A large model driven building knowledge graph construction method and system

Abstract

This invention discloses a large-model-driven method and system for constructing a building knowledge graph. It generates a terminology feature library by differentially annotating unstructured building document data with semantics; based on the terminology feature library, it detects cross-document conflicts to generate a set of conflicting knowledge points; and constructs a building knowledge graph through cross-source semantic difference quantification and difference-oriented knowledge fusion using a large language model. It identifies knowledge backlog nodes and marks weak knowledge areas through topological structure analysis, and injects graph knowledge vectors in a targeted manner to generate a knowledge-enhanced control model. Based on the knowledge-enhanced control model, it executes bidirectional causal reasoning of the HVAC system to generate a decision reasoning chain, and outputs interpretable control decisions after violation of aggregation patterns identification and weight reduction processing. It generates behavioral deviation values ​​based on the comparison between equipment behavior data and the deviation of the decision reasoning chain, and triggers adaptive reconstruction of the building knowledge graph through hierarchical over-limit identification, achieving continuous accumulation of building knowledge and steady improvement in the quality of HVAC control decisions.

Description

Technical Field

[0001] This invention relates to the fields of building intelligence and knowledge engineering technology, and in particular to a method and system for constructing a building knowledge graph driven by a large model. Background Technology

[0002] During the entire life cycle operation of buildings, a large amount of documents and data from various stages such as design and construction, equipment operation and maintenance, and energy consumption management are accumulated. These documents are scattered across different departments and systems, with different formats and inconsistent language styles. The same equipment or system often appears in different descriptions in documents from different sources, making it difficult for building managers to efficiently extract consistent and accurate equipment relationships and operating rules. As a result, the utilization rate of knowledge assets is generally low.

[0003] Building operations involve complex interdisciplinary collaborations among HVAC, electrical, fire protection, and water supply and drainage systems. Existing building management methods often rely on manual experience or simple rule engines for equipment control decisions, making it difficult to dynamically perceive the causal dependencies between these systems. When equipment malfunctions or operating conditions change, rule engines cannot adapt to the dynamic evolution of the knowledge structure, compromising the rationality and interpretability of control decisions and impacting the overall energy efficiency and operational quality of the building. Summary of the Invention

[0004] This invention discloses a large-model-driven method and system for constructing a building knowledge graph. Addressing issues such as inconsistent terminology descriptions, sparse cross-disciplinary knowledge structures, and lack of interpretability in control decisions within multi-source heterogeneous building documents, the invention constructs a building knowledge graph through differentiated semantic annotation and difference-oriented knowledge fusion. Based on graph topology analysis, it selectively reinforces weak knowledge areas, drives bidirectional causal reasoning to output interpretable control decisions, and triggers adaptive graph reconstruction based on hierarchical over-limit identification of deviations between actual equipment behavior and inference expectations. This achieves continuous accumulation of building knowledge and a steady improvement in the quality of HVAC control decisions.

[0005] The first aspect of this invention proposes a method for constructing a building knowledge graph driven by a large model, comprising the following steps: Collect unstructured document data of buildings, and perform text cleaning and domain semantic annotation on the unstructured document data to generate a term feature library; Based on the terminology feature library, multi-source description conflict detection is performed on the unstructured building document data to generate a conflict knowledge point set. The conflict knowledge point set is then input into a large language model to perform cross-source semantic difference quantification to generate a semantic difference matrix. Finally, difference-oriented knowledge fusion is performed on the semantic difference matrix to construct a building knowledge graph. Extract graph knowledge vectors from the building knowledge graph, identify knowledge backlog nodes and determine knowledge-weak areas based on the topological structure of the building knowledge graph, and inject the graph knowledge vectors into the knowledge-weak areas to generate a knowledge enhancement control model; Based on the knowledge-enhanced control model, bidirectional causal reasoning of the HVAC system is executed to generate a decision reasoning chain. Direction violation detection is performed on the decision reasoning chain to generate low-confidence nodes. The low-confidence nodes are then deweighted to output an interpretable control decision. Based on the interpretable control decision-driven operation of the HVAC system, the system generates equipment behavior data. The equipment behavior data is compared with the decision reasoning chain to generate a behavior deviation value. Based on the behavior deviation value, the system performs hierarchical over-limit identification to trigger the adaptive reconstruction of the building knowledge graph and complete the construction of the building knowledge graph.

[0006] A second aspect of this invention proposes a large-model-driven building knowledge graph construction system, comprising: The corpus annotation module is used to collect unstructured building document data, and to perform text cleaning and domain semantic annotation on the unstructured building document data to generate a term feature library; The knowledge graph construction module is used to perform multi-source description conflict detection on the unstructured building document data based on the terminology feature library to generate a conflict knowledge point set, input the conflict knowledge point set into a large language model to perform cross-source semantic difference quantification to generate a semantic difference matrix, and perform difference-oriented knowledge fusion on the semantic difference matrix to construct a building knowledge graph. The knowledge injection module is used to extract graph knowledge vectors from the building knowledge graph, identify knowledge backlog nodes and determine knowledge-weak areas based on the topological structure of the building knowledge graph, and inject the graph knowledge vectors into the knowledge-weak areas to generate a knowledge enhancement control model. The reasoning and decision-making module is used to perform bidirectional causal reasoning of the HVAC system based on the knowledge-enhanced control model to generate a decision reasoning chain, perform directional violation detection on the decision reasoning chain to generate low-confidence nodes, and perform weight reduction processing on the low-confidence nodes to output an interpretable control decision. The knowledge graph reconstruction module is used to generate equipment behavior data based on the interpretable control decision-driven operation of the HVAC system, compare the equipment behavior data with the decision reasoning chain to generate behavior deviation values, and perform hierarchical over-limit identification based on the behavior deviation values ​​to trigger the adaptive reconstruction of the building knowledge graph to complete the construction of the building knowledge graph.

[0007] The beneficial effects of this invention are reflected in the following points: 1. In multi-source documents of a building, the same equipment terminology assumes different semantic roles due to different document types. Existing knowledge graph construction methods lack systematic identification and quantitative evaluation methods for such cross-document role conflicts. This invention establishes a quantitative association between terminology and fault semantics by applying a differentiated annotation strategy to abnormal and normal corpora. On this basis, it uses a terminology role distribution matrix to drive cross-document conflict intensity evaluation. The high-conflict knowledge point set is quantified for cross-source semantic differences through a large language model to generate a semantic difference matrix. Through difference-oriented knowledge fusion, the differences in descriptions from multiple sources are transformed into the basis for graph construction weights, eliminating the interference of inconsistent descriptions from multiple sources on the accuracy of the graph's causal structure from the source. 2. Unlike existing methods that uniformly process all regions of the knowledge graph, this invention introduces dual indicators of in-degree-out-degree ratio and propagation coverage insufficiency rate to actively locate knowledge accumulation nodes in sparse cross-professional areas such as fire protection and water supply and drainage. It uses a gradient-oriented approach to directionally migrate high-density feature dimensions of strong professional domains to weak areas, thereby providing targeted reinforcement to the reasoning coverage capability of the knowledge enhancement control model under cross-professional linkage conditions. 3. This invention establishes a complete closed loop from control decision output to adaptive knowledge graph update: the knowledge-enhanced control model generates interpretable control decisions carrying complete causal evidence chains through the dual channels of forward control reasoning and reverse causal tracing. The severity of the missing response between the actual behavior of the equipment and the expected reasoning is mapped to the behavior deviation value through graded weighting. The graded over-limit identification adaptively determines the reconstruction strategy based on the deviation level, so that the knowledge graph continuously and dynamically aligns with the actual operating state of the HVAC system rather than remaining in static construction. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a large-model-driven method for constructing a building knowledge graph according to the present invention.

[0009] Figure 2 This is a structural block diagram of a large model-driven building knowledge graph construction system according to the present invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0012] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0013] The technical solutions of the embodiments of this application will be described below.

[0014] like Figure 1 As shown, this embodiment of the invention provides a method for constructing a building knowledge graph driven by a large model, including the following steps S110-S150: Step S110: Collect unstructured building document data, perform text cleaning and domain semantic annotation on the unstructured building document data to generate a terminology feature library.

[0015] Specifically, unstructured building document data is collected. This data is gathered from various sources, including equipment operation and maintenance records, construction acceptance reports, energy consumption inspection logs, equipment failure work orders, and HVAC system design specifications. The formatting and language styles of these documents vary significantly. Operation and maintenance records account for 55%-65% of the total unstructured building document data, primarily using colloquial and abbreviated language, while design specifications use standardized terminology. This coexistence of styles results in the same equipment concept often appearing in three to five different expressions, directly interfering with subsequent terminology annotation and knowledge extraction. The data collection channels for unstructured building documents include three types: building automation system export interfaces, work order management platform APIs, and scanned engineering archives. These three channels exhibit differences in data completeness and timeliness. Data exported from automation systems offers the strongest real-time performance but has the loosest structure, while scanned engineering archives contain the most authoritative content but are susceptible to OCR transcription errors. The collection module configures a unified formatting strategy for each of the three sources before storing the data uniformly. Noise filtering and semantic cleaning are performed centrally in the subsequent text cleaning stage. After the unstructured document data of the building is collected, it is organized and stored by dual indexes according to the document source type and the professional domain to which it belongs, covering four major professional domains: HVAC, electrical, fire protection and water supply and drainage. The document data collection volume of a medium-sized building is usually 800-2000 documents, totaling 1.5-4 million words. When the number of documents is less than 500, the collection integrity is insufficient to support the statistical reliability of subsequent cross-document conflict detection, and historical archived documents need to be added to expand the collection scope.

[0016] In some embodiments, the step of performing text cleaning and domain semantic annotation on the unstructured building document data to generate a terminology feature library includes: performing noise filtering and format normalization on the unstructured building document data to generate a cleaned corpus; identifying abnormal descriptive paragraphs using the cleaned corpus to generate an abnormal corpus set and a normal corpus set; performing priority semantic annotation on the abnormal corpus set and de-weighting annotation on the normal corpus set to generate a differentiated annotated corpus; and extracting and encoding terms based on the differentiated annotated corpus to generate a terminology feature library.

