identifiers and the preset inventory list, an optimal cabinet allocation scheme is output; according to the optimal cabinet allocation scheme, the modular cabinet group actuator is driven to adjust the physical structure, and the item status data is simultaneously encrypted and written into the distributed ledger, triggering the edge computing nodes to update the global inventory topology map, thereby enabling intelligent, adaptive, and distributed collaborative management of safes while ensuring security. Attached Figure Description

[0046] Figure 1 A hardware structure block diagram of a computer terminal for a method of storing an intelligent safe provided in an embodiment of the present invention;

[0047] Figure 2 A flowchart illustrating a method for storing an intelligent safe in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the storage system of an intelligent safe deposit box provided in an embodiment of the present invention. Detailed Implementation

[0049] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0050] This invention first provides a method for storing smart safes, which can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0051] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a method of storing an intelligent safe provided in an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0052] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any of the safekeeping methods available in the smart safe.

[0053] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0054] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform any kind of safekeeping method of the smart safe.

[0055] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0056] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0057] See Figure 2 The present invention provides a method for storing an intelligent safe deposit box, which may include the following steps:

[0058] S201, Receive the permission parameters and user biometrics of the online approval request, verify the compliance of the approval process through the blockchain consensus algorithm, and generate an encrypted digital token, wherein the blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints;

[0059] Specifically, based on the permission parameters of the online approval request, the node permission thresholds in the approval hierarchy topology can be parsed to generate a directed graph of the approval process with timestamps.

[0060] Online approval requests typically include permission parameters such as approval type (e.g., financial approval, material procurement), approval level (e.g., department manager → director → CEO), approver IDs at each level, and permission thresholds (e.g., maximum amount). During parsing, the permission parameters are first encoded into a structured data model, for example, using JSON format to define the attributes of each approval node.

[0061] {

[0062] "node_id": "A001",

[0063] "role": "Department Manager",

[0064] "threshold": 50000, / / Unit: Yuan

[0065] "next_nodes": ["B002", "B003"]

[0066] }

[0067] The approval hierarchy is modeled using a directed acyclic graph (DAG), where nodes represent approvers and edges represent the approval flow. For example, a procurement approval process might sequentially go through "applicant → department manager → CFO → CEO," forming a chain-like topology. Timestamp constraints are reflected in the operation time records of each approval node, accurate to the millisecond level (Unix timestamp, such as 1630453200000), ensuring a strict temporal order for the approval process.

[0068] Generation process of directed graph of approval process with timestamps:

[0069] Node initialization: Create an initial node based on permission parameters and set the node status to "unprocessed";

[0070] Edge connection: Directed edges are established between nodes based on the next_nodes field. For example, the edge of node A001 points to B002 and B003.

[0071] Timestamp embedding: Each node is appended with a timestamp field to record the approval arrival time (e.g., A001.receive_time=1630453200123).

[0072] Loop detection: Depth-first search (DFS) is used to detect whether there are circular dependencies in the graph. If they are found, an exception is triggered and the graph is interrupted.

[0073] Example: A company defines the permission parameters for a purchase approval request as requiring CEO approval for purchases exceeding 500,000 yuan. The resulting directed graph contains 4 nodes, with edge weights representing approval jump conditions (e.g., pointing to the CEO node when the amount is ≥ 500,000 yuan). Timestamps ensure that if a department manager submits the request at 10:00:00, the CFO must process it after 10:00:00, preventing time sequence reversal.

[0074] The user's biometric hash value is spatiotemporally correlated with the directed graph of the approval process, and the legality of the signature sequence of each level of approval node is verified by the Byzantine fault-tolerant algorithm.

[0075] User biometrics (such as fingerprints and iris scans) are converted into fixed-length hash values ​​(such as 0x3a7d...f2c1) using a hash function (such as SHA-3-256). Spatiotemporal associative encoding binds the hash values ​​to the spatiotemporal attributes (timestamps, node positions) of the approval process:

[0076] Time dimension: Concatenate the hash value with the timestamp of the approval node, for example, hash = SHA3(user_bio_hash|| node_A.timestamp);

[0077] Spatial dimension: Introduce the geographical location (such as GPS coordinates) of the approval node. If the node is approved in a different location, the location hash value must be added.

[0078] The Byzantine Fault Tolerance (BFT) algorithm employs the Practical Byzantine Fault Tolerance (PBFT) protocol to ensure consensus can still be reached even when a maximum of f malicious nodes exist (total number of nodes 3f+1). The verification process is as follows:

[0079] Proposal stage: The master node (such as the first approver) broadcasts the approval request, which includes the user's biometric hash, the directed graph of the approval process, and the timestamp;

[0080] Preparatory phase: Each node verifies the legality of the proposal, checks the validity of the signature (such as ECDSA signature) and the time sequence compliance (incrementing timestamp);

[0081] Preparation and submission phase: Nodes exchange signed messages and reach a consensus after collecting 2f+1 valid responses.

[0082] Signature sequence validity verification:

[0083] Incrementing timestamps: The processing time of each approval node must be later than that of the predecessor node, for example, node_CEO.timestamp > node_CFO.timestamp;

[0084] Signature chain integrity: A chain signature is used, where each approver uses their private key to re-sign the signature of the predecessor node, forming an immutable signature chain.

[0085] Example: During an approval process, it was detected that the timestamp of the CFO node (1630453200500) was earlier than that of the Department Manager node (1630453200123). The BFT consensus determined that the approval sequence was illegal and triggered the process termination.

[0086] An approval chain hash tree is constructed based on the verification results. A timestamp-weighted voting mechanism is used to prune conflicting approval branches to generate a compliant approval path hash chain.

[0087] The approval chain hash tree adopts a Merkle tree structure. The leaf nodes are the hash values ​​of each approval node (including biometric hash, timestamp, and signature), and the non-leaf nodes are the concatenated hashes of the child node hashes.

[0088] When conflicting branches exist (such as approvals by two department managers simultaneously), pruning is required:

[0089] Conflict detection: Traverse the hash tree to identify branches with multiple child nodes at the same level;

[0090] Timestamp-weighted voting: Each branch is assigned a weight, where the weight is the normalized value of the reciprocal of the timestamp (e.g., branch A has a timestamp of t1 and a weight of 1 / t1). Branches with higher total weights are retained.

[0091] Pruning operation: Remove low-weight branches and update the Merkle tree structure.

[0092] Compliance approval path hash chain generation:

[0093] Path backtracking: Backtrack from the leaf node along the Merkle tree path to the root node, collecting all hash values ​​along the path;

[0094] Chain concatenation: Concatenate path hashes in timestamp order to form a hash chain (e.g., H1→H2→RootHash).

[0095] Digital signature: The root hash is signed by the system's master private key, ensuring the immutability of the chain.

[0096] Example: Due to network latency, an approval process generates two conflicting branches (department managers A and B approve simultaneously), with timestamps of 1630453200123 and 1630453200150 respectively. After weighted voting, the branch with the earlier timestamp (A) has a higher weight, its path is preserved, and branch B is pruned.

