An ai internet of things multi-domain data sharing method and system based on multi-source perception fusion

Abstract

The application relates to the technical field of artificial intelligence and Internet of Things, and discloses an AI Internet of Things multi-domain data sharing method and system based on multi-source perception fusion, which is applied to an Internet of Things edge device provided with a consortium chain and a distributed storage subsystem. The method comprises the following steps: acquiring heterogeneous perception data and identifying entity targets; extracting feature mapping into dynamic space-time perception graph nodes, aligning vision and sensing state through a cross-modal space-time attention mechanism to generate fusion features; inputting a request task, bandwidth and trust level into a reinforcement learning network to output a strategy, performing hierarchical extraction and noise addition on the fusion features to generate desensitization features; storing the desensitization features into the distributed storage to obtain a hash value and generate a symmetric key, generating cross-domain ciphertext credentials through attribute-based encryption; publishing the ciphertext credentials to a smart contract, verifying whether the attributes meet the strategy, and issuing the credentials and chaining the credentials for storage when the attributes meet the strategy. The application realizes high-precision fusion, low communication cost and fine-grained privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and Internet of Things (IoT) technology, and in particular to an AI IoT multi-domain data sharing method and system based on multi-source sensing fusion. Background Technology

[0002] With the deep integration of artificial intelligence and the Internet of Things (AIoT) technologies, massive amounts of multi-source sensing data have been generated by numerous edge devices. However, existing AIoT data sharing systems typically employ a centralized "aggregate first, then share" model when processing massive amounts of heterogeneous sensing data, facing multiple technical bottlenecks: First, multi-source heterogeneous data (such as visual images and high-frequency time-series sensor data) are difficult to align precisely in terms of timestamps and spatial dimensions, and simple splicing and fusion can easily lead to semantic loss and decision bias; second, there are natural trust barriers between different business domains (such as traffic management and commercial insurance), and directly sharing raw sensing data not only incurs huge communication overhead but also easily leads to serious privacy leaks and data misuse risks; finally, existing systems generally lack dynamic data desensitization and adaptive transmission mechanisms based on data value, requester trust levels, and real-time network status, resulting in a serious waste of edge computing power and network bandwidth resources. Therefore, there is an urgent need for a multi-domain data sharing solution that can achieve deep integration of heterogeneous data, provide fine-grained privacy protection, and dynamically adapt to the network environment. Summary of the Invention

[0003] This invention provides an AI IoT multi-domain data sharing method and system based on multi-source sensing fusion. It not only achieves high-precision spatiotemporal fusion of multi-source heterogeneous sensing data, but also ensures absolute security, tamper-proof traceability, and fine-grained privacy protection for cross-domain data sharing while significantly reducing communication overhead through reinforcement learning dynamic desensitization and blockchain attribute-based encryption mechanisms.

[0004] This invention provides an AI IoT multi-domain data sharing method based on multi-source sensing fusion, applicable to IoT edge devices deployed with consortium blockchains and distributed storage subsystems. The method includes:

[0005] S1. Acquire multi-source heterogeneous sensing data in the target physical environment and identify entity targets; wherein, the multi-source heterogeneous sensing data includes image and video streams containing the entity targets and time-series sensing signals reflecting the state of the entity targets;

[0006] S2. Extract the features of the image video stream and the time-series sensing signal respectively, and map them to the node features of the dynamic spatiotemporal perception map constructed with the entity target as the node. Calculate the dynamic cross-correlation between the node features through a cross-modal spatiotemporal attention mechanism to align the visual motion state and sensing abnormal state of the entity target and generate a multi-source fusion feature sequence.

[0007] S3. Receive a cross-domain sharing request initiated by the target request domain node, carrying business domain attributes and task type, obtain the current communication bandwidth status and query the trust evaluation level of the target request domain node on the consortium blockchain, input the task type, current communication bandwidth status and trust evaluation level into a preset reinforcement learning decision network, and output a target semantic extraction strategy to perform hierarchical extraction and noise addition on the multi-source fusion feature sequence to generate desensitized semantic features.

[0008] S4. Store the desensitized semantic features into the distributed storage subsystem, obtain the data addressing hash value and generate a symmetric decryption key, construct an access control tree policy according to the business domain attributes, and encrypt the data addressing hash value and the symmetric decryption key using the ciphertext policy attribute base encryption algorithm to generate a cross-domain ciphertext credential.

[0009] S5. Publish the shared transaction containing the cross-domain encrypted certificate and the access control tree policy to the smart contract of the consortium blockchain, so that when the smart contract verifies that the attributes of the target request domain node meet the access control tree policy, it issues the cross-domain encrypted certificate and packages the data cross-domain sharing flow record onto the blockchain for evidence storage.

[0010] Furthermore, S1 specifically includes:

[0011] S101. Perform global clock synchronization on visual sensors and IoT sensing devices deployed in the target physical environment, and set a global time reference.

[0012] S102. Based on the global time reference, the continuous image and video stream sequence with a first sampling frequency output by the visual sensor and the time-series sensing signal sequence with a second sampling frequency output by the IoT sensing device are acquired in real time and synchronously. The second sampling frequency is higher than the first sampling frequency.

[0013] S103. Using the low-frequency sampling timestamp of the continuous image and video stream sequence as the reference anchor point, the high-frequency time-series sensing signal sequence is subjected to denoising and windowing aggregation processing to obtain a preliminary time-aligned time-series feature segment.

[0014] S104. Input the video frames of the continuous image video stream sequence into a preset deep learning target detection model, identify the two-dimensional bounding box coordinates of the entity target, and spatially associate and anchor the entity target with its dedicated IoT sensing device according to the pre-calibrated physical environment topology mapping relationship.

[0015] Furthermore, in S4, the spatial association anchoring specifically refers to:

[0016] Obtain the global three-dimensional coordinate system of the target physical environment, and the calibration physical coordinates of each IoT sensing device in the global three-dimensional coordinate system;

[0017] Obtain the camera intrinsic and extrinsic parameter matrices of the visual sensor, and based on the camera intrinsic and extrinsic parameter matrices, back-project the two-dimensional bounding box coordinates output by the deep learning object detection model into three-dimensional space estimated coordinates;

[0018] Calculate the spatial distance between the estimated three-dimensional spatial coordinates of the entity target and the calibrated physical coordinates of each of the IoT sensing devices;

[0019] The nearest IoT sensor device with a spatial distance less than a preset threshold is identified as the dedicated sensor device attached to the physical target, thus completing the spatial topological association between the physical target and the time-series sensor data.