[0017] Noise filtering and format normalization were performed on unstructured building document data to generate a clean corpus. Documents transcribed from scanned OCR sources have a character-level recognition error rate of approximately 3%-8%, with errors concentrated in equipment number paragraphs containing a mix of numbers and letters. These paragraphs account for about 12% of the unstructured building document data and are the key area for noise filtering. The noise filtering rule base contains four categories of rules: filtering meaningless symbol sequences, deduplication of repeated paragraphs, replacement of OCR error characters, and removal of garbled paragraphs. These four categories of rules are used to scan unstructured building document data segment by segment in priority order. Each segment of text is processed by all rules before entering the format normalization stage. Format normalization converts full-width punctuation to half-width, replaces variant Chinese characters with standard characters, and uniformly removes spaces before numerical units. The normalization operation is based on a character replacement mapping table. The quality assessment of the cleaned corpus is measured by the paragraph-level readability index. Paragraphs with a readability index below the threshold of 0.75 are marked as low-quality. These low-quality marked paragraphs have their classification confidence weights automatically reduced when fed into the pre-trained language model during the anomaly identification phase. The classification results of low-quality marked paragraphs are not included in the statistical baseline for the normal corpus. The effective text retention rate after noise filtering of unstructured building documents is typically 88%-94%. A retention rate below 85% triggers a document quality warning and prompts a backtracking check of data entry standards. The cleaned corpus is stored in partitions according to document source category. The number of paragraphs and the total number of tokens in each partition are written into the metadata index of the cleaned corpus, forming the data basis for the partition-level statistical baseline during the anomaly description paragraph identification phase.

[0018] Anomaly description paragraphs are identified by cleaning the corpus, generating both anomaly and normal corpora. The negative word density in the cleaned corpus is the most significant quantitative difference between anomaly and normal paragraphs; the negative word density in paragraphs from fault work orders is 2.3-3.1 times that of normal operation records. This density difference constitutes the core discriminative signal for anomaly description paragraph identification. Anomaly description paragraph identification uses a pre-trained language model to perform binary classification inference on each paragraph in the cleaned corpus. The classification confidence threshold is set to 0.82. Paragraphs with a confidence level higher than the threshold and a predicted label of "anomaly" are included in the anomaly corpus, while the remaining paragraphs are included in the normal corpus. The anomaly identification recall rate for three paragraph types in the cleaned corpus—descriptions of equipment parameters exceeding limits, explanations of fault codes, and maintenance operation records—is typically above 92%. However, the recall rate for critical state descriptions in daily equipment inspection records is relatively low, at approximately 76%. The latter requires supplementary identification using a pre-set critical state keyword list. The ratio of the abnormal corpus to the normal corpus is approximately 1:4 to 1:6 in typical building document sets. When the ratio is below 1:8, paragraph-level data augmentation is performed on the abnormal corpus. The augmentation method is synonym replacement. The replacement words are derived from high-frequency device synonym pairs obtained from the cleaned corpus after noise filtering and format normalization. Synonym pairs are identified based on an edit distance of less than 2 and a semantic similarity higher than 0.85. Both types of corpora retain their original document source annotations. These source annotations are used in the differential annotation stage to trace the correlation between annotation confidence and document channel, providing numerical basis for calculating the source weights of the terminology feature library.

[0019] Prioritized semantic annotation is performed on the abnormal corpus, while de-weighted annotation is applied to the normal corpus to generate differentiated annotated corpora. Segments in the abnormal corpus are assigned to a priority processing queue, with the annotation throughput of the priority queue set to three times that of the normal queue, ensuring sufficient annotation input for the abnormal corpus despite limited overall annotation resources. Prioritized semantic annotation employs a fine-grained annotation strategy for fault state descriptors and equipment anomaly parameter quantifiers in the abnormal corpus. Fine-grained annotation labels fault state descriptors word-by-word according to two scenarios: single-device and multi-device concurrent. Multiple anomaly descriptions occurring concurrently within the same segment are annotated with entity relationships to record the direction of influence transmission between devices. De-weighted annotation of the normal corpus is performed using a coarse-grained strategy, only annotating top-level categories for equipment nouns and operating parameter quantifiers, without distinguishing subcategories. The annotation time for each segment of normal corpus text is approximately 35% of that for the corresponding segment in the abnormal corpus. This differentiated annotation input concentrates limited resources on the abnormal corpus, which has a greater impact on the quality of the terminology feature library. The differentiated annotated corpus is formed by merging the annotation results of the abnormal corpus set and the normal corpus set. During the merging process, the consistency of the annotation labels for the same term in both corpora is checked. Term entries with a label conflict rate exceeding 15% are marked as pending review. Entries pending review have their weight reduced during the terminology feature library construction phase. The annotation density ratio of abnormal paragraphs to normal paragraphs in the differentiated annotated corpus is approximately 2.8:1 on a typical medium-sized building document set. This ratio is a core indicator for measuring the effectiveness of the differentiated annotation strategy. If it deviates from the range of 1.5-4.0, the annotation strategy parameters for the two corpora sets need to be adjusted and the annotation process re-executed.

[0020] A terminology feature library is generated based on differentially annotated corpora. When a device term has an annotation density of 4.2 times per 1000 words and accounts for 68% of the abnormal corpus, it is identified as a high-value fault-related term, and its encoded vector receives a high-weight label in the terminology feature library. Term extraction uses a conditional random field model to decode the differentially annotated corpus. The decoded candidate term sequences are filtered by confidence, retaining entries with a confidence score higher than 0.78. Confidence is calculated based on the weighted average of the model output probability and the annotation consistency score, with a weight ratio of 0.6:0.4. Before entering the encoding stage, the candidate term sequences undergo a three-level deduplication and merging process: complete string matching for deduplication, merging of similar words with an edit distance less than 2, and aggregation of synonyms with a semantic similarity higher than 0.92. This three-level merging process compresses the redundancy rate of the candidate term sequences from the original 23%-31% to less than 8%. Term encoding uses a domain-pre-trained word vector model to generate 256-dimensional feature vectors for each term in the merged term list. Term contributions from anomalous corpora in the differentially annotated corpus are weighted with anomaly weights during encoding. The weighting formula is w_abnormal = 1 + λ × (f_abnormal / f_total), where w_abnormal is the anomaly weight, f_abnormal is the frequency of the term in the anomalous corpus, f_total is the frequency of the term in the entire differentially annotated corpus, and λ is an adjustment coefficient with a value of 1.5. Typical high-fault-related terms have w_abnormal values ​​between 1.8 and 2.3. After all terms are encoded, a snapshot of the term feature library is created. This snapshot includes two statistical summaries: the total number of term entries in the current term feature library and the distribution percentage of terms in each professional domain.

[0021] Step S120: Based on the terminology feature library, perform multi-source description conflict detection on unstructured building document data to generate a conflict knowledge point set. Input the conflict knowledge point set into a large language model to perform cross-source semantic difference quantification to generate a semantic difference matrix. Perform difference-oriented knowledge fusion on the semantic difference matrix to construct a building knowledge graph.

[0022] In some embodiments, the step of generating a conflict knowledge point set by performing multi-source description conflict detection on the unstructured building document data based on the terminology feature library includes: performing cross-document terminology role recognition on the unstructured building document data based on the terminology feature library to generate a terminology role distribution; identifying multi-role conflict locations of the same term on the terminology role distribution to generate a role conflict set; evaluating the conflict intensity of the role conflict set to generate a conflict intensity distribution; and identifying high-conflict areas based on the conflict intensity distribution to generate a conflict knowledge point set.

[0023] Based on a terminology feature library, cross-document terminology role recognition is performed on unstructured building document data to generate terminology role distribution. The same HVAC equipment term typically assumes the role of the monitored object in construction acceptance reports, while in operation and maintenance records it often appears as the operating entity. This difference in role positioning for the same term between the two document types is prevalent in unstructured building document data and is a core pattern that cross-document terminology role recognition needs to capture. The source document type distribution information carried by each term entry in the terminology feature library serves as a priori constraint for role recognition. When processing unstructured building document data, the role recognition model injects this priori constraint into the context encoding stage, allowing the model to simultaneously consider the document source type and local syntactic features when judging the term role. Cross-document terminology role recognition outputs a role label for each term occurrence position in the unstructured building document data. The role label set includes five categories: master control equipment, controlled equipment, monitoring object, operation action initiator, and status description object, covering all major types of roles assumed by the term. The term role distribution matrix is ​​constructed using terms as row indices and role label types as column indices. Matrix elements represent the frequency of occurrence of the term under its corresponding role label. Row vectors corresponding to high-weight term entries typically exhibit a multi-column non-zero distribution, reflecting the diversity of roles played by high-value terms across various document types. The size of the term role distribution matrix is ​​consistent with the number of valid entries in the term feature database. With 800-1500 term entries, the matrix contains 4000-9000 non-zero elements, with a non-zero element density of approximately 12%-18%. If the density is too low, the document sources need to be expanded to improve the diversity of term role observations.

[0024] For example, the step of identifying multiple role conflict positions of the same term in the term role distribution to generate a role conflict set includes: performing cross-document role comparison on each term in the term role distribution to generate a role comparison sequence; identifying document positions where the same term assumes opposite roles from the role comparison sequence to generate a role reversal position set; extracting local contextual semantic features from the role reversal position set to generate conflict context features; and confirming conflicts based on the conflict context features to generate a role conflict set.

[0025] Cross-document role comparison is performed on each term in the terminology role distribution matrix to generate a role comparison sequence. The term "chilled water pump," with three non-zero columns in the terminology role distribution matrix, is labeled as controlled equipment in the construction acceptance report, a monitoring object in the energy consumption inspection log, and an operation initiator in the maintenance record. These three non-zero columns directly trigger the cross-document role comparison process for this term. For each highly diverse term in the terminology role distribution matrix, the cross-document role comparison iterates through all role label occurrence records according to document source type. The role label sequences from each document source are concatenated into a time-ordered role comparison sequence. Each element in the sequence carries three pieces of information: document source type, paragraph position offset, and role label. The length of the role comparison sequence is consistent with the total number of times the term appears in the building's unstructured document data. High-frequency term sequences can reach 200-400 elements, while low-frequency terms are typically 15-50 elements. Terms appearing less than 10 times are marked with a low reliability flag after the role comparison sequence is generated. After generation, the role comparison sequences are grouped according to the document source type. The consistency of role labels within the same source type is measured by the mode ratio of the labels within the group. Source groups with a mode ratio of less than 0.6 are determined to have internal discrepancies in the role definition of the term in the source document. The role labels of this group are downweighted in the subsequent role inversion recognition. The downweighting coefficient is calculated proportionally based on the difference between the mode ratio and the threshold of 0.6.

