An echo state bayesian neural network-based running state evaluation method for a mine drilling rig drilling system

Abstract

This invention discloses a method for evaluating the operational status of a mining drilling rig system based on an echo-state Bayesian neural network. This method addresses the challenges of dynamic feature extraction, model overfitting, and the lack of uncertainty quantification in deterministic assessments of drilling rig status under conditions of strong vibration and high noise in underground environments. It employs an echo-state network with a leakage integral mechanism to suppress noise and extract slowly varying dynamic features from multi-source sensor time-series data, constructing a high-dimensional feature vector through dual-view feature aggregation. A Bayesian neural network with Concrete Dropout is introduced to achieve state probability mapping based on variational inference, and Monte Carlo sampling and prediction entropy are combined to quantify cognitive uncertainty and provide early warning of high-entropy anomalies. This invention effectively filters out high-frequency noise, suppresses overfitting in small samples, and enables proactive early warning of abnormal operating conditions. It achieves high classification accuracy on hard coal seam drilling datasets, and its assessment accuracy and decision reliability are significantly superior to traditional deep learning models.

Description

Technical Field

[0001] This invention relates to the fields of coal mine electromechanical equipment condition monitoring and artificial intelligence technology, and in particular to a method for evaluating the operating status of a mining drilling rig drilling system based on an echo state Bayesian neural network. Background Technology

[0002] Mining drilling rigs are key equipment for preventing rockbursts in coal mines. Their drilling systems operate for extended periods in harsh underground environments characterized by strong vibrations and high noise. The multidimensional data collected by sensors is accompanied by strong random noise interference and exhibits complex nonlinear coupling characteristics, posing significant challenges to the accurate identification of the equipment's operating status. Accurate and reliable assessment of the drilling system's operating status is a core element in ensuring underground operational safety and preventing equipment damage and accidents.

[0003] Existing technologies for assessing the operational status of mining drilling rig systems suffer from the following drawbacks: First, extraction of long-term dynamic features is difficult under high-noise environments: multi-dimensional sensor data collected by the drilling rig is mixed with high-frequency random noise, easily masking the slow-changing dynamic trends of the equipment's actual operating conditions. Second, static neural networks lack memory units, and conventional recurrent neural networks face training difficulties and gradient vanishing problems when processing long-sequence signals with strong noise, failing to effectively reconstruct dynamic features. Third, narrow data distribution in single-condition drilling scenarios easily leads to model overfitting: the sample data distribution is limited in single-condition drilling scenarios. Existing models often use deterministic point estimation to update weights, easily leading to memory-based overfitting for specific noise patterns, resulting in poor model generalization ability and inability to adapt to complex real-world conditions. Fourth, deterministic models lack uncertainty quantification capabilities: existing assessment models are mostly deterministic, only outputting deterministic classification results. When faced with abnormal samples deviating from the training distribution, they cannot identify their applicable boundaries, easily giving high-confidence erroneous assessments, posing serious decision-making risks. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the operating status of a mining drilling rig based on an echo state Bayesian neural network. This method has the functions of noise suppression, temporal feature extraction, probabilistic evaluation, and uncertainty early warning, and solves the problems of weak feature extraction capability, easy overfitting, lack of uncertainty quantification, and low decision reliability of traditional models.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] A method for evaluating the operational status of a mining drilling rig based on an echo-state Bayesian neural network includes:

[0007] Step S1. Data preprocessing: Collect multi-source sensor time-series data during the drilling process of the mining drilling rig, perform sliding window truncation and downsampling preprocessing on the time-series data, and construct a classification dataset for drilling status.

[0008] Step S2. Dynamic Feature Extraction: The preprocessed time series data is input into the echo state network reservoir with leakage integration mechanism to suppress high-frequency random noise and extract the slow-varying dynamic features of the drilling rig operation. A high-dimensional dynamic feature vector is obtained through dual-view feature aggregation.

[0009] Step S3. Probabilistic evaluation model training: The high-dimensional dynamic feature vector is input into the Bayesian neural network of ConcreteDropout, and the nonlinear mapping from the feature space to the probability space of the running state is completed based on variational inference;

[0010] Step S4. Online Inference Early Warning: Perform Monte Carlo sampling forward inference on the input sample, calculate the mean value of the state prediction probability to obtain the evaluation result, and at the same time, use the prediction entropy quantification model to recognize uncertainty and trigger a high entropy anomaly early warning.

[0011] In some embodiments, the specific process of step S1 is as follows:

[0012] Step S11, Data Acquisition and Extraction: Define the system at discrete time points. The received input signal variable is ,in The number of multi-dimensional sensor channels is determined; sensor time series data of simulated drilling in hard coal seam drilling rig are selected, and invalid drilling data at the start and end stages are removed. The original time series samples are extracted using a sliding window, and the length of the original time window is set to the preset number of sampling points.

[0013] Step S12, Downsampling: While preserving key frequency domain features, downsampling is performed on the original samples to reduce the dimensionality of the model input.

[0014] Step S13, Dataset Construction and Partitioning: After downsampling, obtain valid samples, and divide the valid samples into training set, validation set and test set according to a preset ratio to obtain standardized time series samples.

