Coal mining operation multi-source time series data attention perception and dynamic memory method

Abstract

The present application relates to a kind of coal mining operation multi-source time series data attention perception and dynamic memory method, belong to coal mine intelligent technology field.It includes: the three-dimensional feature dynamic extraction and scene information attention fusion of the multi-source time series data of coal mining face, generate scene multi-dimensional feature dataset;Multi-agent driven coal mining scene multi-dimensional dynamic information hybrid storage mode is used to store the multi-source time series data of coal mining face and scene multi-dimensional feature dataset, form data base;When obtaining coal mining scene task instruction, based on data base, in combination with equipment, process, environmental information, the related coal mining scene of it is mapped out, and the content indicated by the coal mining scene task instruction is completed by the cooperation of various types of Agent.The present application introduces Agent intelligent agent architecture and attention mechanism, in combination with dynamic memory management strategy, can provide core technical support for coal mining operation state real-time monitoring, fault intelligent diagnosis and production optimization decision.

Description

Technical Field

[0001] This invention relates to the field of intelligent coal mining technology, and in particular to a method for attention perception and dynamic memory of multi-source time-series data in coal mining operations. Background Technology

[0002] Driven by both the green and low-carbon development concept and the intelligent transformation of the coal industry, coal mining is rapidly shifting from a traditional "experience-driven" model to a "data-driven" model. As the core link in coal production, coal mining operations involve dozens of key equipment types, such as coal mining machines, scraper conveyors, hydraulic supports, and tunneling machines, as well as multi-dimensional monitoring indicators like gas concentration, roof pressure, coal flow velocity, and ambient temperature and humidity. This data is continuously generated at millisecond to minute-level frequencies, forming a massive, complex, and closely interconnected multi-source time-series dataset. This type of data is not only the core basis for reflecting the operating status of the coal mining system, predicting equipment failures, and optimizing production efficiency, but also the fundamental support for achieving "unmanned mining" and "intelligent decision-making." Its effective processing and in-depth utilization have become key indicators for measuring the level of intelligence in coal mines. However, current multi-source time-series data processing technology for coal mining operations still faces three major bottlenecks, severely restricting the release of data value:

[0003] First, data sensing suffers from insufficient generalization. Traditional data sensing methods often employ fixed thresholds or single feature extraction patterns, making it difficult to adapt to the dynamic complexity of coal mining scenarios. Existing technologies cannot dynamically focus on key data dimensions, frequently resulting in problems such as "effective information being overwhelmed by redundant data" or "key abnormal signals being interfered with by noise," making it difficult to meet production needs in terms of accuracy and real-time performance of data sensing.

[0004] Secondly, there is a lack of "coordination" in data association. There are strong coupling relationships between the multi-source time-series data of the coal mining system (such as the working resistance of the hydraulic support being directly related to the roof pressure, and the load of the scraper conveyor being highly linked to the coal flow velocity), but existing technologies mostly adopt the "single-source data independent processing" mode, lacking a cross-equipment and cross-dimensional data collaborative perception mechanism.

[0005] Third, the data storage lacks dynamism. Coal mining operation data is characterized by strong time-series dependence and dynamically changing importance. Existing data storage and storage technologies mostly adopt "equalized storage" or "static priority" models, which cannot dynamically adjust the priority and lifecycle of data storage according to the operating status of the coal mining system. This results in insufficient key data storage resources, redundant data consuming a large amount of storage costs, and also reduces the efficiency of subsequent data retrieval and analysis.

[0006] Against this backdrop, there is an urgent need to develop a technology that can achieve accurate perception, collaborative correlation, and dynamic memory of multi-source time-series data from coal mining operations. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method for attention perception and dynamic memory of multi-source time-series data from coal mining operations. The technical solution of this invention is as follows:

[0008] A method for attention perception and dynamic memory of multi-source time-series data in coal mining operations, comprising:

[0009] S1, dynamically extract three-dimensional features and fuse scene information attention into multi-source time-series data of the coal mining face to generate a scene-based multi-dimensional feature dataset, wherein the three-dimensional features include equipment pose features, process features and environmental features;

[0010] S2 uses a multi-agent-driven hybrid storage method to store multi-source time-series data and scenario-based multi-dimensional feature datasets of the coal mining face, forming a data foundation.

[0011] S3, after obtaining the coal mining scenario task instruction, based on the data base, combined with equipment, process and environmental information, maps out the relevant coal mining scenario, and according to the roles and functions of various types of agents pre-deployed at the coal mining face, the various types of agents cooperate to complete the content indicated by the coal mining scenario task instruction.

[0012] Preferably, S1 includes:

[0013] S11. A pose coordinate system is established with the center of the scraper conveyor head in the initial mining state of the coal face as the origin, and the X-axis, Y-axis and Z-axis as the advancing direction of the coal face, the direction perpendicular to the coal face and the vertical direction, respectively.

[0014] S12: Obtain the position data and attitude angle data of key equipment in the coal mining face, and after mapping the position data of the key equipment to the pose coordinate system, combine the attitude angle data to form the pose vector of the key equipment.

[0015] S13, extract the temporal evolution features of the pose vectors of key equipment in the time dimension, and integrate the pose vectors and temporal evolution features of all key equipment to form the spatiotemporal evolution feature vector E(t) of the pose of key equipment.

[0016] S14, construct the feature vector Vt of the coal mining face advance and the feature vector Vs of the coal mining face production process parameters, and integrate them to obtain the dynamic causal feature vector P(t);

[0017] S15: Obtain various types of environmental parameters, extract the time-domain features, trend features, and mutation features of each environmental parameter, and integrate them with the preset weights of the time-domain features, trend features, and mutation features of each environmental parameter to obtain the environmental feature vector C(t).

[0018] S16 uses a weighted and multi-head attention mechanism to fuse the spatiotemporal evolution feature vector E(t) of key equipment pose, the dynamic causal feature vector P(t) and the environmental feature vector C(t) to obtain a scenario-based multidimensional feature dataset.

[0019] Preferably, S14 includes:

[0020] S141, calculate the correlation strength between each element in the feature vector Vt of the coal mining face advancement and the feature vector Vs of the coal mining face production process parameters using a similarity algorithm, and use the correlation strength as the weight of the two.

