Surrounding rock identification method based on TBM cutterhead vibration signal sampling under angle domain

Abstract

The invention discloses a TBM cutterhead vibration surrounding rock recognition method based on angle domain downsampling, belongs to the technical field of TBM construction monitoring, and solves the problems that traditional surrounding rock recognition depends on manpower, the precision is low and the real-time performance is poor. According to the technical scheme, the method comprises the steps that cutter head three-axis vibration signals are collected and preprocessed in real time; extracting a tangential component, and performing mean filtering and sine fitting to obtain a cutterhead rotation period; calculating a phase, intercepting a single-circle vibration signal, and converting the single-circle vibration signal from a time domain to an angle domain; resampling the sample into a standardized sample with fixed length and uniform angle interval through interpolation; and finally, inputting the sample into the trained one-dimensional convolutional neural network model, and directly outputting the surrounding rock grade of the current tunnel face. The method is mainly used for real-time automatic identification of the surrounding rock grade in the TBM tunneling process so as to guide tunneling parameter optimization and construction safety control.

Description

Technical Field

[0001] This invention relates to the field of TBM construction monitoring technology. More specifically, this invention relates to a method for intelligent identification of surrounding rock based on TBM cutterhead vibration signals sampled in the angular domain. Background Technology

[0002] In TBM tunneling construction, real-time and accurate identification of the surrounding rock grade ahead of the tunnel face is crucial for ensuring construction safety and optimizing tunneling parameters. Traditional geological exploration methods, such as core drilling, are insufficient to meet the need for continuous and real-time acquisition of rock mass information during tunneling. Therefore, utilizing data generated during construction, especially cutterhead vibration signals that directly reflect the interaction between the cutterhead and the rock mass, for intelligent identification of the surrounding rock has become an important research direction.

[0003] Existing identification methods based on cutterhead vibration signals still face some fundamental technical challenges in the entire process from raw data to effective model input, which restricts the accuracy and reliability of their practical applications. These challenges are mainly concentrated in two interrelated aspects: data preprocessing and model construction.

[0004] At the data preprocessing level, the core challenge lies in transforming non-stationary time-domain vibration signals into model inputs with clear physical meaning and uniform specifications. In actual tunneling, the cutterhead rotation speed is not constant, leading to inherent defects in directly acquired, time-interval vibration signals: signal segments of the same length correspond to different numbers of cutterhead rotations, meaning the spatial scanning range of the tunnel face they represent is inconsistent. This inconsistency causes the data samples to lose their unified spatial semantics; a single data point cannot stably correspond to a specific azimuth region on the tunnel face. To adapt to neural network models requiring fixed-length inputs, it is usually necessary to standardize the original signal segments of varying lengths. However, conventional random sampling or mean sampling methods are not suitable for TBM cutterhead vibration signals. These signals exhibit typical characteristics of frequent fluctuations, near-zero mean, and alternating positive and negative values. Key impact events containing rock mass information are sparsely and irregularly distributed along the time axis. The aforementioned conventional sampling methods easily destroy the waveform integrity of these key impact events or introduce distorted amplitude representations, resulting in the loss of a large amount of detailed information reflecting rock mass characteristics. Therefore, designing a preprocessing method that can eliminate the influence of rotational speed and normalize the sample length while preserving the impact waveform characteristics associated with spatial location in the original signal to the greatest extent possible is a difficult problem that has not yet been systematically solved in this field.

[0005] At the model building level, the effectiveness of the model itself depends on its ability to process the aforementioned signals. Current research often uses recurrent neural networks (such as LSTM and GRU) to model vibration signal sequences in order to capture their temporal dependencies. However, cutterhead vibration signals are information-intensive data; to fully capture the information from one rotation of the cutterhead, the required sequence length is extremely long. When processing such long sequences, recurrent neural networks are prone to gradient vanishing or exploding problems, making it difficult for the model to effectively learn long-distance dependencies within the sequence, thus limiting further improvements in model performance. Furthermore, since the aforementioned data preprocessing stage failed to provide standardized samples with clear spatial semantics, the model is forced to learn patterns from semantically ambiguous raw sequences. This undoubtedly increases the difficulty of model learning, making it difficult for the model to stably establish a mapping relationship between vibration patterns and surrounding rock categories.

