Energy storage RAG knowledge base system based on time sequence-semantic double channel

Abstract

The application discloses a kind of energy storage RAG knowledge base systems based on time sequence-semantic dual channel, belong to knowledge base management technical field. Including: time sequence processing module, the time sequence data of energy storage power station is acquired, and time sequence key of corresponding working condition is generated according to time sequence data;Mixed index module, the semantic vector of each knowledge item in knowledge base is associated with the time sequence key of corresponding working condition and is stored, and graph index is obtained by construction;Incremental update control module includes: similarity judgment unit, receives new knowledge item, and calculates similarity;Difference replacement unit, in the case where similarity is greater than first threshold value, difference is calculated, and semantic vector is updated by difference;Question and answer interface module receives user question, obtains time sequence key, retrieves knowledge item, and generates answer.The application eliminates the drawbacks that several reports are triggered for micro-increase in traditional scheme for several hours full reconstruction, while greatly reducing CPU / GPU computing power consumption.

Description

Technical Field

[0001] This invention belongs to the field of knowledge base management technology, specifically relating to an energy storage RAG knowledge base system based on a time-series-semantic dual-channel approach. Background Technology

[0002] Throughout the multi-year lifecycle of a 100MWh-level lithium iron phosphate energy storage power station, massive amounts of heterogeneous data are continuously generated, including drawings, reports, voice recordings, infrared images, and millisecond-level voltage / temperature / power flow data. Owners commonly deploy a "RAG combined with a large model" question-and-answer system to enable frontline maintenance personnel to quickly retrieve technical specifications, fault cases, and handling solutions using natural language. The relevant technologies typically involve offline batch import of unstructured data such as PDFs, JPGs, and audio files, extracting text using OCR or ASR, and then uniformly converting it to Markdown format. Time-series data is stored separately in SCADA or PI historical databases, physically isolated from the text database. During knowledge base construction, a Sentence-BERT or CLIP model is used to perform a one-time vectorization of all documents, writing the data into a FAISS or HNSW index file. This index is manually replaced periodically (weekly or monthly) with each software version update. When a user asks a question, the system retrieves the top-k relevant paragraphs through semantic vector retrieval, concatenates prompts, and then generates the answer using a large model.

[0003] However, the aforementioned technologies have the following drawbacks: Incremental updates employ a strategy of full retraining and overall switching. Whenever a new file arrives or any additions to the knowledge base are needed, the system must perform a full revectorization of the entire knowledge base in the test environment, followed by an overall index switch during off-peak hours at night. For a vector database with hundreds of millions of entries, a single full reconstruction can take several hours. This coarse-grained update method results in: first, even minor additions to a few reports require recalculating the entire knowledge base, leading to a severe waste of CPU / GPU computing resources and extremely high computational overhead; second, the nighttime update window is often occupied by SCADA system backup tasks, causing delays or even forced skips in knowledge base updates, preventing frontline personnel from accessing the latest knowledge in a timely manner; and third, the system must be shut down or in an unstable state during reconstruction, failing to meet the timeliness requirements of 24 / 7 uninterrupted operation and maintenance for energy storage power stations. Summary of the Invention

[0004] The purpose of this invention is to provide an energy storage RAG knowledge base system based on a time-series-semantic dual-channel approach, which can solve the problems existing in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides an energy storage RAG knowledge base system based on a time-series-semantic dual-channel approach, comprising: The timing processing module is used to acquire the timing data of the energy storage power station in real time and generate the timing key of the corresponding operating condition based on the timing data. The hybrid index module is used to associate and store the semantic vector of each knowledge entry in the knowledge base with the time sequence key of the corresponding working condition to construct a graph index. The graph index contains multiple nodes and edges between each node. Each node corresponds to a knowledge entry, and each node stores the semantic vector of the corresponding knowledge entry and the time sequence key of the corresponding working condition. The incremental update control module includes: A similarity judgment unit is used to receive new knowledge entries and calculate the similarity between the semantic vector of the new knowledge entry and the semantic vector of existing nodes in the graph index. The differential replacement unit is used to calculate the difference between the semantic vector of the new knowledge entry and the semantic vector of the existing node when the similarity output by the similarity judgment unit is greater than a first threshold, and update the semantic vector of the existing node with the difference. The question-and-answer interface module is used to receive user questions, obtain the time sequence key of the current working condition, and use the graph index to retrieve knowledge entries that are semantically related to the user question and match the time sequence key of the current working condition, and generate the answer to the user question.

[0006] Optionally, it also includes a drift detection module; the incremental update control module further includes an edge update unit; The drift detection module is used to monitor the distribution changes of the time series data acquired by the time series processing module, and send a trigger signal to the edge update unit when a data distribution drift is detected. The edge update unit is specifically used for: In response to the trigger signal sent by the drift detection module, the affected nodes in the graph index are identified; The connected region formed by the affected nodes and their associated edges in the graph index is taken as the subgraph to be reconstructed, and the edges between each node in the subgraph are recalculated to update the edge relationships in the subgraph.

[0007] The technical solutions provided by the embodiments of the present invention bring at least the following beneficial effects: This invention achieves intelligent differential updates for new knowledge entries by setting up an incremental update control module. When the semantic vector similarity between a new knowledge entry and an existing node exceeds a first threshold, only the vector difference between the two is calculated, and the semantic vector of the existing node is directly updated with this difference, without recalculating the vector index of the entire knowledge base. This invention eliminates the drawback of traditional solutions that trigger a full reconstruction of the entire knowledge base for several hours due to a few small increases in reports, while significantly reducing CPU / GPU computing power consumption. In addition, this invention does not require service interruption or a complete index switch during the update process, solving the problem of knowledge base update delays or skips caused by SCADA backup tasks crowding out the nighttime update window, ensuring the real-time availability and usability of knowledge in 24 / 7 operation and maintenance scenarios for energy storage power stations. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the architecture of an energy storage RAG knowledge base system based on a time-series-semantic dual-channel approach according to the present invention. Figure 2 This is a schematic diagram of the differential replacement timing of the incremental update control module in one embodiment of the present invention; Figure 3 This is a schematic diagram of a local subgraph reconstruction process driven by drift detection in one embodiment of the present invention; Figure 4 This is a schematic diagram of the active evolution state of knowledge entries in one embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] The terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0011] The following detailed description, in conjunction with the accompanying drawings, of a time-series-semantic dual-channel energy storage RAG knowledge base system provided by the present invention through specific embodiments and application scenarios, will be provided in detail.

[0012] Figure 1This is a schematic diagram of the architecture of an energy storage RAG knowledge base system based on a time-series-semantic dual-channel approach, as described in this invention. Figure 1 As shown, it includes: The timing processing module is used to acquire the timing data of the energy storage power station in real time and generate timing keys for the corresponding operating conditions based on the timing data.

[0013] The timing processing module establishes a high-speed data channel with the monitoring and data acquisition system or battery management system of the energy storage power station via the Message Queuing Telemetry Transmission (MQTT) protocol. It receives real-time timing data such as cell voltage, cell temperature, power conversion system power, cooling system flow, and various alarm codes at a preset sampling frequency. To extract key features characterizing the current operating condition from massive, high-dimensional timing data, the timing processing module integrates a timing feature extraction unit. This unit preferably employs a compression model combining a shape extraction algorithm and a Long Short-Term Memory (LSTM) network-autoencoder. The shape extraction algorithm identifies the most discriminative subsequence from the original timing sequence as operating condition features, while the LSM-autoencoder further compresses the variable-length timing window into a fixed-length low-dimensional hidden state. Finally, it generates a compact timing key, such as a 64-bit binary form, through quantization encoding. This timing key serves as a unique digital signature of the current operating condition, highly summarizing the operating state of the energy storage system within the most recent time window, including but not limited to key operating condition information such as state of charge (SOC) range, rate characteristics, temperature distribution, and voltage consistency.

