A cloud edge collaboration-based power protocol adaptive fusion method and system

Abstract

The application provides a power protocol adaptive fusion method and system based on cloud edge cooperation, and relates to the technical field of power Internet of Things.Through the edge intelligent terminal, the classification and identification of the protocol message and the confidence determination are completed, the existing protocol can be quickly identified locally, and the new protocol message is screened and reported to the cloud for processing, so that the full incremental training of the model on the edge side is not needed, and the problems of insufficient edge computing power and decreased identification accuracy of the old protocol are avoided; the local prompt word vector library is dynamically updated through the prompt word vector package returned by the cloud, the model identification range is expanded with a small-scale parameter increment, and the accurate adaptation of the new protocol is completed through secondary identification, so that the identification accuracy of various power protocols is effectively improved under the premise of ensuring the lightweight operation of the edge, and the adaptive fusion and flexible access of heterogeneous terminal protocols are realized.

Description

Technical Field

[0001] This invention relates to the field of power Internet of Things (IoT) technology, and in particular to a cloud-edge collaborative adaptive fusion method and system for power protocols. Background Technology

[0002] With the continuous advancement of the construction of new power systems, the ubiquitous power Internet of Things (IoT) is extending deeply to the edge of the distribution network, with a large number of heterogeneous digital terminals such as low-voltage smart circuit breakers, distributed photovoltaic inverters, electric vehicle charging piles, and energy storage converters being widely connected. To achieve full-dimensional power operation status perception, unified equipment management and control, and digital twin modeling, edge intelligent fusion terminals need to identify, parse, and standardize the power industry protocols used in the southbound communication of various terminals in real time. Protocol adaptive fusion capability has become the core foundation supporting the digitalization of the distribution network edge.

[0003] Current power protocol parsing primarily employs a rule-based deterministic parsing scheme. This scheme pre-configures byte offset rules and semantic mapping templates for each standard protocol to achieve structured parsing of known protocols. While this approach ensures the accuracy of parsing known protocols, it only supports pre-registered standard protocols. When faced with numerous vendor-customized proprietary and variant protocols in the power field, manual reverse engineering of the protocol format and firmware updates are required, resulting in long deployment cycles, high maintenance costs, and an inability to meet the rapid iteration and plug-and-play access requirements of distribution network terminals.

[0004] To improve the generalization ability of protocol recognition, some solutions introduce deep learning models to achieve end-to-end intelligent protocol recognition, automatically extracting message features and completing classification through neural networks. However, edge terminals have limited computing power and memory resources. If incremental training and backpropagation of the model are performed locally, memory overflow is likely to occur, and frequent parameter updates can lead to catastrophic forgetting, causing a significant decrease in the recognition accuracy of the original mainstream protocols.

[0005] Therefore, current power protocol identification schemes for edge devices suffer from low identification accuracy. Summary of the Invention

[0006] This invention provides a cloud-edge collaborative adaptive fusion method and system for power protocols, which solves the problem of low accuracy in power protocol identification schemes for edge devices.

[0007] In a first aspect, the present invention provides an adaptive fusion method for power protocols based on cloud-edge collaboration, applied to an edge intelligent terminal. The method includes: detecting a protocol message to be identified; classifying and scoring the protocol message and its pre-stored protocol identification model on the edge intelligent terminal to obtain a primary identification result; the primary identification result includes existing protocol types or new protocol types; if the identification result is a new protocol type, sending the protocol message to be identified to a cloud platform; receiving a prompt word vector packet returned by the cloud platform after receiving the protocol message; updating the local prompt word vector library of the edge intelligent terminal based on the prompt word vector packet, and analyzing the result of the protocol identification model to obtain a secondary identification result of the protocol message, thereby achieving adaptive fusion of power protocols.

[0008] Secondly, embodiments of the present invention provide a cloud-edge collaborative adaptive fusion method for power protocols, applied to a cloud platform. The method includes: receiving protocol messages to be identified from multiple edge intelligent terminals, wherein the protocol messages to be identified are sent after being determined by the protocol identification model in the edge intelligent terminals to be new protocol types; extracting payload features from the protocol messages to be identified sent by each edge intelligent terminal; performing cluster analysis based on the payload features of each protocol message to be identified to obtain multiple candidate categories and multiple protocol messages to be identified corresponding to each candidate category; based on the multiple candidate categories and the multiple protocol messages to be identified corresponding to each candidate category, and combined with the protocol identification model of the cloud platform, obtaining prompt word vectors and parsing mapping scripts for each candidate category; the global weight parameters of the protocol identification model of the cloud platform and the protocol identification model of the edge intelligent terminals are the same; and transmitting the prompt word vectors and parsing mapping scripts for each candidate category back to the edge intelligent terminals.

[0009] Thirdly, embodiments of the present invention provide a cloud-edge collaborative power protocol adaptive fusion system. The fusion system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0010] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0011] This invention provides a cloud-edge collaborative adaptive fusion method and system for power protocols. The invention performs primary classification and confidence assessment of protocol messages through an edge intelligent terminal, enabling rapid local identification of existing protocols while simultaneously filtering out new protocol messages for cloud processing. This eliminates the need for full incremental model training at the edge, avoiding the problems of insufficient edge computing power and decreased accuracy in recognizing older protocols. The local prompt word vector library is dynamically updated using prompt word vector packets transmitted from the cloud, expanding the model's recognition range with minimal parameter increments. Combined with secondary recognition, this achieves accurate adaptation to new protocols. While ensuring lightweight operation at the edge, this effectively improves the recognition accuracy of various power protocols, enabling adaptive fusion and flexible access of heterogeneous terminal protocols. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an adaptive fusion method for power protocols based on cloud-edge collaboration provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall operating architecture and data flow drag-and-drop topology provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of an edge-end dual-threshold routing and security self-locking process provided by an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0015] As described in the background section, there is a clear inherent contradiction between edge computing resources and incremental learning. Traditional deep learning protocol recognition models (such as 1DCNN and LSTM) face the challenge of continuously evolving and newly integrated proprietary device protocols. If full gradient backpropagation training is run directly on the local edge terminal, the massive computational graph and optimizer state space can easily exhaust the fragile memory and computing resources of ARM-based embedded chips, leading to memory overflow (OOM) and gateway crashes. Simultaneously, directly updating parameters can easily cause "catastrophic forgetting" in the neural network, significantly reducing the classification accuracy for existing general protocols.

