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

Abstract

The application discloses a digital equipment real object ID full life cycle management system based on deep learning, relates to the technical field of automatic control and intelligent equipment management, and comprises the following steps: collecting equipment operation state and physical property data in a multi-channel mode, pre-processing the same, extracting multi-level semantic information by using a feature fusion method, identifying the equipment operation state by using a pre-training model, and performing abnormality detection; the system combines equipment attribute information and abnormality results, constructs a corresponding relationship between the equipment and the digital identity thereof in a hash mode, generates an equipment digital identity label which can be dynamically updated and verified, continuously monitors state changes of the equipment in the process of operation of the equipment, automatically judges whether the identity information needs to be updated, and completes recording and verification, so that dynamic management and full-process tracing of the equipment identity are realized. The application supports high-precision identification and flexible expansion, and is suitable for intelligent supervision scenes of various digital industrial equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automation control and intelligent device management, and particularly relates to a digital device physical ID full life cycle management system based on deep learning. BACKGROUND

[0002] With the rapid development of intelligent manufacturing and industrial automation technology, digital devices are increasingly widely used in various production and control systems. Unique identification, running state monitoring and full life cycle management of devices have become the core link in the process of industrial digital transformation. In the current device identity management and running state analysis process, there are still significant technical bottlenecks, which cannot effectively meet the actual needs of intelligent, dynamic and accurate management.

[0003] In the prior art, traditional device identity management and abnormality detection schemes mainly rely on static coding, rule matching and manual threshold setting methods. These methods have the following significant defects in complex operating environments:

[0004] 1. Static identity recognition means: Most existing systems use fixed ID, barcode, MAC address and other methods to identify devices, lack of fusion of dynamic factors such as running state and environmental changes, and are difficult to ensure the timeliness and reliability of device identification.

[0005] 2. Limited accuracy of abnormality detection mechanism: Traditional methods usually rely on fixed thresholds or single scale features to judge the state of the device, which cannot capture the multi-scale semantic feature changes in the running of the device, and is prone to false positives or false negatives.

[0006] 3. Lack of cross-domain adaptability: In different working conditions or deployment environments, existing systems lack domain alignment mechanisms, resulting in poor model generalization ability and invalid detection results in new scenarios.

[0007] 4. Life cycle management fragmentation: Most schemes only focus on a certain stage (such as device registration or maintenance), and fail to form a full-process closed-loop management architecture covering ID generation, abnormality identification, ID update and traceability.

[0008] Therefore, how to provide a digital device physical ID full life cycle management system based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0009] One purpose of the present application is to propose a deep learning-based digital equipment physical ID full life cycle management system, which integrates equipment operating state data and physical attribute data, constructs a multi-scale feature fusion and abnormal state recognition mechanism, realizes two-way mapping between equipment physical attributes and digital identifiers through joint hash coding, and realizes dynamic updating and self-verification processing of digital ID during equipment operation, with the advantages of high recognition accuracy, wide adaptability to working conditions and full life cycle traceability.

[0010] The deep learning-based digital equipment physical ID full life cycle management system according to the embodiment of the present application comprises the following steps:

[0011] S1, a data acquisition and preprocessing module, which acquires equipment operating state data and equipment physical attribute data by using a multi-channel real-time sampling method, and denoises and normalizes the acquired equipment operating state data to generate preprocessed data;

[0012] S2, a feature fusion module, which receives the preprocessed data, extracts and fuses local and global semantic features by using a dynamic convolution kernel and an attention mechanism, and forms fusion feature data;

[0013] S3, a pre-training module, which constructs a deep neural network, constructs a reconstruction task and a contrast learning task based on the fusion feature data, performs multi-round iterative training, and outputs an optimized pre-training model;

[0014] S4, an abnormality detection module, which constructs a domain adaptive mapping, performs domain adaptive mapping on the equipment operating state data under different operating conditions, judges whether there is an abnormal state based on the recognition result of the optimized pre-training model on the fusion feature data, and generates an abnormality detection result;

[0015] S5, a digital ID generation module, which constructs two-way mapping between equipment physical attributes and digital identifiers by using hybrid hash coding in combination with equipment physical attribute data and abnormality detection results, and dynamically generates, updates and self-verifies digital equipment physical ID;

[0016] S6, a full life cycle management module, which triggers updating and recording operations of the digital equipment physical ID based on the abnormality detection result, and realizes maintenance response processing under abnormal state.

[0017] Optionally, the S2 specifically comprises:

[0018] S21, a preprocessed data access unit, which receives preprocessed data, the preprocessed data being denoised and normalized equipment operating state time series;

[0019] S22, a multi-scale convolution extraction unit, inputs the preprocessed data into a parallel convolution branch structure, each branch adopts a dynamic convolution kernel with different expansion rates, the expansion rates are adaptively configured according to the change rate of the equipment running state time sequence, and an initial feature map set is output, the initial feature map set represents local feature representations at multiple scales;

[0020] S23, an attention enhanced feature construction unit, receives the initial feature map set, introduces an attention mechanism for each feature map, the attention mechanism includes channel attention and spatial attention, wherein the channel attention generates a channel attention weight vector according to the response intensity distribution of each channel of the feature map, and the spatial attention generates a position weight map according to the response features of different positions in the feature map, and outputs an attention weighted feature map set;

[0021] S24, a fusion tensor construction unit, splices the attention weighted feature map set and the initial feature map set, and integrates the spliced result by applying a linear combination function, to generate a fusion feature tensor, the fusion feature tensor contains local features and global context information at the same time;

[0022] S25, a fusion feature normalization unit, applies a standardization function to the fusion feature tensor, and outputs fusion feature data:

[0023] ;

[0024] wherein, is the fusion feature data, is a feature value in the fusion feature tensor, and are the mean and standard deviation of the feature values in the fusion feature tensor, respectively, is a positive real number, used to prevent numerical instability caused by a zero denominator.

[0025] Optionally, the S23 specifically includes:

[0026] S231, a feature map input unit, receiving the initial feature map set, the initial feature map set is composed of multiple feature maps with different scale characteristics, each feature map represents the feature expression of the preprocessed data in a specific receptive field range;

[0027] S232, a channel attention calculation unit, modeling the response in the channel dimension of each feature map in the initial feature map set and calculating the statistical distribution of the response values of each channel of the feature map, to generate a channel attention weight vector, the channel attention weight vector is used to weight the initial feature map in the channel dimension, and output a channel enhanced feature map set;

[0028] S233, a spatial attention calculation unit, receives the channel enhanced feature map set, calculates the response intensity of each position point in the two-dimensional spatial dimension of each feature map for each feature map, and generates a position weight map, wherein the position weight map is multiplied by the channel enhanced feature map set element by element, and a spatial enhanced feature map set is output.

