Storage cabinet input method and system based on cloud platform

Abstract

The invention provides a storage cabinet input method and system based on a cloud platform, and the method comprises the steps: collecting the multi-modal data of a storage cabinet, and the multi-modal data comprises an article image, an environment parameter, and user authentication information; performing multi-modal article segmentation processing on the multi-modal data to obtain a segmentation result; performing quantitative mapping processing on the time sequence features, and mapping the time sequence features to a standard feature space to obtain standardized article feature representation; performing multi-level security verification on the standardized article feature representation to obtain a verification result; wherein when the verification result is in a verification passing state, cloud storage processing is executed; and based on the verification result, generating a digital file of the articles in the storage cabinet and executing cloud synchronization processing. According to the technical scheme, the problems that in the prior art, the capability of accurately recognizing and segmenting stored articles is lacked, and safety certification and service management under the heterogeneous Internet of Things environment cannot be effectively solved are solved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence, cloud computing and the Internet of Things, and in particular to a savings cabinet entry method and system based on a cloud platform. Background Art

[0002] At present, the following methods are mainly used to enter information into savings cabinets: the traditional manual entry method requires staff to manually input the information of stored items, which is inefficient and prone to errors; the semi-automatic entry system based on simple image recognition, although it has introduced computer vision technology, is not effective in identifying items in complex scenarios; the existing cloud platform entry solution lacks an effective security protection mechanism and there is a risk of data leakage.

[0003] In addition, the existing technology has the following technical problems: lack of accurate identification and segmentation capabilities for stored items, especially in multi-item scenarios with complex backgrounds; and failure to effectively solve security authentication and service management issues in heterogeneous IoT environments. These problems seriously restrict the actual application effect of the savings cabinet entry system. Summary of the invention

[0004] The present application provides a savings cabinet entry method and system based on a cloud platform. By introducing technologies such as adversarial training, conditional random field optimization, and time-aware diffusion transformer, the present application solves the problem that the prior art lacks the ability to accurately identify and segment stored items, and fails to effectively solve the security authentication and service management problems in a heterogeneous Internet of Things environment.

[0005] In a first aspect, a method for entering data into a savings cabinet based on a cloud platform is provided, comprising: Collecting multimodal data of the savings cabinet, wherein the multimodal data includes an image of an item, environmental parameters, and user authentication information; Performing multimodal item segmentation processing on the multimodal data to obtain a segmentation result; wherein the multimodal item segmentation processing includes: performing segmentation processing on the multimodal data using an adversarial training network architecture to generate an initial segmentation map; optimizing the initial segmentation map using a conditional random field to optimize the spatial consistency of pixel labels in the initial segmentation map to obtain the optimized segmentation result; Inputting the segmentation results into a time-aware diffusion transformation model for processing to extract temporal features; Performing quantization mapping processing on the time series features, mapping the time series features to a standard feature space, and obtaining a standardized feature representation of the item; Performing multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, performing cloud storage processing; Based on the verification results, a digital file of the savings cabinet items is generated and cloud synchronization is performed.

[0006] In a possible implementation, the segmentation process of the multimodal data using an adversarial training network architecture includes: Constructing an adversarial training network including a generator and a discriminator; wherein the generator adopts a U-Net structure including an encoder and a decoder module, and the discriminator adopts a fully convolutional network structure; Inputting the multimodal data into the generator for processing to generate the initial segmentation map; Using the discriminator to perform authenticity evaluation on the initial segmentation map to obtain an evaluation result; Based on the evaluation results, parameters of the generator and the discriminator are optimized.

[0007] In a possible implementation, the optimizing the initial segmentation map using a conditional random field includes: Construct an energy function including data terms and smooth terms; The energy function is used to process the initial segmentation map to obtain a preliminary optimized segmentation result; wherein the data item calculates the label probability that each pixel of the initial segmentation map belongs to a different label category, and the smoothing item calculates the label association degree between adjacent pixels of the initial segmentation map; Using an inference algorithm of an iterative conditional mode to optimize the energy function, thereby obtaining an optimized energy function; The optimized energy function is used to optimize the preliminary optimized segmentation result, and finally the segmentation result after multiple rounds of iterative optimization is output; Combined with the structured learning method, various weight parameters in the energy function are dynamically adjusted.

[0008] In a possible implementation, the step of inputting the segmentation result into a time-aware diffusion transformation model for processing to extract temporal features includes: Constructing a multi-head self-attention mechanism, converting the features of the input segmentation result into a query vector, a key vector and a value vector respectively through linear mapping, and obtaining the features after attention processing; The features after the attention processing are processed by combining position coding and time coding to obtain features before coding; wherein the position coding uses sine-cosine functions to calculate the spatial position information of the features, and the time coding uses a learnable parameter matrix to calculate the temporal dependency of the features; The pre-encoded features are input into a hierarchical diffusion structure for processing to obtain the temporal features; wherein the hierarchical diffusion structure includes multiple layers of encoders and decoders, the encoder converts the input features into potential representations by adding noise, and the decoder reconstructs the temporal features through a denoising process.

[0009] In a possible implementation, the performing quantization mapping process includes: Constructing a vector quantizer, the vector quantizer comprising a codebook of a preset number of codewords and a feature mapping network; Performing normalization preprocessing on the time series features, and using a nearest neighbor search algorithm to retrieve matching codewords in the codebook; A soft allocation mechanism is used to quantize and map the temporal features, allowing a feature vector to be matched with multiple codewords at the same time, and generating the standardized item feature representation according to the matching weight.

[0010] In a possible implementation, performing multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, performing cloud storage processing includes: Construct a Q-learning service framework, which includes: Set up a state space containing system resource usage and network load, and an action space containing service scheduling strategies; Build a dual Q network structure including the main network and the target network; Set up a priority experience replay mechanism to dynamically adjust the sample sampling probability according to the time series difference error; Based on the Q-learning service framework, multi-level security verification is performed on the standardized item feature representation; wherein the multi-level security verification includes identity authentication, authority authentication and abnormal behavior monitoring verification.

[0011] In a possible implementation, the performing cloud storage processing further includes: The received standardized item feature representation is integrity checked using hash chain technology. If the integrity check passes, the standardized item feature representation is sharded using a consistent hash algorithm, and the data is distributed and stored in cloud storage nodes. Based on the data security level represented by the standardized item characteristics, a corresponding encryption algorithm is used to encrypt the data.

[0012] In a possible implementation, generating a digital archive of the items in the savings cabinet based on the verification result and performing cloud synchronization processing includes: Based on the verification result, a multi-level archive structure of the digital archive is generated; wherein the multi-level archive structure of the digital archive includes a basic information layer, a feature representation layer and an associated information layer; Based on the multi-level archive structure of the digital archive, a multi-dimensional retrieval index is established to support cross-level data query operations.

[0013] In a possible implementation, the performing cloud synchronization processing includes: Inputting the archive data of the digital archive into the distributed transaction framework for processing to ensure the consistency of data operations across nodes and obtain processed archive data; Deploy a multi-level cache architecture, deploy local cache at edge nodes to store frequently accessed processed archive data, use a distributed cache cluster in the cloud and implement load balancing through a consistent hashing algorithm; The master-slave switching mechanism of the Raft consensus algorithm is adopted to perform service recovery when the edge node failure is detected.

