Fingerprint identification method and system for terminal equipment

Abstract

The invention discloses a terminal equipment fingerprint identification method and system, and mainly relates to the technical field of power system intelligence. Comprising the steps of collecting multi-source data and extracting basic features; constructing an equipment physical perception neural network model, and realizing hard constraint embedding of a physical law by adopting perception neurons fusing equipment physical parameter tensors and a physical constraint activation function special for equipment types; training the model through a progressive strategy; extracting physical, behavior, time sequence and network flow multi-dimensional features of the target equipment to form a comprehensive fingerprint; and completing identification through multi-dimensional similarity calculation and physical constraint verification. The system comprises corresponding function modules. The method has the beneficial effects that the physical constraint is deeply embedded into the network architecture, so that the physical consistency and high credibility of an identification result are ensured; various terminals such as an electric meter, an inverter, a charging pile and energy storage equipment can be adaptively identified, and the real-time requirement of an electric power system is met while high accuracy is ensured.

Description

Technical Field

[0001] This invention relates to the field of intelligent power system technology, specifically a fingerprint recognition method and system for terminal equipment. Background Technology

[0002] Power system terminal equipment is characterized by its diverse types, heterogeneous protocols, and decentralized deployment. The large-scale access of smart meters, distributed photovoltaic controllers, charging piles, energy storage devices, and other technologies has not only enhanced the intelligence of the system but also brought severe cybersecurity challenges.

[0003] Traditional power systems primarily rely on static authentication mechanisms such as usernames, passwords, and digital certificates. However, these mechanisms exhibit significant limitations in the current environment: static passwords are easily cracked or leaked; digital certificates are costly and complex to manage; many terminals are devices with limited computing power, making it difficult to support complex encryption operations; and traditional authentication is only a one-time verification, unable to continuously monitor identity, leaving vulnerabilities for malicious devices. Furthermore, power equipment is often located in harsh outdoor environments, where factors such as temperature, humidity, and vibration can easily cause hardware characteristic drift, and the system has extremely high real-time requirements; authentication delays can affect grid stability.

[0004] Current cyberattacks are becoming increasingly professional and organized. Attackers infiltrate internal networks by forging device identities to carry out advanced persistent threats such as data theft and command tampering. Traditional perimeter protection and intrusion detection struggle to identify such legitimate, protocol-level forged traffic. To address this, device fingerprinting technology has emerged. It generates a unique "fingerprint" by extracting device hardware, behavioral, and network characteristics, enabling continuous authentication and anomaly detection, suitable for the security needs of power systems with weak computing terminals. Existing solutions mainly include methods based on deep learning, traditional machine learning, and cloud-edge-device collaboration. While these methods can achieve high recognition accuracy and millisecond-level response times, they still have significant shortcomings. For example, one existing method integrating machine learning only performs surface statistical alignment of physical features, without establishing a verifiable physical degradation model. Online updates lack physical consistency verification, are susceptible to environmental interference and hardware aging, and features may be simulated with high fidelity, resulting in weak security boundaries.

[0005] Therefore, there is an urgent need for a device fingerprint recognition method that can deeply integrate real physical constraints, adapt to terminals of multiple service types, and balance high accuracy and high real-time performance. Summary of the Invention

[0006] The purpose of this invention is to provide a fingerprint recognition method and system for terminal devices. It deeply embeds physical constraints into the network architecture, ensuring the physical consistency and high reliability of the recognition results. It can adaptively recognize various terminals such as electricity meters, inverters, charging piles, and energy storage devices, ensuring high accuracy while meeting the real-time requirements of the power system.

