Distributed power multi-agent cooperative knowledge retrieval enhancement method

Abstract

The application discloses a distributed power multi-agent cooperative knowledge retrieval enhancement method, which extracts and binary encodes features of abnormal working condition data through a graph convolution network to generate a compact semantic abstract, embeds the compact semantic abstract into a protocol expansion header, and delivers the compact semantic abstract to each node with a multicast data packet, so that a receiving end hardware layer realizes line-speed pre-screening through exclusive or operation and Hamming distance quantization, only the node with a matching degree meeting a threshold value wakes up a host processor to perform knowledge graph retrieval and enhancement generation, and a differential alignment and incremental hash mechanism is used to guarantee long-term consistency of cache abstracts and node knowledge evolution. In this way, the invalid wake-up frequency of irrelevant nodes is significantly reduced, bandwidth occupation and calculation resource consumption caused by redundant message processing are reduced, and therefore the real-time performance and accuracy of knowledge matching are improved.

Description

Technical Field

[0001] This application relates to the field of power knowledge retrieval, and more specifically, to a distributed power multi-agent collaborative knowledge retrieval enhancement method. Background Technology

[0002] With the deepening evolution of the multi-interaction pattern of "source, grid, load, and storage" in the new power system, the large-scale deployment of distributed intelligent terminals in the distribution network has made rapid diagnosis and collaborative handling of abnormal operating conditions a core requirement for ensuring power supply reliability. While regional distribution terminal intelligent agents have accumulated a wealth of localized fault handling experience and control knowledge over long-term operation, the knowledge coverage of a single node is limited. Faced with cross-regional related faults or rare complex operating conditions, it is urgent to achieve experience sharing and enhanced decision-making through collaborative knowledge retrieval among multiple intelligent agents, thereby improving the overall situational awareness and emergency response capabilities of the distributed power system.

[0003] Existing distributed power multi-agent collaborative knowledge retrieval schemes typically employ centralized index servers or broadcast flooding-based query mechanisms. Centralized schemes aggregate knowledge summaries from all nodes to a central server for matching and routing. While offering a global retrieval perspective, they are limited by the communication bandwidth bottleneck of the central node and the risk of single-point failures, making it difficult to meet the real-time requirements of millisecond-level response latency in distribution network fault scenarios. Broadcast flooding schemes indiscriminately deliver query messages to all nodes in the network, allowing each node to determine whether it possesses the relevant knowledge and respond accordingly. While this approach avoids centralized dependence, in a distributed heterogeneous computing environment, all nodes must perform a complete semantic parsing and knowledge graph retrieval process for each query request. This results in retrieval requests being blindly broadcast to all nodes globally, causing a large amount of redundant message transmission in the communication link, consuming limited network bandwidth resources. Simultaneously, regardless of whether the queried node's own knowledge base is relevant to the request, it is forced to wake up the host processor to perform a complete semantic unpacking and knowledge graph retrieval process. This forces many nodes without relevant experience to consume valuable computing resources to process invalid requests, resulting in a serious waste of host computing resources. The aforementioned technical deficiencies are particularly prominent in scenarios involving large-scale node deployment in power distribution networks, severely restricting the communication efficiency and computing resource utilization of distributed power multi-agent collaborative knowledge retrieval systems.

[0004] Therefore, we look forward to an optimized knowledge retrieval enhancement method for distributed power multi-agent collaboration. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method for enhancing knowledge retrieval through distributed power multi-agent collaboration.

[0006] According to one aspect of this application, a knowledge retrieval enhancement method for distributed power multi-agent cooperation is provided, comprising: Step 1: Using a graph convolutional network, feature extraction and binarization encoding are performed on the abnormal operating condition measurement data stream of the power distribution terminal to obtain the semantic routing summary to be queried and the abnormal operating condition query message. Step 2: Obtain cross-domain semantic probe data packets by embedding the semantic route summary to be queried into the network protocol extension header and encapsulating the abnormal operation query message into a payload, performing data packet assembly and multicast delivery; Step 3: Perform hardware-level XOR operation and Hamming distance quantization on the header digest of the cross-domain semantic probe data packet and the entries in the node high-frequency knowledge digest cache pool. When the Hamming distance is lower than the adaptive matching threshold, a hardware-level wake-up interrupt signal is generated. Step 4: Based on the hardware-level wake-up interrupt signal, the host processor is triggered to unpack and restore the cross-domain semantic probe data packet and perform retrieval, enhancement and fusion in the knowledge graph running locally to obtain cross-domain collaborative control decision-making schemes and historical fault handling experience texts. Step 5: Differential alignment and incremental hashing are performed on the cross-domain collaborative control decision-making scheme and historical fault handling experience text to obtain the updated node high-frequency knowledge summary cache pool.

[0007] Compared with existing technologies, this application provides a distributed power multi-agent collaborative knowledge retrieval enhancement method. It uses a graph convolutional network to extract features from abnormal operating condition data and perform binarization encoding to generate a compact semantic summary. This summary is embedded in a protocol extension header and delivered to each node along with multicast data packets. The receiving hardware layer performs linear-speed pre-screening using XOR operations and Hamming distance quantization. Only nodes with a matching degree meeting a threshold are awakened by the host processor to perform knowledge graph retrieval and enhancement generation. Differential alignment and incremental hashing mechanisms ensure long-term consistency between the cached summary and the evolution of node knowledge. This approach significantly reduces the frequency of invalid awakenings of irrelevant nodes, lowers bandwidth consumption and computational resource consumption caused by redundant message processing, thereby improving the real-time performance and accuracy of knowledge matching. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of a knowledge retrieval enhancement method for distributed power multi-agent collaboration according to an embodiment of this application; Figure 2This is a schematic diagram of the data flow of the knowledge retrieval enhancement method for distributed power multi-agent collaboration according to an embodiment of this application; Figure 3 This is a flowchart of step two in the knowledge retrieval enhancement method for distributed power multi-agent collaboration according to an embodiment of this application; Figure 4 This is a flowchart of step five in the knowledge retrieval enhancement method for distributed power multi-agent collaboration according to an embodiment of this application. Detailed Implementation

[0010] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0011] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] In the technical solution of this application, a knowledge retrieval enhancement method for distributed power multi-agent collaboration is proposed. Figure 1 This is a flowchart of a knowledge retrieval enhancement method for distributed power multi-agent collaboration according to an embodiment of this application. Figure 2 This is a system architecture diagram of a distributed power multi-agent cooperative knowledge retrieval enhancement method according to an embodiment of this application. Figure 1 and Figure 2As shown, the distributed power multi-agent collaborative knowledge retrieval enhancement method according to an embodiment of this application includes the following steps: S1, using a graph convolutional network, performing feature extraction and binarization encoding on the abnormal operating condition measurement data stream of the distribution terminal to obtain a semantic routing summary to be queried and an abnormal operating condition query message; S2, embedding the semantic routing summary to be queried into a network protocol extended header and encapsulating the abnormal operating condition query message into a payload for data packet assembly and multicast delivery to obtain a cross-domain semantic probe data packet; S3, performing hardware-level XOR operation and Hamming distance quantization on the header summary of the cross-domain semantic probe data packet and the entries in the node high-frequency knowledge summary cache pool, wherein a hardware-level wake-up interrupt signal is generated when the Hamming distance is lower than the adaptive matching threshold; S4, triggering the host processor based on the hardware-level wake-up interrupt signal to unpack and restore the cross-domain semantic probe data packet and perform retrieval enhancement fusion in the locally running knowledge graph to obtain a cross-domain collaborative control decision scheme and historical fault handling experience text; S5, performing differential alignment and incremental hash update on the cross-domain collaborative control decision scheme and historical fault handling experience text to obtain an updated node high-frequency knowledge summary cache pool.

[0015] Specifically, in step S1, a graph convolutional network is used to extract features and binarize the abnormal operating condition measurement data stream of the power distribution terminal to obtain a semantic routing summary and an abnormal operating condition query message. It should be understood that when an abnormal operating condition occurs in the power distribution terminal, a cross-domain knowledge query request needs to be initiated to other nodes in the network with relevant handling experience. However, the abnormal operating condition measurement data stream of the power distribution terminal itself consists of high-dimensional continuous time-series signals of various physical modes, including three-phase transient overvoltage amplitude sequences, transformer winding hotspot temperature values, and circuit breaker non-full-phase operation switching pulses. These raw data are characterized by varying sampling rates, inconsistent data formats, high dimensionality, and large data volume. If these raw multimodal data are directly transmitted as query conditions in the distributed network and fully parsed and matched by each node, it will consume a large amount of communication bandwidth resources. Furthermore, all receiving nodes need to wake up the host processor to perform a complete semantic understanding and knowledge graph retrieval process on the raw data, resulting in the ineffective consumption of computing resources of a large number of irrelevant nodes. Therefore, in the technical solution of this application, feature extraction and binarization encoding of the abnormal operating condition measurement data stream of the power distribution terminal are performed through a graph convolutional network. On the one hand, a compact fixed-length binary semantic route summary to be queried is generated. This summary integrates multimodal features and equipment topology association information, and can be embedded in the network protocol extension header for the receiving end data processing unit to perform low-power Hamming distance pre-screening at the hardware level. On the other hand, a structured abnormal operating condition query message is generated, carrying complete fault service semantic information, which serves as the input basis for the awakened node to perform precise knowledge graph retrieval. Through this step, the high-dimensional continuous multimodal signal is transformed into a formatted output that adapts to the two parallel data paths of subsequent hardware-level route screening and software-level precise retrieval, laying the data foundation for the entire collaborative knowledge retrieval enhancement process.