[0020] Furthermore, S2 specifically includes:

[0021] S201. Construct a neural network structure containing a dual-branch feature extractor, extract visual motion state features from the image video stream to obtain a visual feature vector, and extract sensor abnormal state features from the time-series sensing signal to obtain a time-series sensing feature vector.

[0022] S202. Construct a dynamic spatiotemporal sensing map, with the entity target as the node set of the map, the spatial topological relationship between the entity targets as the edge set of the map, and map the visual feature vector and the temporal sensing feature vector to the initial multimodal feature representation of the node;

[0023] S203. Introduce a cross-modal spatiotemporal attention mechanism on the dynamic spatiotemporal perception map, calculate the dynamic cross-correlation between the visual motion state and the sensing abnormal state, and obtain the sensing feature update representation guided by visual features and the visual feature update representation guided by sensing features, respectively.

[0024] S204. Perform feature-level fusion and dimensionality reduction processing on the depth-aligned sensor feature update representation and the visual feature update representation, and output the multi-source fusion feature sequence.

[0025] Furthermore, S203 specifically includes:

[0026] Introducing the first set of learnable weight matrices The visual feature vector Mapped to query vector matrix The time-series sensing feature vector Mapped to a key vector matrix Sum value vector matrix ; through formula Calculate the dynamic cross-correlation attention weight matrix A, where, The feature dimension of the key vector matrix is... The scaling factor is used; the attention weight matrix A is multiplied by the value vector matrix V to obtain the sensor feature update representation after weighting and aligning by the visual motion state. ;

[0027] Introducing a second set of learnable weight matrices The time-series sensing feature vector Mapped to a sensor query vector matrix The visual feature vector Mapped to a visual key vector matrix and visual value vector matrix ; through formula Calculate the reverse attention weight matrix A'; combine the reverse attention weight matrix A' with the visual value vector matrix. Multiplying these results in a weighted and aligned representation of the visual features, obtained from the sensor anomaly states. .

[0028] Furthermore, S3 specifically includes:

[0029] S301. Parse the target request domain node identity, business domain attribute and request task type carried in the cross-domain sharing request. Based on the identity, query the historical behavior record of the node in the distributed ledger of the consortium blockchain to obtain the trust assessment level, and monitor the physical communication bandwidth status between the current IoT edge node and the request domain node in real time.

[0030] S302. Construct a reinforcement learning decision-making agent, defining the environment state space as a vector containing the trust assessment level, the physical communication bandwidth state, and the request task type, and defining the action space as a discretized set of target semantic extraction deep policies; wherein, the environment state space S consists of normalized state vectors. composition:

[0031]

[0032] in, The trust assessment level is... This refers to the physical communication bandwidth status. The quantified request task type;

[0033] The action space A consists of action vectors The action vector comprises semantic extraction layers. Sum of differential privacy noise budget The semantic extraction level corresponds to the kernel size or step size of the pooling operation on the multi-source fusion feature sequence, and the differential privacy noise budget determines the variance intensity of the injected random noise.

[0034] S303. Input the current environmental state vector into a pre-set reinforcement learning decision network, and output the optimal action at the current moment based on the objective of maximizing the cumulative reward function, i.e., the target semantic extraction depth strategy; wherein, the cumulative reward function Defined as:

[0035]

[0036] in, The target semantic extraction depth strategy selected at the current moment; In strategy The following data availability scoring function; This is a privacy breach risk function, the value of which is related to the trust assessment level. There is a negative correlation, meaning that the lower the trust assessment level, the greater the risk penalty value when choosing a high-precision semantic extraction strategy; This is a transmission delay function, the value of which is related to the physical communication bandwidth status. It shows a negative correlation with strategy The amount of data generated is positively correlated; These are non-negative coefficients used to adjust the weights of data utility, privacy risk, and transmission delay, respectively.

[0037] S304. According to the target semantic extraction depth strategy, perform hierarchical feature extraction operation on the multi-source fusion feature sequence to compress the feature dimension, and use differential privacy mechanism to inject random noise into the extracted feature vector to generate the desensitized semantic features.

[0038] Furthermore, S4 specifically includes:

[0039] S401. Use a pseudo-random number generator to generate a symmetric decryption key, use the symmetric decryption key to symmetrically encrypt the desensitized semantic features to generate ciphertext data blocks, upload the ciphertext data blocks to the distributed storage subsystem and obtain the content addressing hash value.

[0040] S402. Based on the business scenarios and preset security rules involved in the cross-domain sharing request, dynamically construct an access control tree strategy that includes attribute leaf nodes and threshold logic gates.

[0041] S403. The symmetric decryption key and the content addressing hash value are concatenated and combined to form the message to be encrypted. The message to be encrypted is encrypted using the ciphertext policy attribute base encryption algorithm and based on the access control tree policy to generate attribute base ciphertext.

[0042] S404. Encapsulate the attribute-based ciphertext, access control tree policy, and digital signature of the edge computing node into a cross-domain ciphertext credential; wherein, the data structure of the cross-domain ciphertext credential CCC is defined as follows:

[0043]

[0044] in, This serves as a unique identifier for this cross-domain data sharing session; The access control tree policy, in plaintext form, is used by smart contracts to verify the permissions of the requester. The attribute base ciphertext encapsulates the key information; The timestamp used to generate the credentials is used to prevent replay attacks. The edge computing node generates a digital signature for the above fields using its own private key, which is used to prove the authenticity and non-repudiation of the data source.

[0045] Furthermore, S403 specifically includes:

[0046] Obtain the system master public key generated during the system initialization phase. ;

[0047] Construct a message M to be encrypted, wherein the message M is the symmetric decryption key. With content addressing hash value splicing, that is ;

[0048] The attribute-based ciphertext is generated by performing an encryption operation using the ciphertext policy attribute-based encryption algorithm. ;in, This refers to the access control tree policy.

[0049] Furthermore, S5 specifically includes:

[0050] S501. The edge computing node calls the interface of the smart contract to broadcast the data sharing and publishing transaction containing the cross-domain encrypted certificate to the consortium blockchain network. After verifying the digital signature of the edge computing node, the smart contract stores the cross-domain encrypted certificate and broadcasts the event log.

[0051] S502. When the target request domain node listens to the event log, it initiates a data access transaction carrying a digital identity certificate and attribute credential set to the smart contract. The smart contract automatically triggers the permission verification logic and logically matches the attribute credential set with the access control tree policy in the cross-domain encrypted credential.