[0026] Role reversal location sets are generated by identifying document locations in a role comparison sequence where the same term assumes opposite roles. For example, "fresh air handling unit" appears as a controlled device being installed and commissioned in a building construction acceptance report, but as a main control device for indoor temperature and humidity in a winter heating operation record. This type of role reversal, caused by different document perspectives during the usage and construction phases, is most concentrated in the role comparison sequence. When the cosine similarity of the role label frequency distribution vectors between two adjacent document source groups in the role comparison sequence is less than 0.25, the switching location is marked as a candidate role reversal location. The probability of role reversal is highest at the switching locations where building design documents transition to operation and maintenance documents; approximately 58% of these switching locations correspond to true role semantic reversal, while the reversal rate within the same document category is only 12%-18%. When core HVAC equipment terms such as "chilled water pump," "cooling tower fan," and "fresh air valve" suddenly reverse after maintaining the same role label in three or more consecutive document locations, this is identified as a strong signal role reversal, and high-confidence markers are added to the role reversal location set. The role reversal position set includes all candidate reversal positions with a confidence level higher than 0.5. Each record carries the role tag pair before and after the reversal, the source type of the document, and the paragraph index. The size is 80-220 records in a typical building document set, of which HVAC host-related terms contribute about 45% of the records.

[0027] Local contextual semantic features are extracted from the role-reversal location set to generate conflict context features. In the role-reversal location set, the term "air conditioning unit" undergoes a role reversal at the intersection of a building energy consumption inspection log and an equipment failure work order. The context at this location simultaneously contains descriptions of both "abnormal start / stop frequency" and "insufficient cooling capacity." This type of contextualized multi-fault occurrence is a key area of ​​information density that needs to be captured when extracting conflict context features. After locating the target position of each record in the role-reversal location set within the building's unstructured document data, a local context window is created by extending 150 words forward and backward from this position. The text within the window is encoded by a pre-trained language model, outputting a 512-dimensional semantic vector. Contextual paragraphs involving HVAC equipment failure status, exceeding operating parameter limits, and cross-disciplinary linkage triggers exhibit a clearly separated vector distribution from descriptions of normal operation in the 512-dimensional space. Principal component analysis is performed on the 512-dimensional semantic vectors at each position to reduce dimensionality. The first 64 principal components are retained to cover more than 95% of the semantic variance. The 64-dimensional vector after dimensionality reduction serves as the core numerical representation of the conflict context features. Conflict context features are clustered by role label in 64-dimensional space. The role reversal context features of building HVAC host equipment form compact clusters after clustering. The variance within the cluster is less than 0.15 and is judged as a high-reliability conflict mode, which corresponds to a typical scenario in which the role definition of the equipment undergoes a systematic change when the equipment transitions from the construction and commissioning stage to the formal operation stage.

[0028] Conflict confirmation is performed based on conflict context features to generate a role conflict set. The conflict context feature vector is interfaced with a pre-trained conflict discriminator classifier. The classifier input is a 74-dimensional vector obtained by concatenating the 64-dimensional conflict context feature vector with the 10-dimensional one-hot encoded vector of the role label pair. The output is the confidence probability that the position belongs to a real semantic conflict. Positions with a confidence probability higher than 0.65 are included in the role conflict set. The discrimination boundary of the conflict discriminator classifier is tightened to 0.55 within the high-reliability conflict pattern cluster. A more sensitive confirmation criterion is used for conflict context features with stable triggering patterns to prevent strong regularity conflict patterns from being missed due to a fixed threshold. After the conflict positions are included in the role conflict set, multiple conflict records for the same term are deredundant and merged. Two records with a conflict context feature similarity higher than 0.88 are determined to be repeated triggers of the same conflict event. During merging, the record with the higher confidence probability is retained, and the document source information of the other record is appended to the retained entry. The final size of the role conflict set is 60-180 records in a medium-sized building document set. Redundancy removal and merging reduce the size by about 30%-45% compared to the role reversal position set. A lower compression rate indicates that the conflict event is triggered independently in multiple document locations, reflecting the diversity of conflict sources. In this case, the role conflict set retains all records without forced compression.

[0029] A conflict intensity distribution is generated by evaluating the conflict intensity of the role conflict set. The quantification of conflict intensity is based on the semantic distance between conflicting role pairs carried by each record in the role conflict set. The conflict intensity calculation formula is S_conflict=d_semantic×(f_occurrence / N_conflict)×(w_source / w_max), where d_semantic is the cosine distance between conflicting role pairs in the role label embedding space, f_occurrence is the frequency of conflict occurrence of the term in the role conflict set, N_conflict is the total number of entries in the role conflict set, w_source is the difference in authority level between the two types of conflict source documents, and w_max is the maximum value of the authority level difference. All three factors are dimensionless, and S_conflict is a dimensionless conflict intensity index with a value range of 0-1. In building HVAC systems, when the term "refrigeration unit" simultaneously functions as a controlled equipment in the construction acceptance report and a main control equipment in the energy management platform's exported reports, its d_semantic value is typically close to 0.8. Combined with the high frequency of this term in role conflict sets, its overall conflict intensity index is significantly higher than that of similar conflicts involving terminal fan coil units. When the same HVAC equipment term repeatedly conflicts in documents across multiple floors or areas of a building, the average conflict intensity index for each location is taken as the overall conflict intensity of that term. Locations with a conflict intensity index below 0.1 are filtered out before the average is calculated to exclude occasional weak conflicts. The conflict intensity distribution uses the overall conflict intensity of all terms as the statistical object. The conflict intensity distribution exhibits a right-skewed characteristic, with over 75% of terms having an overall conflict intensity below 0.4. Terms with a conflict intensity above 0.7 are concentrated in core HVAC equipment terms that span both the building construction and operation and maintenance phases. After generation, the conflict intensity distribution is output in descending order of intensity value for direct use during the high-conflict area calibration phase.

[0030] High-conflict regions are identified based on the conflict intensity distribution to generate a conflict knowledge point set. The identification threshold for high-conflict regions is set to the 80th percentile of the conflict intensity distribution, ensuring that the size of the term entries entering the conflict knowledge point set is controlled within 20% of the total term feature library, avoiding excessive expansion that could reduce the processing efficiency of the large language model. The identification of high-conflict regions is performed using a sliding window method on the conflict intensity distribution, with a window width of 0.1 intensity units. Sections within the window where the term density exceeds twice the distribution mean are identified as dense conflict zones. All terms within dense conflict zones, regardless of whether they exceed the 80th percentile threshold, are forcibly included in the conflict knowledge point set, ensuring that concentrated local conflict patterns are fully captured. Each knowledge point record in the conflict knowledge point set carries five items: target term, conflict intensity index, conflict role pair, list of source document numbers, and context fragment. The list of source document numbers is used to trace the authority level of document sources during the inter-chain conflict pruning stage, and the context fragment extracts 150 words before and after the conflict position to provide semantic background for the large language model. After generation, the conflict knowledge point set is sorted in descending order of conflict intensity index. The top 50 with the highest intensity are marked as the priority batch. The large language model uses a higher inference temperature for the priority batch to improve the diversity of relation extraction. The size of the conflict knowledge point set is usually 120-300 items on a medium-sized building document set.

[0031] The conflict knowledge point set is input into a large language model for causal trigger path identification and strength assessment, generating causal inference chain groups. The conflict knowledge point set consists of target terms, conflict role pairs, and context fragments as input. The large language model identifies the causal trigger relationship between the target terms and surrounding equipment entities based on the context. The trigger types are constrained by four categories defined by the building domain ontology: start-stop trigger, load transfer, temperature induction, and linkage protection. Paths outside the system are placed in a verification queue for manual confirmation. After path identification, a strength assessment is performed on each candidate causal path. The strength calculation formula is I_causal=p_llm×(f_cooccur / N_kp)×r_direction, where p_llm is the output confidence probability of the large language model for the path, f_cooccur is the co-occurrence frequency of the path's head and tail entities in the conflict knowledge point set, N_kp is the total number of entries in the conflict knowledge point set, r_direction is the consistency score between the path direction and the building's physical laws (0-1), and I_causal is in the range of 0-1. Candidate paths with too low strength are marked with a low confidence level. Taking HVAC as an example, the path that triggers the cooling tower fan speed-up due to high load rate of chiller units is frequently co-occurring in multiple maintenance records and energy consumption logs, and its direction conforms to thermodynamic laws. Therefore, its I_causal value is typically significantly higher than that of the end-device fine-tuning path, which only appears in a single document. After intensity assessment, these two types of paths enter chain group splicing queues with different priorities. After intensity assessment, multiple paths with coherent triggering relationships within the same equipment subsystem are linked together as causal inference chains. Paths involving cross-disciplinary linkage retain cross-domain triggering direction labels during linking. Multiple causal inference chains with similar intensity and involving the same equipment subsystem are grouped together to form a causal inference chain group. After all conflict knowledge point sets are processed, cross-group deduplication is performed on the causal inference chain group. Two groups with highly overlapping path coverage within the chain are merged and their trigger intensity weights are accumulated. Causal inference chains with higher weights receive higher initial edge weights when constructing the building knowledge graph.

[0032] A building knowledge graph is constructed by performing inter-chain contradiction pruning on causal reasoning chains. When there are contradictory triggering directions in causal reasoning chains for the same device, the inter-chain contradiction pruning process is triggered. The determination of a contradictory chain pair is based on the condition that the triggering direction codes (up / down, start / stop) of the two causal reasoning chains involving the same head and tail device entities are completely opposite. A typical scenario for triggering inter-chain contradiction pruning is the simultaneous existence of a "load transfer forward chain" and a "linkage protection reverse chain" between building HVAC chilled water pumps and chiller units. The inter-chain contradiction pruning strategy uses the product of the comprehensive trigger strength I_causal of each chain in the causal reasoning chain group and the confidence score output by the large language model as the pruning score. Causal reasoning chains with higher scores are retained as formal causal edges, while those with lower scores are downgraded to candidate causal edges and marked with a dispute label. Disputed edge markers are recorded as independent fields in the building knowledge graph, retaining complete information for subsequent manual verification. Two conflicting chains with a pruning score difference of less than 5% are determined to be conflicts of equal confidence. The document authority level corresponding to the list of source document numbers in the conflicting knowledge point set is used as the final ruling basis. Construction design document source chains are prioritized over operation and maintenance record document source chains as formal causal edges. After inter-chain conflict pruning, the building knowledge graph is organized and constructed with nodes representing equipment entities and directed edges representing causal triggering paths. Node attributes record entity type labels, and edge attributes record trigger type, pruning score, and comprehensive trigger strength. After the building knowledge graph is constructed, connectivity verification is performed. When isolated nodes account for more than 10% of the total number of nodes in the graph, a causal reasoning chain supplementation identification process is triggered to ensure that the connectivity quality of the building knowledge graph meets the structural requirements for subsequent knowledge injection.