[0016] In practice, firstly, feature extraction and channel-wise splicing and fusion are performed on each modal sub-signal in the abnormal operating condition measurement data stream of the power distribution terminal to obtain a multimodal fusion feature vector. The abnormal operating condition measurement data stream of the power distribution terminal includes a three-phase transient overvoltage amplitude sequence, transformer winding hot spot temperature values, and circuit breaker non-full-phase operation switching pulses. In this process, the abnormal operating condition measurement data stream of the power distribution terminal is first separated according to physical mode categories to obtain three independent modal sub-signals: the three-phase transient overvoltage amplitude sequence, the transformer winding hot spot temperature values, and the circuit breaker non-full-phase operation switching pulses. For the three-phase transient overvoltage amplitude sequence, this modal sub-signal is a high-sampling-rate continuous time-series waveform data, recording the instantaneous amplitude changes of phases A, B, and C during the transient event. A one-dimensional convolutional neural network is used as the feature extractor for this mode. Multiple layers of one-dimensional convolutional kernels are used to perform sliding window convolution operations in the time dimension, extracting multi-scale features from local waveform details to global time-series patterns layer by layer. After downsampling in the time dimension by pooling layers, a fixed-length voltage transient feature vector is output. For the transformer winding hot spot temperature values, this mode sub-signal is a scalar numerical time-series data with a relatively low sampling rate, reflecting the temperature rise process of the highest temperature point of the transformer winding over time. A multi-layer fully connected network is used as the feature extractor for this mode, mapping the temperature sampling numerical sequence into a fixed-length temperature state feature vector after layer-by-layer linear transformation and nonlinear activation. For the circuit breaker non-full-phase operation switching pulses, this mode sub-signal is a discrete binary event sequence, recording the occurrence time, duration, and phase identifier of the circuit breaker three-phase asynchronous operation event. Since the signal is discrete event rather than continuous time-series data, an event embedding coding network is used as the feature extractor for this mode. First, the attribute information of each switching pulse event (including the event occurrence time, pulse duration width and asynchronous phase identification) is mapped into a fixed-dimensional event embedding vector through the embedding layer. Then, the embedding vectors of all events within the current time window are subjected to linear transformation and nonlinear activation, and then time-series aggregation operation (such as mean pooling or maximum pooling) is performed. Finally, a fixed-length switching event feature vector is output.

[0017] After extracting the independent features of the three modal sub-signals, the voltage transient feature vector, temperature state feature vector, and switching event feature vector are spliced ​​and fused according to the channel dimension to obtain the multimodal fused feature vector.

[0018] Next, based on the local device topology adjacency matrix, graph convolution aggregation and low-rank mapping are performed on the multimodal fusion feature vectors to obtain the semantic route summary to be queried. In this process, firstly, a local device topology adjacency matrix is ​​constructed according to the physical electrical connection relationships between devices in the local distribution network where the current distribution terminal is located. The multimodal fusion feature vector of the distribution terminal node currently experiencing an abnormal operating condition is used as the initial input feature of that node in the graph structure. Simultaneously, the feature vectors corresponding to the most recent operating states of its topological neighboring device nodes are used as the input features of the neighboring nodes, together forming the node feature input matrix of the graph convolutional network. Then, self-loops are added to the adjacency matrix to ensure that each node retains its own information during the aggregation process. Then, symmetric normalization is used to eliminate the scale effect caused by differences in degree between different nodes. Finally, the normalized adjacency matrix is ​​multiplied by the node feature matrix and processed by linear transformation of the trainable weight matrix and nonlinear activation function. Specifically, for each node in the graph, the current layer feature vectors of all its direct neighboring nodes (including itself) are first collected, weighted and aggregated according to symmetric normalization weights, and then processed by linear transformation and nonlinear activation to obtain the node feature representation of the next layer. After layer-by-layer aggregation through multi-level graph convolution operations, the output features of each node have incorporated topological association information within its multi-hop neighborhood. The feature vector corresponding to the current abnormal operating condition node in the output of the last layer of graph convolution is extracted. This vector integrates the multimodal features of the abnormal operating condition of this node with the contextual information of the electrical topology neighborhood. Subsequently, a low-rank mapping operation is performed on the high-dimensional node feature vectors output by graph convolution aggregation, compressing them from the high-dimensional feature space to a low-dimensional target space suitable for binary encoding. Low-rank mapping achieves dimensionality reduction through a linear projection matrix with a dimension much smaller than the input dimension, preserving the main semantic discriminative information while compressing the feature dimension to a dimension consistent with the target binary summary bit width, thus providing a continuous real-valued input of appropriate dimension for subsequent binary encoding. Finally, a binary encoding operation is performed on the continuous real-valued vector after low-rank mapping, quantizing the continuous real-valued values ​​of each dimension into binary bit values, thereby obtaining a fixed-length binary form of the semantic route summary to be queried. Binary encoding uses a sign function to determine zero-point thresholds for each dimension, encoding dimensions greater than or equal to zero as bit values ​​of 1, and dimensions less than zero as bit values ​​of 0. The calculation formula is as follows:

[0019] in, A continuous real-valued vector The Each dimension component For the output of the binary digest, the first 1 bit This is the semantic route summary to be queried, which has a length of [length missing]. A fixed-length binary string of bits encodes the multimodal semantic features of the current abnormal operating condition and its associated contextual information in the local power distribution topology in an extremely compact data form. This binary digest will be embedded in the network protocol extension header in subsequent steps, allowing the receiving end data processing unit to perform line-rate Hamming distance matching using hardware-level XOR logic gates.

[0020] Furthermore, semantic translation and structured encapsulation are performed on the extreme values ​​and physical attributes of the abnormal operating condition measurement data stream of the power distribution terminal to obtain the abnormal operating condition query message. In this process, firstly, extreme value detection and extreme value parameter extraction are performed on each modal sub-signal in the abnormal operating condition measurement data stream of the power distribution terminal. For the three-phase transient overvoltage amplitude sequence, the voltage waveforms of phases A, B, and C are judged for exceeding limits based on a preset voltage over-limit threshold, and extreme value parameters such as transient peak amplitude (i.e., the maximum instantaneous voltage value exceeding the normal operating range), overvoltage duration (i.e., the time span during which the voltage amplitude continuously exceeds the threshold), and maximum rate of change of voltage rise edge (i.e., the maximum slope of the voltage from the normal value to the peak value) are extracted for each phase voltage. For the transformer winding hot spot temperature value, the over-limit judgment is performed based on the transformer's rated temperature rise limit, and parameters such as the highest temperature value (i.e., the peak value of the winding hot spot temperature within the current time window), temperature rise rate (i.e., the average rate of temperature increase over time), and over-temperature duration (i.e., the continuous time span during which the temperature exceeds the rated limit) are extracted for the over-limit parameters. For circuit breaker non-full-phase operation switching pulses, discrete event attribute parameters such as phase identifier (i.e., which phase or phases experienced asynchronous operation), non-full-phase duration (i.e., the duration of the three-phase asynchronous state), and action sequence difference (i.e., the time interval between the first and second operating phases) are extracted. Subsequently, all extracted limit-extremum parameters and physical attribute information undergo semantic translation processing to convert the numerical physical quantity measurement results into a structured description with semantic meaning in the power business domain, enabling it to be understood and utilized by subsequent knowledge graph retrieval systems. Specifically, the limit-extremum values ​​of each mode are associated and mapped with their physical meaning, equipment attributes, and fault type classification, forming key-value pairs or structured text descriptions with clear semantic annotations. For example, "A-phase voltage peak 35.2kV" is translated into a structured field containing semantic tags such as equipment type, fault mode, and severity, while supplementing the annotation with equipment terminal equipment number, feeder identifier, substation affiliation, voltage level, and other equipment physical attribute information, as well as the precise timestamp of the fault event and the business processing priority tag. Finally, all the key information after semantic translation is structurally encapsulated according to a predefined query message format template to form a standardized abnormal operating condition query message.The message contains the following structured fields: Fault Type Identifier (indicating the fault category to which the current abnormal condition belongs, such as "lightning overvoltage," "transformer overheating," "circuit breaker incomplete phase operation," or combinations thereof), Over-Limit Parameter Value Field (recording all over-limit extreme values ​​extracted from each mode and their units and dimensions in the form of a structured array), Equipment Identifier Field (recording the location information such as the distribution terminal equipment number, feeder number, and substation number where the abnormality occurred), Timestamp Field (recording the precise time of the abnormal condition, accurate to the millisecond level), and Service Priority Field (a label indicating the urgency of processing based on the degree of over-limit and the scope of the fault impact). This structured abnormal condition query message carries the complete business semantic information of the current fault condition. In the subsequent step two, it will be serialized, encoded, and encapsulated into the payload of a data packet, ultimately used by the host processor of the awakened node to perform high-dimensional semantic vector retrieval and most similar entity matching extraction in the knowledge graph running locally.

