A method for rapid fault attribution and self-optimization of construction equipment

Abstract

The application discloses a kind of building equipment failure fast attribution and self-optimization method, it is related to building equipment intelligent operation and maintenance field, the multidimensional association between equipment, building space, sensor, fault mode is established in the application, simultaneously realizes data-driven flexibility and physical logic rigor in fault attribution, realizes inference from abnormal data to root cause analysis.The application significantly improves optimization efficiency through specific incremental learning / local updating method, adjusts the weight and bayesian network conditional probability table of knowledge graph according to maintenance result, calibrates physical equation parameters, realizes the dynamic updating of knowledge graph node attribute;Through dynamic weight optimization, improve real-time inference confidence evaluation and the adaptive ability of scene context;Through the establishment and deep integration of system dynamics model, physical equation is embedded as executable knowledge unit in knowledge graph, and safety and reliability mechanism is designed for verification.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of building equipment, specifically to a method for rapid attribution and self-optimization of building equipment failures. Background Technology

[0002] Bayesian networks handle uncertain information through probabilistic reasoning and are widely used in fault diagnosis in fields such as power systems and aerospace equipment. By integrating historical maintenance data through knowledge graphs and combining it with Bayesian networks to generate maintenance process recommendations, fault tree models are converted into Bayesian networks, enhancing the systematic nature of fault analysis. Structured representation of equipment relationships (such as physical connections and logical dependencies) supports rapid relational queries and integrates expert experience with historical samples, improving diagnostic interpretability. Existing technology (CN119067628A) only uses static rules to construct knowledge graphs, which cannot handle parameter drift caused by equipment aging.

[0003] Some studies attempt to combine Bayesian networks with machine learning (such as GNNs) to optimize inference paths through dynamic weight allocation. For example, one approach uses reinforcement learning to adjust the weights of knowledge graph nodes to reduce the false positive rate. While existing technologies (CN113779872A) combine Bayesian networks and machine learning, their weight allocation is preset or based on simple rules (such as `def dynamic_weight(fault_type)`), lacking adaptive capabilities based on real-time inference confidence assessment and contextual information.

[0004] A few studies have embedded physical equations (such as fluid dynamics formulas) into knowledge graphs to verify the physical consistency of fault reasoning. Existing technologies (CN113779872A) generally lack the ability to deeply integrate first-principles physical models (such as Bernoulli equations) into the reasoning process for physical consistency verification.

[0005] In summary, existing technologies, which only use static rules to construct knowledge graphs, cannot handle parameter drift caused by equipment aging, lack the ability to adapt based on real-time reasoning confidence assessment and scene context, and lack the ability to deeply integrate first-principles physical models (such as Bernoulli equations) into the reasoning process for physical consistency verification. As a result, it is difficult to achieve rapid attribution and self-optimization of building equipment failures. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a rapid attribution and self-optimization method for building equipment failures. This method solves the problems of existing technologies, which rely solely on static rules to construct knowledge graphs, making it difficult to handle parameter drift caused by equipment aging, lacking adaptive capabilities based on real-time reasoning confidence assessment and scenario context, and lacking the ability to deeply integrate first-principles physical models (such as Bernoulli equations) into the reasoning process for physical consistency verification.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0008] A method for rapid attribution and self-optimization of building equipment failures is provided, which includes the following steps:

[0009] Collect and construct a knowledge graph corresponding to the building equipment based on the structured data of the building equipment:

[0010] The system maps instance parameters of building equipment to the energy consumption rule engine of the knowledge graph to obtain energy consumption rule nodes; converts material properties into physical calculation factors in the knowledge graph to obtain physical calculation attribute nodes; constructs a three-dimensional spatial relationship network based on the spatial coordinates of the equipment; constructs a logical dependency tree based on the connection relationship between the equipment; uses component-level BIM parameters as fault propagation path nodes in the knowledge graph; generates corresponding SWRL inference rules for BIM annotations and fault codes; and imports the FTA structure from the equipment manual into the knowledge graph to obtain OWL rules and Bayesian network initial condition probability tables in the knowledge graph.

[0011] Building a failure attribution engine:

[0012] Based on the topology, logical dependency tree, and OWL rules of the knowledge graph corresponding to building equipment, a Bayesian network is constructed to form a Bayesian network engine; a meta-model is constructed based on SWRL inference rules and OWL rules to form a rule inference engine; and a graph neural network is constructed based on the nodes and edges of the knowledge graph to form a graph neural network engine.

[0013] Establish and deeply integrate the system dynamics model:

[0014] The physical equations are encapsulated into executable physical equation nodes in the knowledge graph; a two-layer triggering mechanism is established based on real-time data and physical equation nodes; and physical equation nodes and device nodes are deeply integrated to obtain dynamic binding relationships in the knowledge graph. The two-layer triggering mechanism is used to obtain real-time physical contradiction markers through physical verification.

[0015] Perform multi-source data extraction:

[0016] Unstructured knowledge related to building equipment is extracted from different data sources, and entities, attributes and relationships are extracted from the unstructured knowledge and populated into the knowledge graph corresponding to the building equipment.

[0017] Perform real-time anomaly detection:

[0018] Based on physical contradiction markers and real-time data calculation of the adaptive threshold of the working condition, it is determined whether there is data that reaches the corresponding adaptive threshold. If so, it is marked as an abnormal label and the fault attribution engine is called to output the probability of fault occurrence; otherwise, no action is taken.

[0019] Perform dynamic adjustment of hybrid weights: real-time weights based on physical contradiction marking and real-time data acquisition rule inference engine, Bayesian network engine and graph neural network engine;

[0020] The failure probability outputs from the rule-based reasoning engine, the Bayesian network engine, and the graph neural network engine are weighted and fused according to real-time weights to generate a comprehensive failure probability and obtain the corresponding candidate root causes.

[0021] Repair plans are determined based on candidate root causes; the weights of the knowledge graph and the conditional probability table of the Bayesian network are adjusted based on the repair results, and the parameters of the physical equations are calibrated.

[0022] The beneficial effects of this invention are as follows: This invention establishes a multi-dimensional correlation between equipment, building space, sensors, and failure modes, deeply integrates the failure attribution engine with system dynamics, and utilizes knowledge graphs and Bayesian networks to simultaneously achieve data-driven flexibility and physical logic rigor in failure attribution, enabling reasoning from abnormal data to root cause analysis. This invention significantly improves optimization efficiency through specific incremental learning / local update methods; it adjusts the weights of the knowledge graph and the conditional probability table of the Bayesian network based on maintenance results, and calibrates the parameters of the physical equations to achieve dynamic updates of knowledge graph node attributes; it enhances real-time reasoning confidence assessment and the adaptive capability of scene context through dynamic weight optimization; and it embeds physical equations as executable knowledge units into the knowledge graph through the establishment and deep integration of a system dynamics model, and designs security and reliability mechanisms for verification. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method.

[0024] Figure 2 This is the real-time physical verification process in the embodiment. Detailed Implementation

[0025] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0026] like Figure 1 As shown, the rapid attribution and self-optimization method for building equipment failures includes the following steps:

[0027] S1. Collect and construct a knowledge graph corresponding to the building equipment based on the structured data of the building equipment;

[0028] S2. Building a failure attribution engine:

[0029] Based on the topology, logical dependency tree, and OWL rules of the knowledge graph corresponding to building equipment, a Bayesian network is constructed to form a Bayesian network engine; a meta-model is constructed based on SWRL inference rules and OWL rules to form a rule inference engine; and a graph neural network is constructed based on the nodes and edges of the knowledge graph to form a graph neural network engine.

[0030] S3. Establish and deeply integrate the system dynamics model;

[0031] S4. Perform multi-source data extraction;

[0032] S5. Perform real-time anomaly detection;

[0033] Based on physical contradiction markers and real-time data calculation of the adaptive threshold of the working condition, it is determined whether there is data that reaches the corresponding adaptive threshold. If so, it is marked as an abnormal label and the fault attribution engine is called to output the probability of fault occurrence; otherwise, no action is taken.

[0034] S6. Perform dynamic adjustment of hybrid weights: real-time weights based on physical contradiction marking and real-time data acquisition rule inference engine, Bayesian network engine and graph neural network engine;

[0035] S7. The failure probability output by the rule reasoning engine, Bayesian network engine and graph neural network engine is weighted and fused according to the real-time weights to generate a comprehensive failure probability and obtain the corresponding candidate root causes.

