Method for predicting residual life of key component of tunneling equipment based on large model

Abstract

The invention provides a method for predicting the remaining life of key parts of tunneling equipment based on a large model. According to the method for predicting the residual life of the key component of the tunneling equipment based on the large model, the prediction precision of the residual life of the key component of the tunneling equipment and the model generalization ability can be remarkably improved, predictive maintenance of the equipment is effectively supported, the risk of non-planned shutdown is reduced, and the safety and efficiency of coal mine production are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of residual life prediction of key components of tunneling equipment, and in particular to a residual life prediction method for key components of tunneling equipment based on a large model. BACKGROUND

[0002] Coal mine tunneling equipment, as the core execution unit in the intelligent system of coal mining, is widely used in roadway tunneling operations under complex geological conditions in the mine. With the development of the coal industry towards intelligence and unmanned, the health management of key components of tunneling equipment has become a key link to ensure production continuity and operation safety. In related technologies, a residual life prediction system for equipment is constructed through the cooperative work of state monitoring, feature extraction and life prediction model. Specifically, the system covers the whole process from data acquisition, feature modeling to prediction decision, including key links such as sensor network deployment, time series data analysis, machine learning modeling, etc. Among them, the prediction method based on deep learning has gradually become the mainstream, but due to the limitation of model structure and training strategy, it still has significant deficiencies in multi-dimensional heterogeneous data fusion, nonlinear degradation modeling and small sample adaptability.

[0003] However, in the existing residual life prediction method for tunneling equipment, the traditional LSTM or SVM model is directly used without fully considering the dynamic evolution characteristics of the equipment operating state, which may lead to insufficient feature extraction, poor model generalization ability, or high prediction error rate, thereby affecting the scientificity of equipment maintenance strategy and production efficiency. In addition, the existing deep learning model usually relies on a large amount of labeled data for training, while the acquisition of failure samples of key components in the coal mine site is costly and time-consuming, which limits the practical application of the model. Based on this, there is an urgent need for a residual life prediction method that can fuse multi-dimensional time series data, has strong generalization ability and is suitable for small sample scenarios, to improve the intelligent operation and maintenance level of tunneling equipment. SUMMARY

[0004] The present application aims to at least partially solve one of the technical problems in the related art.

[0005] To this end, a first object of the present application is to propose a residual life prediction method for key components of tunneling equipment based on a large model.

[0006] A second object of the present application is to propose a residual life prediction device for key components of tunneling equipment based on a large model.

[0007] A third object of the present application is to propose an electronic device.

[0008] A fourth object of the present application is to propose a computer-readable storage medium.

[0009] A fifth object of the present application is to propose a computer program product.

[0010] To achieve the above object, the first aspect of the present application proposes a method for predicting the residual life of a key component of tunneling equipment based on a large model, comprising: S1, replacing the feedforward network layer of a pre-trained large language model, and adopting a sparse MoE architecture to realize dynamic feature routing processing; S2, segmenting the multi-dimensional time series monitoring data of the key component of the tunneling equipment according to a set window size and sliding step to obtain a plurality of data sequence segments; S3, constructing a self-supervised learning data set based on the data sequence segments, calculating the reconstruction error by randomly masking a certain dimension of data and reconstructing the dimension of data using a feedforward neural network for self-supervised fine-tuning; S4, constructing a supervised learning data set based on the data sequence segments, embedding each dimension of data into a D-dimensional space and fusing through a linear layer, and using real residual life data to supervise the fine-tuning of the self-supervised fine-tuned model; S5, segmenting the real-time collected key component data of the tunneling equipment according to the same window size and sliding step as in S2 to form a standardized input sequence segment, and inputting the sequence segment into the large language model after supervised fine-tuning to output the residual life prediction result of the key component.

[0011] In an embodiment of the present application, the replacement of the feedforward network layer of the pre-trained large language model and the adoption of the sparse MoE architecture to realize dynamic feature routing processing further comprise: S11, replacing the standard feedforward network (FFN) layer in the Qwen3 model with a sparse mixed expert (MoE) layer, wherein each expert network has different feature processing capabilities; S12, analyzing the input data through a gating mechanism, and dynamically selecting the expert network participating in the calculation according to the data characteristics to improve the adaptability of the model to multi-dimensional time series data.

[0012] In an embodiment of the present application, the segmentation of the multi-dimensional time series monitoring data of the key component of the tunneling equipment according to the set window size and sliding step to obtain a plurality of data sequence segments further comprises: S21, representing the multi-dimensional condition monitoring data X as a real matrix of DxT, where D is the number of data dimensions and T is the number of time steps; S22, cutting the matrix X using a sliding window of length S to generate N data sequence segments, where N=T / S, and there is a partial overlap between adjacent sequence segments.

[0013] In an embodiment of the present application, the self-supervised learning data set is constructed based on the data sequence segment, the reconstruction error is calculated by randomly masking a certain dimension of data and reconstructing the dimension of data using a feedforward neural network for self-supervised fine-tuning, and the method further comprises: S31, randomly selecting a data dimension, and using a mask matrix to randomly eliminate the data of the dimension to form a partially missing input sequence; S32, inputting the masked sequence into a third feedforward neural network R, outputting the reconstruction data of the masked dimension from the network, and calculating the L1 loss value based on the real data and the reconstruction data.

