Method for carrying out mine operation and maintenance by adopting AI large model

Abstract

The invention discloses a method for carrying out mine operation and maintenance by adopting an AI large model. The method comprises the following steps: 1, carrying out model design according to the requirements of a mine operation and maintenance scene: determining a model structure by adopting a recurrent neural network (RNN) algorithm, and then carrying out feature processing on model parameters; 2, according to a mine operation and maintenance scene, building an AI large model architecture by using a multi-layer Transform encoder; 3, training the established AI large model to obtain a trained AI large model; 4, performing fine tuning on the trained AI large model by utilizing a Fine-tuning technology so as to adapt to a specific natural language processing task; and 5, perfectly establishing a knowledge base client system for multi-terminal online question and answer. According to the method for carrying out mine operation and maintenance by adopting the AI large model provided by the invention, the problems of lack of systematic operation and maintenance management tools, lack of effective technical support means and difficulty in positioning and searching operation and maintenance problems in the existing operation and maintenance technology are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine operation and maintenance, and relates to a method for mine operation and maintenance using an AI large model. Background Art

[0002] Coal mine intelligentization is a new form of coal mine industrial production in China and also the basic support for the sustainable and steady development of China's coal mine industry. Coal mine intelligentization integrates a number of advanced scientific and technological means such as big data information technology, Internet of Things technology, artificial intelligence technology, and computer network technology, making the coal resource extraction process more automated, intelligent, and scientific, realizing the whole-process control and full-process management of the coal resource extraction process, greatly improving the extraction efficiency of coal resources, and minimizing the possibility of safety accidents to ensure the safety of coal resource extraction.

[0003] With the gradual deep integration of industrialization and informatization in the coal industry, it has become an urgent need to use advanced information technology to solve the problems in the intelligent and smart transformation of the coal industry. The centralized online operation of intelligent system equipment has led to a disconnection between the use and operation and maintenance of intelligent equipment. With the normal use of the system, the continuous production efficiency of the mine is seriously insufficient, and the equipment is scattered and independent, resulting in an explosive increase in the maintenance volume of various systems; there are many monitoring devices and sensors for various types of equipment, and the professional requirements for maintenance work are high. The mine side lacks its own professional operation and maintenance personnel and can only rely on system manuals or maintenance manuals for operation and maintenance, with a high degree of dependence on manufacturer's technical personnel. On-site technical personnel are difficult to proficiently operate the system and put forward constructive improvement opinions, and cannot well meet the needs of the rapid advancement of intelligent systems.

[0004] The existing technologies have the following defects and deficiencies: (1) For the centralized online intelligent system equipment, there is a lack of systematic operation and maintenance management tools; (2) The technical levels of operation and maintenance personnel vary, and the maintenance technologies for various systems and equipment that need to be mastered are complex and diverse. When a failure occurs, there is a lack of effective technical support means; (3) Technical documents are mostly in the form of offline documents or product manuals, etc., and it is not easy to locate and find operation and maintenance problems.

[0005] Based on the above problems and deficiencies in mine operation and maintenance during the development of coal mine intelligentization, by using AI technology, an application technology research on the use of an AI large model in mine operation and maintenance is proposed. The aim is to combine the current difficulties and technical barriers in mine operation and maintenance, and through training an operation and maintenance technology knowledge base with artificial intelligence for autonomous update, and in the way of NLP tasks, efficiently solve the thorny problems encountered in mine operation and maintenance. Summary of the Invention

[0006] The objective of the present invention is to provide a method for mine operation and maintenance using an AI large model, which solves the problems in existing operation and maintenance technologies, such as the lack of a systematic operation and maintenance management tool, the lack of effective technical support means, and the difficulty in locating and finding operation and maintenance problems.

[0007] The technical solution adopted by the present invention is a method for mine operation and maintenance using an AI large model, and the specific steps are as follows: Step 1: Conduct model design according to the requirements of the mine operation and maintenance scenario: Use the Recurrent Neural Network (RNN) algorithm to determine the model structure, and then perform feature processing on the model parameters; Step 2: Build an AI large model architecture using a multi-layer Transformer encoder according to the mine operation and maintenance scenario; Step 3: Train the built AI large model to obtain a trained AI large model; Step 4: Use the Fine-tuning technology to fine-tune the trained AI large model to adapt to specific natural language processing tasks; Step 5: Improve and establish a knowledge base client system for multi-terminal online Q&A.

[0008] The characteristics of the present invention also lie in: The feature processing in Step 1 specifically translates the original data of the model parameters into a form understandable by the model, including feature extraction, feature selection, feature combination, and feature transformation; Feature extraction specifically extracts the mapping relationship between operation and maintenance content and operation and maintenance solutions in the model, and improves the mapping relationship by introducing an attention weight mechanism to obtain a better representation; Feature selection specifically uses statistical correlation analysis, recursive feature elimination, or embedded methods for feature selection to reduce the number of features, remove redundant features in the operation and maintenance model, and retain high-quality feature data, thereby improving the performance of the model; Feature combination specifically combines features using mathematical operations such as multiplication, addition, and exponentiation, automatically learns the complex relationships between features using a neural network, forms a new feature vector by combining multiple different features, and associates and combines each step required in the operation and maintenance process; Feature transformation specifically adjusts the distance between features in the operation and maintenance model using methods such as standardization, normalization, discretization, encoding, or dimensionality reduction.