[0029] S234, a joint attention fusion unit, reconstructs features in a serial fusion manner for the channel enhanced feature map set and the spatial enhanced feature map set of each initial feature map, and generates an attention weighted feature map set.

[0030] Optionally, the S3 specifically comprises:

[0031] S31, a fusion feature input unit, receives fusion feature data and constructs a deep neural network, wherein the fusion feature data is a feature expression processed by a dynamic convolution kernel extraction and an attention mechanism enhancement;

[0032] S32, a reconstruction task configuration unit, inputs the fusion feature data into a self-encoder structure, wherein the self-encoder structure comprises an encoder and a decoder network module, and compresses and encodes and restores and decodes the fusion feature data to generate reconstruction fusion feature data , defines a reconstruction loss function based on the tensor difference between the fusion feature data and the reconstruction fusion feature data, and the reconstruction loss function is as follows:

[0033] ;

[0034] wherein, is the fusion feature data, is the reconstruction fusion feature data, is the reconstruction loss function;

[0035] S33, a contrastive learning task configuration unit, constructs a plurality of enhanced views based on the fusion feature data , generates a positive sample data pair and a negative sample data pair through a random disturbance manner, inputs the positive sample data pair and the negative sample data pair into a projection subnetwork to obtain a corresponding feature representation vector pair, and calculates a contrastive loss function according to the Euclidean distance relationship between the feature representation vector pair, which is used to optimize the representation distinguishability between different feature representation vectors.

[0036] S34, a joint loss optimization unit, constructs a joint optimization objective function based on the reconstruction loss function and the contrastive loss function , and the joint optimization objective function is defined as follows:

[0037] ;

[0038] wherein, with is a non-negative real number, controlling the proportion of the influence of the reconstruction task and the contrastive learning task on the total loss;

[0039] S35, an optimization model output unit, performing multiple rounds of iterative training on the deep neural network based on the joint optimization objective function, and outputting an optimized pre-training model.

[0040] Optionally, the S4 specifically comprises:

[0041] S41, a state data access unit, receiving device operating state data collected by the data collection and preprocessing module, the device operating state data being a time series tensor, used to reflect the state change process of the target device under the current operating condition;

[0042] S42, a domain adaptation feature mapping unit, constructing a domain adaptation mapping, reconstructing the semantic space of the device operating state data by introducing an operating condition label, generating a domain-aligned feature tensor, the domain-aligned feature tensor unifying the expression distribution between different operating conditions in structure, the different operating conditions including load state, environment temperature, speed setting, working condition stage, task type, etc., used to represent the distribution characteristics of the device operating state data under different behavior modes;

[0043] S43, a recognition result generation unit, inputting the fusion feature data and the domain-aligned feature tensor into the optimized pre-training model, respectively performing forward inference, and obtaining a fusion path recognition result and a domain path recognition result;

[0044] S44, a state offset discrimination unit, performing vector difference calculation on the fusion path recognition result and the domain path recognition result, and defining a state offset indicator function as follows:

[0045] ;

[0046] wherein, represents the fusion path recognition result, represents the domain path recognition result, is the state offset indicator function;

[0047] S45, an anomaly detection output unit, comparing the state offset indicator function with a preset threshold When , an anomaly detection result is generated, and abnormal marker data is output to the digital ID generation module, otherwise normal marker data is generated.

[0048] Optionally, the S5 specifically comprises:

[0049] S51, a physical attribute vector construction unit, receiving device physical attribute data output by the data acquisition and preprocessing module, the device physical attribute data including device model, manufacturing batch, structure parameter and configuration identification, generating a device physical attribute vector through field standardization and numerical encoding ;

[0050] S52, an abnormal state vector construction unit, receiving abnormality detection results, extracting abnormality markers, state offset indicators and running stages, generating an abnormal state vector , for representing current device running behavior characteristics;

[0051] S53, a joint hash mapping construction unit, based on the device physical attribute vector and the abnormal state vector , forming a joint input vector , and inputting the joint input vector into a hybrid hash encoding to generate a joint hash vector , and constructing a bidirectional mapping structure between the device physical attribute space and the digital identification space;

[0052] S54, a digital ID generation unit, based on the generated joint hash vector constructing a digital device physical ID, defined as follows:

[0053] ;

[0054] wherein, is the joint hash vector, is a mapping function from the hash vector space to the digital identification space, is the digital device physical ID;

[0055] S55, an ID updating and self-verification unit, continuously receiving new device physical attribute vectors and abnormal state vectors during device running, generating a candidate joint hash vector and a candidate digital device physical ID, and performing structure consistency comparison with the current digital device physical ID, if the difference exceeds a preset threshold, triggering ID updating, and using the comparison results for digital ID self-verification record and state maintenance.

[0056] Optionally, the S53 specifically includes:

[0057] S531, a joint vector construction unit, receiving the device physical attribute vector and the abnormal state vector , generating a joint input vector through feature dimension splicing;

[0058] S532, a joint hash calculation unit, calculating a joint hash value for the joint input vector Input into the mixed hash code, respectively, calculate each hash segment , all hash segments are combined to generate a joint hash vector , defined as:

[0059] ;

[0060] Wherein, is the first hash function, is the number of hash functions; S533, hash vector normalization unit, the joint hash vector

[0061] Perform bit width uniformity, feature dimension compression and numerical normalization operation, obtain the standardized hash representation vector ;

[0062] S534, bidirectional mapping definition unit, based on the standardized hash representation vector Establish a digital device physical ID mapping mechanism, define the forward mapping function:

[0063] ;

[0064] And define the reverse mapping function:

[0065] ;

[0066] Wherein, is the digital device physical ID, and are used for bidirectional mapping between hash vector and digital identification.