[0014] In an embodiment of the present application, the supervised learning data set is constructed based on the data sequence segment, each dimension of data is embedded into a D-dimensional space and fused through a linear layer, and the model after self-supervised fine-tuning is supervised fine-tuned using real residual life data, and the method further comprises: S41, mapping each dimension of data in each data sequence segment to a D-dimensional feature space respectively to extract independent features of each channel; S42, splicing the feature matrix after mapping each channel and performing fusion processing through a linear embedding layer to form a unified embedding matrix for supervised fine-tuning.

[0015] To achieve the above purpose, the second aspect embodiment of the present application proposes a large model-based tunneling equipment key component residual life prediction device, comprising: a model architecture replacement module for replacing the feedforward network layer of a pre-trained large language model, adopting a sparse MoE architecture to realize dynamic feature routing processing; a data segmentation module for segmenting the multi-dimensional time series monitoring data of the tunneling equipment key component according to a set window size and sliding step to obtain a plurality of data sequence segments; a self-supervised data construction module for constructing a self-supervised learning data set based on the data sequence segment, calculating the reconstruction error by randomly masking a certain dimension of data and reconstructing the dimension of data using a feedforward neural network for self-supervised fine-tuning; a supervised data construction and fine-tuning module for constructing a supervised learning data set based on the data sequence segment, embedding each dimension of data into a D-dimensional space and fusing through a linear layer, and using real residual life data to supervise the fine-tuning of the model after self-supervised fine-tuning; a real-time prediction input module for segmenting the real-time collected tunneling equipment key component data according to the same window size and sliding step as the data segmentation module to form a standardized input sequence segment, and inputting the sequence segment into the large language model after supervised fine-tuning to output the residual life prediction result of the key component.

[0016] To achieve the above purpose, the third aspect embodiment of the present application proposes an electronic device, comprising: a processor, and a memory in communication connection with the processor; The memory stores computer execution instructions; The processor executes computer-executed instructions stored in the memory to implement the method of any one of the first aspect.

[0017] To achieve the above object, the fourth aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores computer-executed instructions, and the computer-executed instructions are executed by a processor to implement the method of any one of the first aspect.

[0018] To achieve the above object, the fifth aspect of the present application provides a computer program product, wherein the computer program is executed by a processor to implement the method of any one of the first aspect.

[0019] The embodiments of the present application provide at least the following beneficial effects: significantly improving the accuracy and generalization ability of the residual life prediction of the key components of the tunneling equipment, and effectively reducing the sudden failure rate and maintenance cost.

[0020] Additional aspects and advantages of the present application will be in part apparent and in part pointed out below in the description of the application. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the accompanying drawings, in which: Figure 1 is a flowchart of a large model-based residual life prediction method for key components of tunneling equipment according to an embodiment of the present application; Figure 2 is a structural schematic diagram of a large model-based residual life prediction device for key components of tunneling equipment according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0023] Figure 1 is a flowchart of a large model-based residual life prediction method for key components of tunneling equipment according to an embodiment of the present application;

[0024] As shown in Figure 1 , the large model-based residual life prediction method for key components of tunneling equipment includes the following steps: S1, replacing the feedforward network layer of the pre-trained large language model, and adopting a sparse MoE architecture to realize dynamic feature routing processing.

[0025] Specifically, in some implementations, step S1 involves replacing the feed-forward network (FFN) layer in the pre-trained Qwen3 large language model with a sparse mixed expert architecture to implement dynamic feature routing processing. The core technical principle of this step is to introduce a gating mechanism and multiple expert networks, so that the model can selectively activate the most relevant expert network when processing different input features, thereby improving the model's expression ability and computational efficiency.

[0026] Further, S1 includes: S11, replacing the standard feed-forward network (FFN) layer in the Qwen3 model with a sparse mixed expert (MoE) layer, wherein each expert network has different feature processing capabilities Specifically, in some implementations, the present patent method replaces the standard feed-forward network (FFN) layer in the Qwen3 model with a sparse mixed expert (MoE) layer to improve the model's feature extraction ability and computational efficiency for multi-dimensional time series data of key components of tunneling equipment. The MoE layer is composed of multiple expert networks (Experts) and a gating network (Gating Network), each expert network has different feature processing capabilities, and can dynamically select the optimal expert for calculation according to the characteristics of the input data, thereby realizing the sparse activation and efficient reasoning of the model.

[0027] In terms of technical implementation, this replacement process includes the reconstruction of the FFN module of the Transformer structure in Qwen3. Specifically, the two fully connected layers of the original FFN layer are replaced with multiple parallel expert networks, each expert network is composed of two or more fully connected layers, has an independent activation function (such as GELU or ReLU) and parameter initialization method. The gating network usually adopts linear transformation combined with Softmax or Top-k strategy to weight the input features and select the most relevant k experts for calculation. In this patent, the Top-2 strategy is preferred, i.e. two expert networks are activated for each input sample to balance the expression ability and computational overhead of the model.

[0028] In terms of parameter indicators, the number of expert networks is usually set to 8 to 16, the hidden layer dimension of each expert is 2048 to 4096, and the input and output dimensions are consistent with the original FFN layer. The output dimension of the gating network is the same as the number of experts, and its activation function can use Softmax or Top-k sparsification mechanism. In addition, the routing weight calculation of the MoE layer needs to satisfy the sparsity constraint, i.e. only a few experts are activated for each sample to reduce the consumption of computational resources. In the training process, a load balancing loss is introduced to prevent uneven use of experts and improve the overall generalization ability of the model.