[0097] By integrating the compliance approval path hash chain with the user's biometric entropy value, a dynamic digital token containing the validity period of permissions is generated through elliptic curve cryptography.

[0098] User biometric entropy values ​​are calculated using Shannon entropy to quantify the randomness of biometric information. For example, the steps for calculating the entropy value of a fingerprint image are as follows:

[0099] Image segmentation: Divide the fingerprint image into 8×8 pixel blocks;

[0100] Gray-scale histogram statistics: Calculate the pixel value distribution of each block;

[0101] Entropy calculation: For each block, use Shannon's formula H = -Σ(p_i * log2(p_i)) and take the average of the entropy values ​​of all blocks.

[0102] Elliptic Curve Cryptography (ECC) uses the secp256k1 curve to generate dynamic digital tokens.

[0103] Key generation: The system generates a public-private key pair (private key d, public key Q = d * G, where G is the base point of the curve);

[0104] Token content encoding: The compliance approval path hash chain (Hex string) is concatenated with the user's biometric entropy value (floating-point number), for example, 0x3a7d...f2c1||0.87;

[0105] Encryption and Signature:

[0106] Encrypt the token content using the recipient's public key (ECIES algorithm);

[0107] Sign the encrypted content using the system's private key (ECDSA algorithm).

[0108] The validity period of permissions is implemented through the time window field in the token:

[0109] Effective time (not_before): Token generation time;

[0110] Expiry time (not_after): Set according to the approval type (e.g., purchase token is valid for 24 hours).

[0111] Example of a dynamic digital token:

[0112] {

[0113] "token_id": "TK202310001",

[0114] "content": "0x3a7d...f2c1||0.87",

[0115] "signature": "0x5b9e...d4a3",

[0116] "not_before": 1630453200000,

[0117] "not_after": 1630539600000

[0118] }

[0119] Verification process:

[0120] Decrypt the token content: Use the recipient's private key to decrypt and obtain the hash chain and entropy value;

[0121] Signature verification: Verify the validity of the signature using the system's public key;

[0122] Timeliness check: The current time must be between not_before and not_after;

[0123] Entropy comparison: Real-time collection of user biometrics to calculate entropy value, the error between the entropy value and the entropy value in the token must be <5%.

[0124] S202, Based on the encrypted digital token, a multimodal biometric fusion algorithm is used for identity authentication, and a dynamic identity verification code is output by associating facial features and voiceprint features through a spatiotemporal attention mechanism.

[0125] Specifically, pre-stored facial image sequences and voiceprint temporal spectra can be extracted based on encrypted digital tokens, and facial spatiotemporal motion features can be extracted through a three-dimensional convolutional network to generate dynamic facial feature tensors;

[0126] During the authentication phase, the system first decrypts the encrypted digital token to obtain the user's pre-stored biometric data. The facial image sequence consists of five consecutive frames of infrared-visible dual-mode images (1280×720 resolution, 30fps), covering minute head movements of the user (such as blinking and nodding). The voiceprint time spectrum is generated from a 1.5-second speech segment collected by a microphone array at a sampling rate of 16kHz, and then pre-emphasized (pre-emphasis coefficient 0.97) and processed by framing (frame length 25ms, frame shift 10ms).

[0127] 3D Convolutional Networks (3D-CNNs) are an improvement on the ResNet-34 architecture, specifically designed for spatiotemporal feature extraction.

[0128] Input layer: 5 frames of images stacked into a 1280×720×5 tensor;

[0129] Convolution kernel configuration: The first layer uses a 7×7×3 spatiotemporal convolution kernel (3 represents the time dimension), with a stride of 2×2×1, to extract macroscopic motion patterns (such as head rotation trajectory).

[0130] Residual module: contains 4 residual blocks, each block contains 3 3×3×3 convolutional layers, the number of channels increases from 64 to 512 in stages, and gradient vanishing is mitigated by skip connections;

[0131] Activation function: LeakyReLU (negative slope 0.1) is used to prevent neuron death;

[0132] Pooling layer: Spatiotemporal max pooling (2×2×1) compresses spatial dimensions while preserving temporal continuity.

[0133] During training, transfer learning was performed using the CASIA-3D face dataset (containing 100,000 dynamic sequences) with an initial learning rate of 0.001, a batch size of 32, and the Adam optimizer (β1=0.9, β2=0.999). The network output is a 512-dimensional dynamic face feature tensor, capturing features such as the frequency of micro-expression muscle movements (e.g., a blink cycle of 0.2Hz) and illumination invariance features.

[0134] Mel-cepstral coefficient transform is performed on the time spectrum of the voiceprint, and the time sequence features of the voiceprint are extracted by combining a gated recurrent unit, and the voiceprint feature vector sequence is output.

[0135] Voiceprint processing employs an architecture combining MFCC (Mel-Frequency Cepstral Coefficients) and GRU (Gated Recurrent Unit):

[0136] MFCC extraction:

[0137] Preprocessing: The speech signal is windowed through a Hanning window, and after being segmented into frames, the short-time energy is calculated, and silent frames below -40dB are filtered out;

[0138] Mel filter bank: 40 triangular filters cover the 0-8kHz range. After logarithmic energy compression, 13-dimensional MFCC coefficients (including 1st and 2nd order differences, for a total of 39 dimensions) are obtained through DCT (Discrete Cosine Transform).

[0139] Example parameters: When a user pronounces "open", the MFCC spectrum shows that coefficients 3 and 7 fluctuate significantly in the 200-400ms range, reflecting the vocal cord vibration pattern.

[0140] GRU time series modeling:

[0141] Network structure: 3-layer bidirectional GRU, 128 hidden units per layer, dropout rate of 0.2 to prevent overfitting;

[0142] Input sequence: Input the 39-dimensional MFCC features in time steps (150 frames), with each time step corresponding to 10ms of speech;

[0143] Gating mechanism: The update gate controls the retention ratio of historical information, and the reset gate determines the degree to which the current input depends on the historical state. For example, when the plosive / p / appears, the reset gate activation value is reduced to decrease the interference of the preceding vowel on the current feature.

[0144] The training data used was the VoxCeleb2 dataset (containing 6000 voices), the loss function was cross-entropy loss, and the learning rate was 0.0001. The network output was a 256-dimensional voiceprint feature vector sequence with a temporal resolution of 10ms, capturing individual features such as formant migration trajectories (e.g., F1 and F2 changes from / i / to / a / ).

[0145] Construct a spatiotemporal attention alignment matrix, calculate the cross-modal correlation between dynamic face feature tensors and voiceprint feature vector sequences, and generate a feature alignment weight map;

[0146] Spatiotemporal attention mechanisms achieve cross-modal alignment through multi-head attention:

[0147] Feature dimension alignment:

[0148] The facial feature tensor (512-dimensional) is projected to 256-dimensionality through a fully connected layer;

[0149] The voiceprint feature sequence (256 dimensions) retains its original dimensions;

[0150] Timeline alignment: Facial features are downsampled over time to 150 time steps (matching the length of the voiceprint sequence).