[0021] S142, compare the weight with the association strength preset by expert experience, calculate the error c. If the error c is less than the preset error threshold c1, it is determined that the association strength calculated by the similarity algorithm is consistent with the association strength preset by expert experience. Then, directly integrate the elements in the weight, the feature vector Vt of the coal mining face advancement and the feature vector Vs of the coal mining face production process parameters to obtain the dynamic causal feature vector P(t).

[0022] S143, if the error c is not less than the preset error threshold c1, that is, if the correlation strength calculated by the similarity algorithm does not match the correlation strength preset by the expert experience, then the weight calculated is adjusted according to the weight preset by the expert experience. Then, the weight after integrating the feature vector Vt of the coal mining face advance and the feature vector Vs of the coal mining face production process parameters is used to obtain the dynamic causal feature vector P(t).

[0023] Preferably, S15 includes:

[0024] The time-domain features, trend features, and mutation features of each type of environmental parameter are extracted sequentially using a sliding window. The mutation feature of each environmental parameter is assigned a weight of 0.4, and the trend feature and time-domain feature of each environmental parameter are assigned a weight of 0.3 respectively. The time-domain features, trend features, mutation features of all types of environmental parameters and their weights are integrated to obtain the environmental feature vector C(t).

[0025] Preferably, S2 includes:

[0026] S21, the data acquisition agent collects multi-source time-series data and scenario-based multi-dimensional feature datasets from the coal mining face into the hybrid storage system, and transfers the collected data in the form of a Kafka data stream;

[0027] S22, the data cleaning agent uses a sliding window mechanism to calculate the statistical characteristics of various types of data in the multi-source time-series data and the scenario-based multi-dimensional feature dataset within the window, and determines whether the statistical characteristics of each type of data are within the corresponding threshold range. If the statistical characteristics of any type of data are not within the corresponding threshold range, then the data of that type is marked as abnormal data, and the data of that type is filled using a time-weighted average algorithm. After the data is filled, the data cleaning agent uses a time window plus a dynamic time warping algorithm to align the various types of data in the time dimension. If the statistical characteristics of all types of data are within the corresponding threshold range, then the data of that type is determined to be normal data, and the data of that type is directly aligned in the time dimension.

[0028] S23, The data cleaning agent classifies and encodes various types of data after time alignment according to predefined coal mine data classification and coding specifications, forming a high-quality, standardized dataset.

[0029] S24, the data classification agent classifies each data entry in the high-quality standardized dataset according to the data format and data volume; for any target data entry in the high-quality standardized dataset, if the target data entry is vector data, it is directly stored in the database; if the target data entry is not vector data, the sliding window algorithm is used to determine whether the number of data entries in the target data entry within the first n sliding windows is greater than the preset threshold N; if the number of data entries within n×1% sliding windows is greater than N, the target data entry is determined to be massive time series data; otherwise, the target data entry is determined to be a small amount of time series data.

[0030] S25, the data storage agent stores massive amounts of time-series data in a distributed columnar database, stores a small amount of time-series data in a time-series database, and stores vector data in a vector database.

[0031] Preferably, after S25, the method further includes:

[0032] S26, the data lineage agent sorts out the data lineage and metadata of the data collection agent's data collection process, the data cleaning agent's data cleaning process, the data classification agent's data classification process, and the data storage agent's data entry and storage process, in order to sort out the data collection points, data point sources, data units, data types, normal value ranges, processing procedures, and storage locations.

[0033] Preferably, S3 includes:

[0034] S31, When the coal mining scenario command t is detected, the coal mining scenario command t is mapped to the knowledge graph set W;

[0035] S32, if the coal mining scenario instruction t is successfully mapped to the knowledge graph set W, then according to the coal mining scenario task instruction t, the associated equipment entities, process entities, environmental entities and corresponding associated weights are traversed in the knowledge graph set W to form an entity weight set Q;

[0036] S33, the reflective program Agent obtains the real-time dataset D1 of the related entities in the entity weight set Q from the data base through the association weight priority order, while the inference model Agent obtains the historical dataset D2 of the first n coal cutting cycles of the related entities from the data base through the association weight priority order.

[0037] S34, construct a coal mining scenario S with four dimensions: task, equipment, process, and environment, based on the real-time dataset D1 and historical dataset D2 of the related entities;

[0038] S35 uses the K-means clustering algorithm to perform clustering analysis on the coal mining scenario S, and combines the cosine similarity algorithm and expert experience to adjust the weights of the entity datasets associated with the actual coal mining scenario S in real time, and update the entity weight set Q, while also updating the knowledge graph set W.

[0039] S36. After the update is completed, determine the task type of the coal mining scenario instruction t that is successfully mapped in the knowledge graph set W. If the coal mining scenario instruction t is a generative task, then the generative task agent, planning organization agent, planning arbitration agent, reasoning model agent, and reflection program agent will collaboratively complete the generative task based on the coal mining scenario S. If the coal mining scenario instruction t is a learning task, then the learning reality loop agent, learning simulation loop agent, and learning dual loop agent will collaboratively complete the learning task based on the coal mining scenario S, thereby updating the coal mining scenario model. If the coal mining scenario instruction t is an investigation instruction, then the generative investigation agent will collaboratively complete the investigation task based on the coal mining scenario S.

[0040] Preferably, before step S31, the method further includes:

[0041] During the coal mining process at the coal face, task entity T, equipment entity E, process entity P, and environmental entity Env are defined based on expert experience, and the relationships between entities are defined. After the expert experience definition is completed, the cosine similarity algorithm is used to calculate the association strength I between each entity. Based on the relationships between entities and the association strength I between entities, a four-dimensional knowledge graph set W={T、E、P、Env、I} with relation weights is constructed.

[0042] All of the above-mentioned optional technical solutions can be combined arbitrarily, and the present invention will not provide a detailed description of the structure after each combination.

[0043] By means of the above solution, the beneficial effects of the present invention are as follows:

[0044] By dynamically extracting three-dimensional features (equipment pose features, process features, and environmental features) from multi-source time-series data of coal mining faces and fusing them with scene information attention, a scene-based multi-dimensional feature dataset is generated. This not only fully considers the strong coupling relationship between multi-source time-series data of the coal mining system and establishes a cross-equipment, cross-dimensional data collaborative perception mechanism, but also enables the obtained scene-based multi-dimensional feature dataset to adapt to the dynamic complexity of the coal mining scene through multi-modal feature fusion.