[0052] S503. If the match is successful, the smart contract sends the cross-domain encrypted credential to the target requesting domain node. The target requesting domain node calls the decryption function of the encrypted policy attribute-based encryption algorithm and uses its own set of attribute credentials. The attribute-based ciphertext in the cross-domain ciphertext certificate Decrypt and extract the symmetric decryption key. and content addressing hash :

[0053]

[0054] Using the content addressing hash value Download the corresponding encrypted data block from the distributed storage subsystem and use the symmetric decryption key. The encrypted data block is symmetrically decrypted to restore the de-identified semantic features;

[0055] S504. The smart contract generates a data transfer record containing information about the cross-domain data sharing and transfer. After the consensus nodes of the consortium blockchain network verify the data transfer record, they package it into a new block and attach it to the blockchain ledger. The data transfer record includes the identity hash of the data provider, the identity hash of the data requester, the session ID of the accessed data, the timestamp of the access, and the hash result of the smart contract execution.

[0056] This invention also provides an AI IoT multi-domain data sharing system based on multi-source sensing fusion, which, based on the multi-source sensing fusion-based AI IoT multi-domain data sharing method described above, includes:

[0057] The acquisition module is used to acquire multi-source heterogeneous sensing data in the target physical environment and identify entity targets; wherein, the multi-source heterogeneous sensing data includes image and video streams containing the entity targets and time-series sensing signals reflecting the state of the entity targets;

[0058] The extraction module is used to extract features from the image video stream and the temporal sensing signal respectively, and map them to node features of a dynamic spatiotemporal perception map constructed with the entity target as the node. The dynamic cross-correlation between the node features is calculated through a cross-modal spatiotemporal attention mechanism to align the visual motion state and sensing abnormal state of the entity target and generate a multi-source fusion feature sequence.

[0059] The generation module is used to receive cross-domain sharing requests initiated by the target request domain node, which carry business domain attributes and task types, obtain the current communication bandwidth status and query the trust evaluation level of the target request domain node on the consortium blockchain, input the task type, current communication bandwidth status and trust evaluation level into a preset reinforcement learning decision network, and output the target semantic extraction strategy to perform hierarchical extraction and noise addition on the multi-source fusion feature sequence to generate desensitized semantic features.

[0060] The encryption module is used to store the desensitized semantic features into the distributed storage subsystem, obtain the data addressing hash value and generate a symmetric decryption key, construct an access control tree policy according to the business domain attributes, and encrypt the data addressing hash value and the symmetric decryption key using the ciphertext policy attribute base encryption algorithm to generate cross-domain ciphertext credentials.

[0061] The publishing module is used to publish the shared transaction containing the cross-domain encrypted credential and the access control tree policy to the smart contract of the consortium blockchain, so that the smart contract issues the cross-domain encrypted credential when verifying that the attributes of the target request domain node meet the access control tree policy, and packages the flow record of this cross-domain data sharing and puts it on the blockchain for evidence storage.

[0062] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0063] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0064] The beneficial effects of this invention are as follows:

[0065] This invention achieves high-precision semantic-level fusion of multi-source heterogeneous sensing data by constructing a dynamic spatiotemporal perception map and introducing a cross-modal spatiotemporal attention mechanism, effectively solving the problem of spatiotemporal misalignment of heterogeneous data. Simultaneously, by combining a reinforcement learning decision network with a differential privacy mechanism, it can dynamically and adaptively adjust the semantic extraction depth and noise level based on the request domain trust and network bandwidth, physically filtering out detailed information from the original image and video streams. Furthermore, in the cross-domain sharing stage, it employs ciphertext policy attribute-based encryption (CP-ABE) combined with automatic authentication via blockchain smart contracts, achieving not only fine-grained privacy protection that is "usable but not visible," but also significantly reducing the transmission bandwidth and communication overhead of edge networks while breaking down multi-domain data silos and achieving tamper-proof traceability. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0067] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0068] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0069] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0070] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0071] like Figure 1 As shown, this invention provides a multi-domain data sharing method for AIoT based on multi-source sensing fusion, applied to IoT edge devices deployed with consortium blockchains and distributed storage subsystems. The method includes:

[0072] S1. Acquire multi-source heterogeneous sensing data in the target physical environment and identify entity targets; wherein, the multi-source heterogeneous sensing data includes image and video streams containing the entity targets and time-series sensing signals reflecting the state of the entity targets.

[0073] In one embodiment, step S1 includes the following sub-steps:

[0074] S101. In the target physical environment (e.g., smart manufacturing workshop, smart traffic intersection, etc.), pre-deploy IoT edge computing nodes and various sensing terminals that communicate with them. The sensing terminals include at least visual sensors (e.g., high-definition network cameras, infrared thermal imagers) for collecting spatial visual information, and IoT sensing devices (e.g., vibration sensors, temperature and humidity sensors, or sound sensors) for collecting one-dimensional continuous physical quantities.

[0075] Preferably, to ensure the consistency of the time domain reference for multi-source heterogeneous data, the edge computing nodes use the Network Time Protocol (NTP) or Precision Time Protocol (PTP) to perform global clock synchronization on the visual sensors and IoT sensing devices, and set a global time reference. .

[0076] S102. The edge computing node receives data collected by the sensing terminal in real time via wired or wireless IoT protocols (such as MQTT, CoAP, or 5G communication protocols). Specifically:

[0077] Based on global time base Acquire a continuous image / video stream sequence with a first sampling frequency output by the visual sensor. ,in This represents the video frame captured at timestamp t;

[0078] Synchronously acquire the time-series sensing signal sequence with a second sampling frequency output by the IoT sensing device. ,in This represents the sensor state value sampled at timestamp t'; wherein the second sampling frequency is higher than the first sampling frequency.

[0079] Those skilled in the art will understand that, due to the significant difference in sampling frequencies between visual sensors and IoT sensing devices (e.g., a video frame rate of 30 FPS, while a vibration sensor sampling rate may be as high as 1000 Hz), the image video stream and the time-series sensing signal are asymmetric heterogeneous data on the time axis.

[0080] S103. For the acquired multi-source heterogeneous sensing data, the edge computing node performs low-level preprocessing to eliminate environmental noise and achieve preliminary alignment, that is, the time-series sensing signal sequence S is subjected to noise reduction processing using Kalman filtering or moving average algorithm to remove abnormal spike data.

