Mine filling pipeline operation state monitoring method and device

Abstract

This invention discloses a method and device for monitoring the operational status of mine backfill pipelines. The method involves acquiring operational status data of the mine backfill pipelines and creating datasets for different operational statuses. A multi-scale data partitioning mechanism is constructed based on the operational status data, and the data is processed to obtain multiple scale components. A deep learning intelligent model is built to perform real-time analysis and status discrimination of the processed data. An artificial intelligence optimization algorithm is constructed to optimize the model parameters, enabling it to react quickly and accurately monitor the operational status of the backfill pipelines, reducing false alarms. Based on the discrimination results, the operational status of the mine backfill pipelines is predicted in real time, allowing for earlier warnings before anomalies occur, while significantly improving identification accuracy and real-time performance.

Description

Technical Field

[0001] This invention relates to the field of mine backfill monitoring and industrial big data analysis technology, and in particular to a method and device for monitoring the operating status of mine backfill pipelines based on a high-order statistical optimization deep learning model. Background Technology

[0002] Mine backfill pipelines are critical facilities for ensuring safe mine production, controlling surface subsidence, and protecting the ecological environment. During operation, the slurry inside the pipeline flows at high speeds, is highly abrasive, and operates under complex and variable conditions, making it extremely prone to malfunctions such as pipe blockage, pipe bursts, and wear leaks. Once a malfunction occurs, it can not only cause production interruptions and economic losses, but may even lead to major safety accidents.

[0003] Currently, monitoring the status of filled pipelines mainly relies on traditional methods such as pressure detection, pulse echo analysis, resistivity measurement, and acoustic monitoring. However, these methods are mostly post-contamination diagnostic technologies, only detectable and identifiable when blockage has occurred or progressed to a certain extent, making it difficult to provide early warning and trend prediction of blockage risks. Due to the lack of real-time, intelligent sensing and prediction capabilities for pipeline operating status, existing technologies struggle to provide timely guidance for production control and risk intervention, severely impacting mine safety and operational efficiency.

[0004] Some methods can monitor the operating status of filled pipelines, but their accuracy in identification is insufficient. They are generally general anomaly detection methods that do not distinguish between types and can only determine whether there is a blockage. Or their early warning capabilities are still insufficient, and they can only identify blockages when they are on the verge of blockage or when they are already blocked. At the same time, in terms of data acquisition, most existing methods use historical temperature or vibration data, and the accuracy of the identification model is not high enough to achieve multi-fault type identification. In addition, most of them are decomposition, prediction and reconstruction or reconstruction error comparison, which also leads to insufficient real-time performance of identification. Summary of the Invention

[0005] This invention provides a method and device for monitoring the operating status of mine backfilling pipelines, which solves the technical problem that existing methods cannot monitor the operating status of backfilling pipelines in a timely manner, thereby affecting the normal production and safety of mines.

[0006] On the other hand, compared with existing methods, the present invention can significantly improve recognition accuracy and real-time performance.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for monitoring the operational status of mine backfill pipelines includes the following steps: Acquire multi-dimensional time-series data during the operation of mine backfilling pipelines, including vibration signals, temperature signals, and pressure signals; A signal feature coupling effect model is constructed to decouple the multi-dimensional time series data. The decoupled multi-dimensional time series data is then divided into multiple scales to obtain multiple multi-scale subsequence data with different time granularities. A hybrid deep learning model based on an improved Transformer and a bidirectional LSTM is constructed. The hybrid deep learning model is used to extract features and determine the state of the multi-scale subsequence data to obtain the determination result of the pipeline operation state. The hyperparameters of the hybrid deep learning model are iteratively optimized based on the HSO optimization algorithm to obtain the optimized monitoring model. Based on the optimized monitoring model, the operational status of mine backfilling pipelines is monitored and early warnings are issued in real time.

[0008] In the above technical solution, a fiber optic sensor array deployed on the filling pipeline system is used to collect multi-source time-series data of the pipeline in real time under various operating conditions. This includes real-time acquisition of multi-dimensional time-series data such as vibration, temperature, and pressure.

[0009] Preferably, the fiber optic sensor array of the present invention may include at least one of a distributed fiber optic acoustic vibration sensor, a fiber optic grating temperature sensor, and a fiber optic pressure sensor.

[0010] Specifically, multi-dimensional time-series data are collected in real time by an optical fiber sensor array deployed on the filling pipeline system. Vibration acceleration signals, temperature field distribution signals and internal pressure signals during pipeline operation are collected synchronously at a preset sampling frequency to form a multi-dimensional time-series dataset with time labels.

[0011] In the above technical solution, the multi-dimensional time-series data is decoupled to separate the original vibration signal, interference signal, and coupling characteristics of each physical field, resulting in multiple data classification results with different categories and labels, including: Based on historical operation and maintenance records and expert knowledge, a signal feature coupling effect model is constructed, and an intelligent model is built to solve the signal features of the operating status. Define typical operating states of pipelines, including: normal state, slight wear state, critical blockage state, severe blockage state, leakage state, etc., and label the collected data with state labels; The raw data undergoes preprocessing operations such as cleaning, denoising, alignment, and normalization to form a well-organized time-series dataset. The processed data is divided into training set, validation set and test set.

[0012] In the above technical solution, the step of constructing a signal feature coupling effect model to decouple the multi-dimensional time-series data includes: Establish a coupling effect model between vibration signals and temperature and flow velocity signals: An end-to-end deep learning network is constructed to solve the coupling effect model. The deep learning network includes a multimodal feature extraction layer, a cross-modal fusion module, and a signal separation module, and outputs the original vibration signal estimate and the interference signal estimate.

[0013] In the above technical solution, the end-to-end deep learning network solves the coupling effect model, including: A one-dimensional convolutional neural network was used to extract local features from vibration signals, temperature signals, and flow velocity signals to obtain feature representations for each mode. Temperature and flow velocity signals are time-series encoded using a long short-term memory network to obtain temperature and flow velocity characteristics. Multi-head attention mechanism is used to fuse features from various modalities. Attention weights are calculated through query-key-value projection to obtain the fused multi-modal features. The interference signal components are estimated using a gating mechanism, and the interference is removed from the fused features. The original vibration signal is then recovered through residual blocks and deconvolution operations.