[0036] S8. Determine the repair plan based on the candidate root causes; adjust the weights of the knowledge graph and the conditional probability table of the Bayesian network based on the repair results, and calibrate the parameters of the physical equation.

[0037] In this embodiment, the structured data of the building equipment includes:

[0038] Building types: residential, commercial, public buildings, etc.;

[0039] Building materials: steel, concrete, glass, etc.;

[0040] Building components: such as beams, columns, walls, etc.;

[0041] Building structure: floors, rooms, areas, pipe / line topology;

[0042] Equipment systems: physical equipment (such as chillers, cooling towers, air conditioners), elevators, fire protection, lighting, water supply and drainage, etc.

[0043] Equipment components: compressors, valves, sensors, controllers, etc.;

[0044] Failure Modes: Common Failure Types, Symptoms, Scope of Impact, and Repair Solutions.

[0045] Building a knowledge graph for building equipment involves the following operations:

[0046] 1) Add the BuildingType instance parameter to the BIM model → map it to the energy consumption rule engine of the knowledge graph;

[0047] In the BIM modeling software, create a text type parameter named "BuildingType" and assign a value to this parameter (such as "Commercial Building") through a programmatic interface. This parameter will be dynamically mapped to the energy consumption rule engine of the knowledge graph, driving the generation of subsequent constraints.

[0048] 2) Material properties are written into BIM material attributes → converted into physical calculation factors in the knowledge graph:

[0049] In Building Information Modeling (BIM) software, add two key physical performance parameters to concrete material objects:

[0050] Thermal conductivity: The set value is 1.28 W / (m·K).

[0051] Compressive strength: Set to 30 MPa.

[0052] When exporting a BIM model containing this concrete material as an IFC (Industry Foundation Classes) file, it is necessary to ensure that the added material parameters (thermal conductivity and compressive strength) are correctly preserved in the corresponding material property structure IfcMaterialProperties of the IFC file.

[0053] During the subsequent construction of the knowledge graph, the system will automatically perform the following association operations:

[0054] In the knowledge graph database, the material node (mat:Material) with type "concrete" is located.

[0055] In the knowledge graph database, find all wall nodes (wall:Wall) that are connected to this concrete material node (mat) via the MADE_OF (made of...) relationship.

[0056] For each such concrete material node (mat) found, set its thermal_conductivity attribute to 1.28 (i.e., the value passed from the BIM / IFC data).

[0057] The above parameters will be dynamically mapped to the energy consumption rule engine of the knowledge graph, driving subsequent energy consumption anomaly detection and root cause reasoning.

[0058] 3) Utilize BIM component IDs to construct a 3D spatial relationship network in the knowledge graph:

[0059] When exporting BIM components, retain the GlobalId and spatial coordinates;

[0060] The knowledge graph uses the Neo4j Spatial plugin to build a spatial index;

[0061] Fault impact range calculation: The system locates the specified pipe (ID: P102) and the wall it penetrates using a knowledge graph, and obtains the three-dimensional spatial coordinates of the wall. Then, the system searches all building space nodes defined in the knowledge graph and calculates the distance from each space node to the pipe penetration point. Finally, it returns all space nodes with a distance of less than 5 meters; these spaces represent the areas potentially affected by the pipe fault.

[0062] The above three-dimensional spatial relationship network is the core basis for fault propagation path analysis and impact range calculation.

[0063] 4) Extract BIM device connection relationships → Build a logical dependency tree in the knowledge graph:

[0064] Device connection and control logic generation:

[0065] The system first extracts all physical connection points (connectors) of a specific chiller unit (ID: 1023) from the BIM model and traverses these connection points to determine the devices they are connected to. The system automatically creates a "CONNECTS_TO" relationship between the chiller unit and each connected device in the knowledge graph. Subsequently, the system further establishes control logic in the knowledge graph: the chiller unit node is connected to the pump node via a "CONTROLS" relationship, and the pump node is connected to the valve node with the "ON" status via a "DEPENDS_ON" relationship.

[0066] The above logical dependency tree forms the basis for constructing Bayesian network structures and performing system-level cascading failure analysis.

[0067] 5) Component-level BIM parameters → Fault propagation path nodes in the knowledge graph:

[0068] The BIM parameter names include IsCriticalComponent and MeanTimeBetweenFailure; the knowledge graph mapping for IsCriticalComponent is fault_anchor:true, and the fault propagation rule is that a failure of this component triggers a system-level alarm. The knowledge graph mapping for MeanTimeBetweenFailure is mtbf:15000, and the fault propagation rule is used for Bayesian network lifetime prediction.

[0069] Dynamic update mechanism:

[0070] Dynamic updates to valve reliability: When valve (ID: V201) has been repaired, the system automatically locates the valve node in the knowledge graph and increases its mean time between failures (MTBF) attribute value by 20% (i.e., mtbf = mtbf * 1.2), thereby reflecting the improvement in valve reliability.

[0071] The above dynamic update mechanism directly serves the self-optimization of the knowledge graph.

[0072] 6) BIM annotation associated with fault codes → Knowledge graph generation of multi-level diagnostic rules:

[0073] In Revit software, the system creates a fault annotation object for a specified device (identified by elementId). This annotation contains the fault code "FC-203" and the symptom description "abnormally high outlet water temperature". Subsequently, based on this annotation information, the system automatically generates corresponding SWRL inference rules in the knowledge graph: when a component has fault code "FC-203" and its outlet water temperature is higher than 45 degrees Celsius, the system will recommend performing a "clean condenser" maintenance operation on that component.

[0074] In this embodiment, since the collected structured data may have different data sources, multi-source data fusion is required:

[0075] 1) IoT Platform: The Kalman filter algorithm is used to clean real-time sensor data (temperature, humidity, peak and valley electricity consumption, pressure, current, power, etc.). Real-time data is accessed through protocols such as Modbus, MQTT, and OPC UA. The data is bound to knowledge graph nodes in real time based on device ID / UUID mapping.

[0076] The cleaned real-time sensor data is directly bound to the corresponding nodes in the knowledge graph, serving as key evidence input for real-time anomaly detection, physical equation calculation, and Bayesian inference.

[0077] For example, temperature sensor data cleaning, assuming the temperature changes slowly and there is no control input;

[0078] If there are multiple sensors (such as temperature and humidity combined cleaning), expand x, P, F, H, Q, Rx,P,F,H,Q,R into a matrix form.

[0079] Regarding parameter tuning, Q and R are adjusted using actual data, and the Q / R ratio affects the filter's confidence in the model / observations.

[0080] If the sensor data and the state have a non-linear relationship, then use an extended Kalman filter (EKF) or an unscented Kalman filter (UKF).

[0081] Using the methods described above, Kalman filtering can effectively smooth noise and track real-time signals, making it suitable for various sensor data cleaning scenarios.

[0082] 2) Fault Tree Analysis (FTA) Structured Import into Knowledge Graph: A three-layer parsing framework transforms the FTA from the equipment manual into an executable knowledge graph. First, a semantic parsing engine deconstructs the fault tree: ① Root nodes (e.g., "cooling system failure") are mapped to fault class entities in the knowledge graph; ② Logic gates (AND / OR) are converted into OWL axioms, e.g., water pump failure AND valve blockage → cooling failure; ③ Leaf nodes (basic events such as "bearing wear") are associated with equipment nodes and injected with probabilistic attributes (initial probability = 0.05). Key rules are represented using SWRL semantics, e.g., (vibration value > 7.5 mm / s triggers bearing wear risk). To handle dynamic scenarios, a hybrid probability-rule engine is designed: static dependencies (e.g., equipment cascading shutdowns) are constrained by hard SWRL rules; dynamic probabilities (e.g., increased failure rate due to aging) are dynamically updated using a Bayesian network CPT. Implementation Results: A centrifuge unit's Free Analytical Assessment (FTA) (containing 158 event nodes) completed knowledge graph transformation within 12 seconds, generating 203 OWL axioms and 47 SWRL rules, reducing the false diagnosis rate by 32%. The generated OWL axioms and SWRL rules constitute the core knowledge base of the rule-based reasoning engine and provide the basis for the initial assignment of the Bayesian network's CPT. This framework enables machine-readable fault logic and provides deterministic support for hybrid reasoning.

