TBM jamming prediction method and system based on large language model and data optimization

Abstract

The application belongs to the field of tunnel boring machine construction, and provides a TBM jam prediction method and system based on a large language model and data optimization, real-time collection of TBM operation state information and construction information and preprocessing to obtain real-time window data flow; feature extraction of high-dimensional multi-modal representation of each window based on the real-time window data flow, jam detection using a trained jam detection model to obtain a jam detection result; joint early warning according to the jam detection result to obtain a jam risk detection result; semantic reasoning based on the jam risk detection result, semantic reasoning of the risk state in the construction information and a historical jam case library using a large language model to obtain a diagnosis report. The application effectively improves the safety, stability and intelligent level of TBM construction.

Description

Technical Field

[0001] This invention belongs to the field of tunnel boring machine construction technology, specifically relating to a TBM machine prediction method and system based on large language model and data optimization. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Tunnel engineering is rapidly developing in transportation, water conservancy, and energy sectors, with continuous improvements in construction scale and technology. Tunnels constructed using large-scale mechanized methods are increasingly being built towards ultra-long, deep, and complex geological conditions. Compared to traditional drill-and-blast methods, full-face tunnel boring machines (TBMs) have become the mainstream construction technology due to their advantages such as faster excavation speed, higher tunnel quality, lower overall cost, and safety and environmental friendliness. However, when TBMs traverse fault fracture zones, highly aquifers, or areas with high ground stress, their adaptability significantly decreases, making them prone to jamming accidents, which occur frequently and seriously threaten construction safety and progress. Therefore, research on TBM jamming prediction models based on artificial intelligence has made some progress. By selecting geological and excavation parameters and combining various algorithms to establish models, jamming risk prediction can be achieved. However, existing technologies still have shortcomings: most models select only single features and fail to fully integrate multi-source heterogeneous data; limited by the small number of monitoring samples, the accuracy and generalization performance of the models are insufficient to meet the needs of actual engineering projects. Therefore, the lack of an accurate model relating rock mass parameters to tunneling parameters in TBM tunneling makes it difficult to achieve real-time, high-precision, and optimal decision-making prediction of machine jamming. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a TBM tunneling prediction method and system based on a large language model and data optimization. This invention collects and preprocesses structured and unstructured data, constructs and trains a model, establishes a real-time data pipeline, visualizes the prediction results, and periodically fine-tunes the model with new data to continuously improve its performance. This ensures the accuracy and real-time nature of TBM tunneling prediction, achieves a perception-decision closed loop, better addresses challenges under different working conditions, improves TBM tunneling efficiency and safety, and has significant practical implications for tunnel construction.

[0005] According to some embodiments, the first aspect of the present invention provides a TBM card machine prediction method based on a large language model and data optimization, employing the following technical solution: A TBM (Traffic Machine Prediction) method based on large language models and data optimization includes: Real-time acquisition and preprocessing of TBM operation status and construction information to obtain real-time window data stream; High-dimensional multimodal representations of each window are extracted based on real-time window data streams. A pre-trained glitch detection model is used to perform glitch detection, and glitch detection results are obtained. Joint early warning is performed based on the glitch detection results to obtain glitch risk detection results. Based on the results of the machine risk detection, semantic reasoning is triggered. A large language model is used to perform semantic reasoning on the risk status in the construction information and the historical machine case database to obtain a diagnostic report.

[0006] Furthermore, the real-time acquisition and preprocessing of TBM operating status information and construction information to obtain a real-time window data stream includes: Real-time acquisition of TBM operation status information and construction information and time-series synchronization are performed to obtain a synchronized multi-source time-series dataset. Data cleaning is performed on the synchronized multi-source time series dataset. The cleaned multi-source time series dataset is then collected using a fixed time window to obtain a real-time window data stream.

[0007] Furthermore, the high-dimensional multimodal representation of each window based on real-time window data stream feature extraction includes: After feature extraction based on real-time window data stream, state subspaces are divided to obtain state subspace features of all state subspaces within each window; Based on the construction information of the current window, related entities are retrieved from the pre-built knowledge graph. The state subspace features of each window are weighted and fused according to the related entities to obtain a high-dimensional multimodal representation of each window.

[0008] Furthermore, after feature extraction based on the real-time window data stream, the state subspace is divided to obtain the state subspace features of all state subspaces within each window, including: Feature extraction is performed based on real-time window data streams to obtain multimodal features for each window; The multimodal features of each window are categorized according to their data source and mechanism of action. Class state subspace; Scale uniformity is performed on each type of state subspace to obtain the corresponding state subspace features.