[0081] Based on global time base Using the low-sampling timestamps of the image / video stream as anchor points, high-frequency time-series sensing signals are windowed and aggregated (e.g., the mean or peak value within the corresponding time window) to obtain preliminary time-aligned time-series feature segments. In the specific windowing aggregation process, firstly, the low-frequency sampling timestamps of two adjacent video frames in the continuous image / video stream sequence are extracted to construct a time aggregation window; then, the effective sampled values ​​of the time-series sensing signal sequence falling within the time aggregation window are extracted, and the statistical features of the effective sampled values ​​are calculated, including at least the mean, extreme values, or variance; finally, the statistical features are used as the preliminary time-aligned time-series feature segments, ensuring that each segment strictly corresponds to the corresponding image / video stream sequence on the time axis.

[0082] S104. The edge computing node calls a pre-built deep learning object detection model (such as YOLOv8 or Faster R-CNN model) to detect video frames corresponding to timestamp t. Frame-by-frame inference is performed to identify the Region of Interest (ROI) in the image, i.e., the entity target (e.g., a specific type of machine tool, a vehicle traveling on the road). The model outputs a set of two-dimensional bounding box coordinates for the entity target. And the confidence level of the target category;

[0083] Furthermore, based on the pre-defined sensor physical installation mapping relationship, the identified entity target is spatially associated and anchored with its corresponding dedicated temporal sensing signal. For example, if "machine tool A" is visually identified, the vibration sensor signal bound to the spindle of "machine tool A" is used as the dedicated temporal sensing signal reflecting the state of the entity target, thereby providing an accurate entity node and data mapping basis for the subsequent construction of a dynamic spatiotemporal perception map.

[0084] To achieve the aforementioned precise spatial association anchoring at the algorithm's underlying level, the following operations are specifically performed: First, the global three-dimensional coordinate system of the target physical environment and the calibration physical coordinates of each IoT sensing device within this global three-dimensional coordinate system are obtained; second, the camera intrinsic and extrinsic parameter matrices of the visual sensor are obtained, and based on these matrices, the two-dimensional bounding box coordinates output by the deep learning target detection model are inversely projected into estimated three-dimensional spatial coordinates; next, the spatial distance between the estimated three-dimensional spatial coordinates of the entity target and the calibration physical coordinates of each IoT sensing device is calculated; finally, the IoT sensing device with a spatial distance less than a preset threshold and the closest distance is identified as the dedicated sensing device attached to the entity target, thereby completing the spatial topological association between the entity target and the temporal sensing data.

[0085] S2. Extract the features of the image video stream and the time-series sensing signal respectively, and map them to the node features of the dynamic spatiotemporal perception map constructed with the entity target as the node. Calculate the dynamic cross-correlation between the node features through a cross-modal spatiotemporal attention mechanism to align the visual motion state and sensing abnormal state of the entity target, and generate a multi-source fusion feature sequence.

[0086] In one embodiment, step S2 includes the following sub-steps:

[0087] S201, The edge computing node constructs a neural network structure containing a dual-branch feature extractor, which performs deep feature extraction on the initially aligned image / video stream and time-series sensor signal, respectively. Specifically,

[0088] For entity targets (i.e., sensed ROI regions) in the image and video stream sequence, a three-dimensional convolutional neural network (3D-CNN) or a spatial residual network (such as ResNet-50) is used to extract their visual motion state features, resulting in a visual feature vector. This feature vector characterizes the apparent texture, contour deformation, and motion trajectory of the entity target.

[0089] For the associated anchored temporal sensing feature segments, a one-dimensional convolutional neural network (1D-CNN) or a long short-term memory network (LSTM) is used to extract their sensing anomalous state features, resulting in a temporal sensing feature vector. This feature vector represents the implicit physical state of the entity target in the current time period (e.g., high-frequency vibration peaks, temperature gradient abrupt changes).

[0090] S202. After extracting single-modal features, construct a dynamic spatiotemporal awareness map in memory. , where t represents the current time step. The node set V represents the set of all entity targets in the physical environment identified in step S1; the edge set E represents the topological adjacency relationship of each entity target in the physical space. By calculating the Euclidean distance between the estimated 3D coordinates of the entity targets, if the distance is less than the safe interaction threshold, a connected edge is constructed between the two entity nodes; the feature mapping matrix M is used to assign the features extracted in S201 to the graph nodes. Specifically, for any entity target node... , and its corresponding visual feature vector With time-series sensing feature vectors After performing dimensional concatenation or linear mapping, the values ​​are assigned to the node to form the initial multimodal feature representation of the node.

[0091] S203. To deeply couple data relationships between different modalities, a cross-modal spatiotemporal attention mechanism is introduced into the dynamic spatiotemporal perception map. Taking visual modality guiding sensing modality as an example, a first set of learnable weight matrices is introduced. Mapping visual features to a query vector matrix Mapping the sensing features into a key vector matrix Sum value vector matrix Subsequently, the dynamic cross-correlation between visual motion state and sensory abnormal state is calculated, resulting in the attention weight matrix A:

[0092]

[0093] in, The feature dimension of the key vector. This is a scaling factor used to prevent gradient vanishing or exploding.

[0094] By multiplying the attention weight matrix A with the value vector matrix V, the sensor feature update representation after weighted alignment based on visual motion states is obtained:

[0095]

[0096] Similarly, a symmetrical bidirectional attention mechanism is employed. This is done after calculating the feature representation of the vision-guided sensor. Subsequently, a symmetrical network structure is used to compute the visual feature update representation guided by the sensor anomaly state. The specific implementation process is as follows:

[0097] Introducing a second set of learnable weight matrices At this time, the time-series sensing feature vector As the active querying party, it is mapped to a sensor query vector matrix. The visual feature vector As the context being queried, it is mapped to a visual key vector matrix. and visual value vector matrix Subsequently, the reverse attention weight matrix A' for the visual region under the sensor abnormality state is calculated:

[0098]

[0099] In A', higher weight values ​​indicate that when the sensor exhibits a specific abnormal waveform (such as a vibration peak), the corresponding visual image area (such as a specific texture or edge in an image) should receive higher attention. Finally, the reverse attention weight matrix A' is compared with the visual value vector matrix. Multiplying these results in a weighted and aligned representation of the visual features, obtained from the sensor anomaly states. .

[0100] S204. After completing the mutual attention weighting, the visual motion state and sensory abnormal state of the entity target are highly aligned in the semantic feature space. For example, when the machine spindle experiences a visually visible sway, the attention mechanism automatically assigns a higher weight to the abnormal vibration signal at that moment. Finally, the aligned... and Feature-level fusion (such as cascade concatenation or element-wise addition) is performed and fed into a fully connected layer for dimensionality reduction and smoothing, outputting the final multi-source fused feature sequence for the current time step. This fused feature sequence removes redundant pixel information from the original image and highly condenses the comprehensive physical and motion anomaly states of the entity target, serving as the input basis for subsequent reinforcement learning desensitization and cross-domain sharing.

