Rock hardness intelligent identification model and method based on deep learning

Abstract

The invention discloses a deep learning-based rock hardness intelligent identification model and method, a DF (data fusion)-improved CNN (Convolutional Neural Network) model provided by the invention still keeps 97.92% stable accuracy under continuous cutting and different working conditions, and the problems of tedious artificial feature extraction engineering and weak model generalization in traditional rock hardness identification are solved. And the problems of insufficient characterization capability and weak model generalization performance of a traditional method are solved. Different from a feature screening process dominated by expert experience in a traditional mode, the method provided by the invention realizes feature adaptive extraction through a multi-channel convolutional neural network, carries out overlapped sampling data enhancement on an original vibration signal, constructs a time-frequency entropy multi-domain fusion graph through short-time Fourier transform, and obtains a time-frequency entropy multi-domain fusion graph; experimental verification shows that the classification accuracy of the multi-channel structure is greatly improved compared with that of a single-channel model, and time-frequency entropy multi-domain fusion is more accurate than that of a single-domain recognition model.

Description

Technical Field

[0001] This invention relates to a rock hardness identification model and method, and more particularly to a rock hardness intelligent identification model and method based on deep learning. Background Technology

[0002] The construction of intelligent mines encompasses not only automation and informatization but also the utilization of advanced sensors, data analysis, and machine learning technologies to improve the safety, efficiency, and environmental friendliness of mining operations. Among these, the tunneling machine (TBM), as a key piece of equipment in underground mining, directly impacts mining efficiency and costs due to its level of intelligence. By intelligently identifying rock hardness through TBM signals, tunneling parameters, such as cutting speed and thrust, can be adjusted in real time to adapt to rock masses of varying hardness, thereby improving tunneling efficiency. Accurate rock hardness identification allows for better planning of the mining sequence and layout, optimizing resource allocation, and extending the mine's service life. Intelligent rock hardness identification helps identify potential geological risks, such as rock bursts and collapses, in advance, enabling preventative measures and improving mine safety. Intelligent identification systems can predict changes in rock hardness, helping operators adjust tunneling strategies promptly and reducing equipment damage and cutter wear caused by sudden changes in rock hardness. Intelligent identification technology provides mine managers with more accurate data support, facilitating the development of more scientific mining plans and decisions. The construction of intelligent mines has promoted the cross-disciplinary integration of geological exploration, mining engineering, mechanical engineering, and information technology, bringing technological innovation and development to the mining industry. Intelligent identification technology helps achieve efficient utilization of mineral resources and protection of the mining environment, supporting the green and sustainable development of the mining industry. In summary, the relationship between intelligent construction of underground metal mines and intelligent identification of rock hardness using tunneling machine signals is reflected in multiple aspects, including improved mining efficiency, reduced equipment wear and tear, safety management, resource optimization, decision support, technological integration, and sustainable development, making it one of the key technologies for achieving intelligent mining.

[0003] Convolutional Neural Networks (CNNs) are deep learning models used to process data with grid-like patterns, and have gained widespread attention due to their outstanding performance in image recognition, object detection, and natural language processing. CNNs extract features from input data through convolutional operations, significantly reducing the number of parameters that need to be trained while preserving spatial relationships. Their main characteristics are weight sharing, local receptive fields, and hierarchical feature extraction. In the convolutional layers of a CNN, the kernel parameters are shared, eliminating the need to learn weights individually for each input location. This reduction in parameters means a smaller model, lowering the risk of overfitting and improving training efficiency and convergence speed. Defect features in signals or images are often localized, and CNN convolutional operations can accurately capture these features through local receptive fields. Compared to global methods, CNNs are more efficient at extracting local features and avoid redundancy. As the network deepens, different convolutional layers in a CNN can progressively extract low-level to high-level features, demonstrating good stability for cross-scale defect recognition and fine-scale prediction.

[0004] However, traditional rock hardness identification models suffer from problems such as cumbersome manual feature extraction and weak model generalization; traditional machine learning methods rely on manual feature engineering, which has the limitation of failing to mine deep representations of high-dimensional data. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based intelligent rock hardness identification model and method to solve the problems of cumbersome manual feature extraction and weak model generalization in traditional rock hardness identification models, as well as the problems of insufficient representation ability and weak model generalization performance of traditional methods.

[0006] To address the aforementioned technical problems, this invention provides a deep learning-based intelligent rock hardness recognition model. The method for constructing this model includes:

[0007] S1: Acquire the raw vibration signal of the target area;

[0008] S2: Expand the dataset using a sliding window oversampling method;

[0009] S3: Convert a one-dimensional time-series signal into a two-dimensional time-frequency-entropy data fusion signal;

[0010] S4: Replace the Softmax classifier with the ELM classifier to build a DF-improved CNN intelligent rock hardness recognition model.

[0011] Furthermore, in step S2, the specific implementation process of expanding the dataset by sliding window overlapping sampling is as follows: according to the preset fixed window size, the window is slid on the original signal sequence with a specific step size, and the signal segments in each window are extracted as new samples in turn.