[0067] The beneficial effects of the present application are:

[0068] (1) The present application fuses multi-channel device state data and physical attribute data, constructs an adaptive multi-scale feature extraction and fusion mechanism, combines a deep learning pre-training strategy, realizes accurate identification and dynamic response of device abnormal state under complex operating conditions, effectively breaks through the limitations of traditional methods in single feature and low identification precision, and enhances the environmental adaptability and identification stability of the system.

[0069] (2) The present application constructs a joint hash coding structure, maps the device physical attributes and operating state information into a unified digital identity representation, and establishes a bidirectional mapping relationship between the digital identification and the device features, realizes the dynamic generation, update and self-verification of the device digital identity, effectively improves the timeliness and security of the device identity management, and avoids the problems that static identification is easy to be forged and difficult to track.

[0070] ​(3) The digital equipment real object ID full life cycle management system provided by the application covers the whole process of equipment identity generation, abnormality identification, state monitoring, ID updating and identity tracing, constructs an end-to-end information closed loop management architecture, improves the integrity and execution efficiency of the system in the intelligent management of the equipment level, reduces manual intervention, and supports efficient and reliable equipment management automation in industrial scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0071] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0072] Figure 1 The overall architecture diagram of the digital equipment real object ID full life cycle management system based on deep learning provided by the application. DETAILED DESCRIPTION

[0073] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only schematically show the basic structure of the application, and therefore only show the components related to the application.

[0074] REFERENCE Figure 1 The digital equipment real object ID full life cycle management system based on deep learning includes the following steps:

[0075] S1, a data acquisition and preprocessing module, adopts a multi-channel real-time sampling method to acquire equipment running state data and equipment physical property data, and performs denoising and normalization processing on the acquired equipment running state data to generate preprocessed data;

[0076] In this embodiment, a multi-channel real-time sampling method is adopted to synchronously collect equipment operation state data and physical attribute data, wherein the operation state data includes dynamic parameters such as current, voltage, temperature and vibration, and the physical attribute data includes static information such as equipment model, structure size and factory number; the multi-channel real-time sampling method realizes high-frequency and non-delayed collection of multi-dimensional state signals by configuring parallel sensing channels and time synchronization mechanism, and then performs noise filtering and normalization processing on the operation state data. The original equipment operation state data is processed by one-dimensional median filtering, that is, the median value of the data in each time sequence channel is calculated through a sliding window, so as to remove high-frequency interference signals and instantaneous mutation noise generated in the collection process. After filtering, the standardization processing is performed on the operation state data in each dimension, which specifically includes calculating the average value and fluctuation amplitude of each dimensional feature in the current time window, and converting the original data into a standardized form with zero center and balanced amplitude, to generate preprocessed data in a unified format, effectively improving the quality and expression consistency of the original data, providing a stable and reliable data basis for subsequent feature extraction and anomaly recognition, and enhancing the sensitivity and response efficiency of the system to equipment state changes under complex working conditions.

[0077] S2, a feature fusion module, receives the preprocessed data, extracts and fuses local and global semantic features by using a dynamic convolution kernel and an attention mechanism, and forms fusion feature data;

[0078] S3, a pre-training module, constructs a deep neural network, constructs a reconstruction task and a contrast learning task based on the fusion feature data, performs multi-round iterative training, and outputs an optimized pre-training model;

[0079] S4, an anomaly detection module, constructs a domain adaptive mapping, and performs domain adaptive mapping on the equipment operation state data under different operating conditions, judges whether there is an abnormal state based on the recognition result of the fusion feature data by the optimized pre-training model, and generates an anomaly detection result;

[0080] S5, a digital ID generation module, combines the equipment physical attribute data and the anomaly detection result, constructs a bidirectional mapping between the equipment physical attribute and the digital identifier by hybrid hash coding, and dynamically generates, updates and self-verifies the digital equipment physical ID;

[0081] S6, a full life cycle management module, triggers the update and recording operation of the digital equipment physical ID based on the anomaly detection result, and realizes the maintenance response processing under the abnormal state.

[0082] In this embodiment, when the anomaly detection module identifies that the device has a running anomaly, the system will automatically trigger the update and record operation of the digital device physical identity. The system collects the physical attribute data and anomaly detection result of the current device, generates the corresponding device physical attribute vector and anomaly state vector, forms a joint input vector through feature splicing, inputs the joint input vector into the hash coding path, generates a new joint hash vector, and generates a candidate digital device physical identity through a mapping function. The system compares the registration identity of the current device with the candidate identity for structural consistency. If the difference value exceeds the set threshold, the 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, realizes identity update record and historical state tracking, ensures the consistency of the device identity and its running state, and improves the traceability and intelligent maintenance capability of the system.

[0083] In this embodiment, S2 specifically includes:

[0084] S21, a preprocessed data access unit, receives preprocessed data, which is a device running state time series after denoising and normalization processing;

[0085] S22, a multi-scale convolution extraction unit, inputs the preprocessed data into a parallel convolution branch structure. Each branch uses a dynamic convolution kernel with different dilation rates. The dilation rate is adaptively configured according to the change rate of the device running state time series. An initial feature map set is outputted, which represents local feature representations at multiple scales.

[0086] The adaptive configuration specifically includes first calculating the local change rate of the device running state time series within a sliding window. The rate is measured by the value variation amplitude between consecutive time points. For sequences with a faster change rate, a smaller dilation rate is set to obtain more intensive local details. For sequences with a slower change rate or stable sequences, a larger dilation rate is set to expand the receptive field to capture global trends. The system automatically adjusts the dilation rate of the corresponding convolution branch according to the rate characteristics of different channel data, thereby realizing adaptive extraction and semantic expression of multi-scale features. This configuration ensures that the feature extraction process can maintain sensitivity to key patterns and information coverage ability under different dynamic conditions.

[0087] S23, an attention-enhanced feature construction unit, receives the initial feature map set, introduces an attention mechanism to each feature map, and outputs an attention-weighted feature map set. 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 of the feature map. The spatial attention generates a position weight map according to the response features of different positions in the feature map.

[0088] S24, a fusion tensor construction unit, performing feature splicing on the attention weighted feature map set and the initial feature map set, and applying a linear combination function to the splicing result to generate a fusion feature tensor, which contains both local features and global context information;

[0089] The application of the linear combination function to the splicing result for structural integration, the specific integration process is: the system inputs the splicing result into a linear mapping structure with trainable weights, weights the feature channels of different sources according to the weights, and combines channel normalization and activation processing to realize the reconstruction and expression compression of feature dimensions. This structural integration process not only makes the fusion feature tensor have a unified scale and distribution form, but also effectively retains the complementary characteristics between local and global information.