[0016] The feature space is prone to collapse under the constraint of extremely small sample sizes. The sample size of the tail variant protocol under the long-tail distribution is extremely small. Due to the lack of sufficient data manifold support, the feature gradient of traditional classification models is easily dominated and submerged by the large class flow during the optimization process of multi-branch network, which leads to a sharp degradation in generalization performance and serious false negatives and false positives.

[0017] Blind semantic parsing poses a catastrophic security risk to industrial control systems. When existing gateways encounter completely unknown, unregistered private protocol messages, if they forcibly use end-to-end large models or fully connected neural networks for crude semantic guessing and field machine translation, the risk of "hallucination" that the model itself cannot eliminate can easily lead to the misjudgment of unknown measurement and control loads as sensitive power control commands (such as misinterpreting routine temperature over-limit telemetry data as high-voltage tripping commands), causing catastrophic malfunctions in the primary power physical system and seriously violating the bottom line of power system network security protection.

[0018] To address the challenges of limited edge computing power, small-sample forgetting, and industrial control system (ICS) safety against malfunctions, this invention provides an adaptive fusion method for power protocols based on cloud-edge collaborative prompting learning and transparent encapsulation. Breaking away from the traditional limitations of "modifying and stacking model architectures" and "heavy-asset training at the edge," it constructs a cloud-edge collaborative architecture system characterized by "lightweight edge inference and secure transparent self-locking, high-power cloud-based small-sample learning and micro-prompt word hot updates." Through a collaborative closed loop of lightweight edge self-locking inference, high-power cloud-based small-sample fine-tuning, and micro-prompt word hot updates, it achieves near-zero parameter growth for the dynamic access of massive heterogeneous devices while ensuring 100% anti-misinterpretation safety for the power grid control system. While guaranteeing absolute anti-misinterpretation safety for the power industry control system, it achieves "plug-and-play" adaptive seamless fusion of massive, diverse, heterogeneous, long-tail protocols with near-zero edge parameter growth energy efficiency.

[0019] like Figure 1 As shown, the present invention provides a cloud-edge collaborative adaptive fusion method for power protocols, which is applied to edge intelligent terminals. The method includes steps S101-S105. S101. Detect the protocol message to be identified.

[0020] S102. Based on the protocol message to be identified and the protocol identification model pre-stored in the edge intelligent terminal, perform classification identification and classification confidence score to obtain the first identification result of the protocol message to be identified.

[0021] In this embodiment of the application, the identification result includes an existing protocol type or a new protocol type.

[0022] In some embodiments, a deep pre-trained Transformer encoder base is deployed at the edge, and its parameter matrix W is fully frozen, allowing the model to perform only forward propagation at the edge gateway. The gateway extracts 80-dimensional statistical features of the pure application layer load and performs Z-score normalization. Subsequently, known protocol soft cue vectors P stored in memory are dynamically concatenated to construct a deep input tensor X with prior indications. input The base model is then used to calculate the logits of each protocol class.

[0023] The system constructs a dual-threshold routing isolation mechanism based on the maximum confidence value S output by the Softmax layer (setting a threshold T). h =0.85). For S≥T h The message goes through branch A to perform deterministic semantic structure parsing, generating IoT JSON format conforming to the CIM standard and reporting it; for S <T h For high-risk unknown messages, branch B executes industrial-grade anti-misinterpretation self-locking: forcibly blocking semantic guessing, performing Base64 secure encoding on the original binary stream, packaging it into a generalized text security envelope, and transmitting it to the cloud.

[0024] When the accumulated data of unknown envelopes in the cloud reaches a small sample size, a GPU cluster with the same freezing base is activated. The cloud uses Class-Balanced Focal Loss to back-calculate the gradient of a single specific cue vector, accurately extracting the high-dimensional spatial mapping relationship of the unknown protocol. After training, only this KB-level lightweight soft cue vector is distributed to the edge gateway via OTA wireless private network. The gateway implements hot-swappable memory registration, instantly closing the loop for the recognition capability of the new protocol.

[0025] like Figure 2 and Figure 3As shown, the edge intelligent fusion terminal is equipped with a low-power quad-core ARM processor, and the cloud-integrated demonstration platform relies on a data center GPU cluster. In baseline testing, the long-tail distribution of protocols is extremely significant; for example, the proportion of IEC-104 category samples is only 0.98%. The base network uses a lightweight Transformer encoder with a hidden layer dimension of D=128. The soft cue vector Pm dimension designed for each protocol is set to 8×16, so the storage space for the cue vector of a single protocol is only about 2KB, which greatly optimizes memory resources. The latency of a single forward inference mathematical calculation during ONNX runtime at the edge terminal is controlled at approximately 0.0497 milliseconds, perfectly supporting line-speed packet capture detection.

[0026] Implementation Scenario A: High-confidence awareness and structured mapping of known protocols.

[0027] When an edge gateway intercepts a packet from a pole-mounted switch via mirroring on a switch's physical port, the raw hexadecimal stream sequence of the application layer is obtained after stripping the outer shell: 68 04 01 00 02 00. Feature extraction and normalization yield an 80-dimensional dense vector. This vector is then concatenated with a soft cue vector library and input into the base network. The final fully connected classification layer of the model outputs the raw Logits vector y=[12.4,-2.1,5.3,...].

