Intelligent preventive maintenance system and method for coal machine equipment multi-modal data

Abstract

The present application relates to a coal machine equipment multi-modal data intelligent preventive maintenance system and method. The system includes a multi-source data acquisition layer, a large language model core processing layer, and an intelligent decision output layer, capable of simultaneously processing structured and unstructured data. The maintenance system of the present application has strong data processing capability and high fault prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal mine production equipment; specifically, the present application relates to a coal machine equipment multi-modal data intelligent preventive maintenance system and method. BACKGROUND

[0002] Coal machine equipment, i.e., coal production equipment operation and maintenance technology, has experienced a development stage from manual experience dominance to informationized monitoring. In the early stage, regular inspection by operation and maintenance personnel was relied on to judge equipment failure through senses; with the application of Internet of Things (IoT) technology, real-time sensor data acquisition and failure early warning were gradually realized to achieve real-time monitoring and maintenance of coal machine equipment.

[0003] The existing coal machine equipment operation and maintenance technology has limited data processing capability and insufficient failure prediction accuracy. SUMMARY

[0004] Therefore, the present application provides a coal machine equipment multi-modal data intelligent preventive maintenance system and method to solve or at least alleviate one or more of the above problems and other problems in the prior art.

[0005] To achieve the foregoing purpose, a first aspect of the present application provides a coal machine equipment multi-modal data intelligent preventive maintenance system, wherein the system comprises:

[0006] a multi-source data acquisition layer having a data cleaning module and a distributed database, the data cleaning module is connected to equipment sensors and a coal mine operation and maintenance management system, the equipment sensors are used to provide real-time running structured data of the coal machine equipment to the data cleaning module, the coal mine operation and maintenance management system is used to provide unstructured data related to the coal machine equipment to the data cleaning module, and the data cleaning module stores the cleaned and desensitized structured data and unstructured data as a multi-modal data set to the distributed database;

[0007] a large language model (LLM) core processing layer, the LLM core processing layer includes a coal machine failure dynamic prediction model and a coal machine knowledge reasoning module, the coal machine failure dynamic prediction model is connected to a multi-modal data fusion module and a reinforcement learning engine, the multi-modal data fusion module establishes an associated mapping of the structured data and the unstructured data and constructs a coal machine equipment state feature matrix, the reinforcement learning engine is used to optimize model parameters of the coal machine failure dynamic prediction model, the coal machine failure dynamic prediction model is used to generate a failure prediction result based on the coal machine equipment state feature matrix and the optimized model parameters of the coal machine failure dynamic prediction model, and the coal machine knowledge reasoning module generates technical points of a maintenance task based on the unstructured data and the failure prediction result;

[0008] An intelligent decision output layer, comprising a coal machine fault maintenance strategy generator and a coal machine operation and maintenance work order system, the coal machine fault maintenance strategy generator being configured to generate a maintenance strategy based on technical points of the maintenance task, the coal machine operation and maintenance work order system being configured to generate an operation and maintenance work order based on the maintenance strategy, and the intelligent decision output layer being connected to an operation and maintenance personnel mobile terminal and pushing the operation and maintenance work order to the operation and maintenance personnel mobile terminal.

[0009] In the system as described above, optionally, the structured data comprises a current signal, a power signal, a vibration signal, a temperature signal and / or a speed signal, and the unstructured data comprises a coal machine equipment operation manual, a repair log and / or an inspection record.

[0010] In the system as described above, optionally, the multi-modal data fusion module performs standardization processing on the structured data to convert the structured data into a vector sequence, performs vectorization processing on the unstructured data by using a text encoder of a large language model, extracts semantic features related to fault description and repair measures, and then establishes an association mapping between the structured data and the unstructured data by using a cross-modal attention mechanism.

[0011] In the system as described above, optionally, the coal machine fault dynamic prediction model inputs the coal machine equipment state feature matrix and historical operation data of the coal machine equipment, outputs a fault probability distribution of a component of interest within a preset future time based on a sequence generation capability of a large language model, and introduces the reinforcement learning engine to dynamically adjust model parameters of the coal machine fault dynamic prediction model according to actual operation and maintenance feedback to realize continuous self-optimization.

[0012] In the system as described above, optionally, the maintenance strategy comprises an optimal maintenance time window, and the maintenance strategy is divided into emergency repair, planned maintenance and state monitoring by the maintenance time window.

[0013] To achieve the foregoing object, a second aspect of the present application provides a coal machine equipment multi-modal data intelligent preventive maintenance method using the coal machine equipment multi-modal data intelligent preventive maintenance system according to any one of the foregoing first aspects.

[0014] In the method as described above, optionally, the method comprises:

[0015] a data input stage, at which the collection and timestamp alignment processing of the structured data and the unstructured data are performed to form the multi-modal data set;

[0016] a large language model processing stage, at which cross-modal feature fusion, fault probability sequence generation and maintenance specification matching are performed on the structured data and the unstructured data to generate the technical points of the maintenance task.