[0090] S25, a fusion feature normalization unit, applying a standardization function to the fusion feature tensor to output fusion feature data:

[0091] ;

[0092] wherein, is the fusion feature data, is the feature value in the fusion feature tensor, and are the mean and standard deviation of the feature value in the fusion feature tensor, is a positive real number, used to prevent numerical instability caused by a zero denominator.

[0093] The 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 by the standard deviation (plus a small positive number to prevent division by zero error), which helps to alleviate the gradient instability problem caused by numerical deviation between features, and improves the model convergence speed and feature expression consistency.

[0094] The embodiment inputs the device running state time series after denoising and normalization into a multi-branch dynamic convolution structure, uses convolution kernels with different dilation rates to realize multi-scale local feature extraction, and combines channel attention and spatial attention mechanisms to enhance feature map responses and guide the network to focus on key channels and position areas. Then, the attention weighted features and the original feature maps are spliced, and a fusion feature tensor is constructed by linear combination to fuse local and global semantic information. Finally, the fusion feature tensor is standardized to output a deep feature representation with a unified scale. This method effectively enhances the sensitivity and expression ability of the model to device state changes, improves the accuracy and robustness of feature extraction, and provides a stable and discriminative input basis for subsequent anomaly detection and ID generation.

[0095] The S23 specifically includes:

[0096] S231, a feature map input unit, receiving an initial feature map set composed of multiple feature maps with different scale characteristics, each feature map representing feature expression of the preprocessed data in a specific receptive field range;

[0097] S232, a channel attention calculation unit, modeling responses in the channel dimension of each feature map in the initial feature map set and calculating statistical distribution of channel response values of the feature map, generating a channel attention weight vector for weighting the initial feature map in the channel dimension, and outputting a channel enhanced feature map set;

[0098] The response modeling in the channel dimension specifically 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, inputting the vector into a channel weight generation network composed of two fully connected layers, outputting weight coefficients corresponding to each channel, and performing a channel-by-channel weighting operation on the weight coefficients and the original feature map channel to enhance the expression ability of key channels and suppress irrelevant or redundant channel interference information. Through the above response modeling process, the system can improve the contribution of channels with distinguishing ability in the feature map to subsequent tasks while maintaining the original structure, thereby optimizing the overall feature expression effect.

[0099] S233, a spatial attention calculation unit, receiving the channel enhanced feature map set, calculating the response intensity of each position point in the two-dimensional spatial dimension of each feature map, and generating a position weight map, multiplying the position weight map and the channel enhanced feature map set element by element, and outputting a spatial enhanced feature map set;

[0100] The calculation of the response intensity of each position point in the two-dimensional spatial dimension of each feature map specifically includes: compressing each channel enhanced feature map in the channel dimension, using pixel-by-pixel channel average pooling and maximum pooling to extract spatial statistical features, thereby obtaining two single-channel images representing spatial response features, reflecting the performance of each spatial position in average response and maximum activation, stacking the two images in the channel dimension to construct a two-dimensional spatial attention feature map, and inputting the spatial attention feature map into a spatial attention mapping network composed of convolution 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.

[0101] The position weight map is specifically that the value of each pixel position represents the importance of the current feature map at the spatial point, and the system element-wise multiplies the position weight map with the original channel enhanced feature map set according to the position, that is, adjusts the weight for each spatial position point, thereby suppressing the spatial information expression of irrelevant areas and highlighting the key area, and generates a spatial enhanced feature map set.

[0102] S234, the joint attention fusion unit adopts a series fusion mode to perform feature reconstruction on the channel enhanced feature map set and the spatial enhanced feature map set of each initial feature map, and generates an attention weighted feature map set.

[0103] The embodiment sequentially performs channel attention calculation and spatial attention calculation, jointly enhances the local channel response in the initial feature map set and the two-dimensional spatial response intensity modeling result, adopts a series fusion strategy to realize reconstruction and optimization of the feature map, and effectively improves the attention ability of the model to the significant area and key features in the preprocessed data. Specifically, the channel attention is used to highlight the importance of different channels in feature expression, the spatial attention is used to position the activation degree of the key area, and finally the fused attention feature map set is generated through the attention weighting mechanism. This mode enhances the expression ability of the model to multi-scale detail changes, significantly improves the accuracy and deep semantic recognition ability of the device state feature extraction, and provides a more stable and reliable feature basis for subsequent anomaly detection and digital ID generation.

[0104] The embodiment, the S3 specifically comprises:

[0105] S31, a fusion feature input unit receives fusion feature data and constructs a deep neural network, the fusion feature data being a feature expression processed by a dynamic convolution kernel extraction and an attention mechanism enhancement;

[0106] S32, a reconstruction task configuration unit inputs the fusion feature data into a self-encoder structure, the self-encoder structure including an encoder and a decoder network module, and performs compression encoding and restoration decoding on the fusion feature data to generate reconstruction fusion feature data , and defines a reconstruction loss function based on the tensor difference between the fusion feature data and the reconstruction fusion feature data as follows:

[0107] ;

[0108] wherein, the fusion feature data is the reconstruction fusion feature data is the reconstruction loss function is;

[0109] The formula measures the reconstruction error of the autoencoder in the process of restoring the feature information by calculating the square difference between the original fusion feature tensor and its reconstruction result, and the principle is to minimize the Euclidean distance between the input tensor and the output tensor in the feature space, so as to promote the encoder to effectively compress the key features and the decoder to accurately restore the original structure. The reconstruction loss function can guide the network to focus on the preservation of semantic information and the improvement of feature fidelity, and improve the expression integrity and compression reconstruction ability of the model to the device state features.

[0110] The compression encoding and restoration decoding are specifically: the autoencoder includes an encoder network and a decoder network two substructures, wherein the encoder is used to extract key feature information from the high-dimensional fusion feature tensor and map it to a low-dimensional latent space representation, and the decoder is used to restore the low-dimensional representation to a reconstruction output consistent with the dimension of the original fusion feature data. The encoder is composed of multiple linear mapping layers and activation functions, and the input is the fusion feature data and the output is the compressed encoding representation. The representation retains the main semantic information of the input data, but the dimension is significantly reduced. The decoder structure is mirror-symmetrical with the encoder. The encoding representation is restored to the reconstruction feature tensor of the original dimension through layer-by-layer linear mapping and nonlinear transformation. The goal of this process is to make the reconstruction output as close as possible to the original fusion feature data. By comparing the difference between the two, a reconstruction loss is constructed to drive the model to learn accurate representation of key semantic features and compression and removal of redundant information during the training process. The compression encoding process emphasizes feature extraction, and the decoding process focuses on structure restoration, which together constitute an unsupervised learning path for fusion feature expression.

