Digital device physical ID full life cycle management system based on deep learning

Abstract

The invention discloses a digital equipment physical ID full life cycle management system based on deep learning, and relates to the technical field of automatic control and intelligent equipment management, and the system comprises the following steps: collecting equipment operation state and physical attribute data through a multi-channel mode, carrying out the preprocessing of the equipment operation state and physical attribute data, extracting multi-level semantic information through a feature fusion method, and carrying out the recognition of the multi-level semantic information; the equipment operation state is recognized through a pre-training model, anomaly detection is carried out, a system combines equipment attribute information and an anomaly result, the corresponding relation between equipment and the digital identity of the equipment is constructed in a hash mode, an equipment digital identity label capable of being dynamically updated and verified is generated, and in the equipment operation process, the system continuously monitors the state change of the equipment, so that the equipment is identified. Whether the identity information needs to be updated or not is automatically judged, recording and verification are completed, and therefore dynamic management and whole-process tracing of the identity of the equipment are achieved. The method supports high-precision identification and flexible expansion, and is suitable for intelligent supervision scenes of various digital industrial devices.

Description

Technical Field

[0001] The present invention relates to the technical field of automated control and intelligent device management, and in particular to a full life cycle management system for physical IDs of digital devices based on deep learning. Background Art

[0002] With the rapid development of intelligent manufacturing and industrial automation technologies, digital devices are increasingly used in various production and control systems. The unique identification, operation status monitoring and full life cycle management of equipment have become the core links in the process of industrial digital transformation. Currently, there are still significant technical bottlenecks in the process of equipment identity management and operation status analysis, which cannot effectively meet the actual needs of intelligent, dynamic and precise management.

[0003] In existing technologies, traditional device identity management and anomaly detection solutions mainly rely on static coding, rule matching, and manually set thresholds. These methods have the following significant drawbacks in complex operating environments: 1. Static identification methods: Most existing systems use fixed IDs, barcodes, MAC addresses, etc. to identify devices. They lack the integration of dynamic factors such as operating status and environmental changes, making it difficult to ensure the timeliness and reliability of device identification.

[0004] 2. Limited accuracy of anomaly detection mechanisms: Traditional methods typically rely on fixed thresholds or single-scale features to judge device status. They are unable to capture the multi-scale semantic feature changes during device operation, which can easily lead to false positives or missed positives.

[0005] 3. Lack of cross-domain adaptability: Existing systems lack domain alignment mechanisms under different working conditions or deployment environments, resulting in poor model generalization and invalid detection results in new scenarios.

[0006] 4. Fragmentation of lifecycle management: Most solutions focus only on a single stage (such as device registration or maintenance), failing to form a closed-loop management architecture covering the entire process, including ID generation, anomaly identification, ID update, and traceability.

[0007] Therefore, how to provide a full life cycle management system for digital device physical ID based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a full life cycle management system for the physical ID of digital devices based on deep learning. The present invention integrates device operating status data and physical property data, constructs a multi-scale feature fusion and abnormal state recognition mechanism, realizes a two-way mapping between the physical properties of the device and the digital identification through joint hash coding, and realizes dynamic update and self-verification processing of the digital ID during the operation of the device. It has the advantages of high recognition accuracy, wide adaptability to working conditions and traceability throughout the entire life cycle.

[0009] The deep learning-based digital device physical ID full lifecycle management system according to an embodiment of the present invention includes the following steps: S1, data acquisition and preprocessing module, uses a multi-channel real-time sampling method to collect equipment operating status data and equipment physical property data, and performs denoising and normalization on the collected equipment operating status data to generate preprocessed data; S2, the feature fusion module, receives preprocessed data, uses dynamic convolution kernel and attention mechanism to extract and fuse local and global semantic features to form fused feature data; S3, pre-training module, builds a deep neural network, constructs reconstruction tasks and comparative learning tasks based on fused feature data, performs multiple rounds of iterative training, and outputs an optimized pre-training model; S4, anomaly detection module, builds domain adaptive mapping, and performs domain adaptive mapping on the equipment operation status data under different operating conditions. Based on the recognition results of the fused feature data by the optimized pre-trained model, it determines whether there is an abnormal state and generates anomaly detection results; S5, the digital ID generation module, combines the device physical attribute data and anomaly detection results, constructs a bidirectional mapping between the device physical attributes and the digital identity through hybrid hash coding, and dynamically generates, updates, and self-verifies the digital device physical ID; S6, the full life cycle management module, triggers the update and recording operation of the physical ID of the digital device based on the abnormal detection results, and realizes maintenance response processing under abnormal conditions.

[0010] Optionally, the S2 specifically includes: S21, a pre-processed data access unit receives pre-processed data, wherein the pre-processed data is a time series of device operation status after denoising and normalization; S22, a multi-scale convolution extraction unit, inputting the preprocessed data into a parallel convolution branch structure, wherein each branch uses a dynamic convolution kernel with a different expansion rate, wherein the expansion rate is adaptively configured according to the rate of change of the device operation state time series, and outputting an initial feature map set, wherein the initial feature map set represents local feature representations at multiple scales; S23, an attention-enhanced feature construction unit receives an initial feature map set and introduces an attention mechanism for each feature map. The attention mechanism includes channel attention and spatial attention. The channel attention generates a channel attention weight vector according to the response intensity distribution of each channel in the feature map. The spatial attention generates a position weight map according to the response features of different positions in the feature map, and outputs a set of attention-weighted feature maps. S24, a fusion tensor construction unit, which performs feature splicing on the attention-weighted feature map set and the initial feature map set, and applies a linear combination function to structurally integrate the splicing results to generate a fused feature tensor, wherein the fused feature tensor contains both local features and global context information; S25, fusion feature normalization unit, applies the normalization function to the fusion feature tensor and outputs the fusion feature data: ; in, To fuse feature data, is the eigenvalue in the fused feature tensor, and are the mean and standard deviation of the eigenvalues ​​in the fused feature tensor, A positive real number used to prevent numerical instability caused by the denominator being zero.

