A multi-modal dynamic knowledge graph construction method and intelligent operation and maintenance system for mine crushing equipment

Abstract

The application discloses a kind of multi-modal data knowledge graph construction method and system for mine crushing equipment, for the scene characteristics of mine crushing equipment heavy load, high dust, fault conduction fast, through "working condition adaptation space-time alignment + crushing scene special fuzzy logic credibility evaluation" to realize multi-modal data high-quality fusion, solve the data distortion problem caused by data island, space-time misplacement and working condition interference;Through "crushing equipment special event triggering- typical fault rule reasoning-graph real-time updating", realize the dynamic evolution of knowledge graph, solve the problem that traditional graph is not matched with the real-time response demand of crushing equipment fault due to staticization. The application improves the accuracy of data fusion and the timeliness of knowledge, the fault positioning accuracy of core components reaches 95%, the fault prediction accuracy is improved by more than 28%, provides reliable support for intelligent operation and maintenance of mine crushing equipment, reduces operation and maintenance cost, improves production efficiency, has significant practical value.

Description

Technical Field

[0001] This invention belongs to the field of industrial big data and knowledge engineering technology, specifically involving a method for constructing a multimodal dynamic knowledge graph for mining crushing and grinding equipment and an intelligent operation and maintenance system. Background Technology

[0002] Mining grinding equipment (including key equipment such as mineral mills, high-pressure roller mills, and crushers, with core components including mill liners, crusher eccentric shafts, high-pressure roller mill roller surfaces, and bearing assemblies) is a core link in the mining production process. Its operational stability directly determines the mining production efficiency, safety level, and operating costs, which also places high demands on operation and maintenance efficiency.

[0003] During equipment operation, massive amounts of multimodal data are generated, including high-frequency vibration signals collected by vibration sensors (reflecting component impact and wear), text reports generated by oil analysis equipment (reflecting lubrication and component corrosion), numerical data recorded by temperature / pressure sensors (characterizing equipment thermal load and operating load), equipment process parameters (such as feed rate and grinding media filling rate), and maintenance records entered by maintenance personnel (containing information on the entire fault handling process). However, currently, there are two core problems in the data management and knowledge application of mining crushing and grinding equipment.

[0004] 1. Multimodal data silos and poor fusion quality Various types of data are scattered across different systems (e.g., mill sensor data is stored in the SCADA system, and maintenance records are stored in the operation and maintenance management system), forming "data silos." Furthermore, the data differs significantly in type (mill operation data / text), acquisition frequency (50-10kHz vibration signal / weekly oil report), and spatiotemporal dimensions. Moreover, the data is easily distorted due to interference from mine dust and equipment vibration impacts. Traditional fusion methods often employ simple splicing, failing to consider the impact of grinding equipment operating conditions on data reliability, and neglecting to prioritize core component data. This results in spatiotemporal misalignment and inconsistent quality of the fused data, failing to provide a reliable data foundation for accurate fault diagnosis of grinding equipment.

[0005] 2. Knowledge graphs lack staticity and timeliness. Traditional knowledge graphs rely on manual input and updates, with update cycles lasting weeks or even months. This makes it difficult to capture real-time changes in the operating status of grinding equipment (such as sudden liner detachment failures) and new failure modes (such as asymmetric roller surface wear failures in high-pressure roller mills), resulting in significant knowledge lag. Furthermore, grinding equipment failures develop rapidly and have complex causes, demanding extremely high real-time and accurate maintenance responses. Static knowledge graphs cannot meet the needs of intelligent maintenance for real-time knowledge support and continuously improving diagnostic accuracy, easily leading to false faults and unplanned downtime.

[0006] Therefore, there is an urgent need for a technical solution that can adapt to the harsh working conditions of grinding equipment, achieve efficient fusion of multimodal data, and dynamically update the knowledge graph, so as to provide high-precision data support and real-time knowledge services for equipment health management, fault diagnosis and predictive maintenance, and solve the above-mentioned industry pain points. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for constructing a multimodal knowledge graph for mining grinding equipment, addressing the technical challenges in knowledge graph construction of existing mining grinding equipment. This method solves the problems of "low data fusion quality" and "static knowledge graph", while also adapting to the characteristics of grinding equipment with complex structure, harsh operating environment, heavy load, high dust, and large fluctuations in operating conditions.

[0008] The technical solution of this invention is: A method for constructing a multimodal knowledge graph for mining crushing and grinding equipment includes the following steps: S1. Multimodal data acquisition and input: Acquire multi-dimensional monitoring data of mining crushing and grinding equipment, including high-frequency vibration signals, oil analysis reports, temperature and pressure, feed data and operation and maintenance records; S2. High-quality multimodal data fusion: including spatiotemporal alignment processing, credibility assessment, and priority-weighted fusion to generate high-quality multimodal fused datasets; S3. Knowledge representation of high-quality multimodal fusion datasets: including entity recognition, relation extraction, and structured knowledge construction, among which, Entity recognition: Reverse engineering named entity recognition for grinding equipment using the high-quality fusion dataset, extracting physical entities, indicator entities, fault entities, and maintenance / operating condition entities; Relation extraction: Extract relations from the extracted entities and establish compositional relations, monitoring relations, representational relations, and causal / disposal relations between entities; Structured knowledge creation: Encapsulate entities and relations into triples (entity-relation-entity) and assign confidence attributes to triples to form structured knowledge; S4. Construction of Dynamic Knowledge Graph: This includes event monitoring, graph storage, fault-oriented knowledge reasoning, dynamic graph updating and tracing, and graph output and application. Event monitoring: Based on the structured knowledge, equipment event monitoring is performed to identify abnormal events and maintenance events; Knowledge graph storage: The triples and the monitored events are stored as a knowledge graph; Fault-oriented knowledge reasoning: Based on the built-in fault rule library for grinding equipment, abnormal events are associated and matched with fault types to deduce the correspondence between equipment abnormalities and faults such as component wear, loosening, and blockage. Knowledge graph dynamic update and traceability: Perform update operations such as adding, deleting, and modifying on the knowledge graph, record the update time, interface, specific content, and update effect, and form a traceable update log; Knowledge Graph Output and Application: Outputs a dynamic knowledge graph that is updated in real time, enabling real-time diagnosis and rapid response to faults in mining crushing and grinding equipment.