[0029] At the application scenario level, this step is mainly applied in the model fine-tuning process of S3 and S4 stages. In self-supervised learning, the MoE layer can more effectively capture the context dependency of the masked channels; in supervised learning, its multi-expert structure helps the model learn feature representations under different fault modes, improving the accuracy of residual life prediction.

[0030] At the technical effect level, this replacement significantly enhances the model's adaptability to complex, high-dimensional time series data, while reducing the occupation of computing resources through the sparse activation mechanism, improving the inference efficiency and prediction accuracy of the model, and providing stronger technical support for the health management of key components of tunneling equipment.

[0031] S12, the input data is analyzed through a gating mechanism, and the expert network participating in the calculation is dynamically selected according to the data characteristics, to improve the adaptability of the model to multi-dimensional time series data.

[0032] Specifically, this step analyzes the input data through a gating mechanism and dynamically selects the expert network participating in the calculation according to the data characteristics, thereby improving the adaptability of the model to multi-dimensional time series data. In some implementations, the gating mechanism is integrated into the model architecture, and a sparse gating strategy based on MoE (Mixture of Experts) is used, replacing the feedforward network (FFN) layer in the traditional Transformer model with multiple expert networks, each responsible for processing a specific type or characteristic of data patterns.

[0033] At the technical implementation level, the gating mechanism is usually composed of a learnable gating network (Gating Network), which receives the embedded representation of the input data and outputs a probability distribution to determine which expert network should be assigned to each time step or feature dimension for processing. In this patent, the gating network optionally adopts a multi-layer perceptron (MLP) structure, with a D-dimensional embedding vector as input and K weight coefficients of K expert networks as output, where K is the number of experts and D is the data dimension. The output of the gating network is normalized by the Softmax function to ensure that only a small number of expert networks (such as 1-2) are activated at each time step, thereby achieving sparse activation, reducing computational complexity, and enhancing the model's interpretability.

[0034] At the parameter index level, a self-supervised learning task is introduced during the training process of the gating mechanism, which is optimized by masking part of the channel data and reconstructing the loss (L1). The masking ratio can be set to 10%-30%, and the masked channels are randomly selected to enhance the model's robustness to missing data. In addition, the activation threshold of the gating network, the width of the expert network (such as the number of neurons per layer), and the sparsity parameter (such as Top-1 or Top-2) can be adjusted according to the actual data distribution to achieve optimal performance.

[0035] At the application scenario level, this step is suitable for multi-dimensional state monitoring data processing of key components of tunneling equipment, such as joint modeling of vibration, temperature, pressure, current, and other sensor signals. In the real-time prediction stage, the gating mechanism can dynamically select the optimal expert network according to the characteristics of the current input sequence, thereby improving the adaptability of the model to different working conditions and fault modes.

[0036] At the technical effect level, this step effectively enhances the model's ability to process complex and variable multi-dimensional time series data, improves the accuracy and generalization of residual life prediction, and provides key technical support for intelligent operation and maintenance of tunneling equipment.

[0037] S2, the multi-dimensional time series monitoring data of the key components of the tunneling equipment is segmented according to the set window size and sliding step, obtaining a plurality of data sequence segments Specifically, in some implementations, step S2 involves segmenting the multi-dimensional time series monitoring data of the key components of the tunneling equipment according to the set window size and sliding step, thereby obtaining a plurality of data sequence segments. This step is a key data preprocessing link in the entire residual life prediction method, and its technical implementation is based on the sliding window mechanism, aiming to convert the continuously collected high-dimensional time series data into discrete sequence segments suitable for model input, to improve the accuracy and stability of subsequent model training and prediction.

[0038] Further, S2 includes: S21, representing the multi-dimensional state monitoring data X as a D×T real matrix, where D is the number of data dimensions and T is the number of time steps Specifically, in some implementations, step S2 involves representing the multi-dimensional state monitoring data X of the key components of the tunneling equipment as a D×T real matrix, where D is the number of data dimensions and T is the number of time steps. The core technical principle of this step is to structure the original multi-dimensional time series data into an input format suitable for deep learning model processing, thereby providing a standardized data basis for subsequent self-supervised learning and supervised learning.

[0039] In terms of specific operation, the tunneling equipment will collect the running state data of the key components through various sensors during operation, such as vibration signals, temperature, pressure, current, etc. These data are continuously collected in the time dimension, forming multi-channel time series signals. These signals are uniformly organized into a D×T real matrix, where each row corresponds to a monitoring dimension and each column corresponds to a time step, thereby constructing a two-dimensional time series data representation. In the data preprocessing stage, the original data is usually normalized to meet the numerical range requirements of model input, such as mapping the data to the [0, 1] interval or using the Z-score standardization method.

[0040] Further, to construct the training dataset, the matrix X is divided using a sliding window mechanism. The length S of the sliding window is a preset sequence segment length, and the sliding step B determines the interval of the window movement. In this scheme, the sliding step B is usually set to 1 or 2 to ensure the continuity of the time series and the data utilization. Through this mechanism, the original data can be divided into N sequence segments, where N = (T-S+B) / B. Each sequence segment is a D×S matrix, which is used as an input sample of the model for subsequent feature learning and life prediction.