[0151] Attention calculation:

[0152] Query: Facial feature vector;

[0153] Key and Value: Voiceprint feature vector;

[0154] Scaling dot product attention: Calculate the similarity score at each time step, with a scaling factor of √256≈16;

[0155] Multi-head mechanism: Eight attention heads compute in parallel, each with 32 dimensions, capturing association patterns at different granularities. For example, head 1 focuses on the synchronization between lip movements and speech fundamental frequency, while head 2 detects the association between blink rhythm and sentence pauses.

[0156] Weighted graph generation:

[0157] Softmax normalization: converts similarity scores into a probability distribution;

[0158] Time axis smoothing: Transient noise interference is eliminated through Gaussian filtering (σ=3);

[0159] Example: When a user says "apple", the facial features show a slight tongue tip movement within 300-400ms, which corresponds to a sudden increase in high-frequency energy in the voiceprint features, with an attention weight of 0.92.

[0160] The final output feature alignment weight map has a dimension of 150×150 (time step × time step), with the weight of the diagonal region being higher than that of the off-diagonal region, reflecting the temporal synchronization of the audiovisual signal.

[0161] Based on the feature alignment weight map, an adaptive gated fusion algorithm is used to dynamically weight and fuse face and voiceprint features, and output a spatiotemporally consistent fused biometric feature vector.

[0162] Adaptive gating fusion employs a dynamic weight allocation strategy:

[0163] Gating signal generation:

[0164] The attention weight graph is compressed into a 150-dimensional vector using global average pooling.

[0165] The fusion weights for each time step are generated using a two-layer fully connected network (128→64→1), with the activation function being Sigmoid.

[0166] Example: When the voiceprint signal-to-noise ratio is <15dB, the gating weight is biased towards facial features (weight 0.7); in a quiet environment, the weight is biased towards voiceprint (weight 0.6).

[0167] Feature fusion:

[0168] Time-step weighted: Fusion vector = weight × face feature + (1 - weight) × voiceprint feature;

[0169] Residual connection: Adds original facial features to prevent information loss;

[0170] Layer normalization: stabilizes the training process and accelerates convergence.

[0171] Spatiotemporal consistency constraints:

[0172] Motion continuity loss: Calculate the cosine similarity of feature vectors from adjacent time steps to force a smooth transition;

[0173] Cross-modal consistency loss: Compare the KL divergence of face-voiceprint features within the same time window, with a threshold set to 0.1.

[0174] The final output fused biofeature vector is 256-dimensional with a time step of 150, which can characterize cross-modal correlation features such as "the nonlinear relationship between the degree of mouth corner raising and the fundamental frequency rise during pronunciation".

[0175] The fused biometric vectors are input into a two-stream Transformer network, and a one-time dynamic authentication code is generated through a timestamp-bound dynamic hash function.

[0176] A two-stream Transformer network consists of a temporal stream and a spatial stream:

[0177] Time Flow:

[0178] To handle the temporal dependencies of the fused vectors, a 12-layer Transformer encoder is used.

[0179] Position encoding uses a learnable sinusoidal embedding with a dimension of 256;

[0180] The number of self-attention heads is 8, and the feedforward network dimension is 1024;

[0181] Capture long-range dependencies such as "synchronization between speech word boundaries and facial expression turning points".

[0182] Space Flow:

[0183] By folding the time dimension, we can focus on key biological characteristics through spatial attention.

[0184] Using 4 layers of Transformer, with 4 attention heads;

[0185] Extract statistical features across time steps (such as the mean of voiceprint MFCC and the variance of facial motion).

[0186] Dynamic hash function:

[0187] Input: A 512-dimensional joint feature vector from dual-stream outputs;

[0188] Timestamp binding: Convert the current UTC time (in milliseconds) into 32-bit binary code and concatenate it with the feature vector;

[0189] Hash calculation: The SHA-3-256 algorithm is used, and the first 128 bits are extracted as a one-time verification code;

[0190] Example: The verification code 0x3a7d...f2c1 was generated at 14:05:30.123 on 2023-10-01 and is valid for 30 seconds.

[0191] Security mechanisms:

[0192] Anti-replay attack: The validity period of the verification code is strictly limited and it automatically expires after that.

[0193] Quantum-resistant: Employs a lattice-based encryption algorithm (Kyber-512) to protect the hashing process;

[0194] Real-time blacklist: Security lock is triggered when three consecutive verification failures are detected.

[0195] S203, activate the target storage cabinet group according to the dynamic identity verification code, identify the items in the cabinet through the anti-interference RFID signal optimization model, and use a spatiotemporal convolutional network to separate the RFID signal of the items from electromagnetic noise to generate a unique set of item identifiers.

[0196] Specifically, it can parse the space permission tag in the dynamic authentication code, activate the radio frequency antenna array of the target safe deposit box group, and transmit multi-band RFID scanning signals;

[0197] The dynamic authentication code is a 256-bit encrypted string, where bits 32-64 are the Space Access Tag (SAT), used to identify the range of vault groups that a user can operate. During parsing, an asymmetric decryption algorithm (such as RSA-2048) is used to extract the SAT field. For example, the SAT value "0x3A7D" corresponds to the 12 cabinets (numbered B3-01 to B3-12) in vault group B3.

[0198] When activating the target cabinet's RF antenna array, a multi-band frequency-hopping (FHSS) technique is employed, switching frequencies 50 times per second within the UHF band (860MHz~960MHz) to avoid environmental electromagnetic interference. The antenna array consists of eight directional microstrip patch antennas, each with a radiation angle of 60 degrees and an adjustable power range of 10dBm~30dBm. For example, when signal attenuation due to the cabinet's metal structure is detected, the system automatically increases the transmit power from 20dBm to 28dBm and switches to a dedicated frequency for anti-metal tags (such as 920MHz).

[0199] The transmission mode of multi-band RFID scanning signals is dynamically adjusted according to the cabinet structure:

[0200] Low frequency band (125kHz): Used for close-range (<10cm) object recognition, with high accuracy but weak penetration;

[0201] High frequency band (13.56MHz): Supports ISO / IEC 14443 protocol for reading encrypted item tags;

[0202] Ultra-high frequency band (UHF): Wide coverage (1~5 meters), suitable for batch scanning.

[0203] For example, when a user needs to access high-value items, the system prioritizes high-frequency scanning to ensure data security; while for ordinary items, the UHF band is used to improve efficiency.

[0204] The time-frequency characteristics of the reflected signals inside the cabinet are collected, and the item tag signals and environmental electromagnetic noise are separated by a spatiotemporal convolutional network to construct a denoised signal tensor.

[0205] The reflected signals inside the cabinet were acquired using a software-defined radio (SDR) device (such as the USRP B210), with a sampling rate of 2MHz and a resolution of 12 bits. The raw signal included:

[0206] Valid tag signal: The coded signal backscattered from the RFID tag (such as EPC code);

[0207] Ambient noise: Broadband interference from Wi-Fi routers (2.4GHz), Bluetooth devices (2.4GHz) and industrial motors.