[0009] Specifically, the spatiotemporal alignment processing includes millisecond-level UTC timestamp calibration of the collected multimodal data, segmented matching based on the spatial dimensions of equipment, components, and monitoring points, and phased sampling frequency adjustment according to the operating characteristics of the grinding equipment; the credibility assessment includes constructing a fuzzy logic model specific to the grinding scenario, introducing dust concentration and operating condition interference factors to calculate data credibility, and outputting dynamic credibility weights in the 0-1 range; the priority weighted fusion includes adopting a strategy of weighted summation of numerical data, weighted fusion of textual data, and keyword confidence calculation to give weight enhancement to core monitoring components, complete multimodal data fusion, and generate a high-quality fused dataset.

[0010] Specifically, the input parameters of the fuzzy logic model include the dust concentration and operating condition fluctuation range of the mining crushing and grinding equipment operating environment, and the output credibility weight is used to reduce the weight or filter low-quality data.

[0011] Specifically, in the aforementioned priority-weighted fusion, a 20% weight increase is assigned to the core monitoring components based on sensitivity analysis of historical fault data.

[0012] The composition relationship is used to characterize the inclusion / subordination relationship between entities, the monitoring relationship is used to characterize the association between sensors and indicator entities, the characterization relationship is used to characterize the association between indicator entities and equipment health status, and the causal / disposal relationship is used to characterize the triggering and disposal logic between faulty entities and abnormal events and maintenance events.

[0013] Specifically, the confidence attribute is used to measure the reliability of triple knowledge and is calculated by comprehensively considering the data source, the accuracy of the extraction algorithm, and historical verification results.

[0014] Specifically, the abnormal equipment events include abnormalities in vibration, temperature, current, and feed rate parameters.

[0015] Specifically, the dedicated fault rule library for the crushing and grinding equipment includes: high-frequency vibration amplitude ≥ threshold corresponds to bearing wear fault, abnormal temperature ≥ threshold corresponds to bearing wear fault, uneven pressure corresponds to liner wear fault, and abnormal feed rate corresponds to feed channel blockage fault. The rule library supports iterative updates based on actual operation and maintenance data.

[0016] Specifically, the update log records include the update operation type, trigger event number, and graph node / relationship information before and after the update, which are used for subsequent fault tracing and model optimization, as well as periodically adjusting the rule confidence in the fault rule base to achieve adaptive optimization of the rule base.

[0017] This invention also provides an intelligent operation and maintenance system for mining crushing and grinding equipment, comprising: The data acquisition module is used to acquire high-frequency vibration signals, oil analysis reports, temperature and pressure data, and operation and maintenance records of the mining crushing and grinding equipment. The data fusion module is used to perform high-quality multimodal data fusion steps and generate a high-quality fused dataset; The knowledge representation module for multimodal fusion data is used to represent multimodal fusion data in a knowledge-based manner to form structured knowledge. The knowledge graph construction module is used to build dynamic knowledge graphs and output dynamically updated knowledge graphs in real time. The fault diagnosis module, based on a dynamic knowledge graph, enables fault identification, location, and early warning for mining grinding equipment. The update and traceability module is used to perform add, delete, and modify operations on the knowledge graph and record update logs containing working condition information.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. To address the extremely rapid transmission of faults in grinding equipment (such as liner detachment occurring within seconds) and the excessively long traditional event triggering cycle (minutes), this invention employs millisecond-level UTC calibration combined with real-time edge-side inference. 2. To address the large fluctuations in the operating conditions of grinding equipment (frequent changes in feed rate and filling rate), which cause fixed rules to fail, this invention introduces operating condition labeling and dynamic adjustment of rule confidence. 3. The newly added "traceable update log + operating condition query" is not merely a simple record, but rather serves subsequent rule base iterations and model optimization, forming a closed loop.

[0019] 4. Targeted Improvement in Data Fusion Quality: Compared to solutions that do not employ condition-adaptive fusion and static graphs, this invention addresses the spatiotemporal misalignment of data in grinding equipment through condition-adaptive spatiotemporal alignment. It introduces fuzzy logic evaluation of condition interference factors and component priorities to ensure scientific fusion weights. The consistency between fused data and the actual equipment status is improved by over 35%, the misdiagnosis rate of typical faults in grinding equipment is reduced by 40%, and the fault location accuracy of core components reaches 95%. 5. Enhanced Timeliness and Adaptability of Knowledge Graph: Event-triggered dynamic updates shorten the update cycle from weeks to minutes. New fault modes in grinding equipment are included in 100% of cases on time, fault prediction accuracy is improved by over 28%, and unplanned downtime is reduced by 22%, effectively addressing the rapid transmission of faults in grinding equipment. Attached Figure Description

[0020] Figure 1 Multimodal data fusion flowchart; Figure 2 Flowchart of spatiotemporal alignment processing; Figure 3 : Structure diagram of fuzzy logic credibility assessment model; Figure 4 Flowchart for knowledge representation of multimodal data; Figure 5 Flowchart for constructing a dynamic knowledge graph; Figure 6 Fault-oriented knowledge reasoning flowchart for grinding equipment. Detailed Implementation