[0006] In summary, the shortcomings of existing technologies can be attributed to the lack of a standardized preprocessing method capable of assigning clear and consistent spatial semantics to cutterhead vibration signals, and a robust identification model architecture suitable for processing the resulting long sequences of high-fidelity signals. Solving these problems is a key prerequisite for achieving high-precision, real-time identification of surrounding rock grades. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0008] Another objective of this invention is to provide a method for identifying surrounding rock based on TBM cutterhead vibration signals sampled in the angle domain. This method aims to solve the technical problems of inconsistent spatial range of the tunnel face represented by the collected vibration signals due to changes in cutterhead rotation speed, ambiguous spatial semantics, and the tendency for gradient vanishing or exploding and poor model performance when using recurrent neural networks for modeling due to excessively long signal sequences.

[0009] To achieve these objectives and other advantages of the present invention, a method for identifying surrounding rock based on angle-domain downsampling TBM cutterhead vibration signals is provided, comprising the following steps: Step 1: Real-time acquisition of triaxial vibration acceleration signals at the cutterhead during TBM tunneling, and preprocessing of the triaxial vibration acceleration signals to obtain preprocessed signals; Step 2: Extract the tangential vibration component from the preprocessed signal, perform mean filtering on the tangential vibration component to extract its periodic gravity component, and fit the gravity component using a sine function model to obtain the periodic information of the cutter head rotation. Step 3: Based on the periodic information of the cutter head rotation, calculate the phase angle corresponding to each moment of the preprocessed signal, and starting from the 0 phase, cyclically extract the vibration signal segments corresponding to the phase intervals of length 2π to obtain the vibration sample of one rotation of the cutter head. Step 4: Convert the vibration samples from the time domain to the angle domain and establish the mapping relationship between vibration acceleration and the rotation angle of the cutter head; resample the vibration samples into standardized angle domain samples with fixed length and uniform angle intervals through interpolation. Step 5: Input the standardized angle domain samples into the pre-trained one-dimensional convolutional neural network model and output the surrounding rock grade corresponding to the current working face.

[0010] Preferably, in the surrounding rock identification method based on angle-domain downsampling of TBM cutterhead vibration signal, step two, mean filtering of the tangential vibration component, is achieved by equation (1). (1); Where, represents the filtered data value. n For window length, x i Indicates the first in the window i One data point; The model of the sine function is shown in equation (2). g(t)=A×sin(2πft+φ)+B(2); Where g(t) is the gravitational component at time t; A is the amplitude in g; f is the frequency in Hz; t is the time in s; φ is the initial phase in rad; and B is the offset in g.

[0011] Preferably, in the surrounding rock identification method based on TBM cutterhead vibration signal sampling in the angle domain, step three involves calculating the phase angle θ corresponding to each moment of the preprocessed signal using equation (3). m (t); θ m (t)=(2πft+φ)-k×2π(3) Where, θ m (t) represents the mechanical phase angle of the sensor relative to the true north direction of the cutter head at time t, in rad; f is the frequency, in Hz; t is the time, in s; φ is the initial phase, in rad; k is the value satisfying 0 ≤ θ m Integers whose t is less than 2π are obtained by k=floor((2πft+φ) / (2π)), where floor() represents the floor function.