[0017] a policy output stage, in which a production downtime window is obtained and the maintenance policy is generated, the operation and maintenance order is generated and pushed;

[0018] a feedback optimization closed loop stage, in which operation and maintenance results are collected and the model parameters are fine-tuned.

[0019] In the method as described above, optionally, in the data input stage:

[0020] by the equipment sensor at a fixed sampling frequency timely collection of structured data of the coal mining equipment to form a numerical time series wherein is a timestamp, is a sensor feature vector ;

[0021] the unstructured data is imported and denoted as wherein is a timestamp of the unstructured data, is a text string of the unstructured data;

[0022] a bilinear interpolation method is used to align the timestamps of the multi-modal data to generate a unified time axis wherein for any time point , the aligned value of the structured data is , the unstructured data is aligned to T by matching, wherein , thereby forming an aligned multi-modal data set .

[0023] In the method as described above, optionally, in the large language model processing stage,

[0024] the step of cross-modal feature fusion includes: converting the unstructured data into a semantic vector by an Encoder of a large language model, splicing the numerical vector of the structured data, and inputting a Transformer network for cross-modal feature extraction, wherein a pre-trained large language model is used as the Encoder to encode the unstructured data and output a text embedding vector wherein is a model hidden layer dimension, and the structured data is standardized by sensor numerical coding The historical mean and standard deviation are mapped to numerical vectors through a fully connected layer. ,in For learnable parameters, and spliced ​​as The input is processed through a Transformer network to achieve cross-modal feature extraction. The multi-head attention mechanism of the Transformer network is calculated as follows: Each of the heads The attention function is Q represents the query, K represents the key, and V represents the value. generate, For header-related parameters, To output the projection matrix and output cross-modal features. ;

[0025] The steps for generating the fault probability sequence include:

[0026] A fault probability sequence for future time periods is generated using the decoder of a large language model. Real-time rolling prediction is achieved using a sliding window technique. The input window is constructed using a sliding window of W = 24 hours with a window step size S = 1 hour. For time point t, the input window is... Fault probability sequence generation involves inputting window features into the decoder of a large language model and outputting future... Hourly Failure Probability Sequence ,in express The probability of failure at time step, the Decoder generation process satisfies ,in For Decoder in The hidden state at all times For the classification layer parameters, the model training loss uses the binary cross-entropy loss function, and the supervision label is the historical fault marker. Where 0 represents normal and 1 represents fault, and the loss function is... ,

[0027] The steps for matching maintenance specification knowledge include:

[0028] Based on a coal mining machinery maintenance standard knowledge base, a retrieval enhancement mechanism using a large language model is employed to match the maintenance rules corresponding to the current state of the coal mining machinery equipment. This involves constructing the coal mining machinery maintenance standard knowledge base. ,in Keywords for knowledge entries For the corresponding maintenance rules, the cross-modal characteristics of the current coal mining machine status will be... As a query vector, the cosine similarity of its embedding vector with the keyword embedding vectors of each entry in the knowledge base is calculated The top entries with the highest similarity are selected as the search results

[0029] The maintenance rule generation is to splice the and the current state description into a prompt word , and input the large language model to generate the final maintenance rule .

[0030] In the method as described above, optionally, in the policy output stage, in combination with the downtime window of the production scheduling system, high-risk fault prediction results are preferentially processed, maintenance work orders containing three levels of priority are generated, and coal machine equipment fault related spare parts inventory early warning is triggered synchronously, wherein

[0031] The priority classification rule divides three levels of priority based on fault probability

[0032]

[0033] Among them , ,

[0034] The downtime window matching is realized by acquiring the downtime window of the production scheduling system , and the red priority fault is allocated to the nearest available window ,

[0035] The spare parts inventory early warning is realized according to the fault type , querying the spare parts inventory , if , wherein is the average spare parts consumption of the same type of historical fault, is a safety factor, then the early warning is triggered .

[0036] The coal machine equipment multi-modal data intelligent preventive maintenance system and method of the present application adopts a multi-source data acquisition layer, a large language model core processing layer and an intelligent decision output layer, can simultaneously process structured and unstructured data, and improves the data processing capability and fault prediction accuracy of the maintenance system.