[0111] S33, a contrast learning task configuration unit, based on the fusion feature data A plurality of enhanced views are constructed, positive sample data pairs and negative sample data pairs are generated by random disturbance, and the positive sample data pairs and the negative sample data pairs are input into the projection sub-network to obtain corresponding feature representation vector pairs. The contrast loss function is calculated according to the Euclidean distance relationship between the feature representation vector pairs , for optimizing the representation distinguishability between different feature representation vector pairs;

[0112] The construction process of the enhanced view is completed by applying various disturbance strategies to the original fusion feature data, mainly including channel shielding, spatial occlusion, feature replacement, random noise injection and data cropping, etc. Each disturbance strategy selectively changes the local structure or value distribution without destroying the overall semantics of the fusion feature, thereby constructing feature views with expression differences but consistent semantics.

[0113] S34, a joint loss optimization unit, based on the reconstruction loss function and the contrast loss function A joint optimization objective function is constructed , defined as follows:

[0114] ;

[0115] wherein, and are non-negative real numbers, controlling the proportion of the influence of the reconstruction task and the contrastive learning task on the total loss;

[0116] The formula combines the reconstruction loss function and the contrastive loss function in a weighted manner to construct a unified joint optimization objective function, thereby optimizing the feature reconstruction capability and semantic discrimination capability simultaneously in the pre-training process. Wherein, respectively represent the reconstruction error and the contrast error between positive and negative data features, by adjusting the weight coefficients and, the balance between maintaining the input feature structure restoration and enhancing the representation discrimination can be flexibly controlled, which helps to improve the overall generalization capability and convergence efficiency of the model in the abnormal state recognition task.

[0117] S35, an optimized 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-training model.

[0118] The multiple rounds of iterative training are specifically that each round of training includes two stages of forward propagation and backward propagation. In the forward propagation stage, the fusion feature tensor is input into the encoder, the reconstruction decoder and the projection subnetwork in turn to generate the reconstruction feature and the feature representation vector, and the system calculates the reconstruction error between the fusion feature and the reconstruction feature, and the distance difference between the positive and negative sample feature vectors to generate the joint loss value. In the backward propagation stage, the trainable parameters in the network are calculated and updated based on the joint loss, and the update range includes the encoder, the decoder and the projection network. The entire training process controls the number of iterations through the set number of training rounds, each round uses a batch of fusion feature data for training, and records the loss value after each round ends, and the training continues to execute 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.

[0119] The embodiment jointly constructs a reconstruction task and a contrast learning task by inputting the fusion feature tensor into a self-encoder structure and a contrast learning path respectively, realizes multi-perspective modeling of the equipment running state, and the self-encoder structure is used to learn the intrinsic compression representation of the fusion feature and reconstruct the original tensor, so as to depict the integrity of the data on its own structure; meanwhile, the enhanced view of the fusion feature is constructed and input into the projection network for contrast learning, so that the model can identify the feature expression with discriminative characteristics in the state change. Finally, the pre-training model trained with the reconstruction error and the contrast loss as the joint optimization target has the sensitivity of abnormality identification and the generalization ability of feature distinction, and still maintains efficient and stable abnormality detection performance under different running conditions, significantly improving the accuracy and robustness of digital equipment state perception.

[0120] The S4 specifically includes:

[0121] S41, a state data access unit, receiving equipment running state data collected by a data collection and preprocessing module, the equipment running state data being a time series tensor, used to reflect the state change process of the target equipment under the current running condition;

[0122] S42, a domain adaptive feature mapping unit, constructing a domain adaptive mapping, reconstructing the semantic space of the equipment running state data by introducing a running condition label, generating a domain-aligned feature tensor, the domain-aligned feature tensor unifying the expression distribution between different running conditions in structure, the different running conditions including load state, environment temperature, speed setting, working condition stage, task type, etc., used to represent the distribution characteristics of the equipment running state data under different behavior modes;

[0123] The semantic space reconstruction specifically includes that the system first introduces a running condition label corresponding to the equipment running state data, the running condition label including parameters such as load state, environment temperature, speed setting, task type, etc., and inputs the equipment running state data and the running condition label into a semantic mapping network, the semantic mapping network being composed of multiple layers of nonlinear transformation units, the input end of the network being a splicing vector of the original state data and the condition label, the network learning the state feature distribution difference under different running conditions through training, and constructing a condition-dependent feature expression mechanism inside the network, so that the data distribution of the same equipment under different running conditions tends to be consistent in the semantic space, and the output result of the system is a domain-aligned feature tensor, which retains the main dynamic characteristics of the original running state and adjusts the distribution difference caused by the running condition in the feature space.

[0124] S43, a recognition result generation unit, inputting the fusion feature data and the domain-aligned feature tensor into an optimized pre-training model, respectively performing forward inference, and obtaining a fusion path recognition result and a domain path recognition result;

[0125] The execution of the forward inference specific process is that firstly, the fusion feature tensor is taken as a first input path, the feature conversion is performed through the encoder and the projection sub-network of the neural network, the fusion path recognition result is output, the domain alignment feature tensor is taken as a second input path, the same neural network structure and parameters are adopted, the feature mapping and expression conversion are sequentially completed, the domain path recognition result is generated, the two inference processes are independently executed under the premise of network parameter sharing, the semantic mapping of the input features is ensured in the same representation space, and the output results respectively represent the deep representation vectors of the fusion features and the domain alignment features in the current state.

[0126] S44, a state offset discrimination unit performs vector difference calculation on the fusion path recognition result and the domain path recognition result, and defines a state offset index function as follows:

[0127] ;

[0128] wherein, the fusion path recognition result is represented by f, the domain path recognition result is represented by d, and the state offset index function is represented by f-d;

[0129] The formula calculates the Euclidean distance between the fusion path recognition result and the domain path recognition result, quantifies the deviation degree of the two recognition paths in the feature space, and thereby constructs the state offset index function. The function is based on the vector difference and reflects the response consistency of the current device state under different semantic mappings. If the deviation is significant, it indicates that the device state has abnormal fluctuations, which helps to improve the discrimination sensitivity and reliability of the abnormal detection, and realizes more refined state monitoring and abnormal recognition.