[0012] Preferably, in the surrounding rock identification method based on angle-domain downsampling of TBM cutterhead vibration signals, step four, establishing the mapping relationship between vibration acceleration and cutterhead rotation angle, and resampling, are achieved through the following methods: S41. For the nth vibration signal, establish the mapping relationship between its time-domain acquisition time and the rotation angle of the cutter head, which is defined by the forward mapping function (4) and the inverse mapping function (5): θn (t)=[(t-t n,start ) / T n ]×2π, t∈[t n,start ,t n,start +T n (4); t n (θ)=(θ×T n ) / 2π+t n,start (5); Where, θ n (t) represents the rotation angle of the nth cutterhead at time t (corresponding to the angle range for the cutterhead to complete one revolution, its physical meaning being to mark the spatial rotational attitude of the cutterhead relative to its initial position at that moment), in rad, with a value range of [0, 2π). n,start The starting time of the nth rotation of the cutterhead (calibrated by the phase angle calculation result, it is the time-domain starting reference point for dividing the single-rotation vibration data, and is the time-domain starting reference point for the nth rotation that satisfies θ) m At the moment (t)=0, the phase angle θ is calculated continuously by detecting... m For a sequence (t), find its value that jumps from near 2π to near 0 (or satisfies θ). m (t)<θ m The point at which the difference between (t-1) and (t-1) is greater than π is determined as the starting time t of one revolution of the cutter head. n,start (, the unit is s; T) n The nth rotation period of the cutter head, in seconds, is obtained by converting the cutter head rotation frequency f obtained from the fitting of Formula 2. The conversion relationship is T. n =1 / f; t n (θ) represents the time-domain acquisition time corresponding to angle θ, in seconds; t represents the time-domain acquisition time of the signal (the time dimension reference when the original vibration signal was acquired), in seconds; θ represents the rotation angle of the nth rotation of the cutter head, in rad, with a value range of [0, 2π). S42. Based on the inverse mapping function, the vibration acceleration value a at angle θ is obtained through formula (6). n (θ); a n (θ)=S n (t n (θ))(6); Among them, a n (θ) represents the vibration acceleration value at angle θ during the nth rotation of the cutterhead (a rock mass response data associated with a specific spatial location on the face in the angle domain), in g; S n The nth cycle is the preprocessed time-domain vibration signal vector (which is the original effective data source for angular domain mapping), with units of g; S43. Within a complete rotation cycle of 0 to 2π radians ([0, 2π)), a fixed number of angle points M are preset; based on the mapping relationship, the angle points {θ1, θ2, ..., θ3} are calculated using an interpolation method. M The corresponding vibration acceleration value sequence {a} (in rad) n (θ1),a n (θ2),...,a n (θ M This sequence is a standardized angular domain sample with a fixed length and uniform angular spacing.

[0013] Preferably, the surrounding rock identification method based on angle-domain downsampled TBM cutterhead vibration signal involves the following sequential processing steps performed by a one-dimensional convolutional neural network model: S51. Input the standardized angle domain samples into the first convolutional layer for feature extraction, and input the output of the first convolutional layer into the first max pooling layer for downsampling; the number of channels of the first convolutional layer is 64, the kernel size is 7, the stride is 1, the activation function is Tanh, and the kernel size of the first max pooling layer is 10, the stride is 10. S52. Input the output of the first max pooling layer into the second convolutional layer for feature extraction, and input the output of the second convolutional layer into the second max pooling layer for downsampling; the second convolutional layer has 32 channels, a kernel size of 7, a stride of 1, and an activation function of Tanh; the second max pooling layer has a kernel size of 10 and a stride of 10. S53. Input the output of the second max pooling layer into the flattening layer to convert it into a one-dimensional feature vector; S54. The one-dimensional feature vector is sequentially input into the fully connected layer of the classification module for decision-making. The classification module outputs the classification probability of the surrounding rock grade. The fully connected layer of the classification module includes a first fully connected layer and a second fully connected layer. The first fully connected layer has 256 neurons and the activation function is ReLU. The second fully connected layer has 4 neurons and the activation function is Softmax.