[0009] The model parameters include the learning rate, batch size, and optimizer; The operation and maintenance content in feature extraction is the daily or special operation and maintenance content required for personnel, equipment, etc. during the mine production process, specifically including equipment monitoring data, environmental parameters, and video images; The operation and maintenance plan is to provide a suitable maintenance plan according to the operation and maintenance content. The AI large model will generate different operation and maintenance plans according to different operation and maintenance contents and combinations. Both the operation and maintenance content and the operation and maintenance plan are formulated according to the safety production regulations.

[0010] The equipment monitoring data specifically refers to the operating parameters extracted from the sensors of the mining equipment; the environmental parameters are the temperature, humidity and gas concentration in the mining environment; the video images specifically refer to the images or videos captured by the cameras at the mining operation site, and the key features extracted through image processing algorithms, including open flames, smoke, and abnormal personnel behavior.

[0011] The specific method of feature selection is as follows: Evaluate the importance of features according to the correlation between features and target variables, the redundancy between features, and the degree of improvement of features on model performance; and use automated feature engineering technology to simplify the feature selection process, improve efficiency and accuracy, remove some redundant features in the operation and maintenance model, reduce computational complexity, and improve the generalization ability of the model.

[0012] The specific method of step 2 is as follows: Step 2.1: First, calculate the weight coefficient using the query Query and the key Key, and then normalize the weights using the SoftMax operation to obtain softmax(f(Q, K)): (1) In the formula, Q is the query for model elements, K is the key of model elements, W is the first weight coefficient, U is the second weight coefficient, tanh() is the cosine similarity function; (2) In the formula, Softmax() is the weight normalization function; Step 2.2: Perform a weighted sum on the value Value to calculate the output of the attention: (3) In the formula, Attention() is the attention function, α is the attention weight parameter, V is the value corresponding to the model prototype; Step 2.3: Introduce the "encoder-decoder" framework into the attention mechanism, use the encoder to encode the input sequence, first obtain the semantic vector, and then obtain the weighted sum of all semantic vectors, which is called the context vector. Using the weight parameter representing the attention, the context vector can be expressed as: (4) In the formula, c is the context vector, α is the attention weight parameter, h is the semantic vector of the input sequence; During the decoding process, the probability of each output word is jointly determined by the context vector and the hidden state of the previous layer: (5) (6) In the formula, p and g respectively represent non-linear transformations; s represents a word in the output sequence, y represents the corresponding hidden state, c represents the context vector; in addition, the attention weight parameter is calculated using another neural network: (7) The model with AI computing power is built.

[0013] The specific method of step 3 is as follows: The training of the AI large model is to adjust and determine the ideal values of all weights Weights and biases Bias through labeled samples; Take one or more features as input and then return a prediction as output; for simplicity, a model that takes one feature and returns one prediction is used, as shown in the following formula: (8) In the formula b is the Bias, w is the Weights; Through the loss function, calculate the loss loss of Bias and Weights under this set of parameters; use the gradient descent method to find the direction with smaller loss and iterate to obtain the trained AI large model.

[0014] The specific method of step 4 is as follows: Copy the source model of the AI large model created in step 2 to create a new neural network model, that is, the target model; the target model copies all the model designs and their parameters on the source model except the output layer; Assume that the model parameters contain the knowledge learned from the source dataset, and this knowledge is also applicable to the target dataset; assume that the output layer of the source model is closely related to the labels of the source dataset, so it is not adopted in the target model; add an output layer with an output size equal to the number of categories in the target dataset to the target model, and randomly initialize the model parameters of this layer. Train the target model on the target dataset, retrain the output layer, while the parameters of the remaining layers are fine-tuned based on the parameters of the source model; Fine-tuning specifically includes reusing the classifier weights, and there are also some labels in the source data that are in the target data; use the vectors corresponding to the labels in the pre-trained model classifier as the initial values; secondly, the neural network learns hierarchical feature representations; fix the relatively bottom layers and do not participate in parameter updates.

[0015] The specific method of step 5 is as follows: Step 5.1, Model training process: Through the data collection system, integrate multi-source data such as web, files, manual operations, and databases, use the AI engine to build the basic knowledge base system, and at the same time use these data resources as the data materials for model training to perform pre-training of the model; Step 5.2, Data retrieval process: During the operation and maintenance process, when the user encounters a problem, a retrieval request is initiated from the client. If there is a retrieval result in the prefabricated library, it is directly fed back to the user. If there is no matching retrieval result, then the retrieval is performed through the AI model, and the retrieval result is fed back to the user through the engine. After the user successfully processes the problem through the feedback result, this processing record will also be recorded into the model library as the data material for model training to iteratively update the model; at the same time, the result will also be stored in the prefabricated library module for more efficient and convenient finding of the result next time.