[0037] The present application further provides a corresponding coal machine equipment multi-modal data intelligent preventive maintenance method, which also has the above advantages. BRIEF DESCRIPTION OF DRAWINGS

[0038] ​​​​The disclosure of the present application will be more apparent with reference to the drawings. It should be understood that these drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the present application. In the drawings:

[0039] Figure 1 a schematic architecture diagram of one embodiment of the coal equipment multi-modal data intelligent preventive maintenance system according to the present application; and

[0040] Figure 2 a schematic flow chart of one embodiment of the coal equipment multi-modal data intelligent preventive maintenance method according to the present application. DETAILED DESCRIPTION

[0041] With reference to the drawings and specific embodiments, the structure, composition, features and advantages of the coal equipment multi-modal data intelligent preventive maintenance system and method according to the present application will be described below in an exemplary manner, however, all descriptions should not be used to form any limitation on the present application.

[0042] In addition, for any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the drawings, the present application still allows any combination or deletion to be continued between these technical features (or their equivalents) without any technical obstacles, so it should be considered that more embodiments according to the present application are also within the scope of the description herein.

[0043] Figure 1 a schematic architecture diagram of one embodiment of the coal equipment multi-modal data intelligent preventive maintenance system according to the present application.

[0044] As shown in the drawings, in this embodiment, the coal equipment multi-modal data intelligent preventive maintenance system can include a multi-source data acquisition layer, a large language model core processing layer and an intelligent decision output layer. It can be understood that this is not exhaustive, and in some embodiments, the coal equipment multi-modal data intelligent preventive maintenance system can also include other functional layers or modules to adapt to different working environments and working requirements.

[0045] In the illustrated example, the multi-source data collection layer can have a data cleaning module and a distributed database. As shown in the figure, the data cleaning module can be connected with device sensors and a coal mine operation and maintenance management system. The device sensors are used to provide real-time running structured data of coal machine equipment to the data cleaning module, and the coal mine operation and maintenance management system is used to provide unstructured data related to coal machine equipment to the data cleaning module. The data cleaning module can store the cleaned and desensitized structured data and unstructured data as a multi-modal data set to the distributed database. The multi-modal data set is a data set containing two or more different types (modalities) of data, such as structured data and unstructured data in this embodiment. These data are collectively used to describe the maintenance information of coal machine equipment, and the different modalities of data are semantically associated.

[0046] The device sensors here can include current, power, vibration, temperature, pressure, speed, etc. related sensors, which can collect real-time running data of coal machine equipment. These real-time running data are structured data, and the related data signals are provided to the intelligent preventive maintenance system in real time. The coal mine operation and maintenance management system here is used to store and provide non-structured data such as coal machine equipment manuals, maintenance logs, and inspection records, in order to fully mine the coal machine equipment operation and maintenance knowledge. In some embodiments, these unstructured data can be text, pictures, and even videos, audio, or converted text data of pictures, videos, and audio.

[0047] The device sensors and the coal mine operation and maintenance management system can be a component of the coal machine equipment multi-modal data intelligent preventive maintenance system of the present application itself, or can be an external device or module of the coal machine equipment multi-modal data intelligent preventive maintenance system of the present application. It provides structured data and unstructured data to the data cleaning module for further cleaning, desensitization, etc. by the data cleaning module, and stores them uniformly to the distributed database. According to the foregoing description, the structured data is highly organized and fixed format data, which can include current signals, power signals, vibration signals, temperature signals, pressure signals, and / or speed signals, etc. The unstructured data is data without a predefined format or organization model, which can include operation manuals, maintenance logs, and / or inspection records of coal machine equipment, etc.

[0048] The data cleaning module is an important link of data preprocessing, which can process structured data and unstructured data, identify and correct / remove inaccurate, incomplete, format error, duplicate or irrelevant data in the data set, so as to improve the data quality and make it suitable for subsequent analysis, modeling or application. The functions of the data cleaning module may, for example, include supplementing missing values, correcting outliers, deleting duplicate values, data normalization, etc. For cleaning of unstructured data, it can also include noise removal, format conversion, information structuring, etc. Data cleaning can effectively improve data reliability and usability, and lay a solid and reliable foundation for downstream data analysis, modeling and application.

[0049] In the illustrated example, the large language model core processing layer can include a coal machine fault dynamic prediction model and a coal machine knowledge reasoning module. As shown in the figure, the coal machine fault dynamic prediction model can be connected to a multi-modal data fusion module and a reinforcement learning engine, wherein the multi-modal data fusion module can establish an association mapping of structured data and unstructured data and construct a coal machine equipment state feature matrix, and the reinforcement learning engine can optimize the model parameters of the coal machine fault dynamic prediction model. Further, the coal machine fault dynamic prediction model can generate a fault prediction result based on the coal machine equipment state feature matrix and the optimized model parameters of the coal machine fault dynamic prediction model, and the coal machine knowledge reasoning module can generate technical points of maintenance tasks based on unstructured data and fault prediction results.

[0050] Specifically, the coal machine fault dynamic prediction model can input the coal machine equipment state feature matrix and the historical operation data of the coal machine equipment, based on the sequence generation capability of the large language model, output the fault probability distribution of the concerned components within the preset future time, and introduce the reinforcement learning engine to dynamically adjust the model parameters of the coal machine fault dynamic prediction model according to the actual operation and maintenance feedback, realizing continuous self-optimization.