[0021] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Example 1

[0022] This embodiment provides a method for constructing a multimodal knowledge graph for mining crushing and grinding equipment, including the following steps: S1. Multimodal data acquisition and input: Acquire multi-dimensional monitoring data of mining crushing and grinding equipment, including high-frequency vibration signals, oil analysis reports, temperature and pressure data, and operation and maintenance records; S2. High-quality multimodal data fusion: This includes spatiotemporal alignment, reliability assessment, and priority-weighted fusion to generate high-quality multimodal fused datasets, such as... Figure 1 As shown; S3. Knowledge representation of high-quality multimodal fusion datasets: including entity recognition, relation extraction, and structured knowledge construction, among which, Entity recognition: The high-quality fusion dataset is used for named entity recognition, extraction of physical entities, indicator entities, fault entities, and maintenance / operating condition entities for grinding equipment; Relation extraction: Extract relations from the extracted entities and establish compositional relations, monitoring relations, representational relations, and causal / disposal relations between entities; Structured knowledge creation: Encapsulate entities and relations into triples (entity-relation-entity) and assign confidence attributes to triples to form structured knowledge; S4. Construction of Dynamic Knowledge Graph: This includes event monitoring, graph storage, fault-oriented knowledge reasoning, dynamic graph updating and tracing, and graph output and application. Event monitoring: Based on the structured knowledge, equipment event monitoring is performed to identify abnormal events and maintenance events; Knowledge graph storage: The triples and the monitored events are stored as a knowledge graph; Fault-oriented knowledge reasoning: Based on the built-in fault rule library for grinding equipment, abnormal events are associated and matched with fault types to deduce the correspondence between equipment abnormalities and faults such as component wear, loosening, and blockage. Knowledge graph dynamic update and traceability: Perform update operations such as adding, deleting, and modifying on the knowledge graph, record the update time, interface, specific content, and update effect, and form a traceable update log; Knowledge Graph Output and Application: Outputs a dynamic knowledge graph that is updated in real time, enabling real-time diagnosis and rapid response to faults in mining crushing and grinding equipment. Example 2

[0023] This embodiment provides specific operations for spatiotemporal alignment processing. This spatiotemporal alignment processing, tailored to the operational characteristics of core components in grinding equipment, unifies the spatiotemporal stamp framework and employs a phased sampling frequency adjustment strategy to eliminate data spatiotemporal misalignment. The process is as follows: Figure 2 As shown, this includes performing millisecond-level UTC timestamp calibration on the collected multimodal data, segmenting and matching based on the spatial dimensions of equipment, components, and monitoring points, and adjusting the sampling frequency in stages according to the operating characteristics of the grinding equipment.

[0024] Timestamp design: UTC time format is adopted, accurate to the millisecond level (e.g., 2024-05-20T10:30:00.123Z), to unify the time reference standard of all data and avoid time misalignment caused by time zone and precision differences; in view of the large data fluctuations in the start-up phase (0-30 minutes) of the grinding equipment, an additional start-up flag is added to facilitate differentiated processing during subsequent fusion.

[0025] Spatial stamp design: Based on "equipment type-equipment number-core component number-monitoring point number", a unique spatial identifier is established (such as "MM-001-LB-001-01" represents the No. 1 monitoring point of the liner LB-001 of the mineral mill MM-001), to ensure that the data accurately corresponds to the monitoring points of the core components of the crushing and grinding equipment, and to adapt to the needs of multi-component collaborative fault diagnosis.

[0026] Phased sampling frequency adjustment: Differentiated frequency adjustment strategies are adopted for different operating stages (start-up / stable / shutdown) and data types of the grinding equipment. The frequency is uniformly set at 0.5Hz during stable operation; increased to 2Hz during the start-up / shutdown stage (when data fluctuations are large); mean downsampling is used for 10000Hz high-frequency vibration signals (bearing and roller surface monitoring); and cubic spline interpolation is used for daily oil reports and weekly process parameter reports to ensure that the data frequency matches the equipment's operating status.

[0027] The credibility assessment includes constructing a fuzzy logic model specific to the grinding and crushing scenario. This model is built by combining the operating characteristics of the grinding and crushing equipment, incorporating dust concentration and operating condition interference factors to calculate data credibility, scientifically evaluating the credibility of each data source, and outputting credibility weights in the 0-1 range. The model structure is as follows: Figure 3 As shown.

[0028] Among them, the input variable optimization is as follows: in addition to signal-to-noise ratio (numerical data), text integrity (text data), and data acquisition equipment stability (historical faults / calibration cycles), a new operating condition interference factor (dust concentration > 50 mg / m³) is added. 3 "High interference", 30-50 mg / m³ 3 "Medium interference", <30mg / m³ 3 (For "low interference"), adapted to the impact of high dust environment in mines on data acquisition.

[0029] Fuzzy processing: Optimize the membership function based on the data characteristics of crushing and grinding equipment. Vibration signal signal-to-noise ratio > 45dB (adapting to effective signal identification under heavy-load impact) is "high" (membership 1), 35-45dB is "high / medium", and < 35dB is "low"; text completeness must include key information specific to crushing and grinding equipment (faulty parts, monitoring points, feed rate and operating conditions, and handling measures), 4 or more items are "high", 2-3 items are "medium", and 1 or less items are "low"; the operating condition interference factor "low interference" corresponds to a 0.1 increase in credibility weight, and "high interference" corresponds to a 0.1 decrease in credibility weight.

[0030] Fuzzy reasoning: Based on the dedicated rule base of grinding equipment (such as "high signal-to-noise ratio + high equipment stability + low operating condition interference → high credibility", "liner monitoring data + text containing the keyword 'leakage powder' → credibility weight increased by 0.05"), the corresponding rules are activated by the Mamdani reasoning method to synthesize a fuzzy set of credibility.

[0031] Defuzzification: The centroid of the fuzzy set is calculated using the centroid method to obtain the precise credibility weight (e.g., 0.92, 0.78). The lower limit of the credibility weight of the core component (liner, roller surface, bearing) data is set to 0.6 to ensure the priority of key data.