[0009] Furthermore, the step of weighted fusion of the state subspace features of each window based on the associated entities to obtain a high-dimensional multimodal representation of each window includes: The state subspace features are jointly mapped with the knowledge entity embeddings of associated entities to obtain the state knowledge subspace features; We obtain a high-dimensional multimodal representation by weighted linear fusion of features in each state knowledge subspace.

[0010] Furthermore, the step of using the trained SIM card detection model to perform SIM card detection, obtaining SIM card detection results, and performing joint early warning based on the SIM card detection results to obtain SIM card risk detection results includes: Based on the high-dimensional multimodal representation of each window, the pre-trained machine detection model is used to perform machine detection, and the machine detection risk probability and uncertainty are obtained as the machine detection result. When the probability of machine malfunction is less than the first risk threshold and the uncertainty is less than the first uncertainty threshold, the current window risk level is a low-level warning. When the probability of a machine malfunction is between the first risk threshold and the second risk threshold, and the uncertainty is between the first uncertainty threshold and the second uncertainty threshold, the current window risk level is a medium warning. When the probability of machine malfunction exceeds the second risk threshold and the uncertainty exceeds the second uncertainty threshold, the risk level of the current window is set to high-level warning. The probability, uncertainty, and warning level of the machine jam risk for each window are used as the results of the machine jam risk detection.

[0011] According to some embodiments, a second aspect of the present invention provides a TBM card machine prediction system based on a large language model and data optimization, employing the following technical solution: A TBM (Traffic Machine Prediction) system based on large language models and data optimization includes: The data acquisition and processing module is configured to collect TBM operating status information and construction information in real time and perform preprocessing to obtain a real-time window data stream. The glitch detection module is configured to extract high-dimensional multimodal representations of each window based on real-time window data streams, perform glitch detection using a trained glitch detection model, and obtain glitch detection results; based on the glitch detection results, a joint early warning is performed to obtain glitch risk detection results; The detection result reasoning module is configured to trigger semantic reasoning based on the card machine risk detection results. It uses a large language model to perform semantic reasoning on the risk status in the construction information and the historical card machine case library to obtain a diagnostic report.

[0012] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the TBM card machine prediction method based on large language models and data optimization as described in the first embodiment above.

[0014] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the TBM card machine prediction method based on large language models and data optimization as described in the first embodiment above.

[0016] According to some embodiments, a fifth aspect of the present invention provides a computer program product or computer program.

[0017] A computer program product or computer program includes computer instructions stored in a computer-readable storage medium, wherein a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the TBM card machine prediction method based on large language model and data optimization as described in the first embodiment above.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention significantly improves the accuracy and system stability of TBM (Turbine Machine Tool) jamming prediction. Through systematic data governance and feature optimization of multi-source monitoring data, it achieves a dual improvement in information quality and feature representation capabilities. By performing time-series alignment, noise removal, missing data repair, and data consistency verification on the original monitoring data, it effectively reduces the interference of abnormal data on the prediction results. Simultaneously, through a multi-modal representation fusion mechanism, it integrates equipment operating parameters and construction-related information into a unified model, solving the information gap problem caused by traditional models relying solely on single numerical data. This enhances the model's overall adaptability to complex geological conditions and changing construction environments. Introducing prior knowledge from a knowledge graph can compensate for insufficient data, improving stability under small sample sizes and complex conditions, making the jamming risk prediction results more accurate, stable, and reliable. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart of a TBM card machine prediction method based on large language model and data optimization in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0025] Example 1 like Figure 1 As shown, this embodiment provides a TBM card machine prediction method based on large language models and data optimization. This embodiment uses the application of this method to a server as an example for illustration. It can be understood that this method can also be applied to terminals, and can also be applied to a system including terminals, servers, and systems, and can be implemented through the interaction between terminals and servers. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected through wired or wireless communication, which is not limited in this application. In this embodiment, the method includes the following steps: Real-time acquisition and preprocessing of TBM operation status and construction information to obtain real-time window data stream; High-dimensional multimodal representations of each window are extracted based on real-time window data streams. A pre-trained glitch detection model is used to perform glitch detection, and glitch detection results are obtained. Joint early warning is performed based on the glitch detection results to obtain glitch risk detection results. Based on the results of the machine risk detection, semantic reasoning is triggered. A large language model is used to perform semantic reasoning on the risk status in the construction information and the historical machine case database to obtain a diagnostic report.