[0021] Specifically, in step S2, a cross-domain semantic probe data packet is obtained by embedding the semantic routing summary to be queried into the network protocol extended header and encapsulating the abnormal operating condition query message into a payload for data packet assembly and multicast delivery. It should be understood that step one has already completed the feature extraction and binarization encoding of the abnormal operating condition measurement data stream of the distribution terminal, obtaining two output results: the semantic routing summary to be queried and the abnormal operating condition query message. The semantic routing summary to be queried is a compact fixed-length binary string used for fast matching and pre-screening at the receiving end hardware level; the abnormal operating condition query message is structured data carrying complete fault service semantics, used by the awakened node to perform accurate knowledge graph retrieval. However, these two data outputs are still in the form of local application layer data and cannot be directly transmitted across domains to other intelligent agent nodes in the distributed distribution network. To achieve cross-domain knowledge query network delivery, these two data need to be encapsulated and assembled according to the specifications of the network communication protocol, so that they can be transmitted to intelligent agent nodes in multiple target domains through the distribution communication network. Specifically, the semantic routing summary to be queried needs to be embedded into the extended header of the network protocol. This allows the data processing unit at the receiving end to extract the summary directly from the header at the network interface card (NIC) hardware layer for line-rate matching, without sending the entire data packet to the host processor for software-level parsing. Simultaneously, the abnormal operation query message needs to be encapsulated as the payload of the data packet. After hardware pre-screening, this payload is then unpacked and extracted by the awakened host processor for knowledge retrieval. Furthermore, since the target node for cross-domain knowledge queries is unknown beforehand, multicast delivery is required to simultaneously send the data packet to multiple potential knowledge-providing nodes. Each node's hardware-level matching mechanism then autonomously determines whether a response is needed.

[0022] Figure 3This is a flowchart of step two in the knowledge retrieval enhancement method for distributed power multi-agent cooperation according to an embodiment of this application. Figure 3 As shown, S2 includes: S21, performing serialization encoding and compression on the abnormal working condition query message, and concatenating the timestamp and service priority tag to obtain a custom data payload; S22, initializing the hop-by-hop option extension header and writing the semantic probe-specific option type field and bit width length field, and embedding and aligning the header data fields of the semantic route digest to be queried to obtain the extended header with embedded digest; S23, constructing a basic header containing the multicast target address, and concatenating the basic header, the extended header with embedded digest, and the custom data payload according to the protocol field order and appending a frame verification sequence, and obtaining a cross-domain semantic probe data packet after multicast delivery.

[0023] Specifically, in step S21, the abnormal operating condition query message is serialized, encoded, and compressed, and then a custom data payload is obtained by concatenating a timestamp and a service priority tag. In this process, firstly, a serialization encoding operation is performed on the abnormal operating condition query message. The abnormal operating condition query message exists in the form of structured data in step one, containing multiple hierarchical data fields such as a fault type identifier field, an out-of-limit parameter value field, an equipment identifier field, a timestamp field, and a service priority field. Serialization encoding converts this structured data object into a continuous byte stream sequence according to predefined encoding rules. Specifically, according to the field definition order, the field identifier, data type tag, and field value of each field are sequentially encoded into a byte sequence and concatenated sequentially. For numeric fields, a fixed-width binary encoding is used; for variable-length string fields, a length prefix plus content bytes encoding method is used, ultimately generating a complete serialized byte stream. Subsequently, a compression operation is performed on the serialized byte stream. Specifically, lossless data compression algorithms (such as LZ4 fast compression or Zstandard compression) are used to encode and compress the serialized byte stream, utilizing repetitive patterns and statistical redundancy in the data to reduce data volume and generate a compressed byte stream. Then, a timestamp field and a business priority tag field are appended to the compressed byte stream to form a complete custom data payload.

[0024] Specifically, in step S22, the semantic probe-specific option extension header is initialized and the semantic probe-specific option type field and bit width length field are written. The header data field of the semantic routing digest to be queried is then embedded and aligned to obtain the extended header of the embedded digest. In this process, firstly, the hop-by-hop option extension header initialization operation is performed. The initialization process allocates a contiguous memory buffer to store the extension header data and writes two fixed preamble fields to the extension header according to the protocol specification: the next header field (occupying 1 byte) and the header extension length field (occupying 1 byte). The value of the next header field is set to a protocol type identifier pointing to the subsequent custom data payload (if the payload is directly used as upper-layer data, the corresponding upper-layer protocol number is filled in; if there is no subsequent extension header, the payload type identifier value is filled in). The value of the header extension length field is calculated based on the total length of the subsequently filled option data, representing the remaining length of the extension header excluding the first 8 bytes in 8-byte units. Subsequently, the semantic probe-specific option type field and bit width length field are written into the initialized extension header. The semantic probe-specific option type field is a 1-byte unsigned integer, whose value is a pre-allocated custom option type identifier code, used to uniquely identify that the data content carried in the current hop-by-hop option is a semantic probe digest. The high 2 bits of this option type field are encoded according to the protocol specification as an action indicator bit (indicating the processing strategy a node should take when it cannot recognize the option, such as skipping and continuing to process subsequent options), the 3rd bit is encoded as a data change indicator bit (indicating whether the option data is allowed to be modified by intermediate nodes during forwarding), and the low 5 bits are encoded as the specific option number. The bit width length field is a 1-byte unsigned integer, immediately following the option type field, whose value is the binary bit width length of the semantic route digest to be queried (i.e., the total number of bits in the digest), used to inform the receiving end's data processing unit how many bits should be extracted from subsequent data fields as the semantic digest. Next, the header data field embedding and alignment operation is performed on the semantic route digest to be queried. Specifically, first, the number of bytes occupied by the semantic digest is calculated. Then, this number of bytes is added to the 4 bytes occupied by the preamble fixed field to obtain the total number of bytes of existing content. Next, the remainder when this total number of bytes is divided by 8 is calculated. If the remainder is zero, no padding is needed. If the remainder is not zero, the number of padding bytes is equal to 8 minus the remainder, thus ensuring that the total length of the extended header is exactly an integer multiple of 8 bytes. After alignment and padding are completed, the value of the header extension length field is updated to the result of the total length of the extended header divided by 8 and then minus 1 (the length excluding the first 8 bytes is represented in 8-byte units according to the protocol specification). At this point, the extended header with embedded digest is obtained, and its byte structure is as follows: next header field (1 byte), header extension length field (1 byte), semantic probe-specific option type field (1 byte), bit width length field (1 byte), semantic routing digest data to be queried (the number of bytes obtained by dividing the digest bit number by 8 and rounding up), and zero-value padding bytes (the number of padding bytes calculated according to alignment).

[0025] Specifically, in step S23, a basic header containing the multicast target address is constructed. The basic header, the extended header with embedded digest, and the custom data payload are concatenated according to the protocol field order, and a frame verification sequence is appended. After multicast delivery, a cross-domain semantic probe data packet is obtained. In this process, firstly, the network layer basic header is constructed. The basic header is a fixed 40-byte data structure (taking IPv6 as an example). The following fields are filled in sequentially according to the protocol specification: version number field (4 bits, filled with the protocol version identifier value 6), traffic category field (8 bits, setting the corresponding service quality flag value according to service priority), flow label field (20 bits, filled with a unique label value to identify the query flow so that network devices can perform flow-level forwarding processing), payload length field (16 bits, its value is calculated by adding the length of the extended header with embedded digest to the length of the custom data payload), and next header field (8 bits, filled with the hop-by-hop option extended header). The header consists of a protocol type identifier value of 0, indicating that the basic header is followed by a hop-by-hop options extension header; a hop count limit field (8 bits, specifying the maximum number of hops a packet is allowed to traverse in the network; packets exceeding this limit will be discarded to prevent infinite loop forwarding); a source address field (128 bits, specifying the network layer address of the distribution terminal node initiating the query); and a destination address field (128 bits, specifying the pre-assigned multicast destination address, corresponding to a multicast group joined by all potential knowledge-providing nodes in the distributed distribution network, enabling the packet to be copied and forwarded by network devices to all member nodes of that multicast group). Subsequently, the basic header, the extended header with embedded digest, and the custom data payload are concatenated according to the network protocol's field hierarchy. Next, a frame check sequence is appended to the end of the concatenated packet body. The frame check sequence calculates a fixed 4-byte (32-bit) checksum by performing a cyclic redundancy check algorithm (such as CRC-32) on all bytes of the packet body. A frame check sequence (FPS) is appended to the end of the data packet body. The receiving end re-performs the same cyclic redundancy check (CRC) calculation on the received data packet body and compares the result with the appended FPS. If they match, the data packet is confirmed to have no bit errors during network transmission and is intact and usable. If they do not match, a transmission error such as bit flipping has occurred during transmission, and the data packet is discarded. Finally, the complete data packet with the appended FPS is sent to the power distribution communication network via multicast delivery. Specifically, the sending end's network interface submits the complete data packet to the local network device's transmission queue. The local network device queries the multicast routing table based on the multicast destination address in the data packet's base header to determine the downstream forwarding interface list for the multicast group, copies the data packet, and forwards it from the corresponding physical interface.Each intermediate forwarding node the data packet passes through in the network also replicates and forwards it according to the multicast routing table until the data packet is delivered to the network interface of all target domain agent nodes that have joined the multicast group. After being delivered by multicast to the network interface of each target node, the data packet becomes a cross-domain semantic probe data packet, waiting for the data processing unit of each target node to intercept and extract the semantic digest in its extended header and perform Hamming distance matching at the network card hardware layer.