[0016] The beneficial effects of the present invention are: (1) The method of using the AI large model for mine operation and maintenance in the present invention innovatively uses the AI large model as the technical cornerstone. In view of the complex current situation and technical challenges in the field of mine operation and maintenance, by constructing an operation and maintenance technology knowledge base with autonomous update and intelligent decision-making, through the NLP task method, it has realized a comprehensive innovation in operation and maintenance management, not only solving many problems in mine operation and maintenance, but also leading a new trend of the intelligent transformation of mine operation and maintenance; (2) The method of using the AI large model for mine operation and maintenance in the present invention has unified modeling and centralized management of the operation and maintenance models of multiple mine systems, completely breaking the previous situation where the operation and maintenance of each system were independent and lacked coordination. This change makes the operation and maintenance management more efficient and orderly, and greatly improves the operation and maintenance efficiency and quality; (3) The method of using the AI large model for mine operation and maintenance in the present invention, AI autonomous update and experience iteration: By establishing an AI autonomous update model, the present invention integrates operation and maintenance experience into the model for iterative upgrade, completely changing the traditional mode that relied on the word-of-mouth transmission of operation and maintenance personnel in the past. This innovative mechanism not only accelerates the dissemination and accumulation of operation and maintenance knowledge, but also ensures the continuous progress and optimization of operation and maintenance technology; (4) The method of using the AI large model for mine operation and maintenance in the present invention realizes precise retrieval and efficient feedback. With the convenient and precise retrieval function of multiple clients, the present invention can quickly locate and solve fault problems, directly feedback the operation and maintenance results to users, significantly improving the solution efficiency of operation and maintenance problems. This function not only reduces the work burden of operation and maintenance personnel, but also enhances the satisfaction and trust of users; (5) The method of using the AI large model for mine operation and maintenance in the present invention realizes online management and resource conservation. Through the online client, the present invention avoids the storage requirements of a large number of offline documents such as technical documents, user manuals, or equipment maintenance manuals, effectively saving resource costs. This transformation not only conforms to the sustainable development concept of green and environmental protection, but also brings a more convenient and efficient online experience for mine operation and maintenance management. Brief Description of the Drawings

[0017] Figure 1 It is a working mode diagram of the long short-term memory network in step 1 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 2 It is a schematic diagram of the Transformer network structure in step 2 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 3 It is a basic framework diagram of the encoder-encoder in step 2 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 4 It is a schematic diagram of the principle of the attention mechanism in step 2 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 5 It is a schematic diagram of the model training structure in step 3 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 6 It is a model training flowchart in step 5 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 7 It is an example diagram of the application effect of the model in the client in step 5 of the method of using the AI large model for mine operation and maintenance in the present invention; Figure 8 It is a schematic diagram of the business application process in step 5 of the method of using the AI large model for mine operation and maintenance in the present invention. Detailed Embodiments

[0018] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0019] The method of the present invention for using an AI large model for mine operation and maintenance is as follows: Step 1: Design the model according to the requirements of the mine operation and maintenance scenario: Use the Recurrent Neural Network (RNN) algorithm to determine the model structure, and then perform feature processing on the model parameters; The design of this model closely revolves around the specific business requirements of the operation and maintenance scenario, and uses the Recurrent Neural Network (RNN) algorithm as the core for constructing the model structure. RNN, with its unique directional cyclic connection feature, can feed the output of the Long Short-Term Memory network (LSTM) as the input of the current stage. This mechanism enables the model to fully utilize and remember past input information.

[0020] As Figure 1 shown, the working mechanism of RNN can be detailedly decomposed into four key steps: 1. Time-delayed input: The output at time t-1 is processed as the input at time t. This step ensures that the model can make predictions or decisions at the current moment based on the previous state; 2. Recursive iteration: Similarly, the output at time t is used as the input at time t+1. This recursive iteration process enables RNN to process data with time series characteristics, thereby capturing the dynamic change rules in the data; 3. Processing of arbitrarily long inputs: One of the core advantages of RNN is its ability to process input sequences of any length. This feature enables the model to handle various complex and changeable input situations in the operation and maintenance scenario without the need for additional preprocessing or truncation of the input data; 4. Fusion of historical information: During the calculation process, RNN fully considers the influence of historical information. By integrating past input information into the calculation at the current moment, the model can more accurately understand and predict the dynamic changes in the operation and maintenance scenario.

[0021] In addition, it is worth noting that the size of the RNN model does not increase with the increase in the input size. This feature ensures the efficiency and stability of the model when processing large-scale data.

[0022] The model structure mainly consists of the following parts: (1) Input layer: Receives data points arranged in chronological order (equipment operation status, energy consumption data, personnel location information, environmental monitoring data) as input; the input data undergoes preprocessing steps, including standardization, normalization, and encoding, to adapt to the requirements of the model; (2) Hidden layer (RNN / LSTM / GRU layer): Each RNN unit is responsible for processing the input at one time step; each unit has a memory mechanism inside to control the flow and storage of information. The hidden layer can capture the temporal dependencies and long-term trends in the input data because the output of each unit depends not only on the input at the current time step but also on the hidden state (or memory) of the previous time step; (3) Output layer: Generates outputs according to specific business requirements, such as equipment failure warnings, personnel scheduling suggestions, and safety risk assessments. The output undergoes post-processing steps and is inverse-normalized into a readable format; (4) Training and evaluation: The model is trained using a labeled training dataset through the backpropagation algorithm to minimize the prediction error; the performance of the model is evaluated using an independent test dataset, and the metrics are accuracy and recall; (5) Application deployment and monitoring: The trained RNN model is deployed into the coal mine operation and maintenance system to achieve the processing and analysis of real-time data; according to the output results of the model, corresponding operation and maintenance strategies are formulated, such as equipment maintenance plans, personnel scheduling plans, and safety warning measures; continuously monitor the model performance, collect feedback data in a timely manner, and iterate and optimize the model.