[0032] Priority-weighted fusion includes strategies that combine weighted summation of numerical data and weighted fusion of textual data with keyword confidence calculation. This approach assigns increased weights to core monitoring components, completes multimodal data fusion, generates a high-quality fused dataset, and employs differentiated fusion strategies based on the data type and component importance of the grinding equipment. The fusion weight of core component data is increased by 20%.

[0033] Numerical data (vibration / temperature / pressure): Calculated using the formula "fusion value = Σ (data source value × credibility weight × component priority coefficient)", with the priority coefficient for core components set to 1.2 and non-core components set to 1.0. Example: Mill bearing temperature sensor data (85℃, weight 0.9, priority coefficient 1.2) + infrared detection data (83℃, weight 0.7, priority coefficient 1.2), fusion value = (0.9 × 85 + 0.7 × 83) × 1.2 = 161.52℃ (this value is a weighted median value; after correction with a mill temperature calibration coefficient of 0.95, it becomes 153.44℃).

[0034] Text-based data (oil reports / maintenance records): A weighted bag-of-words model is used, with a new keyword library for grinding equipment faults (loose liner, roller wear, bearing seizure, feed overload, etc.). The weighted frequency of keywords is doubled. Example: In the oil report (weight 0.8), the keyword "excessive metal particles" has a frequency of 5, while in the maintenance record (weight 0.6), the frequency of the same keyword is 3. After merging, the weighted frequency = (0.8 × 5 + 0.6 × 3) × 2 = 11.6. Example 3

[0035] This embodiment provides a reliability assessment. Combining the operating characteristics of grinding and crushing equipment, a fuzzy logic model is constructed, introducing an operating condition interference factor to scientifically evaluate the reliability of each data source. The model outputs a reliability weight in the 0-1 range. The model structure is as follows: Figure 3 As shown, the entire process from multi-source data input to final credibility weight output is fully demonstrated.

[0036] The model quantifies and evaluates the features of the original data from four dimensions: Signal-to-noise ratio: Vibration signal quality assessment, graded as >45dB (high), 35-45dB (medium), and <35dB (low); Text integrity: Oil fluid reports / maintenance records are classified into three levels: all 4 key information items (high), 2-3 items (medium), and 1 item (low); Equipment stability: Collect equipment status data and assess it based on historical fault records and calibration cycles; The input variables were optimized: in addition to signal-to-noise ratio (numerical data), text integrity (text data), and data acquisition equipment stability (historical faults / calibration cycles), a new operating condition interference factor (dust concentration > 50 mg / m³) was added. 3 "High interference", 30-50 mg / m³ 3 "Medium interference", <30mg / m³ 3 (For "low interference"), adapted to the impact of high dust environment in mines on data acquisition.

[0037] Fuzzy processing: Optimize the membership function based on the data characteristics of crushing and grinding equipment. Vibration signal signal-to-noise ratio > 45dB (adapting to effective signal identification under heavy-load impact) is "high" (membership 1), 35-45dB is "high / medium", and < 35dB is "low"; text completeness must include key information specific to crushing and grinding equipment (faulty parts, monitoring points, feed rate and operating conditions, and handling measures), 4 or more items are "high", 2-3 items are "medium", and 1 or less items are "low"; the operating condition interference factor "low interference" corresponds to a 0.1 increase in credibility weight, and "high interference" corresponds to a 0.1 decrease in credibility weight.

[0038] Fuzzy reasoning: Based on the dedicated rule base of grinding equipment (such as "high signal-to-noise ratio + high equipment stability + low operating condition interference → high credibility", "liner monitoring data + text containing the keyword 'leakage powder' → credibility weight increased by 0.05"), the corresponding rules are activated by the Mamdani reasoning method to synthesize a fuzzy set of credibility.

[0039] Defuzzification: The centroid of the fuzzy set is calculated using the centroid method to obtain the precise credibility weight (e.g., 0.92, 0.78). The lower limit of the credibility weight of the core component (liner, roller surface, bearing) data is set to 0.6 to ensure the priority of key data.

[0040] Output credibility weights: Credibility weights are ultimately assigned to different data sources: Vibration signal: 0.92, Oil report: 0.78, Temperature data: 0.85. An additional core component priority protection strategy is implemented: Liner, roller surface, and bearing data have a lower limit of 0.6 weight, with a 20% weight increase.

[0041] This model integrates multi-source heterogeneous data through fuzzy logic, solving the problem of reliability assessment in mining crushing and grinding equipment with complex environments and high data noise, and providing a reliable data foundation for equipment condition monitoring and fault diagnosis. Example 4

[0042] This embodiment provides a specific operation for component priority weighted fusion. Based on the data type of the grinding equipment and the importance of the components, a differentiated fusion strategy is adopted, and the fusion weight of core component data is increased by 20%.

[0043] Numerical data (vibration / temperature / pressure): Calculated using the formula "fusion value = Σ (data source value × credibility weight × component priority coefficient)", with a priority coefficient of 1.2 for core components and 1.0 for non-core components. Example: Mill bearing temperature sensor data (85℃, weight 0.9, priority coefficient 1.2) + infrared detection data (83℃, weight 0.7, priority coefficient 1.2), fusion value = (0.9 × 85 + 0.7 × 83) × 1.2 = 161.52℃ (corrected to 153.44℃ after adjusting for mill temperature calibration coefficient 0.95).