[0007] To achieve the above objectives, the present invention employs the following technical solution: On one hand, the present invention provides a fingerprint recognition method for a terminal device, comprising the following steps: Collect multi-source data from several types of terminal devices in the power system and extract basic features including identity information, protocol data, and physical operating parameters; Based on the aforementioned fundamental features, an equipment type mapping table is constructed, comprising several equipment types, and a physical parameter library corresponding to each equipment type is initialized. The physical parameter library includes rated parameters and efficiency coefficients characterizing the specialized physical constraints of the equipment. A device physical perception neural network model is constructed, comprising: a device physical perception neuron fused with a device physical parameter tensor, wherein the physical parameter tensor is dynamically invoked according to the device type; and a physical constraint activation function designed for different device types, which is used to embed the laws of electrical physics as hard constraints into the neuron output calculation. Based on the multi-source data and the corresponding device type labels, a progressive physical learning strategy is used to train the device physical perception neural network model. The strategy includes introducing and enhancing physical constraints at different levels in stages. The trained model is used to process the real-time operating data of the target terminal device and extract multi-dimensional device fingerprint features, including physical features, behavioral features, temporal features and network traffic features, to form a comprehensive fingerprint vector. Based on the comprehensive fingerprint vector, the similarity between it and the pre-stored template fingerprint in the fingerprint database is calculated in multiple dimensions, and combined with the physical constraint verification results, the target terminal device is identified and judged.

[0008] Preferably, the construction of the device physical perception neural network model includes designing device physical perception neurons, the calculation process of which is expressed as follows: ; in, This is the weight matrix. For bias vectors, It is a non-linear activation function. For neuron input, For the device physical parameter tensor, , For the number of equipment types, For physical parameter dimensions, For time step, To be based on device context The identified device type This is a physical constraint modulation function used to convert the physical parameter tensor... The corresponding physical constraint information is modulated onto the output of the traditional neuron, so that the output result conforms to the physical laws of the corresponding device type.

[0009] Preferably, the physical constraint activation functions designed for different device types include: The first activation function for smart meter devices is expressed as follows: ; in, Used to constrain measurement accuracy Used to constrain power balance Used to constrain frequency response characteristics; The expression for the second activation function used in photovoltaic inverter equipment is as follows: ; in, Used to constrain maximum power point tracking efficiency Used to constrain temperature coefficient; The third activation function used for charging pile equipment has the following expression: ; in, Used to constrain charging efficiency. Used to constrain current harmonics; The expression for the fourth activation function used in energy storage devices is: , in, Used to constrain the state of charge. Used to constrain the physical properties of thermal management.

[0010] Preferably, the device physical perception neural network model adopts a hierarchical physical modeling architecture, including: A general electrical physical layer is used to embed fundamental electrical constraints, including Ohm's law, power law, and Kirchhoff's laws. A device-specific physical layer, associated with the device type mapping table, is used to establish a dedicated physical characteristic modeling module for different types of terminal devices; The individual differences physical layer is used to model the individualized characteristics of the same type of equipment caused by manufacturing tolerances, aging, and environmental adaptation.

[0011] Preferably, the progressive physics learning strategy includes four phases divided according to the training cycle: In the first phase, the weights of the loss function are configured to focus on learning the fundamental electrical laws of the general power physical layer; In the second stage, the weight of the equipment-specific physical constraints in the loss function is increased, and the focus is on learning the unique physical laws of each equipment type. The third stage involves introducing and learning the coupling relationships between multiple physical fields, including electricity, heat, and magnetism. In the fourth stage, the model parameters are fine-tuned to achieve the target performance indicators.

[0012] Preferably, the extraction of multidimensional device fingerprint features includes: Physical feature fingerprints are extracted, with 128 dimensions, including voltage, current, power characteristics, and frequency response features; Behavioral feature fingerprints were extracted, with 64 dimensions, including load curves, response patterns, and operating condition change characteristics. Temporal feature fingerprints were extracted, with a dimension of 32, including startup characteristics, steady-state characteristics, and dynamic response characteristics; Network traffic feature fingerprints are extracted with 64 dimensions, including protocol features, packet features, transmission behavior features, and network topology features extracted from Modbus, DNP3, MQTT, and IEC61850 protocol data. Physical constraint verification features are generated based on the device physical sensing algorithm. The above features are fused to form a 288-dimensional comprehensive device fingerprint vector.