[0130] S45, an abnormal detection output unit compares the state offset index function with a preset threshold When f-d , an abnormal detection result is generated, and abnormal mark data is output to the digital ID generation module, otherwise, normal mark data is generated.

[0131] The embodiment constructs a field adaptive mapping, combines device running state data with current running condition labels for semantic reconstruction, generates a structure-unified field alignment feature tensor, and inputs the feature tensor into an optimized pre-training model together with a fusion feature tensor to obtain two recognition results, calculate the vector difference value of the two recognition results to form a state offset index, and determine whether the device is in an abnormal state. This method can realize feature alignment and accurate abnormality identification under different running conditions, improve the discrimination accuracy and robustness of the system for abnormal states in complex working conditions, provide a reliable basis for dynamic updating and abnormal response of digital device physical IDs, and significantly enhance the intelligent level and adaptability of the device full life cycle management system.

[0132] The S5 specifically includes:

[0133] S51, a physical attribute vector construction unit, receiving device physical attribute data output by the data acquisition and preprocessing module, the device physical attribute data including device model, manufacturing batch, structure parameter and configuration identification, generating a device physical attribute vector through field standardization and numerical encoding ;

[0134] S52, an abnormal state vector construction unit, receiving an abnormality detection result, extracting an abnormality mark, a state offset index and a running phase, generating an abnormal state vector for representing current device running behavior characteristics;

[0135] S53, a joint hash mapping construction unit, based on the device physical attribute vector and the abnormal state vector , forming a joint input vector , and inputting the joint input vector into a hybrid hash encoding to generate a joint hash vector , and constructing a bidirectional mapping structure between the device physical attribute space and the digital identification space;

[0136] S54, a digital ID generation unit, based on the generated joint hash vector constructing a digital device physical ID, defined as follows:

[0137] ;

[0138] wherein, is the joint hash vector, is a mapping function from the hash vector space to the digital identification space, is the digital device physical ID;

[0139] S55, an ID updating and self-verification unit, continuously receiving new device physical attribute vectors and abnormal state vectors during device running to generate candidate joint hash vectors The system compares the physical ID of the candidate digital device with the current physical ID of the digital device. If the difference exceeds a preset threshold, the ID is updated, and the comparison result is used for the self-verification record and status maintenance of the digital ID.

[0140] This implementation method continuously collects new physical attribute and abnormal state data during device operation, generates corresponding joint hash vectors and candidate digital device physical IDs, and compares them with currently registered IDs for structural consistency. If the difference exceeds a set threshold, an ID update is automatically triggered, and the comparison result is recorded to complete the self-verification process of device identity. This mechanism enables dynamic control and traceability management of device identity as its operating status changes, improves the timeliness, accuracy, and tamper-proof capability of digital IDs, and provides a continuous and reliable identity protection mechanism for the entire lifecycle management of devices.

[0141] This implementation constructs a device physical attribute vector and an abnormal state vector, and merges them into a joint hash vector using hash encoding. This achieves unified modeling of device attribute and state information. Furthermore, a unique digital device physical ID is generated through a mapping function. At the same time, a dynamic update and self-verification mechanism is introduced to continuously sense changes in attributes and fluctuations in state during device operation. If a significant shift in the hash feature structure is detected, an ID update is automatically triggered and a verification log is recorded. This enables accurate characterization, real-time tracking, and trusted management of the device's identity throughout its entire lifecycle, significantly improving the security, real-time performance, and system autonomy of identity management.

[0142] In this embodiment, S53 specifically includes:

[0143] S531, Joint Vector Construction Unit, receives device physical attribute vector. With abnormal state vector A joint input vector is generated by concatenating feature dimensions. ;

[0144] S532, Joint hash calculation unit, for the joint input vector Input into the hybrid hash code and calculate each hash fragment separately. Combine all hash fragments to generate a joint hash vector. Defined as:

[0145] ;

[0146] in, For the first A hash function, The number of hash functions;

[0147] S533, a hash vector normalization unit, normalizing the joint hash vector performing bit width unification, feature dimension compression and numerical normalization operations to obtain a standardized hash representation vector ;

[0148] S534, a bidirectional mapping definition unit, defining a bidirectional mapping mechanism based on the standardized hash representation vector establishing a digital device physical ID mapping mechanism, and defining a forward mapping function:

[0149] ;

[0150] and defining a reverse mapping function:

[0151] ;

[0152] wherein, is a digital device physical ID, and are used for bidirectional mapping between the hash vector and the digital ID.

[0153] The formula maps the standardized hash features to a unique digital ID of the device by establishing a forward mapping function between the hash vector and the digital ID, realizing the conversion of device multi-source information to a unique identity expression; at the same time, a reverse mapping function is defined, so that the system can restore the corresponding hash features from the generated digital ID, realizing the reversible binding between the device identification and the features. The mapping mechanism not only ensures the uniqueness and consistency of the digital ID, but also has traceability and verifiability, supporting subsequent identity resolution and abnormal state tracing.

[0154] The embodiment receives a device physical attribute vector and an abnormal state vector, constructs a joint input vector and inputs it into multiple hash function paths for encoding, generates a joint hash vector, and then performs standardization processing to obtain a hash representation vector in a unified format, and further constructs a forward and reverse mapping function to realize a bidirectional mapping mechanism between the device features and the digital ID. This structure fuses static attributes and dynamic state information into a unique digital identity, which not only improves the discrimination of identity representation, but also supports reverse analysis of device key features from the digital ID, enhances the explainability, uniqueness and traceability of device identity management, and significantly improves the security and intelligent level of device identification.