[0012] Furthermore, in step S3, the one-dimensional vibration signal is converted into a two-dimensional image using a short-time Fourier transform (STFT).

[0013] One aspect provides a deep learning-based intelligent rock hardness identification method, including the following steps:

[0014] S11: The original vibration signal is decomposed by the improved adaptive noise complete set empirical mode decomposition ICEEMDAN, and the signal is reconstructed by screening the effective IMF components in combination with multi-scale permutation entropy MPE. The original signal is then denoised by combining wavelet threshold denoising method.

[0015] S12: Determine the sampling points, number of samples, and sliding window; use overlapping sampling to perform data augmentation, and normalize the collected vibration signal dataset;

[0016] S13: It integrates time domain, frequency domain, and spectral entropy features, and reduces the parameter scale by constructing a multi-channel, multi-branch convolutional architecture and depthwise separable convolution, and embeds a squeeze excitation module to achieve adaptive weighting of feature channels;

[0017] S14: Label the relevant vibration signal feature datasets according to rock hardness, assign corresponding labels, and divide them into training, validation, and test sets in a ratio of 7:1.5:1.5;

[0018] S15: Train the DF-improved CNN model by inputting images from the training set, determine the optimal parameters of the model, and then use the optimized model to identify the test set to obtain the classification and identification results of hardness rocks.

[0019] In summary, the present invention has the following advantages over the prior art:

[0020] The proposed DF-improved CNN recognition model in this invention constructs an end-to-end adaptive recognition architecture for vibration signal processing by deeply fusing the feature self-learning advantages of convolutional neural networks with the efficient classification characteristics of extreme learning machines. This model employs a two-stage collaborative mechanism of "convolutional feature encoding - extreme classification decision": firstly, it automatically extracts the time-frequency local features of vibration signals through multi-scale convolutional kernel groups, replacing traditional manual feature selection; then, the high-dimensional feature map output from the convolutional layer is compressed through global pooling and input into the classification layer of the extreme learning machine with randomly frozen weights for rapid analytical training, achieving decoupling optimization of feature learning and pattern recognition. The network innovatively introduces a multi-channel attention mechanism to optimize the distribution of convolutional kernel weights, and improves the representation ability of key vibration modes through self-calibration of feature channel importance. At the same time, it utilizes the Moore-Penrose generalized inverse matrix analytic property of ELM to replace the original Softmax classifier with ELM while ensuring classification accuracy. Through theoretically guaranteed global optimal solution and adaptive feature recombination, the proposed DF-improved CNN model maintains a stable accuracy of 97.92% under continuous truncation and different working conditions. It solves the problems of cumbersome manual feature extraction and weak model generalization in traditional rock hardness identification, and effectively solves the problems of insufficient representation ability and weak model generalization performance of traditional methods.

[0021] Unlike the traditional feature selection process dominated by expert experience, the method provided by this invention achieves adaptive feature extraction through a multi-channel convolutional neural network, enhances the original vibration signal by overlapping sampling, and constructs a time-frequency entropy multi-domain fusion map through short-time Fourier transform. Experimental verification shows that the classification accuracy of this multi-channel structure is significantly improved compared to the single-channel model, and the time-frequency entropy multi-domain fusion is more accurate than the single-domain recognition model. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0023] Figure 1 This is a schematic diagram of the overlap sampling method;

[0024] Figure 2 A fusion diagram of vibration signal characteristics at different hardness levels;

[0025] Figure 3 This is a flowchart of rock hardness recognition based on DF-improved CNN;

[0026] Figure 4 Optimize the MPE structure graph for the genetic algorithm;

[0027] Figure 5 The signal-to-noise ratios after denoising using three wavelet bases;

[0028] Figure 6 Signal-to-noise ratio after denoising under different thresholds and decomposition levels;

[0029] Figure 7 Here is a flowchart of the noise reduction method;

[0030] Figure 8 To extract vibration acceleration signals from four rock layers of different hardness;

[0031] Figure 9 For temporal graph CNN accuracy and loss rate;

[0032] Figure 10 For frequency domain graph CNN accuracy and loss rate;

[0033] Figure 11 For DF-CNN accuracy and loss rate;

[0034] Figure 12 This is the confusion matrix of the CNN model;

[0035] Figure 13 The confusion matrix for DF-improved CNN models. Detailed Implementation

[0036] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form may also include the plural form unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0038] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0039] This invention provides a deep learning-based intelligent rock hardness recognition model, the construction method of which includes:

[0040] S1: Acquire the raw vibration signal of the target area;

[0041] S2: Expand the dataset using a sliding window oversampling method;

[0042] In the field of deep learning, the size of the data sample is one of the core factors determining model performance. Especially when processing complex data such as images, audio, or time series, insufficient sample size can lead to overfitting and reduced generalization ability. To effectively address this issue, data augmentation techniques utilize coordinated transformations such as rotation, translation, cropping, and scaling on the original dataset to maintain the invariance of key features, constructing diverse training sets and effectively enhancing the model's generalization ability under domain shift and noise interference. For example, considering the unique periodicity and nonlinearity of the vibration signal from the cutting head during rock cutting by a tunnel boring machine, a sliding window overlapping sampling method is used to expand the dataset to maximize the preservation of the signal's temporal continuity and integrity.