[0033] Step S130: Extract graph knowledge vectors from the building knowledge graph, identify knowledge backlog nodes and determine knowledge-weak areas based on the topological structure of the building knowledge graph, and inject graph knowledge vectors into knowledge-weak areas to generate a knowledge-enhancing control model.

[0034] Specifically, graph knowledge vectors are extracted from the building knowledge graph. In the building knowledge graph, HVAC domain nodes account for 42%-48% of all nodes, electrical and fire protection domain nodes account for 38%-43%, and water supply and drainage domain nodes account for the lowest proportion, approximately 10%-15%. This uneven distribution of node density across domains determines a systematic difference in the expressive power of graph knowledge vectors across different domains. Graph knowledge vectors use a graph neural network to perform message aggregation encoding on each node and its first-order neighbors in the building knowledge graph. The aggregation process weightedly fuses the node's own entity type embedding with the relation types of neighboring nodes, outputting a 128-dimensional node embedding vector as the basic unit of the graph knowledge vector. In building knowledge graphs, cross-disciplinary collaborative entity nodes need to be extended to second-order neighborhood aggregation when extracting knowledge vectors. First-order neighborhoods only cover directly related nodes within the same professional domain and cannot capture cross-domain dependencies between HVAC chilled water networks and water supply and drainage systems. Second-order neighborhood expansion incorporates such cross-domain relationships into the vector encoding scope, making the graph knowledge vectors more complete in expressing the collaborative characteristics of multi-professional systems in a building. After the graph knowledge vectors are extracted, a full embedding matrix is ​​constructed using node numbers as row indices. The matrix size is 2000-5000 rows × 128 columns in medium-sized building knowledge graphs. Nodes with a norm lower than 0.5 times the standard deviation of the full matrix mean have their propagation coverage weight increased by 1.2 times in the subsequent propagation coverage insufficiency rate calculation and are preferentially included in the knowledge weakness area labeling range.

[0035] In some embodiments, identifying knowledge backlog nodes and determining knowledge-weak areas based on the topological structure of the building knowledge graph includes: calculating the in-degree to out-degree ratio of each node in the building knowledge graph to generate a knowledge flow ratio distribution; identifying high-in-degree and low-out-degree nodes from the knowledge flow ratio distribution to generate a knowledge backlog node set; assessing the propagation impact range of the knowledge backlog node set to generate a propagation coverage insufficiency rate; and identifying knowledge-weak areas based on the propagation coverage insufficiency rate.

[0036] The in-degree to out-degree ratio of each node in the building knowledge graph is calculated to generate a knowledge flow ratio distribution. Nodes of the "refrigeration unit" type in the building knowledge graph typically have 8-15 in-degree edges and only 2-4 out-degree edges, resulting in an in-degree to out-degree ratio of approximately 3.0-5.5. This ratio is the highest among all node types in the graph, and this imbalance creates a noticeable right-tail elongation characteristic in the knowledge flow ratio distribution. The in-degree to out-degree ratio calculation separately processes isolated nodes (both in-degree and out-degree are 0) in the building knowledge graph. The ratio of isolated nodes is assigned a special marker value of -1 instead of participating in the numerical calculation, preventing division by zero from contaminating the statistical results of the knowledge flow ratio distribution. Isolated nodes typically account for 3%-8% of the building knowledge graph, concentrated in construction and acceptance documents with incomplete equipment number entries. The knowledge flow ratio distribution is plotted using the ratio on the horizontal axis and the number of nodes on the vertical axis to construct a frequency histogram. The bin width of the histogram is set to 0.5, covering the ratio range of 0-10. Nodes with ratios exceeding 10 are separately classified into the overflow interval. In the building knowledge graph, the ratio of cross-disciplinary collaborative nodes in the knowledge flow ratio distribution typically falls within the range of 3.5-7.0. The node density in this range is much higher than the average level of the entire graph, and the knowledge flow ratio distribution exhibits local peaks in this range. Nodes in the peak interval are the main candidate sources of knowledge backlog node sets.

[0037] A knowledge backlog node set is generated by identifying high in-degree and low out-degree nodes from the knowledge flow ratio distribution. The 80th percentile value of the knowledge flow ratio distribution is taken from the distribution statistics and used as the identification threshold for high in-degree and low out-degree nodes. Nodes in the building knowledge graph whose ratio exceeds this threshold and whose absolute in-degree value is greater than 5 are identified as knowledge backlog nodes. This dual-condition filtering excludes low-in-degree nodes with high ratios but sparse actual connections. During the identification process, the directed edges of the building knowledge graph are traced back to their source. Nodes whose incoming edges are concentrated in high-authority design documents are prioritized for inclusion in the knowledge backlog node set. High-ratio nodes whose sources are scattered across multiple types of low-authority documents are downgraded, with their weight reduced by 0.3 and then compared with the threshold again. When the same equipment entity in a building knowledge graph corresponds to multiple nodes in different professional domain contexts, each node independently participates in the identification of knowledge backlog nodes. The identification results show that multiple backlog nodes belonging to the same equipment entity are merged into a single backlog node set record. During merging, the weighted average of the ratios of each node is taken as the overall backlog intensity of the entity, with the weights based on the in-degree of each node in the building knowledge graph. The size of the knowledge backlog node set is typically 80-200 nodes in a medium-sized building knowledge graph, accounting for 4%-8% of the total number of nodes in the graph. When the proportion exceeds 12%, it is necessary to backtrack during the inter-chain contradiction pruning stage to check whether too many cross-professional relationship edges have been incorrectly removed.

[0038] The propagation impact range of a knowledge backlog node set is assessed to generate a propagation coverage insufficiency rate. The propagation impact range is calculated by performing a third-order neighborhood expansion on the building knowledge graph, with each node in the knowledge backlog node set as the source. Nodes reachable within the third-order neighborhood constitute the propagation impact range of that backlog node. The ratio of the number of reachable nodes to the total number of nodes in the building knowledge graph is defined as the propagation coverage rate of that node. In the building knowledge graph, the third-order reachability range of the "Chiller Unit" node to cross-disciplinary nodes such as fire alarm circuits and water supply and drainage networks is often less than 18% of the total number of nodes. However, the third-order reachability range of the "HVAC Master Controller" node, due to its bridging relationships connecting multiple professional domains, reaches over 42% of the total number of nodes. The difference in propagation coverage rates between these two types of nodes produces drastically different assessment results in the propagation coverage insufficiency rate calculation. The formula for calculating the propagation coverage deficiency rate is R_deficit = 1 - min(1, (N_reachable / N_total) × (d_max / d_avg)), where N_reachable is the number of reachable nodes within the 3rd-order neighborhood, N_total is the total number of nodes in the building knowledge graph, d_avg is the average path length from the knowledge backlog node to all reachable nodes, d_max is the diameter of the building knowledge graph (i.e., the maximum path length in the entire graph), d_max / d_avg is the dimensionless path ratio, and R_deficit is a dimensionless ratio with a value of 0-1. A higher value indicates a more severe inhibition of the propagation of professional knowledge around the building by the node. The average propagation coverage deficiency rate of knowledge backlog nodes concentrated in the fire protection professional domain is usually 0.12-0.18 higher than that of the HVAC professional domain, followed by the water supply and drainage professional domain. Backlog nodes in these two professional domains are prioritized for processing during the knowledge-weak area identification phase. The insufficient coverage rate is calculated for each node in the knowledge backlog node set to form a node-level insufficient rate sequence. Nodes with an insufficient rate higher than 0.75 are marked as severely backlogged nodes. The building subsystems corresponding to severely backlogged nodes are given priority in the spatial delineation of knowledge-weak areas.

[0039] Knowledge-weak regions are identified based on the propagation coverage insufficiency rate. The spatial distribution of nodes with a propagation coverage insufficiency rate higher than 0.6 in the building knowledge graph is aggregated using the Louvain community detection algorithm. The aggregation results typically group nodes from the three professional domains—fire protection linkage control, water supply and drainage network monitoring, and building electrical distribution—into their respective independent high-insufficiency-rate communities. HVAC domain nodes, due to their numerous interconnected edges with other professional domains, are less likely to form isolated communities. The final identification of knowledge-weak regions involves connectivity verification based on candidate communities. Regions with weak connectivity are defined as those with an average path length exceeding 4.5 within the building knowledge graph. The intersection of these weakly connected regions and regions with high propagation coverage insufficiency rates defines the official boundaries of knowledge-weak regions. The inadequate coverage rate provides a severity classification of knowledge gaps during the calibration process. A mean inadequate coverage rate between 0.6 and 0.75 is classified as Level 1, between 0.75 and 0.9 as Level 2, and above 0.9 as Level 3. Nodes connecting the fire sprinkler system and the water supply and drainage subsystem in a building typically fall into Level 2 or 3 gaps, receiving the highest density of targeted injection during the knowledge vector injection phase to prioritize filling cross-knowledge gaps. After calibration, knowledge gaps are segmented and stored from the building knowledge graph as subgraphs. These subgraphs retain the node identification information of the original graph. The number of nodes in each subgraph after segmentation is typically 15-80. When a subgraph is too large, it is further subdivided using the inadequate coverage rate as a weight, ensuring that each knowledge gap subgraph can be fully processed in a single forward propagation during the targeted injection phase.

[0040] In some embodiments, the step of injecting the graph knowledge vector into the knowledge-weak region to generate a knowledge enhancement control model includes: partitioning the graph knowledge vector according to the knowledge-weak region to generate a regional knowledge feature vector; identifying high-density feature dimensions from the regional knowledge feature vector to generate a knowledge-dense feature set; redistributing knowledge in high-density regions based on the knowledge-dense feature set to generate a targeted knowledge injection sequence; and executing gradient-guided knowledge injection to generate a knowledge enhancement control model according to the targeted knowledge injection sequence.