[0014] The hybrid index module is used to associate and store the semantic vector of each knowledge entry in the knowledge base with the time sequence key of the corresponding working condition to construct a graph index. The graph index contains multiple nodes and edges between each node. Each node corresponds to a knowledge entry, and each node stores the semantic vector of the corresponding knowledge entry and the time sequence key of the corresponding working condition.

[0015] The hybrid indexing module receives structured textual knowledge entries from external processing flows, such as design drawings parsed by optical character recognition, maintenance voice converted by speech recognition, various test reports, and fault cases. For each knowledge entry, the hybrid indexing module calls a pre-trained semantic vectorization model to convert it into a high-dimensional semantic vector. Simultaneously, the hybrid indexing module obtains the time sequence key generated by the time sequence processing module within the same time window or for the same device state. Subsequently, the hybrid indexing module encapsulates the semantic vector and the time sequence key together as a node in a graph index. The edges between nodes represent the proximity relationships of different knowledge entries in the semantic space; the weights of these edges are pre-calculated and fixed based on the similarity of the semantic vectors between nodes. In this way, the hybrid indexing module expands the traditional flat vector index into a graph structure with operating condition labels, enabling knowledge retrieval not only to navigate based on semantic relevance but also to filter based on the operating condition through the node's time sequence key attribute.

[0016] The incremental update control module includes: A similarity judgment unit is used to receive a new knowledge entry and calculate the similarity between the semantic vector of the new knowledge entry and the semantic vector of an existing node in the graph index; a difference replacement unit is used to calculate the difference between the semantic vector of the new knowledge entry and the semantic vector of the existing node when the similarity output by the similarity judgment unit is greater than a first threshold, and update the semantic vector of the existing node with the difference.

[0017] Figure 2 This is a schematic diagram of the differential replacement timing of the incremental update control module in one embodiment of the present invention, as shown below. Figure 2 As shown, when a new knowledge entry arrives, the incremental update control module performs a hot update interaction process through the similarity judgment unit and the difference replacement unit. The core concept is: first, calculate the semantic similarity between the new entry and the existing nodes; if it exceeds the first threshold, directly calculate the vector difference and update the node in place, avoiding full reconstruction.

[0018] Specifically, the incremental update control module is used to realize dynamic, low-overhead expansion of the knowledge base, including a similarity judgment unit and a differential replacement unit. The similarity judgment unit receives newly arrived knowledge entries and calculates the similarity between the semantic vector of the new knowledge entry and the semantic vectors of existing nodes in the graph index constructed by the hybrid indexing module. The calculation can use cosine similarity or Euclidean distance as a metric. When the calculated similarity is greater than a preset first threshold, it indicates that the new knowledge entry is highly similar to the knowledge represented by an existing node in the graph index, representing only a minor revision or version update of the old knowledge. At this time, the differential replacement unit is triggered. The differential replacement unit calculates the difference (Δ-embedding) between the semantic vector of the new knowledge entry and the semantic vector of the existing node when the similarity judgment unit determines high similarity, and directly updates the semantic vector value of the existing node using the difference. The above update operation is performed in situ, requiring only modification of the vector data stored in the node, without affecting the node's neighbor relationships and edge connections in the HNSW graph. Therefore, the entire update process can be completed in a very short time, achieving hot updates of the knowledge base without the need to rebuild the entire index.

[0019] The question-and-answer interface module is used to receive user questions, obtain the time sequence key of the current working condition, and use the graph index to retrieve knowledge entries that are semantically related to the user question and match the time sequence key of the current working condition, and generate the answer to the user question.

[0020] The question-and-answer interface module, acting as the front end for system-user interaction, receives questions posed by users in natural language, obtains the time sequence key of the current operating condition, and uses the graph index in the hybrid index module to retrieve knowledge entries semantically relevant to the user's question and matching the time sequence key of the current operating condition, ultimately generating an answer to the user's question. Specifically, upon receiving a user's question, the question-and-answer interface module first calls the same semantic vectorization model used during index construction to convert the user's question into a question semantic vector. Simultaneously, the question-and-answer interface module sends a request to the time sequence processing module to obtain the time sequence key reflecting the real-time operating status of the current energy storage system. Subsequently, the question-and-answer interface module performs a retrieval operation in the graph index of the hybrid index module. The retrieval process involves two levels: first, a nearest neighbor search is performed in the graph index based on the question semantic vector to find the top K knowledge nodes with the highest semantic similarity; second, the obtained current operating condition time sequence key is used to filter the preliminary retrieval results, eliminating nodes whose time sequence keys do not match the current operating condition. The filtered and retained knowledge entries are high-quality knowledge that is semantically relevant to the user's question and consistent with the current system's actual state in terms of operating conditions. Finally, the question-answering interface module combines these retrieved knowledge entries with the user's question according to a preset prompt template to form a complete input context, and then calls a large language model to generate a fluent and accurate answer to return to the user. In this way, the question-answering interface module ensures the timeliness and contextual relevance of the system's answers, fundamentally avoiding the problem of knowledge and operating conditions being misaligned in traditional retrieval-enhanced generation systems.

[0021] In one optional implementation, a drift detection module is also included; the incremental update control module further includes an edge update unit. The drift detection module is used to monitor the distribution changes of the time-series data acquired by the time-series processing module, and send a trigger signal to the edge update unit when a data distribution drift is detected.

[0022] The drift detection module is communicatively connected to the timing processing module and is used to continuously monitor the distribution changes of the timing data acquired by the timing processing module.

[0023] In one optional implementation, the drift detection module is specifically used for: The time series data obtained by the time series processing module is divided into univariate time series data and multivariate time series data.

[0024] The drift detection module splits and reassembles the real-time data stream from the time-series processing module at the data access layer. Specifically, based on data attributes and physical meaning, the time-series data acquired by the time-series processing module is divided into two categories: univariate time-series data and multivariate time-series data. Univariate time-series data focuses on characterizing the independent evolution of a single physical quantity of the device, including but not limited to at least one of cell voltage, cell temperature, PCS power, and cooling flow rate. Multivariate time-series data is used to depict the joint operating state of the system under multiple coupling dimensions. Its specific construction is a high-dimensional feature vector sequence composed of battery operating condition variables (such as cell maximum temperature, cell minimum voltage, and SOC value), PCS operating condition variables (such as active power, reactive power, and DC bus voltage), and cooling operating condition variables (such as coolant inlet temperature and cooling pump speed). In this way, the drift detection module can comprehensively capture the evolution of operating conditions from both single-point anomalies and system coupling anomalies.

[0025] The Page-Hinkley algorithm is used to monitor univariate time-series data, which includes at least one of cell voltage, cell temperature, PCS power, and cooling flow rate.