[0083] 3) Unstructured Data Processing: A multimodal fusion architecture is employed to dynamically update the knowledge graph. For text processing, a domain-fine-tuned RoBERTa model is used to parse maintenance logs. Entity recognition (extracting faulty equipment, components, phenomena, etc.) is performed using BiLSTM-CRF, and a dynamic fault dictionary (e.g., terms like "bearing seizure") is constructed using incremental clustering algorithms, achieving an accuracy of 92.1%. For maintenance images, a YOLOv8-Seg segmentation model is used to quantify pipeline corrosion area (calculating percentage levels), a ResNet50+GRU time-series model is used to identify liquid leakage features, and ViT contrastive learning is used to detect component defects (outputting similarity to standard images). Extracted semantic and visual features (e.g., corrosion level 4) are transformed into structured attributes and updated to corresponding nodes in the knowledge graph through automated mapping rules (e.g., updating pipeline remaining life attributes). The extracted structured attributes (e.g., corrosion level) dynamically update knowledge graph node attributes, directly impacting RUL prediction and maintenance strategy risk labeling.

[0084] 4) Resolving multi-source data conflicts: (e.g., text records are normal but images show corrosion). The system introduces DS evidence theory to fuse confidence scores (0.7 * image confidence score + 0.3 * text confidence score). If the score exceeds a threshold of 0.8, a knowledge graph update is triggered. After each repair loop closure, the system dynamically adjusts model weights based on the results (+0.1 confidence score for success, -0.2 for failure). This solution achieves three major breakthroughs: ① Domain-adaptive NLP addresses the lack of industry terminology, improving fault extraction accuracy by 22.5%; ② Visual feature quantification converts corrosion images into graded parameters with an error of <5%; ③ Multimodal conflict decision-making accuracy reaches 89.3%, making unstructured data the core fuel driving the self-optimization of the knowledge graph.

[0085] 5) Sub-second updates of sensor data to the knowledge graph: Implemented through a stream processing architecture, using Apache Kafka to access real-time data streams (temperature, pressure, etc.) from the IoT platform. After Kalman filtering to remove outliers, the attribute mapping engine dynamically binds to device nodes: ① Basic values ​​(e.g., current water temperature 32°C) are directly written to the node's currentValue attribute; ② Statistical indicators (e.g., pressure variance over the past hour) are calculated and updated to the stats attribute. The update frequency is controlled in three levels: critical equipment (e.g., fire pumps) undergoes high-priority stream processing every 200ms, ordinary equipment (air conditioners) uses 1-second window aggregation updates, and historical statistical indicators are calculated in 15-minute batches. To reduce storage overhead, a dynamic hot and cold partitioning strategy is designed: real-time data retains only 72 hours of original values, while long-term trends are sampled and compressed for storage at the minute level. In implementation, 38,000 sensor data points in a commercial complex were updated across the entire knowledge graph within 1.2 seconds, reducing database load by 70%. This mechanism makes the knowledge graph the dynamic nerve center of building equipment, supporting millisecond-level fault response. Real-time updated data is the foundation for achieving millisecond-level response and real-time physical verification.

[0086] In this embodiment, the Bayesian network structure is designed as follows:

[0087] Node types: device entities, sensors, fault modes, physical equation entities (as a special type of node / attribute), location space, device parameters (such as temperature, pressure), fault types (such as valve blockage, sensor drift), environmental variables (such as humidity, load rate).

[0088] Node definition: Three types of random variables: 1) Device failure status (e.g., Pump_Failure); 2) Environmental parameters (e.g., Cooling_Load); 3) Sensor anomalies (e.g., Temperature_Sensor_Drift).

[0089] Relationship types (physical connection hasPhysicalConnectionTo, logical dependency dependsOn, fault association causes / symptomOf, equation association givenBy...), attributes (device specifications specs, real-time sensor values ​​currentValue, historical statistics stats, equation parameters params, dynamic weights dynamicWeight, confidence...).

[0090] Edge relationships: Based on triple dependencies: 1) Physical causality (causes: valve blockage → abnormal flow); 2) Logical dependency (dependsOn: cooling tower operation requires water pump startup); 3) Statistical association (correlation coefficient between vibration value and bearing failure in historical data > 0.7). This is expanded into a DBN through time slicing to dynamically model the evolution of equipment states.

[0091] Edge connection rules: Directed edges are established based on physical dependencies (such as "abnormal cooling water flow → water pump failure") and historical fault data.

[0092] 1) Dynamic binding of entity nodes in physical equations:

[0093] Equation nodes store executable code: lambda ρ,v,L,D,f: f*L / D*ρ*v**2 / 2;

[0094] Automatic parameter mapping: ρ ← water temperature sensor, v ← flow meter;

[0095] Output as parent node: Bernoulli equation. ΔP → Pump node. Theoretical pressure.

[0096] 2) The triple dependency modeling is shown in Table 1 below.

[0097] Table 1

[0098]

[0099] 3) Time-slice evolution model:

[0100] For example, the probability rules for the state transition of a water pump:

[0101] Under normal conditions: there is an 85% probability of remaining normal, a 10% probability of transitioning to a wear-out state, and a 5% probability of transitioning to a fault state.

[0102] When in a worn state: there is a 15% probability of returning to a normal state, a 60% probability of remaining in a worn state, and a 25% probability of transitioning to a fault state.

[0103] In a faulty state: It cannot automatically recover (the probability of returning to a normal or worn state is 0), and there is a 100% probability that it will remain in a faulty state (manual maintenance is required).

[0104] 4) Dynamically weighted edges:

[0105] Relationship confidence dynamic adjustment mechanism:

[0106] The system dynamically updates the confidence weights (w) of the relationships between nodes in the knowledge graph based on actual repair results and environmental factors.

[0107] Positive adjustment: The increase in weight is proportional to the repair success rate triggered by this relationship (number of successful repairs / total number of triggers), and the adjustment strength is controlled by the coefficient α;

[0108] Negative adjustment: The reduction in weight is proportional to the environmental change coefficient characterizing changes in external conditions, and the degree of influence is controlled by the coefficient β;

[0109] Update formula: The new confidence weight (w_new) equals the old weight (w_old) plus the increment based on the repair success rate, minus the reduction based on environmental changes.

[0110] 5) Real-time injection of knowledge graph data;

[0111] 6) Collaborative computation of fault propagation paths:

[0112] Knowledge graph locates the physical path: valve blockage → increased pipeline pressure → water pump overload;

[0113] Bayesian network quantification probabilities: P(pump failure|valve blockage) = 0.75, P(pipe rupture|pressure increase) = 0.62.

[0114] Generate the optimal detection sequence: check valves → measure pressure → diagnose water pumps.

[0115] In this embodiment, the establishment and deep integration of the system dynamics model includes the following operations:

[0116] 1. Executable encapsulation of physical equations:

[0117] The knowledge graph node design creates the EquationEntity entity type, which includes the core attributes shown in Table 2.

[0118] Table 2

[0119]

[0120] `lambda_func` and `params_binding` enable the physics equations to be invoked and executed dynamically in real time. Device-equation dynamic binding connects device nodes and equation nodes via the `GOVERNED_BY` relationship:

[0121] For example, a "GOVERNED_BY" association can be established between pump device nodes and physical equation nodes in a knowledge graph. The equation node stores the unique identifier "EQ_Bernoulli_PressureDrop" of the Bernoulli equation, achieving a deep binding between the device and the physical model.

[0122] 2. Real-time physical verification process (e.g.) Figure 2 (as shown)

[0123] If parameter binding fails (e.g., sensor is offline), automatically switch to historical mean filling and mark a decrease in confidence of 0.3.

[0124] When the calculated result of the equation deviates from the measured value by more than 5%, a physical contradiction marker is generated and passed to step 3. The physical contradiction marker will serve as a key input for the dynamic weight allocation of the hybrid inference engine (especially the conflict arbitration mechanism). When the marker is true or the deviation exceeds the threshold, the weight redistribution logic will be directly triggered, and some candidate root causes that violate physical laws may be rejected.