[0208] The architecture of the Spatiotemporal Convolutional Network (ST-CNN) is as follows:

[0209] Input layer: Receives I / Q signals (1024 sampling points each), forming a 2×1024 time-frequency matrix;

[0210] Spatiotemporal convolutional layer: using a 3×3 convolutional kernel, features are extracted simultaneously in the temporal dimension (sampling point sequence) and the spatial dimension (frequency channel), and the activation function is LeakyReLU (negative slope 0.01).

[0211] Attention mechanism: The weights of effective signal frequency bands are dynamically enhanced through the SENet (Squeeze-and-Excitation Network) module;

[0212] Output layer: Generates a denoised signal tensor with dimensions of 1×1024 (time domain waveform) + 64×64 (frequency domain energy spectrum).

[0213] For example, in a certain scan, the signal-to-noise ratio (SNR) of the original signal was -5dB, which was improved to 15dB after ST-CNN processing, effectively separating the three overlapping label signals.

[0214] An adaptive threshold decoding algorithm is used to parse the EPC code in the signal tensor and to reconstruct the overlapping label signals by phase difference separation.

[0215] The core of the adaptive threshold decoding algorithm is to dynamically adjust the signal decision threshold:

[0216] Baseline estimation: Perform a moving average on the time-domain waveform of the signal tensor (window width 100 sampling points) to calculate the background noise level (e.g. -65dBm);

[0217] Peak detection: A dynamic threshold (baseline + 6dB) is set based on the baseline to identify pulse signals that exceed the threshold;

[0218] Encoding and parsing: According to the EPC Gen2 protocol, the pulse width is converted into binary data (e.g., a width of 2μs represents "0" and 4μs represents "1").

[0219] For overlapping tag signals (signal aliasing caused by multiple tags responding simultaneously), the following method is used for separation:

[0220] IQ quadrature demodulation: Extracts the phase information of the signal through Hilbert transform to distinguish the reflection phase difference of different tags;

[0221] Direction of Arrival (DOA) estimation: Calculate the spatial location of the tag using the phase difference of the antenna array, and separate spatially overlapping signals using the MUSIC (Multiple Signal Classification) algorithm;

[0222] Time-domain framing: The aliased signal is windowed and framed (Hamming window, frame length 256 sampling points), and the time-frequency fingerprint of each tag is extracted by short-time Fourier transform (STFT).

[0223] For example, the EPC codes of two tags are “0x5A3F” and “0xB7E2”, respectively. Their reflected signals have a phase difference of 45 degrees and a spatial position difference of 0.3 meters. After separation, the complete codes of the two tags can be accurately deciphered.

[0224] The EPC code is matched with the pre-stored item fingerprint database using Hamming distance to generate a set of unique item identifiers with confidence scores.

[0225] The pre-stored item fingerprint database contains the following information for each item:

[0226] EPC encoding: A 96-bit unique identifier (e.g., “0x3014BEEF00000000000001”).

[0227] Physical fingerprint: Tag backscattering characteristics (such as resonant frequency offset ±50kHz), installation location (cabinet coordinates);

[0228] Historical access records: last access time, operator ID, etc.

[0229] The Hamming distance matching process is as follows:

[0230] Preliminary screening: Calculate the Hamming distance (with different numbers of bits) between the input EPC code and all codes in the fingerprint database, and screen candidate codes with a distance ≤ 4;

[0231] Physical fingerprint verification: Compare the tag's resonant frequency characteristics, with an allowable deviation of ±10kHz;

[0232] Confidence score: Calculated by combining Hamming distance (weight 0.6), frequency matching (weight 0.3), and spatial consistency (weight 0.1).

[0233]

[0234] Where HD is the Hamming distance, Δf is the frequency deviation, and Δd is the position deviation (meters).

[0235] For example, the matching result for a certain EPC code "0x3014BEEF..." is:

[0236] Hamming distance 2 (98% match);

[0237] Frequency deviation 8kHz (matching degree 92%);

[0238] Positional deviation 0.2 meters (matching degree 83%);

[0239] The overall score is calculated as follows: 0.6 × 0.98 + 0.3 × 0.92 + 0.1 × 0.83 = 0.953. The confidence level is marked as "high".

[0240] The final generated set of unique item identifiers is formatted as follows: [

[0242] {EPC: "0x3014BEEF...", Score: 0.953, Position: "B3-05"},

[0243] {EPC: "0x5A3F...", Score: 0.872, Position: "B3-08"}

[0244] ].

[0245] This collection is encapsulated in JSON format and transmitted to the upper-level system for inventory management and access control.

[0246] S204. Based on the difference analysis between the unique identifier set of the items and the preset inventory list, a reinforcement learning algorithm is used to dynamically optimize the spatial layout of the cabinet group. A three-dimensional spatial value function is constructed according to the storage and retrieval frequency and volume characteristics of the items, and the optimal cabinet allocation scheme is output.

[0247] Specifically, based on the differences between the item's unique identifier set and the inventory list, the volume, weight, and historical access frequency of newly added items can be extracted to construct an item feature matrix;

[0248] After comparing the unique identifier set of an item (e.g., the EPC code set identified by RFID) with the pre-set inventory list, the system first filters out newly added items (such as items not registered in the inventory) or items whose location has changed. For each newly added item, the following features are extracted:

[0249] Physical characteristics:

[0250] Volume: The length, width and height of an item are measured using a 3D vision sensor (such as Intel RealSense D415). For example, a package with dimensions of 30cm × 20cm × 15cm has a volume of 9000cm³.

[0251] Weight: Weight data is obtained using a pressure sensor (range 0~50kg, accuracy ±10g), such as an item weighing 2.3kg;

[0252] Access behavior characteristics:

[0253] Access frequency: Counts the number of times the item has been accessed in the past 30 days. For example, if a tool is accessed 5 times a week, it is marked as a high-frequency item.

[0254] Access time period: Record the timestamp of each access and analyze whether there is a time period preference (e.g., a certain medical supplies are mostly accessed between 9 and 11 am).

[0255] Item feature matrix construction:

[0256] The above features are encoded into a matrix form, with each row representing an item and the columns corresponding to the feature dimensions, as shown in Table 1.

[0257] Table 1

[0258]

[0259] This matrix serves as the input for subsequent space optimization, and its dimension is N×4 (N is the number of items).

[0260] Define a three-dimensional spatial value function, divide the cabinet into voxel units, and calculate the access efficiency weight, space utilization rate and item correlation weight of each voxel.

[0261] The three-dimensional spatial value function guides the optimization of item layout by quantifying the comprehensive value of each cabinet element (i.e., the smallest unit of space, such as 5cm×5cm×5cm). The specific definition is as follows:

[0262] Voxel division:

[0263] Assume the internal space of the storage cabinet is 2m×1m×2m (length×width×height), divided into 40×20×40 voxels, each voxel having a volume of 5cm³;

[0264] The position coordinates of each voxel are encoded as (x, y, z), for example (10, 5, 30) represents the 10th column, the 5th row, and the 30th layer.