[0026] Specifically, in S3, a hardware-level XOR operation and Hamming distance quantization are performed on the header digest of the cross-domain semantic probe data packet and the entries in the node's high-frequency knowledge digest cache pool. A hardware-level wake-up interrupt signal is generated when the Hamming distance is lower than the adaptive matching threshold. It should be understood that the cross-domain semantic probe data packet is delivered to the network interfaces of all target domain agent nodes that have joined the multicast group in the distributed power distribution network via a multicast routing mechanism. However, not all nodes receiving this data packet possess the experience or knowledge to handle or regulate the current abnormal operating condition. In large-scale power distribution network deployment scenarios, a multicast group may contain dozens or even hundreds of agent nodes, of which only a few nodes store fault handling experience highly relevant to the current query in their local knowledge graph. If all nodes receiving the data packet wake up the host processor to execute the complete protocol stack unpacking, semantic parsing, and knowledge graph retrieval process, it will cause a large number of nodes without relevant knowledge to consume valuable computing resources to process invalid requests, resulting in a serious waste of the host processor's computing power and increasing the network burden caused by unnecessary response message return transmissions. Therefore, before the host processor is woken up, the received cross-domain semantic probe data packets are quickly relevance predicted at a lower hardware level. Only query requests that are highly relevant to the knowledge base of this node are triggered to complete the host processor's processing flow, while irrelevant query requests are silently discarded at the hardware level.

[0027] In practice, the data processing unit first intercepts cross-domain semantic probe packets at the network interface card (NIC) hardware layer. A bitmask extraction circuit then strips the binary bit string from the extended header according to a predefined offset address to obtain the probe source semantic digest. During this process, when the cross-domain semantic probe packet arrives at the receiving node through the physical network interface, the data processing unit's network interface controller first performs physical layer decoding and link layer processing on the inbound data frame, verifying the correctness of the frame check sequence to confirm error-free data transmission. After the frame check sequence verification is successful, the data processing unit's packet classification engine quickly parses the basic header of the packet, checking if the value of the next header field indicates the presence of a hop-by-hop option extension header (value 0). When a hop-by-hop option extension header is detected, the packet classification engine further reads the option type field in the extension header to determine if it is a semantic probe-specific option type identifier. When the option type field matches the semantic probe-specific identifier, the data processing unit confirms that the current packet is a cross-domain semantic probe packet, intercepts it from the normal protocol stack processing path, and transfers it to a dedicated hardware matching processing pipeline.

[0028] After interception is confirmed, the data processing unit activates the bitmask extraction circuit to precisely extract the binary bit string of the semantic digest from the extended header according to the predefined offset address. The logic for determining the predefined offset address is as follows: the basic header occupies a fixed 40 bytes, followed by the hop-by-hop options extended header. The field arrangement order within the extended header is: next header field (offset 0 bytes, 1 byte), header extension length field (offset 1 byte, 1 byte), semantic probe-specific option type field (offset 2 bytes, 1 byte), and bit width length field (offset 3 bytes, 1 byte). Therefore, the starting position of the semantic digest data is the 4th byte after the end of the basic header (i.e., the 44th byte calculated from the start position of the data packet), and the length of the digest data is specified by the value of the bit width length field. The bitmask extraction circuit performs bit string extraction on the continuous byte region in the extended header starting from the offset address based on this offset address (44th byte) and the bit width length (the value read from the bit width length field, for example, 64 bits, i.e., 8 bytes). The extraction calculation principle is as follows: The circuit loads the corresponding byte region of the data packet into an internal shift register, and then performs a bitwise AND operation with the data in the shift register using a pre-configured bitmask mode (the position corresponding to the digest data bit in the mask is set to 1, and the rest are set to 0). The result of the AND operation retains only the value of the digest data bit while masking all other irrelevant bits, thereby accurately extracting the binary bit string of the probe source semantic digest and latching it into a dedicated digest register. Thus, the probe source semantic digest stored in the on-chip register of the data processing unit is obtained. This digest is a fixed-length binary bit string that is exactly the same as the semantic routing digest to be queried generated by the sending end, and is used for subsequent Hamming distance matching with entries in the knowledge digest cache pool of this node.

[0029] Next, based on the on-chip XOR logic gate array and adder tree, fully parallel bitwise XOR and group number statistics operations are performed on each entry in the probe source semantic summary and node high-frequency knowledge summary cache pool. The optimal matching Hamming distance value is obtained after minimum value filtering. In this process, firstly, the data processing unit reads all entries from the node high-frequency knowledge summary cache pool from the on-chip hardware solid-state storage. Assume that the cache pool stores M knowledge summary entries, each entry being a fixed-length binary string with the same bit width as the probe source semantic summary. The data processing unit simultaneously loads the M cache entries into the input registers of the M groups of parallel processing units, and simultaneously broadcasts the probe source semantic summary to the other input port of all M groups of processing units, forming a one-to-many parallel matching operation structure. Then, the M groups of processing units simultaneously start fully parallel bitwise XOR operations. Each group of processing units contains... Each of the following is a separate XOR logic gate, connecting the i-th bit of the probe source semantic digest to the i-th bit of the corresponding cache entry. Each XOR gate performs its operation simultaneously within the same clock cycle. For each pair of corresponding bits, the XOR gate outputs 1 when the two input bit values ​​are different (one is 0 and the other is 1, or one is 1 and the other is 0), and outputs 0 when the two input bit values ​​are the same (both are 0 or both are 1). M groups of arithmetic units execute this operation independently and in parallel, with each group outputting one... The XOR result string of the bits is equal to the Hamming distance between the probe source semantic digest and the entry. Subsequently, a group number calculation operation is performed on each XOR result string to obtain a specific Hamming distance value. The adder trees of M operation units perform group number calculations in parallel, resulting in M ​​Hamming distance values, each representing the semantic difference between the probe source semantic digest and the M entries in the cache pool. Then, a minimum value filtering operation is performed on the M Hamming distance values ​​to obtain the optimal matching Hamming distance value. The optimal matching Hamming distance value is latched into a dedicated result register for threshold comparison in subsequent sub-steps.

[0030] Then, a hardware digital comparator compares the optimal matching Hamming distance value with the adaptive matching threshold stored in the register using high and low level comparisons. When the distance value is lower than the adaptive matching threshold, the interrupt controller is driven to generate a hardware-level wake-up interrupt signal. In this process, firstly, the data processing unit reads the currently stored adaptive matching threshold from the on-chip configuration register. This threshold is dynamically updated and written by the host processor during system operation based on the evolution of the knowledge base, historical matching hit rates, and business requirements through register write operations. Once written, it is stored in the hardware register for use by all subsequent arriving data packets until it is updated and overwritten by the host processor. Subsequently, the hardware digital comparator compares the optimal matching Hamming distance value with the adaptive matching threshold. Input A of the digital comparator is connected to the result register storing the optimal matching Hamming distance value, and input B is connected to the configuration register storing the adaptive matching threshold. Within a single clock cycle, the comparator determines the relationship between A and B through cascaded comparison logic from high to low bits, and generates the corresponding logic level signal at the output. Specifically, if the optimal matching Hamming distance is strictly less than the adaptive matching threshold (i.e., A is less than B), the comparator's "less than" output is set to a high level, indicating that there is at least one sufficiently similar entry in the probe source semantic digest and the local node's knowledge digest cache pool, and the local node is highly likely to have experience in handling remote abnormal conditions. If the optimal matching Hamming distance is greater than or equal to the adaptive matching threshold (i.e., A is greater than or equal to B), the comparator's "less than" output remains low, indicating that there is no knowledge in the local node's knowledge base that is sufficiently relevant to remote abnormal conditions, and there is no need to wake up the host processor. Furthermore, the output level of the comparator determines whether to generate a hardware-level wake-up interrupt signal. The comparator's "less than" output is directly connected to the interrupt request input pin of the interrupt controller inside the data processing unit. When this output is set to a high level (i.e., the optimal matching Hamming distance is lower than the adaptive matching threshold), the interrupt controller detects a valid interrupt request input. After internal interrupt arbitration and priority queuing logic, it generates an edge-triggered pulse signal that transitions from low to high on its interrupt output pin, i.e., a hardware-level wake-up interrupt signal. The signal is transmitted to the interrupt input pin of the host processor through a dedicated interrupt request line. After the host processor detects the rising edge of the interrupt signal, it immediately interrupts the currently executing low-priority task, saves the current execution context, and jumps to the preset high-priority interrupt service routine entry point to start the process of unpacking and restoring the cross-domain semantic probe data packet and enhancing the fusion of knowledge graph retrieval in step four.When the "less than" output of the comparator remains in a low-level invalid state (i.e., the optimal matching Hamming distance value is greater than or equal to the adaptive matching threshold), the interrupt controller does not generate an interrupt signal, the host processor is not woken up, the cross-domain semantic probe data packet is silently discarded inside the data processing unit, and this node does not generate any response to the query request, thereby realizing zero-wake-up hardware-level filtering of irrelevant queries.