[0013] Preferably, before identifying and determining the target terminal device, an adaptive reasoning step is also included: The complexity of the input real-time device data is evaluated, and a comprehensive complexity score is calculated. ; According to the above The value of dynamically selects network branches of different depths for inference: when When, select the first inference branch with the fewest layers and the lowest latency; when When, select the second inference branch with the middle layer and delay; when When the time comes, select the third inference branch with the most layers and the highest accuracy.

[0014] Preferably, the calculation of multi-dimensional similarity uses the following formula: ; in, , , , The similarity is calculated based on physical characteristics, behavioral characteristics, temporal characteristics, and network traffic characteristics, respectively. For compliance scoring based on physical constraint verification, , , , , These are the weighting coefficients for each dimension; When making a judgment, the calculated Compared with the preset high confidence threshold Compare the scores, and if the maximum similarity score is... If the result is positive, a high-confidence identification result will be output; otherwise, the device will be marked as unknown or further verification will be initiated.

[0015] On the other hand, the present invention provides a fingerprint recognition system for a terminal device, for implementing the fingerprint recognition method for a terminal device as described above, comprising: The data acquisition and preprocessing module is used to acquire multi-source data from various terminal devices in the power network and perform data cleaning, normalization and encoding preprocessing. The feature extraction and fusion module is used to extract physical features, behavioral features, temporal features and network traffic features from the preprocessed data and fuse them to form a comprehensive fingerprint vector. The model building and training module is used to build device physical sensing neurons and physical constraint activation functions to form a device physical sensing neural network model, and to train the model using a progressive physical learning strategy. The fingerprint database management module is used to store and manage standard fingerprint templates for known device types, update fingerprints and historical fingerprints in real time, and build an efficient retrieval index. The real-time identification and decision module is used to receive real-time data from the target device, call the trained model to extract its fingerprint features, calculate the multidimensional similarity with templates in the fingerprint database, and combine the physical constraint verification results to complete the final identification and decision of the device identity.

[0016] Preferably, the model building and training module specifically includes: Neuron computing unit, used to perform device-physical sensing neuron computing that incorporates physical parameter tensors; A physical constraint activation function library that stores different physical constraint activation functions for smart meters, photovoltaic inverters, charging piles, and energy storage devices; The physical modeling unit is used to implement the hierarchical physical constraint embedding of the general power physical layer, the device-specific physical layer, and the individual difference physical layer; The training scheduling unit is used to execute the progressive physical learning strategy, configure the loss function in stages, and schedule the model training process.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention offers significant advantages in ensuring physical consistency. By directly embedding physical constraints into a neural network architecture through a device physical perception algorithm, it achieves a shift from soft constraints to hard constraints, ensuring that the identification results strictly meet the fundamental physical laws of power systems. This solves the problem that identification results in traditional methods may violate physical laws. Based on a three-layer progressive physical modeling architecture, from general power laws to device-specific characteristics and then to individual differences, it achieves comprehensive modeling of the physical characteristics of power equipment. This ensures that the entire identification process strictly follows physical laws and has good interpretability and reliability.

[0018] 2. This invention achieves a significant breakthrough in technical performance. Based on the fusion of 288-dimensional multi-dimensional fingerprint features, this device fingerprint recognition method, compared to the traditional PINN method which uses soft-constraint regularization terms, directly embeds power balance hard constraints into the activation function. Under the same training rounds, the proportion of valid solutions inevitably increases, thus logically improving recognition accuracy. Simultaneously, because the adaptive inference branch skips deep networks directly when the input complexity is below a threshold, it reduces recognition latency compared to traditional methods and algorithm stacking schemes. By supporting multiple mainstream power communication protocols such as Modbus, DNP3, MQTT, and IEC61850, this invention has strong multi-protocol adaptability and can effectively handle complex industrial communication environments. The network-physical joint constraint verification algorithm ensures that the recognition results simultaneously satisfy network communication laws and power physics laws, achieving true zero physical constraint violation. Although multi-dimensional features increase computational load, optimization through the DPA (Device-Physics-Aware) algorithm effectively controls resource consumption while maintaining high accuracy, supports efficient deployment of edge computing devices, and significantly shortens the adaptation time for new devices. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the data preprocessing process of the present invention. Detailed Implementation