[0101] S3. Receive a cross-domain sharing request initiated by the target request domain node, carrying business domain attributes and task type. Obtain the current communication bandwidth status and query the trust evaluation level of the target request domain node on the consortium blockchain. Input the task type, current communication bandwidth status, and trust evaluation level into a preset reinforcement learning decision network. Output a target semantic extraction strategy to perform hierarchical extraction and noise addition on the multi-source fusion feature sequence to generate desensitized semantic features.

[0102] In one embodiment, step S3 includes the following sub-steps:

[0103] S301, the edge computing node listens for cross-domain data access requests from the network, and when it receives a cross-domain sharing request initiated by the target request domain node... First, the metadata carried in the request is parsed, including: the identity (ID) of the target request domain node, business domain attributes (e.g., traffic management department or commercial insurance company), and request task type (TaskType, e.g., congestion analysis or accident liability determination). Then, multi-dimensional state awareness is executed, specifically including:

[0104] Trust Status Awareness: Using the identity identifier, query the node's historical behavior records in the distributed ledger of the consortium blockchain to obtain its latest trust assessment level. ;

[0105] Network status awareness: Real-time monitoring of the physical communication link between the current IoT edge node and the request domain node to obtain the current available communication bandwidth status. (Unit: Mbps);

[0106] Task requirement awareness: Based on the requested task type, determine the minimum threshold for data accuracy required for that task.

[0107] S302. A reinforcement learning decision agent based on a Deep Q-Network (DQN) or Proximal Policy Optimization (PPO) is pre-configured. To achieve adaptive policy output, the state space and action space are defined. Specifically,

[0108] The state space S consists of the environment state vector at the current time t. Composition, that is The vector is then normalized and input into the decision network.

[0109] Action space A is defined as a discretized set of deep strategies for extracting target semantics, and actions Decision parameters include two dimensions (semantic extraction level) Sum of differential privacy noise budget Specifically, semantic extraction layers In, such as This indicates that all spatiotemporal features are preserved (high precision). This indicates that only keyframe features are retained (medium precision). This indicates that only statistical features are retained (low precision); differential privacy noise budget. This determines the intensity of subsequent noise addition. The smaller the value, the greater the added noise, and the stronger the privacy protection.

[0110] S303, the reinforcement learning decision network makes decisions based on the current input state. Output the optimal action This refers to a deep strategy for extracting target semantics. The training objective of this network is to maximize the cumulative reward function. In this embodiment, the reward function Designed as a comprehensive metric that balances data utility, privacy risks, and transmission latency:

[0111]

[0112] in, Indicates action The following data availability score; This indicates a risk of privacy breaches, which is related to trust levels. Inversely proportional (i.e., the lower the level of trust, the greater the penalty if high-precision data is selected). Indicates transmission delay, relative to current bandwidth. It is inversely proportional to the amount of data; The weighting coefficients are adjustable. Through this mechanism, if the requester has low trust and the network is congested, a low-precision extraction + high-noise strategy will be automatically output; if the requester is a highly trusted public security department handling an urgent task, a high-precision extraction + low-noise strategy will be output.

[0113] S304. Based on the target semantic extraction depth strategy (including the determined semantic extraction level) output in step S303. and noise budget The edge computing nodes generate the multi-source fusion feature sequence in step S2. Perform the specific operations:

[0114] Hierarchical extraction: If the strategy is low precision Then for Max pooling or average pooling operations are performed to compress high-dimensional feature vectors into low-dimensional statistical vectors, thereby physically filtering out the detailed information of the original image and video streams and retaining only the semantic skeleton.

[0115] Differential privacy noise injection: Utilizing the Laplacian or Gaussian mechanism, injecting a distribution that conforms to the target function into the extracted feature vector. random noise (where (For sensitivity), generate the final desensitized semantic features. . It retains the statistical patterns required for the task (such as traffic flow trends) while ensuring that the original physical targets (such as specific faces or license plate numbers) cannot be reconstructed through reverse engineering, thus achieving usability without visibility.

[0116] S4. Store the desensitized semantic features into the distributed storage subsystem, obtain the data addressing hash value, generate a symmetric decryption key, construct an access control tree policy based on the business domain attributes, and encrypt the data addressing hash value and the symmetric decryption key using the ciphertext policy attribute base encryption algorithm to generate a cross-domain ciphertext credential.

[0117] In one embodiment, step S4 includes the following sub-steps:

[0118] S401. Due to the high computational complexity of attribute-based encryption algorithms, directly encrypting large volumes of desensitized semantic features would lead to excessive computational latency at edge nodes. Therefore, this embodiment employs a hybrid mechanism of symmetric and asymmetric encryption.

[0119] First, the edge computing node uses a pseudo-random number generator to generate a high-strength symmetric decryption key. (For example, a 256-bit AES key). Secondly, a symmetric decryption key is used. The desensitized semantic features generated in step S3 Encryption is performed using Advanced Encryption Standard (AES) to generate ciphertext data blocks. Finally, the encrypted data block... Upload to a pre-configured distributed storage subsystem (such as IPFS or Swarm). Upon completion of the upload, the distributed storage subsystem returns a unique content-addressed hash value. (For example, a Multihash string starting with Qm). At this point, only... And without Users are unable to parse the data content.

[0120] S402. To achieve fine-grained access control for "data finding people," the edge computing node dynamically constructs an access control tree strategy based on the business scenarios involved in the cross-domain sharing request. The access control tree policy It is a logical structure tree where leaf nodes correspond to predefined system attributes, and non-leaf nodes correspond to threshold logic gates (such as AND, OR, or (k,n) threshold gates). For example, for data involving "traffic accident scene," the constructed policy logic might be: Policy=("Public Security Department" AND "Criminal Investigation Department") OR("Insurance Company" AND "Senior Claims Adjuster" AND "Case ID Matching"). This means that only request domain nodes whose attribute set satisfies the above logical expression can decrypt subsequent content.

[0121] S403, the encryption function for edge computing nodes to run the Ciphertext Policy Attribute-Based Encryption (CP-ABE) algorithm. Its inputs include: ① the master public key. (Generated and published by the trusted authorization center during the system initialization phase); ② The core confidential information M to be encrypted, in this embodiment M is the symmetric decryption key. With the data addressing hash value The splicing and combination, that is ③ The access control tree policy constructed in step S402 .