[0014] Preferably, in the rock identification method based on TBM cutterhead vibration signal sampling in the angle domain, the step two of mean filtering of the tangential vibration component to extract its periodic gravity component is specifically implemented through the following steps: S21. Use mean filtering to process the tangential component of the cutterhead vibration and obtain the gravitational acceleration variation curve of the tangential component of the cutterhead vibration as a preliminary estimate of the periodic gravitational component. S22. Perform Hilbert transform on the preliminary estimate of the periodic gravity component, calculate its instantaneous amplitude and instantaneous phase, and evaluate the reliability based on the stationarity of the instantaneous amplitude and the linearity of the instantaneous phase. S23. When the confidence assessment result is lower than the preset threshold, the initial periodic gravity component estimate is resynchronized and corrected based on the pulse signal of the cutter head drive encoder, and the corrected signal is used as the final periodic gravity component. Otherwise, the initial periodic gravity component estimate is directly used as the final periodic gravity component for sine function model fitting.

[0015] Preferably, in the surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal, step S43 involves calculation using an interpolation method, specifically including: a) Using cubic spline interpolation, the vibration acceleration value sequence corresponding to M uniformly distributed angle points is initially calculated; b) Identify the angle sub-intervals in the preliminarily calculated vibration acceleration value sequence that correspond to the drastic changes in signal amplitude caused by tool impact; c) Within the angle sub-interval, a piecewise cubic Hermite interpolation polynomial algorithm is used to locally recalculate the vibration acceleration values, replacing the initial calculation results within the sub-interval, to obtain optimized standardized angle domain samples for model input.

[0016] Preferably, the surrounding rock identification method based on angle-domain downsampled TBM cutterhead vibration signal is obtained using a one-dimensional convolutional neural network model in the following manner: Historical tunneling data was acquired and processed to obtain a standardized angle domain sample training set; Data augmentation processing is performed on the training set, including: based on the geomechanical parameters associated with the surrounding rock grade corresponding to the sample, amplitude modulation related to rock strength and phase modulation related to rock mass structure are applied to the sample signal, and noise that conforms to the spectral characteristics of rock mass fracture is superimposed to generate augmented samples. The one-dimensional convolutional neural network model is trained using a training set of augmented samples to obtain a pre-trained model.

[0017] Preferably, the surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal further includes step six: confidence evaluation and feedback optimization of the surrounding rock identification results, specifically including: S61. Record the surrounding rock grade classification probability vector output by the one-dimensional convolutional neural network model in step five, and take the maximum probability value as the confidence score of the identification result. S62. Set a confidence threshold. When the confidence score is lower than the confidence threshold, determine that the confidence of the current recognition result is insufficient. Extract the statistical features of the current standardized angle domain sample and perform similarity matching with the features of high-confidence samples in the historical database. S63. Bind the final confirmed surrounding rock grade on site with its corresponding standardized angle domain sample to form a new training data pair, and periodically add it to the training dataset to trigger incremental learning and parameter fine-tuning of the one-dimensional convolutional neural network model.

[0018] The present invention has at least the following beneficial effects: 1. This invention extracts the periodic gravity component of the tangential component and accurately fits the rotation period to calculate the phase angle at each moment, using this as a basis to extract a single-cycle vibration signal. Subsequently, a strict mapping relationship is established between the time domain and the cutterhead rotation angle, and this is converted into standardized angle domain samples with fixed lengths and uniform angle intervals through interpolation and resampling. This process ensures that each sample point is stably associated with a specific spatial orientation of the face, completely eliminating the influence of rotational speed fluctuations. This provides the model with inputs that have clear and unified spatial semantics, fundamentally improving the reliability and consistency of subsequent identification.

[0019] 2. For angular domain samples with spatial sequence characteristics after standardization, this invention employs a specially designed one-dimensional convolutional neural network for end-to-end recognition. This model, through a double-layer convolution and pooling structure, can automatically extract local waveform details from the signal level by level and aggregate them to form high-level semantic features (such as impact patterns), effectively capturing key information reflecting rock mass characteristics. Compared with traditional recurrent neural networks, the one-dimensional convolutional structure avoids gradient problems when processing long sequences, has higher computational efficiency, and is better at mining spatial local correlations in the signal, thus significantly improving the model's accuracy and stability in judging the surrounding rock grade.