[0044] Textual data (oil fluid report / operation and maintenance record): Using the weighted bag-of-words model, a new fault keyword library for comminution equipment (such as liner looseness, roll surface wear, bearing seizure, ore feed overload, etc.) is added, and the weighted word frequency of keywords is calculated by doubling. Example: In the oil fluid report (weight 0.8), the word frequency of "excessive metal particles" (keyword) is 5, and in the operation and maintenance record (weight 0.6), the word frequency of this keyword is 3. After fusion, the weighted word frequency = (0.8×5 + 0.6×3)×2 = 11.6. Example 5

[0045] This example provides a specific method for the knowledge representation of multimodal fusion data. The high-quality fused multimodal data (numerical vectors and text features) needs to be further transformed into structured knowledge in order to be effectively managed and applied by the knowledge graph. This method uses entity recognition and relationship extraction technologies exclusive to comminution equipment to map the fused data into "entity-relationship-entity" triples, providing standardized knowledge input for subsequent dynamic graph updates. The process is as Figure 4 shown.

[0046] 1) Exclusive entity recognition for comminution equipment For the scenario of mine comminution equipment, a dedicated entity type system is constructed to accurately identify knowledge carriers from the fused data. The entity types include: Physical entities: Equipment level (such as "MM-003 mineral mill"), component level (such as "B-003 bearing", "LB-001 liner"), monitoring point level (such as "MM-003-B-003-01 vibration measurement point").

[0047] Index entities: Monitoring indicators (such as "vibration acceleration", "temperature", "pressure"), process parameters (such as "ore feed amount", "filling rate").

[0048] Fault entities: Fault modes (such as "bearing wear fault", "liner looseness fault"), fault phenomena (such as "abnormal noise", "powder leakage").

[0049] Maintenance entities: Maintenance actions (such as "replace bearing", "optimize ore feed amount"), maintenance results (such as "normal operation"), maintenance resources (such as "ZB-001 bearing").

[0050] Operating condition entities: Environmental conditions (such as "dust concentration 25mg / m 3 "), operating stages (such as "start-up stage").

[0051] Recognition method: For numerical fused data (such as fused temperature 153.44°C), rule-based ontology mapping is used: associating high-dimensional numerical values with preset physical entities and index entities to form attribute triples. For example: <MM-003-B-003-01, monitoring indicator_temperature, 153.44°C> For text-based fusion data (such as weighted bag-of-words vectors and original maintenance records), a natural language processing model fine-tuned for mining crushing and grinding equipment is used for extraction: BERT-BiLSTM-CRF model: Used to identify entity boundaries and types in text. It is fine-tuned using crushing and grinding equipment corpora (such as maintenance reports, fault case libraries) to ensure that the model can accurately identify professional terms such as "liner loosening", "roll pitting", etc.

[0052] Enhanced keyword matching: Combining the "fault keyword library for crushing and grinding equipment" (liner loosening, roll wear, bearing seizure, etc.) established in the first part, the model recognition results are weighted and corrected to improve the recall rate.

[0053] 2) Relationship extraction oriented to crushing and grinding scenarios After identifying entities, it is necessary to determine the semantic relationships between entities to form complete knowledge triples. This method constructs a relationship system for the operation and maintenance scenarios of crushing and grinding equipment, mainly including: Composition relationship: such as <Mineral mill, contains component, Bearing>, <Bearing, contains measurement point, Vibration measurement point 01> Monitoring relationship: such as <Vibration measurement point 01, monitoring index, Vibration acceleration>, <Vibration measurement point 01, collected value, 5.6m / s 2 > Characterization relationship: such as <Vibration acceleration 5.6m / s 2 , characterizes fault, Bearing wear>, <Fusion temperature 153.44°C, characterizes state, Overheating> Causal / association relationship: such as <Oversized ore feed, causes, Bearing wear>, <Bearing wear, shows phenomenon, Abnormal noise> Disposal relationship: such as <Bearing wear, take maintenance, Replace ZB-001 bearing>, <Replace ZB-001 bearing, produces result, Normal vibration> Extraction method: Rule template matching: For structured fusion data (such as fixed fields in event records), predefined rule templates are used for relationship extraction. For example, from the event record "MM-003-B-003-01, Vibration acceleration 5.6m / s 2 , Ore feed 125t / h", the triple can be extracted based on the template: <MM-003-B-003-01, monitoring index_vibration acceleration, 5.6m / s 2 > <MM-003, current working condition_ore feed, 125t / h> Remote supervised learning: For unstructured text data, existing relation instances in a knowledge graph are used as seeds to remotely annotate text corpora and train a relation classifier. For example, if the graph already contains <bearing wear, maintenance measures, bearing replacement>, then when both "bearing wear" and "bearing replacement" appear in maintenance text, they are used as positive examples to train the model, enabling relation extraction from new text.

[0054] 3) Generation of knowledge triples with enhanced fusion features The identified entities and relationships are combined to generate standardized knowledge triples. Simultaneously, the data confidence weights and component priority coefficients calculated in the first part are used as confidence attributes of the triples for labeling, thus achieving the transfer of data quality to knowledge quality.

[0055] Example 1 (Numerical): Fusion data: MM-003-B-003-01 has a fusion temperature of 153.44℃, a confidence level of 0.95, and a priority of 1.2. Generate triplet: (MM-003-B-003-01, hasMeasurement, Measurement_Temp_001)(Measurement_Temp_001, value, “153.44”)(Measurement_Temp_001, unit, “℃”)(Measurement_Temp_001, confidence, “0.95”)(Measurement_Temp_001, priority,"1.2") Example 2 (Text): Combined text features (weighted bag-of-words): {"Excessive metal particles": 10.72, "bearing": 5, "wear": 4} Based on the original text "Severe bearing wear, excessive metal particles in the oil", the following triplet was extracted: (Bearing_B-003, hasFault, WearFault_001)(WearFault_001, faultType, "Bearing wear")(WearFault_001, evidence, "Excessive metal particles in the oil")(WearFault_001, keywordWeight, "10.72") Through the above knowledge representation steps, the originally isolated and unstructured fused data is transformed into structured knowledge triples that are rich in semantics and support reasoning. These triples are not only the basic building blocks of the knowledge graph, forming the basic static knowledge graph, but also the direct operational objects for subsequent event monitoring and rule reasoning for dynamic updates of the knowledge graph. Example 6