[0265] Value dimension calculation:

[0266] Access efficiency weight: Based on the distance from the voxel to the cabinet door (e.g., the closer the voxel is to the door, the higher the weight), the calculation formula is as follows: Where d is the Euclidean distance from the voxel to the gate, and D_max is the distance to the farthest voxel (e.g., a diagonal distance of 3m). Therefore, the weight of a voxel 0.5m from the gate is 0.83.

[0267] Space utilization rate: Calculated based on the volume ratio of items that a voxel can hold. For example, if a voxel already holds items with a volume of 4000 cm³ (voxel capacity 5000 cm³), then the utilization rate is 80%.

[0268] Item association weight: The frequency of coexistence of similar items in historical data. For example, tools often coexist with repair manuals, so the association weight is set to 0.7.

[0269] Comprehensive value function: Value = 0.5 × W access +0.3 × space utilization rate +0.2 × item relevance weight.

[0270] Each voxel has a value rating between 0 and 1, and high-scoring voxels are given priority for allocation to high-frequency or related items.

[0271] A deep reinforcement learning environment is constructed, and the item feature matrix and three-dimensional spatial value function are input into the policy network. The action sequence of cabinet spatial layout is explored through the proximal policy optimization algorithm.

[0272] The deep reinforcement learning environment is customized based on the OpenAI Gym framework and includes the following core components:

[0273] State space:

[0274] Current cabinet space status (voxel occupancy matrix, 40×20×40);

[0275] Item feature matrix (N×4);

[0276] Three-dimensional value matrix (40×20×40×3, including access efficiency, utilization rate, and correlation).

[0277] Action space:

[0278] Discrete action: Select item ID and target voxel coordinates (e.g., "Place EPC001 into (10,5,30)");

[0279] Continuous action: Adjust the object's posture (such as rotating the angle from 0 to 90°) to fit the space.

[0280] Reward function:

[0281] Positive Reward: Upon successful placement of an item, the reward value is the total value of the target voxel (e.g., 0.8).

[0282] Negative Reward: Penalty of -0.5 when the item's volume exceeds the remaining space of the voxel;

[0283] Time penalty: If each step takes more than 1 second, the penalty is -0.1.

[0284] Policy Networks and PPO Algorithm:

[0285] Policy network: A 3-layer convolutional network (32-64-128 channels) is used to extract spatial features, followed by a fully connected layer to output the action probability distribution;

[0286] Proximity Policy Optimization (PPO):

[0287] Set key parameters: learning rate: 0.00025; discount factor γ: 0.99; shear threshold ε: 0.2, to limit the policy update amplitude to prevent oscillation.

[0288] Training process:

[0289] Exploration phase: The agent randomly attempts to place actions to accumulate initial experience (approximately 10,000 actions).

[0290] Strategy optimization: In each iteration, 512 data points are sampled from the experience pool, the advantage function (GAE, λ=0.95) is calculated, and the network is updated;

[0291] Convergence condition: The average reward fluctuation is less than 1% over 10 consecutive rounds.

[0292] The course learning mechanism is designed to first train the optimal positioning strategy for a single item, then gradually increase the number of items to the complexity of the actual scenario, and generate a set of candidate layout schemes.

[0293] Curriculum learning improves model performance by gradually increasing the difficulty in stages:

[0294] Phase 1 (Single Item):

[0295] Task: Place a single item into the optimal voxel;

[0296] Example: EPC001 (volume 9000cm³) should be placed near the door in a voxel with low utilization;

[0297] Training objective: To enable the model to match the volume of an object with the remaining space of a voxel, while maximizing access efficiency.

[0298] Phase 2 (Multiple Items, No Conflicts):

[0299] Task: Place 5 items in sequence, ensuring no overlap.

[0300] Challenges: The relationships between items need to be considered (e.g., frequently used items should be stored together).

[0301] Reward adjustment: Increase the correlation weight to 0.3.

[0302] Phase 3 (Total Complexity):

[0303] Task: Process 50 items simultaneously, allowing for dynamic adjustments (such as replacing old items);

[0304] Introduce random events: simulate sudden access demands (such as the urgent need to retrieve a certain item) to test the robustness of the strategy.

[0305] Candidate solution generation:

[0306] After each training phase, the 10 best-performing strategies (e.g., single-item positioning error <5cm, multi-item collision rate <2%) are saved to form a candidate layout scheme library. For example:

[0307] Option A: Place frequently used small items near the cabinet door, and large items at the back;

[0308] Option B: Divide the space by item type (tools area, document area, etc.).

[0309] Candidate layout schemes are screened based on Pareto front analysis, and the comprehensive score is calculated by the entropy weight-TOPSIS algorithm to output the optimal cabinet allocation layout scheme that balances space utilization and access efficiency.

[0310] Pareto Front Analysis:

[0311] Objective: To maximize space utilization, access efficiency, and connectivity;

[0312] Filtering: Eliminate dominated solutions from the candidate solutions (i.e., there exists another solution that is better on all objectives);

[0313] Example: Option A has a utilization rate of 85% and an efficiency of 0.7; Option B has a utilization rate of 80% and an efficiency of 0.75. If neither option is mutually dependent, both should be retained.

[0314] Entropy weight method for determining weights:

[0315] Calculate the information entropy of each target; the smaller the entropy value (the greater the data fluctuation), the higher the weight.

[0316] For example: the space utilization entropy value is the lowest (with significant fluctuations), with a weight of 0.5; access efficiency is 0.3; and correlation is 0.2.

[0317] TOPSIS (Top-Solution Ranking Method for Approximating Ideal Solutions):

[0318] Ideal solution: the maximum value of each objective (e.g., utilization rate 90%, efficiency 0.8, correlation degree 0.9);

[0319] Negative ideal solution: Minimum of each objective (e.g., utilization rate 70%, efficiency 0.5, correlation degree 0.4);

[0320] Proximity calculation: ,in, Let i be the distance between solution i and the ideal solution. This is the distance from the negative ideal solution.

[0321] Sorting: The closer the similarity (closer to 1), the better the solution.

[0322] Final solution selection:

[0323] Assuming a proximity score of 0.72 for option A and 0.68 for option B, option A is selected as the optimal layout. This layout achieves 85% space utilization and reduces average access time by 30% compared to the baseline.

[0324] S205, according to the optimal cabinet allocation scheme, drive the modular cabinet group actuator to adjust the physical structure, synchronously encrypt and write the item status data into the distributed ledger, and trigger the edge computing node to update the global inventory topology map.

[0325] Specifically, the optimal cabinet allocation scheme can be transformed into a sequence of motion control instructions for the robotic arm, and a collision-free path trajectory can be planned using inverse kinematics algorithms;

[0326] The optimal cabinet allocation scheme defines the target position of items in three-dimensional coordinates (e.g., the coordinates of cabinet B-3 are x=1.2m, y=0.8m, z=0.5m). The generation of robotic arm motion control commands needs to address the following issues:

[0327] Inverse kinematics calculation: Converting the target coordinates into the angles of each joint of the robotic arm. Taking a six-DOF robotic arm as an example, numerical iterative methods (such as the Newton-Raphson algorithm) are used to solve for the joint angles θ1~θ6. For example, the target position corresponds to θ1=30°, θ2=45°, θ3=60°, θ4=90°, θ5=15°, and θ6=0°.