[0043] As a preferred method, the specific implementation process of expanding the dataset using the sliding window overlap sampling method is as follows: Following a preset fixed window size, a sliding window is applied to the original signal sequence with a specific step size, and signal segments within each window are extracted sequentially as new samples. For example... Figure 1 As shown, the solid black box represents the nth sampling window, and the dashed red box corresponds to the (n+1)th sampling window. The overlapping orange area between the two windows visually reflects the continuity of adjacent samples, while the moving distance of the window's starting point (i.e., the step size indicated by the orange part) determines the degree of sample overlap and sampling density. By controlling the window size and sliding step size, this method significantly increases the number of effective training samples while avoiding signal feature fragmentation, providing richer temporal feature inputs for subsequent model training.

[0044] The expression for overlap sampling is:

[0045] Where: T represents the total number of samples; N represents the signal length; L represents the sample length; and S represents the sliding step size.

[0046] S3: Convert a one-dimensional time-series signal into a two-dimensional time-frequency-entropy data fusion signal;

[0047] The cutting vibration signal of a tunneling machine exhibits typical nonlinear dynamic characteristics and non-stationary time-varying properties, and its complexity is reflected in three aspects:

[0048] (1) The multi-fault coupling effect results in the simultaneous presence of composite components such as periodic impulses, modulation side frequencies and random resonances in the signal;

[0049] (2) The time-varying propagation path of the impulse response causes the signal to exhibit non-uniform attenuation characteristics;

[0050] (3) Background noise and operating condition fluctuations further exacerbate the difficulty of extracting transient features of the signal.

[0051] For the analysis of such complex signals, the time-frequency joint analysis method shows significant advantages. Through data fusion (DF), that is, constructing a two-dimensional image of the signal in time-frequency-entropy, it can simultaneously capture the time-domain transient characteristics, frequency-domain modulation law and entropy of the event to describe the randomness and complexity of the signal, thus breaking through the limitations of traditional single-domain analysis.

[0052] The introduction of spectral entropy further enhances the physical interpretability and classification accuracy of features. By extracting signal segments through a sliding time window and calculating the spectrum, a time-frequency-energy mapping is generated, intuitively displaying the non-stationary characteristics of the signal. Simultaneously, spectral entropy quantifies the complexity of the signal's frequency domain energy distribution, supplementing the overall frequency domain features that cannot be reflected by local details in time-frequency plots. Furthermore, while time-frequency plots capture dynamic details, spectral entropy characterizes the overall frequency domain pattern; the combination of the two forms a complementary feature system of "local details + global statistics." The fusion of time-frequency plots and spectral entropy, through the synergistic complementarity of local dynamic characteristics in the time-frequency domain and global statistical characteristics in the frequency domain, enables multi-dimensional analysis of complex vibration signals. As a gold standard indicator of frequency domain complexity, spectral entropy not only enhances the physical interpretability of features but also significantly improves the generalization performance of intelligent recognition models by suppressing noise interference and the influence of operating condition fluctuations, demonstrating significant application value.

[0053] As a preferred method, a common approach to converting one-dimensional vibration signals into two-dimensional images is to use Short-Time Fourier Transform (STFT). This method divides the signal into very short time intervals and represents the spectrum of each interval as a two-dimensional image. In STFT, the signal is first segmented into overlapping time windows, and a Fourier transform is performed on each window to obtain its spectral representation. By projecting the one-dimensional signal into the two-dimensional time-frequency domain through STFT, key dynamic information is preserved while data volume is compressed. Spectral entropy further extracts frequency domain statistical features, avoiding the computational burden of directly processing high-dimensional time-frequency matrices. Simultaneously, spectral entropy, calculated based on probability distribution, naturally suppresses the energy spread of random noise, while the time-frequency map, smoothed by a window function, attenuates high-frequency noise; both work synergistically to improve feature robustness. Reconstructing a two-dimensional fused feature map from a one-dimensional vibration signal using consecutive STFTs effectively preserves the spatiotemporal evolution pattern of signal features, providing structured input for deep learning models, such as... Figure 2 As shown.

[0054] The color mapping in the fusion graph is essentially a visualization coding system for the time-frequency energy density of the signal. It quantitatively characterizes the energy distribution intensity of the signal in the joint time-frequency entropy domain through color gradient. Specifically, the hue-brightness mapping function normalizes the STFT power spectral density of the signal, where high-energy regions are mapped to warm hues or high brightness, and low-energy regions correspond to cool hues. This coding mechanism has multiple advantages in engineering recognition: the periodic micro-impacts caused by the signal are represented as equally spaced bright yellow short stripes in the fusion graph; their time intervals correspond to the reciprocal of the characteristic frequency, and the vertical widening of the stripes reflects the bandwidth modulation feature analysis of the impact response; the sideband modulation phenomenon of the frequency is presented in the fusion graph as symmetrically distributed alternating bright and dark vertical bands centered on the frequency, and its modulation depth can be quantified by the rate of change of the color gradient between adjacent bright bands. By quantitatively analyzing the hue-energy relationship of the fusion graph, feature vectors strongly correlated with hardness type can be constructed, and combined with deep learning models, adaptive classification of patterns can be achieved, improving the accuracy of traditional single analysis.