[0056] This embodiment provides a method for constructing a dynamic knowledge graph, the process of which is as follows: Figure 5 As shown, the dynamic knowledge graph centers on "grinding equipment malfunction event triggering - dedicated rule reasoning - real-time graph updates," enabling real-time knowledge evolution and adapting to the needs of rapid fault response. 1) Event monitoring specifically for grinding equipment Real-time capture of two types of triggering events, focusing on typical faults and maintenance scenarios of grinding equipment, provides accurate basis for map updates: Equipment malfunction events: Preset specific thresholds for core components of the grinding equipment (bearing vibration acceleration > 5m / s²). 2 The following conditions must be met: liner temperature > 90℃, high-pressure roller mill pressure > 12MPa, feed rate fluctuation > 30%. Real-time comparison of operating data with these thresholds is required. If the threshold is exceeded, an event is triggered, and the following information is recorded: "Equipment Type - Equipment Number - Core Component - Monitoring Point - Abnormal Value - Operating Condition - Time" (e.g., "MM-002-B-002-03, Vibration Acceleration 5.8m / s²"). 2 Ore feed rate 120t / h, 2024-05-21T09:15:30.456Z”).

[0057] New maintenance record event: The monitoring operation and maintenance record entry interface will trigger an event when a record containing the specific fields of crushing and grinding equipment (faulty component-monitoring point-operating condition parameter-fault phenomenon-maintenance measure-maintenance result) is submitted (e.g., "MM-003-G-003-02, abnormal noise, feed rate overload, replace gear, normal operation").

[0058] 2) Fault-oriented knowledge reasoning for grinding and crushing equipment Employing a rule-based reasoning method, combined with a typical fault database of grinding and crushing equipment and existing diagrams, it outputs accurate reasoning results. The process is as follows: Figure 6 As shown Domain rule base optimization: Added typical fault reasoning rules for crushing and grinding equipment, including "high frequency vibration (1000-2000Hz) + temperature > 85℃ → bearing wear fault", "abnormal vibration at liner monitoring point + text containing 'powder leakage' → liner loosening fault", "uneven pressure in high pressure roller mill + roller surface temperature difference > 10℃ → asymmetric roller surface wear fault", "feed fluctuation > 30% + abnormal motor current → overload leading to component fatigue fault", etc.

[0059] Inference Execution: After an event is triggered, key information such as "component-operating condition-abnormal characteristics" is extracted and matched against a dedicated rule base. Example: "MM-002-B-002-03 vibration acceleration 5.8m / s²" was detected. 2 (Exceeding the threshold of 5m / s) 2"+temperature 86℃", matching the rule "high frequency vibration + temperature > 85℃ → bearing wear failure", it was found that the existing graph does not have the relationship "MM-002-B-002-03-wear failure", and the reasoning result is "add this relationship".

[0060] 3) Map updates and log recording Based on the reasoning results, a precise update operation is performed, and a dedicated log for the grinding equipment is recorded for easy maintenance and traceability. The update operation is as follows: Add new relationship: If the entity does not exist, create a node and add relationship edges (e.g., add a "fault characterization" edge between "MM-002-B-002-03" and "wear failure", with a confidence level of 0.85 and labeled with the working condition "feed rate 120t / h").

[0061] Correcting relationships: Adjusting relationship edge attributes (e.g., when verifying the maintenance record "Replace ZB-001 bearing + optimize feed rate to 100t / h → vibration disappears", the confidence of "ZB-001 bearing - wear fault elimination" is increased from 0.7 to 0.95).

[0062] Delete relationship: Remove the relationship that conflicts with the new event (e.g., if the original "vibration frequency 800Hz → bearing failure" is negated by the new data, delete the relationship and mark it "grinding equipment condition adaptability correction").

[0063] Update log: Records "Update time (UTC milliseconds) - Trigger event ID - Grinding equipment information - Inference rule - Update content - Operating condition label", example: "2024-05-21T09:20:00.123Z, EV-001, MM-002 (mineral mill) - B-002-03, rule 'high frequency vibration + temperature > 85℃ → bearing wear failure', added MM-002-B-002-03-wear failure relationship, operating condition: feed rate 120t / h". Example 7

[0064] This embodiment provides a multimodal data fusion and dynamic knowledge graph system for mining grinding equipment that matches the above-mentioned method, adapts to the harsh environment of mines and the operation and maintenance needs of grinding equipment, and achieves full-process automation, including the following modules: The functions of each module are as follows: 1) Data Acquisition Module: Used to acquire high-frequency vibration signals, oil analysis reports, temperature and pressure data, and operation and maintenance text records of mining crushing and grinding equipment; collects multi-source data of crushing and grinding equipment (vibration 1000-10000Hz, oil report once / day-week, temperature / pressure 1-10Hz, process parameters 0.1Hz, etc.), supports RS485, Ethernet, and API interfaces, and is equipped with a dust-protected acquisition terminal. The core component monitoring points adopt a dual-sensor redundant acquisition design.

[0065] Includes a spatiotemporal alignment module: implements UTC millisecond-level timestamps, adds unique spatial stamps for "device type-device number-core component-monitoring point", performs phased sampling frequency adjustment, and increases the frequency to 5Hz during the start / stop phase.

[0066] Credibility assessment module: Runs a fuzzy logic model specific to the grinding scenario, introduces working condition interference factors, and outputs credibility weights including component priority.

[0067] Data fusion module: performs weighted summation of numerical values ​​(including component priority coefficients) and weighted bag-of-words fusion of text (including rules for doubling fault keywords).