[0028] Calculated using the Softmax function: ; The confidence score S is 0.9682. The system detected that S≥T. h (0.85), and the corresponding category 1 is "IEC-104 protocol". The gateway hits branch A, automatically calls the deterministic mapping module, parses the remote signaling closure state according to the protocol specification, and outputs the complete SG-CIM format object model: { "timestamp":"2026-05-24T12:00:03Z", "recognized_protocol":"iec-104", "confidence_score": 0.9682, "conversion_strategy":"Deterministic_Mapping", "data":{ "ASDU_Type":"Measured_Value", "Telemetry":{ "Voltage_V":220.4, "Switch_Status":"Closed" } } } Implementation Scenario B: Security blocking and incremental pass-through of unknown private garbled messages.

[0029] When a new type of distributed charging pile is connected to the charging station area, it sends an unregistered private binary message stream: AA BB CC DD11 22 33 44 FF. After the same process of inputting it into the base for inference, because the model has no prior indication of this data distribution, the variance of the generated log-probability is extremely small. After Softmax calculation, the maximum probability peak S is only 0.4521.

[0030] Because S <T h (0.85), the gateway triggers a high-level industrial security self-locking response (branch B): forcibly blocking all decoding rules attempting to disassemble the underlying bytes, completely abandoning semantic guessing. The gateway extracts the 9 bytes of raw binary data, performs Base64 encoding to convert it into a string with no execution harm (such as "qrvM3REiMwT / "), and encapsulates and sends it up: { "timestamp":"2026-05-24T12:00:08Z", "recognized_protocol": "Unknown", "confidence_score": 0.4521, "conversion_strategy":"Transparent_Tunneling", "data":{ "Raw_Payload_Base64":"qrvM3REiMwT / ", "Warning":"Trigger cloud-edge interactive prompt tuning", "Metadata":{"Source_Port":"eth1","Byte_Length":9} } } The cloud-integrated platform captures a large number of packets of the aforementioned "Unknown" type, triggering cloud computing power. While strictly maintaining the base parameters frozen, the cloud network calculates the partial derivative gradient ∇Pnew with respect to the newly created cue vector Pnew according to the category balance focus loss formula, and updates its weights using the AdamW optimizer. After training, the platform pushes a Pnew file, only 1.8KB in size, over-the-air (OTA) via the 5G network. The edge gateway instantly registers it into memory, gaining a high-confidence awareness capability of 0.90 level for the new protocol without needing to re-flash the firmware, thus completing the adaptive full-lifecycle fusion evolution of cloud-edge collaboration.

[0031] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1025.

[0032] S1021. Extract the message payload of the protocol message to be identified.

[0033] S1022. Perform flow-level and packet-level spatiotemporal statistical analysis on the message payload, extract fixed-length high-dimensional statistical feature vectors and perform normalization processing to obtain normalized high-dimensional feature vectors.

[0034] S1023. Based on the normalized high-dimensional feature vector and the currently activated known protocol prompt vector in the local prompt word vector library of the edge intelligent terminal, construct a deep prompt input sequence.

[0035] S1024. Based on the deep cue input sequence and the protocol recognition model, perform forward inference to obtain a first classification confidence score for the protocol message to be identified.

[0036] In some embodiments, the protocol identification model is a deep pre-trained encoder base network deployed on an edge smart terminal with all global weight parameters frozen.

[0037] S1025. Based on the classification confidence score of the protocol message to be identified, a judgment is made to obtain an identification result.

[0038] In some embodiments, the present invention can deploy a deep pre-trained Transformer encoder network as a base network in an edge intelligent fusion terminal and freeze all global weight parameters of the base network; and construct a prompt vector library and a deterministic semantic parsing rule tree of known power protocols locally in the edge intelligent fusion terminal.

[0039] Subsequently, the edge intelligent fusion terminal captures southbound physical communication traffic in real time, strips the protocol headers of the data link layer, network layer, and transport layer, and extracts the binary byte stream of the pure application layer message payload. Statistical spatiotemporal feature analysis is performed on the message payload to extract a fixed-length high-dimensional feature vector, and feature space normalization is performed. The normalized high-dimensional feature vector is transformed into a feature embedding vector, and k known protocol soft cue vectors activated in the current cue vector library are dynamically concatenated at its front end to construct a deep cue input sequence. This sequence is then input into the base network for forward inference, outputting the original logarithmic probabilities for each category. After normalization using the Softmax activation function, the maximum probability value is extracted as the classification confidence score of the current message.

[0040] For example, embodiments of the present invention may introduce a set industrial security confidence threshold Th, and dynamically drive a dual-branch self-locking routing mechanism based on the confidence score: if the confidence score is greater than or equal to the security threshold Th, it is determined to be a known power protocol and enters the deterministic semantic mapping branch; if the confidence score is strictly less than the security threshold Th, it triggers anti-misinterpretation self-locking and enters the zero-trust transparent security encapsulation branch.

[0041] For example, the feature space normalization formula performed after feature extraction is as follows: (2-1); In the formula, Represents the j-th dimension of the original feature value of the i-th data stream, and σ and σ represent the statistical mean and standard deviation of the j-th dimension feature calculated based on the baseline of the total traffic of the large-scale distribution network, respectively. These are the normalized eigenvalues.

[0042] In some embodiments, the mathematical constructive expression for the depth cue input sequence is: (3-1); In the formula, P m ∈R (Lp×D) L represents the m-th consecutive soft cue vector. p Where D is the length of the cue word tag and D is the feature embedding dimension; ∈ R (1×D) This represents the dense feature embedding matrix obtained by projecting a normalized high-dimensional feature vector through a linear layer.