[0020] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0021] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0022] Example: like Figure 1 As shown, this embodiment provides a fingerprint recognition method for a terminal device, including the following steps: Collect multi-source data from several types of terminal devices in the power system and extract basic features including identity information, protocol data, and physical operating parameters; Based on the basic features, an equipment type mapping table is constructed, which includes several equipment types, and the physical parameter library corresponding to each equipment type is initialized. The physical parameter library includes the rated parameters and efficiency coefficients that characterize the specialized physical constraints of the equipment. A device physical perception neural network model is constructed, which includes: a device physical perception neuron that integrates a device physical parameter tensor, the physical parameter tensor being dynamically called according to the device type; and a physical constraint activation function designed for different device types, which is used to embed the laws of electrical physics as hard constraints into the neuron output calculation. Based on multi-source data and corresponding device type labels, a progressive physical learning strategy is adopted to train the device physical perception neural network model. The progressive physical learning strategy includes introducing and enhancing physical constraints at different levels in stages. The trained model is used to process the real-time operating data of the target terminal device and extract multi-dimensional device fingerprint features, including physical features, behavioral features, temporal features and network traffic features, to form a comprehensive fingerprint vector. Based on the comprehensive fingerprint vector, the similarity between the target terminal device and the pre-stored template fingerprint in the fingerprint database is calculated in multiple dimensions, and combined with the physical constraint verification results, the target terminal device is identified and judged.

[0023] like Figure 2 As shown, this embodiment also provides a fingerprint recognition system for a terminal device, including: The data acquisition and preprocessing module is used to acquire multi-source data from various terminal devices in the power network and perform data cleaning, normalization and encoding preprocessing. The feature extraction and fusion module is used to extract physical features, behavioral features, temporal features and network traffic features from the preprocessed data and fuse them to form a comprehensive fingerprint vector. The model building and training module is used to build device physical sensing neurons and physical constraint activation functions to form a device physical sensing neural network model, and to train the model using a progressive physical learning strategy. The fingerprint database management module is used to store and manage standard fingerprint templates for known device types, update fingerprints and historical fingerprints in real time, and build an efficient retrieval index. The real-time identification and judgment module is used to receive real-time data from the target device, call the trained model to extract its fingerprint features, calculate the multidimensional similarity with the templates in the fingerprint database, and combine the physical constraint verification results to complete the final identification and judgment of the device identity. The model building and training module specifically includes: Neuron computing unit, used to perform device-physical sensing neuron computing that incorporates physical parameter tensors; A physical constraint activation function library that stores different physical constraint activation functions for smart meters, photovoltaic inverters, charging piles, and energy storage devices; The physical modeling unit is used to implement the hierarchical physical constraint embedding of the general power physical layer, the device-specific physical layer, and the individual difference physical layer; The training scheduling unit is used to execute the progressive physical learning strategy, configure the loss function in stages, and schedule the model training process.

[0024] Specifically: 1. Multi-source data acquisition and basic feature extraction for power equipment: Communication data is collected from multi-service type terminal equipment in the new power system, and basic equipment characteristics are extracted, including equipment ID, type identifier, rated parameters, MAC address and other identity information, as well as communication protocol data such as Modbus, DNP3, MQTT, IEC61850, etc. The data are then classified and labeled according to equipment type to establish an initial equipment dataset.

[0025] 2. Equipment type mapping table construction and physical parameter library initialization: The collected equipment data is classified and preprocessed according to equipment type, and an equipment type mapping table is established, including smart meters, distributed photovoltaic controllers, charging piles, energy storage equipment, etc. At the same time, the physical parameter library corresponding to each equipment type is initialized, including equipment-specific physical constraint parameters such as rated voltage coefficient, metering accuracy coefficient, MPPT (Maximum Power Point Tracking) efficiency coefficient, and charging efficiency coefficient.

