Coal industry chain multi-modal data grading and ai training method, device and equipment

Abstract

The application provides a coal industry chain multi-modal data grading and AI training method, device and equipment, and relates to the technical field of coal. The method comprises the following steps: constructing a knowledge graph ontology model based on the elements of the coal industry chain; performing semantic mapping on the multi-modal data of the coal industry chain based on the knowledge graph ontology model; performing data grading on the multi-modal data by using the knowledge graph ontology model to obtain the grading labels of each kind of coal data in the multi-modal data; the grading labels of the coal data at least include the safety level and the quality score of the coal data; performing partition isolation on the multi-modal data based on the grading labels; and determining the model training strategy corresponding to the coal data in each partition after the partition isolation based on a set partition training strategy. The application solves the problem of how to provide a grading mechanism capable of fusing the semantic structure of the knowledge graph and the credibility features of the multi-modal data, and driving the partition training of the AI model.

Description

Technical Field

[0001] This invention relates to the field of coal technology, specifically to a method for multimodal data classification and AI training in the coal industry chain, a device and electronic equipment for multimodal data classification and AI training in the coal industry chain. Background Technology

[0002] With the deepening of the digital transformation of the energy industry, the coal industry chain, as a fundamental pillar industry of the country, generates massive amounts of heterogeneous data across its exploration, mining, washing, coal chemical, storage, transportation, and sales processes. This includes structured production reports, unstructured equipment logs, image and video surveillance, Geographic Information System (GIS) information, and supply chain text records—multimodal data. This data contains crucial information reflecting the operational status, risks, and optimization potential of the industry chain, and is of great value for improving safe production, energy efficiency management, and intelligent decision-making. However, coal enterprises currently generally lack the systematic governance capabilities for multimodal data, making it difficult to achieve cross-modal semantic association and knowledge accumulation. This results in data resources remaining dormant for a long time, unable to effectively support the training and application of artificial intelligence models. A knowledge graph-based industry data organization method provides a new approach to solving these problems. This method aims to map multi-source data scattered across various business systems into a unified knowledge network through entity extraction, relation modeling, and semantic reasoning, thereby revealing the deep logical connections between upstream and downstream elements of the industry chain.

[0003] However, existing technologies still face multiple challenges in constructing knowledge graphs for the coal industry chain: On the one hand, multimodal data lacks a unified grading standard, and data from different sources, formats, and levels of credibility are treated equally, leading to severe noise interference during graph construction and diluting key high-value information with low-quality data; on the other hand, AI (Artificial Intelligence) model training typically adopts an "end-to-end" black-box approach, directly using raw or simply pre-processed data for learning, failing to fully utilize the semantic hierarchy and causal logic inherent in knowledge graphs. This results in weak model generalization ability, poor interpretability, and difficulty in adapting to the rapid response needs of complex and dynamic scenarios in the coal industry chain (such as sudden geological disasters, abnormal chemical processes, and transportation disruptions). Especially in high-risk scenarios involving safety supervision and production scheduling, if the model cannot distinguish the reliability level of data and reason based on credible knowledge, it is prone to making misleading decisions, severely restricting the safety and effectiveness of deep application of artificial intelligence in the coal industry. Therefore, there is an urgent need for a grading mechanism that can integrate the semantic structure of knowledge graphs and the credibility characteristics of multimodal data, and use this to drive the partitioned training of AI models.

[0004] It should be noted that the information disclosed in the background section of this invention is intended only to enhance the understanding of the overall background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method, apparatus, and equipment for multimodal data grading and AI training in the coal industry chain, in order to solve the problem of how to provide a grading mechanism that can integrate the semantic structure of knowledge graphs and the credibility features of multimodal data, and thereby drive the partitioned training of AI models.

[0006] To achieve the above objectives, embodiments of the present invention provide a method for multimodal data classification and AI training in the coal industry chain, including: A knowledge graph ontology model is constructed based on the elements of the coal industry chain; the knowledge graph ontology model includes entities constructed based on the elements of the coal industry chain and the attribute relationships between the entities; Semantic mapping is performed on multimodal data of the coal industry chain based on the knowledge graph ontology model. The knowledge graph ontology model is used to classify the multimodal data to obtain a classification label for each type of coal data in the multimodal data; the classification label of the coal data includes at least the safety level and quality score of the coal data; The multimodal data is partitioned and isolated based on the hierarchical labels; Based on the set partition training strategy, the model training strategy corresponding to the coal data of each partition after partition isolation is determined.

[0007] Optionally, the semantic mapping of multimodal data in the coal industry chain based on the knowledge graph ontology model includes: Based on the data attributes of each type of coal data in the multimodal data, a set of candidate entities in the knowledge graph ontology model that matches the data attributes is determined; The numerical field names of the coal data are compared with the attribute name list of each entity in the candidate entity set; Entities whose comparison results are consistent and whose units of the coal data conform to preset rules are identified as target entities; The coal data is connected to the target entity using preset attribute relationships to establish a semantic link between the coal data and the target entity.

[0008] Optionally, the step of using the knowledge graph ontology model to perform data classification on the multimodal data to obtain classification labels for each type of coal data in the multimodal data includes: The knowledge graph ontology model is used to classify each type of coal data in the multimodal data for business purposes, thereby obtaining the business domain of each type of coal data in the multimodal data. The knowledge graph ontology model is used to perform sensitivity assessment on each type of coal data in the multimodal data to obtain the security level of each type of coal data in the multimodal data. The knowledge graph ontology model is used to evaluate the data quality of each type of coal data in the multimodal data, and a quality score for each type of coal data in the multimodal data is obtained. The knowledge graph ontology model is used to set a time window for each type of coal data in the multimodal data, thereby obtaining the time window for each type of coal data in the multimodal data. The business domain, the security level, the quality score, and the timeliness window serve as the grading labels for each type of coal data.

[0009] Optionally, the step of using the knowledge graph ontology model to perform data quality assessment on each type of coal data in the multimodal data to obtain a quality score for each type of coal data in the multimodal data includes: The knowledge graph ontology model is used to determine the time freshness score, integrity score, measurement accuracy score, and data consistency score for each type of coal data in the multimodal data. Based on the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score, a quality score for each type of coal data in the multimodal data is determined; Among them, the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score are all positively correlated with the quality score.