[0055] S4: Replace the Softmax classifier with the ELM classifier to build a data fusion-improved CNN-based intelligent rock hardness recognition model.

[0056] Converting a one-dimensional time-series signal into a two-dimensional image that integrates time, frequency, and spectral entropy via Short-Time Fourier Transform (STFT) offers significant advantages in signal analysis and pattern recognition. Convolutional Neural Networks (CNNs) demonstrate remarkable advantages in signal processing and image recognition. Through their hierarchical feature representation architecture and spatial correlation learning mechanism, they overcome the reliance on manual feature engineering inherent in traditional shallow machine learning models. Compared to traditional methods such as SVM and Random Forest, the core advantages of CNNs lie in: automatically extracting multi-scale features of signals using convolution-pooling operations; and mapping one-dimensional vibration signals to a two-dimensional time-frequency image via STFT, allowing for the full extraction of sensitive information implicit in the spatiotemporal domain through image-based input, thereby further improving model accuracy.

[0057] To further enhance feature completeness, a dual-channel time-frequency domain fusion strategy is adopted, fusing the time-frequency plot generated by the original vibration signal through STFT with the spectral entropy to form a multi-channel feature input. This fusion mechanism relies on the complementary advantages of cross-domain features to effectively enhance the model's recognition performance; by fully considering the correlation and discriminative power between features, it further improves the model's generalization ability and accuracy.

[0058] In summary, this invention proposes a DF-improved CNN rock hardness recognition model, which mainly consists of an input layer, convolutional layers, pooling layers, a fusion layer, a fully connected layer, and a classifier. First, vibration signals are mapped to temporal entropy multi-domain fusion maps, and normalization is used to eliminate dimensional differences and enhance feature comparability. Then, a parallel dual-branch CNN structure is adopted, with the pooling layer using max pooling to compress feature dimensions and suppress noise interference. In the fusion layer, temporal and frequency domain features are integrated through a channel attention-weighted concatenation strategy, and the contribution of the two channels is dynamically calibrated using the SENet module to generate a joint feature vector. A Dropou layer is introduced to prevent overfitting, and regularization is used to enhance generalization ability. Finally, ELM replaces the traditional Softmax classifier to improve the model's recognition accuracy and precision.

[0059] join Figure 3 As shown, a deep learning-based intelligent rock hardness recognition method is used in the DF-improved CNN rock hardness recognition model proposed in this invention, including the following steps:

[0060] S11: The original vibration signal is decomposed using an improved adaptive noise complete set empirical mode decomposition (ICEEMDAN), and the signal is reconstructed by screening effective IMF components using multi-scale permutation entropy (MPE). The original signal is then denoised using a wavelet threshold denoising method.

[0061] I. ICEEMDAN+GA-MPE Combined Noise Reduction Method

[0062] ICEEMDAN is an optimized improvement of the CEEMDAN algorithm. The traditional CEEMDAN method enhances decomposition stability by directly adding Gaussian white noise, but its performance is limited by pre-defined assumptions about the noise distribution (such as Gaussianity and uniformity), making it difficult to adapt to the noise characteristics of complex non-stationary signals. ICEEMDAN, however, proposes an adaptive noise reconstruction mechanism, selecting the k-th IMF component after EMD decomposition of the white noise as an adaptive noise source to replace the original Gaussian white noise.

[0063] Given an original signal x(t), feature enhancement is achieved by iteratively superimposing normalized white Gaussian noise in each step of signal processing. The decomposition process of the ICEEMDAN algorithm is as follows:

[0064] (1) The extreme points of the input signal x(t) are calculated by interpolation, and the upper envelope x is obtained. max (t) and lower envelope x min (t).

[0065] (2) Find the upper and lower envelope mean lines:

[0066] Subtracting the two results gives the remaining component h1(t), which is:

[0067] h1(t)=x(t)-m1(t) (3)

[0068] Determine if h1(t) satisfies the two conditions of the aforementioned IMF. If it does, save h1(t) as an IMF. If it does not, treat h1(t) as the original signal and repeat the above steps to continue the calculation and filtering. If h1(t) obtained after the k-th filtering meets the IMF conditions, then h1(t) is retained as a first-order IMF, denoted as c1(t), i.e.:

[0069] c1(t)=h1(t)=IMF1 (4)

[0070] (3) Subtracting c1(t) yields the first residual, i.e.:

[0071] r1(t)=x(t)-c1(t) (5)

[0072] Treating r1(t) as a new signal again, repeat steps (1) to (3) to obtain c2(t) that satisfies the condition again. Continue this process until r... n The function stops when (t) is a monotonic function or when the termination condition is less than the threshold, which yields:

[0073] Where: r n(t) represents the residual component after signal decomposition; c n (t) represents all IMF components that successively satisfy the conditions.