[0015] In some embodiments, the specific process in step S2 is as follows:

[0016] Step S21. Reservoir Network Construction and Initialization: Construct a reservoir network containing a large number of sparsely connected neurons, with the number of internal neuron nodes being... Control the weight matrix of the reserve pool The spectral radius satisfies the following constraints:

[0017]

[0018] in, The number of neurons within the reservoir determines The dimension; For the reserve pool weight matrix The One eigenvalue; For the reserve pool weight matrix The spectral radius, in its physical sense, is the effective memory length of the system for historical information.

[0019] During the initialization phase, a random sparse matrix following a standard normal distribution is generated. And then perform spectral radius scaling on it, as shown in the following formula:

[0020]

[0021] in, This represents the maximum value of the modulus of the eigenvalues ​​of the matrix. This is a preset spectral radius hyperparameter;

[0022] Step S22. Leakage Integral State Update: Input the acquired time-series signal vector into the storage pool, perform state evolution in the discrete-time domain, and define... The input signal represents the dynamic intermediate state quantity generated by the evolution of the signal in the reservoir. The transitive relationship over time follows:

[0023]

[0024] in, For a moment Preprocessed multi-sensor input signal vector; This represents the state vector within the reservoir, with the dimension corresponding to the size of the neurons in the reservoir. This is the connection weight matrix from the input layer to the reservoir; This is the recursive connection weight matrix within the reserve pool; It is the hyperbolic tangent activation function; Defined as leakage rate; For the reserve pool at the previous moment ; internal state vector;

[0025] Step S23. Dual-view feature aggregation: After processing by the leakage integral reservoir, the original time-series signal is transformed into a high-dimensional state trajectory sequence that evolves over time. Extract the final state vectors of the reserve pools respectively. With time-averaged state vector The two are then concatenated in a high-dimensional manner to form the final dynamic feature vector. The formula is as follows:

[0026]

[0027] in, This represents a vector concatenation operation; the concatenated feature vectors It contains complete system evolution information; This represents the total number of discrete time steps for the timing signal within the current sampling window.

[0028] In some embodiments, the specific process of step S3 is as follows:

[0029] Step S31. BNN Network Architecture and Reparameterization: Construct a fully connected BNN structure containing an input layer, two Bayesian hidden layers, and an output layer; introduce the Concrete Dropout mechanism into the network layers and use reparameterization techniques to continuously relax the binary mask generation process; define the log-odds ratio. The model utilizes a continuous relaxation form of the Gumbel-Softmax distribution to generate intermediate values ​​of the neuron-preserving mask. Its generation relationship combines structural parameters and random noise, as shown in the following formula:

[0030]

[0031] in, Use the Sigmoid activation function; This is a temperature coefficient used to control the shape of the approximate distribution and ensure effective gradient propagation. For random noise that follows a standard uniform distribution;

[0032] Step S32. Loss Function Construction and ELBO Optimization: The training objective of the model is to search for the optimal variational posterior distribution of the weights. Approximates the true posterior distribution ;Transfer cost function Cost of the cross-entropy of the fitted input signal Together with the regularization cost that penalizes network complexity, its optimization transitivity is as follows:

[0033]

[0034] in: Cross-entropy loss represents the likelihood cost and encourages the sampling parameters in the variational distribution to accurately fit the current observation data. For the first Layer weights Norm; This is the weight decay coefficient; For network layer index; To match the mask rate The relevant structure regularization term, representing the complexity cost, is used to drive the model to find the optimal sparsity; for the th Layer The formula for calculating the number of neurons is:

[0035]

[0036] in, For the first The number of neurons in a layer; The mask rate learned by the current model; The initial mask rate;

[0037] Through the above regularization penalty mechanism, the network is forced to maintain the sparsity of parameters as much as possible while interpreting the current observation data, thereby fundamentally suppressing the risk of model overfitting caused by sudden high-frequency noise in small samples.

[0038] Step S33. End-to-end model training: The network training uses the Adam optimizer for variational parameter training. and weight Iterative updates;

[0039] The training set is input into the network in batches, and then passed through the ESN state mapping and dual-view feature aggregation before entering the BNN for forward propagation. The parameters are iteratively updated through the backpropagation algorithm. In each round, the validation set is used to monitor the generalization performance and save the best model weights until the training process converges.

[0040] In some embodiments, the specific process of step S4 is as follows:

[0041] Step S41. Monte Carlo Sampling and Parallel Inference: For new test samples Preserve the Concrete Dropout random mask mechanism in the network and execute it. Sub-independent Monte Carlo sampling forward propagation, employing a Bayesian model averaging strategy, calculates... The average of the predicted probabilities is used as the final classification prediction result, and its calculation formula is as follows:

[0042]

[0043] in, This represents the predicted probability distribution of the model's final output. For the training dataset; For new test samples; The predicted category corresponding to the model; For the first Network variational weight configuration during the next sampling; This represents the optimal variational posterior distribution of the network weights. To configure specific variational weights Below, the model prediction probability corresponding to a single sampling; Index of the number of Monte Carlo samples;

[0044] Step S42. Prediction Entropy Calculation and High Entropy Early Warning Mechanism: While outputting the state evaluation classification result, the model calculates the prediction entropy. To quantify the cognitive uncertainty regarding the current input sample, the calculation formula is as follows:

[0045]

[0046] in: The total number of categories; For the model to the first The average predicted probability of the class;

[0047] The system pre-sets a safety threshold for predicted entropy: during real-time monitoring, if the calculated predicted entropy... The system will proactively determine if the current operating condition is abnormal and immediately trigger a manual intervention alarm.