[0041] The graph knowledge vectors are partitioned and mapped according to knowledge-weak regions to generate regional knowledge feature vectors. The full node embedding matrix of the graph knowledge vectors is 2000-5000 rows × 128 columns on a medium-sized building knowledge graph. The row subsets of graph knowledge vectors corresponding to knowledge-weak regions are extracted separately during the partitioning and mapping stage. The node embedding row vectors of each weak region subgraph are aggregated into a region-level feature matrix according to the weak region number. Partitioning and mapping performs intra-region aggregation on the 128-dimensional embedding vectors of nodes in each weak region of the graph knowledge vectors. The aggregation method is a weighted average of the embedding vectors of all nodes within the weak region. The weights are based on the propagation coverage insufficiency rate of each node; nodes with higher insufficiency rates have greater weights in the regional aggregation. The aggregation result forms the regional knowledge feature vector for that weak region. The dimension of the regional knowledge feature vector is the same as the single-node embedding dimension of the graph knowledge vector, 128 dimensions. The number of knowledge-weak regions determines the number of entries in the regional knowledge feature vector. On a typical medium-sized building knowledge graph, 8-20 regional knowledge feature vectors are usually generated, each vector corresponding to a labeled knowledge-weak region subgraph. After partitioning and mapping, the regional knowledge feature vectors corresponding to the Level 3 weak areas are appended with severity level labels. These labels are used to adjust the confidence threshold for feature dimension selection during the knowledge-intensive feature set identification stage. The Level 3 weak areas use a more lenient threshold to expand the identification range of high-density feature dimensions, ensuring more comprehensive feature coverage for severely missing knowledge regions. The regional knowledge feature vectors generated for each weak area are concatenated into a joint feature matrix. The number of rows in the joint feature matrix equals the number of knowledge-weak areas, and the number of columns is 128. The cosine similarity between the row vectors in the matrix reflects the semantic similarity between different weak areas.

[0042] High-density feature dimensions are identified from regional knowledge feature vectors to generate a knowledge-intensive feature set. The column variance sequence of the joint feature matrix reveals the degree of expression difference between different weak areas from the statistical distribution of 128 dimensions. Dimensions with high column variance indicate that the semantic content carried by that dimension differs significantly in different knowledge-weak areas, and are key feature dimensions for distinguishing knowledge types in weak areas. The identification of high-density feature dimensions is based on the column mean of the joint feature matrix of regional knowledge feature vectors. The identification process calculates the absolute value of the column mean for each of the 128 dimensions and compares it with a threshold of 1.5 times the standard deviation of the full matrix mean. Dimensions that pass the threshold are sorted in descending order of the absolute value of the column mean to form a candidate list of high-density feature dimensions. The dimension indexes in the candidate list are written into the knowledge-intensive feature set. High-density feature dimensions typically occupy 20-35 in the 128-dimensional space, corresponding to the knowledge subspace with the most active semantic expression in the knowledge-weak areas. The knowledge-intensive feature set is constructed after the high-density feature dimension identification is completed. Each record in the set carries a dimension index, the numerical distribution of that dimension in each weak area of ​​the regional knowledge feature vector, and the corresponding professional domain affiliation label. The professional domain affiliation is inferred by projecting the high-density dimension onto the professional domain label space of the building knowledge graph. The professional domain distribution of the knowledge-intensive feature set is basically consistent with the professional domain node proportion distribution of the building knowledge graph, verifying that the partitioning mapping result of the regional knowledge feature vector effectively represents the professional domain structure of the building knowledge graph.

[0043] Based on a knowledge-intensive feature set, a targeted knowledge injection sequence is generated through high-density area knowledge redistribution. When the value of a high-density dimension in a certain HVAC domain within the knowledge-intensive feature set is only 0.31 times the global mean in the regional knowledge feature vector corresponding to a level-three weak area, this dimension is identified as a severely knowledge-deficient dimension in that weak area. High-density area knowledge redistribution proportionally migrates the value of this dimension from the weak area of ​​a strong professional domain to the severely knowledge-deficient weak area. The migration ratio of high-density area knowledge redistribution is determined by the difference between the insufficient coverage rate of the source weak area and the insufficient coverage rate of the target weak area for each dimension in the knowledge-intensive feature set. The larger the difference, the higher the migration ratio, with an upper limit of 40%. The migration operation only updates the target dimension value in the row vector corresponding to the target weak area; the row vector corresponding to the source weak area is not deducted, ensuring that the knowledge representation capability of the source weak area is not affected. High-density dimensions belonging to the same professional domain within the knowledge-intensive feature set undergo batch migration during redistribution. Batch migration ensures that the relative proportions of dimensions within the same professional domain remain unchanged after migration, avoiding single-dimensional migration from disrupting the internal consistency of professional domain knowledge. The targeted knowledge injection sequence is organized in the form of a migration operation list. Each operation record in the list includes the source weak area number, the target weak area number, the set of high-density dimension indices involved, and the migration ratio value. The operation records are arranged in descending order of the severity level of the target weak area. The migration operation corresponding to the third-level weak area is placed at the beginning of the sequence and executed first. The total number of operation entries in the targeted knowledge injection sequence is usually 40-120 when the knowledge-intensive feature set has 20-35 dimensions.

[0044] A gradient-guided knowledge injection generation and knowledge enhancement control model is executed based on the targeted knowledge injection sequence. Each operation record in the targeted knowledge injection sequence sequentially drives the knowledge injection module to update the regional knowledge feature vector. The injection direction is applied along the gradient ascent direction on the row vector corresponding to the target weak region, with an injection step size of 0.01 multiplied by the migration ratio. Gradient ascent ensures that the cosine similarity between the feature vector of the target weak region and the features of the strong region continuously increases after injection. The targeted knowledge injection sequence is executed in rounds. After each round, all operation records are executed completely, and the update magnitude of the regional knowledge feature vector of each weak region is checked. The update magnitude is measured by the ratio of the change in L2 norm before and after injection to the initial L2 norm. When the ratio is less than 0.5%, the knowledge injection of the current weak region is considered converged, and the converged weak region exits the subsequent rounds. The targeted knowledge injection sequence is completed when all weak regions have converged. The knowledge-enhanced control model uses the injected regional knowledge feature vector as the enhanced knowledge representation. This enhanced representation is mapped back to the node level according to the weak zone number to which the node belongs, and then concatenated node-by-node with the original graph knowledge vector. The weak zone affiliation is based on the node-weak zone attribution table in sub-step 4. The concatenated vector of each node has a dimension of 128 + 128 = 256. With 256-dimensional input, the inference module can more accurately classify the operating status of HVAC equipment. During the execution of the targeted knowledge injection sequence, if the norm of a weak zone continuously increases by more than three times the initial value, it is determined to be an injection divergence. The injection step size is automatically halved, and the iteration counter is reset to reconverge. This divergence detection mechanism ensures that the knowledge-enhanced control model maintains numerical stability.

[0045] Step S140: Based on the knowledge-enhanced control model, perform bidirectional causal reasoning of the HVAC system to generate a decision reasoning chain, perform directional violation detection on the decision reasoning chain to generate low-confidence nodes, perform weight reduction processing on the low-confidence nodes, and output interpretable control decisions.

[0046] Specifically, a knowledge-enhanced control model is used to perform bidirectional causal reasoning in the HVAC system to generate a decision-making reasoning chain. Upon receiving real-time sensor data from the building's HVAC system, the knowledge-enhanced control model simultaneously initiates two reasoning channels: forward causal reasoning and reverse causal tracing. These two channels share sensor inputs encoded with a 256-dimensional spliced ​​vector, collaboratively advancing to form the decision-making reasoning chain. The forward causal reasoning channel, based on the enhanced graph knowledge semantics, predicts the response state changes of each downstream device from the current operating state. Each step of reasoning uses the device state variables output from the previous step as input to trigger the next causal inference, chain-like advancement forming a sequence of control action suggestions from the cold source side to the terminal side. The reverse causal tracing channel, upon detecting a deviation signal in the sensor data, starts with the deviation event and traces back layer by layer along the causal edges in the building knowledge graph to identify the upstream device state anomaly node causing the current deviation. The reverse tracing path and the forward reasoning path are cross-labeled in the decision-making reasoning chain, providing a complete causal evidence chain for interpretable control decisions. The weighted harmonic average of the two channel confidence scores at the merging node is used as the comprehensive decision confidence score for this reasoning step. The weight ratio is dynamically adjusted according to the current operating condition. Under stable operating conditions, the forward channel has a dominant weight, while under scenarios triggered by deviation events, the reverse channel weight is increased to enhance the influence of tracing evidence on the decision. The bidirectional causal reasoning of the building HVAC system under summer cooling conditions involves the linkage analysis of three subsystems: the cold source side, the distribution side, and the terminal side. When the chiller unit load rate is too high, the forward channel triggers a suggestion to increase the cooling tower fan speed, while the reverse channel traces the large temperature difference between the chilled water supply and return water to the failure node of the water pump frequency converter control. The two reasoning paths are merged in the decision reasoning chain into a reasoning record with a causal closed-loop structure.

[0047] In some embodiments, the step of detecting directional violations in the decision reasoning chain to generate low-confidence nodes includes: extracting causal directions step by step from the decision reasoning chain to generate a reasoning causal sequence; performing directional violation clustering analysis using the reasoning causal sequence to generate a violation clustering pattern; distinguishing between local random violations and global systematic violations based on the violation clustering pattern to generate a violation classification label; and extracting random violation class nodes based on the violation classification label to generate low-confidence nodes.