[0023] After determining the model structure, feature processing is performed on the model parameters. Among them, the model parameters include the learning rate, batch size, and optimizer. The feature processing specifically translates the original data of the model parameters into a form that the model can understand, including feature extraction, feature selection, feature combination, and feature transformation; Furthermore, feature extraction specifically extracts the mapping relationship between the operation and maintenance content and the operation and maintenance plan in the model, and improves the mapping relationship by introducing an attention weight mechanism to obtain a better representation; Among them, the operation and maintenance content is the daily or special operation and maintenance content required for personnel, equipment, etc. during the mine production process, specifically including: (1) Equipment monitoring data: Specifically, the operation parameters extracted from the sensors of mine equipment; (2) Environmental parameters: Temperature, humidity, and gas concentration in the mine environment; (3) Video images: Images or videos of the mine operation site captured by cameras, and the key features extracted through image processing algorithms, including open flames, smoke, and abnormal personnel behavior.

[0024] The operation and maintenance plan is to give a suitable maintenance plan according to the operation and maintenance content. The AI large model will generate different operation and maintenance plans according to different operation and maintenance content and combinations. Both the operation and maintenance content and the operation and maintenance plan are formulated according to the safety production regulations, and there is a large amount of content, which will vary slightly according to the actual situation of different mines.

[0025] Feature selection specifically involves using statistical - based correlation analysis, recursive feature elimination, or embedded methods for feature selection, and evaluating the importance of features according to the correlation between features and target variables, the redundancy between features, and the criterion of the degree of improvement of features on model performance; and using automated feature engineering technology (AutoML) to simplify the feature selection process, improve efficiency and accuracy, reduce the number of features, remove some redundant features in the operation and maintenance model, reduce computational complexity, and improve the generalization ability of the model; Feature combination specifically involves using mathematical operations such as multiplication, addition, exponentiation, etc. to combine features, and using neural networks to automatically learn the complex relationships between features. By combining multiple different features, a new feature vector is formed to associate and combine each step required in the operation and maintenance process; Feature transformation specifically involves using methods such as standardization, normalization, discretization, encoding, or dimensionality reduction to adjust the distance between features in the operation and maintenance model, as follows: (1) Standardization / Normalization: Scale the collected feature values to the same scale to ensure that different features have equal weights in model training. Standardization usually involves subtracting the mean from the feature values and dividing by the standard deviation, while normalization scales the feature values between 0 and 1; (2) Discretization: For continuous features, such as device temperature or energy consumption, they can be converted into discrete features to better capture their change trends or category information. Discretization can be achieved by setting thresholds or using clustering algorithms; (3) Encoding: Encode categorical features, converting device types, personnel identities into numerical features; (4) Dimensionality reduction: Use principal component analysis (PCA), linear discriminant analysis (LDA) to reduce the feature dimensions. Dimensionality reduction can help identify which features have the greatest impact on the operation and maintenance process.

[0026] Step 2: According to the mine operation and maintenance scenario, use a multi - layer Transformer encoder to build an AI large - model architecture, Such as Figure 2As shown in the figure, the main components of the Transformer model are the self-attention mechanism and the feedforward neural network. When using the self-attention mechanism, the model generates a vector representation of the same length as the sequence length based on the information at each position in the input sequence. This vector captures the relationship between each position in the input sequence and other positions, thus providing the model with a better way to understand the input information. In Transformer, the input sequence is composed of multiple encoders stacked together. In each encoder, the self-attention mechanism and the feedforward neural network form a block, and multiple blocks form the complete encoder. To maintain the information of the sequence, Transformer also uses an attention mechanism to transfer the information at each position in the input sequence to the output sequence. As Figure 3 shown, the concept of the encoder is that the input and output of the network are regarded as different sequences respectively, and then the mapping from sequence to sequence is realized. Its classic implementation structure is the "encoder-decoder" framework. For the input, the encoder will first traverse the input sequence. According to the characteristics of the recurrent neural network, the input of each layer contains the input of the current layer and the hidden state of the previous layer, and the new hidden state will be used as the input of the next layer. The encoder generally only retains the hidden state of the last layer as the semantic vector of the entire input and sends it into the decoder. After this semantic vector is input into the decoder, it will be regarded as the initial hidden state of the decoder. However, due to the structural characteristics of the recurrent neural network itself, if the input sequence is too long, the model performance will be significantly reduced. This is because the recurrent neural network only uses the hidden state of the last layer as the semantic vector representing the entire input text, and information loss will occur in the sequences at the front of the sentence. To solve this problem, the attention mechanism is introduced. The attention mechanism essentially draws on the ability of the human eye to process information visually and can be summarized into two stages: determining which part of the input needs special attention; then preferentially allocating resources to the important part. In the neural network, the attention mechanism can be understood as focusing on different features when predicting the result. Mathematically speaking, as Figure 4 shown, the calculation of the attention mechanism can be described as a mapping from a query (Query) to a series of key-value pairs (Key-Value), and the specific method is as follows: Step 2.1: First, calculate the weight coefficient using the query Query and the key Key, and then normalize the weights using the SoftMax operation to obtain softmax(f(Q, K)): (1) In the formula, Q is the query for the model element, K is the key of the model element, W is the first weight coefficient, U is the second weight coefficient, tanh() is the cosine similarity function; (2) In the formula, Softmax() is the weight normalization function; Step 2.2: Perform a weighted sum on the value Value to calculate the output of the attention: (3) In the formula, Attention() is the attention function, α is the attention weight parameter, V is the value corresponding to the model prototype; Step 2.3: Introduce the "encoder-decoder" framework into the attention mechanism. Use the encoder to encode the input sequence to first obtain semantic vectors, and then obtain the weighted sum of all semantic vectors, which is called the context vector. Using the weight parameter representing the attention, the context vector can be expressed as: (4) In the formula, c is the context vector, α is the attention weight parameter, h is the semantic vector of the input sequence; During the decoding process, the probability of each output word is jointly determined by the context vector and the hidden state of the previous layer: (5) (6) In the formula, p and g respectively represent non-linear transformations; s represents a word in the output sequence, y represents the corresponding hidden state, c represents the context vector; in addition, the attention weight parameter is calculated using another neural network: (7) Through the above steps, the model with AI computing power is built.