[0026] For univariate time-series data, the drift detection module employs the Page-Hinkley algorithm for real-time monitoring. The Page-Hinkley algorithm is a sequential analysis technique whose core principle is to detect abrupt changes in the sequence mean by accumulating offsets. In practice, the drift detection module maintains an independent Page-Hinkley tester for each univariate time-series data stream. The Page-Hinkley tester calculates the cumulative deviation between the current observation and the historical mean reference value in real time and compares the cumulative deviation with a preset dynamic threshold. When the cumulative deviation exceeds the threshold, it indicates that the variable (e.g., the temperature of a battery cell) has experienced a statistically significant mean drift, meaning a substantial change has occurred in the device's operating point or health status. Because the Page-Hinkley algorithm has low computational complexity and is sensitive to abrupt changes, it can promptly detect abnormal changes in a single parameter, such as precursors to battery thermal runaway or PCS overload.

[0027] The maximum mean difference algorithm is used to monitor drift in multivariate time series data, which includes battery operating condition variables, PCS operating condition variables, and cooling operating condition variables.

[0028] For multivariate time-series data, the drift detection module employs the Maximum Mean Difference (MMD) algorithm for drift monitoring. The MMD algorithm is a non-parametric test method that measures the difference between two distributions, and is particularly suitable for drift detection of high-dimensional joint distributions. In specific implementation, the drift detection module sets a sliding time window containing the most recently acquired set of multivariate feature vectors (e.g., a joint vector composed of battery, PCS, and cooling variables within the last 24 hours). Simultaneously, the drift detection module maintains a historical reference window, representing the joint distribution of the system under normal operating conditions. The drift detection module calculates the MMD distance between the current sliding window data distribution and the historical reference window data distribution in real time. This is achieved by mapping the data to a regenerating kernel Hilbert space and comparing the distance between the mean vectors of the two distributions in the regenerating kernel Hilbert space. When the calculated MMD distance exceeds a preset threshold, it indicates that the system's distribution in the high-dimensional joint state space has deviated from the normal operating range, i.e., multivariate joint distribution drift has been detected. Multivariate joint distribution drift means that the system has entered a completely new operating condition range (such as the coupling state of high-rate charging and discharging and high temperature), or that the cooperative relationship between devices has changed.

[0029] If either the Page-Hinkley algorithm or the maximum mean difference algorithm triggers an alarm, it confirms that a data distribution drift has been detected.

[0030] When the Page-Hinkley algorithm generates an alarm for any single-variable time-series data, or the maximum mean difference algorithm generates an alarm for a multivariate joint distribution, the drift detection module confirms the detection of data distribution drift. The single-variable channel ensures a rapid response to any abnormal change in a single key parameter, while the multivariate channel compensates for the inability of single-variable monitoring to detect changes in the coupling relationship between variables. Once drift is confirmed, the drift detection module immediately generates a trigger signal containing information such as the drift start time, drift type (single-variable / multivariable), the identifier of the affected device, and the drift intensity, and sends the trigger signal to the side update unit of the incremental update control module.

[0031] The edge update unit is specifically used for: In response to the trigger signal sent by the drift detection module, the affected nodes in the graph index are identified; the connected region formed by the affected nodes and their associated edges in the graph index is taken as the subgraph to be reconstructed, and the edges between each node in the subgraph are recalculated to update the edge relationships in the subgraph.

[0032] The edge update unit responds to the trigger signal sent by the drift detection module and performs drift-aware local graph index reconstruction. The edge update unit parses the trigger signal to extract the drift period and the affected equipment or operating condition range. Based on this, the edge update unit traverses the graph index in the hybrid index module, identifying all knowledge nodes whose associated time keys overlap with the operating condition characteristics of the drift period, or whose knowledge DNA tags (e.g., operating condition range tags, equipment model tags, which will be explained in detail later) are directly related to the post-drift operating condition. These nodes are marked as affected nodes. Using these affected nodes as seed points, the edge update unit expands outward using a graph traversal algorithm (such as breadth-first search or depth-first search), dividing the connected region formed by each affected node and all its directly associated neighboring nodes into a subgraph to be reconstructed. Subsequently, the edge update unit recalculates the edge relationships between nodes within this subgraph. After recalculation, the edge update unit updates the original edge relationships within the subgraph using the newly generated edge set.

[0033] In one optional implementation, the edge update unit is specifically used for: In response to the trigger signal sent by the drift detection module, the affected nodes in the graph index are determined.

[0034] Figure 3 This is a schematic diagram of a local subgraph reconstruction process driven by drift detection in one embodiment of the present invention. Please refer to [link / reference]. Figure 3 After receiving the trigger signal from the drift detection module, the edge update unit parses the trigger signal and extracts key information such as the drift period, drift type (univariate or multivariate), affected device identifier, and drift intensity. Based on this information, the edge update unit traverses the currently used graph index in the hybrid index module to determine the nodes affected by the drift. Specifically, the edge update unit marks all nodes in the graph index that meet any of the following conditions as affected nodes: First, the time sequence key stored by the node overlaps with the operating condition characteristics of the drift period, that is, the operating condition corresponding to the node is directly related to the current operating condition drift; Second, the knowledge DNA tag associated with the node matches the operating condition parameters after the drift.

[0035] The connected regions formed by the affected nodes and their associated edges are separated from the graph index and used as subgraphs to be reconstructed. The number of nodes in the subgraph does not exceed a preset proportion of the total number of nodes in the entire database.

[0036] Please see Figure 3The edge update unit uses each affected node as a seed node and employs a graph traversal algorithm to expand the graph index. It partitions all connected regions formed by nodes directly or indirectly connected to the affected nodes through one or more edges, along with their associated edges, from the complete graph index, creating subgraphs to be reconstructed. Understandably, after subgraph partitioning, all knowledge entries semantically adjacent to the affected knowledge can be included in the reconstruction scope, thus maintaining the structural consistency of the knowledge graph after local updates. Simultaneously, to control the computational cost of a single update, the edge update unit introduces a size constraint during subgraph partitioning, ensuring that the number of nodes in the partitioned subgraph does not exceed a preset proportion of the total number of nodes in the entire database. If the size of the subgraph partitioned based on connected regions exceeds the preset proportion, the edge update unit will sort the affected nodes according to their drift intensity, prioritizing the seed node with the highest drift intensity and its neighborhood to form the subgraph, or splitting the large-scale subgraph into multiple smaller subgraphs for batch processing, thereby keeping the computational load of each local reconstruction within an acceptable range.

[0037] A copy of the graph index is created as a shadow index, and the edge relationships within the subgraph are recalculated in the shadow index.

[0038] Please see Figure 3 To ensure the continuity and zero interruption of knowledge retrieval services during the update process, after the subgraph to be reconstructed is partitioned, the edge update unit first creates a copy of the currently used graph index, called the shadow index. The shadow index copies all node data, edge relationships, and metadata of the original graph index during creation. Subsequently, the edge update unit recalculates the edge relationships between only the nodes corresponding to the subgraph to be reconstructed within the shadow index. The recalculation process includes: extracting the semantic vectors of all nodes within the subgraph, recalculating the similarity matrix between pairs of nodes based on the similarity between vectors, and reselecting the optimal neighbor nodes and establishing new edge connections for each node within the subgraph according to the graph index structure construction strategy adopted by the hybrid index module. During the recalculation process, the edge relationships between nodes within and outside the subgraph remain unchanged, thus strictly limiting the calculation scope to within the subgraph. Because the size of the subgraph is controlled to a small proportion, the entire edge relationship recalculation process can be completed in a short time.