[0041] This step plays a key role in the entire technical solution. On the one hand, it provides a structured and standardized input format for the model, facilitating subsequent embedding mapping and feature extraction; on the other hand, by reasonably setting the window length and sliding step, the modeling capability of the model for time dependence can be enhanced, and the accuracy and robustness of the remaining life prediction can be improved. In practical applications, this method is suitable for the state monitoring system of tunneling equipment under complex working conditions in coal mines, and can effectively support the development of equipment health management and predictive maintenance strategies.

[0042] S22, the matrix X is divided using a sliding window with a length of S to generate N data sequence segments, where N = T / S, and there is a partial overlap between adjacent sequence segments.

[0043] Specifically, in some implementations, step S2 involves sliding window division of the multi-dimensional state monitoring data X of the key components of the tunneling equipment to generate N data sequence segments. Specifically, the original monitoring data X is usually represented in the form of a D×T real matrix, where D is the number of data dimensions (such as vibration, temperature, pressure, etc. sensor signals), and T is the number of time steps, representing the length of the time series during equipment operation. To adapt to subsequent feature extraction and prediction tasks based on large language models, the matrix needs to be divided into multiple sliding window sequence segments with a length of S, each sequence segment being a D×S sub-matrix, thereby forming N data sequence segments, where N = T / S, and there is a partial overlap between adjacent sequence segments, with an overlap length of usually S / 2 or S / 4 to enhance the ability to capture time continuity and local features.

[0044] In terms of parameter settings, the length S of the sliding window needs to be selected according to the time domain characteristics of the equipment operating state changes, generally in the range of 50 to 500 time steps, and can be adjusted according to the sampling frequency (such as 100 Hz) and the degradation period of the key components of the equipment. For example, if the sampling frequency is 100 Hz and S is 500, it corresponds to 5 seconds of operating data, which helps to capture short-term dynamic characteristics. In addition, to ensure the stability and generalization ability of the model training, the division of the data sequence segments needs to meet the requirements of time alignment and data integrity, and a fixed step sliding strategy is usually used, with a step length of S / 2 or S / 4 to achieve information continuity between adjacent sequence segments.

[0045] This step is usually deployed in an industrial Internet of Things platform or an edge computing node in practical applications, and is used for online or offline processing of multi-dimensional sensor data collected in real time during the operation of the tunneling equipment. By sliding window segmentation, the original data can be converted into a sequence input format suitable for processing by a large language model, thereby providing structured data support for subsequent self-supervised learning and supervised learning stages.

[0046] Further, the technical effect of this step is that by introducing a sliding window mechanism, the time sequence characteristics of the equipment operating state can be effectively extracted, enhancing the model's ability to recognize local degradation patterns. At the same time, the partial overlap between sequence segments helps to improve the model's sensitivity to temporal continuity changes, thereby improving the accuracy and robustness of the remaining life prediction and providing a reliable data foundation for the intelligent operation and maintenance of the tunneling equipment.

[0047] S3, constructing a self-supervised learning data set based on the data sequence segments, calculating reconstruction error for self-supervised fine-tuning by randomly masking a certain dimension of data and reconstructing the dimension of data using a feed-forward neural network.

[0048] Specifically, in step S3, a self-supervised learning data set is constructed based on the data sequence segments of the key components of the tunneling equipment, and the core is to randomly mask a certain dimension of data and reconstruct it using a feed-forward neural network (FFN), thereby achieving self-supervised fine-tuning of the model. The technical implementation of this step is based on the multi-dimensional time sequence characteristics of the data, aiming to enhance the model's perception ability and robustness to key component state features.

[0049] In some implementations, the original data X is a real number matrix of D x T, where D is the dimension number of the monitored data (such as vibration, temperature, current, etc.), and T is the time step number. By setting a fixed window length S and a sliding step Δ, X is divided into N data sequence segments, N = (T-S) / Δ+1. Each sequence segment is a D x S sub-matrix, serving as an input unit for subsequent self-supervised learning.

[0050] Further, the construction of the self-supervised learning data set adopts a "masking-reconstruction" mechanism. Specifically, for each data sequence segment, a dimension (such as the i-th channel) is randomly selected, and a masking matrix M_i ∈ {0,1}^S is used to set part of the data in that dimension at the time step to zero or random noise, forming a partially missing input sequence. The masking ratio is usually set to 10% to 30% to ensure that the model has sufficient reconstruction challenges during the learning process.

[0051] Subsequently, the masked data sequence segment is input into the pre-trained Qwen3 model, and the context representation output by the model is sent to a third feedforward neural network R, which is responsible for reconstructing the masked dimension data. The reconstruction error is quantified by an L1 loss function, whose calculation formula is: wherein is the real data, is the reconstructed data. The loss function can effectively measure the model's ability to capture local features.

[0052] Further, S3 includes: S31, randomly selecting a data dimension and using a mask matrix to randomly eliminate the data in the dimension to form a partially missing input sequence.