[0125] 3. Intelligent dual-layer triggering mechanism:

[0126] 1) Basic triggering layer (dynamic threshold monitoring):

[0127] When three consecutive sensor readings exceed the device's current adaptive dynamic range, the system triggers a physical equation verification. The formula for calculating this dynamic range is:

[0128] Threshold = Current operating condition baseline value (μ) ± 2 × Historical data standard deviation (σ)

[0129] in:

[0130] μ (baseline value): Calculated using a sliding window based on the real-time operating status of the equipment.

[0131] σ (standard deviation): Calculated based on the historical data of this sensor, reflecting the degree of data dispersion.

[0132] If the physical equation verification performed by calling the basic trigger layer passes, the hybrid weights are dynamically adjusted to improve the adaptability of the static rules. If the physical equation verification result performed by calling the basic trigger layer is a transient anomaly, it is determined that there is a device start-up / shutdown disturbance. If the physical equation verification result performed by calling the basic trigger layer is a persistent deviation, it is determined that there is device aging. If the physical equation verification result performed by calling the basic trigger layer is a system conflict, it is determined that there is model parameter inaccuracy. The adaptive dynamic range is the current operating condition baseline value ± twice the standard deviation of historical data. The current operating condition baseline value is calculated through a sliding window based on the real-time operating status of the device.

[0133] 2) Arbitration Triggering Layer (Hybrid Reasoning Conflict Handling):

[0134] When multiple candidate faults output by the Bayesian network simultaneously satisfy the following two conditions, the system will trigger the verification of the physical equations:

[0135] The confidence level of the highest probability fault is insufficient: among all candidate faults, the highest probability fault has a value below 0.7 (max(P(fault_i))<0.7).

[0136] There are other significantly possible faults: the probability of at least one candidate fault is greater than 0.3 (∃ P(fault_j)>0.3).

[0137] Response Action: The system will invoke relevant physical equations (such as Bernoulli's equation and laws of thermodynamics) according to preset priorities to verify the physical rationality of candidate fault assumptions. Specifically, invoking the arbitration trigger layer to verify physical equations involves: invoking relevant physical equations according to preset priorities to verify the physical rationality of candidate faults; if the physical equation verification performed by the arbitration trigger layer passes, dynamic adjustment of the mixed weights is performed to improve the adaptability of static rules; otherwise, the following operations are performed:

[0138] Reject all candidate faults that are verified to violate the laws of physics;

[0139] The system indicates a physical contradiction and invokes the fault attribution engine to output the probability of the fault occurring.

[0140] When the persistent deviation exceeds 24 hours, calibrate the physical equation parameters.

[0141] The arbitration triggering layer solves the decision-making dilemma of Bayesian networks in scenarios with multiple candidates and low confidence.

[0142] By deeply embedding first-principles physical equations into a knowledge graph, real-time physical consistency verification for fault diagnosis is achieved, addressing the core defect of "lack of physical logic" in traditional methods.

[0143] In this embodiment, the establishment and deep integration of the system dynamics model includes the following operations:

[0144] 1. Deep integration of physical equations:

[0145] The integration here involves reusing the defined executable equation nodes and binding mechanisms.

[0146] 1) Executable equation nodes:

[0147] Create EquationEntity entity nodes in the knowledge graph to store metadata (ID, parameter definition) and executable code (Lambda format) of physical equations (such as Bernoulli's equation).

[0148] Dynamic parameter binding: Device nodes are associated through the GOVERNED_BY relationship, and real-time data (such as sensor value → fluid density ρ) is automatically mapped.

[0149] Example: Pump node correlation equation ΔP = f(L / D)(ρv 2 / 2), Parameter binding rules: ρ←water temperature sensor, v←flow meter.

[0150] When the hybrid inference engine detects a physical contradiction marker (output deviation > 5%), it automatically triggers weight reassignment:

[0151] "IF physical contradiction THEN, rule weight = 0.1, GNN weight = 0.7, Bayesian weight = 0.2".

[0152] 2) Two-layer trigger verification:

[0153] Basic layer: Equation calculation is triggered when sensor data exceeds the dynamic threshold (reference value ± 2σ) three times consecutively.

[0154] Arbitration layer: When the Bayesian network outputs multiple candidate faults (probability > 30%), the equation is forcibly invoked to verify physical consistency (if the deviation > 5%, the hypothesis is rejected).

[0155] 2. Dynamic optimization of the hybrid inference engine

[0156] 1) Weight Adaptive Algorithm:

[0157] Define the real-time weights of the rule-based reasoning engine (R), the Bayesian network engine (B), and the GNN engine (Graph Neural Network Engine) (G):

[0158] An adaptive weight algorithm is designed to calculate the real-time weights of three types of engines—rule-based inference, Bayesian networks, and graph neural networks—using a normalized exponential function. The calculated real-time weight W... i(Rules, Bayesian methods, and GNNs) are the core basis for the root cause reasoning engine to select and execute different reasoning paths. Calculation factors include:

[0159] α: Confidence of the current inference result (based on Bayesian posterior probability). It may come from the posterior probability of Bayesian inference, the confidence output of rule / GNN inference, the credibility of the physical verification result, etc.

[0160] β: Matching degree between the fault scenario and historical cases (calculated through the similarity of knowledge graph subgraphs). Based on the similarity matching between the current abnormal scenario and the historical case subgraphs in the knowledge graph.

[0161] γ: Historical accuracy of the engine in similar scenarios.

[0162] The final weight allocation result is used to dynamically select the optimal inference path.

[0163] α, β, and γ are optimized online through reinforcement learning: the repair result is used as the reward, and the loss function L = -log(P(actual fault|predicted fault)) is minimized using gradient descent.

[0164] Conflict arbitration mechanism:

[0165] The generated physical contradiction markers have the highest priority in this mechanism, meaning physical verification takes precedence. The dynamic weights (rule, GNN, Bayesian weight ratio) and physical contradiction markers output in this step will serve as direct inputs for root cause ranking and maintenance recommendation generation.

[0166] 2) Inference engine co-design:

[0167] It mainly originates from the constructed knowledge graph (node / edge attributes, topology) and real-time abnormal states. Its output serves as one of the candidates for root cause reasoning and participates in weighted fusion.

[0168] GNN architecture (GraphSAGE):

[0169] Input: Node attributes (device type / status value), edge attributes (relationship type / dynamic weight).

[0170] Output: Node-level failure probability + edge-level propagation path probability.

[0171] Bayesian network incremental update mechanism:

[0172] After equipment maintenance, the system performs local updates only on the faulty node and its directly associated parent / child nodes (Markov blanket). The Dirichlet prior Bayesian update algorithm is used to correct the conditional probability table (CPT).

[0173] New probability = (historical data weight × old probability + actual observation count) / (historical data weight + total observation count)

[0174] Triggering conditions typically arise from maintenance feedback. The updated CPT is directly used for subsequent Bayesian inference calculations. Emphasis is placed on local updates, which significantly improve inference efficiency.

[0175] Example of effect: After the "bearing wear" fault is repaired, the system only updates the local probability of the related pump node, reducing the computational overhead by 40%.

[0176] 3. Closed-loop driven self-optimization of knowledge graphs:

[0177] 1) Dynamic parameter tuning for reinforcement learning:

[0178] State: <fault type, environmental parameters, knowledge graph substructure hash>.

[0179] Action: Adjust node weight (±0.1) / Add faulty edge / Delete redundant node.

[0180] Rewards: Successful repair (+1.0), false positive (-1.5), reduced response time (+0.3 / 10%).

[0181] State information (fault type, environmental parameters, subgraph structure) comes from anomaly detection, environmental monitoring, and knowledge graph.

[0182] The results of actions (adjusting weights, adding or deleting edges) directly affect the knowledge graph, influencing all subsequent steps that depend on the knowledge graph.

[0183] The calculation of the reward depends on the execution results of the maintenance suggestion (from the work order system) and the response time.

[0184] 2) Localized incremental update:

[0185] The system uses Cypher queries to precisely locate the current faulty node (n.fault=true) and all related nodes within its three-level relationship neighborhood (i.e., nodes directly connected to the faulty node, indirectly connected (two steps), or further away but still within the three-step relationship range). Update operations are performed only on the subgraph structure within this defined range, focusing on optimizing the association weights between nodes within that subgraph.