[0010] Optionally, the security levels include Level 1 Sensitive, Level 2 Sensitive, and Level 3 Non-Sensitive; The partitioning and isolation of the multimodal data based on the hierarchical labels includes: If the security level of the coal data is Level 1 Sensitive and the quality score of the coal data is greater than or equal to a first set threshold, then the coal data is determined to be located in the safe training zone. If the security level of the coal data is Level 2 Sensitive and the quality score of the coal data is greater than or equal to the second set threshold, then the coal data is determined to be located in the restricted training area. If the security level of the coal data is Level 3 Non-sensitive and the quality score of the coal data is greater than or equal to the third set threshold, then the coal data is determined to be located in the open training area. Wherein, the first set threshold is greater than the second set threshold, the second set threshold is greater than the third set threshold; the data access difficulty of the secure training area is higher than the data access difficulty of the restricted training area, and the data access difficulty of the restricted training area is higher than the data access difficulty of the open training area.

[0011] Optionally, the step of determining the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy includes: The coal data of the designated safe training area is used to train an AI model using a federated learning method. The coal data in the restricted training area is used for AI model training using a differential privacy mechanism; The coal data in the open training area is used to train the AI ​​model using a supervised learning method.

[0012] Optionally, after determining the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy, the method further includes: In the knowledge graph ontology model, coal sample data is determined based on the attribute set and attribute relationship path of the entity corresponding to the target AI model; The method for constructing target features is determined based on the type of the coal sample data, and the target feature construction method is used to construct features on the coal sample data to obtain target input features; the target feature construction method is one of the following: sliding window method, inverse distance weighting method, and event counting and time-series coding combination method; The target AI model is trained based on the target input features and the model training strategy corresponding to the coal sample data.

[0013] Optionally, the loss function in the training process of the target AI model is constructed based on the data fitting loss term and the semantic consistency regularization term; The data fitting loss term and the semantic consistency regularization term are positively correlated with the loss function, respectively; the data fitting loss term is constructed using mean squared error or cross-entropy; the semantic consistency regularization term characterizes the degree of logical violation between the model prediction results and the attribute relationship paths of the knowledge graph ontology model.

[0014] On the other hand, embodiments of the present invention also provide a multimodal data classification and AI training device for the coal industry chain, comprising: A construction module is used to build a knowledge graph ontology model based on the elements of the coal industry chain; the knowledge graph ontology model includes entities built based on the elements of the coal industry chain and the attribute relationships between the entities; The mapping module is used to perform semantic mapping on multimodal data of the coal industry chain based on the knowledge graph ontology model. The grading module is used to perform data grading on the multimodal data using the knowledge graph ontology model, and to obtain a grading label for each type of coal data in the multimodal data; the grading label for the coal data includes at least the safety level and quality score of the coal data. An isolation module is used to partition and isolate the multimodal data based on the hierarchical labels; The determination module is used to determine the model training strategy corresponding to the coal data of each partition after partition isolation, based on the set partition training strategy.

[0015] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned coal industry chain multimodal data classification and AI training method.

[0016] Through the above technical solutions, this invention, by constructing a knowledge graph ontology model covering all elements of the coal industry chain, achieves for the first time a unified semantic description and structured association of multimodal data, including coal mines, coal chemicals, and railway and highway transportation, solving the fundamental problems of data silos and semantic fragmentation in traditional systems. Based on this, this invention proposes a multi-dimensional data grading mechanism driven by a knowledge graph ontology model, organically combining security levels and quality scores to establish a unified grading standard, providing refined management strategies for coal data with different sensitivity and quality levels. Simultaneously, it partitions and isolates the multimodal data and determines the model training strategies corresponding to the coal data in each partition. This invention not only ensures the security of highly sensitive industrial data but also fully utilizes the training value of open data, improving the model's generalization ability, interpretability, and adaptability to complex dynamic scenarios. Therefore, this invention proposes a grading mechanism that integrates the semantic structure of knowledge graphs and the credibility features of multimodal data, and uses this to drive the partitioned training of AI models.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is one of the flowcharts illustrating the multimodal data grading and AI training method for the coal industry chain provided by this invention; Figure 2This is the second flowchart illustrating the multimodal data grading and AI training method for the coal industry chain provided by this invention. Figure 3 This is a schematic diagram of the process for partitioning and isolating multimodal data provided by the present invention; Figure 4 This is the third flowchart of the multimodal data classification and AI training method for the coal industry chain provided by the present invention; Figure 5 This is a schematic diagram of the structure of the coal industry chain multimodal data classification and AI training device provided by the present invention; Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0020] In existing technologies, various subsystems in the coal industry chain, such as coal mine monitoring, coal quality testing, chemical reaction process control, railway scheduling, and vehicle route planning, generally adopt independent data acquisition and processing mechanisms, lacking a unified data semantic framework and cross-domain correlation capabilities. At the same time, the lack of systematic hierarchical management of data leads to the improper exposure of highly sensitive data during model training, or low-value noise data interfering with model convergence, seriously affecting the application of artificial intelligence in complex industrial scenarios.

[0021] Therefore, the purpose of this invention is to provide a method, apparatus, and equipment for multimodal data grading and AI training in the coal industry chain, in order to solve the problem of how to provide a grading mechanism that can integrate the semantic structure of knowledge graphs and the credibility features of multimodal data, and thereby drive the partitioned training of AI models.

[0022] Method Implementation Examples Please refer to Figure 1 This invention provides a method for multimodal data classification and AI training in the coal industry chain, including: Step 100: Construct a knowledge graph ontology model based on the elements of the coal industry chain.

[0023] The knowledge graph ontology model based on coal industry chain elements is jointly implemented by electronic devices (including processors and memory) and human experts. Human experts are responsible for defining core entities covering coal industry chain elements and their attribute relationships, then inputting these core entities and their attribute relationships into the memory of the electronic devices. The processor of the electronic devices then constructs the knowledge graph ontology model based on these core entities and their attribute relationships. Therefore, the knowledge graph ontology model in this embodiment includes entities constructed based on coal industry chain elements and the attribute relationships between these entities. In some embodiments, entities (or entity types) constructed based on coal industry chain elements may include coal mine faces, raw coal batches, washed coal products, gasification furnace units, ammonia synthesis units, methanol reactors, railway marshalling yards, freight trains, heavy-duty trucks, drivers, meteorological stations, geological structural units, etc. Attribute relationships between entities may include "originating from," "input to," "located at," "belonging to," "controlled by," "transported via," "influenced by," etc. The knowledge graph ontology model in this embodiment describes the causal, temporal, spatial, and technological constraint relationships between entities through formal logical rules, forming a reasonable semantic network. The formal logic rules are expressed using a descriptive logic language. For example, if entity A is a coal mine working face, entity B is a batch of raw coal, and the relation "produced from" exists, then the calorific value of B must not exceed the historical maximum calorific value of the coal seam to which A belongs. If entity C is a gasification furnace unit, entity D is a synthetic ammonia unit, and the relation "input to" exists, then the output gas composition ratio of C must meet the feed specification constraints of D. The knowledge graph ontology model is stored in the graph database of electronic devices in the form of triples. Each triple contains a subject, predicate, and object, and supports SPARQL query and inference engine calls.