[0122] Perform encryption operations:

[0123]

[0124] The final output is the attribute base ciphertext containing the attribute access control logic. .

[0125] S404. Encapsulate the above calculation results into a standardized cross-domain encrypted credential. The data structure of the cross-domain encrypted credential is defined as follows:

[0126]

[0127] in, This serves as a unique identifier for this sharing session; This is a plaintext access control tree policy used by smart contracts to pre-verify the requester's permissions without decryption. As the core encrypted carrier; The edge computing node generates a digital signature for the aforementioned fields using its own private key, which serves to prove the authenticity and non-repudiation of the data source. At this point, the sensitive semantic feature data has been securely transformed into a policy-protected on-chain credential, ready to enter the smart contract interaction phase.

[0128] S5. Publish the shared transaction containing the cross-domain encrypted certificate and the access control tree policy to the smart contract of the consortium blockchain, so that when the smart contract verifies that the attributes of the target request domain node meet the access control tree policy, it issues the cross-domain encrypted certificate and packages the data cross-domain sharing flow record onto the blockchain for evidence storage.

[0129] In one embodiment, step S5 includes the following sub-steps:

[0130] S501, the edge computing node generates the cross-domain encrypted credential CCC (containing session ID, plaintext access control tree policy) and sets it accordingly. Attribute base ciphertext Encapsulated in a standard data sharing and publishing transaction (including digital signatures). By calling the API interface of the IoT data sharing smart contract pre-installed on the consortium blockchain, the edge computing node will process the transaction. The transaction is broadcast to the consortium blockchain network. Upon receiving the transaction, the smart contract first verifies the digital signature of the edge computing node. To confirm the legitimacy of the data source, the cross-domain encrypted credential CCC is then stored in the state variable of the smart contract, and an event log of "data is ready" is broadcast to the entire network.

[0131] S502. When the target requesting domain node listens to the above event log and intends to obtain the cross-domain data, it needs to initiate a data access transaction to the smart contract. In this transaction, the target requesting domain node must submit a digital identity certificate and a set of attribute credentials issued by the consortium blockchain's trusted attribute authorization center. (For example, it includes attributes such as public security department and high-level authority).

[0132] Upon receiving an access transaction, the smart contract automatically triggers the permission verification logic, which retrieves the set of attribute credentials submitted by the target requesting domain node. With the plaintext access control tree policy contained in the credentials Perform Boolean logic matching. If the matching fails (i.e., the requester's attributes do not meet the set threshold logic), the smart contract automatically rejects the request and throws an exception; if the matching succeeds, proceed to step S503.

[0133] S503. The smart contract verifies that the attributes of the target request domain node satisfy the access control tree policy. Subsequently, the smart contract securely encapsulates the cross-domain encrypted credential CCC within a transaction receipt and sends it to the target requesting domain node. The target requesting domain node then performs a reverse decryption and reconstruction process locally, specifically as follows:

[0134] First, using the private key and attribute key it holds, the decryption function (CP-ABE.Decrypt) of the ciphertext policy attribute-based encryption algorithm is called to decrypt the attribute-based ciphertext. Decryption was performed, and the symmetric decryption key was successfully extracted. and content addressing hash :

[0135]

[0136] Subsequently, the target request domain node utilizes the The encrypted data block is precisely addressed and downloaded from a distributed storage subsystem (such as IPFS). Finally, the extracted symmetric decryption key is used. right Perform AES decryption to finally obtain the de-identified semantic features generated in step S3. .

[0137] S504. In order to achieve "availability, controllability, and traceability" of multi-domain data interaction, the smart contract will automatically generate a data flow record after completing the above-mentioned certificate issuance action. This record contains: the identity hash of the data provider (edge ​​node), the identity hash of the data requester (target requesting domain node), the session ID of the accessed data, the timestamp of the access, and the hash result of the smart contract execution.

[0138] Each consensus node in the consortium blockchain network (e.g., using the Practical Byzantine Fault Tolerant Consensus Algorithm PBFT) stores the transfer record. Consensus verification is performed. Once verification is successful, the record is packaged into a new block and appended to the blockchain ledger. Due to the immutable nature of the blockchain, this transaction record will serve as legally valid electronic evidence, providing an absolutely credible basis for subsequent data auditing, security accountability, or data value settlement.

[0139] like Figure 2 As shown, this invention provides an AI IoT multi-domain data sharing system based on multi-source sensing fusion. Based on the aforementioned AI IoT multi-domain data sharing method based on multi-source sensing fusion, the system includes:

[0140] Acquisition module 1 is used to acquire multi-source heterogeneous sensing data in the target physical environment and identify entity targets; wherein, the multi-source heterogeneous sensing data includes image and video streams containing the entity targets and time-series sensing signals reflecting the state of the entity targets;

[0141] Extraction module 2 is used to extract features from the image video stream and the time-series sensing signal respectively, and map them to node features of a dynamic spatiotemporal perception map constructed with the entity target as the node. The dynamic cross-correlation between the node features is calculated through a cross-modal spatiotemporal attention mechanism to align the visual motion state and sensing abnormal state of the entity target and generate a multi-source fusion feature sequence.

[0142] The generation module 3 is used to receive a cross-domain sharing request initiated by the target request domain node, which carries business domain attributes and task type, obtain the current communication bandwidth status and query the trust evaluation level of the target request domain node on the consortium blockchain, input the task type, current communication bandwidth status and trust evaluation level into a preset reinforcement learning decision network, and output the target semantic extraction strategy to perform hierarchical extraction and noise addition on the multi-source fusion feature sequence to generate desensitized semantic features.

[0143] Encryption module 4 is used to store the desensitized semantic features into the distributed storage subsystem, obtain the data addressing hash value and generate a symmetric decryption key, construct an access control tree policy according to the business domain attributes, and encrypt the data addressing hash value and the symmetric decryption key using the ciphertext policy attribute base encryption algorithm to generate cross-domain ciphertext credentials.

[0144] The publishing module 5 is used to publish the shared transaction containing the cross-domain encrypted certificate and the access control tree policy to the smart contract of the consortium blockchain, so that the smart contract issues the cross-domain encrypted certificate when verifying that the attributes of the target request domain node meet the access control tree policy, and packages the flow record of this cross-domain data sharing and puts it on the blockchain for evidence storage.

[0145] Each of the above modules is used to execute the respective steps in the above-mentioned AIoT multi-domain data sharing method based on multi-source sensing fusion. The specific implementation method is as described in the above-mentioned method embodiment, and will not be repeated here.