[0328] Collision-free path planning:

[0329] Environmental modeling: Construct a 3D point cloud map of the cabinet group (accuracy ±5mm) and mark obstacles (such as adjacent cabinets and stored items).

[0330] Path search: The initial path is generated using RRT* (Fast Exploratory Random Tree Star Algorithm), with a node expansion step size of 50mm and a collision detection frequency of 100Hz.

[0331] Trajectory optimization: The path is smoothed by using B-spline curves, and the joint angular acceleration is constrained to ≤50 rad / s² to avoid mechanical vibration.

[0332] Control command sequence encoding:

[0333] Command format: Includes target coordinates, joint angles, motion speed (e.g., 200mm / s), and acceleration curve (S-shaped acceleration / deceleration).

[0334] Example command:

[0335] MoveTo(x=1.2, y=0.8, z=0.5),

[0336] JointAngles(30,45,60,90,15,0),

[0337] Velocity=200mm / s, AccelerationType=S-Curve.

[0338] After the robotic arm controller (such as EtherCAT bus driver) parses the instructions, it executes the action through servo motors (such as Panasonic MINASA6 series), with a positioning accuracy of ±0.1mm.

[0339] Pressure sensor data is collected in real time during the drive process, and the motor torque is dynamically adjusted using a PID closed-loop control algorithm.

[0340] Pressure sensors (such as the TE Connectivity MS5837-30BA) are deployed on the contact surface between the robotic arm gripper and the cabinet compartment to monitor the gripping force in real time (range 0~50N, accuracy ±0.5N). PID control flow:

[0341] Setting value: Set the target clamping force according to the weight of the item (e.g., 5N corresponds to a 500g item).

[0342] Feedback signal: The sensor uploads the real-time clamping force F_actual at a frequency of 100Hz.

[0343] PID calculation:

[0344] Proportional term (P): Kp=2.0, fast response deviation (e.g., when F_actual=4N, the output adjustment Δ=2×(5-4)=2N);

[0345] Integral term (I): Ki=0.5, to eliminate steady-state error (when the cumulative error reaches 2N·s, Δ+=0.5×2=1N);

[0346] Differential term (D): Kd=0.1, suppressing overshoot (if the clamping force change rate is -0.5N / s, Δ-=0.1×0.5=0.05N).

[0347] Torque adjustment: The PID output Δ is converted into motor current (e.g., 1N=0.2A), and the motor is driven by the PWM signal.

[0348] Adaptive optimization: When the object is detected to be sliding (by a gripper displacement sensor, such as a grating ruler, which reports a position change of >1mm), Kp is automatically increased to 3.0 and the integral term is reset to ensure stable gripping.

[0349] Synchronize the item displacement data with the digital twin model, and generate a state change proof chain using a zero-knowledge proof algorithm;

[0350] Digital twin synchronization:

[0351] Data acquisition: The robotic arm encoder (such as Heidenhain EQN1325) provides real-time feedback on joint angles, which, combined with a vision positioning system (such as Intel RealSense D455), captures the position of the object and generates a displacement data stream (format: timestamp|x|y|z|attitude angle).

[0352] Model updates: Data is synchronized to a digital twin platform (such as Siemens MindSphere) via the OPC UA protocol, with a 3D model update frequency of 30Hz, ensuring that the virtual-to-real error is <2mm.

[0353] Zero-knowledge proof (ZKP):

[0354] Proof generation: The zk-SNARKs (zero-knowledge concise non-interactive knowledge proof) algorithm is used to generate proof chains for item position changes. For example, to prove the truth that "item A was moved from cabinet B-3 to C-5", no details of the movement path need to be revealed.

[0355] Verification process:

[0356] Trusted settings: Pre-generated common parameters (CRS, Common Reference String);

[0357] Proof generation: Input private data (displacement trajectory) and public constraints (such as coordinate range) to generate proof π;

[0358] Verification: Other nodes use CRS to verify the validity of π (time < 10ms).

[0359] Example proof of chain structure:

[0360] Block #1234:

[0361] - Item ID: A-01

[0362] - Original location hash: 0x3a7d...

[0363] - New location hash: 0x5b9e...

[0364] - ZKP proof: 0x8c2f...

[0365] - Previous block hash: 0x4d1a...

[0366] Trigger the Byzantine Broadcast Protocol between edge computing nodes to write encrypted change data into the distributed ledger and update the spatiotemporal state of the global inventory topology.

[0367] Byzantine Broadcasting Agreement:

[0368] Node roles: There are 5 edge nodes (N1~N5), which can tolerate a maximum of f=2 malicious nodes.

[0369] Broadcast process:

[0370] Preparatory phase: Master node N1 sends<PRE-PREPARE, blk, v, n, sig> Where v = view number, n = sequence number, and sig = digital signature;

[0371] Preparation phase: Each node broadcasts the signature after verification.<PREPARE, blk, v, n, sig> After receiving 2f+1 valid messages, the submission phase begins.

[0372] Submission Phase: Broadcast<COMMIT, blk, v, n, sig> After receiving 2f+1 commits, the block is confirmed to be valid.

[0373] Data encryption and writing:

[0374] Encryption algorithm: The displacement data is encrypted using AES-256-GCM mode, and the key is dynamically distributed by KMS (Key Management System);

[0375] Ledger structure: Each block contains:

[0376] Block header: version number, timestamp, previous block hash, Merkle root;

[0377] Transaction list: Encrypted item movement records;

[0378] Consensus signature: Joint signature of PBFT nodes.

[0379] Global inventory topology update:

[0380] Spatiotemporal index: Based on the GeoHash algorithm, the cabinet location is encoded as a 12-bit string (e.g., wx4g0b), supporting fast range queries;

[0381] Incremental update: Only synchronizes changes to specific areas (such as the status of cabinet C-5), reducing network load. For example, if the original topology diagram shows cabinet C-5 as "empty", the updated status will be "occupied | item A-01".

[0382] Example execution result:

[0383] Latency: From the completion of the robotic arm's action to the update of the global topology graph, the end-to-end latency is <500ms;

[0384] Reliability: Byzantine fault tolerance ensures consensus can still be reached even when two nodes fail;

[0385] Security: Zero-knowledge proofs and AES-256 encryption ensure data privacy and integrity.

[0386] As can be seen, the system receives online approval requests based on permission parameters and user biometrics, generating an encrypted digital token. Based on this token, a multimodal biometric fusion algorithm is used for identity authentication, linking facial and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code. The target storage locker group is activated based on this dynamic identity verification code, generating a unique set of item identifiers. Based on the difference analysis between this unique item identifier set and the preset inventory list, an optimal locker allocation scheme is output. This optimal allocation scheme drives the modular locker group's execution mechanism to adjust its physical structure, simultaneously encrypting and writing item status data into a distributed ledger, triggering edge computing nodes to update the global inventory topology. This enables intelligent, adaptive, and distributed collaborative management of storage lockers while ensuring security.