[0074] (4) That is to say, the first residual r1(t) and the IMF component are:

[0075]

[0076] (5) When k = 2, 3, ..., K, the k-th residual is:

[0077]

[0078] (6) The k-th IMF component is calculated as follows:

[0079] IMF k (t)=r k-1 (t)-r k (t) (9)

[0080] (7) Repeat steps (5) and (6) above until the final residual can no longer be decomposed, then end the loop.

[0081] Where: the signal added for the i-th time (i = 1, 2, ..., I) is denoted as v. i (t); ε i E represents the standard deviation of the noise. k (·) represents the k-th IMF component generated using the EMD algorithm; M(·) represents the mean of the EMD decomposed signal.

[0082] ICEEMDAN employs a staged noise injection strategy, dynamically adjusting the noise amplitude before each IMF extraction layer, rather than injecting noise at a fixed ratio. Therefore, after each decomposition, the residual signal undergoes adaptive filtering to remove residual noise, rather than directly superimposing noise, avoiding the problems of high-frequency noise overload or low-frequency noise insufficiency. Addressing the issues of residual mode mixing, noise coupling interference, and feature information loss inherent in traditional EMD, EEMD, and CEEMDAN decomposition processes, ICEEMDAN systematically solves the inherent defects of the EMD series of methods, achieving a significant leap forward in modal purity, computational efficiency, and engineering applicability.

[0083] The ultimate goal of decomposing vibration signals using the ICEEMDAN method is not simply to obtain multi-scale components, but to achieve targeted enhancement of information components and noise removal through feature-guided reconstruction. During signal reconstruction, it is necessary to efficiently select signal components rich in characteristic frequencies, rather than noise components. Therefore, how to effectively select and denoise the noise components in the IMF is the next issue to consider.

[0084] Entropy holds irreplaceable theoretical value and engineering practicality in signal processing. Its importance lies in its ability to quantitatively describe signal complexity, enhance the robustness of feature extraction, and deeply explore nonlinear dynamic characteristics. Its necessity stems from the analytical needs of modern signal processing for nonstationary and nonlinear systems. Entropy values, by calculating the probability distribution differences of signal sequences, can objectively measure the randomness and orderliness of signals. Traditional frequency domain analysis cannot capture transient changes and nonlinear couplings in signals, while entropy can quantify changes in the dynamic patterns of signals. Furthermore, in low signal-to-noise ratio scenarios, entropy metrics are naturally robust to noise interference.

[0085] Permutation entropy (PE), as a nonlinear complexity measure based on symbolic dynamics, exhibits unique advantages in signal processing, particularly suitable for the dynamic characteristic analysis and feature extraction of nonstationary and nonlinear signals. To achieve quantitative analysis of the nonlinear characteristics of IMF components, a joint feature mapping model is constructed using entropy theory. However, single-scale entropy methods face fundamental limitations when analyzing vibration signals with significant nonstationary characteristics; the single-scale analysis framework struggles to effectively characterize the dynamically changing amplitude-frequency characteristics of signals. Multiscale entropy (MSE), by quantifying the complexity characteristics of time series across multiple scale dimensions, significantly expands the application scope of entropy theory.

[0086] To address the limitation of single-scale entropy in capturing multi-resolution features of non-stationary vibration signals, multi-scale permutation entropy (MPE) quantifies the complex dynamic characteristics of non-stationary signals by constructing multi-scale coarse-grained time series and combining them with the principles of symbolic dynamics. Compared to single-scale permutation entropy, MPE overcomes the limitations of traditional entropy methods in terms of scale uniformity and noise sensitivity by introducing hierarchical analysis of the scale factor s, providing a standardized quantitative tool for multi-scale dynamic analysis of complex systems. Its calculation process is as follows:

[0087] (1) For a given time series X = {x i For the sequence {i = 1, 2, ..., N}, a coarse-grained processing method is used to obtain the following transformation sequence.

[0088] Where: s is the scaling factor, s = 1, 2, ...; [N / s] represents rounding down N / s. When the scaling factor s is 1, the formula degenerates into the original time series.

[0089] (2) Time series reconstruction yields:

[0090] Where: m represents the embedding dimension; l represents the l-th reconstructed component; τ represents the delay time.

[0091] (3) Arranging the obtained time reconstruction sequence in ascending order yields:

[0092] S(g)=(j1,j2,...,j m (12)

[0093] Where: g = 1, 2, ..., k (k ≤ m!). Each subsequence of S(g) is symbolically mapped to m! possible permutations. For any permutation, calculate its probability P. g The complexity of quantizing the time series {x(i), i = 1, 2, ..., N} is now:

[0094]

[0095] (4) If P g =1 / m!, and the permutations generated by its mapping satisfy an equal probability distribution. At this point, the permutation entropy reaches its maximum value, lnm!. Furthermore, to eliminate the influence of the embedding dimension on the entropy dimension, a normalization operation is introduced, namely:

[0096] H P =H P (m) / ln(m!) (14)

[0097] Among them, H P The value ranges from 0 to 1, and its magnitude reflects the dynamic characteristics of the time series, i.e., its complexity.