[0011] Optionally, the S23 specifically includes: S231, a feature map input unit receives an initial feature map set, wherein the initial feature map set is composed of a plurality of feature maps with different scale characteristics, each feature map representing a feature expression of the preprocessed data within a specific receptive field range; S232, a channel attention calculation unit performs response modeling on the channel dimension for each feature map in the initial feature map set and calculates the statistical distribution of the response values ​​of each channel of the feature map, generates a channel attention weight vector, and the channel attention weight vector is used to weight the initial feature map in the channel dimension, and outputs a channel enhanced feature map set; S233, a spatial attention calculation unit receives a set of channel-enhanced feature maps, calculates the response strength of each position point in the two-dimensional spatial dimension of each feature map for each feature map, and generates a position weight map, multiplies the position weight map by the set of channel-enhanced feature maps element by element, and outputs a set of spatial-enhanced feature maps; S234, a joint attention fusion unit, which uses a serial fusion method to reconstruct the channel enhancement feature map set and the spatial enhancement feature map set of each initial feature map to generate an attention weighted feature map set.

[0012] Optionally, the S3 specifically includes: S31, a fusion feature input unit, receiving fusion feature data and constructing a deep neural network, wherein the fusion feature data is a feature expression after dynamic convolution kernel extraction and attention mechanism enhancement processing; S32, reconstruction task configuration unit, inputs the fused feature data into the autoencoder structure, the autoencoder structure includes two network modules, encoder and decoder, for the fused feature data Perform compression encoding and restoration decoding to generate reconstructed fusion feature data , based on the tensor difference between the fused feature data and the reconstructed fused feature data, the reconstruction loss function is defined for: ; in, To fuse feature data, To reconstruct the fused feature data, is the reconstruction loss function; S33, contrastive learning task configuration unit, based on fusion feature data Construct multiple enhanced views, generate positive sample data pairs and negative sample data pairs by random perturbation, and input the positive sample data pairs and negative sample data pairs into the projection subnetwork to obtain the corresponding feature representation vector pairs, and calculate the contrast loss function based on the Euclidean distance relationship between the feature representation vector pairs. , used to optimize the representational discriminability between different feature representation vector pairs; S34, joint loss optimization unit, based on the reconstruction loss function Compared with the contrast loss function Constructing joint optimization objective function , defined as follows: ; in, and is a non-negative real number that controls the proportion of the impact of the reconstruction task and the contrastive learning task on the total loss; S35, an optimization model output unit, performs multiple rounds of iterative training on the deep neural network based on the joint optimization objective function, and outputs an optimized pre-trained model.

[0013] Optionally, the S4 specifically includes: S41, a status data access unit receives the device operating status data collected by the data acquisition and preprocessing module, wherein the device operating status data is a time series tensor used to reflect the state change process of the target device under the current operating conditions; S42, a domain-adaptive feature mapping unit, constructing a domain-adaptive mapping, reconstructing the semantic space of the device operating status data by introducing operating condition labels, and generating a domain-aligned feature tensor. The domain-aligned feature tensor structurally unifies the expression distribution between different operating conditions, including load status, ambient temperature, speed setting, operating stage, task type, etc., and is used to characterize the distribution characteristics of the device operating status data under different behavior modes; S43, a recognition result generation unit, inputting the fusion feature data and the domain alignment feature tensor into the optimized pre-training model, performing forward reasoning respectively, to obtain a fusion path recognition result and a domain path recognition result; S44, state offset discrimination unit, calculates the vector difference between the fusion path recognition result and the domain path recognition result, and defines the state offset index function as follows: ; in, represents the fusion path identification result, Indicates the domain path recognition result, is the state shift indicator function; S45, anomaly detection output unit, state offset index function With preset threshold For comparison, when When an abnormality detection result is generated, the abnormality mark data is output to the digital ID generation module, otherwise the normal mark data is generated.

[0014] Optionally, the S5 specifically includes: S51, a physical property vector construction unit receives the device physical property data output by the data acquisition and preprocessing module, wherein the device physical property data includes the device model, manufacturing batch, structural parameters and configuration identification, and generates a device physical property vector through field standardization and numerical coding. ; S52, abnormal state vector construction unit, receives the abnormal detection result, extracts the abnormal mark, state offset index and operation stage, and generates the abnormal state vector , used to characterize the current equipment operation behavior characteristics; S53, joint hash map construction unit, based on the device physical attribute vector and abnormal state vector , forming a joint input vector , and input the joint input vector into the hybrid hash code to generate a joint hash vector , and build a bidirectional mapping structure between the device physical attribute space and the digital identification space; S54, digital ID generation unit, based on the generated joint hash vector Construct a digital device physical ID, defined as follows: ; in, is the joint hash vector, is the mapping function from hash vector space to digital identity space, It is the physical ID of the digital device; S55, ID update and self-verification unit, continuously receives new device physical attribute vectors and abnormal state vectors during device operation, and generates candidate joint hash vectors and the candidate digital device physical ID, and performs a structural consistency comparison with the current digital device physical ID. If the difference exceeds a preset threshold, an ID update is triggered, and the comparison result is used for self-verification records and status maintenance of the digital ID.

[0015] Optionally, the S53 specifically includes: S531, joint vector construction unit, receiving device physical attribute vector and abnormal state vector , generate a joint input vector by concatenating feature dimensions ; S532, joint hash calculation unit, joint input vector Input to the hybrid hash code and calculate each hash fragment separately , combine all hash fragments to generate a joint hash vector , defined as: ; in, For the A hash function, is the number of hash functions; S533, hash vector normalization unit, for the joint hash vector Perform bit width unification, feature dimension compression and numerical normalization operations to obtain a standardized hash representation vector ; S534, bidirectional mapping definition unit, based on standardized hash representation vector Establish a digital device physical ID mapping mechanism and define the forward mapping function: ; And define the reverse mapping function: ; in, It is the physical ID of the digital device. and They are used for bidirectional mapping between hash vectors and digital identifiers respectively.

[0016] The beneficial effects of the present invention are: (1) The present invention constructs an adaptive multi-scale feature extraction and fusion mechanism by fusing multi-channel equipment status data with physical property data. Combined with a deep learning pre-training strategy, it achieves accurate recognition and dynamic response to abnormal equipment status under complex operating conditions, effectively breaking through the limitations of traditional methods in terms of single features and low recognition accuracy, and enhancing the system's environmental adaptability and recognition stability.

[0017] (2) The present invention maps the physical properties of the device and the operating status information into a unified digital identity representation by constructing a joint hash coding structure, and establishes a two-way mapping relationship between the digital identification and the device characteristics, thereby realizing the dynamic generation, update and self-verification of the device digital identity, effectively improving the timeliness and security of device identity management, and avoiding the problem that static identification is easy to forge and difficult to track.