[0020] 3. This invention introduces enhancement mechanisms in multiple stages to ensure reliability in complex engineering environments. In the periodic extraction stage, the robustness of rotational information acquisition is improved by combining Hilbert transform confidence assessment and encoder signal correction. In the angle domain resampling stage, a hybrid strategy combining global smooth interpolation and local conformal interpolation is adopted to ensure signal continuity while preserving key high-frequency abrupt changes caused by tool impact, thus improving the fidelity of the input signal. Furthermore, the system includes confidence assessment and an incremental learning loop based on on-site feedback, which provides auxiliary prompts for low-confidence results and continuously optimizes the model using newly confirmed samples, giving the system excellent adaptability and continuous evolution capabilities, ensuring accuracy for long-term applications.

[0021] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0022] Figure 1This is a flowchart illustrating the surrounding rock identification method based on TBM cutterhead vibration signal sampling in the angle domain as described in this invention. Figure 2 This is a schematic diagram showing the installation position of the triaxial accelerometer on the TBM cutter head of the present invention; Figure 3 This is a schematic diagram illustrating the effect of mean filtering and sine function fitting on the tangential vibration component of the cutter head according to the present invention; Figure 4 This is a schematic diagram illustrating the definition of the phase angle during the rotation of the cutter head in this invention; Figure 5 This is a schematic diagram illustrating the process of generating a vibration sample (angle domain downsampling sample) from a continuous signal to extract the vibration sample of the cutter head rotating one revolution. Figure 6 This is a structural framework diagram of the one-dimensional convolutional neural network (1DCNN) model used in this invention; Figure 7 This is a visualization of the activation response of the intermediate layer in the one-dimensional convolutional neural network model of this invention when processing specific samples. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments, so that those skilled in the art can implement it based on the description.

[0024] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0025] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0026] In the description of this invention, the terms "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0027] like Figure 1As shown, this paper provides a method for identifying surrounding rock based on TBM cutterhead vibration signals sampled in the angle domain. It uses a database as the data support foundation and goes through signal acquisition and preprocessing, periodic extraction, single-cycle sample truncation, angle domain standardization, intelligent identification and feedback optimization in sequence to form a complete technical closed loop. It effectively solves the problems of relying on manual labor, low accuracy and poor real-time performance in existing technologies for surrounding rock identification. At the same time, it overcomes the defects of traditional recurrent neural networks in processing long sequence signals, such as gradient vanishing or exploding and semantic ambiguity in sample space. Figure 1 The database is used to store various key data accumulated during the historical tunneling process of TBM, including raw data of cutterhead triaxial vibration acceleration corresponding to different surrounding rock grades, preprocessed standardized angle domain samples, field-confirmed surrounding rock grade labels, geomechanical parameters (such as rock strength and rock mass structure related data), cutterhead operating parameters (such as historical rotation speed and drive encoder pulse signal records), as well as enhanced samples and training logs generated during model training. These data not only provide sufficient sample support for the initial training of the one-dimensional convolutional neural network model, but also lay the data foundation for subsequent confidence assessment of recognition results, similarity matching, and incremental model learning, realizing the effective linkage between historical experience and real-time recognition.