[0031] Specifically, in step S4, the host processor is triggered by a hardware-level wake-up interrupt signal to unpack and restore the cross-domain semantic probe data packet and perform retrieval enhancement fusion in the local knowledge graph to obtain cross-domain collaborative control decision-making schemes and historical fault handling experience texts. It should be understood that step three has already completed the semantic relevance prediction of the cross-domain semantic probe data packet at the hardware level of the data processing unit. When the optimal matching Hamming distance value is lower than the adaptive matching threshold, a hardware-level wake-up interrupt signal is generated, indicating that the knowledge base of this node is highly likely to store handling experience related to abnormal operating conditions of the remote power distribution terminal. However, the hardware-level matching in step three is only a coarse-grained semantic similarity judgment based on binary summaries. Its role is to filter out obviously irrelevant nodes to save host computing resources. The actual knowledge retrieval, experience matching, and decision-making scheme generation require the host processor to perform a more refined and complete semantic understanding and reasoning process at the software level. Therefore, after the hardware-level wake-up interrupt signal is triggered, the data packet is first restored from the network protocol encapsulation form to the original abnormal working condition query message. Then, based on the complete business semantic information of the message, a precise semantic retrieval is performed in the local knowledge graph to find highly relevant historical fault handling experience. Finally, the retrieved experience is fused with the original query information and a targeted cross-domain collaborative control decision-making scheme is generated through a large language model.

[0032] In practice, firstly, a high-priority interrupt response from the host processor is triggered based on a hardware-level wake-up interrupt signal. This involves stripping the protocol stack header of the cross-domain semantic probe data packet, reverse decompressing and deserializing it to obtain the decompressed abnormal condition query message. During this process, when the hardware-level wake-up interrupt signal reaches the host processor's interrupt input pin, the host processor's interrupt controller detects the rising edge of the interrupt request. Based on the pre-configured interrupt priority table, it confirms that the interrupt is a high-priority interrupt (its priority is higher than the currently executing regular computing task), and then triggers the high-priority interrupt response sequence. Specifically, the current program counter value and the state of all general-purpose registers are first saved to the memory stack for later restoration. Then, the interrupt service routine entry address corresponding to the interrupt number is found in the interrupt vector table, the program counter is set to that entry address, and the interrupt service routine execution begins.

[0033] After the interrupt service routine starts, it first reads the complete byte content of the intercepted cross-domain semantic probe data packet from the receive buffer of the data processing unit into the host processor's memory space. Then, following the network protocol hierarchy from the outermost layer, it first identifies and removes the 40-byte basic header at the very beginning of the data packet (containing network layer routing fields such as version number, source address, and destination address; these fields are no longer needed in this step). Next, it identifies and removes the hop-by-hop options extension header immediately following the basic header (containing semantic probe-specific option type fields, bit width and length fields, embedded semantic digest data, and alignment padding bytes; these fields were used by the DPU for hardware-level matching in step three and are not needed again in this step). After the protocol stack header is stripped, the remaining data is the custom data payload, containing a compressed query message byte stream, an 8-byte timestamp field, and a 1-byte service priority label field.

[0034] The interrupt service routine then performs reverse decompression and deserialization on the custom data payload. Specifically, it first separates a 1-byte business priority tag and an 8-byte timestamp field from the end of the payload and performs timeliness verification—comparing the timestamp with the current system time. If the time since the query request was generated exceeds a preset valid time window (e.g., more than 30 seconds), the query request is deemed expired and processing is terminated; otherwise, processing continues. After the timeliness verification passes, the remaining compressed byte stream in the payload undergoes reverse decompression, i.e., it calls the decompression algorithm corresponding to the compression algorithm used in step two (e.g., LZ4 or Zstandard decompression algorithm) to restore the compressed byte stream to its original serialized byte stream. Finally, it performs deserialization on the restored serialized byte stream. Specifically, following the field definition order, the field identifiers, data type markers, and field values ​​of each field are read sequentially from the byte stream. The binary encoding in the byte stream is restored to the original data type and data value of each structured field. An abnormal operating condition query message data object is then reconstructed, containing all structured fields such as fault type identifier, out-of-limit parameter value, device identifier, timestamp, and service priority. This yields the decompressed abnormal operating condition query message, which is completely identical to the abnormal operating condition query message generated at the sending end, carrying complete fault service semantic information of the abnormal operating condition of the remote power distribution terminal.

[0035] Next, the decompressed abnormal operating condition query messages are subjected to high-dimensional semantic vector retrieval and most similar entity matching extraction within the locally running knowledge graph to obtain historical fault handling experience text. In this process, firstly, semantic embedding encoding is performed on the decompressed abnormal operating condition query messages, converting the structured query message content into a high-dimensional semantic vector representation to obtain the semantic embedding vector of the query message. Subsequently, nearest neighbor search is performed in the semantic vector index of the locally running knowledge graph. Each knowledge entity in the locally running knowledge graph (such as a record of handling experience for a historical fault) has its corresponding high-dimensional semantic vector pre-calculated and stored using the same semantic embedding model during database entry. Assume there are N knowledge entities in the knowledge graph, and their pre-stored semantic vector set is a matrix composed of the embedding vectors of each entity. Cosine similarity is calculated between the query vector and the pre-stored semantic vector of each knowledge entity in the knowledge graph. After calculating the cosine similarity for all N knowledge entities in the knowledge graph, the similarity values ​​are sorted in descending order, and the top K knowledge entities with the highest similarity are selected as the most similar entity matching results. Finally, the top K ranked matches are further processed with relevance filtering and redundancy removal. After removing highly semantically overlapping redundant entries, the detailed text content corresponding to the final retained matching entities is extracted. This text content is the historical fault handling experience text, which records the complete handling process, operation steps, effect verification, and other detailed information of the fault cases that are most semantically similar to the current remote abnormal operating conditions that have been handled by this node in the past.

[0036] Then, the decompressed abnormal operating condition query message and historical fault handling experience text are used to fill the retrieval enhancement to obtain the prompt word template. The prompt word template is then generated using an autoregressive conditional probability decoding method based on a large language model to obtain a cross-domain collaborative control decision-making scheme. In this process, firstly, the retrieval enhancement prompt word template is constructed. The prompt word template is a predefined text structure template containing multiple placeholder areas to be filled, organized in a logical order from contextual background to specific task instructions. The first part of the template is the system role instruction area, filled with a description of the large language model's role positioning (e.g., "You are a professional power system fault handling expert, skilled in formulating control decision-making schemes based on historical experience"). The second part of the template is the reference knowledge context area, where the historical fault handling experience text retrieved in sub-step two is filled, providing the model with domain-specific knowledge references. The third part of the template is the current task description area. All structured fields from the decompressed abnormal operating condition query message (including fault type identifier, modal limit parameter values, equipment identifier, timestamp, etc.) are filled into this area in natural language to fully describe the details of the abnormal operating condition faced by the remote power distribution terminal. The fourth part of the template is the output requirement instruction area, where specific requirements for the output format and content are filled in (e.g., "Based on the above reference experience, please generate a specific cross-domain collaborative control decision-making scheme for the current abnormal operating condition, including emergency handling steps, equipment operation instruction sequences, and subsequent monitoring suggestions"). Subsequently, the completed search enhancement prompt word template is input into the large language model deployed locally on this node, and autoregressive conditional probability decoding is performed to generate it. The large language model first performs word segmentation on the input prompt word template text, converting the text into a token sequence, and then gradually generates the output text according to the autoregressive generation mechanism. Specifically, let the word sequence after input prompt word template is segmented be the preceding condition sequence. At each generation time step t, the model calculates the conditional probability distribution of the current time step on the entire vocabulary through forward propagation, based on all the preceding condition words of the input and all the preceding output words that have been generated. Then, the word with the highest probability is selected from this probability distribution (greedy decoding) or the word is selected according to a specific sampling strategy as the output of the current step. The output word is appended to the generated sequence as the conditional input of the next step. This process is repeated word by word until a special end marker is generated or the preset maximum generation length is reached.

[0037] The output text is generated word-by-word according to the aforementioned autoregressive decoding mechanism until an end marker is generated. After generation, all output words are concatenated and decoded to restore natural language text, which is the cross-domain collaborative control decision-making scheme. This scheme integrates the general reasoning ability of the large language model based on its pre-trained knowledge with the domain-specific historical fault handling experience injected through retrieval enhancement, providing specific control operation suggestions, handling step sequences, and monitoring and recovery guidance for the abnormal operating conditions currently faced by the remote power distribution terminal.