[0048] In some embodiments, the method further includes a model validation and evaluation step:

[0049] Model testing and evaluation: Input the trained evaluation model into the test set samples, and calculate the model's classification accuracy, recall, and the harmonic mean score of precision and recall;

[0050] Ablation experiments were conducted to verify the results: control models with the ESN module removed and control models with BNN replaced by MLP were constructed and tested on the same dataset, and the performance metrics were compared.

[0051] Benchmark Model Comparison: This model is compared with benchmark models such as LSTM, GRU, 1D-CNN, and BiLSTM on the same test set to verify the effectiveness of the model.

[0052] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a leakage integral mechanism into the echo state network, this invention can effectively suppress strong random noise downhole, accurately extract slowly varying dynamic features, and avoid the problems of gradient vanishing and training difficulties in traditional recurrent neural networks. Employing Bayesian variational inference and a Concrete Dropout adaptive sparsity structure, it can suppress model overfitting in single-condition, small-sample scenarios, significantly improving generalization ability. Simultaneously, this invention achieves probabilistic state evaluation and prediction entropy uncertainty quantification, accurately identifying abnormal samples and proactively triggering high-entropy warnings, fundamentally solving the defects of deterministic models that are blindly confident and prone to high-confidence misjudgments. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0054] Figure 1 An overall architecture diagram of the echo state Bayesian neural network runtime state evaluation model is provided for the embodiments of this application.

[0055] Figure 2 This application provides an end-to-end training flowchart for the model in the Del embodiment.

[0056] Figure 3 A schematic diagram of test set evaluation results is provided for embodiments of this application. Detailed Implementation

[0057] 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. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] This application proposes a method for evaluating the operating status of a mining drilling rig based on an echo-state Bayesian neural network, which mainly includes the following modules:

[0059] (1) Construction of the overall architecture of the ESN-BNN hybrid evaluation model;

[0060] To address the highly nonlinear and complex temporal dynamic characteristics exhibited by drilling rigs as large electromechanical-hydraulic coupled systems, and the stringent requirements for reliability and uncertainty quantification in downhole safety monitoring scenarios, this invention constructs a hybrid evaluation model based on echo state networks and Bayesian neural networks. Multi-source sensor time-series data are used as input signal variables. First, the input is fed into the echo state network at the front end, where the low-dimensional observation signal is projected into a dynamic intermediate feature vector through a dynamic reservoir. Subsequently, the aggregated high-dimensional feature vectors are transmitted to the backend Bayesian neural network, where the mapping from the feature space to the class probability space is completed within the variational inference framework. Finally, the system generates and outputs a classification probability containing the running state. With uncertainty measurement The output of the evaluation result of (predicted entropy).

[0061] (2) A time-series feature extraction module based on a leakage integral reservoir;

[0062] To address the issues of mechanical inertia and time lag during drilling, this module defines the temporal propagation relationship of signals in the front-end network. Specifically, it involves the acquisition of multi-sensor input signal variables. The input is fed into the echo state network reservoir, which serves as a dynamic feature extractor; during state transfer, a leakage integration mechanism is introduced, i.e., the input signal... Controlled by the quantitative leakage rate parameter, it is superimposed and transmitted with the internal hysteresis state of the reservoir at the previous moment, thereby generating an intermediate quantity of the reservoir's internal state equivalent to the characteristics of a first-order low-pass filter. This transmission mechanism effectively smooths out high-frequency random noise in the input signal while preserving the slow-varying dynamic trend. To avoid information loss, this module further employs a dual-view feature aggregation strategy, concatenating the final-time state vector from the storage pool with the time-averaged state vector in a high-dimensional manner to form a core intermediate feature vector that comprehensively characterizes the current operating state of the drilling rig. And pass it to the downstream modules.

[0063] (3) Probabilistic state assessment and uncertainty quantification module based on variational inference;

[0064] To address the issue of deterministic models exhibiting overconfidence when faced with outlier samples, this module receives intermediate feature vectors from the front end. A Bayesian neural network is used as a probabilistic decision maker to output the posterior result. In the downward propagation mapping path of the signal, a Concrete Dropout mechanism is introduced, utilizing a continuous relaxation reparameterization technique to generate intermediate quantities of the neuron-preserving mask. This enables the network to adaptively eliminate interfered sensor paths based on the signal-to-noise ratio differences in the transmission channels. During the model optimization and propagation phase, a total loss objective is constructed, consisting of the fitting likelihood cost and the network parameter structure regularization cost. This drives the weight distribution to converge stably. In the final inference phase, the model performs multiple independent Monte Carlo samplings and parallel inferences on the new input samples, ultimately generating two core output quantities: one is the running state probability output quantity, which serves as the final classification result. Secondly, it involves quantifying the early warning output (predictive entropy) of cognitive uncertainty. When the warning output volume When the set safety threshold is exceeded, the system will proactively trigger a manual intervention alarm, completely compensating for the decision-making deficiencies of the deterministic model.