[0026] 3. Design and implementation of physical sensing neurons for equipment: This paper proposes a device-level physical sensing neuron structure that integrates physical constraints, based on the weight matrix of a traditional neuron. and bias vector Based on this, add the device physical parameter tensor. ,in Number of device types For physical parameter dimensions, Given the time step, the formula for neuron calculation is: This enables the transformation of physical constraints from soft constraints to hard constraints; 4. Design and deployment of the physical constraint activation function family: For different device types, a dedicated family of physical constraint activation functions is designed. Smart meters use the following formula to ensure metering accuracy and frequency response characteristics: , The following formula is used to describe the MPPT characteristics of photovoltaic inverters: , The charging pile uses the following formula to ensure charging efficiency and harmonic constraints: , Energy storage devices use the following formula to achieve State of Charge (SOC) management and thermal constraint: .

[0027] 5. Construction of layered physical modeling architecture: A three-layer progressive physical modeling architecture is constructed, with the first layer being a general electrical physics layer that embeds Ohm's law. Power Law The first layer is a basic electrical constraint such as Kirchhoff's laws; the second layer is a device-specific physical layer, which establishes dedicated physical modules for different equipment types; the third layer is an individual difference physical layer, which models individualized characteristics such as manufacturing tolerances, aging characteristics, calibration offsets, and environmental adaptations.

[0028] 6. Multidimensional feature extraction and encoding of device fingerprints: Three types of device fingerprint features are extracted from the preprocessed device data: ① 128-dimensional physical feature fingerprint, including voltage characteristics, current characteristics, power characteristics, frequency response, etc.; ② 64-dimensional behavioral feature fingerprint, including load curve, response mode, operating condition changes, etc.; ③ 32-dimensional time-series feature fingerprint, including startup characteristics, steady-state characteristics, dynamic response, etc.; ④ 64-dimensional network traffic feature fingerprint, including communication protocol characteristics, data packet characteristics, transmission behavior characteristics, network topology characteristics, etc.; ⑤ Physical constraint verification features based on the DPA algorithm, generating a device signature based on physical constraints, forming a 288-dimensional comprehensive device fingerprint vector.

[0029] 7. Implementation of a progressive physics learning and training strategy: A four-stage progressive physics learning strategy is implemented: Stage 1 (Epoch 1-50) focuses on learning basic electrical laws, with loss weights configured as data fitting (1.0) + basic physics (2.0) + device specialization (0.1) + individual differences (0.05); Stage 2 (Epoch 51-100) focuses on device-specific physics learning, emphasizing the learning of the specific physical laws of each device type; Stage 3 (Epoch 101-150) focuses on multi-physics coupling learning, handling complex coupling relationships such as electro-thermal-magnetic; Stage 4 (Epoch 151-200) focuses on fine-tuning to achieve the target performance indicators (thresholds can be preset and adjusted). 8. Device fingerprint database construction and index optimization: Based on the device fingerprint features obtained through training, a hierarchical device fingerprint database is constructed, including a main fingerprint database that stores standard device fingerprint templates, an online fingerprint database that stores real-time updated fingerprints, and a historical fingerprint database that stores the fingerprint evolution history. An efficient fingerprint retrieval index is established to support physical constraint distance metrics and multi-dimensional feature fusion matching.

[0030] 9. Real-time complexity intelligent evaluation mechanism: The complexity of the input real-time device data is intelligently evaluated, and a comprehensive complexity score is calculated: , in Based on signal spectrum analysis Based on signal-to-noise ratio estimation Based on historical identification confidence variance Based on the integrity of contextual information.

[0031] 10. Adaptive inference path dynamic selection: Dynamically select the inference path based on the complexity score calculated in step 9: when When <0.3, select the fast inference branch (3-layer network, latency ≤15ms, accuracy ≥92%); when 0.3 ≤ When <0.7, select the standard inference branch (5-layer network, latency ≤45ms, accuracy ≥96%); when When the accuracy is ≥ 0.7, the full inference branch (8-layer network, latency ≤ 80ms, accuracy ≥ 98%) is selected to achieve the optimal trade-off between accuracy and real-time performance. In a preferred embodiment, the first and second thresholds of the complexity evaluation output are preset to 0.3 and 0.7, respectively. These values ​​are only for interface demonstration purposes and can be adjusted arbitrarily by those skilled in the art according to the real-time requirements of the site. The adjustment process only requires re-execution of the adaptive branch selection process and does not require retraining the model. Therefore, this application does not limit the specific value of the threshold.