[0038] Specifically, in step S5, the updated node high-frequency knowledge summary cache pool is obtained by differential alignment and incremental hashing of the cross-domain collaborative control decision-making scheme and historical fault handling experience text. It should be understood that step four has completed the unpacking and restoration of the cross-domain semantic probe data packet and the enhanced fusion of knowledge graph retrieval, generating two outputs: the cross-domain collaborative control decision-making scheme and the historical fault handling experience text. However, the node high-frequency knowledge summary cache pool, as the basis for hardware-level Hamming distance matching in step three, requires that the stored binary summary entries reflect the node's latest knowledge state. Each successful cross-domain knowledge collaborative interaction brings new knowledge increments to this node—the control execution feedback information contained in the cross-domain collaborative control decision-making scheme and the deviation evaluation information generated after the historical fault handling experience text is actually applied together constitute the incremental update content that the node's knowledge base needs to absorb. If the summary entries in the cache pool are not updated synchronously with the evolution of node knowledge, the cached summary will gradually deviate from the true semantic state of the node over time. This will cause subsequent cross-domain semantic probe data packets to fail to obtain correct matching judgments in step three—related nodes that should be awakened will be missed due to outdated summaries, and irrelevant nodes that should not be awakened will be falsely reported due to summary drift. Therefore, in the technical solution of this application, by evaluating the execution deviation of cross-domain collaborative control and performing differential alignment with historical experience, knowledge increment information is extracted. Then, the new knowledge is integrated into the binary summary in the cache pool through incremental hash encoding, thereby achieving continuous synchronous evolution between the node's semantic state of knowledge and the hardware cache matching basis.

[0039] Figure 4 This is a flowchart of step five in the knowledge retrieval enhancement method for distributed power multi-agent cooperation according to an embodiment of this application. Figure 4As shown, in the first embodiment of this application, step S5 includes: S51, parsing the cross-domain collaborative control decision scheme into equipment execution instructions and collecting the actual state feedback vector after execution, and using a dimensional weight matrix to weighted normalize the error between the expected and actual states to obtain a control execution deviation evaluation matrix; S52, performing text feature embedding and differential alignment residual extraction on the control execution deviation evaluation matrix and historical fault handling experience text to obtain a knowledge increment residual feature vector; S53, performing feature increment hashing and cache pool overlay update on the knowledge increment residual feature vector to obtain an updated node high-frequency knowledge summary cache pool.

[0040] Specifically, in step S51, the cross-domain collaborative control decision-making scheme is parsed into equipment execution instructions, and the actual state feedback vector after execution is collected. A dimensional weight matrix is ​​then used to weighted normalize the error between the expected and actual states to obtain a control execution deviation evaluation matrix. In this process, firstly, the cross-domain collaborative control decision-making scheme is structurally parsed, decomposing the natural language text of the decision-making scheme into several specific equipment execution instructions. Specifically, through predefined instruction template matching and entity extraction, the target equipment identifier, operation type, operation parameters, and expected target state of each operation instruction are identified from the scheme text. Assuming a total of n equipment execution instructions are obtained after parsing, the expected target state of each instruction can be represented as a multi-dimensional state vector, containing the target values ​​of various physical quantities that the equipment should achieve after executing the instruction (such as target voltage value, target temperature value, target switch state, etc.). The expected target state vectors of the n instructions are arranged in rows to form an expected state matrix. Subsequently, after the equipment execution instructions are actually issued and executed, the actual operating state of each device is collected through sensors and monitoring devices in the power distribution terminal to obtain the actual state feedback vector. The actual state feedback of the device corresponding to each instruction is a multi-dimensional vector with the same dimension as the expected state vector, recording the actual physical quantity values ​​reached by the device after instruction execution. The actual state feedback vectors corresponding to n instructions are arranged row-wise to form an actual state matrix. Next, the error matrix between the expected and actual states is calculated. Each element of the error matrix is ​​the difference between the element value at the corresponding position in the expected state matrix and the element value at the corresponding position in the actual state matrix, representing the execution deviation of each instruction in each physical quantity dimension. However, since different physical quantities have different dimensions and orders of magnitude (e.g., voltage deviation is in kilovolts with values ​​in the single to double digits, temperature deviation is in degrees Celsius with values ​​in the single digits, while switch state deviation is a binary quantity of 0 or 1), directly using the original error values ​​for subsequent processing will lead to an unbalanced evaluation due to the inconsistency in dimensions. Therefore, a dimensional weight matrix is ​​used to perform weighted normalization on the error matrix. The dimensional weight matrix is ​​a pre-calibrated diagonal matrix, where each element on the diagonal is a normalized weight factor for the corresponding physical quantity dimension (i.e., the reciprocal of the nominal range of the physical quantity or an importance weight determined by engineering experience). The error matrix is ​​right-multiplied by the dimensional weight matrix to achieve normalized scaling of the deviations in each dimension. This yields the control execution deviation evaluation matrix, which quantifies the actual deviation of each executed command in each state dimension during this cross-domain collaborative control using a unified dimensionless scale. This provides quantitative feedback information on the execution effect for subsequent knowledge increment residual extraction.

[0041] Specifically, in step S52, text feature embedding and differential alignment residual extraction are performed on the control execution deviation evaluation matrix and historical fault handling experience text to obtain a knowledge increment residual feature vector. In this process, firstly, feature embedding encoding is performed on the control execution deviation evaluation matrix. Specifically, each row of the deviation evaluation matrix is ​​mapped to a fixed-dimensional feature vector through a fully connected encoding network, and then weighted aggregation is performed on the feature vectors of all rows, finally outputting a fixed-dimensional control deviation feature vector. Subsequently, text feature embedding encoding is performed on the historical fault handling experience text. Specifically, the historical fault handling experience text is input into a semantic embedding model to obtain a high-dimensional semantic feature vector of the historical experience. This vector encodes the most relevant experiential semantic information in the existing knowledge of this node to the current interaction. Finally, differential alignment residual extraction operation is performed on the control deviation feature vector and the historical experience feature vector. Specifically, the difference between the regulation deviation feature vector and the historical experience feature vector is calculated element by element in the same feature space. The resulting difference vector is the knowledge increment residual. Positive values ​​indicate aspects where the regulation effect is better than expected (knowledge in this direction can be strengthened), negative values ​​indicate aspects where the regulation effect is worse than expected (knowledge in this direction needs to be corrected), and zero values ​​indicate aspects that are completely consistent with historical experience (no need to update).

[0042] Specifically, in step S53, the knowledge increment residual feature vector is subjected to feature increment hashing and cache pool overlay update to obtain the updated node high-frequency knowledge summary cache pool. In this process, firstly, the knowledge increment residual feature vector is scaled by a learning rate factor and added to the baseline historical feature vector. Then, the accumulated quantization residual vector is read from the quantization residual memory buffer and fused element-wise to obtain the quantization error-compensated evolutionary feature vector. The learning rate factor is a small scalar value between 0 and 1, used to control the step size of incremental updates, avoiding knowledge mutations or instability caused by excessively large single updates. The knowledge increment residual feature vector is multiplied by the learning rate factor to obtain the scaled increment vector, which is then added element-wise to the baseline historical feature vector (i.e., the continuous spatial feature vector corresponding to the cache entry in the previous update) to obtain the preliminary incrementally fine-tuned feature vector. Subsequently, the accumulated quantization residual vector from the previous update cycle is read from the locally maintained quantization residual memory buffer. This residual vector is then fused with the feature vector after initial incremental fine-tuning using element-wise addition compensation. This process re-injects the fine-grained bias information discarded due to binarization truncation in previous update cycles into the current update's computational path. This enables the cross-cycle binarization process to adaptively compensate for historical quantization distortion, preventing cumulative drift between the binary summary and the node's true knowledge state after multiple incremental updates. Furthermore, based on an adaptive threshold using the statistical median, bit-balanced binarization encoding and element-wise difference extraction are performed on the quantization error-compensated evolutionary feature vector to obtain the updated fixed-length binary semantic summary and the current quantization residual vector. The binarization encoding uses an adaptive median threshold as the decision boundary—the statistical median of all element values ​​across all dimensions of the evolutionary feature vector is calculated as the threshold. Dimensions greater than the median are encoded as bit values ​​1, and dimensions less than or equal to the median are encoded as bit values ​​0. The use of an adaptive median threshold ensures that the number of 1s and 0s in the output binary digest tends to be balanced (each accounting for about half), maximizing the information entropy and discriminative power of the digest in Hamming space, and avoiding the hash code convergence problem caused by a fixed zero-point threshold when there is a systematic shift in the feature distribution.