[0065] The specific implementation process of this method includes four stages: data preprocessing, dynamic feature extraction, probabilistic evaluation model training, and online inference and early warning. Details are as follows:

[0066] Step S1. Data preprocessing: Collect multi-source sensor time-series data during the drilling process of the mining drilling rig, perform sliding window truncation and downsampling preprocessing on the time-series data, and construct a classification dataset for drilling status.

[0067] Step S1 includes the following steps:

[0068] Step S11, Data Acquisition and Extraction: Define the system at discrete time points. The received input signal variable is ,in To determine the number of multi-dimensional sensor channels (i.e., the dimension of the input signal), time-series sensor data simulating drilling in a hard coal seam drilling rig is selected. After removing invalid drilling data from the start and end stages, a sliding window is used to extract the original time-series samples, setting the original time window length to a preset number of sampling points. For example, the original time window length is set to 20,000 sampling points, corresponding to a physical duration of 2 seconds. Downsampling refers to removing one sample point every few seconds, reducing the sampling rate. For example, if the original sample length is one second and corresponds to 10,000 points, downsampling will result in only 8,000 or 5,000 points, reducing data complexity and the difficulty of model training.

[0069] Step S12, Downsampling Processing: While preserving key frequency domain features, downsampling processing is performed on the original samples to reduce the dimensionality of the model input. Specifically, given that the high sampling rate of 10kHz for the original signal would lead to significant data redundancy, and that excessively long sequences would significantly increase the memory overhead and time cost of subsequent reservoir computation, downsampling processing is performed on the original samples while preserving key frequency domain features to effectively reduce the dimensionality of the model input.

[0070] Step S13, Dataset Construction and Partitioning: After downsampling, obtain valid samples, and divide the valid samples into training set, validation set and test set according to a preset ratio to obtain standardized time series samples.

[0071] For example, after downsampling, a total of 461 valid samples were obtained. The sample performance was divided into three categories: excellent, average, and poor. Specific statistical information is shown in Table 1.

[0072] Table 1. Statistical Table of Sample Data for Single-Condition Drilling Dataset

[0073]

[0074] To evaluate the model's generalization ability, the dataset was divided into training, validation, and test sets in a ratio of 0.7:0.15:0.15. The training set contains approximately 322 samples and is used for model parameter optimization; the validation set contains 69 samples and is used to monitor generalization performance during training and preserve the optimal model weights; the test set contains approximately 70 samples and is used for model performance evaluation and uncertainty quantification analysis.

[0075] Step S2. Dynamic Feature Extraction: The preprocessed time series data is input into the echo state network reservoir with leakage integration mechanism to suppress high-frequency random noise and extract the slow-varying dynamic features of the drilling rig operation. A high-dimensional dynamic feature vector is obtained through dual-view feature aggregation.

[0076] The specific process in step S2 is as follows:

[0077] Step S21. Reservoir Network Construction and Initialization: Construct a reservoir network containing a large number of sparsely connected neurons, with the number of internal neuron nodes being... (For example, the number of internal neuron nodes is set to 200, i.e.) ); Control reserve pool weight matrix The spectral radius satisfies the following constraints:

[0078]

[0079] in, The number of neurons within the reservoir determines The dimension; For the reserve pool weight matrix The One eigenvalue; For the reserve pool weight matrix The spectral radius, in its physical sense, represents the effective length of the system's memory of historical information.

[0080] It refers to the number of network nodes. and matrix It is related to the format. The dimension of this matrix is The number of nodes determines the matrix dimension, which is the complexity of the features extracted by the network.

[0081] During the initialization phase, a random sparse matrix following a standard normal distribution is generated. And then perform spectral radius scaling on it, as shown in the following formula:

[0082]

[0083] in, This represents the maximum value of the modulus of the eigenvalues ​​of the matrix. The preset spectral radius hyperparameter can be set to 0.95 in this model to ensure that the network has long-term memory capability.

[0084] Step S22. Leakage Integral State Update: Input the acquired time-series signal vector into the storage pool and perform state evolution in the discrete-time domain. Definition This refers to the dynamic intermediate state quantities generated by the signal evolution in the reservoir. Input signal The transitive relationship (state update equation) over time follows:

[0085]

[0086] in, For a moment Preprocessed multi-sensor input signal vector; This represents the state vector within the reservoir, with the dimension corresponding to the size of the neurons in the reservoir. This is the connection weight matrix from the input layer to the reservoir; This is the recursive connection weight matrix within the reserve pool; It is the hyperbolic tangent activation function; Defined as leakage rate, in the methods provided by the embodiments of this application It can be set to 0.5. For the reserve pool at the previous moment . ...

[0087] By introducing leakage rate By assigning tunable time constants to the neurons in the reservoir, the network acquires physical characteristics equivalent to a first-order low-pass filter, thereby effectively smoothing out the input signal. While eliminating mid-to-high frequency random noise, it retains the slow-changing dynamic trend that reflects the essence of the operating condition.