[0018] (3) The digital device physical ID full life cycle management system proposed in this invention covers the entire process of device identity generation, anomaly identification, status monitoring, ID update and identity traceability, builds an end-to-end information closed-loop management architecture, improves the integrity and execution efficiency of the system in device-level intelligent management, reduces manual intervention, and supports efficient and reliable equipment management automation in industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is the overall architecture diagram of the deep learning-based digital device physical ID full life cycle management system proposed in this invention. DETAILED DESCRIPTION

[0020] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0021] refer to Figure 1 The deep learning-based digital device physical ID full lifecycle management system includes the following steps: S1, data acquisition and preprocessing module, uses a multi-channel real-time sampling method to collect equipment operating status data and equipment physical property data, and performs denoising and normalization on the collected equipment operating status data to generate preprocessed data; In this embodiment, a multi-channel real-time sampling method is used to synchronously collect equipment operating status data and physical property data. The operating status data includes dynamic parameters such as current, voltage, temperature, and vibration, while the physical property data includes static information such as equipment model, structural dimensions, and factory serial number. The multi-channel real-time sampling method achieves high-frequency, delay-free acquisition of multi-dimensional status signals by configuring parallel sensing paths and a time synchronization mechanism. The operating status data is then subjected to noise filtering and normalization. The collected raw equipment operating status data is then subjected to one-dimensional median filtering. The median calculation is performed on the data in each time series channel using a sliding window to remove high-frequency interference signals and transient mutation noise generated during the acquisition process. After filtering, the operating status data of each dimension is normalized. Specifically, the average value and fluctuation amplitude of each dimensional feature within the current time window are calculated. The raw data is converted to a zero-centered, amplitude-balanced normalized form to generate pre-processed data in a unified format. This effectively improves the quality and consistency of the raw data, provides a stable and reliable data foundation for subsequent feature extraction and anomaly identification, and enhances the system's sensitivity and response efficiency to equipment status changes under complex working conditions.

[0022] S2, the feature fusion module, receives preprocessed data, uses dynamic convolution kernel and attention mechanism to extract and fuse local and global semantic features to form fused feature data; S3, pre-training module, builds a deep neural network, constructs reconstruction tasks and comparative learning tasks based on fused feature data, performs multiple rounds of iterative training, and outputs an optimized pre-training model; S4, anomaly detection module, builds domain adaptive mapping, and performs domain adaptive mapping on the equipment operation status data under different operating conditions. Based on the recognition results of the fused feature data by the optimized pre-trained model, it determines whether there is an abnormal state and generates anomaly detection results; S5, the digital ID generation module, combines the device physical attribute data and anomaly detection results, constructs a bidirectional mapping between the device physical attributes and the digital identity through hybrid hash coding, and dynamically generates, updates, and self-verifies the digital device physical ID; S6, the full life cycle management module, triggers the update and recording operation of the physical ID of the digital device based on the abnormal detection results, and realizes maintenance response processing under abnormal conditions.

[0023] In this embodiment, when the anomaly detection module identifies an operational anomaly in a device, the system automatically triggers an update and recording operation for the physical identity of the digital device. The system collects the physical attribute data of the current device and the anomaly detection results, and generates a corresponding device physical attribute vector and anomaly state vector. A joint input vector is formed by feature concatenation. The system inputs this joint input vector into the hash encoding path to generate a new joint hash vector, and generates a candidate digital device physical identity through a mapping function. The system compares the structural consistency of the registered identity of the current device with the candidate identity. If the difference value exceeds a set threshold, an identity replacement operation is performed, and the candidate identity is used as the new identity of the device. The system records the updated old identity, candidate identity, difference value, anomaly type, and occurrence time in the device lifecycle database, realizing identity update records and historical status tracking, ensuring that the device identity and its operational status remain consistent, and improving the system's traceability and intelligent maintenance capabilities.

[0024] In this embodiment, S2 specifically includes: S21, a pre-processed data access unit receives pre-processed data, wherein the pre-processed data is a time series of device operation status after denoising and normalization; S22, a multi-scale convolution extraction unit, inputting the preprocessed data into a parallel convolution branch structure, wherein each branch uses a dynamic convolution kernel with a different expansion rate, wherein the expansion rate is adaptively configured according to the rate of change of the device operation state time series, and outputting an initial feature map set, wherein the initial feature map set represents local feature representations at multiple scales; Specifically, the adaptive configuration first calculates the local rate of change of the device operating status time series within a sliding window. This rate is measured by the magnitude of the change in value between consecutive time points. For sequences with faster rates of change, a smaller dilation rate is set to capture denser local details. For sequences with slower or more stable rates of change, a larger dilation rate is set to expand the receptive field and capture global trends. The system automatically adjusts the dilation rate of the corresponding convolution branch based on the rate characteristics of the data from different channels, thereby achieving adaptive extraction and semantic expression of multi-scale features. This configuration ensures that the feature extraction process maintains sensitivity to key patterns and information coverage under different dynamic conditions.

[0025] S23, an attention-enhanced feature construction unit receives an initial feature map set and introduces an attention mechanism for each feature map. The attention mechanism includes channel attention and spatial attention. The channel attention generates a channel attention weight vector according to the response intensity distribution of each channel in the feature map. The spatial attention generates a position weight map according to the response features of different positions in the feature map, and outputs a set of attention-weighted feature maps. S24, a fusion tensor construction unit, which performs feature splicing on the attention-weighted feature map set and the initial feature map set, and applies a linear combination function to structurally integrate the splicing results to generate a fused feature tensor, wherein the fused feature tensor contains both local features and global context information; The linear combination function is used to structurally integrate the concatenated results. Specifically, the system feeds the concatenated results into a linear mapping structure with trainable weights. Feature channels from different sources are weighted and superimposed according to their weights. This process, combined with channel normalization and activation processing, achieves feature dimension reconstruction and expression compression. This structural integration process not only ensures a uniform scale and distribution of the fused feature tensor but also effectively preserves the complementary nature of local and global information.

[0026] S25, fusion feature normalization unit, applies the normalization function to the fusion feature tensor and outputs the fusion feature data: ; in, To fuse feature data, is the eigenvalue in the fused feature tensor, and are the mean and standard deviation of the eigenvalues ​​in the fused feature tensor, A positive real number used to prevent numerical instability caused by the denominator being zero.