[0098] II. Adaptive Parameter Selection for MPE

[0099] As a preferred approach, to analyze the overall trend of the data and reveal the asymmetric distribution characteristics, a multi-scale permutation entropy skewness square objective function is constructed. By minimizing the square of the skewness, the genetic algorithm will find the solution with the most symmetric complexity distribution, thereby ensuring a more comprehensive analysis of multi-scale complexity.

[0100] Skewness, as the third-order normalized moment, is a core indicator for quantifying the asymmetry of probability distributions. This parameter reveals the inherent dynamic characteristics of the system by characterizing the directionality and extension of the tail of the data distribution: zero skewness indicates that the mean and median coincide, i.e., a symmetrical distribution; positive skewness indicates the presence of a significant positive impulse component; negative skewness indicates that the left tail of the data distribution is extended, reflecting that the signal is constrained by physical boundaries.

[0101] For the analytic sequence X = {x i Based on the distribution characteristics of {i = 1, 2, ..., N}, construct the permutation entropy under multi-scale factors and form a sequence H. p (X)={H p (1),H p (2),...,H p(s)}, and calculate its skewness index Ske to quantify asymmetry:

[0102] in: For sequence H p The mean of (X); For sequence H p The standard deviation of the sequence is given by E(·), where E is the expected value of the sequence. The objective function is:

[0103] F(X) = Sk 2 (16)

[0104] However, the performance of MPE is highly dependent on the proper setting of its key parameters. Traditional manual parameter tuning methods suffer from low efficiency and susceptibility to local optima. Multi-scale permutation entropy relies on the appropriate selection of key parameters. If the scale factor *s* is too small, feature scale information is overwhelmed, failing to effectively characterize signal features; if *s* is too large, differences in signal complexity are smoothed out by integral effects. The embedding dimension *m* must adhere to the dual criteria of nonlinear dynamics. If *m* is too small, the ability to capture transient impact features is insufficient; if *m* is too large, it significantly increases data requirements while decreasing sensitivity to weak feature amplitudes. In the MPE analysis framework, the selection of the delay time *t* has a significant impact on algorithm performance.

[0105] Genetic Algorithm (GA) is an optimization algorithm that simulates natural selection and genetic mechanisms, and has demonstrated significant advantages in many complex optimization problems. GA's population diversity, flexible encoding, and multi-objective characteristics make it irreplaceable in high-dimensional hybrid optimization. GA initializes the population through hybrid encoding (scale factor, embedding dimension, delay time), evaluates the performance of various parameter combinations through a fitness function, selects and retains individuals with high fitness, and generates a new generation of population through crossover and mutation until convergence is met. Through multiple iterations, it searches for the optimal parameter combination to achieve efficient optimization of MPE (Multi-Optical Process). Its advantages stem from its biomimetic optimization mechanism and high flexibility. Its "exploration-development" balance strategy, noise resistance, and parallelization potential make it the preferred tool for complex engineering optimization problems. The core value of GA in optimizing MPE lies in the perfect match between its hybrid encoding mechanism and global search capability and the parameter characteristics and optimization challenges of MPE. Through a global parallel search mechanism, GA effectively solves the "curse of dimensionality" and local optimum traps in MPE parameter selection. Its algorithm structure is as follows: Figure 4 As shown.

[0106] III. Wavelet Threshold Denoising

[0107] Wavelet thresholding denoising is a non-stationary signal denoising method based on multi-resolution analysis. This method decomposes the signal into multi-scale wavelet coefficients and effectively separates the signal and noise by utilizing the differences in their distributions in the wavelet domain. Its core process includes:

[0108] (1) Adaptive wavelet decomposition: The original signal is decomposed into high-frequency detail coefficients and low-frequency approximation coefficients at different scales through wavelet transform;

[0109] (2) Quantization decision of the optimal wavelet basis: Select the optimal wavelet basis function based on the signal characteristics and wavelet basis properties;

[0110] (3) Thresholding of wavelet decomposition coefficients: Threshold quantization is performed on high-frequency coefficients to suppress weak coefficients dominated by noise;

[0111] (4) High-fidelity signal reconstruction: The noise-reduced signal is reconstructed by inverse wavelet transform using the retained effective coefficients.

[0112] From a mechanistic perspective, noise-dominated wavelet coefficients are typically concentrated in the high-frequency band and have small amplitudes, while real signals exhibit significant, correlated coefficients. By setting a reasonable threshold, wavelet thresholding denoising can achieve signal-to-noise separation: when the absolute value of a wavelet coefficient is below the threshold, it is identified as a noise component and attenuated or zeroed out; coefficients above the threshold are retained as valid signal components. This process effectively filters out random noise interference while preserving the signal's abrupt change characteristics.