[0053] Specifically, in some implementations, the step of "randomly selecting a data dimension and using a mask matrix to randomly eliminate the data in the dimension to form a partially missing input sequence" in step S3 is a key link of self-supervised learning data augmentation and feature learning. This step aims to artificially introduce data missing to force the model to learn the intrinsic relevance and context information of the data, thereby improving its modeling ability for the state sequence of the key components of the mining equipment.

[0054] In terms of technical implementation, the process first randomly selects a dimension d∈{1,2,...,D} from the input multi-dimensional time series data X, where D is the total number of monitored data dimensions, such as vibration, temperature, current, pressure, and other sensor signals. Subsequently, a mask matrix M∈{0,1}^T is constructed, where T is the number of time steps. The elements in the mask matrix are set to 0 with a certain probability p (e.g., p=0.15-0.3), indicating that the data at that time step is masked and eliminated, and the remaining elements are 1, indicating that the original data is retained. In actual operation, a Bernoulli distribution can be used to randomly generate the mask matrix to ensure the randomness and unpredictability of the masking process.

[0055] In terms of parameter indicators, the selection of the masking probability p needs to consider both the model learning efficiency and the data integrity. If p is too small, the model will have difficulty learning sufficient context information; if p is too large, the input sequence information may be too sparse, affecting model convergence. In this scheme, the masking probability p is usually set to 0.25, and the masking length can be set to continuous or non-continuous random segments, with a specific length range of [5, 20] time steps, to simulate the possible data loss or abnormal situations in actual operation.

[0056] At the application scenario level, this step is suitable for state monitoring data processing of tunneling equipment under complex working conditions. For example, in the coal mine underground environment, due to sensor failure or communication interference, part of the monitoring signal may appear intermittent missing. By simulating such missing in the training stage, the model can enhance the robustness of incomplete data and improve the generalization ability in actual deployment.

[0057] At the technical effect level, this step introduces a mask mechanism, forcing the model to reconstruct the masked dimension data in the self-supervised learning stage, thereby strengthening its ability to extract multi-dimensional time series features. Combined with the subsequent feedforward neural network R and loss function L1, the model can complete preliminary feature learning under unlabeled data conditions, providing higher quality embedding representation for the subsequent supervised learning stage, significantly improving the accuracy and stability of the remaining life prediction.

[0058] S32, input the masked sequence into the third feedforward neural network R, output the reconstructed data of the masked dimension from the network, and calculate the L1 loss value based on the real data and the reconstructed data.

[0059] Specifically, in some implementations, the masked sequence is input into a third feedforward neural network R, which outputs reconstructed data for the masked dimension, and an L1 loss value is calculated based on the true data and the reconstructed data, which is one of the key steps in the self-supervised learning fine-tuning phase of the method. The core technical principle of this step is to combine the masking mechanism with the reconstruction task, forcing the model to learn the internal structure and feature representation of the data, thereby improving its modeling ability for the status data of the key components of the tunneling equipment. Specifically, the input data is the multi-dimensional time series data segment of the key components of the tunneling equipment after sliding window segmentation, which is in the form of a real number matrix of D x S, where D is the number of data dimensions and S is the window length. In the self-supervised learning phase, the system will randomly select a dimension and use a masking matrix to set its data to zero or replace it with a special marker, forming a partially missing input sequence. This masked sequence is then input into the third feedforward neural network R, which is usually composed of several fully connected layers, with an input dimension of D x S and an output dimension consistent with the masked dimension, used to reconstruct the missing data. In terms of parameter settings, the hidden layer dimension of the third feedforward neural network R can be set to 256 or 512, the activation function can use ReLU or GELU, and the output layer uses linear activation to maintain numerical consistency. After reconstruction, the system compares the masked dimension in the true data with the reconstructed data output by the network R element by element and calculates the L1 loss value, which is the sum of the absolute errors of the two. The L1 loss function has the advantages of strong robustness to outliers and stable gradients, making it suitable for noise interference that may exist in the status data of the tunneling equipment. This step is usually run in a GPU-accelerated deep learning framework (such as PyTorch or TensorFlow) in practical applications, and the Adam optimizer is used for training with a learning rate set between 1e-4 and 5e-4 and a batch size of 32 or 64. Through the training of this step, the model can enhance its perception of the status features of the key components, providing higher quality feature representations for the remaining life prediction in the subsequent supervised learning phase, thereby improving the overall prediction accuracy and robustness.

[0060] S4, constructing a supervised learning data set based on the data sequence segment, embedding each dimension data into a D-dimensional space and fusing through a linear layer, and supervising and fine-tuning the self-supervised fine-tuned model using real remaining life data.

[0061] Specifically, in step S4, a supervised learning data set is constructed based on the data sequence segment, and each dimension data is fused through a linear layer, and finally the self-supervised fine-tuned model is supervised and fine-tuned using real remaining life data, which is a key link to realize accurate prediction of the remaining life of the key components of the tunneling equipment. The core of this step is to convert multi-dimensional time series data into supervised signals that can be learned by the model, and to improve the prediction ability of the model through an end-to-end fine-tuning process.