[0027] Step 3: Train the built AI large model to obtain the trained AI large model; The training of the model is to adjust (learn) and determine the ideal values of all weights Weights and biases Bias through labeled samples.

[0028] Take one or more features as input and then return a prediction as output; for simplicity, a model that takes one feature and returns one prediction is as follows: (8) In the formula b is the Bias, w is the Weights; Through the loss function, calculate the loss of Bias and Weights under this set of parameters; use the gradient descent method to find the direction with smaller loss and perform iteration to obtain the trained large AI model.

[0029] As Figure 5 shown, the loss is a numerical value indicating the accuracy of the model prediction for a single sample; if the model prediction is completely accurate, the loss is zero, otherwise the loss will be larger; the goal of training the model is to find a set of weights and biases with a "smaller" average loss from all samples; Since the method of finding the convergence point by calculating the loss function for each possible value in the entire dataset is too inefficient, we use the gradient descent method to confirm the loss.

[0030] The gradient descent method first selects a starting value, and then the algorithm calculates the gradient of the loss curve at the starting point; the gradient is a vector of partial derivatives, and the gradient of the loss with respect to a single weight is equal to the derivative. The gradient is a vector and thus has two characteristics: magnitude and direction, and the gradient always points in the direction where the loss function grows most rapidly; the gradient descent algorithm takes a step along the direction of the negative gradient to reduce the loss as quickly as possible; to determine the next point on the loss function curve, the gradient descent algorithm adds a fraction of the gradient magnitude to the starting point, and then the gradient descent method repeats this process, gradually approaching the lowest point; through continuous iteration of this method, the ideal values of all weights Weights and biases Bias are finally determined to achieve the effect of model training.

[0031] Step 4: Use the Fine-tuning technique to fine-tune the trained large AI model to adapt to specific natural language processing (NLP) tasks; During the Fine-tuning process, the performance of the model is evaluated using metrics such as accuracy and recall based on its performance on the test dataset. After multiple rounds of fine-tuning and performance evaluation, if the model prediction can accurately answer the key points of mine operation and maintenance on the test dataset, it represents the successful training of the model; Specific steps for fine-tuning: Copy the source model of the AI large model created in step 2 to create a new neural network model, i.e., the target model; The target model copies all the model designs and their parameters of the source model except the output layer; Assume that the model parameters contain the knowledge learned from the source dataset, and this knowledge is also applicable to the target dataset; Assume that the output layer of the source model is closely related to the labels of the source dataset, so it is not adopted in the target model; Add an output layer with an output size equal to the number of categories in the target dataset to the target model, and randomly initialize the model parameters of this layer. Train the target model on the target dataset and retrain the output layer, while the parameters of the remaining layers are fine-tuned based on the parameters of the source model; Fine-tuning specifically includes reusing classifier weights, and there are also some labels in the source data that are in the target data; Use the vector corresponding to the label in the pre-trained model classifier as the initial value; Secondly, the neural network learns hierarchical feature representations; Fix the relatively bottom layers and do not participate in parameter updates.

[0032] Step 5. Improve and establish a knowledge base client system for multi-terminal online Q&A.

[0033] Improve and establish a knowledge base client system for multi-terminal online Q&A including web programs, mobile apps, mini-programs, etc.: During the operation and maintenance process, automatically analyze the existing problems based on the effects of operation and maintenance. The model extracts parameters according to the problems and gives a complete set of problem solutions based on the pre-trained results of the model.

[0034] The model client adopts a Q&A dialogue mode, passes the retrieval parameters into the model in the form of questions. The model analyzes the questions through natural language processing NLP tasks and parses them into multiple parameters, matches the results with the model through the processing of the transformer architecture, and finally returns the retrieval results in a streaming manner.

[0035] The specific method is as follows: Step 5.1. Model training process: As Figure 6 shown, through the data collection system, integrate multi-source data such as web, files, manual operations, and databases, use the AI engine to build the basic knowledge base system, and at the same time use these data resources as the data materials for model training to perform pre-training of the model; Step 5.2. Data retrieval process: As Figure 7 、 8As shown in the figure, during the operation and maintenance process, when the user encounters a problem, a retrieval request is initiated from the client. If there is a retrieval result in the prefabricated library, it is directly fed back to the user. If there is no matching retrieval result, the AI model is used for retrieval, and the retrieval result is fed back to the user through the engine. After the user successfully solves the problem based on the feedback result, this processing record will also be recorded in the model library as training material data for the model to iteratively update the model. At the same time, the result will also be stored in the prefabricated library module to facilitate finding the result more efficiently and conveniently next time.