[0014] In a second aspect, a savings cabinet entry system based on a cloud platform is provided, comprising: A data acquisition module, used to collect multimodal data of the savings cabinet, wherein the multimodal data includes an image of an item, environmental parameters, and user authentication information; A segmentation processing module is used to perform multimodal item segmentation processing on the multimodal data to obtain a segmentation result; wherein the multimodal item segmentation processing includes: using an adversarial training network architecture to perform segmentation processing on the multimodal data to generate an initial segmentation map; using a conditional random field to optimize the initial segmentation map, optimize the spatial consistency of pixel labels in the initial segmentation map, and obtain the optimized segmentation result; A feature extraction module, used for inputting the segmentation result into a time-aware diffusion transformation model for processing to extract temporal features; A feature quantization module, used to perform quantization mapping processing on the time series features, map the time series features to a standard feature space, and obtain a standardized feature representation of the item; A security verification module, used to perform multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, cloud storage processing is performed; The file generation module is used to generate a digital file of the savings cabinet items based on the verification result and perform cloud synchronization processing.

[0015] The savings cabinet input method and system provided by the present application have the following beneficial effects: The piggy bank entry method and system provided in the present application, by collecting multimodal data including item images, environmental parameters and user authentication information, utilizes the multimodal item segmentation processing combining adversarial training network architecture and conditional random fields, effectively improving the ability to accurately identify and segment stored items in complex backgrounds and multi-item scenarios; extracts time series features with the help of a time-aware diffusion transformation model and performs quantitative mapping processing to obtain standardized item feature representations, providing strong data support for subsequent precise operations; ensures data security through multi-level security verification, executes cloud storage processing after verification, and generates digital archives based on the verification results and performs cloud synchronization processing, successfully solving the security authentication and service management problems in heterogeneous Internet of Things environments, greatly improving the efficiency and accuracy of piggy bank entry, reducing the risk of data leakage, enhancing the practicality and reliability of the system, and effectively promoting the application and development of piggy bank entry technology in the fields of artificial intelligence, cloud computing and Internet of Things technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.

[0017] Figure 1 It is a schematic diagram of the overall process of the savings cabinet entry method based on the cloud platform provided in the embodiment of the present application; Figure 2 It is a structural diagram of multimodal item segmentation processing provided in an embodiment of the present application; Figure 3 is a schematic diagram of the architecture of a time-aware diffusion transformer provided in an embodiment of the present application; Figure 4 Schematic diagram of a cognitive service framework based on Q learning provided in an embodiment of the present application; Figure 5 It is a structural diagram of the savings cabinet entry system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The specific implementation methods of the embodiments of the present application are described in detail below with reference to the case of writing the specification.

[0019] like Figure 1 As shown, the embodiment of the present application provides a savings cabinet entry method based on a cloud platform, comprising: S1. Collect multimodal data of a savings cabinet, wherein the multimodal data includes an image of an item, environmental parameters, and user authentication information; In the savings cabinet entry system, multimodal data must be collected first. These data include three key parts: item image data collects visual information of items through high-resolution cameras; environmental parameters include sensor data such as temperature, humidity, and light, which affect the storage status and image quality of items; user authentication information includes multiple authentication data such as biometric recognition (such as fingerprints, faces) and traditional password verification. The collaborative collection of the above multimodal data provides a comprehensive information basis for subsequent processing.

[0020] S2. Perform multimodal item segmentation processing on the multimodal data to obtain a segmentation result; wherein the multimodal item segmentation processing includes: performing segmentation processing on the multimodal data using an adversarial training network architecture to generate an initial segmentation map; performing optimization processing on the initial segmentation map using a conditional random field to optimize the spatial consistency of pixel labels in the initial segmentation map to obtain the optimized segmentation result; like Figure 2 , the segmentation process of the multimodal data using the adversarial training network architecture includes: Constructing an adversarial training network including a generator and a discriminator; wherein the generator adopts a U-Net structure including an encoder and a decoder module, and the discriminator adopts a fully convolutional network structure; Inputting the multimodal data into the generator for processing to generate the initial segmentation map; Using the discriminator to evaluate the authenticity of the initial segmentation map to obtain an evaluation result; Based on the evaluation results, parameters of the generator and the discriminator are optimized.

[0021] In the construction phase of the adversarial training network, the generator network needs to be designed and implemented first. The generator adopts the U-Net structure, and its encoder part consists of multiple convolutional layers and pooling layers. The convolution kernel size of each layer is set to 3×3, with a step size of 1. Each pooling operation halves the size of the feature map. The decoder part uses a transposed convolution layer with a step size of 2 to achieve upsampling of the feature map. A jump connection is established between the corresponding layers of the encoder and decoder, and the feature map of the encoder is directly connected to the corresponding layer of the decoder to retain the detailed information of the image.

[0022] The discriminator is constructed using a fully convolutional network structure, which includes multiple convolutional layers and batch normalization layers. Each convolutional layer uses a leaky ReLU activation function with a leakage rate of 0.2. The last layer uses a sigmoid activation function and outputs a probability value between 0 and 1, indicating the possibility that the input image is a true segmentation map.

[0023] In the data processing stage, the system first preprocesses the multimodal data. The image data is standardized and the pixel values ​​are normalized to the range of [-1,1]. For non-image data such as environmental parameters and user authentication information, they are converted into a representation compatible with image features through feature embedding.

[0024] The generator receives the preprocessed multimodal data and generates the initial segmentation map through the encoder-decoder structure. The encoder extracts multi-scale features through continuous convolution and pooling operations, and the decoder gradually restores the spatial resolution through transposed convolutions, while using the detail information retained by jump connections, and finally outputs a segmentation map of the same size as the input image.

[0025] The discriminator evaluates the authenticity of the generated segmentation map. During the evaluation process, the discriminator receives both the generated segmentation map and the real annotated segmentation map as input, and calculates their authenticity scores through the fully convolutional network. For the real segmentation map, the system expects the discriminator to output a score close to 1; for the generated segmentation map, it expects to output a score close to 0.

[0026] The parameter optimization stage uses an alternating training strategy. First, the discriminator parameters are fixed and the generator is optimized so that the segmentation map it generates can deceive the discriminator. The optimization objectives include adversarial loss and losses related to the segmentation task (such as cross entropy loss). Then the generator parameters are fixed and the discriminator is optimized to improve its ability to distinguish between true and false segmentation maps. This process is repeated until the network performance converges.

[0027] During the training process, the system dynamically adjusts the learning rates of the generator and discriminator, and uses the Adam optimizer to update parameters. At the same time, a gradient penalty term is introduced to ensure the stability of the discriminator training. The entire training process is monitored by performance indicators on the validation set, including segmentation accuracy, recall rate and other evaluation indicators.

[0028] The following is a detailed example to illustrate how the adversarial training network works in the segmentation of items in a piggy bank.

[0029] If a jewelry box is placed in the savings cabinet, this jewelry box is made of multiple materials: a transparent glass cover, a metal decorative frame, a velvet lining, and various jewelry. This complex scene just shows the working mechanism of the adversarial training network.

[0030] First, the generator uses a U-Net structure to process the input multimodal data. In the encoder stage, the network performs convolution and downsampling operations on the image layer by layer. For example, the first layer may extract basic edge and texture features to capture the outline of the jewelry box and the reflective features of the glass; the second layer may recognize more complex patterns, such as the patterns of metal decorations and the texture of velvet; the third layer may begin to understand the components of the object, such as the relationship between the lid and the base. In this process, the spatial size of the feature map continues to decrease, but the semantic information contained is becoming richer and richer.