[0045] By employing a multi-agent-driven hybrid storage method for multi-source dynamic information in coal mining scenarios, multi-source time-series data and scenario-based multi-dimensional feature datasets from coal mining faces are stored, forming a data foundation. This allows different types of data to be stored in different ways, enabling appropriate storage based on data type, priority, and lifecycle. This avoids problems such as insufficient critical data storage resources and redundant data consuming large amounts of storage costs, which reduces the efficiency of subsequent data retrieval and analysis.

[0046] Once the coal mining scenario task instructions are obtained, the relevant coal mining scenario is mapped based on the data base and combined with equipment, process, and environmental information. According to the roles and functions of various types of agents pre-deployed at the coal mining face, the various types of agents cooperate to complete the content indicated by the coal mining scenario task instructions. It can complete the content indicated by the coal mining scenario task instructions from a dimensional perspective, and achieve accurate execution of the coal mining scenario task instructions.

[0047] In summary, this invention introduces an agent-based intelligent architecture and attention mechanism, combined with a dynamic memory management strategy, which solves the shortcomings of traditional technologies in terms of data perception generalization, correlation and collaboration, and memory dynamism. It provides core technical support for real-time monitoring of coal mining operation status, intelligent fault diagnosis, and production optimization decision-making, and promotes the development of coal mining towards a higher level of intelligence and unmanned operation.

[0048] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0049] Figure 1 This is a flowchart of the coal mining operation multi-source time-series data attention perception and dynamic memory method provided in the embodiments of the present invention.

[0050] Figure 2 This is a flowchart illustrating the specific implementation of step S1 in this embodiment of the invention.

[0051] Figure 3This is a flowchart illustrating the specific implementation of step S2 in this embodiment of the invention.

[0052] Figure 4 This is a flowchart illustrating the specific implementation of step S3 in this embodiment of the invention.

[0053] Figure 5 This is a schematic diagram of the overall architecture of an embodiment of the present invention. Detailed Implementation

[0054] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0055] like Figure 1 As shown in the embodiment of the present invention, the method for attention perception and dynamic memory of multi-source time-series data in coal mining operation includes the following steps S1 to S3:

[0056] S1. Dynamically extract three-dimensional features and fuse scene information attention into multi-source time-series data of the coal mining face to generate a scene-based multi-dimensional feature dataset. The three-dimensional features include equipment pose features, process features, and environmental features.

[0057] Specifically, in this embodiment of the invention, the spatiotemporal evolution calculation of the pose of key equipment such as coal mining machines and hydraulic supports is performed from three dimensions: equipment pose, process parameters, and environmental parameters at the coal mining face.

[0058] In one specific embodiment, S1 includes:

[0059] S11. A pose coordinate system is established with the center of the scraper conveyor head in the initial mining state of the coal face as the origin, and the X-axis, Y-axis and Z-axis as the advancing direction of the coal face, the direction perpendicular to the coal face and the vertical direction, respectively.

[0060] S12: Obtain the position data and attitude angle data of key equipment in the coal mining face, map the position data of key equipment to the pose coordinate system, and combine the attitude angle data to form the pose vector of key equipment.

[0061] Specifically, key equipment includes coal mining machines and hydraulic supports. The attitude angle data of the coal mining machine includes drum angle data and machine body angle data. The attitude angle data of the hydraulic supports includes the top beam pitch angle and the base roll angle.

[0062] For example, at a certain moment, the coal face has advanced 100m (100m relative to the center of the scraper conveyor head in the initial mining state of the coal face). The pushing stroke of the 10th hydraulic support is 750mm, the width of the hydraulic support is 1.5m, the height of the hydraulic support is 3m, the pitch angle of the top beam is 1.5°, and the roll angle of the base is 0.8°. Then the pose vector of the 10th hydraulic support is Ezj={100.75, 15, 3, 1.5°, 0.8°}. The position of the coal mining machine body center corresponding to the 10th hydraulic support, the pitch angle of the left drum is 15°, the pitch angle of the right drum is -10°, the pitch angle of the machine body is 2°, the machine body height is 1.2m, and the drum radius is 0.6m. Then the pose vector of the coal mining machine is Ecmj={100, 15, 1.2, 0.6, 2°, 15°, -10°}.

[0063] S13. Extract the temporal evolution features of the pose vectors of key equipment in the time dimension, and integrate the pose vectors and temporal evolution features of all key equipment to form the spatiotemporal evolution feature vector E(t) of the pose of key equipment.

[0064] Specifically, a sliding window algorithm is used to extract the temporal evolution features of the pose vectors of key devices in the time dimension by setting the window size and window step size.

[0065] For example, setting the window size to 10 and the window step size to 2, the sliding window algorithm is used to statistically analyze the rate of change per second of the Y-axis coordinate of the coal mining machine and the rate of change per second of the hydraulic support's travel data. Referring to the example in S12 above, the extracted time evolution feature of the hydraulic support in the time dimension is 33, and the time evolution features of the coal mining machine in the time dimension are 3.0, 0.1, and -0.3.

[0066] Furthermore, when integrating the pose vectors and temporal evolution features of key equipment, the temporal evolution features are combined with the pose vectors. For example, the pose spatiotemporal evolution feature vector of the 10th hydraulic support obtained in the above example is Ezj={100.75, 1.5, 3, 1.5°, 0.8°, 33}, and the pose spatiotemporal evolution feature vector of the coal mining machine is Ecmj={100, 1.5, 1.2, 0.6, 2°, 15°, -10°, 3.0, 0.1, -0.3}. By integrating the pose vectors and temporal evolution features of the 10th hydraulic support and the coal mining machine, the pose spatiotemporal evolution feature vector of the key equipment is obtained as E(t)={Ezj, Ecmj}.

[0067] S14. Construct the feature vector Vt of the coal mining face advance and the feature vector Vs of the coal mining face production process parameters, and integrate them to obtain the dynamic causal feature vector P(t).