[0039] Use the shadow index as the new graph index to update the edge relationships within the subgraph.

[0040] Please see Figure 3The edge update unit atomically switches the graph index pointer currently providing retrieval services from the original index to the updated shadow index. After the switch, the shadow index becomes the new, active graph index, used for all subsequent knowledge retrieval requests, while the original old index is released or retained as a rollback backup. Because the switching process is atomic and the shadow index is fully ready before the switch, the entire update process is completely transparent to the upper-layer question-answering interface module, and the knowledge retrieval service experiences no interruption or delay during the index update.

[0041] In an optional implementation, a knowledge entry management module is also included, specifically for: The system receives knowledge entries to be added to the database and generates knowledge DNA tags for each entry. These tags include project stage tags, equipment model tags, operating condition range tags, and document confidence level tags.

[0042] The knowledge entry management module first receives knowledge entries to be added to the database from the upstream processing flow. Upon receiving the knowledge entries, the module generates knowledge DNA tags. Knowledge DNA tags are a set of structured metadata tags used to uniquely characterize the applicable scenarios and credibility of knowledge, specifically including project stage tags, equipment model tags, operating condition range tags, and document confidence tags. The project stage tag identifies the engineering lifecycle stage to which the knowledge entry belongs, including at least one of planning, construction, operation and maintenance, and decommissioning. The knowledge DNA tag can be obtained by parsing the source file path, filename keywords, or specified by the user during upload. The equipment model tag identifies the specific equipment type and model to which the knowledge entry applies, such as PCS-2.5MW, lithium iron phosphate battery cell-280Ah, thermal management system-liquid cooled type, etc. The knowledge DNA tag can be obtained from the equipment field in the knowledge entry content or extracted using rule-based natural language processing. Operating condition range labels are used to describe the range of operating conditions to which the knowledge entry applies, including state of charge (SOC) range, rate range, temperature range, etc. Operating condition range labels can be generated by extracting operating condition description information from the title, body text, or chart description of the knowledge entry; Document confidence labels are used to characterize the reliability of the source of the knowledge entry. Document confidence labels can be preset with different confidence scores according to the document source type (such as official design drawings, third-party test reports, on-site operation and maintenance logs), or automatically calculated through document format standardization, integrity verification, etc. The confidence score range is a real number between 0 and 1.

[0043] The initial lifecycle parameters of the knowledge entry are calculated based on the knowledge DNA tag.

[0044] Based on the generated knowledge DNA tags, the knowledge entry management module further calculates the initial lifecycle parameter (TtL, Time-to-Live) of the knowledge entry according to the knowledge DNA tags. The lifecycle parameter is a time metric used to indicate the expected length of time the knowledge entry will remain valid from the time it is entered into the database. Specifically, the knowledge entry management module is configured with a lifecycle calculation strategy, which defines the lifecycle parameter as the product of a base time and multiple influencing factor coefficients. That is, TtL = T_base × stage coefficient × confidence level, where T_base is a preset base lifecycle, for example, the default setting is 180 days; the stage coefficient is a weighting factor determined according to the project stage tag. Since the knowledge stability and timeliness requirements are different in different stages, for example, knowledge changes are more frequent in the construction stage, so the stage coefficient can be set to less than 1, while knowledge in the operation and maintenance stage is relatively stable, so the stage coefficient can be set to greater than 1; the confidence level adopts the value of the file confidence tag. The lower the confidence level of the knowledge, the shorter its initial lifecycle parameter, thereby prompting it to enter the elimination or review process more quickly.

[0045] The knowledge DNA tag, initial lifecycle parameters, semantic vector of the knowledge entry, and time sequence key of the corresponding working condition of the knowledge entry are associated and stored in the corresponding node of the hybrid index module.

[0046] The knowledge entry management module further associates and stores the aforementioned metadata with the vector representation and operational condition signature of the knowledge content itself. Specifically, the knowledge entry management module first calls the vectorization engine to convert the text or image content of the knowledge entry to be added into the database into high-dimensional semantic vectors. For text-based knowledge, a large-scale pre-trained language model is used to encode the text content into semantic vectors; for image-based knowledge, a visual encoder using a contrastive language-image pre-trained model is used to convert the image content into semantic vectors. Simultaneously, the knowledge entry management module obtains the operational condition time sequence key matching the current knowledge entry time window from the time sequence processing module. The time sequence key reflects the operating condition characteristics of the energy storage system at the time of knowledge generation or acquisition. Finally, the knowledge entry management module treats the knowledge DNA tag, the initial lifecycle parameters, the semantic vector, and the operational condition time sequence key as a whole, and stores this information in the corresponding node of the graph index constructed by the hybrid index module through the node management interface of the hybrid index module. In this node, the semantic vector is used to support subsequent semantic similarity retrieval, the working condition time sequence key is used to support working condition matching filtering, and the knowledge DNA tag and initial lifecycle parameters are stored as additional attributes of the node for subsequent lifecycle management, consistency verification, and knowledge evolution processes. Through the associated storage of multi-dimensional information, this invention encapsulates the semantic content, applicable working conditions, lifecycle, and metadata tags of knowledge uniformly in the nodes of the graph index, enabling the knowledge base to not only have semantic retrieval capabilities but also the ability to filter and manage the lifecycle based on multi-dimensional conditions such as working conditions, stages, and equipment models.

[0047] In an optional implementation, a drift detection module is further included, which is used to monitor the distribution changes of the time-series data acquired by the time-series processing module, and send a trigger signal to the knowledge entry management module when a data distribution drift is detected.

[0048] The drift detection module communicates with the knowledge entry management module to drive the proactive evolution and iteration of knowledge. The drift detection module has already been described in detail above and will not be repeated here.

[0049] The knowledge entry management module is also used for: In response to the trigger signal sent by the drift detection module, the affected time series data range is determined according to the drift period, and the original knowledge entries related to the drift period are retrieved from the hybrid index module.

[0050] Figure 4 This is a schematic diagram of the active evolution state of knowledge items in one embodiment of the present invention, such as... Figure 4As shown, after receiving the trigger signal from the drift detection module, the knowledge entry management module parses the trigger signal to extract the specific range of the drift period. Based on the drift period, the knowledge entry management module determines the affected time-series data range, i.e., the original time-series data sequence collected by the time-series processing module corresponding to the drift period. Subsequently, the knowledge entry management module uses the drift period information and the affected device identifier to retrieve the original knowledge entries related to the drift period in the hybrid index module. Specifically, the knowledge entry management module traverses the graph index in the hybrid index module and filters out those knowledge entries whose operating condition interval labels in their knowledge DNA tags match the drifted operating conditions, or whose timestamps corresponding to their associated time-series keys overlap with the drift period. Simultaneously, to improve the quality of knowledge evolution, the knowledge entry management module further retrieves original materials related to the drift period, including but not limited to unstructured data such as maintenance work orders, voice recordings, and infrared images generated during the drift period. These original materials serve as the factual basis for knowledge generation. Through the above retrieval operations, the knowledge entry management module obtains a set of outdated knowledge and a corresponding set of new working condition raw data.

[0051] The large model is invoked to generate new candidate knowledge based on the original knowledge entries and the time series data. The new candidate knowledge includes new conditions and new conclusions.