[0048] The decision-making reasoning chain is processed step-by-step to extract causal directions and generate a causal sequence. In the decision-making reasoning chain, the causal direction of a three-step reasoning segment such as "chilled water pump speed decreases → supply water temperature increases → terminal equipment cooling capacity decreases" has a clear physical orientation. However, "supply water temperature increases → chilled water pump speed increases" is opposite to the direction of thermodynamic laws. The alternation of these two types of causal directions in the decision-making reasoning chain is a typical trigger scenario for direction violation detection. Causal direction extraction uses the entity state variables of adjacent reasoning steps in the decision-making reasoning chain as the operation object. The extraction module encodes the direction of change between the input and output states of each pair of adjacent steps. The encoding result is expressed as directed sign pairs (+ / () indicates the increase or decrease relationship of causal variables. Positive causality is coded as (+,+) or () , The reverse causal encoding is (+, )or( (+). After step-by-step extraction, the decision reasoning chain forms a causal sequence. Each element in the sequence carries four pieces of information: step number, causal entity pair, direction code, and step confidence. The sequence length is consistent with the number of steps in the decision reasoning chain, typically ranging from 6 to 20 steps, corresponding to 5 to 19 elements. When more than 30% of the elements in the causal sequence have a step confidence below 0.5, the overall quality of the decision reasoning chain is deemed insufficient. A low-quality label is added to the entire causal sequence, and directional violation clustering analysis increases the sensitivity of the violation judgment threshold when this label is present, in order to cope with detection scenarios where noise interference increases in low-quality reasoning chains.

[0049] Directional violation clustering analysis is used to generate violation clustering patterns using inference causal sequences. The distribution of inverse causal coding elements at step positions in the inference causal sequence is not random. In HVAC systems, the inverse coding density in inference steps near the switching between cooling and heating modes is typically 1.8-2.6 times higher than in other step segments. Clustering analysis needs to identify this position-related density inhomogeneity. Directional violation clustering analysis performs DBSCAN clustering on the set of positions formed by the step numbers of all inverse coding elements in the inference causal sequence. The cluster radius ε is set to 2 step position units, and the minimum sample size MinPts is set to 2. The clustering results group densely occurring inverse coding positions within adjacent steps into the same violation cluster. After DBSCAN clustering, the members of each cluster of violation clustering patterns are grouped and archived. The inference steps corresponding to noise points (isolated inverse coding positions not belonging to any cluster) in the violation clustering pattern are considered as candidates for local random violations in subsequent violation classification, while the steps corresponding to cluster members are included in the candidate evaluation for global systematic violations. Pattern violations are organized using the central step number of each cluster, the number of elements within the cluster, the range of steps covered by the cluster, and a list of noise point locations. Clusters with more than 3 elements within a cluster and whose range of steps covers more than 30% of the total number of steps in the decision reasoning chain are labeled as large-scale clusters. After clustering, three statistics are recorded for pattern violations: the total number of clusters, the number of large-scale clusters, and the total number of noise points. The proportional relationship of these three statistics directly participates in the calculation of local and global violations in the subsequent violation classification label generation stage.

[0050] For example, the step of distinguishing between local random violations and global systematic violations and generating violation classification labels for the violation clustering pattern includes: performing cluster size statistics on the violation clustering pattern to generate a cluster size distribution; identifying large-scale clusters across inference steps from the cluster size distribution to generate a global systematic violation set; generating a local random violation set from the residual clusters in the violation clustering pattern after excluding the global systematic violation set; and generating violation classification labels based on the global systematic violation set and the local random violation set.

[0051] Cluster size distribution is generated by statistically analyzing the clustering scale of violations of clustering patterns. Clusters with two elements within a cluster account for approximately 45%-55% of all effective clusters in typical multi-device linkage scenarios in buildings, while large-scale clusters with more than four elements account for only 10%-20%, but their coverage of inference steps typically reaches over 40% of the entire chain length. Cluster size statistics calculate three scale indicators for each cluster in violation of clustering patterns: the number of elements within the cluster, the span of inference steps covered by the cluster, and the professional domain diversity index of causal entity pairs within the cluster. These three indicators together constitute a multi-dimensional description of the cluster size distribution. The cluster size distribution constructs a two-dimensional distribution matrix with the number of elements within the cluster as the primary sorting dimension and the span of inference steps as the secondary sorting dimension. Cells in the matrix with three or more elements and a step span of four or more correspond to high-scale, high-span clustering areas. The inference chain segments of building HVAC system start-stop linkage fire protection and water supply and drainage systems fall into these areas most frequently, making them the core target for identifying global systemic violation sets. Noise points in violation of clustering patterns are not included in the cluster size statistics. The ratio of the total number of noise points to the total number of effective clusters is used to adjust the selection threshold of large-scale clustering candidates when identifying global systematic violation sets. The higher the ratio, the tighter the selection threshold.

[0052] Large-scale clusters spanning multiple inference steps are identified from the cluster size distribution to generate a global systematic violation set. Clusters within high-scale, high-span clusters in the two-dimensional matrix of cluster size distribution are considered large-scale cluster candidates. These candidate clusters require further verification of their domain diversity index. The diversity index is calculated by dividing the number of building domains involved in the causal entity pairs within the cluster by the total number of domains in the building knowledge graph (HVAC, electrical, fire protection, and water supply / drainage, a total of four categories). Clusters with cross-domain linkage typically have a diversity index higher than 0.5 and are formally included in the global systematic violation set if this condition is met. Clusters whose step span exceeds 50% of the total length of the decision inference chain in the cluster size distribution are directly identified as having global systematic violations, regardless of the number of elements. After being included in each large-scale cluster, the global systematic violation set undergoes deduplication and merging. Two clusters with a step coverage overlap exceeding 70% are merged into one record. During merging, the union of coverage areas is taken, and the average of the domain diversity indices of the two clusters is calculated. The size of the global systemic violation set is 1-4 records in a typical building HVAC reasoning scenario. When the number of records exceeds 5, it is necessary to backtrack and check whether the targeted knowledge injection sequence has caused over-injection problems on the regional knowledge feature vectors of the relevant professional domain.

[0053] The residual clusters in the violation clustering pattern, excluding the global systematic violation set, generate local random violation sets. Clusters covered by the global systematic violation set are marked and removed from the complete cluster list of the violation clustering pattern. This removal operation only applies to steps of cluster members already included in the global systematic violation set and does not affect the step assignments of other clusters in the violation clustering pattern. The residual clusters consist of all clusters in the violation clustering pattern not covered by the global systematic violation set. For example, in a building inference chain, steps 7 and 8 respectively contain two brief reverse codes: "Increase the opening of the terminal fan coil unit thermostat valve → decrease the indoor temperature" and "Increase the opening of the fresh air valve → increase the supply air temperature." These two adjacent steps are grouped into the same cluster by DBSCAN. This cluster has only 2 elements and a step span of 1, failing to meet the dual conditions of scale and span for a global systematic violation, and is thus entered into the candidate pool as a typical residual cluster. The local random violation set performs random verification on each cluster in the candidate pool. The verification method is to assume that the uniform distribution of the inverse encoding positions in the inference causal sequence across the entire sequence is the null hypothesis, and to perform a chi-square test on the step position distribution of the candidate clusters. Clusters with a chi-square test p-value higher than 0.05 are judged as having a statistically significant random distribution violation. The local random violation set finally includes the residual cluster members that have passed the randomness verification. In a moderately complex building decision reasoning chain, it typically contains 3-8 valid cluster records, each carrying a cluster number, the range of covered steps, and a chi-square test p-value.

[0054] Violation classification labels are generated based on global systematic violation sets and local random violation sets. In the global systematic violation set, each record covering a reasoning step is uniformly labeled "Global Systematic Violation," with the record's domain diversity index added as a severity quantification. Violation records spanning chiller units, fire water supply, and water supply and drainage domains have higher diversity indices, so their covered steps are excluded from candidate pool construction during subsequent weight reduction processing. In the local random violation set, each cluster covering a step is labeled "Local Random Violation," with the corresponding cluster's chi-square test p-value added as a randomness quantification. Steps involving fine-tuning of end-devices have higher p-values ​​and stronger randomness, so their labels participate in candidate pool construction with a smaller decay coefficient during the low-confidence node extraction stage. Violation classification labels assign a "No Violation" label to steps in the decision-making reasoning chain not covered by either violation set. Steps with the "No Violation" label are skipped during the low-confidence node extraction stage and do not participate in weight reduction processing. After all steps are labeled, the percentage of steps with the three types of labels is calculated. When the percentage of global systemic violations exceeds 20%, an inference quality alarm is triggered in the knowledge-enhanced control model. When the explainable control decision is output, a high-risk warning mark is attached and sent to the building operation and maintenance management platform along with the control decision.

[0055] Random violation nodes are extracted based on violation classification labels to generate low-confidence nodes. Steps of cluster members with the label "local random violation" in the violation classification label set are extracted from the violation classification label. All noise point locations are simultaneously extracted from the noise point location list of the violation classification label. The inference steps of these two types of locations together constitute the candidate pool for low-confidence nodes. Steps labeled "global systematic violation" and "no violation" in the violation classification label are not included in the candidate pool. The confidence decay of each node in the candidate pool is determined based on the specific subtype of the violation classification label. The original step confidence of the node corresponding to the noise point is multiplied by a decay coefficient of 0.5. The decay coefficient of local random violation cluster member nodes is determined by linear interpolation based on the cluster density, ranging from 0.55 to 0.7; higher density results in greater decay. After extraction, low-confidence nodes exclude mode-switching-induced violation nodes. Although these nodes are recorded as violations in the violation classification label, their physical causes are known and reasonable. Mode-switching-induced violation nodes are retained as a separate low-weight copy in the low-confidence node set, used only for interpretability labeling and not participating in decision aggregation calculations. In typical HVAC system reasoning scenarios, the scale of low-confidence nodes is 15%-35% of the total number of steps in the decision reasoning chain. When the proportion exceeds 40%, the local knowledge supplementation process of the corresponding professional domain in the building knowledge graph needs to be triggered.

[0056] Low-confidence nodes are weighted to output interpretable control decisions. The weighting process for low-confidence nodes involves replacing the original confidence of the corresponding reasoning step in the decision inference chain with the diminished confidence of each node. The replaced confidence sequence then re-participates in the inference chain aggregation calculation. The aggregation method involves weighted voting of control action suggestions for each step, using the updated confidence as the weight. The contribution weight of the control action suggestion for the step corresponding to a low-confidence node in the voting is less than 50% of that of normal nodes, ensuring that the interference of low-confidence reasoning steps on the final control decision is effectively suppressed. The interpretable control decision is output as a sequence of control instructions. Each control instruction carries three pieces of information: the corresponding reasoning step number, the execution confidence, and a summary of the causal path. The summary of the causal path is represented by the sequence of key reasoning nodes from the trigger state to the control action in the decision inference chain, providing traceable decision-making basis for building maintenance personnel. Explainable control decisions with a global systemic violation rate exceeding 20% ​​are marked with a high-risk warning tag. This tag includes a list of violation step numbers and corresponding professional domain annotations. These tags are sent along with the control decisions to the building operations and maintenance management platform for maintenance personnel to review before execution. Explainable control decisions are written to the inference decision log after generation. The log is stored with dual indexes of operating condition type and timestamp, accumulating sufficient historical samples to provide a labeled inference quality data foundation for the iterative optimization of the knowledge-enhanced control model.