[0088] Step S23. Dual-view feature aggregation: After processing by the leakage integral reservoir, the original time-series signal is transformed into a high-dimensional state trajectory sequence that evolves over time. To comprehensively characterize the drilling rig's operating status within the current sampling window and avoid information loss caused by a single aggregation method, a transient and steady-state dual-view feature aggregation strategy is adopted.

[0089] Extract the final state vector of the reserve pool respectively With time-averaged state vector The two are then concatenated in a high-dimensional manner to form the final dynamic feature vector. The formula is as follows:

[0090]

[0091] in, This represents a vector concatenation operation; the concatenated feature vectors It contains complete system evolution information, which serves as the feature input for subsequent probabilistic state evaluation; This represents the total number of discrete time steps for the timing signal within the current sampling window.

[0092] Please see Figure 1 , Figure 1 This application provides an overall architecture diagram of an echo-state Bayesian neural network runtime state evaluation model for embodiments of the present application. Figure 1 This paper illustrates the overall macroscopic architecture of the runtime state evaluation model based on Echo State Bayesian Neural Network (ESN-BNN) proposed in embodiments of this application. For example... Figure 1 As shown, the architecture is a serial structure, mainly composed of a front-end time-varying dynamic feature extraction module and a back-end probabilistic state evaluation module. Low-dimensional time-series observation data collected by downhole multi-source sensors are first input into the Echo State Network (ESN) reservoir with a leakage integration mechanism to filter out high-frequency noise and map it into a high-dimensional dynamic feature vector. Subsequently, this feature vector is input into the back-end Bayesian Neural Network (BNN), which completes the nonlinear mapping from the feature space to the category probability space under the variational inference framework, and finally outputs a comprehensive evaluation result containing the probabilities of each operating state and the prediction entropy (a measure of uncertainty).

[0093] The high-dimensional dynamic feature vector extracted in step S2 is input into the backend Bayesian Neural Network (BNN) to construct a probabilistic decision-maker. The specific probabilistic evaluation model training process in step S3 is as follows:

[0094] The high-dimensional dynamic feature vector is input into the Bayesian neural network of Concrete Dropout, and a nonlinear mapping from the feature space to the probability space of the running state is completed based on variational inference.

[0095] The specific process of step S3 is as follows:

[0096] Step S31. BNN Network Architecture and Reparameterization: Construct a fully connected BNN structure containing an input layer, two Bayesian hidden layers, and an output layer (e.g., the network structure can be set to 400-128-64-3). Introduce the ConcreteDropout mechanism into the network layers. To address the non-differentiability problem of discrete Bernoulli mask sampling in the computational graph, the reparameterization trick is used to continuously relax the binary mask generation process.

[0097] Define logarithmic odds During signal propagation, the model utilizes a continuous relaxation form of the Gumbel-Softmax distribution to generate intermediate quantities of the neuron-preserving mask. Its generation relationship combines structural parameters and random noise:

[0098]

[0099] in, Use the Sigmoid activation function; The temperature coefficient (which can be set to 0.1 in this model) is used to control the shape of the approximate distribution and ensure effective gradient propagation; the random noise ϵ follows a uniform distribution in the interval (0,1).

[0100] Step S32. Loss Function Construction and ELBO Optimization: The training objective of the model is to search for the optimal variational posterior distribution of the weights. Approximates the true posterior distribution The cost function of transmission Cost of the cross-entropy of the fitted input signal Together with the regularization cost that penalizes network complexity, its optimization transitivity is as follows:

[0101]

[0102] in: Cross-entropy loss represents the likelihood cost and encourages the sampling parameters in the variational distribution to accurately fit the current observation data. For the first Layer weights Norm; The weight decay coefficient (for example, it can be set to 1×10 in the model provided in the embodiments of this application) -4 ); ; For network layer index; To match the mask rate The relevant structure regularization term, representing the complexity cost, is used to drive the model to find the optimal sparsity; for the th Layer The formula for calculating the number of neurons is:

[0103]

[0104] in, For the first The number of neurons in a layer; The mask rate learned by the current model; This is the initially set mask rate.

[0105] Through the above regularization penalty mechanism, the network is forced to maintain the sparsity of parameters as much as possible while interpreting the current observation data, thereby fundamentally suppressing the risk of model overfitting caused by sudden high-frequency noise in small samples.

[0106] Please see Figure 2 , Figure 2 This application provides an end-to-end training flowchart for the model in the Del embodiment. Figure 2This document illustrates the end-to-end training execution flow of the ESN-BNN hybrid evaluation model according to an embodiment of this application. Figure 2 As shown, after loading the time-series data batches, the data first undergoes state mapping and dual-view feature aggregation through the front-end ESN. During this process, the weights within the reserve pool are fixed and do not participate in gradient updates. Subsequently, the aggregated high-dimensional features enter the BNN learning core. The network uses the ConcreteDropout mechanism for mask reparameterization sampling to achieve adaptive sparsity of the topology. After the model outputs the prediction results, the system calculates the total ELBO loss function, which includes cross-entropy loss (likelihood cost) and variational regularization loss (complexity cost). Finally, backpropagation is performed through the Adam optimizer to iteratively update the variational parameters until the optimal model is saved.