[0024] Step 200: Perform semantic mapping on the multimodal data of the coal industry chain based on the knowledge graph ontology model.

[0025] Multimodal data can be multi-source heterogeneous data. In some embodiments, multi-source heterogeneous data includes raw data from at least five types of data sources: the first type is real-time monitoring data output by underground coal mine sensor networks, including gas concentration, roof displacement, coal mining machine position, and conveyor belt speed; the second type is batch test reports issued by coal quality testing laboratories, including moisture content, ash ratio, volatile matter, fixed carbon, sulfur content, and calorific value; the third type is process variables recorded by coal chemical DCS (Distributed Control System), including reaction temperature, pressure, flow rate, liquid level, and catalyst activity index; the fourth type is operation diagram data generated by railway transportation dispatching systems, including train number, origin and destination stations, marshalling information, planned time, and actual arrival and departure time; the fifth type is location trajectory and vehicle weighing data uploaded by highway logistics GPS / BD (BeiDou) terminals, including latitude and longitude coordinates, driving speed, load tonnage, and start / stop events. Each type of coal data in the multimodal data is the raw data from the above five types of data sources. In some embodiments, semantic mapping of multimodal data in the coal industry chain based on the knowledge graph ontology model includes: determining a set of candidate entities in the knowledge graph ontology model that matches the data attributes of each type of coal data in the multimodal data; comparing the numerical field names of the coal data with the list of attribute names of each entity in the candidate entity set; determining entities whose comparison results are consistent and whose units of the coal data conform to preset rules as target entities; and connecting the coal data with the target entities using preset attribute relationships to establish a semantic link between the coal data and the target entities.

[0026] It should be noted that each type of coal data in the multimodal data has corresponding data attributes, including data source identifier, timestamp, spatial coordinates, equipment number, numerical fields, and units. The electronic equipment performs pattern matching between the extracted data attributes (i.e., data source identifier, timestamp, spatial coordinates, equipment number, numerical fields, and units) and the entity attributes in the knowledge graph ontology model. Pattern matching employs a dual verification mechanism based on regular expressions and ontology attribute definitions: first, a candidate entity set is determined based on the data attributes; then, the numerical field names are compared with the attribute name list of each entity in the set. If the names are completely identical and the units conform to the preset specifications (i.e., preset rules), the match is considered successful. After a successful match, the data record is instantiated as an attribute assertion for the corresponding entity, and a semantic link is established between it and related entities. For example, a record of gas concentration data (i.e., a data attribute) includes the equipment number "WJ-2024-087", the timestamp "June 15, 2024, 2:30 PM", the value "1.2 percent", and the unit "volume percentage". By querying the knowledge graph ontology model through the equipment number, it is determined that the entity to which it belongs is "Coal Mine Working Face_West Wing No. 3". The attribute name "Real-time Gas Concentration" under "Coal Mine Working Face_West Wing No. 3" matches the gas concentration data, and the value of the gas concentration data conforms to preset rules. Simultaneously, based on the "located" relationship, the gas concentration data is linked to the entity "Coal Mine_Datong Mining Area". If a match fails, it is temporarily stored in the annotation buffer, where domain experts intervene to correct the knowledge graph ontology model or expand the attribute definition. The annotation buffer uses a first-in, first-out queue management system, with a manual review process triggered every morning.

[0027] Step 300: Use the knowledge graph ontology model to classify the multimodal data and obtain the classification label for each type of coal data in the multimodal data.

[0028] The classification labels for coal data include at least a security level and a quality score. The knowledge graph ontology model includes a set of sensitivity judgment rules pre-set by human experts and a pre-set data quality assessment method. The electronic device uses the pre-set sensitivity judgment rules to assess the security level of the coal data; and uses the pre-set data quality assessment method to assess the quality of the coal data, obtaining a quality score. The pre-set data quality assessment method assesses the quality of the coal data from at least two dimensions: time freshness, data integrity, data accuracy, and data consistency. In other embodiments, the classification labels may include only one of the security level and quality score.

[0029] Step 400: Partition and isolate the multimodal data based on the hierarchical labels.

[0030] Electronic devices can partition and isolate the multimodal data based on two dimensions: security level and quality score. For example, electronic devices can divide the multimodal data into a secure training area (or high-security training area), a restricted training area, and an open training area by comparing it with preset thresholds for security level and quality score. In other embodiments, electronic devices can also divide the multimodal data into a secure training area, a restricted training area, and an open training area based on a single dimension of security level and quality score by comparing it with preset thresholds for security level or quality score. The data access difficulty in the secure training area is higher than that in the restricted training area, and the data access difficulty in the restricted training area is higher than that in the open training area.

[0031] Step 500: Determine the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy.

[0032] In one embodiment, the electronic device determines the model training strategy corresponding to the coal data of each partition after partition isolation based on a set partition training strategy. This includes: the electronic device determining that the coal data in the secure training area is used for AI model training using a federated learning method; the electronic device determining that the coal data in the restricted training area is used for AI model training using a differential privacy mechanism; and the electronic device determining that the coal data in the open training area is used for AI model training using a supervised learning method. Specifically, the federated learning method completes model forward propagation and gradient calculation locally, only uploading encrypted gradients to the central aggregation node. The central aggregation node performs secure multi-party computation to update the global model parameters. The differential privacy mechanism injects noise conforming to a Laplace distribution into the numerical fields of the training samples, and the noise scale is dynamically adjusted according to the data sensitivity level. This invention constructs a knowledge graph ontology model covering all elements of the coal industry chain, performing structured analysis and semantic mapping on multi-source heterogeneous data from coal mines, coal washing plants, coal chemical plants, railway transportation nodes, and highway logistics terminals. Based on the knowledge graph ontology model, it automatically classifies, assesses sensitivity, and labels the multimodal raw data. According to the classification results, it constructs a hierarchical and isolated AI training data pool and designs a matching model training strategy to achieve highly secure, efficient, and robust intelligent model iteration.

[0033] This invention, through the construction of a knowledge graph ontology model covering all elements of the coal industry chain, achieves for the first time a unified semantic description and structured association of multimodal data, including coal mines, coal chemicals, and railway and highway transportation, solving the fundamental problems of data silos and semantic fragmentation in traditional systems. Building upon this, this invention proposes a multi-dimensional data grading mechanism driven by a knowledge graph ontology model, organically combining security levels and quality scores to establish a unified grading standard, providing refined management strategies for coal data with different sensitivity and quality levels. Simultaneously, it partitions and isolates the multimodal data and determines the model training strategies corresponding to the coal data in each partition. This invention not only ensures the security of highly sensitive industrial data but also fully utilizes the training value of open data, improving the model's generalization ability, interpretability, and adaptability to complex dynamic scenarios. Therefore, this invention proposes a grading mechanism that integrates the semantic structure of a knowledge graph with the credibility features of multimodal data, and uses this to drive the partitioned training of AI models.