[0014] In the above technical solution, the decoupled multi-dimensional time-series data is divided into multiple scales, including: The input time-series signal sequence is segmented into multiple non-overlapping subsequences of length L. The length of each subsequence is denoted as P, and the span between consecutive subsequences is denoted as S. The number of subsequences generated is... ; S repeated values ​​are padded at the end of the sequence, and each subsequence is mapped to the Transformer latent space through a trainable linear projection. Define a set of multiple subsequence size values, with each subsequence size corresponding to a scale division, to obtain multiple subsequence data groups with different time granularities.

[0015] In the above technical solution, the improved Transformer includes a multi-scale parallel encoder and a router mechanism, which are used to capture local features at different granularities and long-distance dependencies across granularities.

[0016] In the above technical solution, the hybrid deep learning model is a hybrid model of temporal Transformer model and Long Short-Term Memory Network (LSTM): it uses self-attention mechanism to dynamically calculate the importance of data at different time points and globally perceives temporal dependencies; it uses LSTM to capture the long-term dependencies of data in the time dimension.

[0017] In the above technical solution, the hybrid deep learning model improves the Transformer model to suit the mining environment, including: The self-attention mechanism is improved by proposing an improved encoder module that can process multi-scale data in parallel and capture features within a single granularity. It can also capture long-distance dependencies and local patterns. The input data is weighted and summed by multiple attention heads to extract key features.

[0018] The above technical solutions improve the Transformer model by including multi-scale parallel encoders and router mechanisms, including: For each scale of subsequence data, its block embedding is concatenated with the corresponding router embedding to form an intermediate sequence, and an intragranular self-attention mechanism is applied to update both the block embedding and the router embedding simultaneously. The initialization method for the embedded router is as follows: ; in: It is a position embedding matrix with dimensions G×D, where G is a sufficiently large number and D is the embedding dimension. It is the first i Learnable granular embeddings at each granularity, with a dimension of 1× D ; It is the first i Number of blocks at each granularity.

[0019] By aggregating feature information at different granularities through a router mechanism, cross-granularity feature interaction and complementarity can be achieved, avoiding the sharp increase in computational complexity caused by performing large-scale self-attention operations on block embeddings of all granularities.

[0020] In the above technical solution, the Long Short-Term Memory (LSTM) network is spliced ​​into two layers, one forward and one backward. The forward and backward dependencies of the sequence are learned by the LSTM layers in the forward and backward directions, respectively. Finally, the outputs of the two directions are merged to obtain a more comprehensive representation of the sequence features.

[0021] In the above technical solution, the bidirectional LSTM includes a forward LSTM layer and a reverse LSTM layer, comprising: The forward LSTM layer processes the input sequence in chronological order, learns the forward dependencies of the sequence, and obtains the forward hidden state sequence. The inverse LSTM layer processes the input sequence in reverse chronological order, learns the backward dependencies of the sequence, and obtains the inverse hidden state sequence. By concatenating or summing the forward and reverse hidden state sequences in the time dimension, the final hidden state representation that integrates the bidirectional dependencies is obtained. The final hidden state is mapped to a state discrimination output through a linear layer.

[0022] In the above technical solution, the parameters of the hybrid model are optimized according to the HSO algorithm to obtain the optimized model, including: During model training, the constructed HSO optimization algorithm is integrated to iteratively optimize the model parameters until the model achieves optimal performance on the validation set.

[0023] In the above technical solution, the iterative optimization of the hyperparameters of the hybrid deep learning model based on the HSO optimization algorithm includes: Initialize the population and randomly generate a matrix of candidate solutions; Calculate the root mean square of all fitness values, and calculate the displacement vector coefficients based on the difference between each fitness value and the root mean square. Search for the individual's location and iteratively update the individual's location; Solution selection is based on an adaptive simulated annealing mechanism, including dynamic temperature updates for each iteration; Perform adaptive mutation operations, dynamically adjust the mutation rate and the variable time, extensively explore the search space in the initial iteration, and focus on local optimization in the later stages of the iteration; Iteratively execute the above steps until the termination condition is met, and output the optimal hyperparameter combination.

[0024] The above technical solution utilizes a hybrid deep learning model to extract features and determine states from multi-scale subsequence data, including: Define typical operating conditions of pipelines, including normal condition, slight wear condition, critical blockage condition, severe blockage condition, and leakage condition; The collected historical data is labeled with status tags to form a training dataset with category labels; The training dataset is input into the hybrid deep learning model for supervised learning, and the model outputs a state discrimination result. The target state is determined based on the state discrimination result. The target state includes normal state, critical blockage state, and severe blockage or leakage state.

[0025] For example, a pipeline transportation status classification level Z can be set for mine backfilling pipelines, with Z divided into 3 levels; based on the pipeline transportation status classification level and the model judgment result Q, where Q takes a value between 1 and 9, the target level is determined. See Table 1 below for the pipeline's full life cycle status correspondence table.

[0026] Table 1. Correspondence between the statuses of the pipeline throughout its entire life cycle.

[0027] In the above technical solution, the operation of the mine backfilling pipeline is monitored based on the pipeline status operation data: The operational status data of the mine filling pipeline is classified and categorized with different targets. Based on the classification and target of the operational status data, determine the target level; The operational status of the mine filling pipeline is monitored according to the target level.

[0028] In the above technical solution, the real-time monitoring and early warning of the operating status of the mine backfilling pipeline based on the optimized monitoring model includes: The multi-dimensional time-series data collected in real time is input into the optimized monitoring model to obtain the state judgment result at the current moment; The current operating status level of the pipeline is determined by comparing the status judgment result with the preset threshold level. When the judgment result indicates that the pipeline is in a state of slight wear, critical blockage, or severe blockage, the graded early warning mechanism is triggered, and the corresponding level of early warning information is sent to the monitoring terminal. Predict the development trend of pipeline faults based on the changing trends of status levels, and provide handling suggestions for maintenance personnel.