[0043] In some embodiments, the mathematical expression for calculating the normalized probability and extracting the confidence score is as follows: (4-1); (4-2); In the formula, Let C be the logits of the model output layer for the m-th protocol class, where C is the total number of currently supported protocol classes. To predict the probability of belonging to the m-th protocol class, S is the classification confidence score of the current message.

[0044] S103. If the identification result is a new protocol type, the protocol message to be identified will be sent to the cloud platform.

[0045] For example, if the identification result is a new protocol type, the anti-misinterpretation self-locking mechanism is triggered, and zero-trust transparent security encapsulation is performed on the protocol message to be identified. The byte-field-level semantic decoding of the message payload is forcibly prohibited. The original binary byte stream of the message payload is directly extracted and transcoded. The transcoded data is combined with the timestamp and source MAC address to package it into a transparent security envelope labeled with the unknown protocol type, and the transparent security envelope is sent to the cloud platform.

[0046] Optionally, after step S102, if the identification result is an existing protocol type, then based on the deterministic semantic parsing rule tree pre-set locally in the edge intelligent terminal, combined with the protocol type of the protocol message to be identified, the byte offset and decoding logic are determined; based on the byte offset and decoding logic, the telemetry field and remote signaling field in the message payload of the protocol message to be identified are extracted and mapped into a standardized object model.

[0047] S104. Receive the prompt word vector packet returned by the cloud platform after receiving the protocol message to be identified.

[0048] In some embodiments, the prompt word vector package includes prompt word vectors corresponding to the protocol messages to be identified as new protocol types by each edge smart terminal.

[0049] As one possible implementation, the method for generating word vectors includes steps A1-A5.

[0050] A1. The cloud platform continuously receives protocol messages to be identified from various edge intelligent terminals.

[0051] A2. Extract the payload characteristics of each protocol message to be identified.

[0052] A3. Based on the payload characteristics of each protocol message to be identified, cluster analysis is performed to obtain multiple candidate categories and multiple protocol messages to be identified corresponding to each candidate category.

[0053] A4. Based on multiple candidate categories and multiple protocol messages to be identified for each candidate category, and combined with the protocol recognition model of the cloud platform, the prompt word vectors and parsing mapping scripts for each candidate category are obtained.

[0054] In some embodiments, the global weight parameters of the protocol identification model of the cloud platform and the protocol identification model of the edge intelligent terminal are the same.

[0055] A5. Send the prompt word vectors and parsing mapping scripts for each candidate category back to the edge intelligent terminal.

[0056] In some embodiments, the cloud-integrated empirical platform continuously receives and aggregates the encapsulated information of unknown protocols reported by each edge intelligent fusion terminal. When the cumulative sample size of a specific unknown category reaches a preset small sample threshold, the platform performs backpropagation-based fine-tuning training for the unknown category data stream, while keeping the global parameters of the base network completely frozen in the cloud, and updates the dedicated soft cue vector corresponding to the (k+1)th unknown protocol. The cloud-integrated empirical platform packages the newly trained dedicated soft cue vector and the corresponding parsing mapping script, and sends it to the edge intelligent fusion terminal via a wireless / wired channel. The edge intelligent fusion terminal dynamically registers the new soft cue vector to the local cue vector library list through a hot-loading mechanism, completing the proactive perception and seamless adaptive fusion of the new protocol.

[0057] In some embodiments, the specific logic of the deterministic semantic mapping branch is as follows: retrieve the rule tree according to the identified protocol type, extract the telemetry and teleindication fields by byte offset, and assemble them into an IoT JSON object model conforming to the SG-CIM specification; the specific logic of the zero-trust transparent security encapsulation branch is as follows: forcibly prohibit byte field-level decoding, directly extract the original binary stream, perform Base64 transcoding, and combine it with timestamp and MAC address to package it into a transparent security envelope marked Unknown.

[0058] S105. Based on the prompt word vector package, update the local prompt word vector library of the edge intelligent terminal, and combine it with the protocol recognition model to analyze and obtain the secondary recognition result of the protocol message to be recognized, so as to realize the adaptive fusion of power protocols.

[0059] As one possible implementation, step S105 can be specifically implemented as steps S1051-S1054.

[0060] S1051. Based on the prompt word vector package, the prompt word vector package is dynamically registered and appended to the local prompt word vector library through a hot loading mechanism to obtain the updated local prompt word vector library.

[0061] S1052. Based on the protocol message to be identified and the updated local prompt word vectors, a new deep prompt input sequence is generated by concatenating them.

[0062] S1053. Input the new deep cue input sequence into the protocol recognition model and perform forward inference to obtain the secondary classification confidence score of the protocol message to be recognized.

[0063] S1054. Based on the secondary classification confidence score and the parsing mapping script returned by the cloud platform, deterministic semantic parsing is performed on the message payload of the protocol message to be identified to obtain the secondary identification result of the protocol message to be identified.

[0064] In some embodiments, the secondary identification result includes the protocol type of the protocol message to be identified and the secondary classification confidence score.

[0065] This invention provides a cloud-edge collaborative adaptive fusion method and system for power protocols. The invention performs primary classification and confidence assessment of protocol messages through an edge intelligent terminal, enabling rapid local identification of existing protocols while simultaneously filtering out new protocol messages for cloud processing. This eliminates the need for full incremental model training at the edge, avoiding the problems of insufficient edge computing power and decreased accuracy in recognizing older protocols. The local prompt word vector library is dynamically updated using prompt word vector packets transmitted from the cloud, expanding the model's recognition range with minimal parameter increments. Combined with secondary recognition, this achieves accurate adaptation to new protocols. While ensuring lightweight operation at the edge, this effectively improves the recognition accuracy of various power protocols, enabling adaptive fusion and flexible access of heterogeneous terminal protocols.