[0186] Equation parameter calibration: When the simulation deviation is >5%, the equipment binding parameters (such as the pipeline friction factor f) are automatically adjusted.

[0187] The scope and target of localized incremental updates (fault neighborhood subgraph, parent / child nodes (Markov blanket)) ensure the efficiency of the optimization, enabling it to support real-time requirements.

[0188] The result of equation parameter calibration updates the parameters or device node attributes related to the equation node params_binding, which are used for subsequent more accurate physical verification.

[0189] In this embodiment, multi-source data extraction includes:

[0190] Knowledge related to the construction field is extracted from various data sources, including text data (such as academic papers, technical reports, and standards), image data (such as architectural drawings and photographs), and sensor data (such as environmental monitoring data and structural health monitoring data). Entities, attributes, and relationships are extracted using technologies such as natural language processing, image recognition, and data analysis, and then populated into a knowledge representation model.

[0191] This step is an ongoing background process whose output (extracted new knowledge) is continuously injected into the constructed knowledge graph, thereby indirectly affecting all subsequent steps that depend on the knowledge graph.

[0192] Knowledge extraction technology:

[0193] Rule Engine: Employs Natural Language Processing (NLP) technology to extract key information such as fault modes and solutions from unstructured maintenance logs, and defines IF-THEN rules for equipment faults (such as "IF pump pressure < threshold AND water flow rate = 0 THEN possible blockage").

[0194] Graph embedding: using Graph Neural Networks (GNN) to uncover hidden correlations (such as anomalous propagation paths of sensor data).

[0195] Because knowledge from different data sources may contain duplicates, contradictions, or incomplete information, knowledge fusion is necessary to remove duplicates, resolve contradictions, supplement incomplete knowledge, and form a unified knowledge system. Simultaneously, reasoning techniques can be used to derive new knowledge from existing knowledge, enriching the content of the knowledge graph.

[0196] 1. Alignment method for positional conflicts:

[0197] When the system detects that the 3D spatial distance deviation between the location of the same device in the BIM model (bim_loc) and its actual sensor location (sensor_loc) exceeds 1 meter, the system automatically performs position calibration.

[0198] New location = 30% × BIM location + 70% × Sensor location;

[0199] The calibrated new coordinates will be updated to the node position of the device in the knowledge graph.

[0200] 2. Physical parameter conflict alignment method:

[0201] The system uses the theoretical flow rate calculated by Bernoulli's equation as a benchmark. When the ratio of the theoretical flow rate to the sensor's measured flow rate is not within the range of 0.95 to 1.05 (i.e., the deviation exceeds ±5%), the system automatically rejects the conclusion that the valve is in a "normal" state.

[0202] The results of “1. Location Conflict Alignment” and “2. Physical Parameter Conflict Alignment” directly update the location and status attributes of devices in the knowledge graph, ensuring the accuracy of the data.

[0203] 3. Confidence-weighted fusion:

[0204] DS Evidence Theory Integrating Conflicting Evidence:

[0205] The system uses the DS evidence theory formula to calculate the final confidence level (Bel(A)) for a specific event (such as "equipment corrosion"). This formula comprehensively considers the support of different evidence sources and the degree of conflict between them.

[0206] The results of confidence-weighted fusion are used to update the confidence attributes of knowledge graph nodes / edges, affecting weight calculation and root cause ranking.

[0207] Implementation Case:

[0208] Evidence Source A (Image Recognition): The probability of the equipment being corroded is 0.85.

[0209] Source of evidence B (maintenance log text): The probability of corrosion in the equipment is 0.75.

[0210] Evidence weighting: The system assigns a weight of 0.7 to image recognition evidence and a weight of 0.3 to maintenance log text evidence.

[0211] Calculation process:

[0212] Calculate the weighted support: (0.85 × 0.7) + (0.75 × 0.3) = 0.595 + 0.225 = 0.82;

[0213] The conflict of evidence metric is calculated as 0.15 (representing a certain degree of inconsistency between pieces of evidence).

[0214] The final confidence level was calculated using the DS formula: Bel (corrosion) = 0.82 / (1 - 0.15) = 0.82 / 0.85 ≈ 0.89.

[0215] Result: After fusing the two sources of evidence, the overall confidence level of the system in the presence of corrosion in the equipment increased to 0.89.

[0216] 4. Rule-based reasoning engine:

[0217] Equipment maintenance recommendation rules:

[0218] For example, when a device simultaneously meets the following two conditions:

[0219] Its vibration value exceeds 7.5 mm / s, and

[0220] The humidity in its environment exceeds 80%.

[0221] The system will then automatically recommend performing "bearing rust prevention treatment" maintenance on the equipment.

[0222] 5. Graph Neural Network Inference:

[0223] The system constructs a fault prediction model based on graph convolutional networks (GCN):

[0224] Input data:

[0225] Node characteristics (x): Includes state parameters such as type, temperature, vibration value and humidity for each device.

[0226] Edge index: Describes the connection relationships between device nodes in the knowledge graph.

[0227] Core processing:

[0228] The model aggregates the feature information of each device node and its neighboring nodes through graph convolutional layer (conv1) operations.

[0229] This operation essentially merges the device's own state with the context information of its associated devices.

[0230] Output result:

[0231] Apply the sigmoid activation function to the processed features to convert the output values ​​into probabilities between 0 and 1.

[0232] The final output is a failure probability vector, where each element corresponds to the failure probability of a device node in the knowledge graph.

[0233] 6. Bayesian probability derivation:

[0234] For example, when the system detects excessively high water pump current (A), it considers this a possible symptom of a fault (B). Based on historical data and a probability model, the system analyzes the specific faults that this symptom may be caused by and their probability of occurrence:

[0235] Bearing wear (C): highest probability, 60%.

[0236] Voltage instability (D): Probability is 30%.

[0237] Impeller corrosion (E): Lowest probability, 10%.

[0238] Furthermore, if the diagnosis confirms that the cause of the fault is bearing wear (C), the system will directly provide specific repair recommendations: the bearing needs to be replaced (F).

[0239] Special note on Bayesian probability derivation: This example demonstrates a concrete reasoning process. In the complete workflow, this reasoning is performed using a locally updated Bayesian network model during the root cause localization phase, guided by assigned weights. Its output (P(bearing wear|excessive current) = 60%) is one of the core bases for ranking multiple candidate root causes.

[0240] Rule-based reasoning engines, graph neural network engines, and Bayesian network engines are specific execution instances of hybrid reasoning engines (rule-based reasoning / R, GNN / G, Bayesian / B).

[0241] In this embodiment, the data is also visualized:

[0242] The constructed knowledge graph is displayed graphically, allowing users to intuitively browse and query the nodes and relationships within it. Various visualization tools and technologies can be used, such as the built-in visualization features of graph databases like Neo4j, or professional visualization software like Echarts.

[0243] In this embodiment, anomaly detection, prediction, and alarm triggering specifically include the following operations:

[0244] 1. Dynamic threshold generation:

[0245] Real-time calculation of adaptive thresholds based on knowledge graphs:

[0246] Threshold = Historical Mean ± Safety Factor × Dynamic Standard Deviation × Load Correction Factor

[0247] For example, when the air conditioning load is >80%, the cooling water temperature threshold is tightened by 20%, and the calculation is highly dependent on the equipment specifications, historical statistics, operating condition information (from step 1), and real-time data (from IoT) provided by the knowledge graph.

[0248] 2. Physical constraint prediction model:

[0249] LSTM prediction layer: Input 23-dimensional features (sensor / environment / knowledge graph data);

[0250] Physical constraint layer: Forces the output to conform to physical laws such as Bernoulli's equation;

[0251] Results: Prediction bias <5%, RUL (Remaining Life) accuracy >92%;

[0252] The physical constraint layer enforces the use of integrated physical equations to constrain the output of the LSTM. Its predictions (such as RUL) are an important basis for alarm classification and maintenance recommendations.

[0253] 3. Three-level alarm linkage:

[0254] Level trigger condition response action

[0255] If Level I RUL < 24h or PUE deviation > 15%, initiate shutdown protocol + SMS alarm within 10 seconds;

[0256] If the probability of a Level II critical parameter exceeding the threshold is >70%, generate a preventative work order within 8 hours.