[0051] In optional embodiments, the multi-modal data fusion module and the reinforcement learning engine can be integral parts of the coal machine equipment multi-modal data intelligent preventive maintenance system of the present application, or can also be external devices or modules of the coal machine equipment multi-modal data intelligent preventive maintenance system of the present application.

[0052] For example, the multi-modal data fusion module can standardize the structured data of sensor time series, convert it into a vector sequence, and use the text encoder of the large language model to vectorize unstructured data such as coal machine equipment manuals, maintenance logs, and on-site instrument photos, and extract deep semantic features such as fault description, maintenance measures, and maintenance specifications. Through the cross-modal attention mechanism (Cross-Attention), the structured data of the sensor (power, current, vibration, temperature, etc. time series) and the feature association mapping of the unstructured data are established, and a full-dimensional feature matrix containing the running state of the coal machine equipment, historical fault patterns, and maintenance knowledge is constructed. Here, it can be seen that the specific implementation process of the coal machine equipment multi-modal data fusion can include text vectorization → cross-modal feature alignment → feature matrix generation; the cross-modal attention mechanism uses specific parameter configuration and optimization methods in the coal machine equipment operation and maintenance scenario.

[0053] Here, a reinforcement learning engine can also be optionally introduced, which combines reinforcement learning algorithms with large language models to modify the model training with actual coal machine equipment operation and maintenance results (such as but not limited to whether a fault occurs, the effectiveness of maintenance measures, etc.) as feedback signals to dynamically adjust the prediction parameters of the coal machine fault dynamic prediction model, achieving self-adaptability of the model to complex working conditions in the coal industry without frequent manual adjustment of feature engineering, and achieving continuous self-optimization. Furthermore, based on the continuous learning mechanism of the coal machine equipment operation and maintenance results feedback, the parameters of the large language model or the maintenance knowledge base are updated through incremental learning technology, so that the system and maintenance strategy dynamically evolve with the aging of the coal machine equipment and changes in working conditions, forming a "prediction-execution-optimization" closed loop to achieve decision-making closed loop optimization. In addition, feedback signals can be defined and quantified by fault prediction error rate, maintenance efficiency improvement rate, etc.; the trigger conditions and execution process of model self-updating can be timed updating or real-time feedback triggering.

[0054] As shown in the figure, the coal machine fault dynamic prediction model can input the real-time state feature matrix of the coal machine equipment based on the sequence generation capability of the large language model, and output the fault probability distribution of the key components in the future period. For example, in some embodiments, based on the sequence generation capability of the large language model, the current state feature matrix and historical operation data of the equipment are input, and the fault probability distribution of each key component such as but not limited to bearings, gearboxes, etc. within the next 72 hours is output.

[0055] As for the coal machine knowledge reasoning module, its role is to use stored knowledge (structured or unstructured) for logical reasoning, derive new knowledge, solve complex problems or make intelligent decisions. For example, in the embodiment of the present application, the maintenance specifications in the equipment manual (such as but not limited to lubrication period, component replacement standard, etc.) can be parsed using a large language model, combined with real-time fault prediction results, i.e. the fault probability distribution obtained by the coal machine fault dynamic prediction model, to generate technical points of the maintenance task, such as repair steps, spare part models, etc.

[0056] In the illustrated example, the intelligent decision output layer can include a coal machine fault maintenance strategy generator and a coal machine operation and maintenance work order system. For example, the coal machine fault maintenance strategy generator can be used to generate a maintenance strategy based on the technical points of the maintenance task. In some embodiments, the maintenance strategy can include an optimal maintenance time window, and the maintenance strategy can be divided into emergency repair, planned maintenance and state monitoring through the maintenance time window. The coal machine operation and maintenance work order system is used to generate an operation and maintenance work order based on the maintenance strategy, and the intelligent decision output layer interfaces with the operation and maintenance personnel mobile terminal to push the operation and maintenance work order to it. The intelligent decision output layer can or can not include the APP of the operation and maintenance personnel mobile terminal, i.e. the APP of the mobile terminal can be classified as an internal module of the intelligent decision output layer, or it can also be classified as an external module of the intelligent decision output layer. In optional embodiments, the operation and maintenance work order can also be pushed to other forms of terminals.

[0057] The intelligent decision output layer can calculate the optimal maintenance time window based on the coal machine equipment fault probability, equipment load, production plan and other parameters through a greedy algorithm, can distinguish between "emergency repair", "planned maintenance" and "state monitoring" three types of tasks, and can output a work order containing maintenance time, component location and operation guide through the coal machine operation and maintenance work order system, and can be pushed to the operation and maintenance personnel mobile terminal at the same time.