[0066] This invention solves the problems of insufficient computing power and catastrophic forgetting at the edge: by globally freezing the base and updating only the micro-prompt, the scale of edge parameters achieves near-zero growth. The edge is free from retraining burdens, completely avoiding the pain point of forgetting old protocols due to new knowledge overwriting. It constructs an absolute physical security defense against malfunctions in industrial control networks: this invention breaks away from the traditional end-to-end guessing AI black box logic and pioneers a confidence-based routing filtering mechanism. Base64 transparent encapsulation is enforced for low-confidence packets, completely blocking the risk of false control commands caused by AI illusions at the mechanistic level. It possesses extremely strong long-tail feature extraction and plug-and-play engineering capabilities: a rigorous mathematical loss function (CB-Focal Loss) enhances the model's ability to perceive small-sample, extremely long-tail protocols. New equipment can be connected on-site without power outages and firmware re-burning; knowledge extracted from the cloud is instantly hot-downloaded, resulting in extremely low overall deployment costs.

[0067] Optionally, the cloud-edge collaborative power protocol adaptive fusion method provided in this embodiment of the invention further includes steps S201-S203 before step S102.

[0068] S201. Deploy a pruned and quantized compressed deep pre-trained encoder network as a base network in the edge intelligent terminal, and freeze all global weight parameters of the base network so that the base network only performs forward inference on the edge side and does not generate gradient calculation and backpropagation overhead.

[0069] S202. Build a local prompt word vector library on the edge intelligent terminal. The local prompt word vector library stores soft prompt vectors corresponding to various existing power protocols.

[0070] S203. Configure the deterministic semantic parsing rule tree for each known power protocol locally on the edge intelligent terminal.

[0071] Thus, this invention deploys a lightweight, pruned and quantized base network on the edge terminal and freezes all parameters, allowing the model to perform only forward inference. This completely avoids gradient calculation and backpropagation overhead, significantly reducing memory usage and computing power consumption on the ARM embedded platform and effectively preventing OOM (Out of Memory) crashes. At the same time, freezing parameters fundamentally eliminates the catastrophic forgetting of existing protocol knowledge during incremental learning, ensuring stable and high-precision recognition of general protocols. The locally pre-built hint vector library and deterministic rule tree provide a solid foundation for subsequent zero-overhead hot-swappable expansion and misjudgment-preventing secure parsing.

[0072] Optionally, embodiments of the present invention also provide a cloud-edge collaborative adaptive fusion method for power protocols, applied to a cloud platform, including steps S301-S305.

[0073] S301: Receive protocol messages to be identified from multiple edge intelligent terminals.

[0074] In some embodiments, the protocol message to be identified is sent after the protocol identification model in the edge smart terminal determines that it is a new protocol type.

[0075] S302. Extract payload features from the protocol messages to be identified sent by each edge intelligent terminal.

[0076] S303. Based on the payload characteristics of each protocol message to be identified, perform cluster analysis to obtain multiple candidate categories and multiple protocol messages to be identified corresponding to each candidate category.

[0077] S304. Based on multiple candidate categories and multiple protocol messages to be identified for each candidate category, and combined with the protocol recognition model of the cloud platform, the prompt word vectors and parsing mapping scripts for each candidate category are obtained.

[0078] In some embodiments, the global weight parameters of the protocol identification model of the cloud platform and the protocol identification model of the edge intelligent terminal are the same.

[0079] In some embodiments, the cloud platform indicates the parameter gradient ∇ of the base network during backpropagation in fine-tuning training. W The value is always 0, and only the parameter P of the dedicated soft cue vector is calculated and updated. (k+1) gradient∇ P (k+1) .

[0080] In some embodiments, the objective optimization function used in the backpropagation-guided fine-tuning training is the Class-Balanced Focal Loss function, whose core mathematical expression is: (7-1); In the formula, To characterize the size of the effective sample set, The number of small samples accumulated for the currently unknown protocol category, γ is a focus adjustment factor used to amplify the classification weight gradient of long-tail protocols with extremely small samples. This represents the predicted probability of the corresponding true label category output by the model.

[0081] S305. Send the prompt word vectors and parsing mapping scripts for each candidate category back to the edge intelligent terminal.

[0082] This invention discloses an adaptive fusion method for power protocols based on cloud-edge collaborative cue learning and transparent encapsulation, belonging to the fields of power Internet of Things and industrial control security. The method deploys a Transformer base with completely frozen parameters on the edge intelligent terminal, strips away the transport layer shell of physical traffic to extract pure application layer 80-dimensional spatiotemporal statistical features, and normalizes them; hot-swappable cue vectors are concatenated before the features and input into the base for forward inference to calculate probability scores. The system innovatively constructs an industrial isolation mechanism based on dual-threshold routing: deterministic parsing is performed on high-confidence mainstream protocols to generate IoT CIM format JSON; for uncertain low-confidence messages, zero-trust anti-misinterpretation self-locking is triggered, forcibly performing Base64 secure encoding and encapsulating it into a transparent envelope for transparent transmission to the cloud. After cloud-aggregated samples, the soft cue vector with only minimal parameters is fine-tuned based on class-balanced focus loss (CB-Focal Loss), and then hot-loaded at the edge via OTA. This invention not only completely solves the bottleneck of computing power and the pain point of catastrophic forgetting in edge devices, but also 100% shields the risk of tripping and miscontrolling caused by the blind tampering of AI black boxes from the mechanism, realizing the safe 'plug and play' access of long-tail distribution network equipment at extremely low cost.

[0083] Optionally, steps S401-S403 may be included before step S304.

[0084] S401. Determine the number of protocol messages to be identified in each candidate category.

[0085] S402. If the number of protocol messages to be identified in a candidate category is greater than or equal to the preset small sample threshold, then for that candidate category, while keeping the global weight parameters of the protocol recognition model completely frozen, perform prompt fine-tuning training to generate prompt word vectors and parsing mapping scripts corresponding to that candidate category.