[0052] Please see Figure 4 The knowledge entry management module constructs a prompt word template, comprising three parts: the first part is the "condition-conclusion" statement from the retrieved original knowledge entry, i.e., the specific content of the old knowledge; the second part is the key time-series data features within the drift period, such as voltage change trends, temperature distribution statistics, alarm code sequences, etc., presented in the form of text summaries or structured data; the third part is the generation instruction, requiring the large language model to revise or regenerate the old knowledge based on the new operating condition data, outputting candidate knowledge in the form of "new condition-new conclusion". The knowledge entry management module sends the above prompt words to the large language model and receives the generated results returned by it. The new candidate knowledge is presented in a structured form, where the new conditions describe the operating condition characteristics applicable to the knowledge, including but not limited to SOC range, rate range, temperature range, equipment status, etc.; the new conclusion describes the operational suggestions, protection settings, fault handling schemes, etc., to be taken under this operating condition.

[0053] A causal inference algorithm is used to verify the causal relationship between the new conditions and the new conclusion in the new candidate knowledge.

[0054] Because candidate knowledge generated by large language models may contain causal fallacies or be inconsistent with physical laws, such as... Figure 4As shown, the knowledge entry management module further employs a causal inference algorithm to verify the causal relationship between the new conditions and the new conclusion in the new candidate knowledge. Specifically, the knowledge entry management module integrates a causal inference engine, which preferably uses a causal graph model based on the DoWhy framework or a causal effect estimation method based on a structural causal model. During the verification process, the knowledge entry management module uses the "new conditions" in the new candidate knowledge as processing variables, the "new conclusion" as the outcome variable, and time-series data within the drift period as observation data to construct a causal graph model. The causal inference engine adjusts the confounding factor using a backdoor criterion or a frontdoor criterion, and calculates the average causal effect of the new conditions on the new conclusion under given conditions. If the estimated causal effect exceeds a preset threshold and the statistical significance test passes, it indicates that there is a reliable causal relationship between the new conditions and the new conclusion, and the new candidate knowledge is determined to pass the verification. If the causal inference fails, the knowledge entry management module rejects the new candidate knowledge and may choose to re-call the large language model for correction or transfer the candidate knowledge to the manual review queue.

[0055] The verified new candidate knowledge is added to the hybrid index module as a new knowledge entry, and the corresponding original knowledge entry is marked as obsolete.

[0056] Please see Figure 4 For new candidate knowledge verified through causal inference, the knowledge entry management module adds it as a new knowledge entry to the hybrid index module. The addition process follows the aforementioned knowledge addition procedure: generating a corresponding knowledge DNA tag for the new candidate knowledge, calculating initial lifecycle parameters, calling the vectorization engine to generate a semantic vector, and obtaining the time key of the current operating condition. Finally, this information is associated and stored in the graph index of the hybrid index module, forming a new knowledge node. Simultaneously, the knowledge entry management module marks the corresponding original knowledge entry as "obsolete." Marking it as obsolete involves setting an obsolete flag in the metadata of the original knowledge node while retaining its knowledge DNA, semantic vector, time key, and historical version tracking chain. In this way, the present invention achieves the ability of the knowledge base to actively evolve with changes in system operating conditions, ensuring that the knowledge content remains synchronized with the actual operating state of the equipment, and solving the problem of expired knowledge remaining in the traditional RAG system for a long time and failing to be automatically updated.

[0057] In an optional implementation, the knowledge entry management module is further configured to: Determine the file type of the knowledge entry to be added to the database; if the knowledge entry is text, use a text encoder to convert the text content of the knowledge entry into a semantic vector; if the knowledge entry is image, use an image encoder to convert the image content of the knowledge entry into a semantic vector.

[0058] The knowledge entry management module is used to determine the file type of knowledge entries to be added to the database before they are added, and selects the corresponding encoder for semantic vector conversion based on the file type. Specifically, when the knowledge entry management module receives a knowledge entry to be added, it first identifies the file type by parsing the file extension (e.g., .pdf, .jpg, .png, .mp3, etc.) or by checking the file header information, determining whether the knowledge entry is text or image. If the knowledge entry is text, the knowledge entry management module calls a pre-trained text encoder to convert the text content of the knowledge entry into a semantic vector. The text encoder preferably uses the large-scale Chinese pre-trained model bge-large-zh-v1.5, based on the Transformer architecture, which can map text sequences of arbitrary length into dense vectors of fixed dimensions, preserving the semantic similarity between texts in the semantic space. For drawing text parsed by optical character recognition, maintenance voice text converted from speech recognition, and structured text of various test reports and fault cases, the knowledge entry management module first preprocesses the text content, including segmentation, noise reduction, and splicing. Then, the preprocessed text is input into the text encoder to obtain the corresponding semantic vector. When the knowledge entry is an image, the knowledge entry management module calls the image encoder to convert the image content of the knowledge entry into a semantic vector. The image encoder preferably uses a contrastive language-image pre-trained model visual encoder, based on a Vision Transformer or convolutional neural network architecture, which can convert the input image into a semantic vector with the same output dimension as the text encoder, aligning the image and the text describing its content in the same vector space. For image-based knowledge entries such as infrared images, visible light inspection images, and equipment nameplate photos, the knowledge entry management module first preprocesses the image, including size normalization and color space conversion, and then inputs the processed image into the image encoder to obtain the corresponding semantic vector.

[0059] The hybrid index module is specifically used for: The graph index is constructed using a sharded-hierarchical navigable small-world graph structure. The top layer of the graph index is divided into multiple shards according to project stages. Each shard stores nodes using a navigable small-world graph structure. The semantic vector and the time sequence key stored in each node serve as the attributes of that node.

[0060] The hybrid indexing module employs a sharded-hierarchical navigable small-world graph structure to construct the graph index. This structure, combining sharding strategies with a hierarchical graph structure, achieves efficient management and querying of large-scale knowledge bases while maintaining retrieval accuracy. Specifically, the hybrid indexing module first divides the top layer of the graph index into multiple shards based on project phases. Project phases include at least one of the following: planning, construction, operation and maintenance, and decommissioning phases, with each shard corresponding to a specific project phase. Storing knowledge from these different phases in different shards achieves both physical and logical isolation, preventing cross-phase interference and facilitating phase-based index maintenance and lifecycle management.

[0061] Within each shard, the hybrid indexing module employs a shard-hierarchical navigable small-world graph (HNSW) structure to store nodes. The shard-hierarchical navigable small-world graph structure is a graph-based approximate nearest neighbor index structure. It constructs a multi-layered graph structure, where the bottom layer contains all nodes, and the higher layers serve as navigation indexes for the bottom layer. Efficient nearest neighbor retrieval is achieved through layer-by-layer search from the higher to the lower layers. Specifically, in the HNSW structure of each shard, each node corresponds to a knowledge entry processed by the knowledge entry management module. The semantic vector and temporal key stored in each node serve as the node's attributes. The semantic vector is used to calculate the similarity between nodes in the HNSW graph, thereby constructing edge connections between nodes; the temporal key is stored as the node's condition attribute, used for condition matching filtering during retrieval. In addition to the semantic vector and temporal key, each node also stores the knowledge DNA tag generated by the knowledge entry management module and initial lifecycle parameters. These additional attributes together constitute the node's complete metadata. When constructing the HNSW graph, the hybrid indexing module uses a greedy search and heuristic edge selection strategy based on the semantic vectors of each node to establish edge connections between each node and several semantically nearest neighbor nodes, forming a multi-layered graph structure. It can be understood that this invention first limits the retrieval scope of the knowledge index to the shards corresponding to the project stage, and then performs efficient graph traversal search within the HNSW graph of each shard, ensuring both retrieval accuracy and significantly improving retrieval efficiency.

