CN-121526073-B - Attention sensing and dynamic memory method for multi-source time sequence data of coal mining operation

CN121526073BCN 121526073 BCN121526073 BCN 121526073BCN-121526073-B

Abstract

The invention relates to a method for awareness and dynamic memory of multi-source time sequence data of coal mining operation, and belongs to the technical field of coal mine intellectualization. The method comprises the steps of carrying out three-dimensional feature dynamic extraction and scene information attention fusion on multi-source time sequence data of a coal face to generate a scene multi-dimensional feature data set, storing the multi-source time sequence data of the coal face and the scene multi-dimensional feature data set in a multi-Agent driven coal scene multi-dynamic information mixed storage mode to form a data base, and mapping out related coal scenes according to equipment, process and environment information based on the data base after a coal scene task instruction is acquired, wherein the content indicated by the coal scene task instruction is completed by various types of agents in a matched mode. The invention introduces an Agent architecture and an attention mechanism, combines a dynamic memory management strategy, and can provide core technical support for coal mining running state real-time monitoring, fault intelligent diagnosis and production optimization decision.

Inventors

FU XIANG
ZHANG ZHIXING
QIN YIFAN
YAN MING
LI HAOJIE
XING KEKE
JIA YIFAN

Assignees

太原理工大学

Dates

Publication Date: 20260512
Application Date: 20251119

Claims (6)