[0031] In the decoder stage, the network gradually restores the spatial resolution of the feature map through transposed convolution (deconvolution) operations. The unique jump connection mechanism ensures the preservation of fine details. For example, the edge information passed from the first layer of the encoder can help the decoder accurately reconstruct the fine contours of the jewelry box; the texture information of the second layer helps to distinguish between metal and glass areas. Finally, the generator outputs an initial segmentation map, marking the various components of the jewelry box.

[0032] The discriminator uses a fully convolutional network structure to evaluate the segmentation result. It focuses on several key points: whether the segmentation boundary accurately matches the actual outline of the jewelry box; whether the transparent glass area is correctly identified rather than misjudged as the background; whether the reflection of the metal decoration affects the continuity of the segmentation; whether the segmentation of the velvet lining is complete; through a series of convolutional layers, the discriminator finally gives an evaluation score, pointing out the advantages and problems in the segmentation result.

[0033] Based on the discriminator's evaluation, the system begins to optimize the network parameters. For example, if the discriminator finds that there is an error in the segmentation of the glass area, the parameters of the generator will be adjusted accordingly, perhaps to enhance the sensitivity of specific convolution kernels to transparent materials, or to adjust the weight distribution in the jump connection. If the reflection of the metal frame causes the segmentation to be broken, the system may strengthen the ability of the decoder to handle highly reflective areas. At the same time, the parameters of the discriminator itself are also constantly optimized to improve its accuracy in identifying segmentation defects.

[0034] Through the repeated game between the generator and the discriminator, the network gradually improves its segmentation ability. Eventually, the system can accurately segment the various parts of the jewelry box, and maintain a stable segmentation effect even in the face of complex material combinations and lighting changes. This lays a solid foundation for subsequent feature extraction and item management.

[0035] Optionally, the using a conditional random field to optimize the initial segmentation map includes: Construct an energy function including data terms and smooth terms; The energy function is used to process the initial segmentation map to obtain a preliminary optimized segmentation result; wherein the data item calculates the label probability that each pixel of the initial segmentation map belongs to a different label category, and the smoothing item calculates the label association degree between adjacent pixels of the initial segmentation map; Using an inference algorithm of an iterative conditional mode to optimize the energy function, thereby obtaining an optimized energy function; The optimized energy function is used to optimize the preliminary optimized segmentation result, and finally the segmentation result after multiple rounds of iterative optimization is output; Combined with the structured learning method, various weight parameters in the energy function are dynamically adjusted.

[0036] In the multimodal item segmentation process, the system adopts an innovative dual optimization strategy. The first is the preliminary segmentation based on the adversarial training network architecture. The generator in this architecture adopts the U-Net structure, which is characterized by symmetrical downsampling (encoder) and upsampling (decoder) paths, and jump connections between corresponding layers. The encoder gradually extracts features through multi-layer convolution and pooling operations, while the decoder reconstructs the segmented image through deconvolution (transposed convolution) operations. The discriminator adopts a fully convolutional network structure, which avoids the spatial information loss problem caused by the traditional fully connected layer and can better maintain the spatial structure information of the image.

[0037] During adversarial training, the generator continuously generates segmentation maps, and the discriminator evaluates the authenticity of these segmentation maps. Specifically, the discriminator learns to distinguish between the real segmentation map and the segmentation map generated by the generator. This process is achieved by alternately optimizing the parameters of the generator and the discriminator. The goal of updating the generator parameters is to generate more realistic segmentation maps to deceive the discriminator, while the goal of updating the discriminator parameters is to improve the ability to distinguish. This adversarial mechanism can prompt the generator to produce higher quality segmentation results.

[0038] In order to further improve the segmentation quality, the system introduces conditional random fields for optimization. The energy function of the conditional random field contains two key components: data term and smoothing term. The data term calculates the label probability of each pixel belonging to each label category. The label probability is obtained through a deep learning model based on the feature vector of the pixel. The smoothing term considers the label association between adjacent pixels and is defined by calculating the feature similarity and spatial distance between pixel pairs. The above design can effectively maintain the spatial consistency of the segmentation results.

[0039] In the optimization process of conditional random fields, the system uses the iterative conditional mode (ICM) algorithm for inference. The algorithm updates the label of each pixel by iteration, and selects the label that minimizes the local energy in each iteration. In order to improve the optimization effect, the system also introduces a structured learning method to dynamically adjust the weight parameters in the energy function. This adjustment is based on the annotation information in the training data, and the optimal weight parameters are learned by methods such as maximizing likelihood estimation or structured support vector machines.

[0040] The multi-level segmentation optimization strategy makes full use of the feature extraction capabilities of deep learning and the structured modeling capabilities of conditional random fields. It can ensure the spatial consistency of the results while maintaining the segmentation accuracy. It is particularly suitable for processing multi-item scenes in complex backgrounds in savings cabinets.

[0041] For example, if there are multiple valuables in a savings cabinet, including a metal jewelry box, a stack of cash, and a watch, the system first obtains multimodal data of these items, including high-definition images, environmental temperature and humidity data, and user authentication information. In the above scenario, the metal material of the jewelry box will produce reflections in the image, the cash may overlap due to stacking, and the watch will bring recognition challenges due to its fine structure.

[0042] The adversarial training network first processes this complex scene. After the generator receives the input data, it extracts and reconstructs features through the U-Net structure. In the downsampling stage, the network gradually extracts the features of the items, such as the metal texture of the jewelry box, the edge contour of the cash, and the dial features of the watch. While the spatial resolution of these features is gradually reduced, the semantic information is continuously enriched. In the upsampling stage, the network uses jump connections to restore the spatial details of these features and generate an initial segmentation map.

[0043] The discriminator evaluates the authenticity of the segmentation result. For example, it will pay attention to whether the segmentation of the edge of the jewelry box is accurate, whether the overlapping area of ​​the cash is correctly identified, and whether the segmentation of the fine parts of the watch is complete. If the discriminator finds that the reflective area of ​​the jewelry box is not accurately segmented, or the boundary of the overlapping cash is blurred, these problems will be used to continuously improve the performance of the generator through feedback from adversarial training.

[0044] Next, the conditional random field optimizes the initial segmentation results. Taking the overlapping area of ​​cash as an example, the data term of the energy function calculates the probability of each pixel belonging to the category of "cash", "background", or other items. The smoothing term ensures that the labels of adjacent pixels remain consistent to avoid noisy segmentation results at the edges of cash. For example, if a pixel is judged to be in the "cash" category and the pixels around it also have similar color and texture features, then the smoothing term will tend to label these pixels as "cash" as well.

[0045] During the iterative optimization process, the system may find that the initial segmentation of the reflective areas of the jewelry box is not good. Through the iterative conditional pattern algorithm, combined with the information of the surrounding pixels and the global constraints, the system can gradually correct the segmentation results of these areas. At the same time, the structured learning method dynamically adjusts the weight parameters in the energy function based on the annotation information of similar scenes in the training data, so that the system can better handle the reflective characteristics of metal objects.

[0046] After such multi-level optimization, the system can finally accurately segment the items stored in the savings cabinet, and maintain high segmentation accuracy even under complex lighting conditions or when there are overlaps between items. This segmentation result provides a reliable basis for subsequent feature extraction and digital file generation.