[0058] The greedy algorithm is an algorithm strategy that chooses the best or optimal (i.e. most advantageous) choice in each step, hoping to lead to a globally best or optimal result. Unlike dynamic programming, the greedy algorithm does not backtrack or reconsider previous choices, nor does it sacrifice the local optimal solution for long-term interests. It is based on the assumption that through local optimal selection at each step, a global optimal solution can be achieved.

[0059] In summary, the coal machine equipment multi-modal data intelligent preventive maintenance system of the present application can include three major modules, i.e. a multi-source data acquisition layer, a large language core processing layer and an intelligent decision output layer.

[0060] In the multi-source data collection layer, vibration sensors, temperature sensors, and pressure sensors are installed and deployed on coal machine equipment to collect real-time operation data (structured data). At the same time, the coal mine operation and maintenance management system is connected to obtain historical maintenance logs and equipment fault reports. Combined with the corresponding type of coal machine equipment operation manual, operation and maintenance manual, and other text data (unstructured data), ETL tools are used for data cleaning, desensitization, and unified storage in a distributed database.

[0061] In the large language model core processing layer, it includes: 1) a multi-modal data fusion module that standardizes sensor time series data and converts it into a vector sequence; uses a text encoder of a large language model to vectorize coal machine equipment maintenance logs, device field instrument photos, etc., and extracts semantic features such as fault description and maintenance measures; establishes an association mapping between structured data and unstructured data through a cross-modal attention mechanism, and constructs a coal machine equipment state feature matrix. 2) A dynamic prediction model based on the sequence generation capability of a large language model, which inputs the current state feature matrix and historical operation data of the equipment, and outputs the fault probability distribution of each key component (such as bearings and gearboxes) within the next 72 hours; introduces a reinforcement learning (RL) algorithm to dynamically adjust model parameters based on actual operation and maintenance feedback (such as whether a fault occurs or the maintenance effect), and realizes continuous self-optimization. 3) A knowledge reasoning module that uses a large language model to analyze maintenance specifications (such as lubrication period and component replacement standards) in the equipment manual, and generates technical points (such as maintenance steps and spare part models) for maintenance tasks based on real-time fault prediction results.

[0062] In the intelligent decision output layer, based on coal machine equipment fault probability, equipment load, production plan, etc., the optimal maintenance time window is calculated through a greedy algorithm, and three types of tasks are distinguished: "emergency maintenance", "planned maintenance", and "state monitoring"; output a work order containing maintenance time, component location, and operation guide, and push it to the operation and maintenance personnel mobile terminal.

[0063] Figure 2 A schematic flowchart of an embodiment of the coal machine equipment multi-modal data intelligent preventive maintenance method according to the present application. The method can be performed using the coal machine equipment multi-modal data intelligent preventive maintenance system of any of the preceding embodiments.

[0064] As shown in the figure, the coal machine equipment multi-modal data intelligent preventive maintenance method can include a data input stage, a large language model processing stage, a strategy output stage, and a feedback optimization closed loop stage.

[0065] In the data input stage, the collection and timestamp alignment of structured data and unstructured data are performed to form a multi-modal data set. The data input stage can be implemented by the multi-source data acquisition layer of the coal machine equipment multi-modal data intelligent preventive maintenance system, which receives structured data and unstructured data from coal machine equipment sensors and coal mine operation and maintenance management systems, cleans the data, and stores it in a distributed database for use in the next stage of large language model processing.

[0066] In the data input stage, real-time working condition data of coal machine equipment is collected in a timely manner, and historical operation and maintenance text data is imported in batches periodically, e.g., daily, to form a timestamp-aligned multi-modal data set.

[0067] Real-time working condition data collection is performed by sensors (such as but not limited to vibration sensors, temperature sensors, pressure sensors, etc.) deployed on coal machine equipment to collect structured data of coal machine equipment at a fixed sampling frequency (unit: Hz) in a timely manner, forming numerical time series wherein is the timestamp, is the sensor feature vector .

[0068] Historical operation and maintenance text import is the batch import of historical operation and maintenance records (such as fault description, maintenance log, maintenance report, etc.) and other unstructured data, and the unstructured data is denoted as wherein is the timestamp of the text and other unstructured data, is the text string of the unstructured data.

[0069] Timestamp alignment can use the bilinear interpolation method to align the timestamps of multi-modal data, generating a unified time axis wherein for any time point , the aligned value of structured data of sensors, etc. is , and unstructured data can be aligned to T by matching wherein , thereby forming an aligned multi-modal data set .