1. The method for awareness and dynamic memory of the multi-source time sequence data of coal mining operation is characterized by comprising the following steps: s1, carrying out three-dimensional feature dynamic extraction and scene information attention fusion on multi-source time sequence data of a coal face to generate a scene multi-dimensional feature data set, wherein the three-dimensional features comprise equipment pose features, process features and environment features; S2, storing multi-source time sequence data and a scene multi-dimensional characteristic data set of a coal face in a multi-agent driven coal mining scene multi-element dynamic information mixed storage mode to form a data base; The step S2 comprises the following steps: S21, a data acquisition Agent acquires multi-source time sequence data and a scene multidimensional characteristic data set of a coal face into a mixed warehousing system, and performs data flow on the acquired data in a kafka data flow mode; S22, calculating statistical characteristics of various data in a window in a multi-source time sequence data and a scene multidimensional characteristic data set by a sliding window mechanism, judging whether the statistical characteristics of various data are in a corresponding threshold range, if the statistical characteristics of any type of data are not in the corresponding threshold range, marking any type of data as abnormal data, filling any type of data through a time weighted average algorithm, and carrying out alignment processing on various types of data in a time dimension by using a time window and a dynamic time warping algorithm by the data cleaning Agent after the data filling is finished; S23, carrying out classification coding on various data subjected to time alignment processing by the data cleaning Agent according to a predefined coal mine data classification and coding specification to form a high-quality standard data set; S24, classifying each data item in the high-quality standard data set according to the data format and the data volume by the data classification Agent, directly warehousing and storing any target data item in the high-quality standard data set if the target data item is vector data, judging whether the number of the data item of the target data item in the first N sliding windows is larger than a preset threshold value N through a sliding window algorithm if the target data item is not vector data, determining the target data item as massive time sequence data if the number of the data item in N multiplied by 1% sliding windows is larger than N, otherwise, determining the target data item as a small amount of time sequence data; s25, the data storage Agent stores mass time sequence data in a distributed column type database, stores a small amount of time sequence data in the time sequence database and stores vector data in a vector database; S3, after the coal mining scene task instruction is acquired, mapping out a coal mining scene related to the coal mining scene based on the data base in combination with equipment, process and environment information, and completing the content indicated by the coal mining scene task instruction by the various types of agents in a matching way according to roles and functions of the various types of agents which are arranged in advance on the coal mining working face; The step S3 comprises the following steps: S31, after the input of the coal mining scene instruction t is detected, mapping the coal mining scene instruction t to a knowledge graph set W; S32, if the coal mining scene instruction t is successfully mapped to the knowledge graph set W, traversing the associated equipment entity, the process entity, the environment entity and the corresponding associated weight in the knowledge graph set W according to the coal mining scene task instruction t to form an entity weight set Q; S33, acquiring a real-time data set D1 of an associated entity in the entity weight set Q from the data base by the reflective program Agent through the associated weight priority sequence, and acquiring a historical data set D2 of the previous n coal cutting cycles of the associated entity from the data base by the inference model Agent through the associated weight priority sequence; S34, constructing a coal mining scene S of four dimensions of a task, equipment, a process and an environment according to a real-time data set D1 and a historical data set D2 of an associated entity; S35, carrying out cluster analysis on the coal mining scene S through a K-means clustering algorithm, combining a cosine similarity algorithm and expert experience, adjusting the weight of an entity data set associated with the coal mining scene S in real time according to the actual coal mining scene, updating an entity weight set Q, and updating a knowledge graph set W; S36, judging task types successfully mapped in the knowledge graph set W to the coal mining scene instruction t after updating is completed, if the coal mining scene instruction t is a generating task, the generating task Agent, the planning organization Agent, the planning arbitration Agent, the reasoning model Agent and the reflection type program Agent cooperatively complete the generating task according to the coal mining scene S, if the coal mining scene instruction t is a learning task, the learning reality ring Agent, the learning simulation ring Agent and the learning double-ring Agent cooperatively complete the learning task according to the coal mining scene S, so that model updating of the coal mining scene is realized, and if the coal mining scene instruction t is a investigation type instruction, the generating investigation Agent cooperatively completes the investigation task according to the coal mining scene S.
2. The method for attention sensing and dynamic memorization of coal mining operation multisource time series data according to claim 1, wherein S1 comprises: S11, taking the center of a machine head of the scraper conveyor in a primary mining state of the coal face as an origin, and taking the advancing direction of the coal face, the trend vertical to the coal face and the vertical direction as an X axis, a Y axis and a Z axis respectively to establish a pose coordinate system; S12, acquiring position data and attitude angle data of key equipment of a coal face, mapping the position data of the key equipment into a pose coordinate system, and combining the attitude angle data to form pose vectors of the key equipment; s13, extracting time evolution features of pose vectors of key equipment in a time dimension, and integrating the pose vectors and the time evolution features of all the key equipment to form a pose time-space evolution feature vector E (t) of the key equipment; s14, constructing a coal face propulsion condition feature vector Vt and a coal face production process parameter feature vector Vs, and integrating the feature vectors to obtain a dynamic causal feature vector P (t); S15, acquiring multiple types of environment parameters, extracting time domain features, trend features and mutation features of each environment parameter, and integrating preset weights of the time domain features, the trend features and the mutation features of each environment parameter to obtain an environment feature vector C (t); s16, fusing the pose space-time evolution feature vector E (t), the dynamic causal feature vector P (t) and the environment feature vector C (t) of the key equipment by adopting a weighting and multi-head attention mechanism to obtain a scene multidimensional feature data set.
3. The method for attention sensing and dynamic memorization of coal mining operation multisource time series data according to claim 2, wherein S14 comprises: S141, calculating the association strength between each element in the coal face propulsion condition feature vector Vt and the coal face production process parameter feature vector Vs through a similarity algorithm, and taking the association strength as the weight of the two elements; S142, comparing the weight with the association strength preset by expert experience, calculating an error c, if the error c is smaller than a preset error threshold c1, determining that the association strength calculated by a similarity algorithm is consistent with the association strength preset by expert experience, and directly integrating the weight, the elements in the coal face propulsion situation feature vector Vt and the coal face production process parameter feature vector Vs to obtain a dynamic causal feature vector P (t); S143, if the error c is not smaller than the preset error threshold c1, that is, it is determined that the correlation strength calculated by the similarity algorithm does not accord with the correlation strength preset by the expert experience, the calculated weight is adjusted based on the weight preset by the expert experience, and then the coal face propulsion situation feature vector Vt and the coal face production process parameter feature vector Vs are integrated to obtain the dynamic causal feature vector P (t).
4. A method of attention sensing and dynamic memorization of coal mining operation multisource time series data according to claim 2 or 3, wherein S15 comprises: Sequentially extracting time domain features, trend features and mutation features of each type of environmental parameters through a sliding window, giving weight of 0.4 to the mutation features of each type of environmental parameters, giving weight of 0.3 to the trend features and the time domain features of each type of environmental parameters respectively, and integrating the time domain features, the trend features, the mutation features and the weights of all types of environmental parameters to obtain an environmental feature vector C (t).
5. The method for attention sensing and dynamic memorization of coal mining operation multisource time series data according to claim 1, further comprising, after S25: S26, the data blood edge Agent carries out data blood edge and metadata carding on a data acquisition process of the data acquisition Agent, a data cleaning process of the data cleaning Agent, a data classification process of the data classification Agent and a data warehousing storage process of the data storage Agent so as to comb data acquisition points, data point sources, data units, data types, normal value ranges, processing processes and storage positions.
6. The method for attention sensing and dynamic memorization of coal mining operation multisource time series data according to claim 1, further comprising, before S31: In the coal mining production process of the coal mining working face, defining a task entity T, a device entity E, a process entity P and an environment entity Env through expert experience, defining the relation among the entities, calculating the association strength I among the entities through a cosine similarity algorithm after the expert experience definition is completed, and constructing a knowledge graph set W= { T, E, P, env, I } with relation weight in four dimensions through the relation among the entities and the association strength I among the entities.