[0047] S3, inputting the segmentation result into a time-aware diffusion transformation model for processing to extract temporal features; like Figure 3 , the segmentation result is input into a time-aware diffusion transformation model for processing to extract temporal features, including: Constructing a multi-head self-attention mechanism, converting the features of the input segmentation result into a query vector, a key vector and a value vector respectively through linear mapping, and obtaining the features after attention processing; The features after the attention processing are processed by combining position coding and time coding to obtain features before coding; wherein the position coding uses sine-cosine functions to calculate the spatial position information of the features, and the time coding uses a learnable parameter matrix to calculate the temporal dependency of the features; The pre-encoded features are input into a hierarchical diffusion structure for processing to obtain the temporal features; wherein the hierarchical diffusion structure includes multiple layers of encoders and decoders, the encoder converts the input features into potential representations by adding noise, and the decoder reconstructs the temporal features through a denoising process.

[0048] Specifically, in the construction phase of the multi-head self-attention mechanism, the system first determines the number of attention heads, which is usually set to 8 or 16. For each attention head, three independent linear transformation matrices are implemented to generate query vectors, key vectors, and value vectors. The dimensions of these linear transformations are usually set to the feature dimension divided by the number of attention heads. The input segmentation features are first transformed to ensure that their dimensions are divisible by the number of attention heads. Then, through matrix multiplication operations, the input features are converted into corresponding query, key, and value vectors respectively. Finally, the weighted features are obtained by scaling the dot product attention calculation, and the outputs of all attention heads are concatenated to obtain the final attention processing features through a linear projection layer.

[0049] In the encoding stage, the system realizes the fusion of position encoding and time encoding. Position encoding uses sine and cosine functions. For each dimension d of the feature vector, sine and cosine functions of different frequencies are used to calculate the position encoding. Specifically, for the position pos and dimension d, the calculation formula of the position encoding is: PE(pos,2i) = sin(pos / 10000^(2i / d)), PE(pos,2i+1) = cos(pos / 10000^(2i / d)).

[0050] Temporal encoding is achieved by constructing a learnable parameter matrix, where the number of rows corresponds to the maximum time step and the number of columns is the same as the feature dimension. The system retrieves the corresponding temporal encoding vector from the parameter matrix based on the temporal information of the input feature. Finally, the positional encoding, temporal encoding, and attention-processed features are added and fused to obtain the pre-encoding features.

[0051] In the processing stage of the hierarchical diffusion structure, the system first builds a multi-layer encoder-decoder architecture. The encoder part implements the process of adding noise step by step, using a predefined noise schedule to control the noise intensity added at each step. Specifically, at step t, the encoder converts the input feature x into

[0052] Converted to a noisy feature, where α_t is the noise scheduling coefficient and ε is the noise sampled from a standard normal distribution. The encoder contains multiple such noise adding layers, and the α_t value of each layer gradually decreases, making the feature gradually more blurred.

[0053] The decoder part implements the denoising process and adopts a U-Net-like structure, including multiple upsampling and residual connections. Each denoising step predicts the noise ε, and then uses the formula

[0054] Reconstruct clearer features. The decoder receives information from the corresponding encoder layer through the jump connection at each layer, which helps the decoder predict and remove noise more accurately. The entire denoising process is iterative, and each step gradually restores the clarity of the features, and finally outputs the reconstructed temporal feature representation.

[0055] The entire process is optimized by the gradient descent algorithm, using a combination of time series feature reconstruction loss and KL divergence loss as the optimization target. The system also implements a learning rate scheduling strategy, using a larger learning rate in the early stage of training to speed up convergence, and then gradually reducing the learning rate to achieve more refined optimization.

[0056] In this process in step S3, the multi-head self-attention mechanism is first used to process the features of the segmentation results. Suppose a piece of jewelry is placed in the savings cabinet, and the system needs to track its state over time. The multi-head self-attention mechanism is like multiple experts observing the different features of the jewelry at the same time. Through linear mapping, the system converts the features of the jewelry into query vectors (such as "What is the item at this position?"), key vectors (such as "There is a metallic luster here"), and value vectors (such as "This is a gold ring"). Each "head" focuses on a specific feature pattern, some may focus on the shape change of the object, some on the change of the surface material, and some on the movement of the position.

[0057] Next, the system needs to consider the location information and time changes of the items in the piggy bank. Position encoding uses sine and cosine functions to represent spatial positions, which is like labeling each position in the piggy bank with precise three-dimensional coordinates. For example, when jewelry moves from the front to the back of the piggy bank, the sine and cosine functions can smoothly represent this position change. Time encoding captures changes in the time dimension through a learnable parameter matrix, such as recording the state changes of jewelry at different time points, including possible position movements, orientation changes, or changes in relative position relationships with other items.

[0058] After obtaining these codes, the system uses a hierarchical diffusion structure to extract deeper temporal features. This structure is like a complex information extraction system, in which the encoder perturbs the features by adding carefully designed noise. This process can be understood as the system deliberately "blurring" certain features of the object in order to better understand what is important information. For example, when observing a ring, by adding different degrees of noise, the system can learn to distinguish between the essential features of the ring (such as material, style) and secondary features (such as temporary lighting changes, shadow effects).

[0059] The decoder is responsible for reconstructing clear temporal features from these noisy features. This process is like gradually building an understanding of the behavior patterns of objects by observing how they change at different points in time. For example, the system can learn the typical placement of jewelry in a savings cabinet, and understand that some changes are temporary (such as reflections caused by changes in light) and some changes are permanent (such as adjustments in position or the removal of items).

[0060] What’s unique about this whole process is that it can capture the dynamic characteristics of items over time. For example, if a user often takes a piece of jewelry at a certain time, the system can learn this temporal pattern. Or, if a piece of jewelry is often moved with other specific items, the system can also capture this association. These temporal characteristics are valuable in predicting the usage patterns of items, optimizing storage layouts, and providing smarter security monitoring.

[0061] Through this complex feature extraction process, the system can eventually establish a complete feature representation of the object in the time dimension, which includes not only the static characteristics of the object, but also its dynamic change laws, providing an important data foundation for subsequent intelligent management and security monitoring.

[0062] S4, performing quantization mapping processing on the time series features, mapping the time series features to a standard feature space, and obtaining a standardized feature representation of the item; The performing of quantization mapping processing comprises: Constructing a vector quantizer, the vector quantizer comprising a codebook of a preset number of codewords and a feature mapping network; Specifically, the construction process of the vector quantizer first needs to determine the size and structure of the codebook. The system usually presets 1024 or 2048 codewords as the basic configuration, and the dimension of each codeword is consistent with the input feature dimension. The feature mapping network adopts a multi-layer perceptron structure, which contains multiple fully connected layers, and each layer is followed by batch normalization and ReLU activation function. The input dimension of the first layer is the same as the feature dimension, the dimension of the middle layer is gradually adjusted, and the output dimension of the last layer matches the codeword dimension. At the same time, the system optimizes the codewords in the codebook through the back-propagation algorithm so that it can better represent the distribution of the feature space.