[0070] In the large language model processing stage, the technical points of cross-modal feature fusion of structured data and unstructured data, generation of fault prediction results and generation of maintenance tasks in the large language model processing stage. The large language model processing stage can be implemented by the large language model core processing layer of the coal machine equipment multi-modal data intelligent preventive maintenance system. According to the cross-modal attention mechanism of multi-modal data fusion and the optimization parameters of the reinforcement learning engine, the fault probability is predicted by the coal machine fault dynamic prediction model, and the technical points of the maintenance strategy are inferred by the coal machine knowledge reasoning module based on the fault probability. The next strategy output stage receives application. These technical points can include but are not limited to repair steps, spare parts models, etc.

[0071] The large language model processing stage can include step A of cross-modal feature fusion, step B of fault probability sequence generation, and step C of maintenance specification knowledge matching.

[0072] Step A of cross-modal feature fusion can include: converting unstructured data into semantic vectors by the Encoder of the large language model, splicing with the numerical vectors of structured data, and inputting the Transformer network for cross-modal feature extraction, wherein a pre-trained large language model is used as the Encoder to encode the unstructured data , and output text embedding vectors , wherein is the model hidden layer dimension, the structured data is standardized by sensor numerical coding , is the historical mean and standard deviation, and is mapped to a numerical vector by a fully connected layer , wherein is a learnable parameter, and is spliced with to , input the Transformer network for feature interaction, realize cross-modal feature extraction, and the multi-head attention mechanism of the Transformer network is calculated as follows: , wherein each head , the attention function is , Q is the query, K is the key, V is the value, is generated by , is the head-related parameter, is the output projection matrix, and the cross-modal feature is output.

[0073] The steps for generating the fault probability sequence (i.e., the steps for predicting coal mining equipment faults) B may include: generating a fault probability sequence for future time periods using the decoder of a large language model. This fault probability sequence represents the fault probability at each moment within the time window. A sliding window technique (a dynamic adjustment strategy for window size and sliding step size, e.g., a 24-hour window) is used to achieve real-time rolling prediction. The input window is constructed using a sliding window of W = 24 hours, with a window step size S = 1 hour (i.e., updated every hour). For time point t, the input window is... Fault probability sequence generation involves inputting window features into the decoder of a large language model (such as the GPT architecture) and outputting future... Hourly Failure Probability Sequence ,in express The probability of failure at time step, the Decoder generation process satisfies ,in For Decoder in The hidden state at all times The classification layer parameters (the second dimension corresponds to the probability of the "fault" category) are used. The model training loss adopts the binary cross-entropy loss function, and the supervision label is the historical fault marker. Where 0 represents normal and 1 represents fault, and the loss function is... ,

[0074] The steps for matching maintenance specifications knowledge (i.e., coal mining machinery operation and maintenance knowledge matching steps) C may include: based on the coal mining machinery maintenance specifications knowledge base, matching the maintenance rules corresponding to the current state of the coal mining machinery equipment through the retrieval-Augmented Generation mechanism of a large language model, wherein the coal mining machinery maintenance specifications knowledge base is constructed. ,in Keywords for knowledge entries For the corresponding maintenance rules, the cross-modal characteristics of the current coal mining machine status will be... As the query vector, its embedding vector with the keywords of each entry in the knowledge base is calculated. cosine similarity Select the first one with the highest similarity One entry As a search result,

[0075] Maintenance rule generation is to Description of the current state Concatenate as prompt words Input a large language model to generate the final maintenance rules. .

[0076] In step D of the strategy output stage, high-risk failure prediction results are prioritized in combination with the shutdown window of the production scheduling system to generate maintenance work orders containing three levels of priority (red / yellow / green) and trigger coal machine equipment failure-related spare parts inventory warnings in synchronization, wherein

[0077] Priority classification rules are based on failure probability Three levels of priority are divided:

[0078]

[0079] Among them , Thresholds can be optimized according to historical data

[0080] Shutdown window matching is achieved by obtaining the shutdown time window of the production scheduling system , and the red priority failure is assigned to the nearest available window ,

[0081] Spare parts inventory warning is determined according to the failure type (determined by corresponding failure category), query spare parts inventory , if , wherein is the average spare parts consumption of the same type of failure in history, is the safety factor, then trigger the warning .

[0082] Here, the maintenance strategy of the coal machine equipment is generated through multi-objective algorithms (including priority calculation model, shutdown window matching rules, etc.), which advantageously achieves the optimal maintenance strategy result.

[0083] In the strategy output stage, the production shutdown window is obtained and the maintenance strategy is generated, and the operation and maintenance work order is generated and pushed. The strategy output stage can be realized by the intelligent decision output layer of the coal production device multi-modal data intelligent preventive maintenance system, which generates the coal machine failure maintenance strategy according to the technical points of the previous step, generates the operation and maintenance work order by the coal machine operation and maintenance work order system, and pushes it to the mobile terminal APP of the operation and maintenance personnel. Here, the intelligent generation logic of the maintenance work order content through the mapping relationship between the technical points and the state of the coal machine equipment is creatively adopted.