[0155] Example 1

[0156] In order to verify the feasibility of the application in implementation, the application is applied to

[0157] In order to verify the effectiveness of the deep learning-based digital device physical ID full life cycle management system proposed in the present application, the present application is applied to the power equipment automation management scene of a large intelligent manufacturing enterprise. The enterprise is located in Suzhou Industrial Park, Jiangsu Province, and mainly engages in the assembly and operation and maintenance of precision machining equipment. There are nearly 500 sets of key equipment such as numerical control machining centers, servo motors, industrial fans and pressure pumps in the factory area. These devices are in a long-term multi-condition and high-intensity running state, and the abnormal running state of the devices occurs frequently. The traditional manual inspection and device coding management method cannot meet the systematic requirements for accurate identification of device identity, real-time monitoring of running state and early warning of device abnormalities, especially the outstanding problems such as the inability to trace the identity information, the delay in state updating and the inaccuracy in abnormal positioning during the life cycle of the device.

[0158] In this scene, the present system constructs the device physical attribute vector by deploying data acquisition terminals to sample the state data such as current, voltage, vibration frequency and temperature rise parameters during the running of the device, and combining the static physical information such as device model, structure size and manufacturing batch. All the device running state data are preprocessed by the edge computing node, the effective time sequence features are extracted and normalized, then the preprocessed data are input into the feature fusion module, the local and global semantic features are extracted by using multi-scale dynamic convolution kernel combined with channel attention and spatial attention mechanism, and finally the fusion feature tensor is formed for subsequent abnormal state recognition and digital identity generation.

[0159] During the running process, the system pretrains the deep neural network model using the reconstruction and contrast learning task, realizes the automatic judgment of the abnormal state of the device under different running conditions through the intra-domain feature mapping and the difference evaluation of the recognition results. If the system identifies that the device has an abnormal running state, the system will generate a joint hash vector by hashing and splicing the abnormal state vector and the static attribute of the device, construct a bidirectional mapping function to generate a unique digital device physical ID, and record it in the device master data center. Whenever the device is in a running state, the system will regularly obtain the latest state features and attribute information, automatically calculate the candidate hash vector and digital ID, and compare them with the current registered ID for structural consistency. If the difference exceeds the set threshold, the ID update is triggered, and the update record is written into the device identity log, realizing the dynamic maintenance and full-process self-verification management of the device ID.

[0160] In order to verify the actual application effect of the system, the enterprise selected 60 devices in the same workshop for comparative test, of which 30 devices used the traditional ID management and manual state judgment method, and the other 30 devices were connected to the full life cycle management system of the present application, and the test period was 2 months. The following is the comparison data of device state management:

[0161] Table 1: Comparison of effects of traditional management mode and system management mode of the present application

[0162]

[0163] From the comparison results, it can be seen that the system of the present application effectively reduces the device identity recognition error rate and improves the accuracy and response speed of abnormality detection. In the traditional method, there are problems such as device label damage and manual input error, which lead to confusion of part of the device ID information, and the device state cannot be accurately corresponded to the actual entity. However, through the joint hash mechanism and self-verification process in the present application, the device state awareness and dynamic binding of digital identification can be realized, greatly improving the reliability of device identity.

[0164] In a specific test case, a servo motor with high load operation appeared slight abnormal vibration on the 18th day of operation. Since the traditional system cannot identify the change, it does not record the abnormality in time, and finally causes rotor wear. However, another device with the same specifications using the system of the present application, when the vibration changes in the same operation mode, the system identifies the abnormal state based on multi-scale semantic feature extraction and pre-training model, and triggers ID identification update and maintenance personnel work order notification through the hash update mechanism, which intervenes in maintenance in advance, successfully avoids device failure downtime, saves at least 3 hours of production line downtime and about 12000 yuan of direct loss.

[0165] In addition, in the device operation log management, the system records the device ID update frequency and self-verification results every day, controls the ID update frequency through the standard deviation threshold setting, and ensures that the device can maintain the stability of the identification when the operation fluctuates. As of the end of the test period, the system has completed 734 ID self-verification operations, with an error identification rate of 0, and no false positive state has occurred.

[0166] In summary, the present embodiment fully verifies the effectiveness of the present application in accurate identification of device identity, rapid detection of abnormal state and dynamic management of the whole life cycle in high dynamic industrial environment. The system realizes the accurate binding and real-time maintenance from the physical characteristics of the device to the digital identity, significantly improves the operation and maintenance efficiency and intelligent level, and provides a replicable and scalable intelligent solution for device management in complex industrial scenarios.

[0167] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A deep learning-based digital device physical ID full life cycle management system, characterized in that, Comprise: S1, data acquisition and pretreatment module, adopts multi-channel real-time sampling method to collect equipment running state data and equipment physical property data, and carries out denoising and normalization processing to the collected equipment running state data, generates pretreatment data; S2, feature fusion module, receives pretreatment data, extracts and fuses local and global semantic features by using dynamic convolution kernel and attention mechanism, forms fusion feature data; S3, pre-training module, constructs deep neural network, constructs reconstruction task and contrast learning task based on fusion feature data, carries out multi-round iteration training, outputs optimized pre-training model; S4, abnormality detection module, constructs field adaptive mapping, and carries out field adaptive mapping on equipment running state data under different running conditions, judges whether there is abnormal state based on the recognition result of the optimized pre-training model to the fusion feature data, generates abnormality detection result; S5, digital ID generation module, combines equipment physical property data and abnormality detection result, constructs bidirectional mapping between equipment physical property and digital identification through hybrid hash coding, and dynamically generates, updates and self-verifies digital equipment physical ID; S6, full life cycle management module, based on the abnormality detection result, triggers the update and record operation of the digital equipment physical ID, realizes the maintenance response processing under the abnormal state; The S4 specifically comprises: S41, state data access unit, receives the equipment running state data collected by the data acquisition and pretreatment module, the equipment running state data is a time series tensor, used to reflect the state change process of the target equipment under the current running condition; S42, field adaptive feature mapping unit, constructs field adaptive mapping, reconstructs the semantic space of the equipment running state data by introducing the running condition label, generates the field alignment feature tensor, the field alignment feature tensor unifies the expression distribution between different running conditions in structure, the different running conditions include load state, environment temperature, speed setting, working condition stage, task type, used to represent the distribution characteristics of the equipment running state data under different behavior modes; S43, recognition result generation unit, inputs the fusion feature data and the field alignment feature tensor into the optimized pre-training model, respectively executes forward inference, obtains fusion path recognition result and field path recognition result; S44, a state deviation judging unit, performs vector difference calculation on the fusion path recognition result and the domain path recognition result, and defines a state deviation index function As follows: ； wherein, denotes the fusion path recognition result, denotes the domain path recognition result, is a state shift indicator function; S45, Anomaly Detection Output Unit, for the state offset index function With preset threshold When comparing, The system generates anomaly detection results and outputs anomaly marker data to the numeric ID generation module; otherwise, it generates normal marker data.