[0043] After binarization encoding is completed, the current quantization residual vector is extracted synchronously, recording the continuous spatial fine information discarded during this binarization process for residual compensation in the next update cycle. The current quantization residual vector is obtained by calculating the element-wise difference between the evolutionary feature vector and its continuous spatial equivalent expression of the binarization result. Finally, the current quantization residual vector is written back to the corresponding storage slot in the local quantization residual memory buffer, completing the temporal rolling update of the residual memory. This ensures that the latest accumulated residual can be read from the buffer for compensation injection when the next knowledge increment update is initiated, thus forming a complete feedback loop of "read-compensation-quantization-extraction-writeback". Simultaneously, the updated fixed-length binary semantic summary is written to the corresponding entry position in the on-chip hardware solid-state storage through the direct memory access channel of the data processing unit to perform an overwrite operation, with the new summary directly replacing the old summary at that position. At this point, the updated node high-frequency knowledge summary cache pool is obtained, in which the updated entries reflect the knowledge increment information brought about by this cross-domain collaborative interaction. Subsequently, when new cross-domain semantic probe data packets arrive, the data processing unit in step three will use the updated summary entries to perform hardware-level Hamming distance matching, so that the node's hardware-level matching response capability always remains consistent with its latest knowledge state.

[0044] Specifically, as described above, in the distributed power multi-agent collaborative knowledge retrieval enhancement method, step S53 plays a crucial role in transforming the continuously learned incremental knowledge features into a hardware-executable binary summary and writing it into the data processing unit's cache pool. The first embodiment uses a sign function to perform binary hash compression on the incrementally fine-tuned evolutionary feature vector, and then directly overwrites and writes the generated binary summary into the solid-state storage of the data processing unit. However, the sign function, as an extreme nonlinear quantization operation, forcibly truncates each dimension of continuous floating-point features to +1 or -1 with a fixed zero-point as the decision threshold. This truncation process inevitably produces quantization residuals between the true values ​​of continuous features and their binary approximate expressions. In the processing flow of the first embodiment, these quantization residuals are completely discarded after each update cycle, and the system only retains the binary summary while permanently losing the fine offset information in the continuous space.

[0045] Examining this deficiency in the context of actual operation of distributed power systems, the distribution terminal intelligence agent undergoes extremely frequent incremental knowledge updates during daily operation. Scenarios such as protection strategy adjustments triggered by seasonal load structure changes, power flow characteristic evolution resulting from the integration of new distributed photovoltaic or energy storage devices, and continuous fault learning triggered by regional extreme weather events all drive high-frequency incremental iterations of the node knowledge base. Although the magnitude of a single quantization residual is small, when the number of updates accumulates to tens or even hundreds, these successively discarded residuals will continuously accumulate in the same direction, causing a gradual and irreversible drift between the binary digest stored in the data processing unit's hardware storage and the node's actual semantic state of knowledge. Since the hardware-level Hamming distance matching of the data processing unit in step three relies entirely on the binary digest in the cache pool as the judgment basis, the aforementioned cumulative drift will directly lead to two serious engineering consequences: First, relevant knowledge retrieval requests that should have been activated fail due to the digest drift exceeding the matching threshold, resulting in nodes with relevant handling experience failing to be activated (missed detection); Second, originally unrelated nodes are mistakenly triggered due to the digest drift entering the wrong matching range, consuming valuable host processor computing power and introducing noisy knowledge (false alarm). These two consequences are particularly dangerous in distribution network fault emergency scenarios—missed detection means that upstream nodes with mature handling experience cannot provide timely control strategy support to downstream power outage areas, while false alarms may cause the execution end to receive erroneous guidance schemes that do not match the current operating conditions.

[0046] Furthermore, the fixed zero-point sign function used in the first embodiment has another limitation. In distributed power networks, the knowledge feature vectors of distribution terminals in different geographical areas often exhibit a systematic distribution shift due to differences in the user groups and equipment types they serve. For example, the load characteristics of densely populated commercial areas in urban areas are generally biased towards the high-power positive value range, while the power flow characteristics of rural decentralized photovoltaic access areas are concentrated in the low-power fluctuation range. When a fixed zero point is used as the binarization decision boundary, nodes with a distribution bias towards the positive value region will produce a large area of ​​convergent all-1 summaries, losing the ability to distinguish between different fault types and causing the Hamming distance matching of the data processing unit to degenerate into a nearly random invalid decision.

[0047] To address the aforementioned deficiencies, this application further proposes a second embodiment, which introduces a compensation update method based on the cross-cycle cumulative coupling relationship of quantization residuals. This method eliminates the loss of discriminative power caused by semantic drift accumulation and feature distribution shift during continuous learning by coordinating three steps: injecting historical quantization error memory before binarization, replacing the fixed sign function with an adaptive threshold, and establishing a residual buffer rolling update mechanism.

[0048] Specifically, firstly, the knowledge increment residual feature vector is scaled by a learning rate factor and added to the baseline historical feature vector. Then, the accumulated quantization residual vector is read from the quantization residual memory buffer and fused element-wise to obtain the quantization error-compensated evolutionary feature vector. Here, it should be understood that after completing the learning rate scaling of the knowledge increment residual feature vector and the fine-tuning by addition of the baseline historical feature vector, the system needs to address a key issue: the fine-grained bias information truncated and discarded by the binarization operation in previous update cycles has been permanently removed from the main data path. If not compensated, this lost information will continue to degrade the summarization accuracy in subsequent cycles in the form of cumulative drift.

[0049] Therefore, before feeding the initial evolutionary feature vector into the binarization stage, the accumulated quantization residual vector left over from the previous update cycle is read from the locally maintained quantization residual memory buffer and compensated element-wise with the initial evolutionary feature vector. This compensation operation is physically equivalent to a Sigma-Delta modulation error shaping mechanism, making the cross-cycle binarization process no longer a memoryless independent truncation, but possessing adaptive compensation capabilities for historical quantization distortion. During the long-term operation of the distribution terminal agent, which undergoes hundreds of incremental fault learning iterations, each compensation injection reintegrates the sub-threshold fine information discarded in previous cycles into the current quantization decision-making scope. This effectively curbs the gradual deviation between the binary summary and the node's true semantic state, ensuring that the retrieval matching criteria in the data processing unit's hardware cache remain synchronized with the actual evolution trajectory of the node's capabilities.

[0050] This step can be represented as:

[0051] in, This represents the evolutionary feature vector after quantization error compensation. This represents the baseline historical feature vector corresponding to the original empirical text in this map atlas. This represents the exponentially decaying learning rate factor. This represents the knowledge increment residual feature vector output from the upstream cross-attention mechanism. This indicates the first time the data was read from the quantization residual memory buffer. The cumulative quantization residual vector left over from each update cycle.

[0052] It should be noted that the first item inside the parentheses The current incremental fine-tuning has been completed; the second item... Historical error memory injection has been completed, and the addition operation between the two has realized the temporal coupling and fusion of current incremental update and historical quantization error memory.

[0053] Next, based on the adaptive threshold of the statistical median, bit-balanced binarization encoding and element-wise difference extraction are performed on the quantization error-compensated evolutionary feature vector to obtain the updated fixed-length binary semantic summary and the current quantization residual vector. That is, after obtaining the quantization error-compensated evolutionary feature vector, it is compressed into a fixed-length binary summary suitable for line-speed matching at the hardware level of the data processing unit. Unlike the hard sign function with a fixed zero point in the first embodiment, an adaptive median threshold based on the current statistical distribution of the vector is introduced here as the binarization decision boundary—encoding dimensions in the feature vector greater than the median threshold as 1 and dimensions less than the median threshold as 0.

[0054] In distributed power networks, distribution terminals in different areas exhibit significant differences in the user groups and equipment types they serve, often resulting in a systematic distributional shift in their knowledge feature vectors. For example, fault characteristics in high-load urban areas are generally biased towards positive values, while power flow characteristics in rural distributed photovoltaic areas are concentrated in low-value fluctuation ranges. Fixed-zero sign functions generate widely convergent hash codes when faced with such offset distributions, losing the ability to distinguish between different types of fault experiences. Adaptive median thresholding, however, ensures a balance between the number of 1s and 0s in the output digest, maximizing the information entropy of the binary digest in Hamming space. This guarantees that even at nodes with systematically shifted feature distributions, the hardware matching stage of the data processing unit still possesses sufficient distinguishing accuracy.

[0055] Its binary encoding process can be represented as:

[0056] in, This represents the updated fixed-length binary semantic digest. This indicates an indicator function (outputs 1 if the condition is true, otherwise outputs 0). This represents the adaptive decision threshold obtained by taking the median of all dimensions of the evolutionary feature vector after quantization error compensation.

[0057] After completing the binarization encoding, the fine information discarded during this quantization process is extracted simultaneously for use in the next cycle for compensation. By calculating the element-wise difference between the evolutionary feature vector after quantization error compensation and its integer equivalent representation of the binarization result, the continuous spatial information lost in this binarization is accurately recorded.

[0058] This stage can be represented as:

[0059] in, Indicates the first The current quantized residual vector generated in each update cycle This represents the integer representation of a binary vector (taking values ​​0 / 1) as a symmetric binary vector (taking values ​​-1 / +1), i.e., the equivalent approximation of the binarization result in continuous space. This represents a vector consisting entirely of 1s.

[0060] It should be noted that the minuend in the subtraction operation To achieve high-precision continuous representation after compensation, the subtraction is its coarse-grained binarization approximation. The difference between the two precisely captures the subthreshold fine offset discarded in this quantization, providing a quantitative basis for residual compensation in the next cycle.