[0062] In an optional implementation, a consistency verification module is further included, specifically used for: Obtain the current running data signature from the timing processing module.

[0063] The consistency verification module operates at a preset time period or is triggered by an event. First, it obtains the current operational data signature from the time-series processing module. The operational data signature is a compact representation generated by the time-series processing module through its time-series feature extraction unit, which compresses and encodes the latest acquired time-series data window. The operational data signature highly summarizes the operating conditions of the energy storage system within the current time window. By obtaining the current operational data signature, the consistency verification module acquires a digital representation of the system's current real-time state.

[0064] The knowledge statement signature of the target knowledge entry is read from the hybrid index module. The knowledge statement signature is generated by encoding the condition-conclusion text in the knowledge entry.

[0065] The consistency verification module reads the knowledge statement signature of the target knowledge entry from the hybrid index module. The knowledge statement signature is a compact representation generated by encoding the condition-conclusion text in the knowledge entry. Its generation process is independent of the semantic vector of the knowledge entry and is specifically used to represent the applicable working conditions of the knowledge. Specifically, the consistency verification module integrates a knowledge statement signature generator. First, it parses the text content of the target knowledge entry, using a combination of rule matching and natural language processing dependency parsing to extract the "condition-conclusion" structure pairs from the knowledge text. The "condition" describes the applicable working conditions of the target knowledge entry, and the "conclusion" describes the appropriate action or judgment under those conditions. Subsequently, the generator converts the extracted condition text into a fixed-length binary signature using hash encoding or a locality-sensitive hashing algorithm, serving as the knowledge statement signature of the knowledge entry. The knowledge statement signature is stored in association with the corresponding node attribute of the hybrid index module.

[0066] Calculate the consistency score between the running data signature and the knowledge statement signature.

[0067] The consistency verification module calculates a consistency score between the runtime data signature and the knowledge statement signature. This consistency score quantifies the degree of matching between the current real-time operating conditions and the pre-defined applicable operating conditions of the knowledge entry. Preferably, the consistency verification module uses two complementary metrics to calculate the consistency score: first, it calculates the Jaccard coefficient between the runtime data signature and the knowledge statement signature, which is the ratio of the number of bits simultaneously equal to 1 in both binary signatures to the number of bits in their union, reflecting the degree of overlap in their feature dimensions; second, it calculates the KL divergence between the time-series data distribution corresponding to the runtime data signature and the operating condition distribution implied by the knowledge statement signature, measuring the degree of difference between the two probability distributions. The consistency verification module combines the calculation results of the Jaccard coefficient and the KL divergence into a consistency score, or sets independent thresholds for each.

[0068] If the consistency score is lower than the second threshold, the target knowledge entry is marked as mismatched, and the large model is invoked to generate revision suggestions for the target knowledge entry based on the time series data at the current moment.

[0069] If the consistency score is lower than a preset second threshold, it indicates a significant mismatch between the current real-time operating conditions and the applicable operating conditions of the target knowledge item, meaning the target knowledge item may have become outdated or no longer applicable due to changes in operating conditions. In this case, the consistency verification module marks the target knowledge item as "mismatched" and initiates an automatic revision process. Specifically, the consistency verification module calls a large language model to generate revision suggestions for the target knowledge item based on the current time-series data. During the call, the consistency verification module constructs a prompt containing three parts: the first part is the original "condition-conclusion" text of the target knowledge item; the second part is the key time-series data features reflecting the current operating conditions obtained from the time-series processing module; and the third part is a revision instruction, requiring the large language model to determine whether the original knowledge still holds true based on the new operating condition data. If not, it generates a revised "new condition-new conclusion" suggestion. The revision suggestions returned by the large language model include, but are not limited to, types such as "retain the original knowledge," "revise the operating conditions of the original knowledge," "revise the disposal conclusion of the original knowledge," and "abandon the original knowledge." For revision suggestions, specific revision content is also included.

[0070] Update the corresponding nodes in the hybrid index module according to the proposed revisions.

[0071] The consistency verification module updates the corresponding node in the hybrid index module based on the revision suggestions. If the revision suggestion is "Revised" and has a high confidence level, the consistency verification module directly updates the content of the target knowledge node, including modifying its knowledge statement signature, semantic vector, and related knowledge DNA tags, and updating its lifecycle parameters to reflect the revised knowledge validity period. If the revision suggestion is "Discarded," the consistency verification module marks the target knowledge node as discarded, but retains its historical version tracking chain for auditing. If the revision suggestion has a low confidence level, the consistency verification module pushes the revision suggestion to the manual review queue for final confirmation by the operations engineer.

[0072] In an optional implementation, the question-and-answer interface module is further configured to: Receive user questions and obtain the timing key of the current working condition from the timing processing module.

[0073] The question-and-answer interface module receives questions input by users in natural language (i.e., user questions). User questions can be text input or text converted from speech. After receiving a user question, the question-and-answer interface module immediately sends a request to the timing processing module to obtain the timing key reflecting the current real-time operating status of the energy storage system.

[0074] Using the graph index of the hybrid index module, the top K knowledge entries with the highest semantic similarity to the user's question are retrieved, and the search results are filtered according to the time sequence key of the current working condition, retaining the knowledge entries that match the time sequence key; k is an integer greater than 0.

[0075] The question-answering interface module first calls the same semantic vectorization model used during index construction to convert the user's question into a question semantic vector. Then, the module performs an approximate nearest neighbor search in the graph index of the hybrid index module, retrieving the top K knowledge entries with the highest semantic similarity to the user's question, where K is a positive integer, and its value can be preset according to the actual application scenario. Since the graph index adopts a sharded-hierarchical navigable small-world graph structure, the retrieval process first locates the corresponding shard based on the current user's project stage or default configuration, and then performs an efficient graph traversal search in the HNSW graph of that shard to quickly locate the K nodes with the highest semantic similarity. After completing the initial retrieval, the module further filters the retrieval results based on the time sequence key of the current work condition. Specifically, the module traverses the initially retrieved K knowledge nodes, comparing the time sequence key stored in the node attributes with the time sequence key of the current work condition, for example, calculating their Jaccard coefficients or determining whether they belong to the same work condition cluster. Only those knowledge nodes whose time sequence keys match the current working conditions are retained as the final search results.

[0076] The retrieved knowledge entries are concatenated with the user's question to form prompts, and the large model is then used to generate the answer.

[0077] The question-answering interface module concatenates the retrieved knowledge entries with the user's question to form prompt words, and then calls the large language model to generate the answer. Specifically, the question-answering interface module constructs a structured prompt word template, which includes three parts: the first part is the original question posed by the user; the second part is the content of several retrieved knowledge entries, each presented in its original text form; and the third part is an instruction text, requiring the large language model to answer strictly according to the provided knowledge entries, without introducing external knowledge. The question-answering interface module sends the above prompt words to the large language model and receives the generated answer returned by it.