[0028] First, real-time acquisition and preprocessing of TBM cutterhead vibration signals are performed. During the TBM tunneling process, according to... Figure 2 The sensor acquisition method shown involves installing a triaxial accelerometer at the scraper plate of the cutter head. Figure 2 The relative positions of the sensor, cutterhead, and rotation center are clearly marked. The sensor is numbered 1, the cutterhead is numbered 2, and the rotation center is numbered 3. This installation method ensures that the sensor can rotate synchronously with the cutterhead, accurately capturing the vibration response generated by the interaction between the cutterhead and the rock mass. The sensor collects axial, tangential, and radial triaxial vibration signals at the cutterhead in real time at a sampling frequency of 3200Hz, with a single acquisition duration of 2 minutes, comprehensively covering the multi-directional vibration state of the cutterhead. After acquisition, the triaxial vibration acceleration signal is preprocessed. First, detrending is performed to filter out DC values ​​caused by environmental noise and sampling anomalies. The detrending method follows the trend separation logic in mathematics, eliminating linear or nonlinear trend components in the signal and retaining the true vibration response. Next, quantile outlier removal is performed. A quantile threshold 'a' is set, and the portion of the data distributed between 'a' and 100-a% is defined as the valid confidence interval. Data within this interval is retained, and outliers at the ends of the interval are filled with 0. These two preprocessing steps yield a preprocessed signal with optimized quality. Simultaneously, the real-time acquired raw signal and the preprocessed signal are uploaded to the database to supplement the real-time data reserve.

[0029] Next, the cutter head rotation period is extracted. The tangential vibration component is specifically extracted from the preprocessed signal because it contains a periodic gravitational component consistent with the cutter head rotation period; this characteristic is crucial for obtaining the rotation period. The tangential vibration component is then processed using mean filtering. Mean filtering is based on the sliding window averaging algorithm in mathematics. By setting a fixed-length window, data is truncated along the tangential vibration signal, and the mean of all data within the window replaces the data at the current position, achieving low-pass filtering of the signal and thus initially extracting the periodic gravitational component. The effect of mean filtering and subsequent trigonometric function fitting is shown below. Figure 3 As shown, Figure 3 The image clearly shows the comparison between the original data and the fitted curve. The horizontal axis represents time in seconds, and the vertical axis represents the tangential vibration correlation values. The expression corresponding to the fitted curve is y=1.03sin(2π×0.06x+1.13)+0.11, with an error value of 0.09. This expression strictly follows the basic form of the sine function model, where 1.03 corresponds to the amplitude A, 0.06 corresponds to the frequency f, 1.13 corresponds to the initial phase φ, and 0.11 corresponds to the offset B. This intuitively demonstrates the high degree of consistency between the fitted curve and the original gravity component variation trend. The gravitational component was then fitted using a sinusoidal function model. This model originates from the classic description of periodic vibration phenomena in physics. In this implementation, the periodic variation characteristics of the gravitational component during cutterhead rotation were combined, and the model was applied to accurately fit the gravitational component. By adjusting parameters such as amplitude, frequency, initial phase, and offset, the fitting results were made to closely match the actual variation law of the gravitational component. Finally, the periodic information of the cutterhead rotation was accurately obtained through the fitted parameters, providing a core benchmark for subsequent phase calculation and sample extraction. At the same time, the extracted periodic information was associated with the corresponding tunneling condition data and stored in the database to enrich the data dimensions.