[0034] In other aspects of the embodiments of the present invention, the step of using the knowledge graph ontology model to classify the multimodal data and obtain a classification label for each type of coal data in the multimodal data includes: using the knowledge graph ontology model to classify each type of coal data in the multimodal data into a business domain; using the knowledge graph ontology model to perform sensitivity assessment on each type of coal data in the multimodal data to obtain a security level for each type of coal data in the multimodal data; using the knowledge graph ontology model to perform data quality assessment on each type of coal data in the multimodal data to obtain a quality score for each type of coal data in the multimodal data; and using the knowledge graph ontology model to set a timeliness window for each type of coal data in the multimodal data to obtain a timeliness window for each type of coal data in the multimodal data.

[0035] Please refer to Figure 2In one embodiment, to achieve more refined data classification, the electronic device uses business domain, security level, quality score, and timeliness window as classification labels for each type of coal data. First, the electronic device uses the knowledge graph ontology model to classify each type of coal data in the multimodal data into business domains, obtaining the business domain for each type of coal data in the multimodal data. Specifically, the electronic device determines the business domain to which the coal data belongs based on the entity type of the knowledge graph ontology model associated with it. These business domains include safety production, product quality, process control, transportation scheduling, and personnel management. For example, gas concentration data is associated with the entity "coal mine working face" and belongs to the safety production domain; coal calorific value data is associated with the entity "raw coal batch" and belongs to the product quality domain; gasifier temperature data is associated with the entity "gasifier unit" and belongs to the process control domain; train arrival and departure time data is associated with the entity "freight train" and belongs to the transportation scheduling domain; and driver location data is associated with the entity "driver" and belongs to the personnel management domain. Business domain codes can be represented by two digits: 0-1 for safety production, 0-2 for product quality, 0-3 for process control, 0-4 for transportation scheduling, and 0-5 for personnel management. Next, the electronic device assigns a safety level value to each coal data instance based on the sensitivity judgment rule set preset by the knowledge graph ontology model. The sensitivity judgment rule set includes: if the coal data involves underground personnel positioning or gas exceeding limit alarms, it is marked as Level 1 sensitive; if the coal data contains key coal chemical process parameters or unpublished coal quality ratio schemes, it is marked as Level 2 sensitive; if the coal data is public transportation plans or historical meteorological information, it is marked as Level 3 non-sensitive. The safety level assignment process is executed by the rule engine. The rules are encoded in the form of decision trees, with the input being the entity type, attribute name, and numerical range of the data instance, and the output being the safety level. Finally, the electronic device calculates the quality score of its coal data by combining the scoring rules preset by the knowledge graph ontology model for data freshness, numerical integrity, measurement accuracy, and consistency with other data. The quality score ranges from zero to one hundred. The timeframe window is dynamically set according to task requirements: five minutes for safety production data, thirty minutes for process control data, and two hours for transportation scheduling data. Finally, electronic equipment generates a four-dimensional hierarchical label based on the integrated business domain, security level, and data quality score, in the format (business domain code, security level, quality score, timeframe window). This embodiment of the invention achieves refined data management through a four-dimensional hierarchical mechanism.

[0036] In one embodiment, the step of using the knowledge graph ontology model to evaluate the data quality of each type of coal data in the multimodal data to obtain a quality score for each type of coal data in the multimodal data includes: using the knowledge graph ontology model to determine the time freshness score, integrity score, measurement accuracy score, and data consistency score for each type of coal data in the multimodal data; and determining the quality score for each type of coal data in the multimodal data based on the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score; wherein the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score are positively correlated with the quality score.

[0037] Specifically, time freshness refers to the interval between the coal data acquisition time and the current system time; the shorter the interval, the higher the score, and a score of zero is given if the interval exceeds 72 hours. Numerical completeness refers to the proportion of missing fields; full marks are given for no missing fields, and ten points are deducted for each missing key field. Data measurement accuracy is set according to the sensor calibration certificate; full marks are given if the accuracy error is within the allowable range, and points are deducted proportionally if it exceeds the allowable range. Data consistency is achieved by comparing with existing facts in the knowledge graph ontology model. For example, if the sulfur content of a batch of coal deviates significantly from the historical average of the same coal seam by more than three standard deviations, the consistency score is zero. In one embodiment, the formula for calculating the quality score is: ; Where Q is the quality score, T is the time freshness score (0 to 100), C is the integrity score (0 to 100), P is the measurement accuracy score (0 to 100), and A is the measurement consistency score (0 to 100). This embodiment of the invention calculates the quality score of coal data by integrating multiple dimensions such as time freshness score, integrity score, measurement accuracy score, and data consistency score, which helps to improve the comprehensiveness and accuracy of data quality assessment.

[0038] In one specific embodiment, the time freshness score T is calculated as the interval between the data acquisition time and the current system time. The calculation formula is: The time unit is hours; if the collection time exceeds 72 hours, then T=0. The completeness score C is calculated based on the proportion of missing required fields in the data record: C=100-∑(number of missing key fields × 5), with a minimum of 0 points; key fields include, but are not limited to, "sensor number," "timestamp," and "numerical unit." The measurement accuracy score P is determined based on the sensor calibration record: if the measured value error is within ±10% of the certificate's nominal error, then P=100; deduct 5 points for every ±1% exceeding this range, until all points are deducted. The consistency score A is calculated by comparing the current data value with the statistical distribution of similar data in the knowledge graph ontology model: if the current data value is within ±3 standard deviations of the historical data mean, then A=100; otherwise, A=0. The historical data statistical period is the most recent 30 days.

[0039] For example, assuming that a gas concentration data point was collected 2 hours before the current time, with no missing key fields, a sensor error of 0.5% (within ±10%), and the value conforms to historical statistical patterns, then: T=100-(2 / 3600)×100≈99.94, C=100, P=100, A=100, therefore the quality score Q=0.3×99.94+0.25×100+0.25×100+0.2×100≈97.48.

[0040] In other embodiments, the quality score of coal data can also be calculated based on a weighted sum of two or three of the following scores: time freshness score, integrity score, measurement accuracy score, and data consistency score.