[0032] 11. Real-time extraction of device fingerprints and verification of physical constraints: Real-time device fingerprint extraction is performed on new input device data, including the calculation of physical feature fingerprints, behavioral feature fingerprints, temporal feature fingerprints, and network traffic feature fingerprints, generating a 288-dimensional comprehensive fingerprint vector. Physical constraint verification is used to ensure that the fingerprint features meet the physical laws of the corresponding device type, including basic electrical law verification, device-specific physical constraint verification, individual difference rationality verification, and network-physical joint constraint verification.

[0033] 12. Multidimensional fingerprint similarity calculation and device identification decision: The extracted device fingerprint is compared with the standard fingerprint in the fingerprint database using a five-dimensional similarity calculation. , in The physical constraint verification results based on the DPA algorithm calculate the similarity of each dimension using the physical constraint distance metric function, the network traffic feature distance metric function, and the physical consistency verification function. The final device identification decision is then made based on the comprehensive similarity score and a confidence threshold. If a high-confidence identification result is returned, it is marked as an unknown device or further verification is required.

[0034] Step 1 above specifically includes: Step 1.1: Analyze the basic structural components of the new power system and examine its basic characteristics; Step 1.2: Deploy traffic acquisition devices in the new power system network to collect multi-protocol network communication data over a long period of time, including data packets, transmission timing, and communication behavior patterns of protocols such as Modbus, DNP3, MQTT, and IEC61850. Classify the data according to protocol type, device type, and time window to establish a network traffic dataset.

[0035] Step 2 above specifically includes: Step 2.1: Perform binary one-hot encoding on the high-dimensional data to convert it into numerical data; Step 2.2: Normalize all eigenvalues, as shown in the following formula: , in, These are the normalized eigenvalues. These are the original eigenvalues. The maximum value of the feature. The minimum value of the feature.

[0036] Step 3 above specifically includes: Step 3.1: Construct the core computational unit of the device physical sensing neuron, integrating the device physical parameter tensor into the traditional neuron. This enables hard embedding of physical constraints; Step 3.2: Implement an automatic device type identification mechanism Dynamically select the corresponding subset of physical parameters based on the device context information; Step 3.3: Design the physical constraint modulation function physics_modulate to ensure that the neuron output strictly satisfies the physical laws of the power system, realizing the transformation from data-driven to physics-driven; Step 4 above specifically includes...

[0037] Step 4.1: Design a dedicated activation function for smart meters, embedding metering accuracy constraints, power balance constraints, frequency response constraints, and phase accuracy constraints to ensure that the metering error meets the national standard level 0.2 requirements; Step 4.2: Design a dedicated activation function for the photovoltaic inverter, embedding MPPT efficiency constraints, temperature coefficient constraints (-0.4% / °C), DC-AC conversion efficiency constraints (>95%), and grid compatibility constraints; Step 4.3: Design a dedicated activation function for the charging pile, embedding charging efficiency constraints, harmonic distortion rate constraints (THD<5%), safety protection constraints, and power regulation characteristic constraints; Step 4.4: Design a dedicated activation function for energy storage devices, embedding SOC state constraints (10%-90%), charge and discharge efficiency constraints, thermal management physical constraints, and cycle life impact modeling.