[0026] This embodiment provides a TBM (Traffic Machine Machine) intelligent prediction method based on Large Language Model (LLM) and data optimization. The overall process of the TBM intelligent prediction system based on Large Language Model (LLM) and data optimization can be divided into four main stages, forming an intelligent prediction system that can be implemented in an engineering manner.

[0027] Step S1: Collect TBM operating status information and construction information in real time and preprocess them to obtain a real-time window data stream; Step S1.1: Real-time acquisition and synchronization of multi-source data - Real-time acquisition of TBM operation status information and construction information and time-series synchronization to obtain synchronized multi-source time-series dataset; TBM operating status information includes cutter head torque. (range 0–1.2×10) 6 N·m), propulsion (range 0–4.5×10) 6 N), penetration (Average 2.1mm / r), Earth pressure (0.15–0.35 MPa) and vibration signal The sampling frequency is 10Hz, and the daily data volume is approximately 5–8GB.

[0028] Construction information is unstructured text data, including daily construction logs, geological description text, historical jamming event records (including precise time, location, and handling measures), and risk status (such as a 18% increase in torque, a 32% decrease in penetration rate, and a 45% increase in high-frequency vibration energy within the past 15 minutes). The geological description text includes lithological classifications (such as granite, mudstone, and sandstone), fault fracture zone locations, groundwater conditions, friction coefficients, and muck sample density. This type of data is recorded as discrete points and correlated with the tunneling mileage.

[0029] Since the time base and sampling period of each data source are inconsistent, the system establishes a unified clock base (based on GPS clock or NTP server) to align the timestamps of all data sources to millisecond precision. For geological description texts without timestamps, they are associated with the nearest tunneling mileage location to form a location-time mapping.

[0030] Specifically, a linear interpolation method is used to unify all sensor channels to the same sampling time step. Let the multi-channel sensor signal be... Since the original sampling periods of each channel are inconsistent, the linear interpolation formula is as follows:

[0031] in, It is a unified timestamp. Indicates the sensor channel number. Indicates the first The sampling sequence number of the original sampling sequence in each channel. This indicates the total number of original sampling points corresponding to this channel.

[0032] After the above processing, the synchronized multi-source time series dataset is output. ,in, It is a unified timestamp. It is the sensor time-series vector. This is the interpolated geological parameter vector. Index for associated text records.

[0033] Step S1.2: Clean the synchronized multi-source time series dataset, and collect data from the cleaned multi-source time series dataset using a fixed time window to obtain a real-time window data stream; After outlier removal, linear interpolation, and wavelet denoising, a cleaned multi-source time series dataset is obtained. Specifically, a sliding window with a fixed width of W (e.g., 10 minutes) slides forward continuously as new data arrives, inputting the real-time window data stream within the current window into the trained detection model, and the detection model outputs the risk probability corresponding to that window; This step, during the training process, is detailed as follows: By utilizing historical TBM operation status information and construction information, time-series synchronization is performed to obtain a synchronized historical multi-source time-series dataset. Then, data cleaning is performed to obtain a cleaned historical multi-source time-series dataset. Then, using the time of the machine jamming event as the anchor point, a high-risk evolution window is constructed, and the normal operation window and the critical abnormal non-jamming window are hierarchically labeled. A meta-description table and a quality assessment standard system (DQI, Data Quality Index) are established to provide high-quality input for subsequent modeling.

[0034] Based on the criteria for determining machine jamming events, the cleaned historical multi-source time-series dataset is analyzed to determine the time when a machine jamming event occurs. The criteria for determining a machine freeze event can be defined by a combination of the following conditions:

[0035] in: It's the cutter head torque. It is the torque threshold. It is the speed of propulsion. It is the speed threshold. It is a low propulsion duration. It is the duration threshold. It's the previous timestamp.

[0036] Let the first The time when the card-breaking incident occurred was Define the set of events that cause the machine to freeze as follows:

[0037] in, This is a category for lag events, used to distinguish different types of lag events. For the first The incident of the card machine being used. .

[0038] Then, based on the time of the freeze event Centered on this, three types of time windows are constructed: High-risk evolution window Define the prediction advance window length as Window width is ,but:

[0039] High-risk evolution window label is This indicates that within this time period, the system should predict the future. A system freeze will occur after a certain time.