[0029] In addition, this invention also provides a monitoring device for the operating status of mine backfilling pipelines, including: The multimodal data acquisition module is used to acquire multi-dimensional time-series data of the mine backfilling pipeline during operation through an optical fiber sensor array. The multi-dimensional time-series data includes vibration signals, temperature signals, and pressure signals. The signal decoupling processing module is used to construct a signal feature coupling effect model, decouple the multi-dimensional time series data, and separate the original vibration signal, interference signal and coupling characteristics of each physical field. The multi-scale data partitioning module is used to partition the decoupled multi-dimensional time series data at multiple scales to obtain multiple subsequence data with different time granularities. A deep learning model building module is used to build a hybrid deep learning model based on an improved Transformer and a bidirectional LSTM. The improved Transformer includes a multi-scale parallel encoder and a router mechanism. The HSO optimization module is used to iteratively optimize the hyperparameters of the hybrid deep learning model based on the HSO optimization algorithm to obtain the optimized monitoring model. The pipeline monitoring module is used to extract features and determine the status of the multi-scale subsequence data using the optimized monitoring model, and to monitor and warn of the operation status of the mine backfilling pipeline in real time based on the determination results.

[0030] In the above technical solution, the multimodal data acquisition module includes: Distributed fiber optic acoustic vibration sensors are deployed along the filling pipeline system to collect pipeline vibration acceleration signals; Fiber Bragg grating temperature sensor is used to collect temperature field distribution signals in pipelines; Fiber optic pressure sensors are used to collect pressure signals inside pipelines. The synchronous acquisition unit is used to synchronously acquire signals from each sensor at a preset sampling frequency to form a multi-dimensional time-series dataset with time labels.

[0031] In the above technical solution, the signal decoupling processing module includes: The coupling effect model building unit is used to establish a mathematical model of the coupling effect between vibration signals and temperature and flow velocity signals. An end-to-end deep learning network unit, including a multimodal feature extraction layer, a cross-modal fusion module, and a signal separation module, is used to solve the coupling effect model and output the original vibration signal estimate. The multimodal feature extraction layer employs a one-dimensional convolutional neural network, the cross-modal fusion module uses a multi-head attention mechanism, and the signal separation module employs a gating mechanism and a residual network structure.

[0032] In the above technical solution, the multi-scale data partitioning module includes: The multi-scale subsequence segmentation unit is used to segment the input time-series signal into subsequences at multiple scales to obtain subsequences with different time granularities. Linear projection unit, used to map each subsequence to the Transformer latent space through trainable linear projection; The scale set management unit is used to manage a set of multiple subsequence size values ​​to achieve parallel output of multi-scale data.

[0033] In the above technical solution, the improved Transformer in the deep learning model building module includes: A multi-scale parallel encoder is used to process multiple subsequence data with different time granularities in parallel. The router mechanism unit is used to introduce router embeddings to aggregate feature information at different granularities, enabling cross-granularity feature interaction. Intragranular self-attention units are used to apply self-attention mechanisms to block embeddings and router embeddings within each granularity to capture local features and long-range dependencies. The router embeds initialization fusion location embedding and granularity embedding information.

[0034] In the above technical solution, the bidirectional LSTM in the deep learning model construction module includes: The forward LSTM layer processes the input sequence in chronological order and learns the forward temporal dependencies; The reverse LSTM layer processes the input sequence in reverse chronological order to learn backward temporal dependencies. The bidirectional fusion unit is used to fuse the forward hidden state and the reverse hidden state to obtain a comprehensive sequence feature representation; The output mapping unit is used to map the fused hidden states to state discrimination outputs through a linear layer.

[0035] In the above technical solution, the HSO optimization module includes: Population initialization unit, used to randomly generate candidate solution matrix; The higher-order statistics calculation unit is used to calculate the root mean square of the fitness value and the displacement vector coefficients. The location update unit is used to update the individual location based on a weighted differential strategy; An adaptive simulated annealing unit is used to dynamically adjust temperature parameters and accept new solutions probabilistically. Adaptive mutation unit is used to dynamically adjust the mutation rate and mutation length to balance global exploration and local development; The iterative control unit is used to control the iterative execution and termination determination of the optimization process.

[0036] In the above technical solution, the state discrimination and monitoring module includes: The status category definition unit is used to define five typical operating states: normal state, slight wear state, critical blockage state, severe blockage state, and leakage state. The supervised learning unit is used to train the hybrid deep learning model using a labeled training dataset and outputs state discrimination results with values ​​ranging from 1 to 9. The level determination unit is used to determine the target state level based on the state discrimination result value: 1-3 is the normal state, 4-6 is the critical pipe blockage state, and 7-9 is the severe pipe blockage or leakage state. The graded early warning unit is used to trigger the corresponding early warning mechanism according to the status level and send early warning information to the monitoring terminal; The trend prediction unit is used to predict the development trend of the fault based on the changing trend of the status level and generate handling suggestions.

[0037] Based on the above-mentioned method and device for monitoring the operational status of mine backfill pipelines, the present invention also provides the following: A mine backfill pipeline operation status monitoring system includes: The mine backfill pipeline operation status monitoring device described in any of the above items; A data transmission network is used to transmit the data collected by the fiber optic sensor array to the monitoring device; The monitoring terminal is used to receive the judgment results and early warning information from the monitoring device, display the pipeline operating status, and provide a human-machine interface. The cloud server is used to store historical monitoring data, model parameters, and operation and maintenance records, and supports big data analysis and remote model updates.

[0038] An electronic device, comprising: Processor and memory; The memory stores computer program instructions, which, when executed by the processor, implement the steps of the mine backfilling pipeline operation status monitoring method described above.

[0039] A computer-readable storage medium, characterized in that the storage medium stores computer program instructions, which, when executed by a processor, implement the steps of the mine backfill pipeline operation status monitoring method described above.