[0146] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores all the data required for the process of a multi-domain data sharing method for AI IoT based on multi-source sensing fusion. The network interface is used for communication with external terminals via a network connection. The computer program is executed by the processor to implement the multi-domain data sharing method for AI IoT based on multi-source sensing fusion.

[0147] Those skilled in the art will understand that Figure 3 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 equipment on which the present application is applied.

[0148] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described multi-domain data sharing methods for AI IoT based on multi-source sensing fusion.

[0149] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0150] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0151] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-domain data sharing method for AIoT based on multi-source sensing fusion, characterized in that, The method, applied to IoT edge devices deployed with consortium blockchains and distributed storage subsystems, includes: S1. Acquire multi-source heterogeneous sensing data in the target physical environment and identify entity targets; wherein, the multi-source heterogeneous sensing data includes image and video streams containing the entity targets and time-series sensing signals reflecting the state of the entity targets; S2. Extract the features of the image video stream and the time-series sensing signal respectively, and map them to the node features of the dynamic spatiotemporal perception map constructed with the entity target as the node. Calculate the dynamic cross-correlation between the node features through a cross-modal spatiotemporal attention mechanism to align the visual motion state and sensing abnormal state of the entity target and generate a multi-source fusion feature sequence. S3. Receive a cross-domain sharing request initiated by the target request domain node, carrying business domain attributes and task type, obtain the current communication bandwidth status and query the trust evaluation level of the target request domain node on the consortium blockchain, input the task type, current communication bandwidth status and trust evaluation level into a preset reinforcement learning decision network, and output a target semantic extraction strategy to perform hierarchical extraction and noise addition on the multi-source fusion feature sequence to generate desensitized semantic features. S4. Store the desensitized semantic features into the distributed storage subsystem, obtain the data addressing hash value and generate a symmetric decryption key, construct an access control tree policy according to the business domain attributes, and encrypt the data addressing hash value and the symmetric decryption key using the ciphertext policy attribute base encryption algorithm to generate a cross-domain ciphertext credential. S5. Publish the shared transaction containing the cross-domain encrypted certificate and the access control tree policy to the smart contract of the consortium blockchain, so that when the smart contract verifies that the attributes of the target request domain node meet the access control tree policy, it issues the cross-domain encrypted certificate and packages the data cross-domain sharing flow record onto the blockchain for evidence storage.

2. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 1, characterized in that, S1 specifically includes: S101. Perform global clock synchronization on visual sensors and IoT sensing devices deployed in the target physical environment, and set a global time reference. S102. Based on the global time reference, the continuous image and video stream sequence with a first sampling frequency output by the visual sensor and the time-series sensing signal sequence with a second sampling frequency output by the IoT sensing device are acquired in real time and synchronously. The second sampling frequency is higher than the first sampling frequency. S103. Using the low-frequency sampling timestamp of the continuous image and video stream sequence as the reference anchor point, the high-frequency time-series sensing signal sequence is subjected to denoising and windowing aggregation processing to obtain a preliminary time-aligned time-series feature segment. S104. Input the video frames of the continuous image video stream sequence into a preset deep learning target detection model, identify the two-dimensional bounding box coordinates of the entity target, and spatially associate and anchor the entity target with its dedicated IoT sensing device according to the pre-calibrated physical environment topology mapping relationship.

3. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 2, characterized in that, In S4, the spatial association anchoring specifically refers to: Obtain the global three-dimensional coordinate system of the target physical environment, and the calibration physical coordinates of each IoT sensing device in the global three-dimensional coordinate system; Obtain the camera intrinsic and extrinsic parameter matrices of the visual sensor, and based on the camera intrinsic and extrinsic parameter matrices, back-project the two-dimensional bounding box coordinates output by the deep learning object detection model into three-dimensional space estimated coordinates; Calculate the spatial distance between the estimated three-dimensional spatial coordinates of the entity target and the calibrated physical coordinates of each of the IoT sensing devices; The nearest IoT sensor device with a spatial distance less than a preset threshold is identified as the dedicated sensor device attached to the physical target, thus completing the spatial topological association between the physical target and the time-series sensor data.

4. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 2, characterized in that, S2 specifically includes: S201. Construct a neural network structure containing a dual-branch feature extractor, extract visual motion state features from the image video stream to obtain a visual feature vector, and extract sensor abnormal state features from the time-series sensing signal to obtain a time-series sensing feature vector. S202. Construct a dynamic spatiotemporal sensing map, with the entity target as the node set of the map, the spatial topological relationship between the entity targets as the edge set of the map, and map the visual feature vector and the temporal sensing feature vector to the initial multimodal feature representation of the node; S203. Introduce a cross-modal spatiotemporal attention mechanism on the dynamic spatiotemporal perception map, calculate the dynamic cross-correlation between the visual motion state and the sensing abnormal state, and obtain the sensing feature update representation guided by visual features and the visual feature update representation guided by sensing features, respectively. S204. Perform feature-level fusion and dimensionality reduction processing on the depth-aligned sensor feature update representation and the visual feature update representation, and output the multi-source fusion feature sequence.

5. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 4, characterized in that, S203 specifically includes: Introducing the first set of learnable weight matrices The visual feature vector Mapped to query vector matrix The time-series sensing feature vector Mapped to a key vector matrix Sum value vector matrix ; through formula Calculate the dynamic cross-correlation attention weight matrix A, where, The feature dimension of the key vector matrix is... The scaling factor is used; the attention weight matrix A is multiplied by the value vector matrix V to obtain the sensor feature update representation after weighting and aligning by the visual motion state. ; Introducing a second set of learnable weight matrices The time-series sensing feature vector Mapped to a sensor query vector matrix The visual feature vector Mapped to a visual key vector matrix and visual value vector matrix ; through formula Calculate the reverse attention weight matrix A'; combine the reverse attention weight matrix A' with the visual value vector matrix. Multiplying these results in a weighted and aligned representation of the visual features, obtained from the sensor anomaly states. .

6. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 4, characterized in that, S3 specifically includes: S301. Parse the target request domain node identity, business domain attribute and request task type carried in the cross-domain sharing request. Based on the identity, query the historical behavior record of the node in the distributed ledger of the consortium blockchain to obtain the trust assessment level, and monitor the physical communication bandwidth status between the current IoT edge node and the request domain node in real time. S302. Construct a reinforcement learning decision-making agent, defining the environment state space as a vector containing the trust assessment level, the physical communication bandwidth state, and the request task type, and defining the action space as a discretized set of target semantic extraction deep policies; wherein, the environment state space S consists of normalized state vectors. composition: in, The trust assessment level is... This refers to the physical communication bandwidth status. The quantified request task type; The action space A consists of action vectors. The action vector comprises semantic extraction layers. Sum of differential privacy noise budget The semantic extraction level corresponds to the kernel size or step size of the pooling operation on the multi-source fusion feature sequence, and the differential privacy noise budget determines the variance intensity of the injected random noise. S303. Input the current environmental state vector into a pre-set reinforcement learning decision network, and output the optimal action at the current moment based on the objective of maximizing the cumulative reward function, i.e., the target semantic extraction depth strategy; wherein, the cumulative reward function Defined as: in, The target semantic extraction depth strategy selected at the current moment; In strategy The following data availability scoring function; This is a privacy breach risk function, the value of which is related to the trust assessment level. There is a negative correlation, meaning that the lower the trust assessment level, the greater the risk penalty value when choosing a high-precision semantic extraction strategy; This is a transmission delay function, the value of which is related to the physical communication bandwidth status. It shows a negative correlation with strategy The amount of data generated is positively correlated; These are non-negative coefficients used to adjust the weights of data utility, privacy risk, and transmission delay, respectively. S304. According to the target semantic extraction depth strategy, perform hierarchical feature extraction operation on the multi-source fusion feature sequence to compress the feature dimension, and use differential privacy mechanism to inject random noise into the extracted feature vector to generate the desensitized semantic features.

7. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 6, characterized in that, S4 specifically includes: S401. Use a pseudo-random number generator to generate a symmetric decryption key, use the symmetric decryption key to symmetrically encrypt the desensitized semantic features to generate ciphertext data blocks, upload the ciphertext data blocks to the distributed storage subsystem and obtain the content addressing hash value. S402. Based on the business scenarios and preset security rules involved in the cross-domain sharing request, dynamically construct an access control tree strategy that includes attribute leaf nodes and threshold logic gates. S403. The symmetric decryption key and the content addressing hash value are concatenated and combined to form the message to be encrypted. The message to be encrypted is encrypted using the ciphertext policy attribute base encryption algorithm and based on the access control tree policy to generate attribute base ciphertext. S404. Encapsulate the attribute-based ciphertext, access control tree policy, and digital signature of the edge computing node into a cross-domain ciphertext credential; wherein, the data structure of the cross-domain ciphertext credential CCC is defined as follows: in, This serves as a unique identifier for this cross-domain data sharing session; The access control tree policy, in plaintext form, is used by smart contracts to verify the permissions of the requester. The attribute base ciphertext encapsulates the key information; The timestamp used to generate the credentials is used to prevent replay attacks. The edge computing node generates a digital signature for the above fields using its own private key, which is used to prove the authenticity and non-repudiation of the data source.

8. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 7, characterized in that, Specifically, S403 includes: Obtain the system master public key generated during the system initialization phase. ; Construct a message M to be encrypted, wherein the message M is the symmetric decryption key. With content addressing hash value splicing, that is ; The attribute-based ciphertext is generated by performing an encryption operation using the ciphertext policy attribute-based encryption algorithm. ;in, This refers to the access control tree policy.

9. The AIoT multi-domain data sharing method based on multi-source sensing fusion according to claim 7, characterized in that, S5 specifically includes: S501. The edge computing node calls the interface of the smart contract to broadcast the data sharing and publishing transaction containing the cross-domain encrypted certificate to the consortium blockchain network. After verifying the digital signature of the edge computing node, the smart contract stores the cross-domain encrypted certificate and broadcasts the event log. S502. When the target request domain node listens to the event log, it initiates a data access transaction carrying a digital identity certificate and attribute credential set to the smart contract. The smart contract automatically triggers the permission verification logic and logically matches the attribute credential set with the access control tree policy in the cross-domain encrypted credential. S503. If the match is successful, the smart contract sends the cross-domain encrypted credential to the target requesting domain node. The target requesting domain node calls the decryption function of the encrypted policy attribute-based encryption algorithm and uses its own set of attribute credentials. The attribute-based ciphertext in the cross-domain ciphertext certificate Decrypt and extract the symmetric decryption key. and content addressing hash : Using the content addressing hash value Download the corresponding encrypted data block from the distributed storage subsystem and use the symmetric decryption key. The encrypted data block is symmetrically decrypted to restore the de-identified semantic features; S504. The smart contract generates a data transfer record containing information about the cross-domain data sharing and transfer. After the consensus nodes of the consortium blockchain network verify the data transfer record, they package it into a new block and attach it to the blockchain ledger. The data transfer record includes the identity hash of the data provider, the identity hash of the data requester, the session ID of the accessed data, the timestamp of the access, and the hash result of the smart contract execution.

10. An AI IoT multi-domain data sharing system based on multi-source sensing fusion, based on the AI ​​IoT multi-domain data sharing method based on multi-source sensing fusion as described in any one of claims 1 to 9, characterized in that, The system includes: The acquisition module is used to acquire multi-source heterogeneous sensing data in the target physical environment and identify entity targets; wherein, the multi-source heterogeneous sensing data includes image and video streams containing the entity targets and time-series sensing signals reflecting the state of the entity targets; The extraction module is used to extract features from the image video stream and the temporal sensing signal respectively, and map them to node features of a dynamic spatiotemporal perception map constructed with the entity target as the node. The dynamic cross-correlation between the node features is calculated through a cross-modal spatiotemporal attention mechanism to align the visual motion state and sensing abnormal state of the entity target and generate a multi-source fusion feature sequence. The generation module is used to receive cross-domain sharing requests initiated by the target request domain node, which carry business domain attributes and task types, obtain the current communication bandwidth status and query the trust evaluation level of the target request domain node on the consortium blockchain, input the task type, current communication bandwidth status and trust evaluation level into a preset reinforcement learning decision network, and output the target semantic extraction strategy to perform hierarchical extraction and noise addition on the multi-source fusion feature sequence to generate desensitized semantic features. The encryption module is used to store the desensitized semantic features into the distributed storage subsystem, obtain the data addressing hash value and generate a symmetric decryption key, construct an access control tree policy according to the business domain attributes, and encrypt the data addressing hash value and the symmetric decryption key using the ciphertext policy attribute base encryption algorithm to generate cross-domain ciphertext credentials. The publishing module is used to publish the shared transaction containing the cross-domain encrypted credential and the access control tree policy to the smart contract of the consortium blockchain, so that the smart contract issues the cross-domain encrypted credential when verifying that the attributes of the target request domain node meet the access control tree policy, and packages the flow record of this cross-domain data sharing and puts it on the blockchain for evidence storage.

Images