[0062] At the parameter index level, the dimension D of the embedding layer is usually set to be consistent with or slightly higher than the original data dimension (e.g., D = 64, 128) to retain sufficient information. The output dimension of the linear layer can be set to the model hidden layer dimension (e.g., 256 or 512) to adapt to the subsequent prediction module. In the supervised fine-tuning process, the mean square error (MSE) loss function L2 is used, and its calculation formula is:

[0063] where y is the true residual life value of the key component of the tunneling equipment, is the model output equipment key component residual life prediction calculation value. During training, the Adam optimizer is used, the learning rate is set to between 1e-4 and 1e-3, the batch size is 32 or 64, and the training epoch is usually 10-20 rounds to prevent overfitting.

[0064] Further, S4 includes: S41, mapping each dimension data in each data sequence segment to a D-dimensional feature space to extract independent features of each channel Specifically, this step involves mapping each dimension data in the multi-dimensional state monitoring data of the key component of the tunneling equipment to a D-dimensional feature space to extract independent features of each channel. In some implementations, this process is based on an embedding mapping mechanism, which independently converts each channel (i.e., each dimension) in the original data sequence segment to a feature space, thereby enhancing the model's perception ability and representation learning effect for multi-source heterogeneous data.

[0065] From the technical implementation level, this step first regards each data sequence segment as a D x T real matrix, where D is the number of data dimensions (such as vibration, temperature, pressure, etc. sensor channels), and T is the number of time steps. In the data preprocessing stage, the data of each channel will be independently input to the embedding layer (Embedding Layer), which is usually composed of a trainable linear transformation matrix with a dimension of D x D, which is used to map the original data to a D-dimensional feature space. Further, the embedding layer can combine normalization operations (such as Layer Normalization) and position encoding (Positional Encoding) to enhance the model's ability to model time series structures.

[0066] At the parameter index level, the output dimension D of the embedding layer is usually set according to the actual number of sensors and the complexity of features, for example, in tunneling equipment, D can take values of 16-32. In the embedding mapping process, the model optimizes parameters by minimizing reconstruction error (such as mean square error MSE) to ensure that the mapped features can retain the key information of the original data. In addition, the update frequency of the embedding matrix, the learning rate (such as 0.001-0.0001) and the regularization coefficient (such as L2 regularization coefficient λ = 0.01) all need to be adjusted according to the convergence during training.

[0067] In application scenarios, this step is suitable for state monitoring and life prediction of key components of tunneling equipment such as tunneling machines, hydraulic systems, transmission systems, etc. By independently mapping each channel data, the model can more effectively capture the non-linear relationships and local features between different sensor signals, thereby improving the prediction accuracy.

[0068] The technical effect of this step is that through independent channel feature mapping, the model's representation ability for multi-dimensional time series data is enhanced, providing high-quality embedding representation for subsequent self-supervised learning and supervised learning stages, thereby significantly improving the accuracy and robustness of the remaining life prediction.

[0069] S42, the mapped feature matrix of each channel is spliced and fused by a linear embedding layer to form a unified embedding matrix for supervised fine-tuning.

[0070] Specifically, in some implementations, steps S3 and S4 involve splicing the mapped feature matrix of each channel and fusing it through a linear embedding layer to form a unified embedding matrix for supervised fine-tuning, which is based on the principle of multi-modal feature fusion and embedding space alignment. Specifically, this step first independently embeds the state monitoring data (such as vibration, temperature, pressure, etc.) of each dimension in the data sequence segment of the key components of the tunneling equipment, i.e., maps the data of each channel to a unified D-dimensional embedding space. This embedding process can be implemented using a fully connected layer (Fully Connected Layer) or a learnable embedding matrix (Learnable Embedding Matrix) to ensure that data of different physical quantities are comparable and fusible in a unified semantic space.

[0071] Further, the embedding matrices of each channel are kept aligned in the time dimension, and then concatenated in the channel dimension to form a higher-dimensional feature tensor. For example, if the original data is a D x T matrix, and each channel embedding is a d x T matrix, the concatenated feature matrix is K x d x T, where K is the number of channels. The concatenated feature tensor is then reduced in dimension and fused by a linear embedding layer, which is usually composed of a trainable weight matrix and a bias term , and the output is a unified D-dimensional embedding matrix, which facilitates the input processing of subsequent supervised fine-tuning tasks.

[0072] In terms of parameter settings, the output dimension D of the embedding layer is usually set according to the size of the model's hidden layer, for example, it can be set to 768 or 1024 dimensions in the Qwen3 model. The concatenation operation can use depth-wise concatenation or channel-wise concatenation to preserve the timing information of each channel. This step serves as a bridge between self-supervised learning and supervised learning, improving the model's representation ability for multi-dimensional time series data through structured feature fusion, providing high-quality embedding representation for the subsequent remaining life prediction task, and significantly improving the prediction accuracy and robustness.

[0073] The method for predicting the remaining life of key components of tunneling equipment based on large models according to the embodiments of the present application effectively improves the accuracy of predicting the remaining life of key components of tunneling equipment, enhances the equipment health management capability, and improves the safety and efficiency of coal mine production.

[0074] S5, the real-time collected tunneling equipment key component data is segmented according to the same window size and sliding step as in S2 to form a standardized input sequence segment, and the sequence segment is input into the large language model after supervised fine-tuning to output the remaining life prediction result of the key component.

[0075] Specifically, in step S5, the real-time collected tunneling equipment key component data is segmented according to the same window size and sliding step as in S2 to form a standardized input sequence segment, and the sequence segment is input into the large language model after supervised fine-tuning to output the remaining life prediction result of the key component. This step is the final inference stage of the entire prediction method, and its technical implementation is based on the standardized processing of time series data and the online inference ability of the model.