[0086] S403. If the number of protocol messages to be identified in a candidate category is less than the preset small sample threshold, the protocol messages to be identified in that candidate category reported by each edge intelligent terminal will be continuously aggregated, and the prompt fine-tuning training will not be started for the time being.

[0087] Thus, by setting a small sample threshold as the trigger condition for cloud-based soft cue fine-tuning, this invention avoids overfitting and cue vector degradation caused by forced training when samples are extremely scarce. This ensures that the model only starts learning when there is sufficient data to support it, guaranteeing the generalization and reliability of the extracted soft cue vectors. At the same time, a continuous aggregation mechanism is adopted for categories that have not reached the threshold, which not only saves cloud GPU computing resources but also lays the foundation for subsequent high-quality fine-tuning through an accumulation and waiting strategy, achieving the optimal balance between cloud learning efficiency and model accuracy.

[0088] For example, the cloud-edge collaborative power protocol adaptive fusion method provided in this embodiment of the invention can be implemented as steps one to six.

[0089] Step 1: Deployment and initialization of the lightweight base network for the edge gateway.

[0090] In edge devices such as smart converged terminals or smart IoT gateways in distribution areas, a deep pre-trained Transformer encoder network, pruned and quantized for compression, is deployed. To ensure long-term stable operation of the gateway and completely avoid catastrophic forgetting, the global weight parameters of the pre-trained Transformer encoder are completely frozen at the edge (i.e., set to an untrainable state). Simultaneously, a prompt vector library and deterministic semantic parsing rule trees corresponding to known mainstream power protocols (such as IEC-104, Modbus, DL / T-645, etc.) are constructed locally at the edge.

[0091] Step 2: Real-time capture and high-dimensional feature extraction of southbound physical traffic.

[0092] The edge gateway monitors on-site communication traffic through its southbound physical network port, using the port mirroring mechanism of the switching chip to obtain the physical frame data of the underlying packets. The gateway's internal stream processing engine strips away the Ethernet frame header, IP packet header, and TCP / UDP transport layer shell, completely eliminating routing interference noise from the network and transport layers, extracting only the core application layer packet payload binary byte stream (Raw Payload Bytes). Subsequently, the feature engineering component performs flow-level and packet-level spatiotemporal statistical analysis on the intercepted single packet stream, extracting an 80-dimensional high-dimensional statistical feature vector X (covering packet length mean, arrival interval variance, stream duration, etc.) that reflects the essential characteristics of the protocol, and performs feature space normalization using a pre-loaded standardizer (StandardScaler).

[0093] Step 3: Introduce open set protocol intelligent recognition and confidence assessment based on prompting learning.

[0094] The edge gateway transforms the normalized 80-dimensional feature vector X into an input embedding. At the very beginning of the input sequence, it dynamically concatenates the k active learnable soft prompt vectors from the current prompt vector library to construct a deep prompt input tensor: ; Among them, P k The soft cue vectors, ranging in size from a few hundred bytes to a few KB, are specific to various known long-tail protocols. `Efeature` is a dense embedding of the feature vector `X` after a linear mapping. `Xinput` is input into a base Transformer network with frozen parameters for pure forward inference, incurring no gradient overhead. The network output layer undergoes probability normalization using a Softmax activation function, outputting the probability distribution of the current input frame for each protocol category. The model extracts the maximum probability value from the probability distribution as the classification confidence score for that frame, representing the model's certainty regarding its identification conclusion.

[0095] Step 4: Adaptive mapping and zero-trust transparent security encapsulation mechanism based on dual-threshold routing.

[0096] The system pre-sets a hard industrial control safety confidence threshold T. h (Recommended setting: 0.85). Based on the confidence score calculated in step three, the gateway automatically drives the dual-branch self-locking routing mechanism: Branch A (Deterministic Semantic Mapping Routing): If the confidence score is greater than or equal to the security threshold Th, the system determines that the current message belongs to a known power protocol. The gateway immediately calls the corresponding local deterministic semantic parsing dictionary based on the protocol type code corresponding to the identification conclusion. According to the specified byte offset and decoding logic, the actual telemetry, teleindication, and remote control fields (such as the current three-phase voltage value, circuit breaker opening and closing status, active power, etc.) in the message are accurately extracted, and these discrete and heterogeneous fields are adaptively mapped and filled into the standard IoT middleware JSON object model that conforms to the State Grid SG-CIM (Public Information Model). Finally, it is sent to the integrated demonstration platform through the northbound network in the MQTT publish-subscribe mode.

[0097] Branch B (Zero-Trust Transparent Secure Encapsulation Routing): If the confidence score is strictly less than the security threshold Th, the system activates the highest level of industrial security "self-locking anti-misinterpretation mechanism." At this time, the gateway abandons all semantic guessing and end-to-end translation, forcibly prohibiting any structured decomposition of the message. The gateway directly extracts the raw binary byte stream of the application layer message payload and fully converts it into an unstructured secure text string using the Base64 encoding algorithm. Subsequently, the Base64 string, the current packet capture timestamp, the device's physical MAC address, and the southbound port number are encapsulated into a generalized, semantically risk-free secure JSON envelope, labeled 'Unknown', and directly transmitted to the cloud-integrated verification platform via the northbound network, fundamentally eliminating the disaster of mis-control commands caused by AI model illusions.

[0098] Step 5: Cloud-based high-performance computing with small sample sizes provides continuous prompts for fine-tuning and learning.

[0099] The cloud-integrated empirical platform continuously receives transparently encapsulated envelopes with 'Unknown' tags reported from massive amounts of distribution network edge gateways, automatically categorizing and aggregating them into an unknown protocol sample pool. When the sample set of a specific unknown tag accumulates to a small sample threshold, it triggers the cloud computing center to perform deep reverse analysis and incremental learning. 1. Maintain a Transformer encoder base network that is completely consistent with the edge gateway and freeze its global parameters completely.