[0257] Level III related device anomaly probability >50% - AR visual risk warning;

[0258] Alarm triggering is the direct entry point for root cause reasoning. "When the system triggers a Level I or Level II alarm or a specific Level III alarm (such as an associated device anomaly probability > 50%), the knowledge graph-driven dynamic root cause reasoning process will be immediately initiated, and the abnormal device / parameter corresponding to the alarm will be used as the initial anomaly label input."

[0259] 4. Four-fold safety protection:

[0260] Conflicts are handled according to priority: physical equations > manually preset red lines > knowledge graph rules > real-time sensor data;

[0261] All automated decisions must pass four layers of verification; otherwise, execution will be suspended.

[0262] 5. Closed-loop update mechanism:

[0263] After each adjustment, the system records the deviation between actual energy consumption and predicted values, and updates the neural network model accordingly. Through backtesting of historical data, innovative strategies such as pre-cooling circulating water in low-temperature environments during winter were found to further improve energy efficiency.

[0264] Model fine-tuning: When the prediction bias is >10%, update the last two layers of the LSTM online;

[0265] Knowledge graph calibration:

[0266] False alarm → Reduce rule confidence (-0.2);

[0267] Effective strategy → Increase decision weight (+0.3);

[0268] Prediction bias and false alarm information are key inputs for model fine-tuning and knowledge graph calibration in step 10.

[0269] Knowledge graph-driven dynamic root cause reasoning includes the following operations:

[0270] 1. Real-time anomaly monitoring:

[0271] When sensors detect abnormal data (such as excessively high temperature in a certain area), the system automatically associates relevant devices (such as air conditioning units, air supply valves, and power loads) through a knowledge graph and triggers multi-level reasoning. For example, if the air conditioning unit is in normal condition but the air supply valve data is abnormal, the system will prioritize indicating the possibility of valve failure.

[0272] Abnormal temperature → Check the status of the air supply valve of the relevant air conditioning unit → Check if the associated power load exceeds the limit.

[0273] The results of the knowledge graph association query (the list of related devices) are the starting point for generating the candidate root cause set C.

[0274] 2. Optimization of probabilistic reasoning:

[0275] The system constructs and executes dynamic Bayesian network (DBN) inference based on a knowledge graph structure to quantify the probabilities of identified candidate root causes of failures. This process specifically includes the following operations:

[0276] 1) Generation of candidate root cause set: Based on the association query of knowledge graph (such as multi-level reasoning path), identify all potential equipment failure or abnormal status nodes that may cause the current observation anomaly (such as excessive temperature or water leakage), and form a candidate root cause set C = {Cause1, Cause2, ..., CauseN}.

[0277] The generation of the candidate root cause set C is based on knowledge graph topology and anomaly propagation analysis.

[0278] 2) Evidence Variable Mapping: Real-time sensor data (such as temperature, humidity, pressure, flow rate, and current values), equipment status information (such as valve opening and pump start / stop), environmental parameters (such as ambient temperature and humidity), and historical statistical indicators (such as the mean / variance of the past hour) related to the current anomaly, as well as the theoretical output values ​​of physical equation nodes (such as Bernoulli's equation ΔP), are mapped to observed evidence variables E = {Evidence1, Evidence2, ..., EvidenceM} in the Bayesian network. Each evidence variable corresponds to the attribute value of a specific node or edge in the knowledge graph.

[0279] 3) Network structure and parameter loading:

[0280] Load the predefined Bayesian network substructure most relevant to the current anomalous scenario (this substructure contains candidate root cause nodes C, relevant evidence nodes E, and the causal / conditional dependencies between them).

[0281] Load the Conditional Probability Table (CPT) for each node (derived from the build and incremental update mechanism). The initial values ​​of the CPT are derived from the Fault Tree Analysis (FTA) in the equipment manual, expert experience, and historical data statistics (see the "CPT Generation and Update" section).

[0282] For faults involving time evolution (such as gradual performance degradation), activate the Dynamic Bayesian Network (DBN) pattern and load the state transition probabilities between time slices.

[0283] 4) Probability calculation and inference: Utilizing a Bayesian inference engine (built on a Bayesian model specified by the assigned weights W_B) (e.g., integrating PyMC3, HUGIN, or pgmpy libraries):

[0284] The observed specific evidence value E = e is input into the network.

[0285] Perform probabilistic inference calculations (such as exact inference: connection tree algorithm; large-scale approximate inference: particle filter).

[0286] Calculate the posterior probability P(Cause_i | E=e) of each candidate root cause Cause_i given evidence e.

[0287] 5) Result Sorting and Output: All candidate root causes C are sorted in descending order based on the calculated posterior probability P(Cause_i | E=e). The system prioritizes and investigates high-probability root causes. For example, if P(pipe rupture | abnormally high humidity, decreased water pressure) = 0.85 and P(valve seal failure | abnormally high humidity, decreased water pressure) = 0.12, then pipe rupture is prioritized for investigation.

[0288] 3. Generation and updating of conditional probability tables (CPT):

[0289] 1) Initial CPT generation: By parsing the fault tree (FTA) in the equipment manual, combining the experience of domain experts to assign values, and analyzing historical fault record data to perform statistical estimation, the conditional probability table (CPT) of each node in the Bayesian network is initialized.

[0290] 2) Dynamic update mechanism: Establish a closed-loop feedback learning mechanism:

[0291] After each repair work order is completed, the system obtains the actual confirmed root cause of the fault (ActualCause).

[0292] For the Bayesian network substructures involved in this inference (especially the parent nodes, child nodes, and their CPTs directly related to the ActualCause node), the Dirichlet prior Bayesian update algorithm is applied to correct the local CPTs:

[0293] A Bayesian incremental update algorithm is used to revise the conditional probability table. Upon receiving maintenance feedback: the set of parent / child nodes (Markov blanket) of the faulty node is locked; the event count is updated based on the actual fault state; and the new probability value is calculated using the formula:

[0294] New probability = (historical data weight × old probability + actual observation count) / (historical data weight + total observation count), where actual observation count = 1 (fault confirmation) / 0 (false alarm);

[0295] This process only updates a local part of the network, avoiding global retraining.

[0296] The update process only applies to local network nodes related to this fault event, avoiding global retraining and ensuring efficiency.

[0297] 4. Applications of Dynamic Bayesian Networks (DBNs): (Combined with probabilistic reasoning)

[0298] For fault scenarios that require analysis of the evolution of the state over time (such as bearing wear and deterioration, or pipe scaling), the system enables DBN mode.

[0299] The continuous sensor monitoring data stream is divided into discrete time slices t, t+1, ... according to an adaptive time window (e.g., when the abnormal variance of 3 consecutive time slices is greater than the threshold, switch to the outbreak mode, with a fault outbreak period of 1 minute and a stable period of 15 minutes).

[0300] Within each time slice, perform the above probabilistic reasoning.

[0301] Simultaneously, time series prediction is performed using state transition probabilities (which describe the likelihood of equipment states such as [normal, worn, fault] transitioning between adjacent time slices). For example, if the state at time t is "worn," then the probability of transitioning to "fault" at time t+1 is 25% (based on the transition matrix).

[0302] Continuous early warning trigger: When the system predicts, based on DBN inference, that the probability of a critical device reaching a fault state within a finite number of future time steps exceeds a preset threshold (e.g., P(fault@t+3)>0.7), or observes abnormal states in multiple consecutive time slices (e.g., pressure fluctuations exceeding the threshold three times in a row), an early warning is triggered.

[0303] 5. Root cause localization:

[0304] Knowledge graph path analysis: Based on the physical connections and logical dependencies of the knowledge graph, graph algorithms (such as PageRank variants, weighted edge weights such as fault propagation probability, dynamic confidence) are used to calculate the "influence" or "centrality" score of candidate root cause nodes.

[0305] Knowledge graph path analysis (such as PageRank) directly affects the topology and relation weights of the constructed knowledge graph.

[0306] Time series data analysis (DBN output): Provides the evolution probability of candidate root causes over time.

[0307] Bayesian posterior probability: Provides the immediate likelihood of a candidate root cause given current and historical evidence. When calculating the posterior probability, the Bayesian network uses dynamic weight coefficients W_B to weight and correct the basic prior probability value of the cause of the failure, i.e.: Corrected probability = Basic probability × Weight coefficient W_B.