[0030] Based on the acquired cutter head rotation cycle information, the phase angle of the preprocessed signal at each moment is calculated. The calculation logic of the phase angle is based on the phase principle of trigonometric functions. Taking the north direction of the cutter head as the 0 phase starting point, and combining the fitted frequency and initial phase parameters, the mechanical phase angle of the sensor relative to the north direction of the cutter head at each moment is determined through the correlation between time and phase, ensuring that the value range of the phase angle is strictly controlled between 0 and 2π. The definition and calculation logic of the phase angle are as follows: Figure 4 The phase clipping is illustrated in the diagram. Figure 4The diagram clearly shows the positional relationships between the cutter head, sensor, and cutter head center. Angle markings of 0°, 90°, 180°, and 270° are marked on the cutter head, with arrows indicating the direction of rotation. During the sensor's rotation with the cutter head, the gravitational acceleration vector in the world coordinate system remains constant, while its projection component on the sensor's tangential sensitive axis (i.e., the tangential gravity component) varies sinusoidally with the rotation angle. The phase angle θ represents the angle traversed by the sensor relative to the cutter head's true north direction (0° position) in a clockwise direction, visually demonstrating the change in the sensor's phase angle as the cutter head rotates. Subsequently, starting from phase 0, vibration signal segments corresponding to phase intervals of length 2π are cyclically extracted. These signal segments correspond precisely to the vibration response of one rotation of the cutter head, representing the vibration sample for one rotation. The generation process of the angle-domain sampling sample is as follows... Figure 5 As shown, Figure 5 The figure shows the distribution of eight samples, from Sample 1 to Sample 8. The horizontal axis represents the phase angle θ in radians, and the vertical axis represents the axial, radial, and tangential acceleration values. It can be clearly seen from the figure that the phase angle coverage of Sample 1 and Sample 8 is less than 2π, which cannot fully represent the face information of the cutting head rotating one revolution. Therefore, they are considered invalid samples and deleted (invalid samples are the first incomplete data segment at the beginning of data processing or the last incomplete data segment at the end; invalid samples are deleted to wait for the next revolution of data). The phase angle coverage of Samples 2 to Sample 7 is all complete and reaches 2π, so they are valid samples. This ensures that all valid samples can fully reflect the spatial scanning range of the face. The marking information of invalid samples and the truncation parameters of valid samples will also be recorded in the database simultaneously.

[0031] The obtained single-cycle vibration samples are transformed from the time domain to the angle domain, and a mapping relationship between vibration acceleration and cutter head rotation angle is established. This mapping relationship is based on the linear correlation logic between time and angle. According to the cutter head rotation cycle, a one-to-one positive mapping and inverse mapping relationship is formed between the time domain acquisition time and the cutter head rotation angle, so that each time domain moment can accurately correspond to the specific rotation angle of the cutter head, and vice versa. The vibration samples were then resampled using interpolation methods. First, the classic cubic spline interpolation algorithm was used to initially calculate the vibration acceleration value sequence at a fixed number of uniformly distributed angle points within the range of 0 to 2π radians. Cubic spline interpolation has good smoothness and can ensure the continuity of the overall signal. At the same time, the angle sub-intervals corresponding to the drastic changes in signal amplitude caused by the cutter impact in the initially calculated sequence were identified. These sub-intervals contain key impact information reflecting the characteristics of the rock mass. For these sub-intervals, a piecewise cubic Hermite interpolation polynomial algorithm was used for local recalculation. This algorithm can better preserve the steep waveform and amplitude characteristics of the impact signal and avoid the blurring of key features caused by overall interpolation. Finally, standardized angle domain samples with fixed length and uniform angle intervals were obtained, completely eliminating the influence of cutterhead speed changes on the samples, so that all samples have consistent spatial semantics and a unified input dimension. The standardized samples will be bound with information such as acquisition time and tunneling mileage and stored in the database to form reusable sample resources.