[0113] In wavelet denoising, the selection of wavelet basis functions is crucial. Based on the characteristics of vibration signals and computational efficiency considerations, this study uses discrete wavelet basis functions for signal processing. Typical discrete wavelet basis functions include Haar, Coiflet (coif), Daubechies (db), Biorthogonal (bior), and Symlet (sym) wavelets. Their selection requires a systematic trade-off based on key characteristics such as quadrature support, regularity, symmetry, and vanishing moment. As the earliest orthogonal compactly supported wavelet, the Haar wavelet is relatively simple to compute. However, due to its poor smoothness, discontinuity in the time domain, low frequency resolution, and low vanishing moment, it cannot provide sufficient detail. Compared to the Haar wavelet, the Coiflet (coif) wavelet is more computationally complex, with a finite vanishing moment and a larger support width. Due to the characteristics of biorthogonal wavelets, the Bior wavelet requires consideration of two wavelet functions (one for decomposition and one for reconstruction), making its implementation and design more difficult. The increased regularity of the db wavelet helps to obtain smooth results during signal reconstruction. Its high vanishing moment is helpful for representing signals with polynomial phase in the wavelet domain, but its longer support length requires more computation, impacting real-time performance due to its computational complexity. The Symlet wavelet (sym), as an improved algorithm of the Daubechies wavelet (db), effectively reduces phase distortion during wavelet transform by enhancing symmetry.

[0114] For example:

[0115] This invention uses db1-db8, sym1-sym8, and coif1-coif5 as wavelet bases for experiments. The signal-to-noise ratio (SNR) results after denoising using the three wavelet bases are as follows: Figure 5 As shown. The sym7 wavelet basis function was ultimately chosen. SNR is an important indicator of signal quality and system performance; a high SNR indicates good signal quality and low noise impact.

[0116] In the construction of a denoising system based on the sym7 wavelet basis, the optimal selection of the number of decomposition layers follows a time-frequency resolution balance criterion. Experimental studies show that the selection of the number of decomposition layers has a significant impact on the denoising effect: too high a number of decomposition layers will cause excessive signal suppression, while too low a number of layers will lead to noise residue; both will reduce the signal-to-noise ratio. Figure 6 As shown, the optimal denoising effect can be obtained with a 4-level decomposition. To ensure experimental consistency, this invention uniformly uses the sym7 wavelet basis function in conjunction with the parameter configuration of the 4-level decomposition for subsequent processing.

[0117] Besides the wavelet basis and the number of decomposition levels, key parameters for wavelet denoising include threshold selection rules and threshold function type. While hard thresholding functions can better preserve signal amplitude characteristics, they are prone to introducing signal discontinuities and oscillations. Soft thresholding functions, on the other hand, ensure the continuity of the reconstructed signal but suffer from amplitude attenuation. Both thresholding functions have their own characteristics and good engineering applicability. Compared to hard thresholding denoising, soft thresholding overcomes the discontinuity defect, resulting in a smoother denoised signal. Therefore, soft thresholding functions are more commonly used when processing vibration signals from components such as cutting arms in mining equipment.

[0118] Therefore, this invention proposes a joint denoising method based on ICEEMDAN and genetic algorithm-optimized MPE. This method innovatively introduces a genetic algorithm to optimize MPE parameters, accurately extracts the IMF that fully characterizes the original signal features, and combines wavelet thresholding to denoise the vibration signal during the tunnel boring machine cutting process. The specific process is as follows: Figure 7 As shown.

[0119] S12: Given the sampling points, sample size, and sliding window, the overlapping sampling method is used to augment the data, and the collected vibration signal dataset is normalized.

[0120] S13: It integrates time domain, frequency domain, and spectral entropy features, and reduces the parameter scale by constructing a multi-channel, multi-branch convolutional architecture and depthwise separable convolution, and embeds a squeeze excitation module to achieve adaptive weighting of feature channels;

[0121] S14: Label the relevant vibration signal feature datasets according to rock hardness, assign corresponding labels, and divide them into training, validation, and test sets in a ratio of 7:1.5:1.5;

[0122] S15: Train the DF-improved CNN model by inputting images from the training set, determine the optimal parameters of the model, and then use the optimized model to identify the test set to obtain the classification and identification results of hardness rocks.

[0123] Example:

[0124] Vibration acceleration signals of the cutting arm were collected from the signal database of an onboard high-capacity data recording and vibration testing system at a metal mine in Anhui Province. Two hundred sets of vibration signals for different rock hardnesses were collected, including vibration signals for cutting rocks of varying hardness. Figure 8 As shown.

[0125] A vibration signal dataset was constructed using the overlapping sliding window sampling technique, resulting in 400 sets of experimental data for each type. There were a total of 4 different types of vibration signals, which were divided into training set, validation set, and test set in a ratio of 7:1.5:1.5.

[0126] To verify the effectiveness of the multi-domain fusion model proposed in this invention, time-domain and frequency-domain plots of single-channel input were selected as controls for comparative experiments. The hardware and software environment for this experiment was as follows: MATLAB language environment; CPU: Intel(R) Core(TM) i7-12700H@2.30GHz; GPU: NVIDIA GeForce RTX3060 Laptop; RAM: 16GB; OS: Windows 11. The parameters of the improved CNN model were: initial image size 227*227*3, 5 convolutional layers, 7 activation function layers (ReLU activation function), 3 max-pooling layers, 2 drop layers, 2 fully connected layers, and the classification layer replaced the original Softmax with ELM. In the experiment, except for the difference between multi-channel and single-channel in the feature fusion stage, all other structural parameters remained consistent. Iterative curves of loss function versus accuracy were plotted through experiments, with a total of 35 iterations and 1120 iterations.