[0063] Performing normalization preprocessing on the time series features, and using a nearest neighbor search algorithm to retrieve matching codewords in the codebook; Specifically, a step-by-step normalization scheme is used for the normalization preprocessing of time series features. First, the mean and standard deviation of the feature are calculated, and normalization is achieved by subtracting the mean and dividing by the standard deviation. For possible outliers, the system sets a threshold for truncation to ensure that the normalized feature distribution is within a reasonable range. Then, the nearest neighbor search algorithm is used to retrieve matching codewords in the codebook. The specific implementation uses a KD tree or ball tree data structure to improve the search efficiency. The system maintains a priority queue for each input feature, storing the K nearest neighbor codewords and their corresponding distance values.

[0064] A soft allocation mechanism is used to quantize and map the temporal features, allowing a feature vector to be matched with multiple codewords at the same time, and generating the standardized item feature representation according to the matching weight.

[0065] Specifically, the soft assignment mechanism is implemented using a softmax function based on a temperature parameter. The system first calculates the Euclidean distance between the input feature and each codeword in the codebook, scales the distance value by the temperature parameter τ, and then calculates the assignment weight for each codeword using the softmax function. The temperature parameter τ controls the "softness" of the assignment, with larger τ values ​​producing smoother assignments and smaller τ values ​​making the assignments closer to hard assignments. The system dynamically adjusts the temperature parameter, using larger values ​​in the early stages of training to facilitate exploration, and then gradually reducing it to achieve more accurate matching.

[0066] When generating the standardized feature representation, the system performs a weighted combination of the soft-assigned weights and the features of the corresponding codewords. This process is implemented through matrix multiplication, where the weight vector is multiplied by the selected codeword matrix to obtain the final feature representation. In order to improve the robustness of the feature representation, the system also implements a residual coding mechanism to preserve the difference information between the original features and the quantized features, and use this residual information for feature reconstruction during decoding.

[0067] The entire quantization mapping process is optimized in an end-to-end manner, and the loss function includes reconstruction error, quantization error, and regularization term. The system uses a stochastic gradient descent optimizer and implements a learning rate decay strategy. At the same time, by periodically updating the parameters of the codebook and feature mapping network, the quantizer is ensured to adapt to changes in feature distribution. The entire processing flow is designed to be parallelizable, making full use of the parallel computing capabilities of modern hardware.

[0068] In another embodiment, first, the embodiment of the present application needs to understand the construction process of the vector quantizer. Assume that the embodiment of the present application has a database containing a large number of locker item features, and these features include timing information such as the appearance, position changes, and usage frequency of the items. The vector quantizer is like a feature dictionary, in which the "codewords" are a collection of typical feature patterns. For example, for jewelry, one codeword may represent the feature pattern of "small valuables that are often taken out and used on weekends", and another codeword may represent the feature pattern of "valuables that are stored for a long time and rarely moved". These codewords constitute a preset feature dictionary, that is, a codebook.

[0069] Before feature mapping, the system first needs to normalize the time series features. This step is like converting measurements of different scales to the same standard scale. For example, the movement frequency of an item may be measured in times / week, while the storage time may be measured in months. Through normalization, these features of different scales are converted to the same numerical range, usually in the [0,1] interval. Such standardization ensures that different types of features can be compared and matched fairly.

[0070] Next, the system uses the nearest neighbor search algorithm to find matching codewords in the codebook. This process is similar to finding the closest description in a dictionary. Assume that there is a piece of jewelry in the embodiment of the present application, which is taken out once every weekend and stored in a relatively fixed location. The system calculates the distance between this feature vector and all the codewords in the codebook to find the most similar feature pattern. The "distance" here is the mathematical distance in the high-dimensional feature space, and can be measured using Euclidean distance or cosine similarity.

[0071] The most innovative thing is that the system uses a soft assignment mechanism for quantitative mapping. The traditional hard assignment method strictly corresponds a feature to a codeword, but this may oversimplify the actual situation. The soft assignment mechanism allows a feature to be associated with multiple codewords at the same time and assign different weights. For example, the usage pattern of a piece of jewelry may be 70% consistent with the codeword of "weekend use" and 30% consistent with the codeword of "long-term storage". This soft assignment mechanism can more accurately describe the complex characteristics of items.

[0072] The result of this quantitative mapping is a standardized representation of item features, which not only retains the key feature information of the item, but also achieves dimensionality reduction and standardization of the data. This representation format facilitates the system to perform subsequent processing, such as item classification, similarity calculation, anomaly detection, etc. At the same time, due to the use of a soft allocation mechanism, this representation also retains the subtle differences in features, allowing the system to more accurately describe and distinguish the features of different items.

[0073] Through this complex quantitative mapping process, the system finally obtains a standardized and expressive feature representation, which provides a reliable data foundation for subsequent intelligent management and decision-making. This representation method is not only convenient for computer processing, but also convenient for the system to compare and analyze features across time and scenarios.

[0074] S5. Perform multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, perform cloud storage processing; specifically including: Construct a Q-learning service framework, which includes: Set up a state space containing system resource usage and network load, and an action space containing service scheduling strategies; Build a dual Q network structure including the main network and the target network; Set up a priority experience replay mechanism to dynamically adjust the sample sampling probability according to the time series difference error; Based on the Q-learning service framework, multi-level security verification is performed on the standardized item feature representation; wherein the multi-level security verification includes identity authentication, authority authentication and abnormal behavior monitoring verification.

[0075] like Figure 4 In the above system, the construction of the Q learning service framework is the key to understanding the entire security verification process. First, the system needs to define the state space, which includes all aspects of the system operation. For example, system resource usage may include indicators such as CPU utilization, memory usage, and storage space; network load includes parameters such as bandwidth utilization, network latency, and number of connections. At the same time, the action space defines various service scheduling strategies that the system can adopt, such as adjusting processing priorities, allocating computing resources, and selecting verification methods.

[0076] The design of the dual Q network structure reflects the innovation of the system. The main network is responsible for real-time decision-making and selects the optimal verification strategy based on the current state; while the target network remains relatively stable and provides a reference benchmark for decision-making. This structure is particularly suitable for handling tasks such as security verification that require a balance between stability and adaptability. For example, when the system detects an abnormal access pattern, the main network may recommend adding additional verification steps, while the target network helps evaluate the long-term impact of this decision.

[0077] The priority experience playback mechanism is the core of the system's learning ability. The system records the experience of each verification, including the verification measures taken, system state changes, and verification results. By calculating the timing difference error, the system can identify which experiences are more valuable. For example, if a verification successfully intercepts an unauthorized access, the sampling probability of this experience will increase, allowing the system to learn more about how to deal with such situations.

[0078] On this basis, the system performs three layers of security verification. Identity authentication is the first line of defense, verifying the user's identity information, which may include biometric recognition, password verification, etc. Permission authentication ensures that users can only access authorized resources, which requires a precise access control matrix. The most complex is abnormal behavior monitoring verification, where the system analyzes the user's behavior patterns and identifies suspicious operation sequences. For example, if a user accesses a locker at an unconventional time, or changes the access items multiple times in a short period of time, the system will increase its level of vigilance.

[0079] The advantage of the Q-learning framework is that it can dynamically adjust the verification strategy. For example, when the network load is low, the system may choose a more stringent verification process; while during peak hours, a more efficient verification scheme may be adopted. This adaptive mechanism ensures a balance between security and efficiency. At the same time, through continuous learning and optimization, the system can gradually improve its ability to identify and respond to various security threats.