[0040] In summary, the mine backfilling pipeline operation status monitoring method and device provided by this invention acquires multi-source time-series data during the operation of the mine backfilling pipeline to construct a comprehensive and high-quality dataset based on different operating states. Then, a hybrid deep learning model combining temporal Transformer and bidirectional LSTM is built, utilizing an improved self-attention mechanism and multi-scale parallel processing capabilities to capture global features and local patterns of the pipeline data. The model parameters are iteratively optimized using the HSO optimization algorithm to improve model performance and generalization ability. Finally, based on the optimized model, the pipeline operating status is identified and warned in real time. This allows for timely detection of abnormal states before serious wear, blockage, or leakage occurs in the pipeline, enabling staff to take appropriate measures based on the monitoring results, effectively ensuring the safety and continuity of mine backfilling operations.

[0041] Compared with existing technologies, the present invention has the following significant advantages: First, it achieves collaborative sensing and precise decoupling of multi-dimensional physical fields, significantly improving the signal-to-noise ratio and physical interpretability of monitoring data.

[0042] Existing technologies mainly rely on temperature and vibration sensors to collect temperature and vibration data, which has a single data dimension and does not consider the coupling interference of physical fields such as temperature and flow velocity on vibration signals; some existing technologies only collect flow and pressure parameters and cannot obtain vibration response information of pipeline structures.

[0043] This invention innovatively employs an optical fiber sensor array (including a distributed optical fiber acoustic vibration sensor (DAS), a fiber optic grating temperature sensor, and an optical fiber pressure sensor) to achieve synchronous monitoring of vibration, temperature, and pressure in three dimensions, thus constructing a complete sensing system covering the pipe structure response, thermodynamic state, and fluid dynamics characteristics.

[0044] More importantly, this invention addresses the complex physical environment of mine backfill pipelines by establishing a mathematical model of the vibration-temperature-velocity coupling effect and designing an end-to-end deep learning decoupling network. This network, through multimodal feature extraction, cross-modal attention fusion, and gated interference separation, effectively separates the original vibration features, temperature drift components, and velocity disturbance components from the observed vibration signal, overcoming the technical bottleneck of "signal aliasing and feature submersion" in traditional methods. Practical verification shows that the signal-to-noise ratio of the decoupled signal is improved by 15-20 dB, providing high-quality input data for subsequent state recognition.

[0045] Second, a multi-scale spatiotemporal feature collaborative extraction mechanism was constructed, which breaks through the limitations of traditional single-scale analysis.

[0046] Existing technologies mostly employ single-scale signal decomposition methods, such as VMD decomposition, which struggle to simultaneously capture both rapid transient features and slow trend features during pipeline operation. This invention innovatively proposes a Transformer architecture combining multi-scale subsequence segmentation and a router mechanism. By dividing the original time-series data into multiple time granularities, it achieves multi-scale subsequence (Patch) functionality. Fine-grained patch: Captures transient features such as local vibration anomalies and pressure pulsations in the early stages of pipe blockage; Coarse-grained patch: Identifies long-term trend characteristics such as cumulative pipe wear and gradual temperature field changes; Router mechanism: Through learnable granular embedding and cross-granularity attention, it enables adaptive fusion and information complementarity of features at different scales.

[0047] The architecture significantly reduces computational complexity, with the number of subsequences being much smaller than the original sequence length. While maintaining the ability to model long-distance dependencies, it significantly reduces computational overhead. Compared to the complexity of the standard Transformer, the method of this invention improves efficiency by 3-5 times when processing long sequences, meeting the timeliness requirements of real-time monitoring of mine filling pipelines.

[0048] Third, a high-order statistical optimization algorithm for HSO was designed to achieve adaptive and efficient optimization of hyperparameters of deep learning models.

[0049] To address the challenges of hyperparameter tuning in deep learning models and the tendency of traditional optimization algorithms to get trapped in local optima, this invention proposes a swarm intelligence optimization algorithm based on Higher-Order Statistics Optimization (HSO).

[0050] By using the root mean square and difference coefficient of fitness values ​​to guide the search direction, compared with traditional strategies based on mean or optimal values, the population distribution is more accurately characterized and the convergence speed is improved by more than 30%. By using adaptive simulated annealing to dynamically adjust temperature parameters and acceptance probability, population diversity is maintained in the early stage of optimization, and local fine-grained search is strengthened in the later stage, effectively avoiding premature convergence. By utilizing a dynamic mutation mechanism, the mutation rate and step size are adaptively adjusted according to the iteration progress, achieving a smooth transition from "wide-area exploration" to "local development".

[0051] In the task of hyperparameter optimization of the Transformer-LSTM hybrid model, the application of this invention in the mining field improves the optimization efficiency by 25% and the accuracy of the model validation set by 3-5 percentage points compared with the commonly used GOOSE algorithm and crow search algorithm.

[0052] Fourth, a fine-grained status classification system of "five categories and nine levels" has been established, realizing status identification and early warning throughout the entire life cycle of pipelines, and at the same time, it can provide early warnings earlier than traditional methods.

[0053] Existing technologies mainly focus on predicting congestion risks, and the status classification is relatively coarse, generally divided into three levels: normal, warning, and congestion; or they can only distinguish between normal and abnormal, but cannot identify specific fault types.

[0054] Based on the failure mechanisms and operation and maintenance experience of mine backfill pipelines, this invention innovatively establishes a "five-category, nine-level" status classification system. This classification system achieves full life-cycle coverage from early wear to severe failure, enabling operation and maintenance personnel to detect potential problems at the slight wear stage. Compared with traditional methods, it provides a 2-4 hour earlier warning of critical pipe blockage, buying valuable time for emergency response and effectively avoiding serious accidents such as pipe blockage and pipe bursts.

[0055] Fifth, an end-to-end supervised deep learning paradigm has been formed, and the state recognition accuracy is significantly better than that of unsupervised methods.