2. The deep learning based digital device physical ID lifecycle management system of claim 1, wherein, The S2 specifically comprises: S21, pretreatment data access unit, receives pretreatment data, the pretreatment data is the denoised and normalized equipment running state time series; S22, multi-scale convolution extraction unit, inputs the pretreatment data into parallel convolution branch structure, each branch uses dynamic convolution kernel with different expansion rate, the expansion rate is adaptively configured according to the change rate of the equipment running state time series, outputs initial feature map set, the initial feature map set represents local feature representation under multiple scales; S23, an attention enhancement feature construction unit, receives the initial feature map set, introduces an attention mechanism to each feature map, the attention mechanism includes channel attention and spatial attention, wherein the channel attention generates a channel attention weight vector according to the response intensity distribution of each channel of the feature map, and the spatial attention generates a position weight map according to the response features of different positions in the feature map, and outputs an attention weighted feature map set; S24, a fusion tensor construction unit, splices the attention weighted feature map set and the initial feature map set, and integrates the spliced result by using a linear combination function to generate a fusion feature tensor, the fusion feature tensor contains local features and global context information at the same time; S25, a fusion feature normalization unit, applies a standardization function to the fusion feature tensor, and outputs fusion feature data: ； wherein, is the fused 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, respectively, is a positive real number to prevent numerical instability due to a zero denominator.

3. The deep learning based digital device physical ID lifecycle management system of claim 2, wherein, The S23 specifically includes: S231, a feature map input unit, receives the initial feature map set, the initial feature map set is composed of multiple feature maps with different scale characteristics, and each feature map represents the feature expression of the preprocessed data in a specific receptive field range; S232, a channel attention calculation unit, models the response in the channel dimension of 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 to generate a channel attention weight vector, the channel attention weight vector is used for weighting the initial feature map in the channel dimension, and an enhanced channel feature map set is output; S233, a spatial attention calculation unit, receives the enhanced channel feature map set, calculates the response intensity of each position point in the two-dimensional spatial dimension of each feature map for each feature map, and generates a position weight map, the position weight map is multiplied with the enhanced channel feature map set element by element, and an enhanced spatial feature map set is output; S234, a joint attention fusion unit, reconstructs the enhanced channel feature map set and the enhanced spatial feature map set of each initial feature map in a series fusion manner to generate an attention weighted feature map set.

4. The deep learning based digital device physical ID lifecycle management system of claim 1, wherein, The S3 specifically includes: S31, a fusion feature input unit, receives the fusion feature data and constructs a deep neural network, the fusion feature data is a feature expression processed by a dynamic convolution kernel and an attention mechanism enhancement; S32, a task reconstruction configuration unit inputs the fusion feature data into a self-encoder structure, the self-encoder structure includes two network modules of an encoder and a decoder, and the fusion feature data is compressed and decoded by the two network modules to generate reconstructed fusion feature data , a reconstruction loss function is defined based on the tensor difference between the fusion feature data and the reconstructed fusion feature data :​ ； wherein, is the fusion feature data, is the reconstructed fusion feature data, is the reconstruction loss function; S33, the contrast learning task configuration unit is based on the fusion feature data Constructing multiple augmented views, generating positive sample data pairs and negative sample data pairs through random disturbance, inputting the positive sample data pairs and the negative sample data pairs into the projection sub-network to obtain corresponding feature representation vector pairs, and calculating the contrast loss function according to the Euclidean distance relationship between the feature representation vector pairs , for optimizing the representation distinguishability between different feature representation vector pairs; S34, a joint loss optimization unit, based on the reconstruction loss function with the contrastive loss function constructing a joint optimization objective function defined as follows: ； wherein, with is a non-negative real number, controlling the proportion of the influence 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 a joint optimization objective function, and outputs an optimized pre-training model.

5. The deep learning based digital device physical ID lifecycle management system of claim 1, wherein, The S5 specifically includes: S51, a physical attribute vector construction unit, receiving the device physical attribute data output by the data acquisition and preprocessing module, the device physical attribute data including device model, manufacturing batch, structure parameter and configuration identification, generating device physical attribute vector through field standardization and numerical encoding ; S52, an abnormal state vector construction unit, receiving the abnormality detection result, extracting the abnormality mark, the state offset index and the operation stage, and generating an abnormal state vector for representing a current device operation behavior feature; S53, a joint hash mapping construction unit, based on the device physical attribute vector with the abnormal state vector , forming a joint input vector , and inputting the joint input vector to a hybrid hash encoding to generate a joint hash vector , and constructing a bidirectional mapping structure between the device physical attribute space and the digital identifier space; S54, a digital ID generation unit, based on the generated joint hash vector Constructing a digital device physical ID, defined as follows: ； wherein, is a joint hash vector, is a mapping function from the hash vector space to the digital identity space, is a digital device physical ID; S55, ID updating and self-verification unit, continuously receiving new device physical attribute vector and abnormal state vector during device operation, generating candidate joint hash vector and candidate digital device physical ID, and comparing with the current digital device physical ID for structural consistency, triggering ID updating if the difference exceeds the preset threshold, and using the comparison result for digital ID self-verification record and state maintenance.

6. The deep learning-based digital device physical ID lifecycle management system of claim 5, wherein, The S53 specifically includes: S531, a joint vector construction unit receiving the device physical attribute vector with the abnormal state vector , generating a joint input vector through a feature dimension splicing manner ; S532, the joint hash calculation unit calculates a joint hash vector corresponding to the joint input vector The input is mixed hash coding, and each hash segment is calculated All hash segments are combined to generate a joint hash vector , defined as: ； wherein, is a first hash function, is a second hash function, is a number of hash functions; S533, a hash vector normalization unit, normalizes the joint hash vector performing bit width unification, feature dimension compression and numerical normalization operations to obtain a standardized hash representation vector ; S534, a bidirectional mapping definition unit, based on the standardized hash representation vector A digital device physical ID mapping mechanism is established, and a forward mapping function is defined: ； And define the inverse mapping function: ； wherein, is a digital device physical ID, and are used for bidirectional mapping between hash vector and digital identity, respectively.

Images