[0080] When all verifications are passed, the system will execute cloud storage processing. This multi-level security verification mechanism, combined with the intelligent decision-making ability of Q learning, provides a strong security guarantee for the locker system. It can not only effectively prevent common security threats, but also adapt to new attack methods, realizing the continuous evolution of security protection.

[0081] The specific process and technical implementation scheme for performing multi-level security verification based on the Q-learning service framework are as follows: The Q-learning service framework uses a three-layer verification architecture when performing security verification: identity authentication layer, permission authentication layer, and abnormal behavior monitoring layer. In the identity authentication layer, the system first verifies the identity credentials provided by the user. This includes feature extraction and matching of biometric data (such as fingerprints and facial features), as well as verifying the user's password and token information. The system uses a deep neural network to process biometric data and evaluates the degree of matching by calculating the cosine similarity of feature vectors.

[0082] The authorization layer implements a hybrid mechanism of role-based access control (RBAC) and attribute-based access control (ABAC). The system maintains a dynamically updated access control matrix that records the access rights of different roles to different resources. Each access request triggers the authorization verification process, and the system checks the user's role attributes, time restrictions, location restrictions and other constraints in multiple dimensions. The Q learning algorithm dynamically adjusts the strictness of the authorization policy by observing the user's access pattern.

[0083] The abnormal behavior monitoring layer is the most complex verification link. The system builds a user behavior model to identify suspicious activities. This model is based on multiple feature dimensions, including access time patterns, operation sequence characteristics, access location distribution, etc. The Q learning algorithm continuously observes user behavior and learns the characteristic distribution of normal behavior. When a behavior that deviates from the normal pattern is detected, the system will take corresponding security measures based on the degree of deviation.

[0084] In actual execution, the Q-learning algorithm evaluates the effects of different verification strategies through the state-action value function (Q function). The state space includes multiple dimensions such as the current system load, historical user behavior characteristics, and risk assessment results. The action space contains different levels of verification measures, from basic password verification to multi-factor authentication. The system selects the optimal verification strategy based on the current state and continuously optimizes the decision-making process through a reward mechanism.

[0085] In order to improve the efficiency of verification, the system implements a hierarchical verification strategy. For low-risk operations, only basic identity verification may be performed; for high-risk operations, more stringent multi-factor authentication will be triggered. The Q learning algorithm continuously accumulates and learns the processing experience of various verification scenarios through the experience replay mechanism, gradually improving the accuracy and efficiency of the verification strategy.

[0086] The entire verification process is dynamically adaptive, and the system will continuously adjust policy parameters based on the verification results. Successful verification operations will strengthen the corresponding policy selection, while failed verifications will trigger policy adjustments. This adaptive mechanism ensures that the system can optimize user experience and system performance while maintaining security.

[0087] S6. Based on the verification result, a digital file of the items in the savings cabinet is generated and cloud synchronization is performed.

[0088] In step S6, the cloud storage process is further performed: The received standardized item feature representation is integrity checked using hash chain technology. If the integrity check passes, the standardized item feature representation is sharded using a consistent hash algorithm, and the data is distributed and stored in cloud storage nodes. Based on the data security level represented by the standardized item characteristics, a corresponding encryption algorithm is used to encrypt the data.

[0089] Optionally, generating a digital archive of the items in the savings cabinet based on the verification result and performing cloud synchronization processing includes: Based on the verification result, a multi-level archive structure of the digital archive is generated; wherein the multi-level archive structure of the digital archive includes a basic information layer, a feature representation layer and an associated information layer; Based on the multi-level archive structure of the digital archive, a multi-dimensional retrieval index is established to support cross-level data query operations.

[0090] The system uses hash chain technology to verify the integrity of item feature data, which is like creating a unique digital fingerprint for each piece of data. For example, when the feature data of a piece of jewelry is uploaded, the system will generate a hash value based on the data content and link this hash value with the hash value of the previous data. This chain structure ensures that the data cannot be tampered with, because any slight change will cause the subsequent hash value to change.

[0091] After the data passes the integrity check, the system uses the consistent hashing algorithm to shard the data. This process can be compared to the partitioned storage system of a library. Assuming that we have multiple cloud storage nodes, the system will intelligently determine how the data is distributed based on the characteristics of the data and the status of the storage nodes. For example, data on items that are frequently accessed together will be stored on nearby nodes, which can improve access efficiency. The use of the consistent hashing algorithm ensures that even if a storage node fails, the system can quickly redistribute data and maintain service continuity.

[0092] In terms of data security, the system selects encryption algorithms of different strengths according to the security level represented by the characteristics of the items. For example, for the basic information of ordinary items, standard AES encryption may be used; while for the sensitive information of high-value items, stronger encryption schemes may be used, such as quantum-resistant encryption algorithms. This hierarchical encryption strategy optimizes system performance while ensuring security.

[0093] In terms of archive management, the system has established a carefully designed multi-level archive structure. The basic information layer records the basic attributes of the item, such as size, weight, storage time, etc. This is like an ID card for the item, containing the most basic identification information. The feature representation layer stores the deep features of the item, including previously extracted time series features, behavior patterns, etc. This layer is more like a detailed portrait of the item, recording its various characteristics and change patterns. The associated information layer records the relationship between the item and other entities, such as ownership, usage history, related transaction records, etc. This layer of information builds a relationship network of items in a larger system.

[0094] In order to achieve efficient data retrieval, the system has established a multi-dimensional retrieval index on top of this multi-level structure. This index structure allows users to query information from different angles. For example, users can query all items stored in a certain period by time, or query items with similar usage patterns by features, or even query all items of a certain user by association. This flexible retrieval mechanism greatly improves the efficiency of the system.

[0095] The design of this entire storage and archive management system embodies the perfect balance of security, efficiency and ease of use. Through the layered data structure and intelligent storage strategy, the system not only ensures the security and integrity of the data, but also provides powerful retrieval and analysis capabilities, providing a solid technical foundation for the intelligent management of savings cabinets.

[0096] Optionally, the performing of cloud synchronization processing includes: Inputting the archive data of the digital archive into the distributed transaction framework for processing to ensure the consistency of data operations across nodes and obtain processed archive data; Deploy a multi-level cache architecture, deploy local cache at edge nodes to store frequently accessed processed archive data, use a distributed cache cluster in the cloud and implement load balancing through a consistent hashing algorithm; The master-slave switching mechanism of the Raft consensus algorithm is adopted to perform service recovery when the edge node failure is detected.

[0097] First of all, the implementation of the distributed transaction framework is the basis of the entire synchronization process. Assume a scenario: when a user deposits valuables at a savings bank outlet in Beijing, the system needs to update the local database, regional server, and central cloud storage at the same time. These cross-node data operations must maintain consistency, just like a bank's transfer operation, either all succeed or all fail. The distributed transaction framework achieves this through a two-phase commit protocol: in the first phase, the system asks all participating nodes whether they are ready to receive updates; in the second phase, only when all nodes confirm that they are ready, the system will perform the actual data update operation. This ensures the consistency of data on all nodes.

[0098] The design of the multi-level cache architecture reflects the performance optimization concept of the system. Deploying local cache at edge nodes (such as various savings locker outlets) is like setting up a small data center at each outlet. For example, if a customer frequently visits a specific locker, the relevant data of this locker will be cached locally, greatly reducing access latency. At the same time, the distributed cache cluster in the cloud balances the load through a consistent hashing algorithm. This is like setting up multiple toll booths on a highway and ensuring the efficiency of each channel through intelligent diversion. When a cache node is overloaded, the system will automatically transfer some requests to other nodes to maintain a balance in overall performance.