[0084] In the feedback optimization closed loop stage, the operation and maintenance results are collected and the model parameters are optimized. The feedback optimization closed loop stage can be realized by the reinforcement learning engine of the large model core processing layer of the coal production device multi-modal data intelligent preventive maintenance system.

[0085] Some embodiments of the present application can: deeply utilize unstructured data: apply large language models to coal industry coal machine equipment operation and maintenance text analysis, breaking through the limitations of traditional models relying only on sensor data; adopt a dynamic adaptive framework: through the combination of reinforcement learning and large language models, realize the adaptability of the model to the complex working conditions of the coal industry, without frequent manual adjustment of feature engineering; closed-loop optimization decision: based on the continuous learning mechanism of the feedback of the coal machine equipment operation and maintenance results, the maintenance strategy evolves dynamically with the aging of the equipment and the change of the working conditions, forming a "prediction-execution-optimization" closed loop.

[0086] Some embodiments of the coal machine equipment multi-modal data intelligent preventive maintenance system and method of the present application strengthen the "use of unstructured data", "improve fault prediction accuracy" and "dynamic maintenance strategy" in coal machine equipment operation and maintenance, realize multi-modal data deep analysis, accurate fault prediction and intelligent maintenance decision-making, and its beneficial effects can include:

[0087] (1) Full-dimensional data fusion: through large language models to process text data such as coal machine equipment maintenance logs, operation manuals, and operation manuals, combined with sensor time series data, fully excavate equipment operation and maintenance knowledge, and build an equipment operation full-life cycle knowledge base, filling the gap in the application of unstructured data;

[0088] (2) Adaptive prediction ability: using the context understanding and generation ability of large language models, dynamically learning the abnormal patterns of coal machine equipment under different working conditions, enhancing the generalization ability of the model, adapting to the complex working conditions such as dust and vibration in the coal production environment, improving the fault prediction accuracy in complex environments, and reducing the false negative rate by more than 40%;

[0089] (3) Intelligent decision optimization: based on real-time state and remaining life prediction of coal machine equipment, rather than fixed threshold or periodic planning to develop maintenance solutions, generating personalized maintenance strategies (such as maintenance priority, spare parts inventory recommendations), improving operation efficiency by 30%, reducing downtime by 25%, and reducing maintenance costs by 20%.

[0090] The application range of the coal machine equipment multi-modal data intelligent preventive maintenance system and method of the present application can cover the full life cycle management of heavy equipment such as coal mining machines, heading machines, scraper conveyors, and crushers in coal mines.

[0091] The technical scope of the present application is not limited to the content in the above description, and those skilled in the art can make various modifications and changes to the above embodiments without departing from the technical idea of the present application, and these modifications and changes should all be within the scope of the present application.