[0061] Then, the current quantized residual vector is written back to the quantized residual memory buffer to complete the residual memory rolling update, and the updated fixed-length binary semantic digest is written to the hardware storage of the data processing unit to obtain the updated node high-frequency knowledge digest cache pool. That is, after obtaining the current quantized residual vector and the updated fixed-length binary semantic digest, two parallel persistent write operations are completed to close the entire compensation update data loop.

[0062] Specifically, firstly, the current quantized residual vector is written back to the corresponding storage slot in the local quantized residual memory buffer, completing the time-series rolling update of the residual memory. This ensures that when the next knowledge increment update is initiated, the latest accumulated residual can be read from the buffer for compensation injection in step one, thus forming a complete feedback loop of read-compensation-quantization-extraction-writeback. Subsequently, the updated fixed-length binary semantic digest is written to the corresponding entry position in the on-chip hardware solid-state storage through the direct memory access channel of the data processing unit to perform insertion or overwrite operations, ultimately obtaining the updated node high-frequency knowledge digest cache pool. In the actual deployment scenario of the distribution network, this closed-loop mechanism ensures that even if a single distribution terminal undergoes hundreds of seasonal load change learning, equipment access event learning, and fault emergency experience learning, the binary digest cached in its data processing unit hardware can still accurately match the semantic probe data packet transmitted from the remote cross-domain in step three, maintaining the stability and reliability of the network-wide knowledge collaborative retrieval system over its long lifecycle.

[0063] In summary, by introducing a quantitative residual cross-cycle cumulative coupling compensation mechanism, the problem of progressive semantic drift of hardware cache summaries caused by the periodic accumulation of binary truncation error in the continuous learning scenario of distributed power multi-agent is solved. This ensures that the binary matching basis stored in the data processing unit of the distribution terminal agent can still accurately correspond to the actual knowledge semantic state of the node after experiencing high-frequency incremental knowledge updates.

[0064] Meanwhile, the adaptive median threshold binarization strategy replaces the hard sign function with a fixed zero point, eliminating the hash code convergence and discriminative degradation caused by the systematic shift in feature distribution of nodes in heterogeneous regions, thus ensuring the uniformity of the summary distribution and the maximum information entropy in the Hamming space. In terms of engineering performance, this improvement effectively reduces the false negative and false positive rates in the hardware-level retrieval and matching process of the data processing unit, reduces the ineffective computing power consumption of the host processor caused by the erroneous wake-up of irrelevant nodes, and improves the matching accuracy stability and hardware resource utilization efficiency of the distributed power multi-agent collaborative knowledge retrieval system under long-term continuous operation conditions.

[0065] In summary, the distributed power multi-agent collaborative knowledge retrieval enhancement method according to the embodiments of this application is explained. It uses a graph convolutional network to extract features from abnormal operating condition data and perform binarization encoding to generate a compact semantic summary. This summary is embedded in the protocol extension header and delivered to each node along with multicast data packets. The receiving hardware layer performs linear-speed pre-screening through XOR operations and Hamming distance quantization. Only nodes whose matching degree meets the threshold are awakened by the host processor to perform knowledge graph retrieval and enhancement generation. Differential alignment and incremental hashing mechanisms ensure the long-term consistency between the cached summary and the evolution of node knowledge. This approach significantly reduces the frequency of invalid awakenings of irrelevant nodes, lowers the bandwidth consumption and computational resource consumption caused by redundant message processing, and thus improves the real-time performance and accuracy of knowledge matching.

[0066] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A knowledge retrieval enhancement method for distributed power multi-agent collaboration, characterized in that, include: Step 1: Using a graph convolutional network, feature extraction and binarization encoding are performed on the abnormal operating condition measurement data stream of the power distribution terminal to obtain the semantic routing summary to be queried and the abnormal operating condition query message. Step 2: Obtain cross-domain semantic probe data packets by embedding the semantic route summary to be queried into the network protocol extension header and encapsulating the abnormal operation query message into a payload, performing data packet assembly and multicast delivery; Step 3: Perform hardware-level XOR operation and Hamming distance quantization on the header digest of the cross-domain semantic probe data packet and the entries in the node high-frequency knowledge digest cache pool. When the Hamming distance is lower than the adaptive matching threshold, a hardware-level wake-up interrupt signal is generated. Step 4: Based on the hardware-level wake-up interrupt signal, the host processor is triggered to unpack and restore the cross-domain semantic probe data packet and perform retrieval, enhancement and fusion in the knowledge graph running locally to obtain cross-domain collaborative control decision-making schemes and historical fault handling experience texts. Step 5: Differential alignment and incremental hashing are performed on the cross-domain collaborative control decision-making scheme and historical fault handling experience text to obtain the updated node high-frequency knowledge summary cache pool.

2. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 1, characterized in that, The abnormal operating condition measurement data stream of the power distribution terminal includes the three-phase transient overvoltage amplitude sequence, transformer winding hot spot temperature value, and circuit breaker non-full-phase operation switching pulse.

3. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 2, characterized in that, Step one includes: Feature extraction and channel-wise splicing and fusion of each modal sub-signal in the abnormal operating condition measurement data stream of the power distribution terminal are performed to obtain a multimodal fusion feature vector; Based on the local device topology adjacency matrix, graph convolution aggregation and low-rank mapping are performed on the multimodal fusion feature vector to obtain the semantic route summary to be queried; Semantic translation and structured encapsulation are performed on the extreme values ​​and physical attributes of the abnormal operating conditions measurement data stream of the power distribution terminal to obtain the abnormal operating condition query message.

4. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 1, characterized in that, Step two includes: The abnormal operating condition query message is serialized, encoded, and compressed, and then a custom data payload is obtained by concatenating the timestamp and business priority tag. By initializing the hop-by-hop option extension header and writing the semantic probe-specific option type field and bit width length field, the header data field of the semantic route digest to be queried is embedded and aligned to obtain the extended header of the embedded digest; A basic header containing the multicast target address is constructed, and the basic header, the extended header with embedded digest, and the custom data payload are concatenated and spliced ​​according to the protocol field order, and a frame verification sequence is appended. After multicast delivery, a cross-domain semantic probe data packet is obtained.

5. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 1, characterized in that, Step three includes: The data processing unit intercepts cross-domain semantic probe packets at the network card hardware layer, and uses a bit mask extraction circuit to strip the binary bit string from the extended header according to a predefined offset address to obtain the probe source semantic digest. Based on the on-chip XOR logic gate array and adder tree, fully parallel bitwise XOR and group number statistics are performed on each entry in the probe source semantic summary and node high-frequency knowledge summary cache pool. The optimal matching Hamming distance value is obtained after minimum value filtering. The optimal matching Hamming distance value is compared with the adaptive matching threshold fixed in the register by a hardware digital comparator. When the distance value is lower than the adaptive matching threshold, the interrupt controller is driven to generate a hardware-level wake-up interrupt signal.

6. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 1, characterized in that, Step four includes: Based on the hardware-level wake-up interrupt signal triggering the host processor's high-priority interrupt response, the protocol stack header of the cross-domain semantic probe data packet is stripped and decompressed and deserialized to obtain the decompressed abnormal condition query message. The decompressed abnormal operating condition query message is used to perform high-dimensional semantic vector retrieval and most similar entity matching extraction in the local knowledge graph to obtain historical fault handling experience text. The decompressed abnormal operating condition query message and historical fault handling experience text are used together to fill the search enhancement to obtain the prompt word template. The prompt word template is then autoregressive conditional probability decoding is performed on the prompt word template to generate a cross-domain collaborative control decision scheme.

7. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 1, characterized in that, Step five includes: The cross-domain collaborative control decision-making scheme is parsed into equipment execution instructions and the actual state feedback vector after execution is collected. The error between the expected and actual state is weighted and normalized using a dimensional weight matrix to obtain the control execution deviation evaluation matrix. Text feature embedding and differential alignment residual extraction are performed on the control execution deviation assessment matrix and historical fault handling experience text to obtain knowledge increment residual feature vectors; The feature increment hash and cache pool overwrite update are performed on the knowledge increment residual feature vector to obtain the updated node high-frequency knowledge summary cache pool.

8. The knowledge retrieval enhancement method for distributed power multi-agent collaboration according to claim 7, characterized in that, The knowledge increment residual feature vector is subjected to feature increment hashing and cache pool overwrite update to obtain the updated node high-frequency knowledge summary cache pool, including: The knowledge increment residual feature vector is scaled by the learning rate factor and added to the baseline historical feature vector. Then, the accumulated quantized residual vector is read from the quantized residual memory buffer and fused element by element to obtain the evolutionary feature vector after quantization error compensation. Based on the statistical median adaptive threshold, bit-balanced binary encoding and element-by-element difference extraction are performed on the evolutionary feature vector after quantization error compensation to obtain the updated fixed-length binary semantic summary and the current quantization residual vector. The current quantized residual vector is written back to the quantized residual memory buffer to complete the residual memory rolling update, and the updated fixed-length binary semantic summary is written to the hardware storage of the data processing unit to obtain the updated node high-frequency knowledge summary cache pool.

Images