[0041] In other aspects of this invention, the security level includes Level 1 sensitive, Level 2 sensitive, and Level 3 non-sensitive; the partitioning and isolation of the multimodal data based on the hierarchical label includes: determining that the coal data is located in a secure training zone when the security level of the coal data is Level 1 sensitive and the quality score of the coal data is greater than or equal to a first preset threshold; determining that the coal data is located in a restricted training zone when the security level of the coal data is Level 2 sensitive and the quality score of the coal data is greater than or equal to a second preset threshold; determining that the coal data is located in an open training zone when the security level of the coal data is Level 3 non-sensitive and the quality score of the coal data is greater than or equal to a third preset threshold; wherein, the first preset threshold is greater than the second preset threshold, the second preset threshold is greater than the third preset threshold; the data access difficulty in the secure training zone is higher than the data access difficulty in the restricted training zone, and the data access difficulty in the restricted training zone is higher than the data access difficulty in the open training zone.

[0042] Please refer to Figure 3In one embodiment, the electronic device comprehensively utilizes the security level and quality score of coal data to achieve partitioning and isolation of multimodal data. Specifically, the electronic device constructs a hierarchically isolated AI training data pool. The electronic device establishes three physically isolated data storage areas: a high-security training area, a restricted training area, and an open training area. The high-security training area stores only data with a security level of Level 1 sensitivity and a quality score higher than or equal to 90, with access restricted to encrypted training servers deployed in the enterprise's intranet isolation zone; the restricted training area stores data with a security level of Level 2 sensitivity and a quality score higher than or equal to 70, requiring dual authentication and operation auditing before access; the open training area stores data with a security level of Level 3 non-sensitive data and a quality score higher than or equal to 50, accessible to external partners after signing a data usage agreement. Furthermore, each data storage area is further divided into subdirectories according to business domain and task type, with subdirectory naming following the specification of "business domain_task type_time interval". For example, "01_disaster warning_June 2024" represents a subset of data from the coal mine disaster warning task under the safety production domain in June 2024. When writing data, the electronic device automatically parses the four-dimensional hierarchical tags and routes the data to the corresponding region and subdirectory. When reading data, the caller needs to declare the target task type, and the electronic device verifies whether its permissions match the data access policy. Only if the match is successful can it return a data reference handle.

[0043] In other embodiments, the electronic device may also use only the security level or quality score of the coal data to partition and isolate multimodal data. Taking security level as an example, the high-security training area stores only coal data with a security level of Level 1 sensitivity; the restricted training area stores coal data with a security level of Level 2 sensitivity; and the open training area stores coal data with a security level of Level 3 non-sensitive.

[0044] This invention proposes a four-dimensional data grading mechanism driven by a knowledge graph ontology model, which organically combines business domain, security level, quality score and timeliness window, providing a refined management strategy for data with different sensitivity and quality levels.

[0045] In other aspects of the embodiments of the present invention, the step of determining the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy includes: determining that the coal data in the safe training area is used to train the AI ​​model using a federated learning method; determining that the coal data in the restricted training area is used to train the AI ​​model using a differential privacy mechanism; and determining that the coal data in the open training area is used to train the AI ​​model using a supervised learning method.

[0046] This invention designs a model training strategy to match coal data in partitioned and isolated regions. For coal data in high-security training areas, a federated learning architecture is adopted. Forward propagation and gradient calculation are performed locally, with only encrypted gradients uploaded to the central aggregation node. The central aggregation node performs secure multi-party computation to update global model parameters. Homomorphic encryption is used to process the encrypted gradients, ensuring that no single participant can reverse-engineer the original data. For coal data in restricted training areas, a differential privacy mechanism is employed. Noise conforming to a Laplace distribution is injected into the numerical fields of the training samples, with the noise scale dynamically adjusted according to the data sensitivity level. The formula for calculating the Laplace noise scale b is: ; Where Δf is the sensitivity of the query function, that is, the maximum change in output caused by a change in a single record; ε For privacy budgets, Level 1 sensitive data corresponds to ε =0.1, corresponding to Level 2 sensitive data ε =1.0.

[0047] For data in the open training region, a standard supervised learning process is adopted, but adversarial example enhancement and domain adaptation modules are forcibly introduced. Adversarial examples are generated using the fast gradient sign method, with the perturbation amplitude limited to within 5% of the original value; the domain adaptation module uses the maximum mean difference minimization loss to align the feature distributions of the source and target domains.

[0048] This invention achieves a unified semantic framework for multi-source data through a knowledge graph ontology model, implements refined data management through a four-dimensional hierarchical mechanism, and ensures data security and model performance through a layered training strategy, ultimately constructing a safe, efficient, and reliable AI training system for the coal industry chain. In other words, this invention constructs a knowledge graph ontology model covering all elements of the coal industry chain, performing structured parsing and semantic mapping on multi-source heterogeneous data from coal mines, washing plants, coal chemical plants, railway transportation nodes, and highway logistics terminals; based on the knowledge graph ontology model, it automatically classifies, assesses sensitivity, and labels the multimodal raw data; according to the classification results, it constructs a layered and isolated AI training data pool and designs a matching model training strategy to achieve highly secure, efficient, and robust intelligent model iteration.

[0049] In other aspects of the embodiments of the present invention, after determining the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy, the method further includes: determining coal sample data based on the attribute set and attribute relationship path of the entity corresponding to the target AI model in the knowledge graph ontology model; determining a target feature construction method based on the type of the coal sample data; and constructing features of the coal sample data based on the target feature construction method to obtain target input features; the target feature construction method is one of the sliding window method, the inverse distance weighting method, and the event counting and temporal coding combination method; and training the target AI model based on the target input features and the model training strategy corresponding to the coal sample data.

[0050] After determining the model training strategy for each partition, the electronic device executes AI model training and deployment. The target AI models in this embodiment of the invention include a coal mine face disaster early warning model, a coal washing and processing yield prediction model, a gasifier thermal efficiency optimization model, a synthetic ammonia energy consumption prediction model, a railway empty car allocation optimization model, and a highway transportation route dynamic planning model. The electronic device is equipped with a dedicated feature engineering pipeline for each type of model. This pipeline automatically filters relevant data fields and constructs high-order interactive features using the target feature construction method based on the attribute set and attribute relationship paths of the entities involved in the task in the knowledge graph ontology model. For example, taking the coal mine face disaster early warning model as an example, based on the calorific value of the raw coal batches involved in the coal mine face disaster early warning model in the knowledge graph ontology model, the coal sample data corresponding to the coal mine face disaster early warning model can be determined as the calorific value of the raw coal batches. The target feature construction method in this embodiment of the invention includes entity attribute splicing, attribute relationship path aggregation, time window statistics, and spatial proximity interpolation. For time-series data, a sliding window method is used to extract the mean, variance, slope, and number of abrupt change points as derived features. The window length is set according to the task response time requirements: five minutes for coal mine disaster early warning tasks and two hours for railway dispatching tasks. For spatially distributed data, an inverse distance weighting method is used to fuse data from neighboring monitoring points, with a weight decay coefficient set to two. For discrete event data, a combination of event counting and time-series coding is used, with the number of occurrences of similar events in the past 24 hours and the interval between the most recent occurrence and the current time used as input features. The electronic device trains the target AI model based on the target input features constructed using the target feature construction method and the model training strategy corresponding to the coal sample data.