[0036] Embodiment 1 A method for mine operation and maintenance using a large AI model, the specific steps are as follows: Step 1: Design the model according to the requirements of the mine operation and maintenance scenario: Use the Recurrent Neural Network (RNN) algorithm to determine the model structure, and then perform feature processing on the model parameters; Step 2: Build the architecture of the large AI model using a multi-layer Transformer encoder according to the mine operation and maintenance scenario; Step 3: Train the built large AI model to obtain the trained large AI model; Step 4: Use the Fine-tuning technique to fine-tune the trained large AI model to adapt to specific natural language processing tasks; Step 5: Improve and establish a knowledge base client system for multi-terminal online question answering.

[0037] Embodiment 2 Based on the method of Embodiment 1, the specific method of Step 1 in this embodiment is as follows: The design of this model closely revolves around the specific business requirements of the operation and maintenance scenario, and uses the Recurrent Neural Network (RNN) algorithm as the core for constructing the model structure.

[0038] The model structure mainly consists of the following parts: (1) Input layer: Receives data points arranged in chronological order (equipment operation status, energy consumption data, personnel location information, environmental monitoring data) as input; the input data undergoes preprocessing steps, including standardization, normalization, and encoding, to adapt to the requirements of the model; (2) Hidden layer (RNN / LSTM / GRU layer): Each RNN unit is responsible for processing the input at one time step; each unit has a memory mechanism inside to control the flow and storage of information. The hidden layer can capture the temporal dependencies and long-term trends in the input data because the output of each unit depends not only on the input at the current time step but also on the hidden state (or memory) of the previous time step; (3) Output layer: Generates outputs according to specific business requirements, such as equipment fault warnings, personnel scheduling suggestions, and safety risk assessments. The output undergoes post-processing steps and is inversely normalized to a readable format; (4) Training and evaluation: The model is trained using a labeled training dataset through the backpropagation algorithm to minimize the prediction error; the performance of the model is evaluated using an independent test dataset, and the metrics are accuracy and recall; (5) Application deployment and monitoring: The trained RNN model is deployed to the coal mine operation and maintenance system to process and analyze real-time data; according to the output results of the model, corresponding operation and maintenance strategies are formulated, such as equipment maintenance plans, personnel scheduling plans, and safety warning measures; continuously monitor the model performance, collect feedback data in a timely manner, and iteratively optimize the model.

[0039] After determining the model structure, feature processing is performed on the model parameters. Among them, the model parameters include the learning rate, batch size, and optimizer. The feature processing specifically translates the original data of the model parameters into a form that the model can understand, including feature extraction, feature selection, feature combination, and feature transformation; Furthermore, feature extraction specifically extracts the mapping relationship between the operation and maintenance content and the operation and maintenance plan in the model, and improves the mapping relationship by introducing an attention weight mechanism to obtain a better representation; Among them, the operation and maintenance content is the daily or special operation and maintenance content required for personnel, equipment, etc. during the mine production process, specifically including: (1) Equipment monitoring data: Specifically, the operation parameters extracted from the sensors of mining equipment; (2) Environmental parameters: Temperature, humidity, and gas concentration in the mine environment; (3) Video images: Images or videos of the mine operation site captured by cameras, and the key features extracted through image processing algorithms, including open flames, smoke, and abnormal personnel behaviors.

[0040] The operation and maintenance plan is a set of appropriate maintenance plans given according to the operation and maintenance content. The AI large model will generate different operation and maintenance plans according to different operation and maintenance contents and combinations. Both the operation and maintenance content and the operation and maintenance plan are formulated according to the safety production regulations, and there is a large amount of content and it will vary slightly according to the actual situation of different mines.

[0041] Feature selection specifically uses statistical correlation analysis, recursive feature elimination, or embedded methods for feature selection, and evaluates the importance of features according to the correlation between features and target variables, the redundancy between features, and the criterion of the degree of improvement of features on model performance; and uses automated feature engineering technology (AutoML) to simplify the feature selection process, improve efficiency and accuracy, reduce the number of features, remove some redundant features in the operation and maintenance model, reduce computational complexity, and improve the generalization ability of the model; Feature combination specifically uses mathematical operations such as multiplication, addition, and exponentiation to combine features, and uses neural networks to automatically learn the complex relationships between features. By combining multiple different features, a new feature vector is formed to associate and combine each step required in the operation and maintenance process; Feature transformation specifically uses methods such as standardization, normalization, discretization, encoding, or dimensionality reduction to adjust the distance between features in the operation and maintenance model, as follows: (1) Standardization / normalization: Scale the collected feature values to the same scale to ensure that different features have equal weights in model training. Standardization usually involves subtracting the mean from the feature values and dividing by the standard deviation, while normalization scales the feature values between 0 and 1; (2) Discretization: For continuous features, such as device temperature or energy consumption, they can be converted into discrete features to better capture their change trends or category information. Discretization can be achieved by setting thresholds or using clustering algorithms; (3) Encoding: Encode categorical features and convert device types and personnel identities into numerical features; (4) Dimensionality reduction: Use principal component analysis (PCA) and linear discriminant analysis (LDA) to reduce the feature dimensions. Dimensionality reduction can help identify which features have the greatest impact on the operation and maintenance process.