[0068] Specifically, according to the technological parameters such as the advancement of the coal mining face and the production process, a feature vector Vt of the advancement of the coal mining face and a feature vector Vs of the production process parameters of the coal mining face are constructed.

[0069] Among them, the elements of the feature vector Vt of the advancement of the coal mining face include the average advancement speed of the working face, the support quality of the support, and the straightness of the working face, etc. The elements of the feature vector Vs of the production process parameters of the coal mining face include the drum rotation speed of the coal shearer, the pressure condition of the 10th support, the pushing stroke of the 10th support, the gas concentration, etc.

[0070] In specific implementation, the S14 includes: S141, calculating the correlation intensity between each element in the feature vector Vt of the advancement of the coal mining face and the feature vector Vs of the production process parameters of the coal mining face through a similarity algorithm, and taking the correlation intensity as the weight between the two; S142, comparing the weight with the correlation intensity preset by expert experience, calculating the error c. If the error c is less than the preset error threshold c1 (the specific data of the preset error threshold c1 is determined according to experience), it is determined that the correlation intensity calculated by the similarity algorithm conforms to the correlation intensity preset by expert experience, and then directly integrate the weight, the elements in the feature vector Vt of the advancement of the coal mining face and the feature vector Vs of the production process parameters of the coal mining face to obtain the dynamic causal feature vector P(t); S143, if the error c is not less than the preset error threshold c1, that is, it is determined that the correlation intensity calculated by the similarity algorithm does not conform to the correlation intensity preset by expert experience, then adjust the calculated weight based on the weight preset by expert experience, and then integrate the feature vector Vt of the advancement of the coal mining face and the feature vector Vs of the production process parameters of the coal mining face with the adjusted weight to obtain the dynamic causal feature vector P(t).

[0071] For example, the constructed feature vector Vt of the advancement of the coal mining face = {average advancement speed of the coal mining face, support quality of the support, straightness of the working face} = {3.2 m / min, poor support quality of the 10th support, not straight in the area of the 8th - 12th hydraulic supports}, and the constructed feature vector Vs of the production process parameters of the coal mining face = {drum rotation speed of the coal shearer, pressure condition of the 10th support, pushing stroke of the 10th support, gas concentration} = {95 rpm, 31.2 Mpa, 745 mm, 3.2}. Calculate the similarity, that is, the correlation intensity, between each element in Vt and each element in Vs through the cosine similarity algorithm. Taking the non - straightness in the area of the 8th - 12th hydraulic supports and the pushing stroke of the 10th support as an example, the calculated correlation intensity is 0.85, and the correlation intensity preset by expert experience under this working condition is 0.84. Then the error c = 0.01 < c1 = 0.1 (preset error threshold), and a part of the dynamic causal feature vector P(t) obtained by integrating the weight is {non - straightness in the area of the 8th - 12th hydraulic supports, 745, 0.85}.

[0072] S15: Obtain various types of environmental parameters, and extract the time-domain features, trend features, and mutation features of each environmental parameter. Combine the preset weights of the time-domain features, trend features, and mutation features of each environmental parameter to obtain the environmental feature vector C(t).

[0073] Environmental parameters include mining pressure, gas, and dust.

[0074] In specific implementation, S15 includes: sequentially extracting the time-domain features, trend features, and mutation features of each type of environmental parameter through a sliding window, assigning a weight of 0.4 to the mutation feature of each environmental parameter (because the mutation feature has the greatest impact on the environmental parameter, so the mutation feature is assigned the highest weight of 0.4), assigning a weight of 0.3 to the trend feature and time-domain feature of each environmental parameter respectively, and integrating the time-domain features, trend features, mutation features of all types of environmental parameters and their weights to obtain the environmental feature vector C(t).

[0075] For example, taking mine pressure in the stope as an example, the sliding window size is set to 1 minute and the sliding step size is 10 seconds. First, the time-weighted mean is calculated to be 24.5 MPa and the variance is 1.2 MPa, etc., which are time-domain features. Then, the slope trend feature is calculated to be 0.2 MPa / min. The mutation checkpoint is designed as the slope exceeds 3 MPa / min (considered as a mine pressure mutation at this point). There are 0 mutations within the window. The mutation feature is assigned a weight of 0.4, and the trend feature and time-domain feature are assigned a weight of 0.3. Then, the environmental feature vector of mine pressure in the stope is C1(t) = {(0, 0.4)(24.5, 1.2, 0.3), (3, 0.3)}. Then, the environmental feature vector of gas C2(t) and the environmental feature vector of dust C3(t) are integrated to obtain C(t) = {(0, 0.4)(24.5, 1.2, 0.3), (3, 0.3), C2(t), C3(t)}. Here, the specific constituent elements of C2(t) and C3(t) are not specifically exemplified.

[0076] S16 uses a weighted and multi-head attention mechanism to fuse the spatiotemporal evolution feature vector E(t) of key equipment pose, the dynamic causal feature vector P(t) and the environmental feature vector C(t) to obtain a scenario-based multidimensional feature dataset.

[0077] Specifically, a weighted and multi-head attention mechanism is adopted to construct a three-dimensional attention-weighted feature fusion formula, expressed as:

[0078] ;

[0079] In the formula, , , These are the dynamic weights for the dimensions of device pose, process parameters, and environmental parameters, representing the degree of contribution of each dimension of data to the task in the current scenario, calculated by the multi-head attention Transformer.

[0080] For example, using a multi-head attention mechanism, the weights of E(t) are calculated to be 0.45, P(t) to be 0.35, and C(t) to be 0.2. A unified feature representation is then obtained through a three-dimensional attention-weighted feature fusion formula. Output t ={(E(t), 0.45), (P(t), 0.35), (C(t), 0.2)}.

[0081] The specific implementation process of step S1 is as follows: Figure 2 As shown.

[0082] S2 uses a multi-agent-driven hybrid storage method to store multi-source time-series data and scenario-based multi-dimensional feature datasets of the coal mining face, forming a data foundation.