[0076] In terms of technical implementation, the real-time collected data is usually a multi-dimensional time series in the form of a real number matrix D x T, where D represents the dimension of the monitoring parameters (such as vibration, temperature, pressure, current, etc.), and T represents the current timestamp. To maintain consistency with the training phase, the same sliding window mechanism is used for data segmentation. Specifically, by setting the window length S (for example, S = 512 time steps) and the sliding step L (for example, L = 64), the real-time data is divided into several sequence segments with a length of S by sliding in time steps. Each sequence segment can be represented as where t is the current time step, and S is the window length. This operation ensures that the model input format is consistent with the training phase, thereby improving the stability and accuracy of the prediction.

[0077] In terms of parameter indicators, the selection of window size S and sliding step L needs to balance the local feature capture ability of the time series and the computational efficiency. Usually, S is between 256 and 1024, and L is 1 / 8 to 1 / 4 of S, to ensure the continuity of the data and the real-time response ability of the model. In addition, the input sequence segment needs to be standardized, such as Z-score normalization or Min-Max normalization, to eliminate the dimensional differences of different sensor data and improve the model generalization ability.

[0078] In terms of application scenarios, this step is suitable for online health monitoring systems of coal mine underground tunneling equipment. By deploying edge computing devices or cloud servers, multi-source heterogeneous data from vibration sensors, temperature sensors, pressure sensors, etc. are received in real time, and sequence processing and model inference are performed, outputting the remaining life prediction results of key components (such as cutting motors, hydraulic systems, bearings, etc.), providing decision support for equipment maintenance.

[0079] In terms of technical effects, this step realizes the end-to-end mapping from raw monitoring data to remaining life prediction, and through the supervised fine-tuned large language model, it can effectively capture the degradation trend of key components and improve the prediction accuracy. The remaining life prediction value output by the model is usually in hours or working cycles, and the error rate can be controlled within 5%, which is significantly better than traditional statistical models and shallow neural network methods.

[0080] The large model-based remaining life prediction method for key components of tunneling equipment in the embodiments of the present application further improves the real-time performance and generalization ability of the remaining life prediction by dividing the real-time collected data into standardized input sequence segments and inputting them into the supervised fine-tuned large language model, making the prediction results more suitable for complex working condition changes, thereby optimizing the equipment maintenance strategy, reducing the risk of unplanned downtime, and improving the intelligent level and production efficiency of coal mines.

[0081] To realize the above-mentioned embodiments, the present application further proposes a large model-based remaining life prediction device for key components of tunneling equipment. Figure 2A structure schematic diagram of a large model-based tunneling equipment key component residual life prediction device is provided for an embodiment of the present application. As shown in the figure, Figure 2 The device comprises: A model architecture replacement module 100 is configured to replace the feedforward network layer of a pre-trained large language model, and adopt a sparse MoE architecture to realize dynamic feature routing processing. A data segmentation module 200 is configured to segment the multi-dimensional time series monitoring data of the tunneling equipment key component according to a set window size and sliding step, and obtain a plurality of data sequence segments. A self-supervised data construction module 300 is configured to construct a self-supervised learning data set based on the data sequence segments, mask a certain dimension data randomly, reconstruct the dimension data by using a feedforward neural network, and calculate the reconstruction error for self-supervised fine-tuning. A supervised data construction and fine-tuning module 400 is configured to construct a supervised learning data set based on the data sequence segments, embed each dimension data into a D-dimensional space and fuse through a linear layer, and perform supervised fine-tuning on the self-supervised fine-tuned model by using real residual life data. A real-time prediction input module 500 is configured to segment the real-time collected tunneling equipment key component data according to the same window size and sliding step as the data segmentation module, form a standardized input sequence segment, and input the sequence segment into the supervised fine-tuned large language model to output the residual life prediction result of the key component.

[0082] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment related to the method, and will not be described in detail here.

[0083] In order to realize the above-mentioned embodiments, the present application further proposes an electronic device, comprising: a processor, and a memory in communication connection with the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to realize the method provided by the foregoing embodiments.

[0084] In order to realize the above-mentioned embodiments, the present application further proposes a computer readable storage medium, the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by the processor to realize the method provided by the foregoing embodiments.

[0085] In order to realize the above-mentioned embodiments, the present application further proposes a computer program product, comprising a computer program, which is executed by the processor to realize the method provided by the foregoing embodiments.

[0086] The collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the present application comply with relevant laws and regulations and do not violate public order and good customs.

[0087] It should be noted that the personal information from the user should be collected for legal and reasonable purposes, and not shared or sold outside these legal uses. In addition, such collection / sharing should be carried out after the user's informed consent is received, including but not limited to informing the user to read the user agreement / user notice before the user uses the function, and signing the agreement / authorization including authorization of relevant user information. In addition, any necessary steps should be taken to protect and secure access to such personal information data and ensure that other people with access to personal information data comply with their privacy policies and processes.

[0088] The present application is expected to provide embodiments in which the user can selectively prevent the use or access of personal information data. That is, the present disclosure is expected to provide hardware and / or software to prevent or block access to such personal information data. Once the personal information data is no longer needed, the risk is minimized by limiting data collection and deleting data. In addition, such personal information is de-identified, if applicable, to protect the privacy of the user.