[0027] This formula converts the original feature distribution into a normal distribution with zero mean and unit variance by subtracting the mean of the feature tensor and dividing it by the standard deviation (adding a very small positive number to prevent division by zero errors). This helps alleviate the gradient instability problem caused by numerical deviations between features and improves the model convergence speed and feature expression consistency.

[0028] This implementation method inputs the denoised and normalized device operating status time series into a multi-branch dynamic convolution structure, uses convolution kernels with different expansion rates to extract multi-scale local features, and combines channel attention and spatial attention mechanisms to enhance feature map responses, guiding the network to focus on key channels and location areas. The attention-weighted features are then concatenated with the original feature map, and a fused feature tensor is constructed through linear combination, fusing local and global semantic information. Finally, the fused feature tensor is standardized to output a deep feature representation with a uniform scale. This approach effectively enhances the model's sensitivity and expressiveness to device state changes, improves the accuracy and robustness of feature extraction, and provides a stable and discriminative input foundation for subsequent anomaly detection and ID generation.

[0029] In this embodiment, the S23 specifically includes: S231, a feature map input unit receives an initial feature map set, wherein the initial feature map set is composed of a plurality of feature maps with different scale characteristics, each feature map representing a feature expression of the preprocessed data within a specific receptive field range; S232, a channel attention calculation unit performs response modeling on the channel dimension for each feature map in the initial feature map set and calculates the statistical distribution of the response values ​​of each channel of the feature map, generates a channel attention weight vector, and the channel attention weight vector is used to weight the initial feature map in the channel dimension, and outputs a channel enhanced feature map set; The specific modeling process of response modeling in the channel dimension includes: first, performing a global pooling operation on the feature map in the spatial dimension, aggregating the spatial features of each channel into a channel response value, thereby obtaining a vector containing only channel-level responses, and inputting the vector into a channel weight generation network composed of two fully connected layers, outputting the weight coefficient corresponding to each channel, and performing a channel-by-channel weighted operation on these weight coefficients and the original feature map channels to enhance the expressive power of key channels and suppress the interference information of irrelevant or redundant channels. Through the above response modeling process, the system can improve the contribution of channels with distinguishing capabilities in the feature map to subsequent tasks while keeping the original structure unchanged, thereby optimizing the overall feature expression effect.

[0030] S233, a spatial attention calculation unit receives a set of channel-enhanced feature maps, calculates the response strength of each position point in the two-dimensional spatial dimension of each feature map for each feature map, and generates a position weight map, multiplies the position weight map by the set of channel-enhanced feature maps element by element, and outputs a set of spatial-enhanced feature maps; The calculation of the response intensity of each position point in the two-dimensional spatial dimension of each feature map is specifically as follows: each channel enhancement feature map is compressed in the channel dimension, and spatial statistical features are extracted by using pixel-by-pixel channel average pooling and maximum pooling, thereby obtaining two single-channel images representing spatial response features, which respectively reflect the performance of each spatial position in terms of average response and maximum activation. The two images are stacked in the channel dimension to construct a two-dimensional spatial attention feature map, and the map is input into a spatial attention mapping network composed of convolutional layers. The network is used to learn the correlation features between spatial positions and output a position weight map with the same spatial size as the input feature map.

[0031] Specifically, the position weight map is a map where the value of each pixel position represents the importance of the current feature map at that spatial point. The system multiplies this position weight map with the original channel enhanced feature map set element by element according to position, that is, the weight of each spatial position point is adjusted, thereby suppressing the spatial information expression of irrelevant areas and highlighting the key areas, and generating a spatial enhanced feature map set.

[0032] S234, a joint attention fusion unit, which uses a serial fusion method to reconstruct the channel enhancement feature map set and the spatial enhancement feature map set of each initial feature map to generate an attention weighted feature map set.

[0033] This embodiment performs channel attention calculation and spatial attention calculation in sequence to jointly enhance the local channel response and two-dimensional spatial response intensity modeling results in the initial feature map set, and adopts a serial fusion strategy to achieve feature map reconstruction optimization, effectively improving the model's ability to focus on significant areas and key features in the preprocessed data. Specifically, channel attention is used to highlight the importance of different channels in feature expression, and spatial attention is used to locate the activation level of key areas. Finally, a fused attention feature map set is generated through an attention weighting mechanism. This method enhances the model's ability to express multi-scale detail changes, significantly improves the accuracy of device status feature extraction and deep semantic recognition capabilities, and provides a more stable and reliable feature foundation for subsequent anomaly detection and digital ID generation.

[0034] In this embodiment, S3 specifically includes: S31, a fusion feature input unit, receiving fusion feature data and constructing a deep neural network, wherein the fusion feature data is a feature expression after dynamic convolution kernel extraction and attention mechanism enhancement processing; S32, reconstruction task configuration unit, inputs the fused feature data into the autoencoder structure, the autoencoder structure includes two network modules, encoder and decoder, for the fused feature data Perform compression encoding and restoration decoding to generate reconstructed fusion feature data , based on the tensor difference between the fused feature data and the reconstructed fused feature data, the reconstruction loss function is defined for: ; in, To fuse feature data, To reconstruct the fused feature data, is the reconstruction loss function; This formula measures the reconstruction error of the autoencoder during feature restoration by calculating the squared difference between the original fused feature tensor and its reconstruction. Its principle is to minimize the Euclidean distance between the input and output tensors in feature space, thereby enabling the encoder to effectively compress key features and the decoder to accurately restore the original structure. This reconstruction loss function can guide the network to focus on preserving semantic information and improving feature fidelity, thereby improving the model's ability to express device state features completely and compress and reconstruct them.

[0035] The compression encoding and restoration decoding described are specifically as follows: the autoencoder comprises two substructures: an encoder network and a decoder network. The encoder is used to extract key feature information from the high-dimensional fused feature tensor and compress it into a low-dimensional latent space representation. The decoder is used to restore this low-dimensional representation to a reconstructed output with the same dimensionality as the original fused feature data. The encoder is composed of multiple linear mapping layers and activation functions. Its input is the fused feature data, and its output is a compressed encoded representation. This representation retains the main semantic information of the input data but has a significantly reduced dimensionality. The decoder structure is mirror-symmetric to the encoder. It restores the encoded representation to a reconstructed feature tensor of the original dimensionality through layer-by-layer linear mapping and nonlinear transformation. The goal of this process is to make the reconstructed output as close as possible to the original fused feature data. By comparing the differences between the two, a reconstruction loss is constructed, which drives the model to learn to accurately represent key semantic features and compress and remove redundant information during training. The compression encoding process emphasizes feature extraction, while the decoding process focuses on structural restoration. Together, they constitute an unsupervised learning path for fused feature expression.