[0083] Specifically, during coal mining, various monitoring data, environmental monitoring data, ledger data, and personnel and vehicle data are generated in real time, all of which are multi-source time-series data. The content of the scenario-based multi-dimensional feature dataset also varies over time. Therefore, this embodiment of the invention stores the multi-source time-series data and the scenario-based multi-dimensional feature dataset using a multi-agent driven hybrid storage method for multi-dimensional dynamic information in coal mining scenarios. The multi-agent includes a data acquisition agent, a data cleaning agent, a data classification agent, a data storage agent, and a data lineage agent, which are respectively used for data acquisition, data cleaning, data classification, data storage, and data lineage analysis. The specific functions of each agent are detailed in the following embodiments.

[0084] In one specific embodiment, S2 includes:

[0085] S21, the data acquisition agent collects multi-source time-series data and scenario-based multi-dimensional feature datasets from the coal mining face into the hybrid storage system, and transfers the collected data in the form of a Kafka data stream.

[0086] S22, the data cleaning agent uses a sliding window mechanism to calculate the statistical characteristics (mean, variance, etc.) of various types of data in the multi-source time-series data and scenario-based multi-dimensional feature dataset within the window, and determines whether the statistical characteristics of each type of data are within the corresponding threshold range. If the statistical characteristics of any type of data are not within the corresponding threshold range, then the data of that type is marked as abnormal data, and the data of that type is filled by a time-weighted average algorithm. After the data is filled, the data cleaning agent uses a time window plus dynamic time warping (DTW) algorithm to align the various types of data in the time dimension. If the statistical characteristics of all types of data are within the corresponding threshold range, then the data of that type is determined to be normal data, and the data of that type is directly aligned in the time dimension.

[0087] S23, the data cleaning agent classifies and encodes various types of data after time alignment according to predefined coal mine data classification and coding specifications, forming a high-quality standardized dataset.

[0088] Specifically, the coal mine data classification and coding standard mainly targets the classification and coding of intelligent mine data. It distinguishes, classifies and codes intelligent mine data according to certain principles and methods, and establishes a certain classification system, coding system and code element set.

[0089] Taking coal mining production as an example, if the production category code is defined as MJB, the coal mining code is 01, the coal mining machine code is 02, the coal mining machine traction unit code is 003, and the traction unit voltage monitoring code is 0001, then the coal mining machine traction unit voltage monitoring data will be coded as MJB01020030001.

[0090] S24, the data classification agent classifies each data entry in the high-quality standardized dataset according to the data format and data volume; for any target data entry in the high-quality standardized dataset, if the target data entry is vector data, it is directly stored in the database; if the target data entry is not vector data, the sliding window algorithm is used to determine whether the number of data entries in the target data entry within the first n sliding windows is greater than the preset threshold N; if the number of data entries within n×1% sliding windows is greater than N, the target data entry is determined to be massive time series data; otherwise, the target data entry is determined to be a small amount of time series data.

[0091] Specifically, when determining whether a target data entry is vector data, it is determined whether the target data entry is a sequence consisting entirely of numerical values. If it is, the target data entry is determined to be vector data; otherwise, the target data entry is determined not to be vector data.

[0092] S25, the data storage agent stores massive amounts of time-series data in a distributed columnar database, stores a small amount of time-series data in a time-series database, and stores vector data in a vector database.

[0093] Furthermore, after S25, it also includes: S26, where the data lineage agent sorts out the data lineage and metadata of the data collection process of the data collection agent, the data cleaning process of the data cleaning agent, the data classification process of the data classification agent, and the data storage process of the data storage agent, in order to sort out the data collection points, data point sources, data units, data types, normal value ranges, processing procedures, and storage locations (including storage media, storage repositories, and storage tables).

[0094] Through steps S25 and S26, the semantics of smart coal mine data can be unified, ensuring the consistency, integrity, and visibility of key data, while also facilitating data access for each agent.

[0095] The specific implementation process of step S2 is as follows: Figure 3 As shown.

[0096] S3, after obtaining the coal mining scenario task instruction, based on the data base, combined with equipment, process and environmental information, maps out the relevant scenario, and according to the roles and functions of various types of agents pre-deployed at the coal mining face, the various types of agents cooperate to complete the content indicated by the coal mining scenario task instruction.

[0097] Specifically, this step mainly involves summarizing coal mining scenarios and constructing the Agent information space. From four dimensions—"task, equipment, process, and environment"—it integrates statistical analysis, clustering algorithms, and expert experience to abstract and summarize a set of diverse scenario representation information in coal mining. For the input coal mining scenario task instructions, it maps related scenarios, data, and internal states based on equipment, process, and environmental information. According to the roles and functions of each Agent, it designs a multi-layered "task-scenario-data mapping relationship," constructing a linkage architecture between the Agent information space and the task / scenario response mechanism.

[0098] In one specific embodiment, S3 includes:

[0099] S31, when the coal mining scenario command t is detected, the coal mining scenario command t is mapped to the knowledge graph set W.

[0100] It should be noted that before step S31, a knowledge graph set W needs to be constructed first. Constructing the knowledge graph set W includes: during the coal mining production process at the coal face, defining task entities T, equipment entities E, process entities P, and environmental entities Env based on expert experience, and defining the relationships between entities. After the expert experience definition is completed, the correlation strength I between each entity is calculated using a cosine similarity algorithm. Based on the relationships between entities and the correlation strength I, a four-dimensional knowledge graph set with relational weights is constructed: W = {(T1, E1, P1, Env1, I1), (T2, E2, P2, Env2, I2), ...}. Among these, the task entity T includes the task and the task type. The task type includes generative tasks, investigative tasks, and learning tasks.

[0101] For example, based on expert experience, the following entities are defined: task entity T {(face straightening, generative task), (face abnormal condition adjustment, generative task), (judging hydraulic support quality, investigative task), ...}; equipment entity E {coal mining machine, scraper, hydraulic support, emulsion pump, ...}; process entity P {hydraulic support column lowering, hydraulic support column raising, hydraulic support push conveyor, coal mining machine cutting, ...}; and environmental entity Env {mine pressure, gas, hydrology, face geological information, ...}. A knowledge graph set W = {(face straightening, hydraulic support, 0.6), (face straightening, hydraulic support push conveyor, 0.4), (face straightening, coal mining machine startup, 0.2), (face straightening, face geological information, 0.5), (judging hydraulic support quality, mine pressure, 0.7), ...}.