[0057] Step S150: Based on the interpretable control decision-driven operation of the HVAC system, generate equipment behavior data, compare the deviation between the equipment behavior data and the decision reasoning chain to generate behavior deviation value, and based on the behavior deviation value, perform hierarchical over-limit identification to trigger adaptive reconstruction of the building knowledge graph to complete the construction of the building knowledge graph.

[0058] Specifically, interpretable control decisions drive the operation of the HVAC system, generating equipment behavior data. These decisions are issued as a sequence of control commands to the execution layers of each device within the building's HVAC system. The command sequences are grouped by equipment type, with chiller start / stop and load rate adjustment commands having the highest priority and being issued first. Variable frequency speed control commands for chilled water pumps and cooling tower fans are issued next, and commands for adjusting the supply air temperature and fresh air valve opening of terminal air conditioning units are executed last. This staggered timing of the three types of commands ensures that no electrical surges occur when each subsystem of the building's HVAC system responds to interpretable control decisions. After responding to the control commands, each device execution layer transmits operating parameters in real time: chillers transmit actual load rate and outlet water temperature, chilled water pumps transmit actual speed and flow rate, and terminal air conditioning units transmit supply air temperature and indoor temperature and humidity. These parameters are aggregated by the building automation system to form equipment behavior data. The sampling frequency of this data varies from 1Hz to 10Hz depending on the equipment type. During the acquisition process, equipment behavior data synchronously records the time of instruction issuance that can explain control decisions. Both share the same time base to ensure that the timing alignment error does not exceed 100ms when comparing subsequent deviations. When there are abnormal equipment responses in the building, the feedback parameters of the corresponding equipment in the equipment behavior data may show a prolonged state of stillness or a state opposite to the control command. Such abnormal feedback data is marked with an anomaly before the equipment behavior data is entered into the database. The confidence level C_confidence for missing responses during the anomaly-marked period is reduced to 0.4 when response matching is triggered in the subsequent period, to prevent data anomalies caused by hardware failures from being misjudged as knowledge graph inference bias.

[0059] In some embodiments, the step of generating a behavior deviation value by comparing the device behavior data with the decision reasoning chain includes: extracting the expected behavior trigger timing sequence from the decision reasoning chain to generate an expected timing sequence; matching the device behavior data with the expected timing sequence to generate a response missing set; jointly evaluating the response missing set based on the duration of response missing and trigger priority to generate a response missing severity distribution; and performing a weighted mapping of the response missing severity distribution to generate a behavior deviation value.

[0060] The expected behavior trigger timing is extracted from the decision reasoning chain to generate an expected timing sequence. In HVAC systems with multi-device coordinated operation, the decision reasoning chain typically contains 8-14 equipment action triggering steps. The expected response time of each step is estimated by the knowledge-enhanced control model during the reasoning phase based on fixed response delay parameters for each equipment type (8-12 minutes for chillers, 1-3 minutes for chilled water pumps, and 0.5-1 minute for terminal equipment), and written into the intermediate layer output of the reasoning chain. The estimation accuracy of the expected response time is typically better than ±15 seconds after fine-tuning of the knowledge-enhanced control model. The expected behavior trigger timing extraction process reads the expected response time, the involved equipment number, and the expected direction of change of the state variable for each reasoning step containing equipment action instructions in the decision reasoning chain. These three pieces of information are combined into an element of the expected timing sequence, which is arranged in ascending order of expected response time, forming a time-ordered list of expected equipment actions. In the expected timing sequence, pairs of elements with an interval of less than 200ms between two adjacent trigger times are identified as concurrent action pairs. These concurrent action pairs are processed as a whole during the trigger-response matching phase, and the matching window is expanded to the later of the two expected times plus 300ms to ensure that concurrent scheduling delays at the hardware execution layer are not misjudged as missing responses. The completeness of the expected timing sequence is measured by the coverage of device action steps in the decision reasoning chain. When the coverage is less than 85%, an insufficient coverage flag is added to the expected timing sequence. The trigger-response matching of sub-step 2 expands the matching window of the expected timing sequence with the insufficient coverage flag to ±800ms to compensate for the timing offset of implicit action steps.

[0061] The device behavior data is matched with the expected time series to generate a response missing set. The sampling frequency of the device behavior data varies from 1Hz to 10Hz, depending on the device type. The sampling frequency for HVAC main unit devices is typically 10Hz, while for terminal fan coil units it is 1-2Hz. The difference in sampling frequency between the two types needs to be uniformly interpolated to a 5Hz reference frequency when matching with the expected time series. The trigger response matching is centered on the expected response time of each element in the expected time series, with a matching window of 500ms before and after it. The device behavior data is searched for whether the corresponding device number in the window shows a state transition consistent with the expected state variable change direction. The threshold for judging a state transition is 1.5 times the standard deviation of the historical mean of the device behavior data. Elements in the expected time series for which no state transition matching the condition is detected in the matching window are judged as response missing events. The response missing event, along with its expected response time, device number, expected state change direction, and trigger priority, is written into the response missing set. The trigger priority is read from the control instruction priority field corresponding to the expected response time of each record in the response missing set. Data gaps in device behavior data caused by network latency or sensor malfunction during acquisition and transmission are filled in forward before response matching is triggered. Interpolation markers are added to the filled data points, and the confidence level for missing responses within the matching window containing the interpolated data points is reduced to 0.6. Under typical normal operating conditions of a HVAC system, the number of entries in the missing response set is usually 0-3. An immediate alarm is triggered when the number of entries exceeds 5. After maintenance personnel confirm the device is online, the confidence level C_confidence of the corresponding record in the missing response set is set to 0.3, and subsequent deviation calculations continue. When the device is confirmed to be offline, the corresponding record is removed from the missing response set and does not participate in severity calculations.

[0062] A severity distribution of response missing is generated by jointly evaluating the duration of response missing and the trigger priority of the response missing set. The duration of response missing is defined as the time interval from the expected response time to the first occurrence of an expected directional change in the corresponding device state variable in the device behavior data. If the duration of the missing response of the building chilled water pump speed exceeds 3 seconds, it usually means that the inverter communication is interrupted or the control signal is lost, and it is judged as a continuous missing and assigned a maximum penalty value of 3.0 seconds. The trigger priority is transmitted along with each record in the response missing set. Core control actions such as the start and stop of the building HVAC host are assigned a priority of 5, coordinated speed regulation actions of chilled water pumps and cooling tower fans are assigned a priority of 3-4, and low-impact actions such as fine-tuning of terminal fan coil thermostatic valves are assigned a priority of 1-2. The severity calculation formula for a single record in the response missing severity distribution is: Severity = (T_delay / T_max) × P_priority × C_confidence, where T_delay is the duration of the response missing (in seconds), T_max is the maximum acceptable duration of the response missing set by the system (3 seconds), T_delay / T_max is the dimensionless normalized duration, P_priority is the trigger priority, and C_confidence is the confidence level for determining the response missing. All three factors are dimensionless, and the dimension of Severity is the dimensionless severity index, ranging from 0 to 5. After calculating the severity index for each record in the response missing set, they are sorted in descending order of index size. The response missing severity index of HVAC main unit level devices is usually much higher than that of end devices. Records with a severity index higher than 3 are marked as high-severity missing records, and the corresponding device types receive priority reinforcement processing during the building knowledge graph reconstruction phase. The response missing severity distribution is output as an ordered sequence of records arranged in descending order of severity index. Records with high-severity missing in the response missing severity distribution are marked for priority processing in sub-step 4 weighted mapping.

[0063] The severity distribution of missing responses is weighted and mapped to generate behavioral bias values. The severity index of each record in the ordered sequence of the missing response severity distribution is normalized before participating in the weighted mapping. Normalization uses the maximum possible severity of 5 as the denominator, compressing the severity index of each record to 0-1. The weighted mapping assigns a weight coefficient of 2.0 to high-severity missing records, 1.2 to medium-severity records, and 0.6 to low-severity records. Missing responses from core equipment such as building chillers and cooling towers fall into the high-severity range due to their high priority, and their weight amplification effect ensures that the linkage response failure of cold source-side equipment is fully reflected in the behavioral bias value. Missing responses from the coordinated speed regulation of chilled water pumps and cooling tower fans on the distribution side fall into the medium-severity range, and their weight reflects the medium-level impact of distribution-side failures on the overall system's cooling capacity. Occasional missing responses from the temperature control valves of terminal fan coil units fall into the low-severity range due to their low priority, and the low weight coefficient ensures that the brief response delays of such terminal fine-tuning equipment do not excessively amplify the behavioral bias value. The three-tiered weighting design ensures that behavioral deviation values ​​are highly sensitive to anomalies on the cold source side during multi-device coordinated cooling operations in summer buildings, while remaining tolerant of fluctuations within the normal adjustment range of terminal equipment, preventing false exceedances caused by frequent fine-tuning of terminal equipment. The behavioral deviation value is calculated using the formula V_deviation=Σ(Severity_i×w_i) / N_expected, where Severity_i is the normalized severity index of the i-th record, w_i is the corresponding weight coefficient, and N_expected is the total number of elements in the expected time series. Normalization of the denominator ensures that the behavioral deviation value does not exhibit systematic deviations due to the number of HVAC control action steps in the building.