[0068] The knowledge representation module for multimodal fusion data is used to represent multimodal fusion data into structured knowledge, including entity recognition, relation extraction, and structured knowledge establishment. Entity recognition: Perform reverse entity recognition on the high-quality fusion dataset to extract physical entities, indicator entities, fault entities, and maintenance / operating condition entities; Relation extraction: Extract relations from the extracted entities and establish compositional relations, monitoring relations, representational relations, and causal / disposal relations between entities; Structured knowledge creation: Encapsulate entities and relations into triples (entity-relation-entity) and assign confidence attributes to triples to form structured knowledge.

[0069] The knowledge graph construction module is used to build dynamic knowledge graphs and output dynamically updated knowledge graphs in real time; including... Event monitoring module: Monitors abnormal events (triggered by a dedicated threshold) and new maintenance record events (triggered by a dedicated field) of core components of grinding equipment.

[0070] Knowledge Reasoning Module: Based on the rule base of typical faults of grinding and crushing equipment and the existing graph reasoning, it outputs the results of adding / correcting / deleting relationships.

[0071] Knowledge graph update module: Performs graph update operations, marking working condition information and component associations.

[0072] Log recording module: Stores dedicated update logs for grinding equipment, and supports querying and exporting by equipment number, component type, and operating conditions.

[0073] Storage module: Adopting a distributed architecture, it is adapted to the deployment requirements of edge computing in mines, stores raw data, fused data, map data and logs, and has stable storage capabilities in high temperature and high dust environments.

[0074] The update and traceability module is used to perform add, delete, and modify operations on the knowledge graph and record update logs containing working condition information.

[0075] The fault diagnosis module, based on a dynamic knowledge graph, enables fault identification, location, and early warning for mining grinding equipment. Example 8

[0076] This embodiment uses a mineral mill (equipment type: mineral mill, equipment number MM-003) in a mining enterprise as an example to verify the effectiveness of the method provided by the present invention. The implementation steps are as follows: (I) Implementation of Multimodal Data Fusion Data Acquisition: Collect data from core components, including vibration signal (10000Hz) of bearing B-003 (monitoring point 01 / 02), temperature (1Hz) of gearbox G-003 (monitoring point 01), pressure (1Hz) of motor M-003 (monitoring point 01), oil reports (once / day), process parameters (feed rate, grinding media filling rate, 0.1Hz), and operation and maintenance records; dust-protected sensors are used, and dual sensors are used for redundant acquisition at the bearing monitoring points.

[0077] Spatiotemporal alignment: UTC timestamps are accurate to milliseconds, and spatial stamps are “MM-003-part number-monitoring point number” (e.g., “MM-003-B-003-01”); oil reports are upsampled to 1Hz using cubic spline interpolation, vibration signals are downsampled to 1Hz using mean downsampling, and process parameters are interpolated to 1Hz; during the startup phase (0-30 minutes), the frequency is increased to 5Hz, and a startup flag is added.

[0078] Reliability assessment: Vibration signal signal-to-noise ratio 45dB (high, weight 0.95); high oil report completeness (including 4 key information items specific to grinding equipment, weight 0.9); temperature data (gearbox) weight 0.85; pressure data weight 0.8; operating condition interference factor is low (dust concentration 25mg / m³). 3 The weight of each data source is uniformly increased by 0.1; the priority coefficient of core components (bearings, gearboxes) is set to 1.2.

[0079] Weighted fusion: Gearbox temperature fusion value = (0.85×82+0.7×83)×1.2=161.52℃, which is corrected to 153.44℃ after combining the temperature calibration coefficient of the grinding equipment of 0.95 (based on the linear regression result of the measured temperature and sensor data during shutdown maintenance); Weighted frequency of "excessive metal particles" (keyword of grinding equipment failure) = (0.9×4+0.88×2)×2=10.72.

[0080] (II) Implementation of Dynamic Knowledge Graph Event monitoring: Vibration acceleration of 5.6 m / s² was detected in B-003-01. 2 (Exceeding the threshold of 5m / s) 2+ Temperature 87℃ (exceeding the threshold of 85℃) triggered an abnormal event, recording "MM-003-B-003-01, vibration acceleration 5.6m / s²". 2 Temperature 87℃, feed rate 125t / h, 2024-06-01T14:30:00.789Z; the maintenance personnel submitted a record of "Vibration returned to normal after replacing ZB-001 bearing + optimizing feed rate to 100t / h", triggering a maintenance event.

[0081] Knowledge Reasoning: To match the specific rule for grinding equipment, "High-frequency vibration + temperature > 85℃ → bearing wear failure", the reasoning adds the relationship "MM-003-B-003-01-wear failure"; To match the rule "Bearing replacement + failure disappears after working condition optimization → adaptation relationship confirmed", the reasoning adds the adaptation relationship "ZB-001 bearing-MM-003-B-003-01 wear failure", and corrects the confidence of "abnormal vibration - wear failure" to 0.95.

[0082] Map update: Added ZB-001 node, "MM-003-B-003-01-wear failure" relationship edge, and "ZB-001 bearing-MM-003-B-003-01 wear failure" adaptation relationship edge, and marked the working condition "feed rate 125t / h"; recorded update log, including dedicated fields such as equipment type, component monitoring points, and working condition information.