[0100] 2. On the cloud, for this small batch of unknown data streams only, initiate prompt-tuning backpropagation training. During the optimization process, the base does not undergo any parameter changes; only the gradient backpropagation update is performed on the newly created soft cue vector corresponding to the (k+1)th unknown protocol at the input. Training is guided by a class-balanced focal loss function, which is highly robust to long-tail data, forcibly widening the classification decision boundary between the unknown protocol and existing protocols in the latent feature space.

[0101] 3. After training converges, a brand new protocol-specific soft hint vector P(k+1) with a very small parameter space (usually only a few hundred bytes to a few KB) is generated. Cloud-based business personnel or large models assist in generating standardized semantic parsing and mapping scripts for this new protocol.

[0102] Step Six: Closed Loop of Micro-Knowledge Hot Loading and Plug-and-Play.

[0103] The cloud-integrated demonstration platform packages the newly generated micro-prompt vector P_(k+1) and its corresponding parsing and mapping script, and pushes it to the associated intelligent converged terminal at the edge of the distribution network via over-the-air (OTA) technology through a private wireless network / 5G channel. Without shutting down, restarting system container services, or changing any base network structure parameters, the edge terminal executes a hot-loading mechanism, directly registering and appending the received micro-prompt vector to its local prompt vector library list. In subsequent cycles, the gateway instantly gains proactive awareness and seamless, high-precision structured transformation capabilities for this new protocol, completely completing the adaptive convergence closed loop.

[0104] Thus, this invention completely overcomes the industry's persistent problems of limited computing power and catastrophic forgetting at the edge. By globally freezing the Transformer base and training only in the cloud while distributing KB-level micro-hint vectors, this invention leaves heavy training in the cloud and lightweight inference at the edge. When updating to a new protocol at the edge, full retraining is not required, achieving near-zero growth in parameter size. This completely eliminates the influx of new knowledge that washes away and obliterates old protocol knowledge, making it extremely suitable for the ever-changing long-tail evolution environment of power distribution networks. It constructs an absolute physical defense line for industrial control network security. This invention breaks away from the traditional AI's mindset of 'blindly parsing and translating end-to-end,' pioneering a dual-threshold confidence level splitting approach. For unknown or long-tail variant protocols that are uncertain, it adopts self-locking zero-trust pass-through, preferring to package and send the Base64 data intact rather than guessing blindly. At the mechanism level, it completely shields the inherent 'illusion' mistranslation risk of deep learning, safeguarding the safety line against malfunctions in power grid control systems. It achieves true dynamic 'plug-and-play' capability for diverse and heterogeneous edge devices. Relying on a cloud-edge collaborative digital twin network, new distributed photovoltaic, energy storage, and charging pile devices can be connected to the field without frequent manual offline code rewriting or firmware upgrades. Through small-sample extraction of cloud-based knowledge and online fine-tuning and hot-swappable deployment, it achieves high-efficiency, seamless, adaptive, and dynamic integration of distribution network services, resulting in extremely low overall engineering implementation costs.

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0106] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention; the electronic device 500 includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501; when the processor 501 executes the computer program 503, it implements the steps in the above method embodiments; or, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the above device embodiments; For example, computer program 503 may be divided into one or more modules / units, one or more modules / units may be stored in memory 502 and executed by processor 501 to complete the present invention; one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, the instruction segments being used to describe the execution process of computer program 503 in electronic device 500.

[0107] The processor 501 may be a central processing unit, or it may be other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.; the general-purpose processor may be a microprocessor or any conventional processor, etc. The memory 502 can be an internal storage unit of the electronic device 500, such as a hard disk or RAM of the electronic device 500; the memory 502 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., equipped on the electronic device 500; furthermore, the memory 502 can include both internal storage units and external storage devices of the electronic device 500. The memory 502 is used to store computer programs and other programs and data required by the terminal; the memory 502 can also be used to temporarily store data that has been output or will be output.

[0108] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A cloud-edge collaborative adaptive fusion method for power protocols, characterized in that, include: Detect the protocol message to be identified; Based on the protocol message to be identified and the protocol identification model pre-stored by the edge intelligent terminal, classification and identification and classification confidence scoring are performed to obtain the first identification result of the protocol message to be identified; the first identification result includes existing protocol types or new protocol types; If the identification result indicates a new protocol type, the protocol message to be identified will be sent to the cloud platform. The cloud platform receives the prompt word vector packet returned after receiving the protocol message to be identified; Based on the aforementioned prompt word vector package, the local prompt word vector library of the edge intelligent terminal is updated. Combined with the aforementioned protocol recognition model, the secondary recognition result of the protocol message to be recognized is obtained, thereby realizing the adaptive fusion of power protocols.

2. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 1, characterized in that, The process of classifying and scoring the protocol message to be identified based on the protocol message to be identified and the protocol identification model pre-stored by the edge intelligent terminal, to obtain a first identification result of the protocol message to be identified, includes: Extract the message payload of the protocol message to be identified; Perform flow-level and packet-level spatiotemporal statistical analysis on the message payload, extract fixed-length high-dimensional statistical feature vectors and perform normalization processing to obtain normalized high-dimensional feature vectors; Based on the normalized high-dimensional feature vector and the currently active known protocol prompt vector in the local prompt word vector library of the edge intelligent terminal, a deep prompt input sequence is constructed; Based on the deep cue input sequence and the forward inference performed by the protocol recognition model, a first classification confidence score of the protocol message to be identified is obtained; the protocol recognition model is a deep pre-trained encoder base network deployed on an edge intelligent terminal with all global weight parameters frozen. Based on the first classification confidence score of the protocol message to be identified, a determination is made to obtain the first identification result.

3. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 1, characterized in that, The prompt word vector package includes prompt word vectors corresponding to the unidentified protocol messages that each edge intelligent terminal determines to be of a new protocol type; The process of updating the local prompt word vector library of the edge intelligent terminal based on the prompt word vector packet, and analyzing the secondary recognition result of the protocol message to be identified by combining the protocol recognition model, includes: Based on the aforementioned prompt word vector package, the prompt word vector package is dynamically registered and appended to the local prompt word vector library through a hot loading mechanism, resulting in an updated local prompt word vector library; Based on the protocol message to be identified and the prompt word vectors in the updated local prompt word vector library, a new deep prompt input sequence is generated by concatenation; The new deep cue input sequence is input into the protocol identification model to perform forward inference, and a secondary classification confidence score of the protocol message to be identified is obtained. Based on the secondary classification confidence score and the parsing mapping script returned by the cloud platform, deterministic semantic parsing is performed on the packet payload of the protocol message to be identified to obtain the secondary identification result of the protocol message to be identified; the secondary identification result includes the protocol type and secondary classification confidence score of the protocol message to be identified.

4. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 1, characterized in that, After classifying and scoring the protocol message to be identified based on the protocol identification model pre-stored by the edge intelligent terminal, and obtaining a first identification result of the protocol message to be identified, the method further includes: If the identification result is an existing protocol type, then based on the deterministic semantic parsing rule tree pre-set locally by the edge intelligent terminal, combined with the protocol type of the protocol message to be identified, the byte offset and decoding logic are determined. Based on the byte offset and decoding logic, the telemetry field and teleindication field in the payload of the protocol message to be identified are extracted and mapped into a standardized object model.

5. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 1, characterized in that, Before performing classification and confidence scoring based on the protocol message to be identified and the protocol identification model pre-stored by the edge intelligent terminal to obtain a first identification result of the protocol message to be identified, the process further includes: A pruned and quantized compressed deep pre-trained encoder network is deployed as a base network in the edge intelligent terminal, and the global weight parameters of the base network are frozen in full, so that the base network only performs forward inference on the edge side and does not generate gradient calculation and backpropagation overhead. A local prompt word vector library is built locally on the edge intelligent terminal, and the local prompt word vector library stores soft prompt vectors corresponding to various existing power protocols; The edge intelligent terminal is locally configured with deterministic semantic parsing rule trees for each known power protocol.

6. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 1, characterized in that, The method for generating the prompt word vectors includes: The cloud platform continuously receives protocol messages to be identified from various edge intelligent terminals; Extract the payload characteristics of each protocol message to be identified; Based on the payload characteristics of each protocol message to be identified, cluster analysis is performed to obtain multiple candidate categories and multiple protocol messages to be identified for each candidate category; Based on the multiple candidate categories and the multiple protocol messages to be identified corresponding to each candidate category, and combined with the protocol identification model of the cloud platform, the prompt word vectors and parsing mapping scripts for each candidate category are obtained; the global weight parameters of the protocol identification model of the cloud platform and the protocol identification model of the edge intelligent terminal are the same. The prompt word vectors and parsing mapping scripts for each candidate category are sent back to the edge intelligent terminal.

7. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 1, characterized in that, If the identification result is a new protocol type, then the protocol message to be identified is sent to the cloud platform, including: If the identification result is a new protocol type, the anti-misinterpretation self-locking mechanism is triggered, and zero-trust transparent security encapsulation is performed on the protocol message to be identified. Byte-level semantic decoding of the message payload is forcibly prohibited. The original binary byte stream of the message payload is directly extracted and transcoded. The transcoded data is combined with the timestamp and source MAC address to package it into a transparent security envelope labeled with the unknown protocol type, and the transparent security envelope is sent to the cloud platform.

8. A cloud-edge collaborative adaptive fusion method for power protocols, characterized in that, Applied to a cloud platform, the method includes: Receives protocol messages to be identified from multiple edge intelligent terminals, wherein the protocol messages to be identified are sent after the protocol identification model in the edge intelligent terminal determines that the protocol is a new protocol type; Extract payload features from the protocol messages to be identified sent by each edge intelligent terminal; Cluster analysis is performed based on the payload characteristics of each protocol message to be identified to obtain multiple candidate categories and multiple protocol messages to be identified for each candidate category; Based on the multiple candidate categories and the multiple protocol messages to be identified corresponding to each candidate category, and combined with the protocol identification model of the cloud platform, the prompt word vectors and parsing mapping scripts for each candidate category are obtained; the global weight parameters of the protocol identification model of the cloud platform and the protocol identification model of the edge intelligent terminal are the same. The prompt word vectors and parsing mapping scripts for each candidate category are sent back to the edge intelligent terminal.

9. The adaptive fusion method for power protocols based on cloud-edge collaboration according to claim 8, characterized in that, Before obtaining the prompt word vectors and parsing mapping scripts for each candidate category based on the multiple candidate categories and the multiple protocol messages to be identified corresponding to each candidate category, combined with the protocol recognition model of the cloud platform, the process further includes: Determine the number of protocol messages to be identified in each candidate category; If the number of protocol messages to be identified in a candidate category is greater than or equal to the preset small sample threshold, then for that candidate category, while keeping the global weight parameters of the protocol recognition model completely frozen, a prompt fine-tuning training is performed to generate the prompt word vectors and parsing mapping scripts corresponding to that candidate category. If the number of protocol messages to be identified in a certain candidate category is less than the preset small sample threshold, the protocol messages to be identified in that candidate category reported by each edge intelligent terminal will be continuously aggregated, and the prompt fine-tuning training will not be initiated for the time being.

10. A cloud-edge collaborative adaptive fusion system for power protocols, characterized in that, The fusion system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 9.

Images