[0056] Compared to unsupervised GAN methods that can only distinguish between normal and abnormal, the supervised deep learning paradigm of this invention has the following advantages: More accurate state recognition: By training with a large amount of labeled data, the model directly learns the mapping relationship from multimodal signals to state categories, achieving a classification accuracy of over 95%, while the GAN method relies on reconstruction error threshold judgment, resulting in a higher misclassification rate for abnormal samples. It can clearly distinguish different fault types such as wear, blockage, and leakage, providing a basis for targeted maintenance, while the GAN method can only indicate "abnormality" and cannot determine the nature of the abnormality; The output results are quantifiable and support trend analysis and remaining lifetime prediction, while the confidence calculation of GAN is relatively coarse.

[0057] VI. The system architecture is highly integrated, easy to implement, highly applicable to engineering projects, and has good promotional value.

[0058] The monitoring device of this invention adopts a modular design, with each functional unit (data acquisition, signal decoupling, multi-scale partitioning, deep learning model, HSO optimization, and state discrimination) coupled through a standard interface, facilitating flexible configuration according to actual mine conditions. Sensor layer: Fiber optic sensor arrays support distributed deployment along the pipeline, with a monitoring distance of up to tens of kilometers and a spatial resolution of meters, making them suitable for long-distance filling pipelines; Edge computing layer: Model inference can be completed in real time on edge devices with a response latency of less than 100ms, meeting the requirements of online monitoring; Cloud-based collaboration layer: Supports cloud storage of historical data, remote model updates, and big data mining, enabling intelligent operation and maintenance with "single-point monitoring and global optimization".

[0059] In summary, this invention, through system innovation of multimodal perception, physical decoupling, multi-scale deep learning, intelligent optimization, and fine-grained classification, constructs a complete, advanced, and practical solution for monitoring the operational status of mine backfilling pipelines. It significantly outperforms existing technologies in terms of monitoring dimensions, analysis depth, identification accuracy, and early warning timeliness, and is of great value in improving the safety, reliability, and intelligence level of mine backfilling systems. Attached Figure Description

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 A schematic flowchart of an embodiment of the mine backfill pipeline operation status monitoring method provided by the present invention; Figure 2 This is a schematic flowchart of an embodiment of the HSO algorithm provided by the present invention; Figure 3 A schematic flowchart of an embodiment of the mine backfill pipeline operation status monitoring provided by the present invention; Figure 4 A schematic diagram of an embodiment of the mine filling pipeline blockage monitoring device provided by the present invention; Figure 5 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] Example 1 like Figure 1 As shown in the illustration, a specific embodiment of the present invention discloses a method for monitoring the operational status of mine backfilling pipelines, belonging to the field of mine backfilling monitoring technology. The method includes: acquiring operational status data of the mine backfilling pipelines; creating datasets for different operational statuses; constructing a multi-scale data partitioning based on the operational status data; processing the operational status data to obtain multiple scale data components; constructing an encoder-based router mechanism to enable complementary multi-scale data channels and complete data telomeres; constructing a deep learning intelligent model to perform real-time analysis and status discrimination of the processed data; and constructing an artificial intelligence optimization algorithm to optimize the model parameters, enabling it to react quickly and accurately monitor the operational status of the backfilling pipelines, reducing false alarms. Real-time prediction of the operational status of the mine backfilling pipelines based on the discrimination results can provide timely alarms before anomalies occur, allowing staff to take appropriate action based on the monitoring situation, thereby ensuring normal mine production and safety.

[0063] Specifically, it includes: S101. Obtain data on the mine backfilling pipeline during operation and construct a dataset based on data from different operating states; S102. Build an artificial intelligence model based on deep learning and train it using the constructed dataset; S103. Construct an optimization algorithm based on HSO to optimize the model parameters; S104. Based on the optimized model, test and verify the instance data, and provide feedback on the results to ensure timely monitoring and guarantee mine production safety.

[0064] It should be understood that multiple fiber optic sensors can be installed on the mine filling pipeline. These sensors can collect multi-dimensional data such as vibration, temperature, and pressure with time-series information, such as multi-dimensional vibration data at historical moments and multi-dimensional data during the current operation. The specific acquisition process can be set according to the actual situation, and this embodiment of the invention does not impose any limitations on it.

[0065] In specific embodiments of the present invention, various types of fiber optic sensors can be deployed along the mine backfilling pipeline system, including but not limited to distributed fiber acoustic vibration sensors (DAS), fiber optic grating (FBG) temperature sensors, and pressure sensors. These sensors simultaneously collect time-series data of multidimensional physical quantities such as vibration acceleration, temperature field distribution, and internal pressure during pipeline operation at a frequency of tens to hundreds of sampling points per second. Specific sensor models, installation spacing, sampling frequencies, and other parameters can be adaptively adjusted according to the pipeline's diameter, material, laying slope, and slurry transport characteristics; this embodiment of the present invention does not impose rigid limitations on these aspects.

[0066] In one specific embodiment of the present invention, step S101 includes: Fiber optic sensors deployed on the filling pipeline system collect multi-dimensional time-series data in real time, including vibration, temperature, and pressure. Based on historical operation and maintenance records and expert knowledge, a signal coupling effect model is constructed, and an intelligent model is built to solve for the operating status signals.

[0067] Its coupling effect model is set as follows: ; in: For the observed vibration signal, in the time domain; The original vibration signal or the target signal; It is a temperature signal; For flow rate signals; , The nonlinear coupling coefficient between temperature and flow rate; The linear coupling coefficient between temperature and flow rate; Other interfering signals, such as noise, unmodeled dynamics, etc.

[0068] The intelligent model is an end-to-end deep learning model. Its detailed components include: The multimodal feature extraction layer is used to extract local features of data in various dimensions; the cross-modal fusion module is used to fuse multimodal data features; and the signal separation module is used to extract the original vibration signal.