[0036] S33, contrastive learning task configuration unit, based on fusion feature data Construct multiple enhanced views, generate positive sample data pairs and negative sample data pairs by random perturbation, and input the positive sample data pairs and negative sample data pairs into the projection subnetwork to obtain the corresponding feature representation vector pairs, and calculate the contrast loss function based on the Euclidean distance relationship between the feature representation vector pairs. , used to optimize the representational discriminability between different feature representation vector pairs; The enhanced view construction process is completed by applying multiple perturbation strategies to the original fused feature data, mainly including channel shielding, spatial occlusion, feature permutation, random noise injection and data cropping. Each perturbation strategy selectively changes the local structure or value distribution without destroying the overall semantics of the fused features, thereby constructing feature views with different expressions but consistent semantics.

[0037] S34, joint loss optimization unit, based on the reconstruction loss function Compared with the contrast loss function Constructing joint optimization objective function , defined as follows: ; in, and is a non-negative real number that controls the proportion of the impact of the reconstruction task and the contrastive learning task on the total loss; This formula constructs a unified joint optimization objective function by combining the reconstruction loss function and the contrast loss function in a weighted manner, thereby simultaneously optimizing feature reconstruction and semantic differentiation capabilities during pre-training. Here, represents the reconstruction error and the contrast error between positive and negative data features, respectively. By adjusting the weight coefficients and , the model can flexibly control the balance between preserving the input feature structure and enhancing the discriminability of the representation, helping to improve the model's overall generalization ability and convergence efficiency in abnormal state recognition tasks.

[0038] S35, an optimization model output unit, performs multiple rounds of iterative training on the deep neural network based on the joint optimization objective function, and outputs an optimized pre-trained model.

[0039] Specifically, the multiple rounds of iterative training are as follows: each round of training includes two stages: forward propagation and back propagation. In the forward propagation stage, the fused feature tensor is sequentially input into the encoder, reconstruction decoder and projection sub-network to generate reconstructed features and feature representation vectors. The system calculates the reconstruction error between the fused features and the reconstructed features, as well as the distance difference between the positive and negative sample feature vectors, to generate a joint loss value. In the back propagation stage, the gradient of the trainable parameters in the network is calculated and updated based on the joint loss, and the update range includes the encoder, decoder and projection network. The entire training process controls the number of iterations through the set number of training rounds. Each round uses batch fused feature data for training, and the loss value is recorded after each round. The training continues until the preset maximum number of rounds or the loss convergence threshold condition is met, and finally the parameter state of the deep neural network is output as the optimized model result.

[0040] This implementation combines the reconstruction and contrastive learning tasks by inputting the fused feature tensor into an autoencoder structure and a contrastive learning path, achieving multi-perspective modeling of the device's operating status. The autoencoder structure learns the inherently compressed representation of the fused features and reconstructs the original tensor, thereby capturing the data's structural integrity. Simultaneously, by constructing an enhanced view of the fused features and inputting it into a projection network for contrastive learning, the model is able to identify discriminative feature expressions in state changes. Ultimately, with reconstruction error and contrastive loss as the joint optimization targets, the resulting pre-trained model possesses both sensitivity for anomaly recognition and generalization capabilities for feature differentiation. It maintains efficient and stable anomaly detection performance under varying operating conditions, significantly improving the accuracy and robustness of digital device state perception.

[0041] In this embodiment, the S4 specifically includes: S41, a status data access unit receives the device operating status data collected by the data acquisition and preprocessing module, wherein the device operating status data is a time series tensor used to reflect the state change process of the target device under the current operating conditions; S42, a domain-adaptive feature mapping unit, constructing a domain-adaptive mapping, reconstructing the semantic space of the device operating status data by introducing operating condition labels, and generating a domain-aligned feature tensor. The domain-aligned feature tensor structurally unifies the expression distribution between different operating conditions, including load status, ambient temperature, speed setting, operating stage, task type, etc., and is used to characterize the distribution characteristics of the device operating status data under different behavior modes; Specifically, the semantic space reconstruction is as follows: the system first introduces the operating condition label corresponding to the equipment operating status data, and the operating condition label includes parameters such as load status, ambient temperature, speed setting, and task type. The equipment operating status data and the operating condition label are input together into the semantic mapping network. The semantic mapping network is composed of multiple layers of nonlinear transformation units, and its input end is the splicing vector of the original state data and the condition label. The network learns the distribution differences of state features under different operating conditions through training, and constructs a condition-dependent feature expression mechanism within the network, so that the data distribution of the same equipment under different operating conditions tends to be consistent in the semantic space. The result output by the system is a domain-aligned feature tensor, which retains the main dynamic characteristics of the original operating state, and at the same time reconstructs and adjusts the distribution differences caused by the operating conditions in the feature space.

[0042] S43, a recognition result generation unit, inputting the fusion feature data and the domain alignment feature tensor into the optimized pre-training model, performing forward reasoning respectively, to obtain a fusion path recognition result and a domain path recognition result; The specific process of performing forward reasoning is as follows: first, the fused feature tensor is used as the first input path, feature conversion is performed through the encoder and projection subnetwork of the neural network, and the fused path identification result is output; the domain-aligned feature tensor is used as the second input path, and the same neural network structure and parameters are used to complete feature mapping and expression conversion in sequence to generate a domain path identification result. The two reasoning processes are independently executed under the premise of sharing network parameters to ensure that the input features complete semantic mapping in the same representation space. The obtained output results respectively represent the deep representation vectors of the fused features and the domain-aligned features in the current state.

[0043] S44, state offset discrimination unit, calculates the vector difference between the fusion path recognition result and the domain path recognition result, and defines the state offset index function as follows: ; in, represents the fusion path identification result, Indicates the domain path recognition result, is the state shift indicator function; This formula calculates the Euclidean distance between the fusion path identification results and the domain path identification results, quantifying the degree of deviation between the two identification paths in the feature space, thereby constructing a state deviation indicator function. Based on vector differences, this function reflects the consistency of the current device state response under different semantic mappings. Significant deviations indicate abnormal fluctuations in the device state, helping to improve the sensitivity and reliability of anomaly detection and achieve more refined state monitoring and anomaly identification.