[0089] In the foregoing embodiment descriptions, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples, without contradiction.

[0090] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0091] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0092] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0093] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0094] Those skilled in the art can understand that all or part of the steps of the method of the above-mentioned embodiments can be instructed by a program to complete the relevant hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0095] In addition, each functional unit in each embodiment of the present application can be integrated into one processing module, or each unit can exist physically independently, or two or more units can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. When the integrated module is realized in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0096] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application. Those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

[0097] It should be understood that various forms of flow shown above can be used to reorder, add or delete steps. For example, each step described in the present application can be executed in parallel, sequentially or in different order, as long as the desired results of the technical solutions of the present application can be achieved, which is not limited herein.

[0098] The above detailed description does not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and replacements can be made according to design requirements and other factors. Any modification, equivalent replacement and improvement within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for predicting the remaining life of key components of tunneling equipment based on a large model, characterized in that, include: S1 replaces the feedforward network layer of the pre-trained large language model and adopts a sparse MoE architecture to achieve dynamic feature routing processing. S2, the multi-dimensional time series monitoring data of key components of the tunneling equipment is divided according to the set window size and sliding step size to obtain multiple data sequence segments; S3, construct a self-supervised learning dataset based on the data sequence segment, reconstruct the data of a certain dimension by randomly masking the data of a certain dimension and using a feedforward neural network, and calculate the reconstruction error to perform self-supervised fine-tuning; S4. Based on the data sequence segment, construct a supervised learning dataset, embed data of each dimension into a D-dimensional space and fuse them through a linear layer, and use real remaining lifetime data to supervise and fine-tune the self-supervised fine-tuning model. S5 divides the real-time collected key component data of the tunneling equipment into standardized input sequence segments according to the same window size and sliding step size as in S2. These sequence segments are then input into the supervised fine-tuned large language model to output the remaining life prediction results of the key components.

2. The method as described in claim 1, characterized in that, The replacement of the feedforward network layer of the pre-trained large language model with a sparse MoE architecture to achieve dynamic feature routing processing also includes: S11, replace the standard feedforward network layer in the Qwen3 model with a sparse hybrid expert layer, where each expert network has different feature processing capabilities; S12 analyzes the input data through a gating mechanism and dynamically selects the expert network to participate in the calculation based on the data characteristics, so as to improve the model's adaptability to multidimensional time series data.

3. The method as described in claim 1, characterized in that, The process of dividing the multidimensional time-series monitoring data of key components of the tunneling equipment into multiple data sequence segments according to a set window size and sliding step size also includes: S21, Represent the multidimensional state monitoring data X as a D×T real matrix, where D is the number of data dimensions and T is the number of time steps; S22. A sliding window of length S is used to divide matrix X into N data sequence segments, where N = T / S, and there is partial overlap between adjacent sequence segments.

4. The method as described in claim 1, characterized in that, The step of constructing a self-supervised learning dataset based on the data sequence segment, reconstructing the data in a certain dimension by randomly masking the data in a certain dimension and using a feedforward neural network, and calculating the reconstruction error for self-supervised fine-tuning, further includes: S31, randomly select a data dimension and use a mask matrix to randomly eliminate the data in that dimension to form a partially missing input sequence; S32, the masked sequence is input into the third feedforward neural network R, which outputs the reconstructed data of the masked dimension, and calculates the L1 loss value based on the real data and the reconstructed data.

5. The method as described in claim 1, characterized in that, The process of constructing a supervised learning dataset based on the data sequence segments, embedding data from each dimension into a D-dimensional space and fusing them through a linear layer, and using real remaining lifetime data to perform supervised fine-tuning of the self-supervised fine-tuning model also includes: S41, map the data of each dimension in each data sequence segment to the D-dimensional feature space to extract the independent features of each channel; S42 concatenates the feature matrices mapped from each channel and fuses them through a linear embedding layer to form a unified embedding matrix for supervised fine-tuning.

6. A device for predicting the remaining life of key components of tunneling equipment based on a large model, characterized in that, include: The model architecture replacement module is used to replace the feedforward network layer of the pre-trained large language model, and adopts a sparse MoE architecture to achieve dynamic feature routing processing. The data segmentation module is used to divide the multi-dimensional time series monitoring data of key components of tunneling equipment according to the set window size and sliding step size to obtain multiple data sequence segments; The self-supervised data construction module is used to construct a self-supervised learning dataset based on the data sequence segment, reconstruct the data in a certain dimension by randomly masking the data in a certain dimension and using a feedforward neural network, and calculate the reconstruction error for self-supervised fine-tuning. The supervised data construction and fine-tuning module is used to construct a supervised learning dataset based on the data sequence segment, embed data of each dimension into a D-dimensional space and fuse them through a linear layer, and use real remaining lifetime data to perform supervised fine-tuning on the self-supervised fine-tuning model. The real-time prediction input module is used to divide the real-time collected key component data of the tunneling equipment into standardized input sequence segments according to the same window size and sliding step size as in the data segmentation module. These sequence segments are then input into the supervised fine-tuned large language model to output the remaining life prediction results of the key components.

7. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-5.

9. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1-5.

Images