[0038] Step 6 above specifically includes: Step 6.1: Detailed implementation of network traffic feature fingerprint extraction. For different communication protocols such as Modbus, DNP3, MQTT, and IEC61850, device fingerprint features at the network traffic layer are extracted, including protocol feature fingerprints (16-dimensional data frame structure, function code distribution, address field pattern, CRC check features), data packet feature fingerprints (16-dimensional packet size distribution, packet interval statistics, payload length features, fragmentation pattern features), transmission behavior feature fingerprints (16-dimensional communication periodicity, burst transmission pattern, request-response delay, retransmission rate features), and network topology feature fingerprints (16-dimensional number of communication objects, connection mode, routing path, network hop distribution). Step 6.2: For the construction of multi-protocol fusion network fingerprints, a network fingerprint extraction framework supporting multiple power communication protocols is constructed, including the usage frequency of function codes 0x01-0x06, register address access mode, RTU / TCP transmission characteristics, and abnormal response handling mode of Modbus protocol fingerprints; the data link layer confirmation mechanism, application layer function code distribution, time synchronization behavior, and event data transmission mode of DNP3 protocol fingerprints; the topic subscription mode, QoS level distribution, keep-alive heartbeat interval, and last wish message characteristics of MQTT protocol fingerprints; and the GOOSE message publishing frequency, MMS service call mode, SCL configuration characteristics, and sampled value transmission behavior of IEC61850 protocol fingerprints. Step 6.3: For network-physical joint constraint verification, implement joint constraint verification of network traffic characteristics and power physical characteristics, including verification of the correlation between communication frequency and equipment operating conditions to ensure that the communication frequency increases accordingly under high load; verification of the correlation between data transmission volume and power parameter changes to ensure that data transmission increases accordingly when power parameters fluctuate; verification of the consistency between protocol response time and equipment response characteristics to ensure that devices with fast physical response also have correspondingly fast network response; and verification of the correlation between network anomalies and power anomalies to ensure that the network communication mode changes accordingly when there is a power failure.

[0039] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A fingerprint recognition method for a terminal device, characterized in that, Includes the following steps: Collect multi-source data from several types of terminal devices in the power system and extract basic features including identity information, protocol data, and physical operating parameters; Based on the aforementioned fundamental features, an equipment type mapping table is constructed, comprising several equipment types, and a physical parameter library corresponding to each equipment type is initialized. The physical parameter library includes rated parameters and efficiency coefficients characterizing the specialized physical constraints of the equipment. A device physical perception neural network model is constructed, which includes: device physical perception neurons fused with device physical parameter tensors, wherein the physical parameter tensors are dynamically invoked according to the device type; And physical constraint activation functions designed for different device types, used to embed the laws of electrophysics as hard constraints into the neuron output calculation; Based on the multi-source data and the corresponding device type labels, a progressive physical learning strategy is used to train the device physical perception neural network model. The strategy includes introducing and enhancing physical constraints at different levels in stages. The trained model is used to process the real-time operating data of the target terminal device and extract multi-dimensional device fingerprint features, including physical features, behavioral features, temporal features and network traffic features, to form a comprehensive fingerprint vector. Based on the comprehensive fingerprint vector, the similarity between it and the pre-stored template fingerprint in the fingerprint database is calculated in multiple dimensions, and combined with the physical constraint verification results, the target terminal device is identified and judged.

2. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, The construction of the device physical perception neural network model includes designing device physical perception neurons, and the calculation process is expressed as follows: ； in, This is the weight matrix. For bias vectors, It is a non-linear activation function. For neuron input, For the device physical parameter tensor, , For the number of equipment types, For physical parameter dimensions, For time step, To be based on device context The identified device type This is a physical constraint modulation function used to convert the physical parameter tensor... The corresponding physical constraint information is modulated onto the output of the traditional neuron, so that the output result conforms to the physical laws of the corresponding device type.

3. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, The physical constraint activation functions designed for different device types include: The first activation function for smart meter devices is expressed as follows: ； in, Used to constrain measurement accuracy Used to constrain power balance Used to constrain frequency response characteristics; The expression for the second activation function used in photovoltaic inverter equipment is as follows: ； in, Used to constrain maximum power point tracking efficiency Used to constrain temperature coefficient; The third activation function used for charging pile equipment has the following expression: ； in, Used to constrain charging efficiency. Used to constrain current harmonics; The expression for the fourth activation function used in energy storage devices is: ， in, Used to constrain the state of charge. Used to constrain the physical properties of thermal management.

4. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, The device physical perception neural network model adopts a hierarchical physical modeling architecture, including: A general electrical physical layer is used to embed fundamental electrical constraints, including Ohm's law, power law, and Kirchhoff's laws. A device-specific physical layer, associated with the device type mapping table, is used to establish a dedicated physical characteristic modeling module for different types of terminal devices; The individual differences physical layer is used to model the individualized characteristics of the same type of equipment caused by manufacturing tolerances, aging, and environmental adaptation.

5. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, The progressive physics learning strategy comprises four phases divided according to the training cycle: In the first phase, the weights of the loss function are configured to focus on learning the fundamental electrical laws of the general power physical layer; In the second stage, the weight of the equipment-specific physical constraints in the loss function is increased, and the focus is on learning the unique physical laws of each equipment type. The third stage involves introducing and learning the coupling relationships between multiple physical fields, including electricity, heat, and magnetism. In the fourth stage, the model parameters are fine-tuned to achieve the target performance indicators.

6. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, The extraction of multidimensional device fingerprint features includes: Physical feature fingerprints are extracted, with 128 dimensions, including voltage, current, power characteristics, and frequency response features; Behavioral feature fingerprints were extracted, with 64 dimensions, including load curves, response patterns, and operating condition change characteristics. Temporal feature fingerprints were extracted, with a dimension of 32, including startup characteristics, steady-state characteristics, and dynamic response characteristics; Network traffic feature fingerprints are extracted with 64 dimensions, including protocol features, packet features, transmission behavior features, and network topology features extracted from Modbus, DNP3, MQTT, and IEC61850 protocol data. Physical constraint verification features are generated based on the device physical sensing algorithm. The above features are fused to form a 288-dimensional comprehensive device fingerprint vector.

7. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, Before identifying and determining the target terminal device, an adaptive reasoning step is also included: The complexity of the input real-time device data is evaluated, and a comprehensive complexity score is calculated. ; According to the above The value of dynamically selects network branches of different depths for inference: when When, select the first inference branch with the fewest layers and the lowest latency; when When, select the second inference branch with the middle layer and delay; when When the time comes, select the third inference branch with the most layers and the highest accuracy.

8. The fingerprint recognition method for a terminal device according to claim 1, characterized in that, The multi-dimensional similarity is calculated using the following formula: ； in, , , , The similarity is calculated based on physical characteristics, behavioral characteristics, temporal characteristics, and network traffic characteristics, respectively. For compliance scoring based on physical constraint verification, , , , , These are the weighting coefficients for each dimension; When making a judgment, the calculated Compared with the preset high confidence threshold Compare the scores, and if the maximum similarity score is... If the result is positive, a high-confidence identification result will be output; otherwise, the device will be marked as unknown or further verification will be initiated.

9. A fingerprint recognition system for a terminal device, used to implement the fingerprint recognition method for a terminal device as described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source data from various terminal devices in the power network and perform data cleaning, normalization and encoding preprocessing. The feature extraction and fusion module is used to extract physical features, behavioral features, temporal features and network traffic features from the preprocessed data and fuse them to form a comprehensive fingerprint vector. The model building and training module is used to build device physical sensing neurons and physical constraint activation functions to form a device physical sensing neural network model, and to train the model using a progressive physical learning strategy. The fingerprint database management module is used to store and manage standard fingerprint templates for known device types, update fingerprints and historical fingerprints in real time, and build an efficient retrieval index. The real-time identification and decision module is used to receive real-time data from the target device, call the trained model to extract its fingerprint features, calculate the multidimensional similarity with templates in the fingerprint database, and combine the physical constraint verification results to complete the final identification and decision of the device identity.

10. A fingerprint recognition system for a terminal device according to claim 9, characterized in that, The model building and training module specifically includes: Neuron computing unit, used to perform device-physical sensing neuron computing that incorporates physical parameter tensors; A physical constraint activation function library that stores different physical constraint activation functions for smart meters, photovoltaic inverters, charging piles, and energy storage devices; The physical modeling unit is used to implement the hierarchical physical constraint embedding of the general power physical layer, the device-specific physical layer, and the individual difference physical layer; The training scheduling unit is used to execute the progressive physical learning strategy, configure the loss function in stages, and schedule the model training process.

Images