[0040] Critical anomaly window Defined as a status window that is close to freezing but has not yet been triggered:

[0041] The critical exception window label is This indicates an abnormal state that does not meet the criteria for a machine freeze, and is used to improve the model's discrimination ability. This indicates the length of the critical anomaly time window, used to limit the time range of potential abnormal states before a machine malfunction occurs.

[0042] Normal operating window Random sampling is performed within a time interval far removed from any lag event.

[0043] Normal operation window tab is ,in, This is the safety interval length; For any type of time window Feature extraction was performed on the cleaned historical multi-source time-series data to obtain window state features. :

[0044] in, Indicates the time within the time window Cleaned historical multi-source time-series data; It is a feature extraction function; Based on the three types of time windows and corresponding time-crossing real labels, the unified label function is defined as follows:

[0045] Define time window weights for the three types of time windows: ,

[0046] Based on each sample within various time periods Window state features Time window true label and time window weight Construct the training dataset as follows:

[0047] Based on each sample within various time periods Time window features Generate time window prediction labels The calculation is as follows:

[0048] in, It is the feature of time window Mapped to the corresponding time window prediction label The mapping function is based on each sample within various time periods. The time window weights, true labels within the time window, and predicted labels within the time window are used to construct a weighted loss function during training. :

[0049] To quantify the reliability of construction information, construction information quality indicators are defined. The formula is as follows:

[0050] in, It is a data accuracy indicator, obtained by measuring the deviation between construction parameters and physical constraints or empirical ranges. yes The weights; It is a data consistency indicator that measures the continuity of changes between adjacent time points or related variables. yes The weights; It is a data integrity indicator, quantified based on the proportion of missing data or the imputation ratio. yes The weight.

[0051] For each time window Window-level quality scores are calculated based on the construction information data it contains. The formula is as follows:

[0052] in, It is an aggregation function representing the quality indicators of construction information within a time window, used to summarize the quality indicators of construction information at each moment within the window. It is every time window middle Construction information quality indicators at any given time; if When the sample size falls below a set threshold, the corresponding window sample is removed or its weight in the training process is reduced.

[0053] Then, As an additional feature, it is input together with the window state features in the training dataset into the untrained card detection model to improve the model's ability to perceive data reliability.

[0054] Step S2: Based on the real-time window data stream, perform feature extraction for each window to obtain a high-dimensional multimodal representation. Use the trained glitch detection model to perform glitch detection and obtain the glitch detection results. Perform joint early warning based on the glitch detection results to obtain the glitch risk detection results.

[0055] Step S2.1: After extracting features based on the real-time window data stream, divide the state subspace to obtain the state subspace features of all state subspaces within each window; Specifically, feature extraction is performed based on the real-time window data stream to obtain multimodal features for each window, including: Time-domain representation (mean, variance, volatility); Frequency domain characterization (dominant frequency energy ratio, high-frequency energy distribution); Text features (construction information is transformed into a 768-dimensional semantic vector using an LLM embedding model); Geological parameters (such as friction coefficient, slag density, etc.) are used as static inputs.

[0056] The multimodal features of each window are categorized according to their data source and mechanism of action. Class state subspaces, including but not limited to: Dynamic states (such as propulsion force, cutterhead torque, etc.); Kinematic states (such as propulsive speed, rotational speed, etc.); Vibration and energy-related (such as vibration signals, high-frequency energy, etc.); Geological environment (such as stratigraphic parameters, earth pressure, etc.); Construction control parameters (such as operating parameters, tunneling modes, etc.); Scale uniformity is performed on each type of state subspace to obtain the corresponding state subspace features. Class state subspace features constitute a comprehensive operational state representation. ; Set at time Comprehensive operational status characterization ,as follows:

[0057] in, Indicates the first Features of class-state subspace.

[0058] Step S2.2: Based on the construction information of the current window, retrieve related entities from the pre-built knowledge graph, and perform weighted fusion of the state subspace features of each window based on the related entities to obtain a high-dimensional multimodal representation of each window; To incorporate prior engineering knowledge, a TBM construction knowledge graph is constructed:

[0059] in: It is a collection of entities (e.g., formation type, cutterhead components, construction conditions, risk events). It is a set of relationships (e.g., causing, associating, exacerbating, depending); It is a ternary set (e.g., highly plastic clay, which is prone to causing cutterhead mud). It is a set of results, such as sludge buildup on the cutterhead; To achieve joint modeling of knowledge and data, a knowledge graph embedding method is used to map entities and relations to a continuous semantic space. :

[0060] in, It is a data point in the set of triples. , , And satisfy structural constraints (taking the translation model as an example):

[0061] Then, the corresponding training objective function in the knowledge graph construction process for:

[0062] This process encodes causal relationships and dependencies in the knowledge graph into computable geometric constraints.