[0107] Step S33. End-to-end model training: The network training uses the Adam optimizer for variational parameter training. and weight The iterative updates, specifically the training hyperparameter configurations are as follows: Initial learning rate: Batch size: 32. Number of training epochs: 100.

[0108] The training set is batch-input into the network, sequentially passing through the ESN state mapping and dual-view feature aggregation before entering the BNN for forward propagation. Parameters are iteratively updated using the backpropagation algorithm. In each round, the validation set is used to monitor generalization performance and save the optimal model weights until the training process converges.

[0109] Please see Figure 3 , Figure 3 A schematic diagram of test set evaluation results is provided for embodiments of this application. Figure 3 This paper visually demonstrates the final evaluation performance and uncertainty quantification capabilities of the model in the embodiments of this application on a hard coal seam drilling test set. It mainly includes three key subgraphs: a state classification confusion matrix, a prediction entropy distribution, and a Monte Carlo sampling distribution of misjudged samples. The confusion matrix shows that the model can accurately classify most test samples into the true operating state category, proving its high accuracy in feature recognition under strong noise. The prediction entropy distribution reveals the significant difference in cognitive certainty between correctly predicted samples and misjudged samples, confirming the scientific validity of proactively triggering unknown risk warnings by setting a prediction entropy safety threshold (e.g., 0.8). The Monte Carlo sampling distribution subgraph further analyzes misjudged abnormal samples, showing the large fluctuations and high dispersion of their single prediction probability across multiple samples. This combined set of graphs, from macroscopic accuracy to microscopic probability distribution, fully demonstrates that the Bayesian network in the embodiments of this application can proactively issue warnings lacking confidence evidence through a broad probability distribution, completely breaking the limitations of blind confidence in traditional deterministic models.

[0110] Step S4. Online Inference Early Warning: Perform Monte Carlo sampling forward inference on the input sample, calculate the mean value of the state prediction probability to obtain the evaluation result, and at the same time, use the prediction entropy quantification model to recognize uncertainty and trigger a high entropy anomaly early warning.

[0111] The specific process of step S4 is as follows:

[0112] Step S41. Monte Carlo Sampling and Parallel Inference: For new test samples Preserve the Concrete Dropout random mask mechanism in the network and execute it. Sub-independent Monte Carlo sampling forward propagation, this process is equivalent to using Parallel inference is performed on sub-models with different weight configurations.

[0113] Using a Bayesian model averaging strategy, calculate The average of the predicted probabilities is used as the final classification prediction result, and its calculation formula is as follows: ;

[0114] in, This represents the predicted probability distribution of the model's final output. For the training dataset; For new test samples; The predicted category corresponding to the model; For the first Network variational weight configuration during the next sampling; This represents the optimal variational posterior distribution of the network weights. To configure specific variational weights Below, the model prediction probability corresponding to a single sampling; Index of the number of Monte Carlo samples.

[0115] This strategy smooths out the random prediction bias of different sub-models, thereby obtaining more robust evaluation results than a single deterministic network.

[0116] Step S42. Prediction Entropy Calculation and High Entropy Early Warning Mechanism: While outputting the state evaluation classification result, the model calculates the prediction entropy. To quantify the cognitive uncertainty regarding the current input sample, the calculation formula is as follows:

[0117]

[0118] in: The total number of categories to be classified (in the problem of this embodiment) (corresponding to three operating states: excellent, medium, and poor). For the model to the first The average predicted probability of the class;

[0119] Predicting entropy The value directly reflects the degree of disagreement in the model's decision-making: The closer the value is to 0, the better. The evaluation results tend to be consistent, and the model has a high degree of certainty in its understanding of the current sample; A larger value indicates a significant discrepancy between the results of multiple evaluations, suggesting that the current input may be an outlier sample deviating from the training distribution.

[0120] The system pre-sets a safety threshold for the predicted entropy (e.g., 0.8): During real-time monitoring, if the calculated predicted entropy... The system will proactively determine if the current operating condition is abnormal and immediately trigger a manual intervention alarm. This mechanism completely avoids the problem of traditional deterministic models blindly giving high-confidence misjudgments when faced with unreliable data, effectively improving the reliability of on-site safety decisions.

[0121] To verify the effectiveness and superiority of the hybrid evaluation model based on echo state Bayesian neural network (ESN-BNN) proposed in the embodiments of this application under strong noise and single-condition environments, performance evaluation, ablation experiments, and benchmark model comparison experiments were conducted on a hard coal seam drilling test set. The specific process is as follows:

[0122] Test set performance results:

[0123] The end-to-end trained ESN-BNN model was finally evaluated on a test set containing 70 samples. Quantitative evaluation results show that the model proposed in this application performs excellently in multi-class classification tasks, achieving an overall classification accuracy of 98.55%, a weighted average recall, and an F1-score of 0.98. Based on the confusion matrix statistics of the test set classification results, the model correctly classified the vast majority of samples into the true categories (38 "excellent," 15 "medium," and 16 "poor" were all correctly classified), with only one anomaly where a "medium" state was misclassified as "poor." Sufficient quantitative metrics demonstrate that this method possesses extremely high state recognition accuracy and reliability even in single-condition, high-noise environments.