[0044] S45, anomaly detection output unit, state offset index function With preset threshold For comparison, when When an abnormality detection result is generated, the abnormality mark data is output to the digital ID generation module, otherwise the normal mark data is generated.

[0045] This implementation constructs a domain-adaptive mapping, semantically reconstructing device operating status data in conjunction with current operating condition labels. This generates a uniformly structured domain-aligned feature tensor, which is then fed into an optimized pre-trained model along with the fused feature tensor. After obtaining two recognition results, the vector difference is calculated to form a state offset indicator, thereby determining whether the device is in an abnormal state. This method enables feature alignment and accurate anomaly identification under different operating conditions, improving the system's accuracy and robustness in identifying abnormal states in complex operating conditions. It provides a reliable basis for the dynamic updating of digital device physical IDs and abnormal responses, significantly enhancing the intelligence and adaptability of the device's full lifecycle management system.

[0046] In this embodiment, S5 specifically includes: S51, a physical property vector construction unit receives the device physical property data output by the data acquisition and preprocessing module, wherein the device physical property data includes the device model, manufacturing batch, structural parameters and configuration identification, and generates a device physical property vector through field standardization and numerical coding. ; S52, abnormal state vector construction unit, receives the abnormal detection result, extracts the abnormal mark, state offset index and operation stage, and generates the abnormal state vector , used to characterize the current equipment operation behavior characteristics; S53, joint hash map construction unit, based on the device physical attribute vector and abnormal state vector , forming a joint input vector , and input the joint input vector into the hybrid hash code to generate a joint hash vector , and build a bidirectional mapping structure between the device physical attribute space and the digital identification space; S54, digital ID generation unit, based on the generated joint hash vector Construct a digital device physical ID, defined as follows: ; in, is the joint hash vector, is the mapping function from hash vector space to digital identity space, It is the physical ID of the digital device; S55, ID update and self-verification unit, continuously receives new device physical attribute vectors and abnormal state vectors during device operation, and generates candidate joint hash vectors and the candidate digital device physical ID, and performs a structural consistency comparison with the current digital device physical ID. If the difference exceeds a preset threshold, an ID update is triggered, and the comparison result is used for self-verification records and status maintenance of the digital ID.

[0047] This implementation continuously collects new physical attribute and abnormal status data during device operation, generates a corresponding joint hash vector and candidate digital device ID, and compares the structural consistency with the currently registered ID. If the difference metric exceeds a set threshold, an ID update is automatically triggered and the comparison results are recorded to complete the device identity self-verification process. This mechanism enables dynamic control and traceability of device identity as it changes with operating status, improving the timeliness, accuracy, and tamper-proofing of digital IDs, and providing a continuous and trusted identity assurance mechanism for device lifecycle management.

[0048] This implementation method achieves unified modeling of device attributes and status information by constructing a device physical attribute vector and an abnormal state vector, and using hash coding to merge the two into a joint hash vector. It further generates a unique digital device physical ID through a mapping function, and introduces a dynamic update and self-verification mechanism to continuously perceive its attribute changes and state fluctuations during device operation. If a significant deviation in the hash feature structure is detected, the ID update is automatically triggered and the verification log is recorded, thereby achieving accurate characterization, real-time tracking and full life cycle trusted management of the device identity, significantly improving the security, real-time nature and system autonomy of identity management.

[0049] In this embodiment, the S53 specifically includes: S531, joint vector construction unit, receiving device physical attribute vector and abnormal state vector , generate a joint input vector by concatenating feature dimensions ; S532, joint hash calculation unit, joint input vector Input to the hybrid hash code and calculate each hash fragment separately , combine all hash fragments to generate a joint hash vector , defined as: ; in, For the A hash function, is the number of hash functions; S533, hash vector normalization unit, for the joint hash vector Perform bit width unification, feature dimension compression and numerical normalization operations to obtain a standardized hash representation vector ; S534, bidirectional mapping definition unit, based on standardized hash representation vector Establish a digital device physical ID mapping mechanism and define the forward mapping function: ; And define the reverse mapping function: ; in, It is the physical ID of the digital device. and They are used for bidirectional mapping between hash vectors and digital identifiers respectively.

[0050] This formula establishes a forward mapping function between hash vectors and digital IDs, compressing and mapping standardized hash features into a device's unique digital identifier. This allows the conversion of multi-source device information into a unique identity. It also defines a reverse mapping function that allows the system to restore the corresponding hash features from the generated digital ID, achieving reversible binding between the device's identity and features. This mapping mechanism not only ensures the uniqueness and consistency of the digital ID, but also provides traceability and verifiability, supporting subsequent identity resolution and abnormal status tracing.

[0051] This implementation receives a device's physical attribute vector and an abnormal state vector, constructs a joint input vector, and feeds it into multiple hash function paths for encoding. After generating the joint hash vector, it performs standardization processing to obtain a hash representation vector in a unified format. Furthermore, by constructing forward and reverse mapping functions, a bidirectional mapping mechanism between device characteristics and digital identification is implemented. This structure fuses static attributes and dynamic state information and maps them into a unique digital identity, improving the distinguishability of identity representation while also supporting reverse parsing of key device features from the digital ID. This enhances the interpretability, uniqueness, and traceability of device identity management, significantly improving the security and intelligence of device identification.

[0052] Example 1: In order to verify the feasibility of the present invention in practice, the present invention is applied to In order to verify the effectiveness of the deep learning-based full-lifecycle management system for physical IDs of digital devices proposed in the present invention, the present invention was applied to the power equipment automation management scenario of a large-scale intelligent manufacturing enterprise. The enterprise is located in Suzhou Industrial Park, Jiangsu Province, and is mainly engaged in the assembly and operation and maintenance of precision processing equipment. There are nearly 500 sets of key equipment such as CNC machining centers, servo motors, industrial fans and pressure pumps in the factory. These devices have been in multi-working conditions and high-intensity operation for a long time, and the equipment operation status is abnormal frequently. The traditional manual inspection and equipment coding management methods are difficult to meet the systematic requirements of accurate equipment identity identification, real-time monitoring of operation status and equipment abnormality warning. In particular, there are prominent problems such as the inability to trace identity information, untimely status updates and inaccurate abnormality positioning during the equipment life cycle.