[0063] Based on the construction information in the current window, retrieve related entities from the pre-built knowledge graph; To achieve deep integration of data and knowledge, the features of the state subspace are... By jointly mapping the knowledge entity embeddings of associated entities, we obtain the features of the state knowledge subspace. :

[0064] in: It is related to the characteristics of the state subspace Embedding of related knowledge entities; It is a nonlinear fusion function (such as splicing mapping or bilinear interaction). To characterize the features of knowledge subspaces in different states To determine the relative importance of card-related risks, a weighted linear fusion method was used to obtain a high-dimensional multimodal characterization. :

[0065]

[0066] in, The feature weights can be adaptively learned during training through the Attention mechanism:

[0067] in, It is the first Features of each state subspace The vector, It is the first Transpose of the weight vector of features of each state subspace It is the first Features of each state subspace The vector, It is the first The weight vector transpose of the features of each state subspace.

[0068] The mapping relationship between high-dimensional multimodal representations and card-related events is modeled. Specifically, the modeling method (taking XGBoost as an example) is optimized, and its overall learning objective function is... Defined as:

[0069]

[0070] Among them, the overall learning objective function It is used to measure the error between the model's prediction results and the actual card tag, including prediction error term and regularization term, and is used to constrain model complexity while improving prediction accuracy; This represents the loss metric function between the predicted output and the actual observation. For the first Individual base learners; This is a model complexity penalty term used to suppress overfitting; The number of leaf nodes in the tree is used to characterize the complexity of the model. The regularization constraint coefficient controls the smoothness of the parameter space.

[0071] Step S2.3: Based on the high-dimensional multimodal representation of each window, use the trained card detection model to perform card detection, and obtain the card detection risk probability and uncertainty as the card detection result; The probability of a checkpoint being set up is , indicating at time The probability of a machine freezing is predicted, and:

[0072] in, For the current moment The high-dimensional multimodal representation of the previous window, This indicates that a machine freeze has occurred; when Then it is determined that: That is, the system in time Issuing early warnings to anticipate the future Risk of system malfunction after a certain time.

[0073] For the same moment At the same time, output the uncertainty of the prediction. , defined as the confidence level of the SIM card prediction model in the current prediction:

[0074] It should be noted that the card detection model here can be, but is not limited to, the XGBoost model, the LSTM model, etc., which will not be elaborated here.

[0075] Each feature is fused using attention weights to generate a high-dimensional multimodal representation. The data is input into the card reader prediction model (such as the TFT-XGBoost hybrid framework). The card reader prediction model is trained using historical tunneling data. 6 With a sample size of [number], the results on the validation set achieved an AUC of 0.93, a recall rate of 0.87, and a mean advance prediction time of 28 minutes. The SIM card prediction model outputs the probability of SIM card malfunction. and combined with uncertainty The alarm threshold is dynamically adjusted based on the risk probability.

[0076] Step S2.4: Based on the set risk threshold and uncertainty threshold, perform joint early warning on the card detection results to obtain the early warning level for each window. Use the card risk probability, uncertainty and early warning level of each window as the card risk detection result. When the probability of machine malfunction is less than the first risk threshold and the uncertainty is less than the first uncertainty threshold, the current window risk level is a low-level warning. When the probability of a machine malfunction is between the first risk threshold and the second risk threshold, and the uncertainty is between the first uncertainty threshold and the second uncertainty threshold, the current window risk level is a medium warning. When the probability of machine malfunction exceeds the second risk threshold and the uncertainty exceeds the second uncertainty threshold, the risk level of the current window is set to high-level warning. The probability, uncertainty, and warning level of the machine jam risk for each window are used as the results of the machine jam risk detection.

[0077] Based on the joint assessment mechanism, the system defines the following three early warning levels according to different combinations of card-related risk probabilities and uncertainties: 1. Normal warning (Level 0), i.e., low-level warning: The trigger condition is < and The risk of machine jamming is low under the current construction conditions and the confidence level of the model prediction is high, so the system does not need to take further measures.

[0078] 2. Warning / Alert (Level 1), i.e., Intermediate Warning: Triggering conditions: and Meaning: The current risk of machine malfunction is high, and the model uncertainty is moderate. The system should issue a warning, and attention should be paid to changes in the construction status, with preparations made for early warning.