[0102] Based on a pre-constructed knowledge graph set W, upon detecting a coal mining scenario command t, the command t is first matched against the task entities of each knowledge graph in the knowledge graph set W to determine if the command t can be successfully mapped to the knowledge graph set W. Specifically, this matching is achieved by calculating the cosine similarity between the command t and the task entities of each knowledge graph in the knowledge graph set W. If the cosine similarity between the command t and a task entity of a certain knowledge graph is greater than a similarity threshold, then the command t is considered successfully matched with that knowledge graph's task entity.

[0103] S32. If the coal mining scenario instruction t is successfully mapped to the knowledge graph set W, then according to the coal mining scenario task instruction t, the associated equipment entities, process entities, environmental entities and corresponding association weights are traversed in the knowledge graph set W to form an entity weight set Q.

[0104] The entity weight set Q can be represented as Q={[t, E1, I1], [t, P1, I2], [t, Env1, I3]...}.

[0105] Based on the example in S31 above, assuming the coal mining scenario instruction t is "perform straightening of working face 31001", after the coal mining scenario instruction "perform straightening of working face 31001" is input, it is mapped to the knowledge graph set W. The task entity in the corresponding knowledge graph set W is {working face straightening}. Traversing the knowledge graph set W yields the entity weight set Q = {(working face straightening, hydraulic support, 0.6), (working face straightening, hydraulic support pushing conveyor, 0.4), (working face straightening, coal mining machine cutting, 0.2), (working face straightening, working face geological information, 0.5), ...}.

[0106] S33, the reflective program Agent obtains the real-time dataset D1 of the associated entities in the entity weight set Q from the data base by association weight priority order, while the inference model Agent obtains the historical dataset D2 of the first n coal cutting cycles of the associated entities from the data base by association weight priority order.

[0107] Using the example above, the reflective agent program sequentially obtains real-time data of the hydraulic support, real-time data of the working face geological information, real-time data of the hydraulic support pushing process, and real-time data of the coal cutting machine, forming a real-time dataset D1, based on the priority order of association weights. The inference model agent sequentially obtains historical data of the hydraulic support for the previous n coal cutting cycles, historical data of the working face geological information for the previous n coal cutting cycles, historical data of the hydraulic support pushing process for the previous n coal cutting cycles, and historical data of the coal cutting machine for the previous n coal cutting cycles, obtaining a historical dataset D2.

[0108] S34. Construct a coal mining scenario S with four dimensions: task, equipment, process, and environment, based on the real-time dataset D1 and historical dataset D2 of the associated entities.

[0109] For example, a coal mining scenario S{Q,D1,D2} can be constructed using the entity weight set Q, the real-time dataset D1, and the historical dataset D2.

[0110] S35 uses the K-means clustering algorithm to perform clustering analysis on the coal mining scenario S, and combines the cosine similarity algorithm and expert experience to adjust the weights of the entity datasets associated with the actual coal mining scenario S in real time, and update the entity weight set Q, while also updating the knowledge graph set W.

[0111] Specifically, K-means clustering is used to cluster mining scenarios S to group scenarios with similar equipment, process, and environmental conditions into one class, facilitating subsequent weight adjustments. Assuming this scenario is clustered as a fully mechanized high-gas top-coal caving mining scenario, and combining cosine similarity algorithm and expert experience, based on mining scenario S, the correlation between the face straightening task and the coal cutting process under the current working environment is calculated to be 0.4. Then, the weights of the entity datasets associated with mining scenario S are adjusted in real-time, and the entity weight set Q is updated, while the knowledge graph set W is also updated.

[0112] S36. After the update is completed, determine the task type of the coal mining scenario instruction t that is successfully mapped in the knowledge graph set W. If the coal mining scenario instruction t is a generative task, then the generative task agent, planning organization agent, planning arbitration agent, reasoning model agent, and reflection program agent will collaboratively complete the generative task based on the coal mining scenario S. If the coal mining scenario instruction t is a learning task, then the learning reality loop agent, learning simulation loop agent, and learning dual loop agent will collaboratively complete the learning task based on the coal mining scenario S, thereby updating the coal mining scenario model. If the coal mining scenario instruction t is an investigation instruction, then the generative investigation agent will collaboratively complete the investigation task based on the coal mining scenario S.

[0113] In specific implementation, the generative task agent, planning organization agent, planning arbitration agent, reasoning model agent, and reflective program agent collaboratively complete the generation task based on the coal mining scenario S. This includes: the generative task agent combining the tasks in the coal mining scenario instruction t and the entities involved in the coal mining scenario S to obtain control instruction I; the planning organization agent and the planning arbitration agent cooperating to determine the optimal target planning scheme based on the tasks successfully mapped to the coal mining scenario instruction t, and transferring the optimal target planning scheme to the reasoning model agent; the reasoning model agent performing dynamic decision reasoning on the optimal target planning scheme, transforming the optimal target planning scheme into specific and implementable equipment execution instructions, and then issuing the equipment execution instructions to the reflective program agent; and the reflective program agent controlling each coal mining entity equipment to execute the equipment execution instructions.

[0114] Specifically, when the planning organization agent and the planning arbitration agent cooperate to determine the optimal target planning scheme, the planning organization agent organizes a set of coal mining production links V (composed of the coal mining production links required to execute the tasks successfully mapped to t) based on the tasks successfully mapped to the coal mining scenario instruction t. It then obtains the production weights of each coal mining production link in set V to form a production weight set A (the weight values ​​are determined by experts based on the importance of the coal mining production links). Based on the coal mining production link set V and the production weight set A, it performs planning on a directed weighted graph G, and determines the initial optimal target planning scheme using the shortest path algorithm of the directed weighted graph. This initial optimal target planning scheme is then submitted to the planning arbitration agent. The planning arbitration agent performs multi-dimensional verification of the initial optimal target planning scheme. If the initial optimal target planning scheme fails verification, the planning arbitration agent reports the conflict point to the planning organization agent. The planning organization agent readjusts the initial target planning scheme and resubmits it to the planning arbitration agent for arbitration until the conflict is resolved, thus obtaining the optimal target planning scheme.