[0069] The specific process can be as follows: Step 1: Define the input and output, including: Input tensor: vibration signal (Batch number B, time step T), temperature signal: Flow rate signal: ; Output: Estimated original signal: Interference signal estimation: ; Step 2: Extract local features using a 1D convolutional neural network: ; in: ; k is the kernel size; Output feature map: ; Temperature and flow rate signals are encoded separately: Temperature characteristics:

[0070] The LSTM unit calculates:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] The flow velocity characteristics use the same structure:

[0077] Step 3: Use multi-head attention to fuse features from different modalities: Query-key-value projection:

[0078] Attention weights:

[0079] Bullish Attention:

[0080]

[0081] Features after fusion:

[0082] Estimating interference using gating mechanisms:

[0083]

[0084]

[0085] Remove interference from fused features:

[0086]

[0087]

[0088] Where ResBlock is the residual block:

[0089] Step 4: Restore the original temporal resolution using deconvolution or interpolation:

[0090] Final output layer: ; Use tanh to ensure the output is in the range [-1, 1].

[0091] In some embodiments of the present invention, the processed operating status data is defined as follows: normal state, slight wear state, critical blockage state, severe blockage state, and leakage state.

[0092] The collected raw data is cleaned, denoised, aligned and normalized to form a regular time series dataset, which is then divided into training set, validation set and test set.

[0093] In some embodiments of the present invention, step S102 includes: An improved and integrated model of Transformer and LSTM was obtained to obtain a combined model; The combined model is trained to obtain the initial weight parameters.

[0094] In a specific embodiment of the present invention, the input module can be improved to obtain multi-scale data sequences. The specific scale division can be adaptively set according to the actual situation, and the embodiments of the present invention do not impose any limitations on this. Then, the encoder module can be improved to enable it to perform diagnostic classification on multi-scale data, thereby obtaining multiple initial prediction results from multiple sequences.

[0095] In some embodiments of the present invention, multi-scale data sequences with at least three dimensions (x, y, z) undergo noise reduction and scale partitioning to obtain multiple sub-data sequences, including: Based on the collected data, noise reduction measures are adopted to preprocess the multi-scale data series in the x, y, and z dimensions to eliminate noise interference and improve data quality.

[0096] Based on scale partitioning, and according to partitioning criteria such as frequency characteristics, the data sequence in each directional dimension after noise reduction is divided into multiple sub-data columns of different scales in order to capture the characteristics and change patterns of the data at different scales.

[0097] In a specific embodiment of the present invention, the data column is divided into different subsequences based on a scale partitioning operation, and these subsequences are used as input to the Transformer. The specific process can be as follows: Step 1: Vibration signal sequence acquired by the sensor For each vibration signal, its length is L. First, it is divided into non-overlapping subsequences (patch). The length of each subsequence (patch) is denoted as L. P The span of the non-overlapping region between two consecutive subsequences is denoted as . S Then the segmentation process will generate a subsequence. ,in N It is the number of subsequences, calculated using the following formula:

[0098] In the formula: L P represents the total length from the starting position of the first patch to the starting position of the last possible patch.

[0099] Step 2: Before segmentation, we pad the end of the original signal. S The last value of the repeating sequence Through this segmentation method, the number of input tokens can be increased from... L Reduce to approximately The patch is mapped to the Transformer latent space via a trainable linear projection, as shown in the following formula:

[0100] Where B is the batch size. Let be the number of blocks for the i-th type of patch.

[0101] After the above steps, we define the set of M patch size values ​​as S = {S1, ..., S2}. M Each patch size S corresponds to one patch segmentation operation. For an input vibration signal X, each patch segmentation operation with a patch size of S divides X into P patches, namely (X1, X2, ..., X...). PDifferent patch sizes in the set resulted in segmented patches of different scales.

[0102] In some embodiments of the present invention, it is necessary to process the subsequences after multi-scale partitioning more efficiently, and to classify and discriminate data with similar frequency features, including: Combining internal attention mechanisms with router mechanisms to better adapt to the field of mine backfill monitoring, the goal of granular self-attention mechanisms is to capture the local correlation between features and timestamps within the same granularity. For each granularity... i Given the block embedding of the input and the corresponding router embedded First, they are concatenated into an intermediate sequence. :

[0103] Here, ‖ represents the concatenation operation. Next, regarding... Apply self-attention mechanism and update block embeddings simultaneously. and router embedded :

[0104]

[0105] in, This represents an inner-granularity self-attention operation, specifically a scaled dot product self-attention operation.

[0106] In multi-granularity self-attention mechanisms, directly performing large-scale self-attention operations on block embeddings of all granularities leads to a sharp increase in computational complexity. Router mechanisms address this by introducing router embeddings. This approach aggregates and interacts features at different granularities, thus avoiding large-scale self-attention operations on all blocks directly. Specifically, router embedding... It is updated in the inner granularity self-attention stage to capture features within each granularity. For each granularity i Router Embedded The initialization is as follows:

[0107] in: It is a position embedding matrix with dimensions G×D, where G is a sufficiently large number and D is the embedding dimension. It is the first i Learnable granular embeddings at each granularity, with a dimension of 1× D . It is the first i Number of blocks at each granularity.

[0108] In a specific embodiment of the present invention, the data processed by the improved Transformer model can be fed into an improved LSTM for further classification and discrimination. For the input layer, the input sequence is received. X =( x 1, x 2,…, x T ),in x t It is a time step t The input vector, T It is the length of the sequence. For a forward LSTM layer, it is at time step... t The calculation can be expressed as:

[0109] in, It is a time step t The positive hidden state, It is a time step t The positive cell state, It is a time step t The input vector.

[0110] The internal computations of LSTM include:

[0111]

[0112]

[0113]

[0114]

[0115]

[0116] in, σ is the sigmoid function, tanh is the hyperbolic tangent function, and ⊙ is element-wise multiplication, which is the Hadamard product (element-by-element multiplication). , , , , , , , It is a weight matrix. , , , It is the bias vector.

[0117] For the reverse LSTM layer, the internal computation of the LSTM is similar to that of the forward LSTM layer, except that the order of the input sequences is reversed:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] For the output layer, at time step t The calculation can be expressed as: ; : indicates a concatenation operation, which connects two vectors end to end to form a new vector.

[0124] Final output From the merged hidden state Obtained through a linear layer: ; in, It is the output weight matrix. It is the output bias vector.

[0125] Therefore, the output results can be obtained through the improved LSTM network, thereby determining the operating status of the traditional pipeline.