[0124] Ablation experiment:

[0125] To verify the core roles of the front-end echo state network in extracting temporal features and the back-end Bayesian network in probabilistic inference, a Flatten-BNN model with the ESN module removed and an ESN-MLP model with the BNN replaced by a multilayer perceptron were constructed for comparative experiments. The quantitative test metrics for each variant model are shown in Table 2.

[0126] Table 2 Quantitative Table of Fusion Experimental Test Performance Indicators

[0127]

[0128] Regarding training stability and convergence accuracy, the Flatten-BNN model, due to its disruption of the time dependence of the data, exhibited severe fluctuations in loss convergence, resulting in highly unstable accuracy, which at one point dropped to 79.71%. The ESN-MLP model's training loss eventually stabilized at around 0.68, and its test accuracy reached a bottleneck at around 78.26%. In contrast, the ESN-BNN model of the embodiment in this application demonstrated the fastest convergence speed and the most robust trajectory, achieving a final accuracy of 98.55%, fully validating the scientific validity and effectiveness of the joint architecture design of the reservoir leakage integral and Bayesian inference.

[0129] Comparative experimental results:

[0130] Existing mainstream deep learning temporal prediction models (LSTM, GRU, 1D-CNN, BiLSTM) were selected as benchmark models, and rigorous quantitative comparison experiments were conducted with the methods provided in the embodiments of this application on the same test set. The quantitative results of the test performance metrics of each model are shown in Table 3:

[0131] Table 3. Quantitative Table of Performance Indicators in Comparative Experiments

[0132]

[0133] Quantitative analysis shows that traditional recurrent neural networks (such as BiLSTM) face significant training difficulties when processing long sequences of signals accompanied by strong noise, with training accuracy only maintaining around 70% and a final test accuracy of only 0.86, the worst performance among all benchmark models. Compared to the suboptimal benchmark model (1D-CNN, accuracy 0.96), the ESN-BNN model of this invention still achieves significant superiority in accuracy, recall, and F1-score, with all performance indicators reaching the optimal level. Comprehensive experimental data thoroughly demonstrates that the method of this invention successfully overcomes the performance bottleneck of benchmark point estimation models under complex nonlinear coupling conditions, possessing excellent diagnostic accuracy and robust generalization ability.

[0134] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the operating status of a mining drilling rig's drilling system based on an echo-state Bayesian neural network, characterized in that, include: Step S1. Data preprocessing: Collect multi-source sensor time-series data during the drilling process of the mining drilling rig, perform sliding window truncation and downsampling preprocessing on the time-series data, and construct a classification dataset for drilling status. Step S2. Dynamic Feature Extraction: The preprocessed time series data is input into the echo state network reservoir with leakage integration mechanism to suppress high-frequency random noise and extract the slow-varying dynamic features of the drilling rig operation. A high-dimensional dynamic feature vector is obtained through dual-view feature aggregation. Step S3. Probabilistic evaluation model training: The high-dimensional dynamic feature vector is input into the Bayesian neural network of ConcreteDropout, and the nonlinear mapping from the feature space to the probability space of the running state is completed based on variational inference; Step S4. Online Inference Early Warning: Perform Monte Carlo sampling forward inference on the input sample, calculate the mean value of the state prediction probability to obtain the evaluation result, and at the same time, use the prediction entropy quantification model to recognize uncertainty and trigger a high entropy anomaly early warning.

2. The method for evaluating the operating status of a mining drilling rig based on an echo-state Bayesian neural network according to claim 1, characterized in that, The specific process of step S1 is as follows: Step S11, Data Acquisition and Extraction: Define the system at discrete time points. The received input signal variable is ,in The number of multi-dimensional sensor channels is determined; sensor time series data of simulated drilling in hard coal seam drilling rig are selected, and invalid drilling data at the start and end stages are removed. The original time series samples are extracted using a sliding window, and the length of the original time window is set to the preset number of sampling points. Step S12, Downsampling: While preserving key frequency domain features, downsampling is performed on the original samples to reduce the dimensionality of the model input. Step S13, Dataset Construction and Partitioning: After downsampling, obtain valid samples, and divide the valid samples into training set, validation set and test set according to a preset ratio to obtain standardized time series samples.