Description

Attention sensing and dynamic memory method for multi-source time sequence data of coal mining operation Technical Field The invention relates to the technical field of coal mine intellectualization, in particular to a method for sensing attention and dynamically memorizing multi-source time sequence data of coal mining operation. Background Under the dual driving of the green low-carbon development concept and the intelligent transformation of the coal industry, coal mining is accelerating iteration from a traditional 'experience driven' mode to a 'data driven' mode. The coal mining operation is used as a core link of coal production, the operation process of the coal mining operation relates to tens of key equipment such as a coal mining machine, a scraper conveyor, a hydraulic support, a heading machine and the like, and multidimensional monitoring indexes such as gas concentration, roof pressure, coal flow speed, environment temperature and humidity and the like, and the data are continuously generated at a frequency of millisecond level to minute level, so that a multi-source time sequence data set with huge volume, complex dimension and close association is formed. The data not only reflects the running state of the coal mining system, prejudges equipment faults and optimizes the core basis of production efficiency, but also realizes the basic support of unmanned mining and intelligent decision making, and the effective processing and deep utilization of the data become key marks for measuring the intelligent level of the coal mine. However, the current multi-source time sequence data processing technology for coal mining operation still faces three major core bottlenecks, and severely restricts the release of data value: First, data perception is "under generalization". The traditional data perception method mostly adopts a fixed threshold value or a single characteristic extraction mode, and is difficult to adapt to the dynamic complexity of a coal mining scene. The prior art cannot dynamically focus on the dimension of key data, and the problem that effective information is submerged in redundant data or key abnormal signals are interfered by noise often occurs, so that the accuracy and instantaneity of data perception are difficult to meet the production requirement. Second, the "synergy" of the data association is lost. Strong coupling correlation exists among multi-source time sequence data of the coal mining system (such as direct correlation of hydraulic support working resistance and roof pressure and high linkage of scraper conveyor load and coal flow speed), but the prior art mostly adopts a single-source data independent processing mode, and lacks a cross-equipment and cross-dimension data collaborative perception mechanism. Third, the "dynamic nature" of data memory is inadequate. The coal mining operation data has the characteristics of strong time sequence dependence and dynamic change of importance. The existing data storage and memorization technology mostly adopts an 'equalization storage' or 'static priority' mode, and the priority and the life cycle of data memorization cannot be dynamically adjusted according to the running state of the coal mining system, so that key data storage resources are insufficient, redundant data occupy a large amount of storage cost, and meanwhile, the efficiency of subsequent data retrieval and analysis is reduced. Under the background, there is a need to develop a technology capable of realizing accurate sensing, collaborative association and dynamic memory of multi-source time series data of coal mining operation. Disclosure of Invention In order to solve the technical problems, the invention provides a method for awareness and dynamic memory of multi-source time sequence data of coal mining operation. The technical scheme of the invention is as follows: A method for awareness and dynamic memory of multi-source time sequence data of coal mining operation comprises the following steps: s1, carrying out three-dimensional feature dynamic extraction and scene information attention fusion on multi-source time sequence data of a coal face to generate a scene multi-dimensional feature data set, wherein the three-dimensional features comprise equipment pose features, process features and environment features; S2, storing multi-source time sequence data and a scene multi-dimensional characteristic data set of a coal face in a multi-agent driven coal mining scene multi-element dynamic information mixed storage mode to form a data base; And S3, after the coal mining scene task instruction is acquired, mapping out the related coal mining scene based on the data base by combining equipment, process and environment information, and completing the content indicated by the coal mining scene task instruction by the various types of agents in a matching way according to roles and functions of the various types of agents which are arranged in advance on the coal mining