[0387] Another embodiment of the present invention provides a storage system for an intelligent safe deposit box, see [link to relevant documentation]. Figure 3 The system may include:

[0388] The receiving module 301 is used to receive the permission parameters and user biometric features of the online approval request, verify the compliance of the approval process through a blockchain consensus algorithm, and generate an encrypted digital token. The blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints.

[0389] The authentication module 302 is used to perform identity authentication based on the encrypted digital token using a multimodal biometric fusion algorithm, and to associate facial features and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code.

[0390] The identification module 303 is used to activate the target storage cabinet group according to the dynamic identity verification code, identify the items in the cabinet through the anti-interference RFID signal optimization model, and use a spatiotemporal convolutional network to separate the RFID signal of the items from electromagnetic noise to generate a unique set of item identifiers.

[0391] The optimization module 304 is used to dynamically optimize the cabinet space layout based on the difference analysis between the unique identifier set of the items and the preset inventory list, and to construct a three-dimensional spatial value function based on the item access frequency and volume characteristics, and output the optimal cabinet allocation scheme.

[0392] The execution module 305 is used to drive the modular cabinet group execution mechanism to adjust the physical structure according to the optimal cabinet allocation scheme, and simultaneously encrypt and write the item status data into the distributed ledger, triggering the edge computing node to update the global inventory topology map.

[0393] As can be seen, the system receives online approval requests based on permission parameters and user biometrics, generating an encrypted digital token. Based on this token, a multimodal biometric fusion algorithm is used for identity authentication, linking facial and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code. The target storage locker group is activated based on this dynamic identity verification code, generating a unique set of item identifiers. Based on the difference analysis between this unique item identifier set and the preset inventory list, an optimal locker allocation scheme is output. This optimal allocation scheme drives the modular locker group's execution mechanism to adjust its physical structure, simultaneously encrypting and writing item status data into a distributed ledger, triggering edge computing nodes to update the global inventory topology. This enables intelligent, adaptive, and distributed collaborative management of storage lockers while ensuring security.

[0394] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0395] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0396] S201, Receive the permission parameters and user biometrics of the online approval request, verify the compliance of the approval process through the blockchain consensus algorithm, and generate an encrypted digital token, wherein the blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints;

[0397] S202, Based on the encrypted digital token, a multimodal biometric fusion algorithm is used for identity authentication, and a dynamic identity verification code is output by associating facial features and voiceprint features through a spatiotemporal attention mechanism.

[0398] S203, activate the target storage cabinet group according to the dynamic identity verification code, identify the items in the cabinet through the anti-interference RFID signal optimization model, and use a spatiotemporal convolutional network to separate the RFID signal of the items from electromagnetic noise to generate a unique set of item identifiers.

[0399] S204. Based on the difference analysis between the unique identifier set of the items and the preset inventory list, a reinforcement learning algorithm is used to dynamically optimize the spatial layout of the cabinet group. A three-dimensional spatial value function is constructed according to the storage and retrieval frequency and volume characteristics of the items, and the optimal cabinet allocation scheme is output.

[0400] S205, according to the optimal cabinet allocation scheme, drive the modular cabinet group actuator to adjust the physical structure, synchronously encrypt and write the item status data into the distributed ledger, and trigger the edge computing node to update the global inventory topology map.

[0401] As can be seen, the system receives online approval requests based on permission parameters and user biometrics, generating an encrypted digital token. Based on this token, a multimodal biometric fusion algorithm is used for identity authentication, linking facial and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code. The target storage locker group is activated based on this dynamic identity verification code, generating a unique set of item identifiers. Based on the difference analysis between this unique item identifier set and the preset inventory list, an optimal locker allocation scheme is output. This optimal allocation scheme drives the modular locker group's execution mechanism to adjust its physical structure, simultaneously encrypting and writing item status data into a distributed ledger, triggering edge computing nodes to update the global inventory topology. This enables intelligent, adaptive, and distributed collaborative management of storage lockers while ensuring security.

[0402] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0403] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0404] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0405] S201, Receive the permission parameters and user biometrics of the online approval request, verify the compliance of the approval process through the blockchain consensus algorithm, and generate an encrypted digital token, wherein the blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints;

[0406] S202, Based on the encrypted digital token, a multimodal biometric fusion algorithm is used for identity authentication, and a dynamic identity verification code is output by associating facial features and voiceprint features through a spatiotemporal attention mechanism.

[0407] S203, activate the target storage cabinet group according to the dynamic identity verification code, identify the items in the cabinet through the anti-interference RFID signal optimization model, and use a spatiotemporal convolutional network to separate the RFID signal of the items from electromagnetic noise to generate a unique set of item identifiers.

[0408] S204. Based on the difference analysis between the unique identifier set of the items and the preset inventory list, a reinforcement learning algorithm is used to dynamically optimize the spatial layout of the cabinet group. A three-dimensional spatial value function is constructed according to the storage and retrieval frequency and volume characteristics of the items, and the optimal cabinet allocation scheme is output.

[0409] S205, according to the optimal cabinet allocation scheme, drive the modular cabinet group actuator to adjust the physical structure, synchronously encrypt and write the item status data into the distributed ledger, and trigger the edge computing node to update the global inventory topology map.

[0410] As can be seen, the system receives online approval requests based on permission parameters and user biometrics, generating an encrypted digital token. Based on this token, a multimodal biometric fusion algorithm is used for identity authentication, linking facial and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code. The target storage locker group is activated based on this dynamic identity verification code, generating a unique set of item identifiers. Based on the difference analysis between this unique item identifier set and the preset inventory list, an optimal locker allocation scheme is output. This optimal allocation scheme drives the modular locker group's execution mechanism to adjust its physical structure, simultaneously encrypting and writing item status data into a distributed ledger, triggering edge computing nodes to update the global inventory topology. This enables intelligent, adaptive, and distributed collaborative management of storage lockers while ensuring security.