[0053] In this scenario, the system deploys data acquisition terminals to perform multi-channel sampling of equipment status data, such as current, voltage, vibration frequency, and temperature rise parameters during operation. This data is then combined with static physical information such as the equipment model, structural dimensions, and manufacturing batch to construct a device physical property vector. All equipment operating status data is preprocessed by edge computing nodes to extract valid time series features and normalize them. This preprocessed data is then input into a feature fusion module, which uses multi-scale dynamic convolution kernels combined with channel and spatial attention mechanisms to extract local and global semantic features. This ultimately forms a fused feature tensor for subsequent abnormal state recognition and digital identity generation.

[0054] During operation, the system uses reconstruction and contrast learning tasks to pre-train the deep neural network model, and automatically judges the abnormal state of the device under different operating conditions through intra-domain feature mapping and recognition result difference evaluation. If the system identifies that the device is in an abnormal operating state, it will combine the abnormal state vector with the static properties of the device, generate a joint hash vector through hash splicing, and construct a bidirectional mapping function to generate a unique digital device physical ID, which is recorded in the device's main data center. Whenever the device is in operation, the system will regularly obtain the latest status characteristics and attribute information, automatically calculate the candidate hash vector and digital ID, and compare the structure consistency with the current registered ID. If the difference exceeds the set threshold, the ID update is triggered, and the update record is written to the device identity log, realizing dynamic maintenance of the device ID and full-process self-verification management.

[0055] To verify the actual application effect of the system, the company selected 60 devices in the same workshop for comparative testing. Among them, 30 devices used traditional ID management and manual status interpretation methods, and the other 30 devices were connected to the full life cycle management system of the present invention. The test period was 2 months. The following is the comparative data of equipment status management: Table 1: Comparison of the effects of traditional management methods and the system management method of the present invention The comparison results show that the system of the present invention effectively reduces the error rate of device identity recognition and improves the accuracy and response speed of anomaly detection. Traditional methods suffer from issues such as damaged device tags and manual data entry errors, which can lead to confusion in some device ID information and prevent the device status from accurately corresponding to the actual entity. However, the present invention combines a hashing mechanism with a self-verification process to dynamically bind device status perception to digital identification, significantly improving the reliability of device identity.

[0056] In one specific test case, a servo motor under heavy load experienced slight abnormal vibration on its 18th day of operation. Because the traditional system failed to recognize this change and failed to promptly record the anomaly, it ultimately caused rotor wear. However, when another device with identical specifications, operating in the same mode, experienced vibration changes, the system identified the abnormality using multi-scale semantic feature extraction and a pre-trained model. Combined with a hash update mechanism, it triggered an ID update and maintenance personnel ticket notification, enabling proactive maintenance intervention. This successfully avoided equipment downtime, saving at least three hours of production line downtime and approximately 12,000 yuan in direct losses.

[0057] Furthermore, the system records the number of device ID updates and self-verification results daily within its device operation log management. Standard deviation thresholds are set to control the ID update frequency, ensuring identification stability despite operational fluctuations. By the end of the trial period, the system had completed 734 ID self-verification operations, with a zero error rate and no false status alarms.

[0058] In summary, this example fully demonstrates the effectiveness of the present invention in accurately identifying device identities, rapidly detecting abnormal conditions, and dynamically managing devices throughout their lifecycle in highly dynamic industrial environments. The system achieves precise binding and real-time maintenance of device physical characteristics to digital identities, significantly improving operational efficiency and intelligence, and providing a replicable and scalable intelligent solution for device management in complex industrial scenarios.

[0059] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A digital device physical ID full lifecycle management system based on deep learning, characterized by: include: S1, data acquisition and preprocessing module, uses a multi-channel real-time sampling method to collect equipment operating status data and equipment physical property data, and performs denoising and normalization on the collected equipment operating status data to generate preprocessed data; S2, the feature fusion module, receives preprocessed data, uses dynamic convolution kernel and attention mechanism to extract and fuse local and global semantic features to form fused feature data; S3, pre-training module, builds a deep neural network, constructs reconstruction tasks and comparative learning tasks based on fused feature data, performs multiple rounds of iterative training, and outputs an optimized pre-training model; S4, anomaly detection module, builds domain adaptive mapping, and performs domain adaptive mapping on the equipment operation status data under different operating conditions. Based on the recognition results of the fused feature data by the optimized pre-trained model, it determines whether there is an abnormal state and generates anomaly detection results; S5, the digital ID generation module, combines the device physical attribute data and anomaly detection results, constructs a bidirectional mapping between the device physical attributes and the digital identity through hybrid hash coding, and dynamically generates, updates, and self-verifies the digital device physical ID; S6, the full life cycle management module, triggers the update and recording operation of the physical ID of the digital device based on the abnormal detection results, and realizes maintenance response processing under abnormal conditions.

2. The deep learning-based digital device physical ID full lifecycle management system according to claim 1 is characterized in that: The S2 specifically includes: S21, a pre-processed data access unit receives pre-processed data, wherein the pre-processed data is a time series of device operation status after denoising and normalization; S22, a multi-scale convolution extraction unit, inputting the preprocessed data into a parallel convolution branch structure, wherein each branch uses a dynamic convolution kernel with a different expansion rate, wherein the expansion rate is adaptively configured according to the rate of change of the device operation state time series, and outputting an initial feature map set, wherein the initial feature map set represents local feature representations at multiple scales; S23, an attention-enhanced feature construction unit receives an initial feature map set and introduces an attention mechanism for each feature map. The attention mechanism includes channel attention and spatial attention. The channel attention generates a channel attention weight vector according to the response intensity distribution of each channel in the feature map. The spatial attention generates a position weight map according to the response features of different positions in the feature map, and outputs a set of attention-weighted feature maps. S24, a fusion tensor construction unit, which performs feature splicing on the attention-weighted feature map set and the initial feature map set, and applies a linear combination function to structurally integrate the splicing results to generate a fused feature tensor, wherein the fused feature tensor contains both local features and global context information; S25, fusion feature normalization unit, applies the normalization function to the fusion feature tensor and outputs the fusion feature data: ； in, To fuse feature data, is the eigenvalue in the fused feature tensor, and are the mean and standard deviation of the eigenvalues ​​in the fused feature tensor, A positive real number used to prevent numerical instability caused by the denominator being zero.