[0083] (III) Implementation Results After three months of application, the various indicators have been verified to meet the operation and maintenance requirements of grinding and crushing equipment: The accuracy of data integration with actual equipment status is improved by 35%, the initial judgment accuracy of typical faults in crushing and grinding equipment reaches 92%, and the fault location accuracy of core components reaches 95%; the knowledge graph update timeliness rate is 100%, the response time for incorporating new fault modes of crushing and grinding equipment is less than 5 minutes, and the fault prediction accuracy is improved by 28%; unplanned downtime is reduced by 22%, the fault judgment efficiency of operation and maintenance personnel is improved by 45%, equipment maintenance costs are reduced by 18%, and the stability of ore feed is improved by 30%.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for constructing a multimodal dynamic knowledge graph for mining crushing and grinding equipment, characterized in that, Includes the following steps: S1. Multimodal data acquisition and input: Acquire multi-dimensional monitoring data of mining crushing and grinding equipment, including high-frequency vibration signals, oil analysis reports, temperature and pressure data, and operation and maintenance records; S2. High-quality multimodal data fusion: including spatiotemporal alignment processing, credibility assessment, and priority-weighted fusion to generate high-quality multimodal fused datasets; S3. Knowledge representation of high-quality multimodal fusion datasets: including entity recognition, relation extraction, and structured knowledge construction, among which, Entity recognition: Named entity recognition for grinding equipment is performed on the high-quality fusion dataset to extract physical entities, indicator entities, fault entities, and maintenance / operating condition entities; Relation extraction: Extract relations from the extracted entities and establish compositional relations, monitoring relations, representational relations, and causal / disposal relations between entities; Structured knowledge creation: Encapsulate entities and relations into triples (entity-relation-entity) and assign confidence attributes to triples to form structured knowledge; S4. Construction of Dynamic Knowledge Graph: This includes event monitoring, graph storage, fault-oriented knowledge reasoning, dynamic graph updating and tracing, and graph output and application. Event monitoring: Based on the structured knowledge, equipment event monitoring is performed to identify abnormal events and maintenance events; Knowledge graph storage: The triples and the monitored events are stored as a knowledge graph; Fault-oriented knowledge reasoning: Based on the built-in fault rule library for grinding equipment, abnormal events are associated and matched with fault types to deduce the correspondence between equipment abnormalities and faults such as component wear, loosening, and blockage. Knowledge graph dynamic update and traceability: Perform update operations such as adding, deleting, and modifying on the knowledge graph, record the update time, interface, specific content, and update effect, and form a traceable update log; Knowledge Graph Output and Application: Outputs a dynamic knowledge graph that is updated in real time, enabling real-time diagnosis and rapid response to faults in mining crushing and grinding equipment.

2. The method for constructing a multimodal dynamic knowledge graph according to claim 1, characterized in that, The spatiotemporal alignment processing includes millisecond-level UTC timestamp calibration of the collected multimodal data, segmented matching based on the spatial dimensions of equipment, components, and monitoring points, and phased sampling frequency adjustment according to the operating characteristics of the grinding equipment; the credibility assessment includes constructing a fuzzy logic model specific to the grinding scenario, introducing operating condition interference factors to calculate data credibility, and outputting dynamic credibility weights in the 0-1 range; the priority weighted fusion includes adopting a strategy of weighted summation of numerical data, weighted fusion of textual data, and keyword confidence calculation to give weight enhancement to core monitoring components, complete multimodal data fusion, and generate a high-quality fused dataset.

3. The method for constructing a multimodal dynamic knowledge graph according to claim 2, characterized in that, The input variables of the fuzzy logic model include the signal-to-noise ratio of numerical data, the text integrity of text data, and the stability of the acquisition device. The output credibility weight is used to reduce the weight or filter low-quality data.

4. The method for constructing a multimodal dynamic knowledge graph according to claim 2, characterized in that, In the aforementioned priority-weighted fusion, a 20% weight increase is determined for core monitoring components through sensitivity analysis of historical fault data.

5. The method for constructing a multimodal dynamic knowledge graph according to claim 1, characterized in that, The composition relationship is used to characterize the inclusion / subordination relationship between entities, the monitoring relationship is used to characterize the association between sensors and indicator entities, the characterization relationship is used to characterize the association between indicator entities and equipment health status, and the causal / disposal relationship is used to characterize the triggering and disposal logic between faulty entities and abnormal events and maintenance events.

6. The method for constructing a multimodal dynamic knowledge graph according to claim 1, characterized in that, The confidence attribute is used to measure the reliability of triple knowledge and is calculated by combining the data source, extraction algorithm accuracy, and historical verification results.

7. The method for constructing a multimodal dynamic knowledge graph according to claim 1, characterized in that, The abnormal equipment events include vibration, temperature, current, and abnormal feed rate parameters.

8. The method for constructing a multimodal dynamic knowledge graph according to claim 1, characterized in that, The dedicated fault rule library for the grinding equipment includes: high-frequency vibration amplitude ≥ threshold corresponds to bearing wear fault, abnormal temperature ≥ threshold corresponds to bearing wear fault, uneven pressure corresponds to liner wear fault, and abnormal feed rate corresponds to feed channel blockage fault. The rule library supports iterative updates based on actual operation and maintenance data.

9. The method for constructing a multimodal dynamic knowledge graph according to claim 1, characterized in that, The update log records include the update operation type, trigger event number, and updated graph node / relationship information, which are used for subsequent fault tracing and model optimization, as well as periodically adjusting the rule confidence in the fault rule base to achieve adaptive optimization of the rule base.

10. An intelligent operation and maintenance system for mining crushing and grinding equipment, characterized in that, include: The data acquisition module is used to acquire high-frequency vibration signals, oil analysis reports, temperature and pressure data, and operation and maintenance records of the mining crushing and grinding equipment. The data fusion module is used to perform high-quality multimodal data fusion steps and generate a high-quality fused dataset; The knowledge representation module for multimodal fusion data is used to represent multimodal fusion data in a knowledge-based manner to form structured knowledge. The knowledge graph construction module is used to build dynamic knowledge graphs and output dynamically updated knowledge graphs in real time. The fault diagnosis module, based on a dynamic knowledge graph, enables fault identification, location, and early warning for mining grinding equipment. The update and traceability module is used to perform add, delete, and modify operations on the knowledge graph and record the update date containing working condition information.

Images