3. The method for evaluating the operating status of a mining drilling rig based on an echo-state Bayesian neural network according to claim 1, characterized in that, The specific process in step S2 is as follows: Step S21. Reservoir Network Construction and Initialization: Construct a reservoir network containing a large number of sparsely connected neurons, with the number of internal neuron nodes being... Control the weight matrix of the reserve pool The spectral radius satisfies the following constraints: in, The number of neurons within the reservoir determines The dimension; For the reserve pool weight matrix The One eigenvalue; For the reserve pool weight matrix spectral radius; During the initialization phase, a random sparse matrix following a standard normal distribution is generated. And then perform spectral radius scaling on it, as shown in the following formula: in, This represents the maximum value of the modulus of the eigenvalues ​​of the matrix. This is a preset spectral radius hyperparameter; Step S22. Leakage Integral State Update: Input the acquired time-series signal vector into the storage pool, perform state evolution in the discrete-time domain, and define... The input signal represents the dynamic intermediate state quantity generated by the evolution of the signal in the reservoir. The transitive relationship over time follows: in, For a moment Preprocessed multi-sensor input signal vector; This represents the state vector within the reservoir, with the dimension corresponding to the size of the neurons in the reservoir. This is the connection weight matrix from the input layer to the reservoir; This is the recursive connection weight matrix within the reserve pool; It is the hyperbolic tangent activation function; Defined as leakage rate; For the reserve pool at the previous moment ; internal state vector; Step S23. Dual-view feature aggregation: After processing by the leakage integral reservoir, the original time-series signal is transformed into a high-dimensional state trajectory sequence that evolves over time; the final state vectors of the reservoir are extracted respectively. With time-averaged state vector The two are then concatenated in a high-dimensional manner to form the final dynamic feature vector. The formula is as follows: in, This represents a vector concatenation operation; the concatenated feature vectors It contains complete system evolution information; This represents the total number of discrete time steps for the timing signal within the current sampling window.

4. The method for evaluating the operating status of a mining drilling rig based on an echo-state Bayesian neural network according to claim 1, characterized in that, The specific process of step S3 is as follows: Step S31. BNN Network Architecture and Reparameterization: Construct a fully connected BNN structure containing an input layer, two Bayesian hidden layers, and an output layer; introduce the Concrete Dropout mechanism into the network layers and use reparameterization techniques to continuously relax the binary mask generation process; define the log-odds ratio. The model utilizes a continuous relaxation form of the Gumbel-Softmax distribution to generate intermediate values ​​of the neuron-preserving mask. Its generation relationship combines structural parameters and random noise, as shown in the following formula: in, Use the Sigmoid activation function; ϵ is a temperature coefficient used to control the shape of the approximate distribution and ensure effective gradient propagation; ϵ is random noise that follows a standard uniform distribution. Step S32. Loss Function Construction and ELBO Optimization: The training objective of the model is to search for the optimal variational posterior distribution of the weights. Approximates the true posterior distribution ;Transfer cost function Cost of the cross-entropy of the fitted input signal Together with the regularization cost that penalizes network complexity, its optimization transitivity is as follows: in: Cross-entropy loss represents the likelihood cost and encourages the sampling parameters in the variational distribution to accurately fit the current observation data. For the first Layer weights Norm; This is the weight decay coefficient; For network layer index; To match the mask rate The relevant structure regularization term, representing the complexity cost, is used to drive the model to find the optimal sparsity; for the th Layer The formula for calculating the number of neurons is: in, For the first The number of neurons in a layer; The mask rate learned by the current model; The initial mask rate; Step S33. End-to-end model training: The network training uses the Adam optimizer for variational parameter training. and weight Iterative updates; The training set is input into the network in batches, and then passed through the ESN state mapping and dual-view feature aggregation before entering the BNN for forward propagation. The parameters are iteratively updated through the backpropagation algorithm. In each round, the validation set is used to monitor the generalization performance and save the best model weights until the training process converges.

5. The method for evaluating the operating status of a mining drilling rig based on an echo-state Bayesian neural network according to claim 1, characterized in that, The specific process of step S4 is as follows: Step S41. Monte Carlo Sampling and Parallel Inference: For new test samples Preserve the ConcreteDropout random masking mechanism in the network and execute it. Sub-independent Monte Carlo sampling forward propagation, employing a Bayesian model averaging strategy, calculates... The average of the predicted probabilities is used as the final classification prediction result, and its calculation formula is as follows: in, This represents the predicted probability distribution of the model's final output. For the training dataset; For new test samples; The predicted category corresponding to the model; For the first Network variational weight configuration during the next sampling; This represents the optimal variational posterior distribution of the network weights. To configure specific variational weights Below, the model prediction probability corresponding to a single sampling; Index of the number of Monte Carlo samples; Step S42. Prediction Entropy Calculation and High Entropy Early Warning Mechanism: While outputting the state evaluation classification result, the model calculates the prediction entropy. To quantify the cognitive uncertainty regarding the current input sample, the calculation formula is as follows: in: In a given training dataset Next, the model applies new samples. Prediction results Predicted entropy; The total number of categories; Assign category index number; For the model to the first The average predicted probability of the class; The predicted category corresponding to the model; The new feature vector input during the testing phase; For observation datasets; The system pre-sets a safety threshold for predicted entropy: during real-time monitoring, if the calculated predicted entropy... The system will proactively determine if the current operating condition is abnormal and immediately trigger a manual intervention alarm.

6. The method for evaluating the operating status of a mining drilling rig based on an echo-state Bayesian neural network according to claim 1, characterized in that, The method also includes model validation and evaluation steps: Model testing and evaluation: Input the trained evaluation model into the test set samples, and calculate the model's classification accuracy, recall, and the harmonic mean score of precision and recall; Ablation experiments were conducted to verify the results: control models with the ESN module removed and control models with BNN replaced by MLP were constructed and tested on the same dataset, and the performance metrics were compared. Benchmark Model Comparison: This model is compared with benchmark models such as LSTM, GRU, 1D-CNN, and BiLSTM on the same test set to verify the effectiveness of the model.

Images