3. The deep learning-based digital device physical ID full lifecycle management system according to claim 2 is characterized in that: The S23 specifically includes: S231, a feature map input unit receives an initial feature map set, wherein the initial feature map set is composed of a plurality of feature maps with different scale characteristics, each feature map representing a feature expression of the preprocessed data within a specific receptive field range; S232, a channel attention calculation unit performs response modeling on the channel dimension for each feature map in the initial feature map set and calculates the statistical distribution of the response values ​​of each channel of the feature map, generates a channel attention weight vector, and the channel attention weight vector is used to weight the initial feature map in the channel dimension, and outputs a channel enhanced feature map set; S233, a spatial attention calculation unit receives a set of channel-enhanced feature maps, calculates the response strength of each position point in the two-dimensional spatial dimension of each feature map for each feature map, and generates a position weight map, multiplies the position weight map by the set of channel-enhanced feature maps element by element, and outputs a set of spatial-enhanced feature maps; S234, a joint attention fusion unit, which uses a serial fusion method to reconstruct the channel enhancement feature map set and the spatial enhancement feature map set of each initial feature map to generate an attention weighted feature map set.

4. The deep learning-based digital device physical ID full lifecycle management system according to claim 1 is characterized in that: The S3 specifically includes: S31, a fusion feature input unit, receiving fusion feature data and constructing a deep neural network, wherein the fusion feature data is a feature expression after dynamic convolution kernel extraction and attention mechanism enhancement processing; S32, reconstruction task configuration unit, inputs the fused feature data into the autoencoder structure, the autoencoder structure includes two network modules, encoder and decoder, for the fused feature data Perform compression encoding and restoration decoding to generate reconstructed fusion feature data , based on the tensor difference between the fused feature data and the reconstructed fused feature data, the reconstruction loss function is defined for: ； in, To fuse feature data, To reconstruct the fused feature data, is the reconstruction loss function; S33, contrastive learning task configuration unit, based on fusion feature data Construct multiple enhanced views, generate positive sample data pairs and negative sample data pairs by random perturbation, and input the positive sample data pairs and negative sample data pairs into the projection subnetwork to obtain the corresponding feature representation vector pairs, and calculate the contrast loss function based on the Euclidean distance relationship between the feature representation vector pairs. , used to optimize the representational discriminability between different feature representation vector pairs; S34, joint loss optimization unit, based on the reconstruction loss function Compared with the contrast loss function Constructing joint optimization objective function , defined as follows: ； in, and is a non-negative real number that controls the proportion of the impact of the reconstruction task and the contrastive learning task on the total loss; S35, an optimization model output unit, performs multiple rounds of iterative training on the deep neural network based on the joint optimization objective function, and outputs an optimized pre-trained model.

5. The deep learning-based digital device physical ID full lifecycle management system according to claim 1 is characterized in that: The S4 specifically includes: S41, a status data access unit receives the device operating status data collected by the data acquisition and preprocessing module, wherein the device operating status data is a time series tensor used to reflect the state change process of the target device under the current operating conditions; S42, a domain-adaptive feature mapping unit, constructing a domain-adaptive mapping, reconstructing the semantic space of the device operating status data by introducing operating condition labels, and generating a domain-aligned feature tensor. The domain-aligned feature tensor structurally unifies the expression distribution between different operating conditions, including load state, ambient temperature, speed setting, operating stage, and task type, and is used to characterize the distribution characteristics of the device operating status data under different behavior modes; S43, a recognition result generation unit, inputting the fusion feature data and the domain alignment feature tensor into the optimized pre-training model, performing forward reasoning respectively, to obtain a fusion path recognition result and a domain path recognition result; S44, state offset discrimination unit, calculates the vector difference between the fusion path recognition result and the domain path recognition result, and defines the state offset index function as follows: ； in, represents the fusion path identification result, Indicates the domain path recognition result, is the state shift indicator function; S45, anomaly detection output unit, state offset index function With preset threshold For comparison, when When an abnormality detection result is generated, the abnormality mark data is output to the digital ID generation module, otherwise the normal mark data is generated.

6. The deep learning-based digital device physical ID full lifecycle management system according to claim 5 is characterized in that: The S5 specifically includes: S51, a physical property vector construction unit receives the device physical property data output by the data acquisition and preprocessing module, wherein the device physical property data includes the device model, manufacturing batch, structural parameters and configuration identification, and generates a device physical property vector through field standardization and numerical coding. ; S52, abnormal state vector construction unit, receives the abnormal detection result, extracts the abnormal mark, state offset index and operation stage, and generates the abnormal state vector , used to characterize the current equipment operation behavior characteristics; S53, joint hash map construction unit, based on the device physical attribute vector and abnormal state vector , forming a joint input vector , and input the joint input vector into the hybrid hash code to generate a joint hash vector , and build a bidirectional mapping structure between the device physical attribute space and the digital identification space; S54, digital ID generation unit, based on the generated joint hash vector Construct a digital device physical ID, defined as follows: ； in, is the joint hash vector, is the mapping function from hash vector space to digital identity space, It is the physical ID of the digital device; S55, ID update and self-verification unit, continuously receives new device physical attribute vectors and abnormal state vectors during device operation, and generates candidate joint hash vectors and the candidate digital device physical ID, and performs a structural consistency comparison with the current digital device physical ID. If the difference exceeds a preset threshold, an ID update is triggered, and the comparison result is used for self-verification records and status maintenance of the digital ID.

7. The deep learning-based digital device physical ID full lifecycle management system according to claim 6 is characterized in that: The S53 specifically includes: S531, joint vector construction unit, receiving device physical attribute vector and abnormal state vector , generate a joint input vector by concatenating feature dimensions ; S532, joint hash calculation unit, joint input vector Input to the hybrid hash code and calculate each hash fragment separately , combine all hash fragments to generate a joint hash vector , defined as: ； in, For the A hash function, is the number of hash functions; S533, hash vector normalization unit, for the joint hash vector Perform bit width unification, feature dimension compression and numerical normalization operations to obtain a standardized hash representation vector ; S534, bidirectional mapping definition unit, based on standardized hash representation vector Establish a digital device physical ID mapping mechanism and define the forward mapping function: ； And define the reverse mapping function: ； in, It is the physical ID of the digital device. and They are used for bidirectional mapping between hash vectors and digital identifiers respectively.

Images