[0051] This invention improves the physical plausibility, generalization ability, and interpretability of AI models by introducing an automated feature engineering pipeline.

[0052] In other aspects of this invention, the loss function during the training of the target AI model is constructed based on a data fitting loss term and a semantic consistency regularization term. The data fitting loss term and the semantic consistency regularization term are positively correlated with the loss function; the data fitting loss term is constructed using mean squared error or cross-entropy; and the semantic consistency regularization term characterizes the degree of logical violation between the model's prediction results and the attribute relationship paths of the knowledge graph ontology model.

[0053] In this embodiment of the invention, a knowledge graph constraint loss function (hereinafter referred to as the loss function) is introduced during model training. In one embodiment, this loss function is composed of a weighted sum of a data fitting loss term and a semantic consistency regularization term. The data fitting loss term uses mean squared error or cross-entropy, depending on the task type. The semantic consistency regularization term is quantified by calculating the degree of logical violation of the model's prediction results on the attribute relationship path of the knowledge graph ontology model. For example, if the model predicts that the calorific value of a batch of coal is 5000 kcal / kg, while the historical maximum value of its coal seam is 4800 kcal / kg, then the numerical boundary constraint of the "originating from" attribute relationship is violated. The formula for calculating the semantic consistency regularization term is: L cons =max(0, -y max )+max(0,y min - ); in, For the model prediction results, y max With y min This represents the upper and lower bounds of the entity's attribute in the knowledge graph ontology model. cons This is the semantic consistency regularization term. The semantic consistency regularization term is a weighted sum of all violations. The total loss function is expressed as: L = L fit +λ·L cons ; Where λ is the weighting coefficient, set according to the task's dependence on physical rationality; the weighting coefficient for process control tasks is 0.8, and the weighting coefficient for transportation scheduling tasks is 0.3. L is the loss function. fit This is the loss term for data fitting.

[0054] The embodiments of the present invention significantly improve the physical rationality, generalization ability and interpretability of AI models by introducing a knowledge graph constraint loss function and an automated feature engineering pipeline.

[0055] In other aspects of this invention, the trained target AI model is deployed on edge computing nodes or a central cloud platform. During deployment, a hierarchical label summary of its training data and a knowledge graph reference index are attached. When the target AI model receives new input for inference, it first verifies whether the hierarchical labels of the input data are compatible with the admission policy of the training data pool. If incompatible, the inference request is rejected and a security policy conflict warning is returned. If compatible, inference is performed, and the inference result is bound to the associated knowledge graph ontology model entity, forming a traceable and interpretable decision record. For example, if a coal mine working face disaster early warning model outputs "High risk of roof collapse in the next ten minutes for the West Wing No. 3 working face," this result is automatically associated with the entity "Coal Mine Working Face_West Wing No. 3" and a new "Risk Assessment" attribute assertion is generated for subsequent use by the scheduling system.

[0056] Please refer to Figure 4 In one exemplary embodiment, the coal industry chain multimodal data classification and AI training method of this invention includes the following steps: S1, Construct a knowledge graph ontology model covering all elements of the coal industry chain; S2 performs structured parsing and semantic mapping on multimodal data (multi-source heterogeneous data); S3 automatically classifies, assesses sensitivity, and labels multimodal raw data based on a knowledge graph ontology model; S4, construct a hierarchical and isolated AI training data pool; S5, design a matching model training strategy; S6 performs AI model training and deployment.

[0057] In summary, this invention provides a knowledge graph-based method for multimodal data grading and AI training in the coal industry chain. It aims to address the problems of low training efficiency, weak generalization ability, and insufficient decision-making credibility of artificial intelligence models in the coal, coal chemical, and transportation industries, caused by heterogeneous data sources, semantic fragmentation, inconsistent data quality, and varying security sensitivities. This invention constructs a knowledge graph ontology model covering the entire coal industry chain, achieving for the first time a unified semantic description and structured association of heterogeneous data from multiple sources, including coal mines, coal chemicals, railways, and highway transportation, thus solving the fundamental problems of data silos and semantic fragmentation in traditional systems. Based on this, an ontology-driven four-dimensional data grading mechanism is proposed, organically combining business domain, security level, quality score, and timeliness window, providing refined management strategies for data with different sensitivities and quality levels. Simultaneously, a hierarchical and isolated AI training data pool and a matching differentiated training strategy are designed, ensuring the security of highly sensitive industrial data while fully utilizing the training value of open data. By introducing a knowledge graph-constrained loss function and an automated feature engineering pipeline, the physical rationality, generalization ability, and interpretability of the model are significantly improved. Ultimately, this invention realizes a paradigm shift in the application of artificial intelligence in the coal industry chain from "data availability" to "data reliability, security, and efficient utilization," providing a replicable and scalable technical framework for the intelligent upgrading of complex industrial systems.

[0058] Device Examples Please refer to Figure 5 On the other hand, embodiments of the present invention also provide a multimodal data classification and AI training device for the coal industry chain, comprising: Module 501 is used to construct a knowledge graph ontology model based on elements of the coal industry chain; the knowledge graph ontology model includes entities constructed based on elements of the coal industry chain and the attribute relationships between the entities; Mapping module 502 is used to perform semantic mapping on multimodal data of the coal industry chain based on the knowledge graph ontology model; The grading module 503 is used to perform data grading on the multimodal data using the knowledge graph ontology model to obtain a grading label for each type of coal data in the multimodal data; the grading label for the coal data includes at least the safety level and quality score of the coal data. Isolation module 504 is used to partition and isolate the multimodal data based on the hierarchical labels; The determination module 505 is used to determine the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy.

[0059] Optionally, the semantic mapping of multimodal data in the coal industry chain based on the knowledge graph ontology model includes: Based on the data attributes of each type of coal data in the multimodal data, a set of candidate entities in the knowledge graph ontology model that matches the data attributes is determined; The numerical field names of the coal data are compared with the attribute name list of each entity in the candidate entity set; Entities whose comparison results are consistent and whose units of the coal data conform to preset rules are identified as target entities; The coal data is connected to the target entity using preset attribute relationships to establish a semantic link between the coal data and the target entity.