[0127] From the experimental results Figure 9 , Figure 10 and Figure 11 It can be seen that the single-channel temporal graph model performs well on both the training and validation sets, but exhibits significant curve fluctuations and insufficient stability in the early iteration stages, potentially leading to local overfitting. The single-channel temporal graph model only converges after 70 iterations, while the single-channel frequency graph model only converges after 150 iterations, revealing the limitations of single-channel deep feature extraction capabilities. In contrast, the model based on multi-channel feature fusion images demonstrates significant advantages: it converges after approximately 40 iterations, the training and validation set curves show a reasonable separation trend, and the final classification accuracy is significantly better than the single-feature input model. This verifies that the multi-domain feature fusion strategy, by effectively integrating the correlations between different features, significantly enhances the model's multi-dimensional feature representation capabilities, thereby achieving superior classification performance in pattern recognition tasks.

[0128] Experimental verification:

[0129] To systematically verify the effectiveness of the multi-domain feature fusion model, this experiment incorporates a controlled ablation study. This invention uses the confusion matrix output by the rock hardness identification method to evaluate the performance of different models. By comparing the predicted results with the actual results, the classification performance of the model can be intuitively measured, clearly reflecting the classification accuracy and misclassification rates under different categories.

[0130] To verify the superiority of the DF-improved CNN rock recognition model proposed in this invention, a comparative analysis was conducted with other single-domain and multi-domain recognition models. The confusion matrices of each model are as follows: Figure 12 and Figure 13 As shown.

[0131] according to Figure 12 and Figure 13 Confusion matrix and accuracy analysis show that the classification performance of single-time-domain graph CNN models, single-frequency-domain graph CNN models, and single-spectral entropy graph CNN models is significantly weaker than other models. In contrast, although both time-frequency-domain graph CNN and DF-CNN models exhibit misclassification, the latter demonstrates better accuracy, while the misclassification distribution of single-channel models is more discrete. Notably, the DF-improved CNN model performs best, achieving the highest accuracy.

[0132] The detailed recognition results of different models are shown in Table 1, which fully demonstrates the superiority and generalization ability of the DF-improved CNN model.

[0133] Table 1 Recognition results of different models

[0134]

[0135] The CNN-SVM-based multi-domain fusion recognition system adopts a three-stage process: First, the original vibration signal undergoes time-frequency entropy multi-domain processing to construct a dataset containing 1600 samples; then, a dynamic learning rate strategy is used to train the CNN network and control model convergence; finally, features from the fully connected layers of the CNN are extracted, and a grid search-optimized SVM classifier is used for decision-making. As shown in Table 1, this DF-improved CNN model maintains a stable accuracy of 97.92% under continuous truncation and different working conditions, verifying the engineering applicability of the feature fusion strategy.

[0136] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent rock hardness recognition model, characterized in that, The method for constructing the recognition model includes: S1: Acquire the raw vibration signal of the target area; S2: Expand the dataset using a sliding window oversampling method; S3: Convert a one-dimensional time-series signal into a two-dimensional time-frequency-entropy data fusion signal; S4: Replace the Softmax classifier with the ELM classifier to build a DF-improved CNN intelligent rock hardness recognition model.

2. The intelligent rock hardness recognition model based on deep learning according to claim 1, characterized in that, In step S2, the specific implementation process of expanding the dataset by sliding window overlapping sampling is as follows: according to the preset fixed window size, the window is slid on the original signal sequence with a specific step size, and the signal segments in each window are extracted as new samples in turn.

3. The intelligent rock hardness recognition model based on deep learning according to claim 1, characterized in that, In step S3, the one-dimensional vibration signal is converted into a two-dimensional image using a short-time Fourier transform (STFT).

4. A deep learning-based intelligent rock hardness recognition method, used in the DF-improved CNN intelligent rock hardness recognition model, comprising the following steps: S11: The original vibration signal is decomposed by the improved adaptive noise complete set empirical mode decomposition ICEEMDAN, and the signal is reconstructed by screening the effective IMF components in combination with multi-scale permutation entropy MPE. The original signal is then denoised by combining wavelet threshold denoising method. S12: Determine the sampling points, number of samples, and sliding window; use overlapping sampling to perform data augmentation, and normalize the collected vibration signal dataset; S13: It integrates time domain, frequency domain, and spectral entropy features, and reduces the parameter scale by constructing a multi-channel, multi-branch convolutional architecture and depthwise separable convolution, and embeds a squeeze excitation module to achieve adaptive weighting of feature channels; S14: Label the relevant vibration signal feature datasets according to rock hardness, assign corresponding labels, and divide them into training, validation, and test sets in a ratio of 7:1.5:1.5; S15: Train the DF-improved CNN model by inputting images from the training set, determine the optimal parameters of the model, and then use the optimized model to identify the test set to obtain the classification and identification results of hardness rocks.

Images