[0078] The answer is returned to the user, along with the knowledge ID of the knowledge entry on which the answer is based and the consistency score calculated by the consistency verification module.

[0079] After generating the answer, the question-and-answer interface module returns the answer to the user, along with the knowledge ID of the knowledge entry on which the answer is based and the consistency score calculated by the consistency verification module. Specifically, the question-and-answer interface module also includes a list of one or more knowledge IDs in the response result, each knowledge ID uniquely identifying a knowledge entry on which the answer was based; simultaneously, for each referenced knowledge entry, the question-and-answer interface module reads its latest consistency score (i.e., the Jaccard coefficient or composite score between the running data signature and the knowledge statement signature) from the consistency verification module's cache and displays the consistency score to the user.

[0080] The system receives user feedback on the answer, including correct and incorrect feedback. Upon receiving incorrect feedback, the system marks the knowledge entry corresponding to the incorrect feedback as pending review and re-verifies the consistency of the knowledge entry through the consistency verification module.

[0081] On the answer display interface, the Q&A interface module provides users with two interactive buttons: "Correct" and "Incorrect." Users can evaluate the accuracy of the answer based on the actual situation. When a user clicks the "Correct" button, the Q&A interface module receives correct feedback and records it to the knowledge item management module. This feedback is used to improve the confidence tag of the knowledge item or extend its lifecycle parameter, forming positive reinforcement. When a user clicks the "Incorrect" button, the Q&A interface module receives incorrect feedback and triggers the knowledge review process. Specifically, upon receiving the incorrect feedback, the Q&A interface module marks the knowledge item corresponding to the incorrect feedback as "pending review" and sends the identification information of the knowledge item and the error type reported by the user to the consistency verification module. After receiving the trigger signal, the consistency verification module re-verifies the consistency of the knowledge item, including obtaining the latest runtime data signature from the time-series processing module, recalculating the consistency score between the runtime data signature and the knowledge item's knowledge statement signature, and calling the large language model to generate revision suggestions if the score is lower than a threshold. If the revision suggestion is to revise or discard, the consistency verification module updates the corresponding node in the hybrid index module.

[0082] In one optional implementation, a federated aggregation relay module is further included; the federated aggregation relay module is connected to the system deployed locally by multiple energy storage power stations; the federated aggregation relay module is specifically used for: The system receives knowledge evolution summaries and knowledge statement signatures uploaded by each energy storage power station. The knowledge evolution summary includes a symptom description, a condition description, and a conclusion description of the knowledge item. The knowledge statement signature is generated by encoding the condition-conclusion text in the knowledge item.

[0083] The federated aggregation relay module acts as a cloud hub connecting the systems deployed locally at multiple energy storage power stations. It establishes secure communication links with these systems via a network. Each local system includes the aforementioned time-series processing module, hybrid indexing module, incremental update control module, question-and-answer interface module, knowledge entry management module, drift detection module, and consistency verification module. The federated aggregation relay module first receives knowledge evolution summaries and knowledge statement signatures uploaded by each energy storage power station. The knowledge evolution summary describes the core content of a knowledge entry, summarizing its key elements in structured text form, specifically including symptom descriptions, condition descriptions, and conclusion descriptions. The symptom description characterizes the abnormal phenomena or operating characteristics of the equipment targeted by the knowledge entry, such as "PCS overvoltage alarm," "cell temperature difference exceeding 5℃," and "abnormal SOC drop rate." The condition description characterizes the specific operating conditions to which the knowledge entry applies, including the state of charge range, rate range, temperature range, and equipment model. The conclusion description characterizes the measures or judgments to be taken under the given symptom and conditions, such as "reduce the overvoltage protection setting to 1070V," "start the equalization charging program," and "reduce power to 0.3C operation." The knowledge statement signature is a compact representation generated by the local system's consistency verification module by encoding the condition-conclusion text in the knowledge entry.

[0084] The similarity of multiple received knowledge statement signatures is calculated, and knowledge entries with similarity higher than the third threshold are classified as candidate consensus knowledge.

[0085] The federated aggregation relay module calculates the similarity of multiple received knowledge statement signatures. Specifically, it uses metrics such as Jaccard coefficients or Hamming distance to calculate the similarity between knowledge statement signatures of different knowledge items. Because knowledge statement signatures are in binary form, similarity calculation is highly efficient, enabling rapid pairwise comparisons across large-scale cross-site data. The federated aggregation relay module categorizes knowledge items with similarity scores higher than a preset third threshold as candidate consensus knowledge. This third threshold can be adjusted based on the specific application scenario.

[0086] The candidate consensus knowledge is subjected to causal consistency verification, and the candidate consensus knowledge that passes the verification is merged into industry common knowledge.

[0087] Since signature similarity alone cannot guarantee the causal correctness and cross-site generalization ability of knowledge, the federated aggregation relay module further verifies the causal consistency of candidate consensus knowledge. Specifically, the federated aggregation relay module uses a structural causal model or a causal inference method based on the DoWhy framework to verify the causal relationship between conditions and conclusions in candidate consensus knowledge. During the verification process, the federated aggregation relay module comprehensively utilizes the symptom description, condition description, and conclusion description in the knowledge evolution summary uploaded by each site, as well as the causal verification records generated by each site during the knowledge evolution process (such as the results of passing local DoWhy causal verification), to construct a cross-site causal graph model. The causal verification engine calculates the average causal effect and its statistical significance to determine whether the causal relationship between the condition and the conclusion is robustly present in different environmental contexts across multiple sites. If the causal relationship passes the verification, it indicates that the knowledge has cross-site universality. The candidate consensus knowledge that passes the verification is merged into industry-wide common knowledge. The merging process includes content fusion of multiple knowledge evolution summaries to generate a comprehensive unified knowledge description covering the applicable conditions of multiple sites, and generating a unified industry-wide common knowledge ID for it.

[0088] The industry-wide knowledge is sent to each energy storage power station so that each energy storage power station stores the industry-wide knowledge in its own hybrid index module.

[0089] The federated aggregation relay module sends industry-wide knowledge to each energy storage power station, enabling each station to store this knowledge in its respective hybrid index module. Upon receiving the industry-wide knowledge, each power station's local knowledge entry management module processes it according to a standard inbound process: generating a knowledge DNA tag for the knowledge (where the device model tag can be marked as "general" or "applicable to multiple models"), calculating initial lifecycle parameters, calling the vectorization engine to generate semantic vectors, obtaining the time sequence key for the current operating condition, and associating and storing this information in the graph index of its local hybrid index module, forming a new knowledge node.

[0090] Knowledge entries with similarity below the third threshold or that fail the causal consistency check are retained as station-level specialized knowledge for each energy storage power station.

[0091] For knowledge entries with similarity below the third threshold or failing the causal consistency check, the federated aggregation relay module retains them as station-level specialized knowledge for each energy storage power station. Specifically, these knowledge entries, because they appear only at a few stations, exhibit significant differences in their manifestation across different stations, or whose causal relationships cannot be robustly verified at the cross-site level, do not meet the conditions for forming industry-wide common knowledge and are therefore not merged and distributed. This station-level specialized knowledge continues to be retained in the local hybrid index module of each station as a unique accumulation of experience for that station, serving the personalized operation and maintenance needs of that station. This can be understood as common knowledge being aggregated in the cloud and differentiated knowledge being retained locally.

[0092] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.