[0308] The system integrates the above information (such as a weighted average of three indicators with different weights: evidence-based Bayesian posterior probability α; PageRank influence score β based on the knowledge graph structure; and risk trend score γ based on time-series data analysis) to finally rank and locate candidate root causes. During the inference process, a path pruning strategy is applied to automatically exclude device nodes marked as "recent maintenance normal" or "current status confirmed normal" in the knowledge graph, narrowing the search scope and improving efficiency.

[0309] The steps for generating repair suggestions (determining repair solutions) in this embodiment are as follows:

[0310] 1. Precise strategy matching:

[0311] Extract historical repair solutions from the knowledge graph (e.g., replacing a bearing takes 30 minutes).

[0312] By combining fault propagation path analysis (PageRank algorithm), a contingency plan with spatiotemporal constraints is generated:

[0313] Example: Close the associated valve → drain the pipeline → disassemble the pump body (avoid adjacent cables).

[0314] 2. AR Augmented Reality Guide:

[0315] When maintenance personnel scan the device using AR glasses:

[0316] Automatically overlay risk warnings (such as "High-pressure pipeline, depressurize first!");

[0317] Visual operation guides (3D animation showing the disassembly sequence) can reduce the rate of human error by 92%.

[0318] The information overlaid in AR guidance (risk warnings, operation instructions) comes from device safety rules, operating procedures, and identified real-time risks stored in the knowledge graph.

[0319] 3. Multi-objective optimization ranking:

[0320] A priority list is generated based on cost, time consumption, and scope of impact, and the result is the final executable solution directly output to the operations and maintenance personnel.

[0321] 4. Dynamic risk assessment:

[0322] Automatically mark operation red lines (e.g., "Do not disassemble when valve front pressure > 0.5MPa");

[0323] Real-time monitoring of associated sensors; alarms are sent immediately when limits are exceeded.

[0324] In this embodiment, the closed-loop verification and knowledge graph self-optimization mechanism (adjusting the weights of the knowledge graph and the conditional probability table of the Bayesian network based on the maintenance results, and calibrating the parameters of the physical equations) includes the following operations:

[0325] 1. Multi-source feedback intelligent data acquisition:

[0326] The automatic synchronization work order system obtains the actual root cause (e.g., "impeller corrosion").

[0327] Analysis of maintenance notes (BERT model extracts key semantics: "Non-bearing wear, water hardness needs to be tested").

[0328] Actual root causes and repair semantics are the gold standard for verifying the accuracy of diagnostic results and are also key to calculating reinforcement learning rewards.

[0329] 2. Localized model update:

[0330] Knowledge Graph:

[0331] Lock the faulty device's 3-layer neighborhood subgraph;

[0332] Reinforcement learning dynamically adjusts the node relationship weights (+0.1 for successful cases, -0.2 for misjudgments).

[0333] The optimized subgraph range (3-layer neighborhood) identifies the root cause nodes of the fault based on location. Reinforcement learning actions directly modify the topology and relation weights of the knowledge graph, affecting subsequent hybrid reasoning weight calculations, reasoning, association analysis, and root cause search.

[0334] Bayesian networks:

[0335] Update only the parent / child nodes of the faulty node (Markov blanket)

[0336] Dirichlet Prior Bayesian Update Algorithm Corrects CPT:

[0337] Example: P (pump malfunction | abnormal flow) changes from 60% to 75%.

[0338] The updated node range (parent / child nodes (Markov blanket)) also locks the located root cause node of the failure. The Dirichlet-updated CPT is directly used for subsequent Bayesian inference.

[0339] GNN model:

[0340] Event-triggered fine-tuning: When the maintenance deviation is >20%, only the parameters of the fault association subgraph are optimized (learning rate 0.001).

[0341] Periodic incremental training: The entire model is lightly updated every 100 maintenance cases (by calculating the parameter importance score (using the EWC algorithm), only the 10% of parameters with the lowest score are updated).

[0342] When the root cause reasoning result does not match the actual maintenance, a reinforcement learning action is triggered:

[0343] If the actual fault does not belong to the Top3 candidate set, then add a fault-related edge ().

[0344] The target subgraph for event-triggered fine-tuning is determined by the fault propagation path or root cause neighborhood. Periodic incremental training ensures the overall adaptability of the model. The updated GNN model is invoked during hybrid inference.

[0345] Physical equations:

[0346] Genetic algorithm calibration of equipment binding parameters (such as pipeline friction factor f, triggered when deviation > 5%).

[0347] Parameter calibration targets parameters or associated device attributes defined in the equation node `params_binding`. The calibration results are used to improve the accuracy of subsequent physical verification and physical constraint prediction.

[0348] 3. Reinforcement Learning Decision Engine:

[0349] Status: <Fault type, ambient humidity, subgraph structure>;

[0350] Actions: Add faulty edge / Delete redundant node / Adjust weight;

[0351] Reinforcement learning integrates fault type, environmental parameters, and subgraph structure into its state information. Actions directly modify core system components. Reward calculation is highly dependent on repair results (execution feedback) and system performance metrics such as response time.

[0352] 4. Overload protection mechanism:

[0353] The historical experience pool retains the most recent 200 records (to prevent overfitting).

[0354] High-frequency fault models are fine-tuned in real time, while low-frequency equipment is updated in batches quarterly.

[0355] These strategies ensure the stability and feasibility of optimization, avoid system performance degradation due to over-optimization or outdated data, and guarantee continuous and reliable operation.

Claims

1. A method for rapid fault attribution and self-optimization of construction equipment, characterized in that, Includes the following steps: Collect and construct a knowledge graph corresponding to the building equipment based on the structured data of the building equipment: The system maps instance parameters of building equipment to the energy consumption rule engine of the knowledge graph to obtain energy consumption rule nodes; converts material properties into physical calculation factors in the knowledge graph to obtain physical calculation attribute nodes; constructs a three-dimensional spatial relationship network based on the spatial coordinates of the equipment; constructs a logical dependency tree based on the connection relationship between the equipment; uses component-level BIM parameters as fault propagation path nodes in the knowledge graph; generates corresponding SWRL inference rules for BIM annotations and fault codes; and imports the FTA structure from the equipment manual into the knowledge graph to obtain OWL rules and Bayesian network initial condition probability tables in the knowledge graph. Building a failure attribution engine: Based on the topology, logical dependency tree, and OWL rules of the knowledge graph corresponding to building equipment, a Bayesian network is constructed to form a Bayesian network engine; a meta-model is constructed based on SWRL inference rules and OWL rules to form a rule inference engine; and a graph neural network is constructed based on the nodes and edges of the knowledge graph to form a graph neural network engine. Establish and deeply integrate the system dynamics model: The physical equations are encapsulated into executable physical equation nodes in the knowledge graph; a two-layer triggering mechanism is established based on real-time data and physical equation nodes; and physical equation nodes and device nodes are deeply integrated to obtain dynamic binding relationships in the knowledge graph. The two-layer triggering mechanism is used to obtain real-time physical contradiction markers through physical verification. Perform multi-source data extraction: Unstructured knowledge related to building equipment is extracted from different data sources, and entities, attributes and relationships are extracted from the unstructured knowledge and populated into the knowledge graph corresponding to the building equipment. Perform real-time anomaly detection: Based on physical contradiction markers and real-time data calculation of the adaptive threshold of the working condition, it is determined whether there is data that reaches the corresponding adaptive threshold. If so, it is marked as an abnormal label and the fault attribution engine is called to output the probability of fault occurrence. Otherwise, no action will be taken; Perform dynamic adjustment of hybrid weights: real-time weights based on physical contradiction marking and real-time data acquisition rule inference engine, Bayesian network engine and graph neural network engine; The failure probability outputs from the rule-based reasoning engine, the Bayesian network engine, and the graph neural network engine are weighted and fused according to real-time weights to generate a comprehensive failure probability and obtain the corresponding candidate root causes. Repair schemes are determined based on candidate root causes; the weights of the knowledge graph and the conditional probability table of the Bayesian network are adjusted based on the repair results, and the parameters of the physical equations are calibrated.