[0042] Embodiment 3 Based on the method of Embodiment 1, the specific method for Step 2 in this embodiment is as follows: Step 2.1: First, calculate the weight coefficients using the query Query and the key Key, and then normalize the weights using the SoftMax operation to obtain softmax(f(Q, K)): (1) In the formula, Q is the query for model elements, K is the key of model elements, W is the first weight coefficient, U is the second weight coefficient, tanh() is the cosine similarity function; (2) In the formula, Softmax()is the weight normalization function; Step 2.2: Perform weighted summation on the value Value to calculate the output of attention: (3) In the formula, Attention() is the attention function, α is the attention weight parameter, V is the value corresponding to the model prototype; Step 2.3: Introduce the "encoder-decoder" framework into the attention mechanism. Use the encoder to encode the input sequence to first obtain semantic vectors, and then obtain the weighted sum of all semantic vectors, which is called the context vector. Using the weight parameter representing attention, the context vector can be expressed as: (4) In the formula, c is the context vector, α is the attention weight parameter, h is the semantic vector of the input sequence; During the decoding process, the probability of each output word is jointly determined by the context vector and the hidden state of the previous layer: (5) (6) In the formula, p and g respectively represent non-linear transformations; s represents a word in the output sequence, y represents the corresponding hidden state, c represents the context vector; in addition, the attention weight parameter is calculated using another neural network: (7) Through the above steps, the model with AI computing power is built.

[0043] Example 4 Based on the method of Example 1, the specific method of step 3 in this example is as follows: The training of the model is to adjust (learn) and determine the ideal values of all weights Weights and biases Bias through labeled samples.

[0044] Take one or more features as input and then return a prediction as output; for simplicity, a model that takes one feature and returns one prediction is used, as shown in the following formula: (8) In the formula b is Bias, wFor Weights; Through the loss function, calculate the losses of Bias and Weights under this set of parameters; use the gradient descent method to find the direction with smaller loss and perform iterations to obtain the trained large AI model.

[0045] Example 5 Based on the method of Example 1, the specific method for step 4 of this example is as follows: During the Fine-tuning process, the model uses metrics such as accuracy and recall to evaluate its performance based on its performance on the test dataset. After multiple rounds of fine-tuning and performance evaluation, if the results predicted by the model can accurately answer the key points of mine operation and maintenance on the test dataset, it means that the model training is successful; Specific steps of fine-tuning: Copy the source model of the large AI model created in step 2 to create a new neural network model, i.e., the target model; the target model copies all the model designs and their parameters of the source model except for the output layer; Assume that the model parameters contain the knowledge learned on the source dataset and that this knowledge is also applicable to the target dataset; assume that the output layer of the source model is closely related to the labels of the source dataset, so it is not adopted in the target model; add an output layer with an output size equal to the number of categories of the target dataset to the target model, and randomly initialize the model parameters of this layer. Train the target model on the target dataset and retrain the output layer, while the parameters of the remaining layers are fine-tuned based on the parameters of the source model; Fine-tuning specifically includes reusing the classifier weights, where the source data also has some labels in the target data; using the vectors corresponding to the corresponding labels in the pre-trained model classifier as the initial values; secondly, the neural network learns hierarchical feature representations; fix the relatively bottom layers and do not participate in parameter updates.

[0046] Example 6 Based on the method of Example 1, the specific method for step 5 of this example is as follows: Step 5.1, Model training process: As Figure 6 shown, through the data collection system, integrate multi-source data of web, files, manual operations, and databases, use the AI engine to build the basic knowledge base system, and at the same time use these data resources as the data materials for model training to perform pre-training of the model; Step 5.2, Data retrieval process: As Figure 7 、 8As shown in the figure, during the operation and maintenance process, when the user encounters a problem, a retrieval request is initiated from the client. If there is a retrieval result in the prefabricated library, it is directly fed back to the user. If there is no matching retrieval result, the AI model is used for retrieval, and the retrieval result is fed back to the user through the engine. After the user successfully solves the problem based on the feedback result, this processing record will also be recorded in the model library as training material data for the model to iteratively update the model. At the same time, the result will also be stored in the prefabricated library module to facilitate finding the result more efficiently and conveniently next time.

Claims

1. A method for mine operation and maintenance using an AI large model, characterized in that, The specific steps are as follows: Step 1: Conduct model design according to the requirements of the mine operation and maintenance scenario: Use the Recurrent Neural Network (RNN) algorithm to determine the model structure, and then perform feature processing on the model parameters; Step 2: Build an AI large model architecture using a multi-layer Transformer encoder according to the mine operation and maintenance scenario; Step 3: Train the built AI large model to obtain the trained AI large model; Step 4: Use the Fine-tuning technique to fine-tune the trained AI large model to adapt to specific natural language processing tasks; Step 5: Improve and establish a knowledge base client system for multi-terminal online Q&A.

2. The method for mine operation and maintenance using an AI large model according to claim 1, wherein, The feature processing in Step 1 specifically translates the original data of the model parameters into a form understandable by the model, including feature extraction, feature selection, feature combination, and feature transformation; Feature extraction specifically extracts the mapping relationship between the operation and maintenance content and the operation and maintenance plan in the model, and improves the mapping relationship by introducing an attention weight mechanism to obtain a better representation; Feature selection specifically uses statistical correlation analysis, recursive feature elimination, or embedded methods for feature selection to reduce the number of features, remove redundant features in the operation and maintenance model, retain high-quality feature data, and thus improve the performance of the model; Feature combination specifically combines features using mathematical operations such as multiplication, addition, and exponentiation, automatically learns the complex relationships between features using a neural network, forms a new feature vector by combining multiple different features, and associates and combines each step required in the operation and maintenance process; Feature transformation specifically adjusts the distance between the features of the operation and maintenance model using methods such as standardization, normalization, discretization, encoding, or dimensionality reduction.