[0411] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for storing items in an intelligent safe, characterized in that, The method includes: The system receives permission parameters and user biometrics from online approval requests, verifies the compliance of the approval process using a blockchain consensus algorithm, and generates an encrypted digital token. The blockchain consensus algorithm integrates the approval hierarchy topology and timestamp constraints. Specifically, based on the permission parameters of the online approval request, it parses the node permission thresholds in the approval hierarchy topology to generate a directed graph of the approval process with timestamps. It then performs spatiotemporal correlation encoding between the user's biometric hash value and the directed graph of the approval process, and verifies the signature sequence legality of each level of approval nodes using a Byzantine fault-tolerant algorithm. Based on the verification results, it constructs an approval chain hash tree, prunes conflicting approval branches using a timestamp-weighted voting mechanism, and generates a compliant approval path hash chain. Finally, it integrates the compliant approval path hash chain with the user's biometric entropy value and generates a dynamic digital token containing permission validity using an elliptic curve cryptography algorithm. Based on the encrypted digital token, a multimodal biometric fusion algorithm is used for identity authentication. A spatiotemporal attention mechanism is used to associate facial features and voiceprint features to output a dynamic identity verification code. The target safe deposit box group is activated according to the dynamic identity verification code. The items in the cabinet are identified by the anti-interference RFID signal optimization model. The RFID signal of the items is separated from the electromagnetic noise by the spatiotemporal convolutional network to generate a unique set of item identifiers. Based on the difference analysis between the unique identifier set of the items and the preset inventory list, a reinforcement learning algorithm is used to dynamically optimize the spatial layout of the cabinet group. A three-dimensional spatial value function is constructed according to the storage and retrieval frequency and volume characteristics of the items, and the optimal cabinet allocation scheme is output. Based on the optimal cabinet allocation scheme, the modular cabinet group actuator is driven to adjust the physical structure, and the item status data is simultaneously encrypted and written into the distributed ledger, triggering the edge computing node to update the global inventory topology map.

2. The method according to claim 1, characterized in that, The method involves using a multimodal biometric fusion algorithm for identity authentication based on the encrypted digital token, linking facial features and voiceprint features through a spatiotemporal attention mechanism, and outputting a dynamic identity verification code, including: Based on the encrypted digital token, the pre-stored face image sequence and voiceprint temporal spectrum are extracted, and the spatiotemporal motion features of the face are extracted through a three-dimensional convolutional network to generate a dynamic face feature tensor. Mel-cepstral coefficient transform is performed on the time spectrum of the voiceprint, and the time sequence features of the voiceprint are extracted by combining a gated recurrent unit, and the voiceprint feature vector sequence is output. Construct a spatiotemporal attention alignment matrix, calculate the cross-modal correlation between dynamic face feature tensors and voiceprint feature vector sequences, and generate a feature alignment weight map; Based on the feature alignment weight map, an adaptive gated fusion algorithm is used to dynamically weight and fuse face and voiceprint features, and output a spatiotemporally consistent fused biometric feature vector. The fused biometric vectors are input into a two-stream Transformer network, and a one-time dynamic authentication code is generated through a timestamp-bound dynamic hash function.

3. The method according to claim 2, characterized in that, The process of activating the target storage locker group based on the dynamic authentication code, identifying items inside the locker using an anti-interference RFID signal optimization model, separating the RFID signal and electromagnetic noise of the items using a spatiotemporal convolutional network, and generating a unique set of item identifiers includes: The space access tag in the dynamic authentication code is parsed to activate the radio frequency antenna array of the target safe deposit box group and transmit multi-band RFID scanning signals. The time-frequency characteristics of the reflected signals inside the cabinet are collected, and the item tag signals and environmental electromagnetic noise are separated by a spatiotemporal convolutional network to construct a denoised signal tensor. An adaptive threshold decoding algorithm is used to parse the EPC code in the signal tensor and to reconstruct the overlapping label signals by phase difference separation. The EPC code is matched with the pre-stored item fingerprint database using Hamming distance to generate a set of unique item identifiers with confidence scores.

4. The method according to claim 3, characterized in that, The method involves analyzing the differences between the unique identifier set of the items and the preset inventory list, using a reinforcement learning algorithm to dynamically optimize the spatial layout of the cabinet groups, constructing a three-dimensional spatial value function based on the item access frequency and volume characteristics, and outputting the optimal cabinet allocation scheme, including: Based on the difference between the unique identifier set of items and the inventory list, the volume, weight, and historical access frequency of newly added items are extracted to construct an item feature matrix; Define a three-dimensional spatial value function, divide the cabinet into voxel units, and calculate the access efficiency weight, space utilization rate and item correlation weight of each voxel. A deep reinforcement learning environment is constructed, and the item feature matrix and three-dimensional spatial value function are input into the policy network. The action sequence of cabinet spatial layout is explored through the proximal policy optimization algorithm. The course learning mechanism is designed to first train the optimal positioning strategy for a single item, then gradually increase the number of items to the complexity of the actual scenario, and generate a set of candidate layout schemes. Candidate layout schemes are screened based on Pareto front analysis, and the comprehensive score is calculated by the entropy weight-TOPSIS algorithm to output the optimal cabinet allocation layout scheme that balances space utilization and access efficiency.

5. The method according to claim 4, characterized in that, The process of driving the modular cabinet group actuator to adjust the physical structure according to the optimal cabinet allocation scheme, simultaneously encrypting and writing the item status data into the distributed ledger, and triggering the edge computing nodes to update the global inventory topology map includes: The optimal cabinet allocation scheme is transformed into a sequence of motion control instructions for the robotic arm, and a collision-free path trajectory is planned using inverse kinematics algorithms. Pressure sensor data is collected in real time during the drive process, and the motor torque is dynamically adjusted using a PID closed-loop control algorithm. Synchronize the item displacement data with the digital twin model, and generate a state change proof chain using a zero-knowledge proof algorithm; Trigger the Byzantine Broadcast Protocol between edge computing nodes to write encrypted change data into the distributed ledger and update the spatiotemporal state of the global inventory topology.

6. A storage system for an intelligent safe deposit box, characterized in that, The system includes: The receiving module receives permission parameters and user biometrics from online approval requests, verifies the compliance of the approval process using a blockchain consensus algorithm, and generates an encrypted digital token. The blockchain consensus algorithm integrates the approval hierarchy topology with timestamp constraints. Specifically, based on the permission parameters of the online approval request, it parses the node permission thresholds in the approval hierarchy topology to generate a directed graph of the approval process with timestamps. It then performs spatiotemporal correlation encoding between the user's biometric hash value and the directed graph of the approval process, verifies the signature sequence legality of each level of approval nodes using a Byzantine fault-tolerant algorithm, constructs an approval chain hash tree based on the verification results, prunes conflicting approval branches using a timestamp-weighted voting mechanism, and generates a compliant approval path hash chain. Finally, it integrates the compliant approval path hash chain with the user's biometric entropy value and generates a dynamic digital token containing permission validity using an elliptic curve cryptography algorithm. The authentication module is used to perform identity authentication based on the encrypted digital token using a multimodal biometric fusion algorithm, and to associate facial features and voiceprint features through a spatiotemporal attention mechanism to output a dynamic identity verification code. The identification module is used to activate the target storage cabinet group according to the dynamic identity verification code, identify the items in the cabinet through the anti-interference RFID signal optimization model, and use the spatiotemporal convolutional network to separate the RFID signal of the items from the electromagnetic noise, and generate a unique set of item identifiers. The optimization module is used to dynamically optimize the cabinet space layout based on the difference analysis between the unique identifier set of the items and the preset inventory list, and to construct a three-dimensional spatial value function based on the item access frequency and volume characteristics, and output the optimal cabinet allocation scheme. The execution module is used to drive the modular cabinet group execution mechanism to adjust the physical structure according to the optimal cabinet allocation scheme, synchronously encrypt and write the item status data into the distributed ledger, and trigger the edge computing node to update the global inventory topology map.

7. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images