[0064] The building knowledge graph is constructed by adaptively reconstructing the building knowledge graph based on the behavioral deviation value for hierarchical over-limit identification. After generation, the behavioral deviation value is compared with the preset three over-limit thresholds level by level. The three thresholds accurately correspond the reconstruction strategy with the severity of the deviation, avoiding unnecessary full-graph update overhead caused by slight deviations. When the behavioral deviation value is in the first level of over-limit, the local causal edge weight correction is performed on the subset of single-professional domain nodes in the deviation concentration. The correction weight is dynamically calculated based on the deviation between the measured state transition direction and the original causal inference chain path strength, and only the edge set directly associated with high-severity missing devices is updated. When the behavioral deviation value is in the second level of over-limit, cross-professional causal edges are supplemented or corrected across professional domain boundaries, and the propagation coverage insufficiency rate assessment of the affected area is re-executed to drive the directional knowledge injection sequence to perform incremental injection on the newly identified weak areas. When the behavioral deviation value exceeds the third level of over-limit, the core control node subgraph is fully reconstructed, and the measured state transition relationship of high-severity missing devices is written into the building knowledge graph in batches. At the same time, the knowledge enhancement control model is fully re-injected to ensure the semantic consistency between the model and the reconstructed graph. After each level of reconstruction is completed, the building knowledge graph performs a global connectivity index recalculation. If the number of connected components does not increase and the average path length of each professional domain subgraph decreases, the reconstruction is deemed valid. The version number is incremented, and a reconstruction summary is generated and written to the graph's history archive. Reconstruction is not triggered under normal fluctuation conditions where the behavior deviation value is below the lower limit. Equipment behavior data is updated incrementally in a lightweight manner to update the running frequency statistics field of each node. After several consecutive inference and decision cycles, a batch edge weight update is triggered to ensure continuous alignment between the building knowledge graph and the actual operating status of the HVAC system.

[0065] To implement the above method embodiments, a large model-driven building knowledge graph construction method is provided to achieve the corresponding functionalities and technical effects. See also Figure 2 , Figure 2 This diagram illustrates a structural block diagram of a large model-driven building knowledge graph construction system 200 provided in an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The large model-driven building knowledge graph construction system 200 provided in this embodiment includes: The corpus annotation module 201 is used to collect unstructured building document data and perform text cleaning and domain semantic annotation on the unstructured building document data to generate a term feature library. The knowledge graph construction module 202 is used to perform multi-source description conflict detection on the unstructured building document data based on the terminology feature library to generate a set of conflict knowledge points, input the set of conflict knowledge points into a large language model to perform cross-source semantic difference quantification to generate a semantic difference matrix, and perform difference-oriented knowledge fusion on the semantic difference matrix to construct a building knowledge graph. The knowledge injection module 203 is used to extract graph knowledge vectors from the building knowledge graph, identify knowledge backlog nodes and determine knowledge weak areas based on the topological structure of the building knowledge graph, and inject the graph knowledge vectors into the knowledge weak areas to generate a knowledge enhancement control model. The reasoning and decision-making module 204 is used to perform bidirectional causal reasoning of the HVAC system based on the knowledge-enhanced control model to generate a decision reasoning chain, perform directional violation detection on the decision reasoning chain to generate low-confidence nodes, and perform weight reduction processing on the low-confidence nodes to output an interpretable control decision. The knowledge graph reconstruction module 205 is used to generate equipment behavior data based on the interpretable control decision-driven operation of the HVAC system, compare the equipment behavior data with the decision reasoning chain to generate behavior deviation values, and perform hierarchical over-limit identification based on the behavior deviation values ​​to trigger the adaptive reconstruction of the building knowledge graph to complete the construction of the building knowledge graph.

[0066] The large-model-driven building knowledge graph construction system 200 described above can implement a large-model-driven building knowledge graph construction method according to the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0067] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for constructing a building knowledge graph driven by a large model, characterized in that, include: Collect unstructured document data of buildings, and perform text cleaning and domain semantic annotation on the unstructured document data to generate a term feature library; Based on the terminology feature library, multi-source description conflict detection is performed on the unstructured building document data to generate a conflict knowledge point set. The conflict knowledge point set is then input into a large language model to perform cross-source semantic difference quantification to generate a semantic difference matrix. Finally, difference-oriented knowledge fusion is performed on the semantic difference matrix to construct a building knowledge graph. Extract graph knowledge vectors from the building knowledge graph, identify knowledge backlog nodes and determine knowledge-weak areas based on the topological structure of the building knowledge graph, and inject the graph knowledge vectors into the knowledge-weak areas to generate a knowledge enhancement control model; Based on the knowledge-enhanced control model, bidirectional causal reasoning of the HVAC system is executed to generate a decision reasoning chain. Direction violation detection is performed on the decision reasoning chain to generate low-confidence nodes. The low-confidence nodes are then deweighted to output an interpretable control decision. Based on the interpretable control decision-driven operation of the HVAC system, the system generates equipment behavior data. The equipment behavior data is compared with the decision reasoning chain to generate a behavior deviation value. Based on the behavior deviation value, the system performs hierarchical over-limit identification to trigger the adaptive reconstruction of the building knowledge graph and complete the construction of the building knowledge graph.

2. The method according to claim 1, characterized in that, The process of text cleaning and domain semantic annotation of the unstructured building document data to generate a terminology feature library includes: The unstructured document data of the buildings is subjected to noise filtering and format normalization to generate a clean corpus; The cleaned corpus is used to identify abnormal descriptive paragraphs and generate an abnormal corpus set and a normal corpus set. Prioritize semantic annotation on the abnormal corpus set and de-weight annotation on the normal corpus set to generate differentiated annotated corpus; Based on the differentiated annotated corpus, terminology is extracted and encoded to generate a terminology feature library.

3. The method according to claim 1, characterized in that, The step of generating a conflict knowledge point set by performing multi-source description conflict detection on the unstructured building document data based on the terminology feature library includes: Based on the terminology feature library, cross-document terminology role recognition is performed on the unstructured building document data to generate a terminology role distribution; The terminology role distribution identifies multiple role conflict locations for the same term, generating a role conflict set; The conflict intensity of the aforementioned conflict set is evaluated to generate a conflict intensity distribution; Based on the aforementioned conflict intensity distribution, high-conflict areas are identified, and a set of conflict knowledge points is generated.

4. The method according to claim 1, characterized in that, The process of identifying knowledge backlog nodes and determining knowledge-weak areas based on the topological structure of the building knowledge graph includes: The in-degree to out-degree ratio of each node in the building knowledge graph is calculated to generate a knowledge flow ratio distribution. A knowledge backlog node set is generated by identifying high in-degree and low out-degree nodes from the knowledge flow ratio distribution. The propagation impact range of the knowledge backlog node set is assessed to generate a propagation coverage insufficiency rate; The knowledge gap area is identified based on the aforementioned insufficient coverage rate.

5. The method according to claim 1, characterized in that, The step of injecting the knowledge vectors from the knowledge graph into the knowledge-weak regions to generate a knowledge-enhancing control model includes: The knowledge vector in the graph is partitioned and mapped according to the knowledge-weak regions to generate regional knowledge feature vectors; A knowledge-intensive feature set is generated by identifying high-density feature dimensions from the region's knowledge feature vector. Based on the knowledge-intensive feature set, a targeted knowledge injection sequence is generated by redistributing knowledge in high-density areas. A knowledge enhancement control model is generated by performing gradient-guided knowledge injection based on the described targeted knowledge injection sequence.

6. The method according to claim 1, characterized in that, The step of generating low-confidence nodes by detecting directional violations in the decision reasoning chain includes: The decision reasoning chain is subjected to step-by-step causal direction extraction to generate a reasoning causal sequence; The inference causal sequence is used to perform directional violation clustering analysis to generate violation clustering patterns; To differentiate between local random violations and global systematic violations based on the aforementioned violation clustering patterns, violation classification labels are generated. Based on the violation classification label, random violation class nodes are extracted to generate low-confidence nodes.

7. The method according to claim 1, characterized in that, The step of comparing the device behavior data with the decision reasoning chain to generate a behavior deviation value includes: Extract the expected behavior trigger timing from the decision reasoning chain to generate the expected timing sequence; The device behavior data is matched with the expected time sequence to generate a missing response set; The severity distribution of response missing data is generated by jointly evaluating the duration of response missing data and the trigger priority of the missing data set. The severity distribution of missing responses is weighted and mapped to generate behavioral bias values.

8. The method according to claim 3, characterized in that, The step of identifying multiple role conflict locations for the same term in the terminology role distribution and generating a role conflict set includes: A cross-document role comparison is performed on each term in the terminology role distribution to generate a role comparison sequence; Generate a set of role-reversal positions by identifying document positions where the same term assumes opposite roles from the role comparison sequence; Local contextual semantic features are extracted from the set of inverted roles to generate conflict contextual features; Based on the conflict context features, conflict confirmation is performed to generate a set of role conflicts.

9. The method according to claim 6, characterized in that, The step of generating violation classification labels to distinguish between local random violations and global systematic violations based on the violation clustering pattern includes: Cluster size distributions are generated by statistically analyzing the cluster size of the violated clustering patterns. From the cluster size distribution, large-scale clustering across inference steps is identified to generate a global set of systematic violations; The residual clustering after excluding the global systematic violation set from the violation clustering pattern generates a local random violation set; Violation classification labels are generated based on the global systematic violation set and the local random violation set.

10. A large-model-driven building knowledge graph construction system, characterized in that, include: The corpus annotation module is used to collect unstructured building document data, and to perform text cleaning and domain semantic annotation on the unstructured building document data to generate a term feature library; The knowledge graph construction module is used to perform multi-source description conflict detection on the unstructured building document data based on the terminology feature library to generate a conflict knowledge point set, input the conflict knowledge point set into a large language model to perform cross-source semantic difference quantification to generate a semantic difference matrix, and perform difference-oriented knowledge fusion on the semantic difference matrix to construct a building knowledge graph. The knowledge injection module is used to extract graph knowledge vectors from the building knowledge graph, identify knowledge backlog nodes and determine knowledge-weak areas based on the topological structure of the building knowledge graph, and inject the graph knowledge vectors into the knowledge-weak areas to generate a knowledge enhancement control model. The reasoning and decision-making module is used to perform bidirectional causal reasoning of the HVAC system based on the knowledge-enhanced control model to generate a decision reasoning chain, perform directional violation detection on the decision reasoning chain to generate low-confidence nodes, and perform weight reduction processing on the low-confidence nodes to output an interpretable control decision. The knowledge graph reconstruction module is used to generate equipment behavior data based on the interpretable control decision-driven operation of the HVAC system, compare the equipment behavior data with the decision reasoning chain to generate behavior deviation values, and perform hierarchical over-limit identification based on the behavior deviation values ​​to trigger the adaptive reconstruction of the building knowledge graph to complete the construction of the building knowledge graph.

Images