[0126] In some embodiments of the present invention, such as Figure 2 As shown, some parameters are optimized, including: The process of automatically tuning the hyperparameters of a deep learning model based on the HSO algorithm is as follows.

[0127] 1. Initialization: Like other swarm optimization algorithms, random initialization is used.

[0128]

[0129]

[0130]

[0131]

[0132] Coefficient Calculation: Calculate the root mean square (RMS) of all fitness values ​​using the following formula:

[0133] Then, the coefficients of the displacement vector are calculated based on the difference between each fitness value and the root mean square; the individual differences are:

[0134] The difference between coefficients c and their root mean square values ​​is positive, and the difference between coefficients c and their root mean square values ​​is negative. The sum of the normalized absolute values ​​is 1.

[0135] 3. Location Update: The location of each individual searched is iteratively updated. Each individual's location is affected by the weighted sum of the differences between its current location and the locations of all other individuals. The update rule is as follows:

[0136] 4. Selection based on adaptive simulated annealing: The selection process based on simulated annealing (SA) includes dynamic temperature updates for each iteration. The updates are as follows:

[0137] If the new fitness value is better than the previous one, the algorithm operates greedily, replacing the old solution with the new one. The following criteria apply to selection:

[0138]

[0139] if

[0140] 5. Adaptive Mutation: To further improve exploration capabilities, an adaptive mutation operator is defined, which dynamically adjusts the mutation rate and mutation step size based on the current iteration. iter

[0141] iter

[0142] This configuration allows HSO to explore the search space more extensively in the initial iterations, while gradually focusing on utilizing the current solution as it approaches the maximum number of iterations.

[0143] If a mutation occurs, a random perturbation is added to the individual's location:

[0144]

[0145] By following these steps, the HSO algorithm effectively balances global and local search strategies to select the optimal parameters.

[0146] In some embodiments of the present invention, such as Figure 3 As shown, the operation of the mine backfilling pipeline is monitored based on data from different operating states, including: S301. Set the classification level Z for the pipeline transportation status of mine backfilling pipelines; S302. Determine the target level based on the classification of pipeline transportation status and the model judgment result Q; S303. Monitor the operation of mine backfill pipelines according to the target level.

[0147] In a specific embodiment of the present invention, the pipeline transportation status classification Z of the mine backfilling pipeline can be set. Z includes three levels: Level 1 represents a severe abnormal state, Level 2 represents a blockage state, and Level 3 represents a normal state. Based on the value determined by the model, we set it as Q. The value of Q ranges from 1 to 9. When Q is between 1 and 3, its Z is Level 3 normal state; when Q is between 4 and 6, its Z is Level 2 blockage state; and when Q is between 7 and 9, its Z is Level 1 severe abnormal state.

[0148] This invention utilizes fiber optic sensors installed on filling pipelines to acquire time-series temperature and vibration data, thereby obtaining data on the mine filling pipeline during operation and constructing a dataset based on data from different operating states. It also builds a deep learning-based artificial intelligence model, trains it using the constructed dataset, and constructs an HSO-based optimization algorithm to optimize the model parameters. Based on the optimized model, the example data was tested and verified, and the results were fed back. With the help of the monitoring results, the quantitative characterization of the pipeline operation status was realized, providing a new approach for mine pipeline monitoring.

[0149] To better implement the mine backfill pipeline blockage monitoring method in this embodiment of the invention, correspondingly, this embodiment of the invention also provides a mine backfill pipeline blockage monitoring device, such as... Figure 4 As shown, the mine backfill pipeline blockage monitoring device 400 includes: Data acquisition module 401 is used to acquire multi-dimensional data of the mine backfilling pipeline at the current moment; The data processing module 402 is used to process multi-dimensional data to obtain the noise-reduced result; The result output module 403 is used by the trained model to make predictions and obtain the target discrimination results. The pipeline monitoring module 404 is used to monitor the blockage of mine filling pipelines based on the target discrimination results.

[0150] The mine filling pipeline blockage monitoring device 400 provided in the above embodiments can realize the technical solution described in the above embodiments of the mine filling pipeline blockage monitoring method. The specific implementation principle of each module or unit can be found in the corresponding content in the above embodiments of the mine filling pipeline blockage monitoring method, which will not be repeated here.

[0151] Example 2 Based on Example 1, such as Figure 5 As shown, the present invention also provides an electronic device 500. The electronic device 500 includes a processor 501, a memory 502, and a display 503. Figure 5 Only some components of the electronic device 500 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0152] In some embodiments, memory 502 may be an internal storage unit of electronic device 500, such as a hard disk or memory of electronic device 500. In other embodiments, memory 502 may also be an external storage device of electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 500.

[0153] Furthermore, the memory 502 may include both internal storage units of the electronic device 500 and external storage devices. The memory 502 is used to store application software and various types of data installed on the electronic device 500.

[0154] In some embodiments, processor 501 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in memory 502 or process data, such as the mine filling pipeline blockage monitoring method of the present invention.

[0155] In some embodiments, display 503 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 503 is used to display information from electronic device 500 and to display a visual user interface. Components 501-503 of electronic device 500 communicate with each other via a system bus.

[0156] In some embodiments of the present invention, when the processor 501 executes the mine filling pipeline blockage monitoring program in the memory 502, the following steps can be implemented: Obtain multi-dimensional data of the mine backfilling pipeline at the current moment; The multi-dimensional data is processed to obtain the noise-reduced result; Based on the trained model, predictions are made to obtain the target discrimination results; The blockage status of mine filling pipelines is monitored based on the target discrimination results.

[0157] It should be understood that when the processor 501 executes the mine filling pipeline blockage monitoring program in the memory 502, in addition to the functions mentioned above, it can also perform other functions, as detailed in the description of the corresponding method embodiments above.

[0158] Furthermore, the embodiments of the present invention do not specifically limit the type of electronic device 500 mentioned. Electronic device 500 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device.

[0159] Exemplary examples of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic devices may also be other portable electronic devices, such as laptops with touch-sensitive surfaces (e.g., touch panels).