[0060] Optionally, the step of using the knowledge graph ontology model to perform data classification on the multimodal data to obtain classification labels for each type of coal data in the multimodal data includes: The knowledge graph ontology model is used to classify each type of coal data in the multimodal data for business purposes, thereby obtaining the business domain of each type of coal data in the multimodal data. The knowledge graph ontology model is used to perform sensitivity assessment on each type of coal data in the multimodal data to obtain the security level of each type of coal data in the multimodal data. The knowledge graph ontology model is used to evaluate the data quality of each type of coal data in the multimodal data, and a quality score for each type of coal data in the multimodal data is obtained. The knowledge graph ontology model is used to set a time window for each type of coal data in the multimodal data, thereby obtaining the time window for each type of coal data in the multimodal data. The business domain, the security level, the quality score, and the timeliness window serve as the grading labels for each type of coal data.

[0061] Optionally, the step of using the knowledge graph ontology model to perform data quality assessment on each type of coal data in the multimodal data to obtain a quality score for each type of coal data in the multimodal data includes: The knowledge graph ontology model is used to determine the time freshness score, integrity score, measurement accuracy score, and data consistency score for each type of coal data in the multimodal data. Based on the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score, a quality score for each type of coal data in the multimodal data is determined; Among them, the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score are all positively correlated with the quality score.

[0062] Optionally, the security levels include Level 1 sensitive, Level 2 sensitive, and Level 3 non-sensitive; the partitioning and isolation of the multimodal data based on the hierarchical labels includes: If the security level of the coal data is Level 1 Sensitive and the quality score of the coal data is greater than or equal to a first set threshold, then the coal data is determined to be located in the safe training zone. If the security level of the coal data is Level 2 Sensitive and the quality score of the coal data is greater than or equal to the second set threshold, then the coal data is determined to be located in the restricted training area. If the security level of the coal data is Level 3 Non-sensitive and the quality score of the coal data is greater than or equal to the third set threshold, then the coal data is determined to be located in the open training area. Wherein, the first set threshold is greater than the second set threshold, the second set threshold is greater than the third set threshold; the data access difficulty of the secure training area is higher than the data access difficulty of the restricted training area, and the data access difficulty of the restricted training area is higher than the data access difficulty of the open training area.

[0063] Optionally, the step of determining the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy includes: The coal data of the designated safe training area is used to train an AI model using a federated learning method. The coal data in the restricted training area is used for AI model training using a differential privacy mechanism; The coal data in the open training area is used to train the AI ​​model using a supervised learning method.

[0064] Optionally, the device further includes: The training module is used to determine coal sample data based on the attribute set and attribute relationship path of the entity corresponding to the target AI model in the knowledge graph ontology model; determine the target feature construction method based on the type of the coal sample data; and construct features of the coal sample data based on the target feature construction method to obtain target input features; the target feature construction method is one of the sliding window method, the inverse distance weighting method, and the event counting and temporal coding combination method; and train the target AI model based on the target input features and the model training strategy corresponding to the coal sample data.

[0065] Optionally, the loss function in the training process of the target AI model is constructed based on the data fitting loss term and the semantic consistency regularization term; The data fitting loss term and the semantic consistency regularization term are positively correlated with the loss function, respectively; the data fitting loss term is constructed using mean squared error or cross-entropy; the semantic consistency regularization term characterizes the degree of logical violation between the model prediction results and the attribute relationship paths of the knowledge graph ontology model.

[0066] The coal industry chain multimodal data grading and AI training device includes a processor and a memory. The aforementioned construction module 501, mapping module 502, grading module 503, isolation module 504, and determination module 505 are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. The processor contains a kernel, which retrieves the corresponding program units from the memory. One or more kernels can be configured. The memory may include non-permanent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. The memory includes at least one memory chip.

[0067] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a method for grading multimodal data in the coal industry chain and training AI. This method includes: constructing a knowledge graph ontology model based on elements of the coal industry chain; the knowledge graph ontology model includes entities constructed based on elements of the coal industry chain and the attribute relationships between the entities; performing semantic mapping on the multimodal data of the coal industry chain based on the knowledge graph ontology model; grading the multimodal data using the knowledge graph ontology model to obtain a grading label for each type of coal data in the multimodal data; the grading label for the coal data includes at least the security level and quality score of the coal data; partitioning and isolating the multimodal data based on the grading labels; and determining the model training strategy corresponding to the coal data in each partition after partitioning and isolating based on a set partition training strategy.

[0068] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0069] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a machine-readable storage medium. When the computer program is executed by a processor, the computer can execute a method for classifying and training multimodal data in the coal industry chain using AI. This method includes: constructing a knowledge graph ontology model based on elements of the coal industry chain; the knowledge graph ontology model includes entities constructed based on elements of the coal industry chain and attribute relationships between the entities; performing semantic mapping on multimodal data of the coal industry chain based on the knowledge graph ontology model; classifying the multimodal data using the knowledge graph ontology model to obtain a classification label for each type of coal data in the multimodal data; the classification label for the coal data includes at least the security level and quality score of the coal data; partitioning and isolating the multimodal data based on the classification labels; and determining the model training strategy corresponding to the coal data in each partition after partitioning and isolating based on a set partition training strategy.

[0070] In another aspect, the present invention also provides a machine-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for classifying and training multimodal data in the coal industry chain using AI. This method includes: constructing a knowledge graph ontology model based on elements of the coal industry chain; the knowledge graph ontology model including entities constructed based on elements of the coal industry chain and attribute relationships between the entities; performing semantic mapping on the multimodal data of the coal industry chain based on the knowledge graph ontology model; classifying the multimodal data using the knowledge graph ontology model to obtain a classification label for each type of coal data in the multimodal data; the classification label for the coal data includes at least the security level and quality score of the coal data; partitioning and isolating the multimodal data based on the classification labels; and determining a model training strategy corresponding to the coal data in each partition after partitioning and isolating based on a set partitioning training strategy.