2. The method of claim 1, wherein, Structured data for building equipment includes: Building types include residential, commercial, and public buildings; Building materials, including steel, concrete, and glass; Building components, including beams, columns, and walls; Building structure, including floors, rooms, areas, and pipe / line topology; Equipment systems, including chillers, cooling towers, air conditioning, elevators, fire protection, lighting, and water supply and drainage; Equipment components, including compressors, valves, sensors, and controllers; Failure modes, including common failure types, symptoms, scope of impact, and repair solutions.

3. The method of claim 1, wherein, Specific methods for mapping instance parameters of building equipment to the energy consumption rule engine of the knowledge graph include: In the BIM modeling software, a text type parameter named "BuildingType" is created, and a value is assigned to this parameter through a programmatic interface. This parameter is then dynamically mapped to the energy consumption rule engine of the knowledge graph. Specific methods for constructing a logical dependency tree based on the connection relationships between devices include: Extract all physical connection points of the target device from the BIM model, traverse the connection points to determine all devices connected to the target device, automatically create a "CONNECTS_TO" relationship between the target device and each connected device in the knowledge graph, and then establish control logic in the knowledge graph to obtain a logical dependency tree; Specific methods for generating corresponding SWRL inference rules for BIM annotations and fault codes include: In Revit software, a fault annotation object is created for a specified device. Based on this fault annotation object, the corresponding SWRL inference rules are automatically generated in the knowledge graph. The fault annotation object contains the fault code and symptom description. The specified device is identified by elementId. The specific methods for importing the FTA structure from the equipment manual into the knowledge graph to obtain the OWL rules and Bayesian network conditional probability tables in the knowledge graph include: A semantic parsing engine is used to deconstruct the fault tree: the root node is mapped to a fault class entity in the knowledge graph, the logic gates are transformed into OWL axioms, the leaf nodes are associated with device nodes, and probabilistic attributes are injected. Dynamic scenarios are handled through a hybrid probability-rule engine: SWRL hard rules are used to constrain static dependencies, and dynamic probabilities are dynamically updated through a Bayesian network conditional probability table.

4. The method of claim 1, wherein, Node types in Bayesian networks include device entities, sensors, fault modes, physical equation entities, location spaces, device parameters, fault types, and environmental variables; The relationship types of nodes include physical connection hasPhysicalConnectionTo, logical dependency dependsOn, fault association causes / symptomOf, and equation association governedBy; Node attributes include device specification parameters (specs), real-time sensor values ​​(currentValue), historical statistics (stats), equation parameters (params), dynamic weights (dynamicWeight), and confidence. Edge relationships include physical causality, logical dependency, and statistical association; Edge connection rules include constructing directed edges based on physical dependencies and establishing directed edges based on historical fault data; The relation confidence in Bayesian networks employs the following dynamic adjustment mechanism: Positive adjustment: The increase in weight is proportional to the repair success rate triggered by this relationship, where the repair success rate is the number of successful repairs divided by the total number of triggers. The adjustment strength is controlled by the coefficient α. Negative adjustment: The reduction in weight is proportional to the environmental change coefficient characterizing changes in external conditions, and the degree of influence is controlled by the coefficient β; Updated formula: The new confidence weight equals the old weight plus the increment based on repair success rate, minus the reduction based on environmental changes.

5. The method of claim 1, wherein, Specific methods for encapsulating physical equations into executable physical equation nodes in a knowledge graph include: An `EquationEntity` entity type is created to store the metadata and executable code of the physical equations, and this `EquationEntity` entity type is used as an executable physical equation node in the knowledge graph; the attributes of the `EquationEntity` entity type include: The attribute name is equation_id, the data type is string, and the attribute description is a unique identifier for the physical equation; The attribute name is lambda_func, the data type is a code object, and the attribute description is an executable equation; The attribute name is params_binding, the data type is JSON, and the attribute description is a dynamic parameter mapping rule; The attribute name is threshold, the data type is floating point, and the attribute description is the theoretical value / actual value deviation alarm threshold. lambda_func and params_binding are used to invoke the physics equations; The method for obtaining dynamic binding relationships in the knowledge graph is to map physical equation parameters to device sensors through the params_binding attribute.

6. The method of claim 1, wherein, The two-layer triggering mechanism includes a basic triggering layer and an arbitration triggering layer; When the sensor readings exceed the device’s current adaptive dynamic range for three consecutive times, the basic triggering layer is invoked to verify the physical equations. If the physical equation verification performed by calling the basic trigger layer passes, the hybrid weights are dynamically adjusted to improve the adaptability of the static rules; if the physical equation verification result performed by calling the basic trigger layer is a transient anomaly, it is determined that there is a device start-up / shutdown disturbance; if the physical equation verification result performed by calling the basic trigger layer is a persistent deviation, it is determined that there is device aging; if the physical equation verification result performed by calling the basic trigger layer is a system conflict, it is determined that there is model parameter inaccuracy. The adaptive dynamic range is the current operating condition baseline value ± twice the standard deviation of historical data; the current operating condition baseline value is calculated through a sliding window based on the real-time operating status of the equipment. When multiple candidate faults output by the Bayesian network simultaneously satisfy condition one and condition two, the arbitration trigger layer is invoked to verify the physical equations. Condition 1: Insufficient confidence of the highest probability fault: Among all candidate faults, the highest probability fault has a confidence value of less than 0.7; Condition 2: Other significantly probable faults exist: the probability of at least one candidate fault is greater than 0.3; The process of calling the arbitration trigger layer to verify the physical equations involves calling the relevant physical equations according to a preset priority to verify the physical rationality of the candidate faults. If the physical equation verification performed by calling the arbitration trigger layer passes, the mixed weights are dynamically adjusted to improve the adaptability of the static rules. Otherwise, perform the following operations: Reject all candidate faults that are verified to violate the laws of physics; The system indicates a physical contradiction and invokes the fault attribution engine to output the probability of the fault occurring. When the persistent deviation exceeds 24 hours, calibrate the physical equation parameters.

7. The method of claim 1, wherein, The specific methods for extracting unstructured knowledge related to building equipment from different data sources, and extracting entities, attributes, and relationships from unstructured knowledge, are as follows: Entities, attributes, and relationships are extracted from academic papers, technical reports, standards, architectural drawings, photographs, environmental monitoring data, and / or structural health monitoring data using natural language processing, image recognition, and / or data analysis techniques. For knowledge that is duplicated, contradictory, or incomplete in different data sources, knowledge fusion is carried out, duplicate knowledge is removed, knowledge contradictions are eliminated, and incomplete knowledge is supplemented to form a unified knowledge system. When data extracted from unstructured knowledge contains a conflict between the location of the same device in the BIM model and its actual sensor positioning, position calibration is performed. The expression for this is: New location = 30% * BIM location + 70% * sensor location; When there are conflicts in physical parameters in the data extracted from unstructured knowledge, the theoretical values ​​calculated by the physical equations are used as the benchmark.

8. The method of claim 1, wherein, The formula for calculating the adaptive threshold for the operating condition is: Threshold = Historical mean ± Safety factor * Dynamic standard deviation * Load correction factor.

9. The method of claim 1, wherein, The expression for calculating real-time weights is: W i = softmax(α*a i +β*b i +γ*c i ) Among them W i α represents the real-time weight of the i-th engine; i = 1, 2, 3, which represent the rule-based reasoning engine, the Bayesian network engine, or the graph neural network engine, respectively; α, β, and γ are all weights. a i is the confidence for the i-th engine; b i is the scenario relevance for the i-th engine; c i is the historical accuracy rate for the i-th engine.

10. The method of claim 1, wherein, It also includes three-level alarm linkage: The trigger conditions for a Level 1 alarm are: the remaining lifespan of the equipment is less than 24 hours or the energy utilization efficiency deviation is greater than 15%; The response action for a Level 1 alarm is: activate the shutdown protocol and send an SMS alarm within 10 seconds; The trigger condition for a Level 2 alarm is: the probability of a key parameter exceeding the threshold is greater than 70%. The response action for a Level 2 alarm is to generate a preventative work order within 8 hours. The trigger condition for a Level 3 alarm is: the probability of an abnormality of the associated device is greater than 50%. The response action for a Level 3 alarm is: AR visual risk alert.

Images