[0160] It should also be understood that, in some other embodiments of the present invention, the electronic device 500 may not be a portable electronic device, but a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0161] Accordingly, this application also provides a computer-readable storage medium for storing computer-readable programs or instructions. When the programs or instructions are executed by a processor, they can implement the steps or functions of the mine filling pipeline blockage monitoring method provided in the above-described method embodiments.

[0162] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for monitoring the operational status of mine backfilling pipelines, characterized in that... Includes the following steps: Acquire multi-dimensional time-series data during the operation of mine backfilling pipelines, including vibration signals, temperature signals, and pressure signals; A signal feature coupling effect model is constructed to decouple the multi-dimensional time series data. The decoupled multi-dimensional time series data is then divided into multiple scales to obtain multiple multi-scale subsequence data with different time granularities. A hybrid deep learning model based on an improved Transformer and a bidirectional LSTM is constructed. The hybrid deep learning model is used to extract features and determine the state of the multi-scale subsequence data to obtain the determination result of the pipeline operation state. The hyperparameters of the hybrid deep learning model are iteratively optimized based on the HSO optimization algorithm to obtain the optimized monitoring model. Based on the optimized monitoring model, the operational status of mine backfilling pipelines is monitored and early warnings are issued in real time.

2. The method for monitoring the operational status of mine backfilling pipelines according to claim 1, characterized in that... By deploying a fiber optic sensor array on the filling pipeline system, multi-source time-series data of the pipeline under various operating conditions are collected in real time.

3. The method for monitoring the operational status of mine backfilling pipelines according to claim 1, characterized in that... The multi-dimensional time-series data is decoupled to separate the original vibration signal, interference signal, and coupling characteristics of each physical field, resulting in multiple data classification results with different categories and labels, including: Based on historical operation and maintenance records and expert knowledge, a signal feature coupling effect model is constructed, and an intelligent model is built to solve the signal features of the operating status. Define typical operating states of pipelines, including: normal state, slight wear state, critical blockage state, severe blockage state, leakage state, etc., and label the collected data with state labels; The raw data undergoes preprocessing operations such as cleaning, denoising, alignment, and normalization to form a well-organized time-series dataset. The processed data is divided into training set, validation set and test set.

4. The method for monitoring the operating status of mine backfilling pipelines according to claim 1, characterized in that... The construction of the signal feature coupling effect model, and the decoupling processing of the multi-dimensional time series data, includes: Establish a coupling effect model between vibration signals and temperature and flow velocity signals: An end-to-end deep learning network is constructed to solve the coupling effect model. The deep learning network includes a multimodal feature extraction layer, a cross-modal fusion module, and a signal separation module, and outputs the original vibration signal estimate and the interference signal estimate.

5. The method for monitoring the operational status of mine backfilling pipelines according to claim 1, characterized in that... Solving the coupling effect model using an end-to-end deep learning network includes: A one-dimensional convolutional neural network was used to extract local features from vibration signals, temperature signals, and flow velocity signals to obtain feature representations for each mode. Temperature and flow velocity signals are time-series encoded using a long short-term memory network to obtain temperature and flow velocity characteristics. Multi-head attention mechanism is used to fuse features from various modalities. Attention weights are calculated through query-key-value projection to obtain the fused multi-modal features. The interference signal components are estimated using a gating mechanism, and the interference is removed from the fused features. The original vibration signal is then recovered through residual blocks and deconvolution operations.

6. The method for monitoring the operational status of mine backfilling pipelines according to claim 1, characterized in that... The hybrid deep learning model is a hybrid model of temporal Transformer and LSTM: it uses a self-attention mechanism to dynamically calculate the importance of data at different time points and globally perceives temporal dependencies; it uses LSTM to capture the long-term dependencies of data in the time dimension.

7. The method for monitoring the operational status of mine backfilling pipelines according to claim 1, characterized in that... The Long Short-Term Memory (LSTM) network is spliced ​​into two layers, forward and backward. The forward and backward dependencies of the sequence are learned by the LSTM layers in both directions, and finally the outputs of the two directions are merged to obtain a more comprehensive representation of the sequence features.

8. The method for monitoring the operating status of mine backfilling pipelines according to claim 1, characterized in that... The parameters of the hybrid model are optimized using the HSO algorithm to obtain the optimized model, including: During model training, the constructed HSO optimization algorithm is integrated to iteratively optimize the model parameters until the model achieves optimal performance on the validation set.

9. The method for monitoring the operating status of mine backfilling pipelines according to claim 1, characterized in that... Feature extraction and state determination of multi-scale subsequence data are performed using a hybrid deep learning model, including: Define typical operating conditions of pipelines, including normal condition, slight wear condition, critical blockage condition, severe blockage condition, and leakage condition; The collected historical data is labeled with status tags to form a training dataset with category labels; The training dataset is input into the hybrid deep learning model for supervised learning, and the model outputs a state discrimination result. The range of the target state is determined based on the state discrimination result, thereby determining the target level.

10. A monitoring device for the operational status of a mine backfilling pipeline, comprising: The multimodal data acquisition module is used to acquire multi-dimensional time-series data of the mine backfilling pipeline during operation through an optical fiber sensor array. The multi-dimensional time-series data includes vibration signals, temperature signals, and pressure signals. The signal decoupling processing module is used to construct a signal feature coupling effect model, decouple the multi-dimensional time series data, and separate the original vibration signal, interference signal and coupling characteristics of each physical field. The multi-scale data partitioning module is used to partition the decoupled multi-dimensional time series data at multiple scales to obtain multiple subsequence data with different time granularities. A deep learning model building module is used to build a hybrid deep learning model based on an improved Transformer and a bidirectional LSTM. The improved Transformer includes a multi-scale parallel encoder and a router mechanism. The HSO optimization module is used to iteratively optimize the hyperparameters of the hybrid deep learning model based on the HSO optimization algorithm to obtain the optimized monitoring model. The pipeline monitoring module is used to extract features and determine the status of the multi-scale subsequence data using the optimized monitoring model, and to monitor and warn of the operation status of the mine backfilling pipeline in real time based on the determination results.

Images