[0115] The multi-dimensional verification includes safety compliance, timing verification, and resource allocation verification. Specifically, the planning arbitration agent first determines whether the task sequence in the initial optimal target planning scheme meets the preset safety procedure set (composed of several coal mine safety procedures, such as the gas concentration during coal mining must not exceed 2.5%). Then, it determines whether the task sequence in the initial optimal target planning scheme meets the preset coal mining process timing arrangement set (composed of several coal mining process timings, such as the hydraulic support relocation process must follow the lowering-relocation-raising sequence). Finally, it calls the hardware system resource data to determine whether the hardware resources required for the operation of the initial optimal target planning scheme are sufficient. Once all three verifications pass, the verification is considered successful.

[0116] The learning-oriented reality loop agent, learning-oriented simulation loop agent, and learning-oriented dual-loop agent collaboratively complete learning tasks based on the coal mining scenario S, realizing model updates for the coal mining scenario. This includes: a reflective program agent collecting real-time production data from the coal mining entity equipment and feeding this data back to the inference model agent; the inference model agent analyzing the production data to generate equipment execution strategies for the coal mining entity equipment and synchronously feeding these strategies and production data back to the learning-oriented reality loop agent; and the learning-oriented reality loop agent performing model self-learning based on the equipment execution strategies and production data. The learning-oriented simulation loop agent, relying on the coal mining simulation system, automatically generates and optimizes model update algorithms based on equipment execution strategies and production data, providing knowledge from the simulation environment for the iteration of the coal mining entity-related models. The learning-oriented dual-loop agent integrates the learning results of the learning-oriented reality loop agent and the learning-oriented simulation loop agent, improving the self-learning process of the coal mining entity-related models and achieving model optimization and model sinking.

[0117] Specifically, the learning-based reality loop agent automatically executes the integration-analysis-modeling process based on the equipment execution strategy and production data to complete the initial update of the coal mining equipment entity-related model; the learning-based dual-loop agent deeply integrates the field experience from the learning-based reality loop agent with the simulation knowledge from the learning-based simulation loop agent through the fusion analysis algorithm to complete the self-learning process of the coal mining equipment entity-related model and generate a new coal mining equipment entity-related model.

[0118] The generative survey agent collaboratively completes the survey task based on the coal mining scenario S, including: the generative survey agent queries the entity dataset related to the task, equipment, process, and environment through the data base based on the coal mining scenario S, and generates feedback results on the working status of the coal mining site.

[0119] The basic information of the various types of Agents involved in this step is shown in Table 1.

[0120]

[0121] The specific implementation process of step S3 is as follows: Figure 4 As shown.

[0122] In summary, the architecture of the coal mining operation multi-source time-series data attention perception and dynamic memory method provided in this embodiment of the invention is as follows: Figure 5As shown in the figure. This invention addresses different stages of coal mining, geological conditions, and equipment combinations, abstracting and summarizing diverse typical scenarios from four dimensions: "task, equipment, process, and environment." Using statistical analysis, clustering algorithms, and expert experience, it extracts information feature combinations from these diverse scenarios, establishes a scenario labeling system and semantic descriptions, and constructs an input / output information space for each agent. A three-dimensional attention mechanism of "equipment status - process parameters - environmental parameters" is designed. A spatiotemporal semantic feature encoding method for multi-source time-series data such as sensor signals, control commands, and human-machine interaction information during the coal mining process is studied. Utilizing time-series, columnar, and vector database technologies, a hybrid architecture of a dynamic memory for representing diverse scenarios is constructed, enabling scenario-level semantic reasoning and digital cognition of the coal mining production process based on multi-source time-series data.

[0123] Through a multi-agent information / scenario linkage architecture, employing a collaborative model of "multi-type agents + multi-dimensional scenario features," dynamic perception and associated response to coal mining operation scenarios are achieved, forming a precise linkage of "information-scenario-agent" and providing "scenario-based cognitive basis." By dynamically extracting three-dimensional features from multi-source time-series data and fusing scene information with attention, the transformation from raw time-series data to effective features is realized, ultimately generating a scenario-based multi-dimensional feature dataset. The hybrid storage technology for diversified dynamic scenario information serves as the core storage foundation for the data, employing a dynamic hierarchical mechanism with multi-database collaboration to achieve efficient management of massive amounts of data in smart coal mines.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for attention perception and dynamic memory of multi-source time-series data in coal mining operations, characterized in that, include: S1, dynamically extract three-dimensional features and fuse scene information attention into multi-source time-series data of the coal mining face to generate a scene-based multi-dimensional feature dataset, wherein the three-dimensional features include equipment pose features, process features and environmental features; S2 uses a multi-agent-driven hybrid storage method to store multi-source time-series data and scenario-based multi-dimensional feature datasets of the coal mining face, forming a data foundation. S2 includes: S21, the data acquisition agent collects multi-source time-series data and scenario-based multi-dimensional feature datasets from the coal mining face into the hybrid storage system, and transfers the collected data in the form of a Kafka data stream; S22, the data cleaning agent uses a sliding window mechanism to calculate the statistical characteristics of various types of data in the multi-source time-series data and the scenario-based multi-dimensional feature dataset within the window, and determines whether the statistical characteristics of each type of data are within the corresponding threshold range. If the statistical characteristics of any type of data are not within the corresponding threshold range, then the data of that type is marked as abnormal data, and the data of that type is filled using a time-weighted average algorithm. After the data is filled, the data cleaning agent uses a time window plus a dynamic time warping algorithm to align the various types of data in the time dimension. If the statistical characteristics of all types of data are within the corresponding threshold range, then the data of that type is determined to be normal data, and the data of that type is directly aligned in the time dimension. S23, The data cleaning agent classifies and encodes various types of data after time alignment according to predefined coal mine data classification and coding specifications, forming a high-quality, standardized dataset. S24, the data classification agent classifies each data entry in the high-quality standardized dataset according to the data format and data volume; for any target data entry in the high-quality standardized dataset, if the target data entry is vector data, it is directly stored in the database; if the target data entry is not vector data, the sliding window algorithm is used to determine whether the number of data entries in the target data entry within the first n sliding windows is greater than the preset threshold N; if the number of data entries within n×1% sliding windows is greater than N, the target data entry is determined to be massive time series data; otherwise, the target data entry is determined to be a small amount of time series data. S25, the data storage agent stores massive amounts of time-series data in a distributed columnar database, stores a small amount of time-series data in a time-series database, and stores vector data in a vector database; S3, after obtaining the coal mining scenario task instruction, based on the data base, combined with equipment, process and environmental information, the relevant coal mining scenario is mapped out, and according to the roles and functions of various types of agents pre-deployed at the coal mining face, the various types of agents cooperate to complete the content indicated by the coal mining scenario task instruction. S3 includes: S31, When the coal mining scenario command t is detected, the coal mining scenario command t is mapped to the knowledge graph set W; S32, if the coal mining scenario instruction t is successfully mapped to the knowledge graph set W, then according to the coal mining scenario task instruction t, the associated equipment entities, process entities, environmental entities and corresponding associated weights are traversed in the knowledge graph set W to form an entity weight set Q; S33, the reflective program Agent obtains the real-time dataset D1 of the related entities in the entity weight set Q from the data base through the association weight priority order, while the inference model Agent obtains the historical dataset D2 of the first n coal cutting cycles of the related entities from the data base through the association weight priority order. S34, construct a coal mining scenario S with four dimensions: task, equipment, process, and environment, based on the real-time dataset D1 and historical dataset D2 of the related entities; S35 uses the K-means clustering algorithm to perform clustering analysis on the coal mining scenario S, and combines the cosine similarity algorithm and expert experience to adjust the weights of the entity datasets associated with the actual coal mining scenario S in real time, and update the entity weight set Q, while also updating the knowledge graph set W. S36. After the update is completed, determine the task type of the coal mining scenario instruction t that is successfully mapped in the knowledge graph set W. If the coal mining scenario instruction t is a generative task, then the generative task agent, planning organization agent, planning arbitration agent, reasoning model agent, and reflection program agent will collaboratively complete the generative task based on the coal mining scenario S. If the coal mining scenario instruction t is a learning task, then the learning reality loop agent, learning simulation loop agent, and learning dual loop agent will collaboratively complete the learning task based on the coal mining scenario S, thereby updating the coal mining scenario model. If the coal mining scenario instruction t is an investigation instruction, then the generative investigation agent will collaboratively complete the investigation task based on the coal mining scenario S.