[0079] 3. Emergency Warning (Level 2), also known as Advanced Warning: Triggering conditions: or Meaning: The risk of system malfunction is high and the model has great uncertainty. The system needs to respond immediately and implement emergency measures.

[0080] This embodiment can also enable a review mechanism for warnings (Level 1) and emergency warnings (Level 2), combining real-time construction status data for secondary evaluation to ensure the accuracy of prediction results and the timeliness of execution.

[0081] The review and judgment are based on the real-time evolution of the construction status and adopt the following rules:

[0082] in, This is for updating information from the engineering knowledge base or in real-time feedback. If the verification results indicate a high probability of a false alarm, the system will cancel the warning and adjust the handling plan.

[0083] Based on different warning levels, the system outputs corresponding response measures and suggestions: Level 0: No action is required.

[0084] Level 1: It is recommended to strengthen monitoring and adjust construction strategies (such as deceleration, adjusting thrust, etc.).

[0085] Level 2: It is recommended to immediately suspend operations, conduct equipment inspections, and take emergency measures (such as reversing the cutterhead, stopping the feed, etc.).

[0086] Step S3: Based on the risk detection results of the machine jammer, trigger semantic reasoning, use a large language model to perform semantic reasoning on the risk status in the construction information and the historical machine jammer case library to obtain a diagnostic report; Semantic reasoning is triggered when the risk warning level in the card machine risk detection result is medium or high warning. The semantic similarity between the risk status in the construction information and the historical machine jamming case library is calculated using a large language model. Historical machine jamming cases and risk statuses that meet the semantic similarity criteria are selected to obtain a diagnostic report.

[0087] Risk conditions (such as a 18% increase in torque, a 32% decrease in penetration rate, and a 45% increase in high-frequency vibration energy within the last 15 minutes) are input into the LLM. This is then combined with retrieval enhancement using approximately 200 historical machine malfunction cases from the knowledge base (RAG architecture). The LLM outputs a structured diagnostic report through semantic reasoning, including information such as "Probability of increased formation friction: 0.72", "Probability of cutterhead wear: 0.65", and "Recommended measures: Reduce advance rate by 15%, inject 100L of lubricant, and briefly reverse the cutterhead to clear clay blockages." This enables the interpretable generation of risk sources and corresponding response strategies.

[0088] Specifically, the vectorized representation and semantic retrieval methods are as follows: First, review each document in the historical card machine case library. and each risk status Generate document embedding representation and risk embedding representation :

[0089]

[0090] in, It is an embedded function.

[0091] Computational document embedding representation and risk embedding representation Semantic similarity:

[0092] Sort by similarity, and select the top... The most relevant documents

[0093] Specifically, the knowledge enhancement generation method is as follows: The LLM input consists of the current risk semantic description and a set of relevant cases:

[0094] in, Explanations and suggestions for the generated content.

[0095] For example, a Prompt template: "Currently, torque is increasing while propulsion speed is decreasing. Please analyze the potential causes based on historical cases and provide suggestions." This embodiment achieves interpretable and intelligent analysis of jamming prediction results. By introducing a Large Language Model (LLM), it performs semantic parsing on unstructured information such as construction logs and geological descriptions, and integrates the prediction model output with domain knowledge for analysis, thereby realizing intelligent interpretation and causal analysis of jamming risks. This invention not only outputs a judgment result indicating the existence of jamming risk, but also provides possible causes and influencing factors, expanding the prediction results from a single numerical output to semantically grounded analytical conclusions. This significantly improves the system's understandability and decision support capabilities in practical engineering applications.

[0096] Step S4: Continuously and adaptively monitor the high-dimensional multimodal representation of the current window. When the preset index threshold is reached, update and train the card detection model and update the historical card case library.

[0097] The system continuously tracks the distribution of input states and changes in model performance, using drift detection algorithms (such as KL divergence and PSI index) to identify state distribution drift. When the distribution difference exceeds 0.15 or the model accuracy drops by more than 10%, the system automatically triggers a fine-tuning mechanism, using data from the most recent two weeks for incremental learning. If the drift persists, full retraining and manual validation are performed to ensure the long-term stability and adaptability of the prediction model.

[0098] The calculation of data distribution is based on construction status characteristic data (such as torque, propulsion speed, vibration, etc.). The probability distribution is obtained by statistical modeling historical data and current time window data; Specifically, the data drift detection (KL divergence) method is as follows:

[0099] like If the current construction status enters a new statistical distribution region, the model will be retrained.