[0093] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0094] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A time-series-semantic dual-channel energy storage RAG knowledge base system, characterized in that, include: The timing processing module is used to acquire the timing data of the energy storage power station in real time and generate the timing key of the corresponding operating condition based on the timing data. The hybrid index module is used to associate and store the semantic vector of each knowledge entry in the knowledge base with the time sequence key of the corresponding working condition to construct a graph index. The graph index contains multiple nodes and edges between each node. Each node corresponds to a knowledge entry, and each node stores the semantic vector of the corresponding knowledge entry and the time sequence key of the corresponding working condition. The incremental update control module includes: A similarity judgment unit is used to receive new knowledge entries and calculate the similarity between the semantic vector of the new knowledge entry and the semantic vector of existing nodes in the graph index. The differential replacement unit is used to calculate the difference between the semantic vector of the new knowledge entry and the semantic vector of the existing node when the similarity output by the similarity judgment unit is greater than a first threshold, and update the semantic vector of the existing node with the difference. The question-and-answer interface module is used to receive user questions, obtain the time sequence key of the current working condition, and use the graph index to retrieve knowledge entries that are semantically related to the user question and match the time sequence key of the current working condition, and generate the answer to the user question.

2. The system according to claim 1, characterized in that, It also includes a drift detection module; the incremental update control module further includes an edge update unit; The drift detection module is used to monitor the distribution changes of the time series data acquired by the time series processing module, and send a trigger signal to the edge update unit when a data distribution drift is detected. The edge update unit is specifically used for: In response to the trigger signal sent by the drift detection module, the affected nodes in the graph index are identified; The connected region formed by the affected nodes and their associated edges in the graph index is taken as the subgraph to be reconstructed, and the edges between each node in the subgraph are recalculated to update the edge relationships in the subgraph.

3. The system according to claim 2, characterized in that, The edge update unit is specifically used for: In response to the trigger signal sent by the drift detection module, the affected nodes in the graph index are determined; The connected region formed by the affected nodes and their associated edges is divided out from the graph index and used as a subgraph to be reconstructed. The number of nodes in the subgraph does not exceed a preset proportion of the total number of nodes in the entire database. Create a copy of the graph index as a shadow index, and recalculate the edge relationships within the subgraph in the shadow index; Use the shadow index as the new graph index to update the edge relationships within the subgraph.

4. The system according to claim 2, characterized in that, The drift detection module is specifically used for: The time series data obtained by the time series processing module is divided into univariate time series data and multivariate time series data; The Page-Hinkley algorithm is used to monitor univariate time-series data, which includes at least one of cell voltage, cell temperature, PCS power, and cooling flow rate. The maximum mean difference algorithm is used to monitor drift in multivariate time series data, which includes battery operating condition variables, PCS operating condition variables, and cooling operating condition variables. If either the Page-Hinkley algorithm or the maximum mean difference algorithm triggers an alarm, it confirms that a data distribution drift has been detected.

5. The system according to claim 1, characterized in that, It also includes a knowledge entry management module, which is specifically used for: Receive knowledge entries to be added to the database, and generate knowledge DNA tags for the knowledge entries. The knowledge DNA tags include project stage tags, equipment model tags, operating condition range tags, and document confidence tags. The initial lifecycle parameters of the knowledge entry are calculated based on the knowledge DNA tag. The knowledge DNA tag, initial lifecycle parameters, semantic vector of the knowledge entry, and time sequence key of the corresponding working condition of the knowledge entry are associated and stored in the corresponding node of the hybrid index module.

6. The system according to claim 5, characterized in that, It also includes a drift detection module, which is used to monitor the distribution changes of the time series data acquired by the time series processing module, and send a trigger signal to the knowledge entry management module when a data distribution drift is detected; The knowledge entry management module is also used for: In response to the trigger signal sent by the drift detection module, the affected time series data range is determined according to the drift period, and the original knowledge entries related to the drift period are retrieved from the hybrid index module; The large model is invoked to generate new candidate knowledge based on the original knowledge entries and the time-series data. The new candidate knowledge includes new conditions and new conclusions. A causal inference algorithm is used to verify the causal relationship between the new conditions and the new conclusions in the new candidate knowledge; The verified new candidate knowledge is added to the hybrid index module as a new knowledge entry, and the corresponding original knowledge entry is marked as obsolete.

7. The system according to claim 5, characterized in that, The knowledge entry management module is also used for: Determine the file type of the knowledge entry to be added to the database; When the knowledge entry is text-based, a text encoder is used to convert the text content of the knowledge entry into a semantic vector; When the knowledge entry is an image, an image encoder is used to convert the image content of the knowledge entry into a semantic vector; The hybrid index module is specifically used for: The graph index is constructed using a sharded-hierarchical navigable small-world graph structure. The top layer of the graph index is divided into multiple shards according to project stages. Each shard stores nodes using a navigable small-world graph structure. The semantic vector and the time sequence key stored in each node serve as the attributes of that node.

8. The system according to claim 1, characterized in that, It also includes a consistency verification module, which is specifically used for: Obtain the current running data signature from the timing processing module; The knowledge statement signature of the target knowledge entry is read from the hybrid index module. The knowledge statement signature is generated by encoding the condition-conclusion text in the knowledge entry. Calculate the consistency score between the runtime data signature and the knowledge statement signature; If the consistency score is lower than the second threshold, the target knowledge item is marked as mismatched, and the large model is invoked to generate revision suggestions for the target knowledge item based on the time series data at the current time. Update the corresponding nodes in the hybrid index module according to the proposed revisions.

9. The system according to claim 8, characterized in that, The question-and-answer interface module is also used for: Receive user questions and obtain the timing key of the current working condition from the timing processing module; Using the graph index of the hybrid index module, the top K knowledge entries with the highest semantic similarity to the user's question are retrieved, and the search results are filtered according to the time sequence key of the current working condition, retaining the knowledge entries that match the time sequence key; k is an integer greater than 0; The retrieved knowledge items are concatenated with the user's question to form prompts, and the large model is then used to generate the answer. The answer is returned to the user, along with the knowledge ID of the knowledge entry on which the answer is based and the consistency score calculated by the consistency verification module. Receive user feedback on the answer, including correct feedback and incorrect feedback; Upon receiving the error feedback, the knowledge item corresponding to the error feedback is marked as pending review, and the consistency verification module is used to re-verify the consistency of the knowledge item.

10. The system according to claim 1, characterized in that, It also includes a federated aggregation relay module; the federated aggregation relay module is connected to the system deployed locally by multiple energy storage power stations; the federated aggregation relay module is specifically used for: The system receives knowledge evolution summaries and knowledge statement signatures uploaded by each energy storage power station. The knowledge evolution summary includes a symptom description, condition description, and conclusion description of the knowledge item. The knowledge statement signature is generated by encoding the condition-conclusion text in the knowledge item. The similarity of multiple received knowledge statement signatures is calculated, and knowledge entries with similarity higher than the third threshold are classified as candidate consensus knowledge. The candidate consensus knowledge is subjected to causal consistency verification, and the candidate consensus knowledge that passes the verification is merged into industry common knowledge. The industry-specific knowledge is sent to each energy storage power station so that each energy storage power station stores the industry-specific knowledge in its own hybrid index module; Knowledge entries with similarity below the third threshold or that fail the causal consistency check are retained as station-level specialized knowledge for each energy storage power station.

Images