2. The method for attention perception and dynamic memory of multi-source time-series data in coal mining operation according to claim 1, characterized in that, S1 includes: S11. A pose coordinate system is established with the center of the scraper conveyor head in the initial mining state of the coal face as the origin, and the X-axis, Y-axis and Z-axis as the advancing direction of the coal face, the direction perpendicular to the coal face and the vertical direction, respectively. S12: Obtain the position data and attitude angle data of key equipment in the coal mining face, and after mapping the position data of the key equipment to the pose coordinate system, combine the attitude angle data to form the pose vector of the key equipment. S13, extract the temporal evolution features of the pose vectors of key equipment in the time dimension, and integrate the pose vectors and temporal evolution features of all key equipment to form the spatiotemporal evolution feature vector E(t) of the pose of key equipment. S14, construct the feature vector Vt of the coal mining face advance and the feature vector Vs of the coal mining face production process parameters, and integrate them to obtain the dynamic causal feature vector P(t); S15: Obtain various types of environmental parameters, extract the time-domain features, trend features, and mutation features of each environmental parameter, and integrate them with the preset weights of the time-domain features, trend features, and mutation features of each environmental parameter to obtain the environmental feature vector C(t). S16 uses a weighted and multi-head attention mechanism to fuse the spatiotemporal evolution feature vector E(t) of key equipment pose, the dynamic causal feature vector P(t) and the environmental feature vector C(t) to obtain a scenario-based multidimensional feature dataset.

3. The method for attention perception and dynamic memory of multi-source time-series data in coal mining operation according to claim 2, characterized in that, S14 includes: S141, calculate the correlation strength between each element in the feature vector Vt of the coal mining face advancement and the feature vector Vs of the coal mining face production process parameters using a similarity algorithm, and use the correlation strength as the weight of the two. S142, compare the weight with the association strength preset by expert experience, calculate the error c. If the error c is less than the preset error threshold c1, it is determined that the association strength calculated by the similarity algorithm is consistent with the association strength preset by expert experience. Then, directly integrate the elements in the weight, the feature vector Vt of the coal mining face advancement and the feature vector Vs of the coal mining face production process parameters to obtain the dynamic causal feature vector P(t). S143, if the error c is not less than the preset error threshold c1, that is, if the correlation strength calculated by the similarity algorithm does not match the correlation strength preset by the expert experience, then the weight calculated is adjusted according to the weight preset by the expert experience. Then, the weight after integrating the feature vector Vt of the coal mining face advance and the feature vector Vs of the coal mining face production process parameters is used to obtain the dynamic causal feature vector P(t).

4. The method for attention perception and dynamic memory of multi-source time-series data in coal mining operation according to claim 2 or 3, characterized in that, S15 includes: The time-domain features, trend features, and mutation features of each type of environmental parameter are extracted sequentially using a sliding window. The mutation feature of each environmental parameter is assigned a weight of 0.4, and the trend feature and time-domain feature of each environmental parameter are assigned a weight of 0.3 respectively. The time-domain features, trend features, mutation features of all types of environmental parameters and their weights are integrated to obtain the environmental feature vector C(t).

5. The method for attention perception and dynamic memory of multi-source time-series data in coal mining operation according to claim 1, characterized in that, Following S25, the following is also included: S26, the data lineage agent sorts out the data lineage and metadata of the data collection agent's data collection process, the data cleaning agent's data cleaning process, the data classification agent's data classification process, and the data storage agent's data entry and storage process, in order to sort out the data collection points, data point sources, data units, data types, normal value ranges, processing procedures, and storage locations.

6. The method for attention perception and dynamic memory of multi-source time-series data in coal mining operation according to claim 1, characterized in that, Before S31, it also includes: During the coal mining process at the coal face, task entity T, equipment entity E, process entity P, and environmental entity Env are defined based on expert experience, and the relationships between entities are defined. After the expert experience definition is completed, the cosine similarity algorithm is used to calculate the association strength I between each entity. Based on the relationships between entities and the association strength I between entities, a four-dimensional knowledge graph set W={T、E、P、Env、I} with relation weights is constructed.

Images