[0099] The application of the Raft consensus algorithm solves the problem of system reliability. The algorithm works like an efficient team collaboration mechanism. Under normal circumstances, the master node is responsible for processing all write operation requests, while the backup node keeps the data synchronized in real time. When the system detects a failure in an edge node (such as a network outage or hardware failure), the Raft algorithm immediately starts the master-slave switching process. This process is divided into three steps: first, the surviving nodes will conduct a leader election; then, the new master node will confirm the consistency of the data; finally, the system resumes normal service. The whole process is like in a symphony orchestra, when the conductor suddenly leaves, the deputy conductor can take over seamlessly and ensure that the concert continues.

[0100] In actual operation, these three mechanisms work together to form a powerful data synchronization and fault recovery system. For example, when the network connection of an edge node is detected to be unstable, the system will actively increase the frequency of data backup and prepare backup nodes in advance. If a failure does occur, the Raft algorithm can complete the master-slave switch in milliseconds, and the distributed transaction framework ensures that no data is lost in the process. At the same time, the multi-level cache architecture can maintain basic service capabilities during fault recovery to minimize the impact on users.

[0101] This complex and sophisticated synchronization mechanism ensures that the savings cabinet system can maintain stable and reliable service quality in the face of various challenges. It not only meets the basic data consistency requirements, but also provides excellent performance and reliability through multi-level optimization, providing strong technical support for the stable operation of the entire savings cabinet system.

[0102] Figure 2 This is a schematic diagram of a savings cabinet entry system based on a cloud platform provided in an embodiment of the present application. Figure 5 The cloud platform-based savings cabinet entry system 500 includes: A data acquisition module 501 is used to collect multimodal data of the savings cabinet, wherein the multimodal data includes an image of an item, environmental parameters, and user authentication information; The segmentation processing module 502 is used to perform multimodal item segmentation processing on the multimodal data to obtain a segmentation result; wherein the multimodal item segmentation processing includes: using an adversarial training network architecture to perform segmentation processing on the multimodal data to generate an initial segmentation map; using a conditional random field to optimize the initial segmentation map, optimize the spatial consistency of pixel labels in the initial segmentation map, and obtain the optimized segmentation result; A feature extraction module 503 is used to input the segmentation result into a time-aware diffusion transformation model for processing to extract temporal features; A feature quantization module 504 is used to perform quantization mapping processing on the time series features, map the time series features to a standard feature space, and obtain a standardized feature representation of the item; The security verification module 505 is used to perform multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, cloud storage processing is performed; The file generation module 506 is used to generate a digital file of the savings cabinet items based on the verification result and perform cloud synchronization processing.

[0103] In some embodiments, the segmentation processing module 502 includes: Constructing an adversarial training network including a generator and a discriminator; wherein the generator adopts a U-Net structure including an encoder and a decoder module, and the discriminator adopts a fully convolutional network structure; Inputting the multimodal data into the generator for processing to generate the initial segmentation map; Using the discriminator to perform authenticity evaluation on the initial segmentation map to obtain an evaluation result; Based on the evaluation results, parameters of the generator and the discriminator are optimized.

[0104] In some embodiments, the segmentation processing module 502 further includes: Construct an energy function including data terms and smooth terms; The energy function is used to process the initial segmentation map to obtain a preliminary optimized segmentation result; wherein the data item calculates the label probability that each pixel of the initial segmentation map belongs to a different label category, and the smoothing item calculates the label association degree between adjacent pixels of the initial segmentation map; Using an inference algorithm of an iterative conditional mode to optimize the energy function, thereby obtaining an optimized energy function; The optimized energy function is used to optimize the preliminary optimized segmentation result, and finally the segmentation result after multiple rounds of iterative optimization is output; Combined with the structured learning method, various weight parameters in the energy function are dynamically adjusted.

[0105] In some embodiments, the feature extraction module 503 includes: Constructing a multi-head self-attention mechanism, converting the features of the input segmentation result into a query vector, a key vector and a value vector respectively through linear mapping, and obtaining the features after attention processing; The features after the attention processing are processed by combining position coding and time coding to obtain features before coding; wherein the position coding uses sine-cosine functions to calculate the spatial position information of the features, and the time coding uses a learnable parameter matrix to calculate the temporal dependency of the features; The pre-encoded features are input into a hierarchical diffusion structure for processing to obtain the temporal features; wherein the hierarchical diffusion structure includes multiple layers of encoders and decoders, the encoder converts the input features into potential representations by adding noise, and the decoder reconstructs the temporal features through a denoising process.

[0106] In some embodiments, the feature quantization module 504 includes: Constructing a vector quantizer, the vector quantizer comprising a codebook of a preset number of codewords and a feature mapping network; Performing normalization preprocessing on the time series features, and using a nearest neighbor search algorithm to retrieve matching codewords in the codebook; A soft allocation mechanism is used to quantize and map the temporal features, allowing a feature vector to be matched with multiple codewords at the same time, and generating the standardized item feature representation according to the matching weight.

[0107] In some embodiments, the security verification module 505 includes: Construct a Q-learning service framework, which includes: Set up a state space containing system resource usage and network load, and an action space containing service scheduling strategies; Build a dual Q network structure including the main network and the target network; Set up a priority experience replay mechanism to dynamically adjust the sample sampling probability according to the time series difference error; Based on the Q-learning service framework, multi-level security verification is performed on the standardized item feature representation; wherein the multi-level security verification includes identity authentication, authority authentication and abnormal behavior monitoring verification.

[0108] In some embodiments, the security verification module 505 further includes: The received standardized item feature representation is integrity checked using hash chain technology. If the integrity check passes, the standardized item feature representation is sharded using a consistent hash algorithm, and the data is distributed and stored in cloud storage nodes. Based on the data security level represented by the standardized item characteristics, a corresponding encryption algorithm is used to encrypt the data.

[0109] In some embodiments, the profile generation module 506 includes: Based on the verification result, a multi-level archive structure of the digital archive is generated; wherein the multi-level archive structure of the digital archive includes a basic information layer, a feature representation layer and an associated information layer; Based on the multi-level archive structure of the digital archive, a multi-dimensional retrieval index is established to support cross-level data query operations.

[0110] In some embodiments, the profile generation module 506 further includes: Inputting the archive data of the digital archive into the distributed transaction framework for processing to ensure the consistency of data operations across nodes and obtain processed archive data; Deploy a multi-level cache architecture, deploy local cache at edge nodes to store frequently accessed processed archive data, use a distributed cache cluster in the cloud and implement load balancing through a consistent hashing algorithm; The master-slave switching mechanism of the Raft consensus algorithm is adopted to perform service recovery when the edge node failure is detected.

[0111] The basic principles of the present disclosure are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present disclosure are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present disclosure. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, and are not limitations. The above details do not limit the present disclosure to the necessity of adopting the above specific details to be implemented.

[0112] In the present disclosure, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. The block diagrams of the devices, devices, equipment, and systems involved in the present disclosure are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, devices, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0113] Additionally, as used herein, "or" used in a list of items beginning with "at least one" indicates a separate list, so that, for example, a list of "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not mean that the example described is preferred or better than other examples.

[0114] It should also be noted that in the system and method of the present disclosure, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present disclosure.