[0100] in, It is a historical reference distribution. This is the current data distribution. It is the KL divergence threshold, used to determine whether data drift has occurred. It is the index of the discretized feature interval; Model performance evaluation is based on the consistency between model prediction results and real labels, which are derived from construction records, equipment status feedback, etc.

[0101] Specifically, the performance evaluation and update criteria methods are as follows:

[0102] in, It's the size of the sliding window. It is a moment The probability of accurate prediction, The card prediction model is in the first Prediction results on a single sample Is the model in the first The true label of each sample It is the current moment. It is the sample index within the window; when Automatically perform fine-tuning or incremental learning:

[0103] in, These are the updated model parameters. It's the learning rate. It is the time window number. These are the current model parameters; It is the first The representation state of each window; It is the first The actual label of each window; It is a loss function; It is the gradient of the parameters; After each model update, the system will synchronously update the knowledge base and case library. The specific steps are as follows: 1. Add new cases to the database: New cases of device malfunction will be labeled as positive samples. =1) or negative samples ( =0), and store it in the construction case library. Each new case will be labeled according to its geological conditions, construction methods, risk characteristics and other information, and a time-series evolution feature will be constructed.

[0104] 2. Vector index update: Whenever a new case is added, the system calculates its high-dimensional multimodal representation and maps it to a vector space. The system uses an incremental learning algorithm to update the vector index database, ensuring that index queries are not affected by the addition of new samples.

[0105] 3. Knowledge Graph Revision and Expansion: Based on the construction techniques, geological features, and machine jamming event types in the newly added cases, the system will automatically correct or expand the entities and relationships in the knowledge graph. For example, if a new case involves a machine jamming event caused by a new type of geological condition, the system will add that geological type as a new entity to the knowledge graph and correct the relevant entity relationships.

[0106] 4. Model version and knowledge version are updated in tandem: Each time the model is updated, the system checks the current model version and the knowledge graph version to ensure that both are synchronized. This version control system ensures that new model versions can depend on the updated knowledge graph, and that all versions are traceable.

[0107] Through the implementation of this system, potential tunneling jam risks can be identified on average 25–30 minutes in advance during continuous tunneling monitoring, with a prediction accuracy of approximately 90% and a false alarm rate controlled below 10%. Practice has shown that the integration of LLM and data optimization technology effectively improves the timeliness, interpretability, and reliability of TBM jam prediction, providing intelligent and adaptive decision support for shield tunneling under complex geological conditions.

[0108] This embodiment features adaptive model updates, making it suitable for long-term prediction needs under complex working conditions. By introducing an adaptive learning mechanism, it continuously monitors changes in the distribution of input data. When a decline in model performance or a significant change in environmental characteristics is detected, it automatically triggers a model adjustment and update process, achieving dynamic optimization and online evolution of the model. This ensures the predictive stability and accuracy of the system during long-term construction. Therefore, this invention possesses stronger engineering adaptability and applicability, and can be applied to TBM truck risk prediction tasks under different geological conditions and construction scenarios.

[0109] Example 2 This embodiment provides a TBM (Traffic Machine Prediction) system based on large language models and data optimization, including: The data acquisition and processing module is configured to collect TBM operating status information and construction information in real time and perform preprocessing to obtain a real-time window data stream. The glitch detection module is configured to extract high-dimensional multimodal representations of each window based on real-time window data streams, perform glitch detection using a trained glitch detection model, and obtain glitch detection results; based on the glitch detection results, a joint early warning is performed to obtain glitch risk detection results; The detection result reasoning module is configured to trigger semantic reasoning based on the card machine risk detection result, and use a large language model to perform semantic reasoning on the risk status in the construction information and the historical card machine case library to obtain a diagnostic report. The monitoring and update module is configured to continuously and adaptively monitor the high-dimensional multimodal representation of the current window. When the preset index threshold is reached, the card detection model is updated and trained, and the historical card case library is updated.

[0110] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0111] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0112] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0113] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the TBM card machine prediction method based on large language models and data optimization as described in Embodiment 1 above.

[0114] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the TBM card machine prediction method based on large language model and data optimization as described in Embodiment 1 above.

[0115] Example 5 This embodiment provides a computer program product or computer program, including computer instructions stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the TBM card machine prediction method based on large language model and data optimization described in Embodiment 1 above.