Claims

1. A multi-modal data-driven intelligent preventive maintenance system for coal mining equipment, characterized in that: The system includes: A multi-source data acquisition layer includes a data cleaning module and a distributed database. The data cleaning module interfaces with equipment sensors and a coal mine operation and maintenance management system. The equipment sensors provide the data cleaning module with real-time structured operational data of the coal mining equipment, while the coal mine operation and maintenance management system provides the data cleaning module with unstructured data related to the coal mining equipment. The data cleaning module stores the cleaned and de-identified structured and unstructured data as a multimodal dataset in the distributed database. The core processing layer of the large language model includes a coal mining machinery fault dynamic prediction model and a coal mining machinery knowledge reasoning module. The coal mining machinery fault dynamic prediction model is connected to a multimodal data fusion module and a reinforcement learning engine. The multimodal data fusion module establishes an association mapping between the structured data and the unstructured data and constructs a coal mining machinery equipment state feature matrix. The reinforcement learning engine is used to optimize the model parameters of the coal mining machinery fault dynamic prediction model. The coal mining machinery fault dynamic prediction model is used to generate fault prediction results based on the coal mining machinery equipment state feature matrix and the optimized model parameters of the coal mining machinery fault dynamic prediction model. The coal mining machinery knowledge reasoning module generates the technical points of the maintenance task based on the unstructured data and the fault prediction results. The intelligent decision output layer includes a coal mining machine fault maintenance strategy generator and a coal mining machine operation and maintenance work order system. The coal mining machine fault maintenance strategy generator is used to generate maintenance strategies based on the technical points of the maintenance task. The coal mining machine operation and maintenance work order system is used to generate operation and maintenance work orders based on the maintenance strategies. The intelligent decision output layer connects to the mobile terminal of operation and maintenance personnel and pushes the operation and maintenance work orders to them. The core processing layer of the large language model is used for cross-modal feature fusion, fault probability sequence generation, and maintenance specification knowledge matching. The cross-modal feature fusion process includes: converting the unstructured data into semantic vectors using the large language model's encoder, concatenating these vectors with the numerical vectors of the structured data, and inputting the concatenated vectors into a Transformer network for cross-modal feature extraction. A pre-trained large language model is used as the encoder to process the unstructured data. Encode and output text embedding vectors ,in For the hidden layer dimension of the model, sensor numerical encoding is used to process structured data. Standardization processing The historical mean and standard deviation are mapped to numerical vectors through a fully connected layer. ,in For learnable parameters, and spliced ​​as The input is processed through a Transformer network to achieve cross-modal feature extraction. The multi-head attention mechanism of the Transformer network is calculated as follows: Each of the heads The attention function is Q represents the query, K represents the key, and V represents the value. generate, For header-related parameters, To output the projection matrix and output cross-modal features. ; The process of generating the fault probability sequence includes: generating a fault probability sequence for future time periods using the decoder of a large language model; and employing a sliding window technique to achieve real-time rolling prediction. The input window is constructed using a sliding window of W = 24 hours with a window step size S = 1 hour. For time point t, the input window is... Fault probability sequence generation involves inputting window features into the decoder of a large language model and outputting future... Hourly Failure Probability Sequence ,in express The probability of failure at time step, the Decoder generation process satisfies ,in For Decoder in The hidden state at all times For the classification layer parameters, the model training loss uses the binary cross-entropy loss function, and the supervision label is the historical fault marker. Where 0 represents normal and 1 represents fault, and the loss function is... ; The maintenance specification knowledge matching process includes: based on the coal mining machinery maintenance specification knowledge base, matching the maintenance rules corresponding to the current state of the coal mining machinery equipment through a large language model retrieval enhancement mechanism, wherein the coal mining machinery maintenance specification knowledge base is constructed. ,in Keywords for knowledge entries For the corresponding maintenance rules, the cross-modal characteristics of the current coal mining machine status will be... As the query vector, its embedding vector with the keywords of each entry in the knowledge base is calculated. cosine similarity Select the first one with the highest similarity One entry As a search result, Maintenance rule generation is to Description of the current state Concatenate as prompt words Input a large language model to generate the final maintenance rules. .

2. The system as described in claim 1, characterized in that, The structured data includes current signals, power signals, vibration signals, temperature signals and / or speed signals, while the unstructured data includes coal mining equipment operation manuals, maintenance logs and / or inspection records.

3. The system as described in claim 1, characterized in that, The multimodal data fusion module standardizes the structured data and converts it into a vector sequence; it uses a text encoder based on a large language model to vectorize the unstructured data and extracts semantic features related to fault descriptions and maintenance measures. Then, it establishes an association mapping between the structured data and the unstructured data through a cross-modal attention mechanism.

4. The system as described in claim 1, characterized in that, The coal mining machine fault dynamic prediction model takes into account the state feature matrix of the coal mining equipment and the historical operating data of the coal mining equipment. Based on the sequence generation capability of the large language model, it outputs the fault probability distribution of the components of interest within a preset time period. Furthermore, it introduces the reinforcement learning engine to dynamically adjust the model parameters of the coal mining machine fault dynamic prediction model according to actual operation and maintenance feedback, thereby achieving continuous self-optimization.

5. The system as described in claim 1, characterized in that, The maintenance strategy includes an optimal maintenance time window, which is divided into emergency repair, planned maintenance, and status monitoring based on the maintenance time window.

6. A method for multimodal data-driven intelligent preventive maintenance of coal mining equipment using the multimodal data-driven intelligent preventive maintenance system for coal mining equipment as described in any one of claims 1 to 5.

7. The method as described in claim 6, characterized in that, The method includes: In the data input phase, the structured data and the unstructured data are collected and timestamp aligned to form the multimodal dataset. In the large language model processing stage, the key technical points of generating the maintenance task are to perform cross-modal feature fusion, fault probability sequence generation, and maintenance specification knowledge matching on the structured data and the unstructured data. In the strategy output phase, the production shutdown window is obtained and the maintenance strategy is generated, and the operation and maintenance work order is generated and pushed. In the feedback optimization closed-loop stage, operation and maintenance results are collected and the model parameters are tuned.

8. The method as described in claim 7, characterized in that, In the strategy output phase, combined with the downtime window of the production scheduling system, high-risk fault prediction results are prioritized, generating maintenance work orders with three priority levels, and simultaneously triggering early warnings for spare parts inventory related to coal mining equipment faults. Priority grading rules are based on the aforementioned failure probability. Three levels of priority: in , , Downtime window matching is achieved by obtaining the downtime window from the production scheduling system. And the red priority fault is assigned the nearest available window. , Spare parts inventory warning based on fault type Check spare parts inventory ,like ,in This represents the average spare parts consumption for similar historical failures. To ensure a safety margin, an early warning will be triggered. .

Images