3. The method for mine operation and maintenance using an AI large model according to claim 2, wherein The model parameters include the learning rate, batch size, and optimizer; The operation and maintenance content in the feature extraction is the daily or special operation and maintenance content required for personnel, equipment, etc. during the mine production process, specifically including equipment monitoring data, environmental parameters, and video images; The operation and maintenance plan is to give a suitable maintenance plan according to the operation and maintenance content. The AI large model will generate different operation and maintenance plans according to different operation and maintenance contents and combinations. Both the operation and maintenance content and the operation and maintenance plan are formulated according to the safety production regulations.

4. The method for mine operation and maintenance using the AI large model according to claim 3, characterized in that The equipment monitoring data is specifically the operation parameters extracted from the sensors of the mine equipment; the environmental parameters are the temperature, humidity, and gas concentration in the mine environment; the video images are specifically the images or videos captured by the camera at the mine operation site, and the key features extracted through image processing algorithms, and the key features include open flames, smoke, and abnormal personnel behavior.

5. The method for mine operation and maintenance using an AI large model according to claim 2, characterized in that, The specific method of the feature selection is as follows: Evaluate the importance of features according to the criteria of the correlation between features and the target variable, the redundancy between features, and the degree of improvement of features on the model performance; and use automated feature engineering techniques to simplify the feature selection process, improve efficiency and accuracy, remove some redundant features in the operation and maintenance model, reduce the computational complexity, and improve the generalization ability of the model.

6. The method for mine operation and maintenance using an AI large model according to claim 1, wherein The specific method of Step 2 is as follows: Step 2.1: First, calculate the weight coefficient using the query Query and the key Key, and then normalize the weights using the SoftMax operation to obtain softmax(f(Q, K)): （1） In the formula, Q is the query for the model element, K is the key of the model element, W is the first weight coefficient, U is the second weight coefficient, tanh() is the cosine similarity function; （2） In the formula, Softmax() is the weight normalization function; Step 2.2: Perform a weighted sum on the value Value to calculate the output of the attention: （3） In the formula, Attention() is the attention function, α is the attention weight parameter, V is the value corresponding to the model prototype; Step 2.3: Introduce the "encoder-decoder" framework into the attention mechanism. Use the encoder to encode the input sequence. First, obtain the semantic vector, and then obtain the weighted sum of all semantic vectors, which is called the context vector. Using the weight parameters representing attention, the context vector can be expressed as: （4） In the formula, c is the context vector, α is the attention weight parameter, h is the semantic vector of the input sequence; During the decoding process, the probability of each output word is jointly determined by the context vector and the hidden state of the previous layer: （5） （6） Wherein, p and g respectively represent non-linear transformations; s represents a word in the output sequence, y represents the corresponding hidden state, c represents the context vector; in addition, the weight parameters of the attention are calculated using another neural network: （7） The model with AI computing power is built.

7. The method for mine operation and maintenance using the AI large model according to claim 6, wherein The specific method of step 3 is as follows: The training of the AI large model is to adjust and determine the ideal values of all weights Weights and biases Bias through labeled samples; Take one or more features as input and then return a prediction as output; for simplicity, consider a model that takes one feature and returns one prediction, as follows: （8） where b is Bias, w is Weights; Through the loss function, calculate the loss loss of Bias and Weights under this set of parameters; use the gradient descent method to find the direction with less loss and perform iteration to obtain the trained AI large model.

8. The method for mine operation and maintenance using an AI large model according to claim 7, characterized in that The specific method of step 4 is as follows: Copy the source model of the AI large model created in step 2 to create a new neural network model, that is, the target model; the target model copies all model designs and their parameters except the output layer on the source model; Assume that the model parameters contain the knowledge learned on the source dataset, and this knowledge is also applicable to the target dataset; assume that the output layer of the source model is closely related to the labels of the source dataset, so it is not adopted in the target model; add an output layer with an output size equal to the number of categories of the target dataset to the target model, and randomly initialize the model parameters of this layer. Train the target model on the target dataset and retrain the output layer, while the parameters of the remaining layers are fine-tuned based on the parameters of the source model; The fine-tuning specifically includes reusing the classifier weights. There are also some labels in the source data that are in the target data; use the vector corresponding to the label in the pre-trained model classifier as the initialization value; secondly, the neural network learns hierarchical feature representations; fix the relatively bottom layers and do not participate in parameter updates.

9. The method for mine operation and maintenance using an AI large model according to claim 8, characterized in that, The specific method of step 5 is as follows: Step 5.1: Model training process: Through the data collection system, integrate multi-source data such as web, files, manual operations, and databases, use the AI engine to build the knowledge base basic system, and at the same time use these data resources as the data materials for model training to perform pre-training on the model; Step 5.2: Data retrieval process: During the operation and maintenance process, when the user encounters a problem, a retrieval request is initiated from the client. If there is a retrieval result in the prefabricated library, it is directly fed back to the user. If there is no matching retrieval result, the retrieval is performed through the AI model, and the retrieval result is fed back to the user through the engine. After the user successfully processes the problem through the feedback result, this processing record will also be recorded into the model library as data for model training materials to iteratively update the model; at the same time, the result will also be stored in the prefabricated library module for more efficient and convenient finding of the result next time.

Images