[0116] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0117] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0121] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A TBM (Traffic Machine Prediction) method based on large language models and data optimization, characterized in that, include: Real-time acquisition and preprocessing of TBM operation status and construction information to obtain real-time window data stream; Based on the real-time window data stream, feature extraction is performed on the high-dimensional multimodal representation of each window, and the trained card detection model is used to perform card detection to obtain the card detection results; Based on the card reader detection results, a joint early warning is issued to obtain the card reader risk detection results; Based on the results of the machine risk detection, semantic reasoning is triggered. A large language model is used to perform semantic reasoning on the risk status in the construction information and the historical machine case database to obtain a diagnostic report.

2. The TBM card machine prediction method based on large language model and data optimization as described in claim 1, characterized in that, The real-time acquisition and preprocessing of TBM operating status information and construction information to obtain a real-time window data stream includes: Real-time acquisition of TBM operation status information and construction information and time-series synchronization are performed to obtain a synchronized multi-source time-series dataset. Data cleaning is performed on the synchronized multi-source time series dataset. The cleaned multi-source time series dataset is then collected using a fixed time window to obtain a real-time window data stream.

3. The TBM card machine prediction method based on large language model and data optimization as described in claim 1, characterized in that, The high-dimensional multimodal representation of each window, based on real-time window data stream feature extraction, includes: After feature extraction based on real-time window data stream, state subspaces are divided to obtain state subspace features of all state subspaces within each window; Based on the construction information of the current window, related entities are retrieved from the pre-built knowledge graph. The state subspace features of each window are weighted and fused according to the related entities to obtain a high-dimensional multimodal representation of each window.

4. The TBM card machine prediction method based on large language model and data optimization as described in claim 3, characterized in that, After feature extraction based on the real-time window data stream, the state subspace is divided to obtain the state subspace features of all state subspaces within each window, including: Feature extraction is performed based on real-time window data streams to obtain multimodal features for each window; The multimodal features of each window are categorized according to their data source and mechanism of action. Class state subspace; Scale uniformity is performed on each type of state subspace to obtain the corresponding state subspace features.

5. The TBM card machine prediction method based on large language model and data optimization as described in claim 3, characterized in that, The step of weighted fusion of the state subspace features of each window based on associated entities to obtain a high-dimensional multimodal representation of each window includes: The state subspace features are jointly mapped with the knowledge entity embeddings of associated entities to obtain the state knowledge subspace features; We obtain a high-dimensional multimodal representation by weighted linear fusion of features in each state knowledge subspace.

6. The TBM card machine prediction method based on large language model and data optimization as described in claim 1, characterized in that, The process involves using a trained SIM card detection model to perform SIM card detection, obtaining SIM card detection results, and then performing joint early warning based on these results to obtain SIM card risk detection results, including: Based on the high-dimensional multimodal representation of each window, the pre-trained machine detection model is used to perform machine detection, and the machine detection risk probability and uncertainty are obtained as the machine detection result. When the probability of machine malfunction is less than the first risk threshold and the uncertainty is less than the first uncertainty threshold, the current window risk level is a low-level warning. When the probability of a machine malfunction is between the first risk threshold and the second risk threshold, and the uncertainty is between the first uncertainty threshold and the second uncertainty threshold, the current window risk level is a medium warning. When the probability of machine malfunction exceeds the second risk threshold and the uncertainty exceeds the second uncertainty threshold, the risk level of the current window is set to high-level warning. The probability, uncertainty, and warning level of the machine jam risk for each window are used as the results of the machine jam risk detection.

7. A TBM (Traffic Machine Prediction) system based on large language models and data optimization, characterized in that: include: The data acquisition and processing module is configured to collect TBM operating status information and construction information in real time and perform preprocessing to obtain a real-time window data stream. The card detection module is configured to extract high-dimensional multimodal representations of each window based on real-time window data streams, and use the trained card detection model to perform card detection to obtain card detection results. Based on the card reader detection results, a joint early warning is issued to obtain the card reader risk detection results; The detection result reasoning module is configured to trigger semantic reasoning based on the card machine risk detection results. It uses a large language model to perform semantic reasoning on the risk status in the construction information and the historical card machine case library to obtain a diagnostic report.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the TBM card machine prediction method based on large language model and data optimization as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the TBM card machine prediction method based on large language model and data optimization as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps in the TBM card machine prediction method based on large language models and data optimization as described in any one of claims 1-6.

Images