[0071] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0072] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0073] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multimodal data classification and AI training in the coal industry chain, characterized in that, include: A knowledge graph ontology model is constructed based on the elements of the coal industry chain; the knowledge graph ontology model includes entities constructed based on the elements of the coal industry chain and the attribute relationships between the entities; Semantic mapping is performed on multimodal data of the coal industry chain based on the knowledge graph ontology model. The knowledge graph ontology model is used to classify the multimodal data to obtain a classification label for each type of coal data in the multimodal data; the classification label of the coal data includes at least the safety level and quality score of the coal data; The multimodal data is partitioned and isolated based on the hierarchical labels; The model training strategy corresponding to the coal data of each partition is determined based on the set partition training strategy. The knowledge graph ontology model includes a set of sensitivity determination rules preset by human experts; the security levels include Level 1 sensitive, Level 2 sensitive, and Level 3 non-sensitive. The partitioning and isolation of the multimodal data based on the hierarchical labels includes: If the security level of the coal data is Level 1 Sensitive and the quality score of the coal data is greater than or equal to a first set threshold, then the coal data is determined to be located in the safe training zone. If the security level of the coal data is Level 2 Sensitive and the quality score of the coal data is greater than or equal to the second set threshold, then the coal data is determined to be located in the restricted training area. If the security level of the coal data is Level 3 Non-sensitive and the quality score of the coal data is greater than or equal to the third set threshold, then the coal data is determined to be located in the open training area. Wherein, the first set threshold is greater than the second set threshold, and the second set threshold is greater than the third set threshold; the data access difficulty of the secure training area is greater than the data access difficulty of the restricted training area, and the data access difficulty of the restricted training area is greater than the data access difficulty of the open training area; The semantic mapping of multimodal data in the coal industry chain based on the knowledge graph ontology model includes: Based on the data attributes of each type of coal data in the multimodal data, a set of candidate entities in the knowledge graph ontology model that matches the data attributes is determined; the data attributes include data source identifier, timestamp, spatial coordinates, equipment number, numerical fields, and units; The numerical field names of the coal data are compared with the attribute name list of each entity in the candidate entity set; Entities whose comparison results are consistent and whose units of the coal data conform to preset rules are identified as target entities; The coal data is connected to the target entity using preset attribute relationships to establish a semantic link between the coal data and the target entity; The model training strategy for determining the coal data corresponding to each partition after partition isolation based on the set partition training strategy includes: The coal data of the designated safe training area is used to train an AI model using a federated learning method. The coal data in the restricted training area is used for AI model training using a differential privacy mechanism; The coal data in the open training area is used to train the AI ​​model using a supervised learning method.

2. The method for multimodal data classification and AI training in the coal industry chain according to claim 1, characterized in that, The step of using the knowledge graph ontology model to classify the multimodal data and obtain classification labels for each type of coal data in the multimodal data includes: The knowledge graph ontology model is used to classify each type of coal data in the multimodal data for business purposes, thereby obtaining the business domain of each type of coal data in the multimodal data. The knowledge graph ontology model is used to perform sensitivity assessment on each type of coal data in the multimodal data to obtain the security level of each type of coal data in the multimodal data. The knowledge graph ontology model is used to evaluate the data quality of each type of coal data in the multimodal data, and a quality score for each type of coal data in the multimodal data is obtained. The knowledge graph ontology model is used to set a time window for each type of coal data in the multimodal data, thereby obtaining the time window for each type of coal data in the multimodal data. The business domain, the security level, the quality score, and the timeliness window serve as the grading labels for each type of coal data.

3. The method for multimodal data classification and AI training in the coal industry chain according to claim 2, characterized in that, The process of using the knowledge graph ontology model to evaluate the data quality of each type of coal data in the multimodal data, and obtaining a quality score for each type of coal data in the multimodal data, includes: The knowledge graph ontology model is used to determine the time freshness score, integrity score, measurement accuracy score, and data consistency score for each type of coal data in the multimodal data. Based on the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score, a quality score for each type of coal data in the multimodal data is determined; Among them, the time freshness score, the integrity score, the measurement accuracy score, and the data consistency score are all positively correlated with the quality score.

4. The method for multimodal data classification and AI training in the coal industry chain according to claim 1, characterized in that, After determining the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy, the method further includes: In the knowledge graph ontology model, coal sample data is determined based on the attribute set and attribute relationship path of the entity corresponding to the target AI model; The method for constructing target features is determined based on the type of the coal sample data, and the target feature construction method is used to construct features on the coal sample data to obtain target input features; the target feature construction method is one of the following: sliding window method, inverse distance weighting method, and event counting and time-series coding combination method; The target AI model is trained based on the target input features and the model training strategy corresponding to the coal sample data.

5. The method for multimodal data classification and AI training in the coal industry chain according to claim 4, characterized in that, The loss function during the training process of the target AI model is constructed based on the data fitting loss term and the semantic consistency regularization term; The data fitting loss term and the semantic consistency regularization term are positively correlated with the loss function, respectively; the data fitting loss term is constructed using mean squared error or cross-entropy; the semantic consistency regularization term characterizes the degree of logical violation between the model prediction results and the attribute relationship paths of the knowledge graph ontology model.

6. A multimodal data classification and AI training device for the coal industry chain, characterized in that, include: A construction module is used to build a knowledge graph ontology model based on the elements of the coal industry chain; the knowledge graph ontology model includes entities built based on the elements of the coal industry chain and the attribute relationships between the entities; The mapping module is used to perform semantic mapping on multimodal data of the coal industry chain based on the knowledge graph ontology model. The grading module is used to perform data grading on the multimodal data using the knowledge graph ontology model, and to obtain a grading label for each type of coal data in the multimodal data; the grading label for the coal data includes at least the safety level and quality score of the coal data. An isolation module is used to partition and isolate the multimodal data based on the hierarchical labels; The determination module is used to determine the model training strategy corresponding to the coal data of each partition after partition isolation based on the set partition training strategy. The knowledge graph ontology model includes a set of sensitivity determination rules preset by human experts; the security levels include Level 1 sensitive, Level 2 sensitive, and Level 3 non-sensitive. The partitioning and isolation of the multimodal data based on the hierarchical labels includes: If the security level of the coal data is Level 1 Sensitive and the quality score of the coal data is greater than or equal to a first set threshold, then the coal data is determined to be located in the safe training zone. If the security level of the coal data is Level 2 Sensitive and the quality score of the coal data is greater than or equal to the second set threshold, then the coal data is determined to be located in the restricted training area. If the security level of the coal data is Level 3 Non-sensitive and the quality score of the coal data is greater than or equal to the third set threshold, then the coal data is determined to be located in the open training area. Wherein, the first set threshold is greater than the second set threshold, and the second set threshold is greater than the third set threshold; the data access difficulty of the secure training area is greater than the data access difficulty of the restricted training area, and the data access difficulty of the restricted training area is greater than the data access difficulty of the open training area; The semantic mapping of multimodal data in the coal industry chain based on the knowledge graph ontology model includes: Based on the data attributes of each type of coal data in the multimodal data, a set of candidate entities in the knowledge graph ontology model that matches the data attributes is determined; the data attributes include data source identifier, timestamp, spatial coordinates, equipment number, numerical fields, and units; The numerical field names of the coal data are compared with the attribute name list of each entity in the candidate entity set; Entities whose comparison results are consistent and whose units of the coal data conform to preset rules are identified as target entities; The coal data is connected to the target entity using preset attribute relationships to establish a semantic link between the coal data and the target entity; The model training strategy for determining the coal data corresponding to each partition after partition isolation based on the set partition training strategy includes: The coal data of the designated safe training area is used to train an AI model using a federated learning method. The coal data in the restricted training area is used for AI model training using a differential privacy mechanism; The coal data in the open training area is used to train the AI ​​model using a supervised learning method.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the coal industry chain multimodal data classification and AI training method as described in any one of claims 1 to 5.

Images