[0115] Various changes, substitutions, and modifications of the techniques described herein may be made without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of the present disclosure is not limited to the specific aspects of the processes, machines, manufactures, compositions of events, means, methods, and actions described above. Currently existing or later to be developed processes, machines, manufactures, compositions of events, means, methods, or actions that perform substantially the same functions or achieve substantially the same results as the corresponding aspects described herein may be utilized. Thus, the appended claims include such processes, machines, manufactures, compositions of events, means, methods, or actions within their scope.

[0116] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present disclosure. Therefore, the present disclosure is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0117] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present disclosure to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for entering data into a savings cabinet based on a cloud platform, characterized in that: include: Collecting multimodal data of the savings cabinet, wherein the multimodal data includes an image of an item, environmental parameters, and user authentication information; Performing multimodal object segmentation processing on the multimodal data to obtain a segmentation result; wherein the multimodal object segmentation processing includes: Using an adversarial training network architecture to segment the multimodal data and generate an initial segmentation map; Using a conditional random field to optimize the initial segmentation map, optimizing the spatial consistency of pixel labels in the initial segmentation map, and obtaining the optimized segmentation result; Inputting the segmentation results into a time-aware diffusion transformation model for processing to extract temporal features; Performing quantization mapping processing on the time series features, mapping the time series features to a standard feature space, and obtaining a standardized feature representation of the item; Performing multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, performing cloud storage processing; Based on the verification results, a digital file of the savings cabinet items is generated and cloud synchronization is performed.

2. The method for entering data into a savings cabinet according to claim 1, characterized in that: The segmentation process of the multimodal data using the adversarial training network architecture includes: Constructing an adversarial training network including a generator and a discriminator; wherein the generator adopts a U-Net structure including an encoder and a decoder module, and the discriminator adopts a fully convolutional network structure; Inputting the multimodal data into the generator for processing to generate the initial segmentation map; Using the discriminator to perform authenticity evaluation on the initial segmentation map to obtain an evaluation result; Based on the evaluation results, parameters of the generator and the discriminator are optimized.

3. The method for entering data into a savings cabinet according to claim 1, characterized in that: The using conditional random field to optimize the initial segmentation map includes: Construct an energy function including data terms and smooth terms; The energy function is used to process the initial segmentation map to obtain a preliminary optimized segmentation result; wherein the data item calculates the label probability that each pixel of the initial segmentation map belongs to a different label category, and the smoothing item calculates the label association degree between adjacent pixels of the initial segmentation map; Using an inference algorithm of an iterative conditional mode to optimize the energy function, thereby obtaining an optimized energy function; The optimized energy function is used to optimize the preliminary optimized segmentation result, and finally the segmentation result after multiple rounds of iterative optimization is output; Combined with the structured learning method, various weight parameters in the energy function are dynamically adjusted.

4. The method for entering data into a savings cabinet according to claim 1, characterized in that: The step of inputting the segmentation result into a time-aware diffusion transformation model for processing to extract temporal features includes: Constructing a multi-head self-attention mechanism, converting the features of the input segmentation result into a query vector, a key vector and a value vector respectively through linear mapping, and obtaining the features after attention processing; The features after the attention processing are processed by combining position coding and time coding to obtain features before coding; wherein the position coding uses sine-cosine functions to calculate the spatial position information of the features, and the time coding uses a learnable parameter matrix to calculate the temporal dependency of the features; The pre-encoded features are input into a hierarchical diffusion structure for processing to obtain the temporal features; wherein the hierarchical diffusion structure includes multiple layers of encoders and decoders, the encoder converts the input features into potential representations by adding noise, and the decoder reconstructs the temporal features through a denoising process.

5. The method for entering data into a savings cabinet according to claim 1, characterized in that: The performing of quantization mapping processing comprises: Constructing a vector quantizer, the vector quantizer comprising a codebook of a preset number of codewords and a feature mapping network; Performing normalization preprocessing on the time series features, and using a nearest neighbor search algorithm to retrieve matching codewords in the codebook; A soft allocation mechanism is used to quantize and map the temporal features, allowing a feature vector to be matched with multiple codewords at the same time, and generating the standardized item feature representation according to the matching weight.

6. The method for entering data into a savings cabinet according to claim 1, characterized in that: The step of performing multi-level security verification on the standardized item feature representation to obtain a verification result; wherein when the verification result is a verification pass state, performing cloud storage processing includes: Construct a Q-learning service framework, which includes: Set up a state space containing system resource usage and network load, and an action space containing service scheduling strategies; Build a dual Q network structure including the main network and the target network; Set up a priority experience replay mechanism to dynamically adjust the sample sampling probability according to the time difference error; Based on the Q-learning service framework, multi-level security verification is performed on the standardized item feature representation; wherein the multi-level security verification includes identity authentication, authority authentication and abnormal behavior monitoring verification.

7. The method for entering data into a savings cabinet according to claim 1, characterized in that: The executing cloud storage processing also includes: The received standardized item feature representation is integrity checked using hash chain technology. If the integrity check passes, the standardized item feature representation is sharded using a consistent hash algorithm, and the data is distributed and stored in cloud storage nodes. Based on the data security level represented by the standardized item characteristics, a corresponding encryption algorithm is used to encrypt the data.

8. The method for entering data into a savings cabinet according to claim 1, characterized in that: The method of generating a digital file of the items in the savings cabinet based on the verification result and performing cloud synchronization processing includes: Based on the verification result, a multi-level archive structure of the digital archive is generated; wherein the multi-level archive structure of the digital archive includes a basic information layer, a feature representation layer and an associated information layer; Based on the multi-level archive structure of the digital archive, a multi-dimensional retrieval index is established to support cross-level data query operations.

9. The method for entering data into a savings cabinet according to claim 1, characterized in that: The executing cloud synchronization process includes: Inputting the archive data of the digital archive into the distributed transaction framework for processing to ensure the consistency of data operations across nodes and obtain processed archive data; Deploy a multi-level cache architecture, deploy local cache at edge nodes to store frequently accessed processed archive data, use a distributed cache cluster in the cloud and implement load balancing through a consistent hashing algorithm; The master-slave switching mechanism of the Raft consensus algorithm is adopted to perform service recovery when the edge node failure is detected.

10. A savings cabinet entry system based on a cloud platform, characterized in that: include: A data acquisition module, used to collect multimodal data of the savings cabinet, wherein the multimodal data includes an image of an item, environmental parameters, and user authentication information; A segmentation processing module is used to perform multimodal item segmentation processing on the multimodal data to obtain a segmentation result; wherein the multimodal item segmentation processing includes: using an adversarial training network architecture to perform segmentation processing on the multimodal data to generate an initial segmentation map; using a conditional random field to optimize the initial segmentation map, optimize the spatial consistency of pixel labels in the initial segmentation map, and obtain the optimized segmentation result; A feature extraction module, used for inputting the segmentation result into a time-aware diffusion transformation model for processing to extract temporal features; A feature quantization module, used to perform quantization mapping processing on the time series features, map the time series features to a standard feature space, and obtain a standardized feature representation of the item; A security verification module, used to perform multi-level security verification on the standardized item feature representation to obtain a verification result; wherein, when the verification result is a verification pass state, cloud storage processing is performed; The file generation module is used to generate a digital file of the savings cabinet items based on the verification result and perform cloud synchronization processing.

Images