[0032] Standardized angle domain samples are input into a pre-trained one-dimensional convolutional neural network (1DCNN) model for surrounding rock grade identification. The framework of the one-dimensional convolutional neural network model is as follows: Figure 6 As shown in Table 1, the parameter settings are as follows; Figure 6The model clearly demonstrates the hierarchical structure from sample input to surrounding rock grade output, in the following order: sample input, convolutional layer 1 (first convolutional layer), max-pooling layer 1 (first max-pooling layer), convolutional layer 2 (second convolutional layer), max-pooling layer 2 (second max-pooling layer), flattening layer, fully connected layer 1 (first fully connected layer), and fully connected layer 2 (second fully connected layer). These layers are sequentially connected to form a complete feature extraction and classification process. The initial training of this model relies on a historical standardized angle domain sample training set stored in the database. Before training, data augmentation processing is performed on the training set. Based on the geomechanical parameters associated with the surrounding rock grade of the sample (retrieved from the database), the sample signal undergoes amplitude modulation related to rock strength and phase modulation related to rock mass structure. Noise consistent with the spectral characteristics of rock mass fracture is then superimposed to generate augmented samples, expanding the diversity and coverage of the training data and improving the model's generalization ability. The one-dimensional convolutional neural network model processes samples according to a fixed procedure. First, preliminary feature extraction is performed through the first convolutional layer, which has 64 channels, a kernel size of 7, a stride of 1, and uses the Tanh activation function. Then, it is input into the first max pooling layer for downsampling, with a kernel size of 10 and a stride of 10. Next, the output of the first max pooling layer is input into the second convolutional layer for further extraction of deep features. The second convolutional layer has 32 channels, a kernel size of 7, a stride of 1, and uses the Tanh activation function. It is then downsampled again through the second max pooling layer, with a kernel size of 10 and a stride of 10. After that, the output of the second max pooling layer is input into a flattening layer to convert it into a one-dimensional feature vector. Finally, the one-dimensional feature vector is sequentially input into the fully connected layers of the classification module. The first fully connected layer has 256 neurons and uses the ReLU activation function, while the second fully connected layer has 4 neurons and uses the Softmax activation function. Through the decision calculation of the fully connected layers, the rock grade corresponding to the current working face is finally output.

[0033] surface 1DCNN parameter settings Activation response during model sample processing, such as Figure 7 As shown, Figure 7This diagram fully illustrates the activation states of vibration signal samples corresponding to Class V rock masses after angular domain downsampling during the network's forward propagation. The left side of the diagram shows the input downsampled samples, the middle column sequentially presents the activation maps of Convolutional Layer 1 (first convolutional layer), Max Pooling Layer 1 (first max pooling layer), Convolutional Layer 2 (second convolutional layer), and Max Pooling Layer 2 (second max pooling layer), and the right side shows the final output rock mass classification, including Class II, III, IV, and V. It is clearly observed from the diagram that in the shallow layers of the network, Convolutional Layer 1 (first convolutional layer) and Max Pooling Layer 1 (first max pooling layer), due to the high spatial resolution of the feature maps, the activation responses are relatively uniformly distributed along the spatial axis, and no obvious class-specific patterns have yet been observed. This is because the shallow network mainly captures local waveform details of the signal and has not yet effectively aggregated semantic features. As the network depth increases, after entering convolutional layer 2 (second convolutional layer) and max pooling layer 2 (second max pooling layer), the spatial dimension of the feature map is gradually compressed, but the semantic information is significantly enhanced. Especially in the area marked by the red dashed line in max pooling layer 2, a significant fluctuation in activation intensity can be clearly seen, and the angle range corresponding to this fluctuation is highly consistent with the impact vibration period of the tool impacting the rock mass in the original vibration signal. This phenomenon fully verifies that the model can effectively capture and compress the key impact features in the encoded signal. These impact features are the core basis for distinguishing different surrounding rock grades, intuitively demonstrating the superiority of the one-dimensional convolutional neural network model in feature extraction.

[0034] After identification, the confidence level of the surrounding rock identification results will be evaluated and optimized. A confidence threshold will be set. When the confidence score is lower than the threshold, the current identification result is considered to have insufficient confidence. At this time, the statistical features of historical high-confidence samples will be retrieved from the database and matched with the statistical features of the current standardized angle domain samples to provide auxiliary reference for on-site personnel. After the final surrounding rock grade is confirmed by on-site geological verification, the confirmation result will be bound with the corresponding standardized angle domain samples to form a new training data pair. This pair will be periodically added to the training dataset in the database to trigger incremental learning and parameter fine-tuning of the one-dimensional convolutional neural network model. This allows the model to continuously absorb new engineering experience and geological cases, adaptively optimize performance, and gradually adapt to changes in different engineering projects and geological environments, thereby achieving continuous improvement in identification accuracy.

[0035] Compared with existing technologies, which mostly rely on manual feature extraction or recurrent neural network modeling, manual feature extraction is prone to disrupting the temporal structure of signals, and recurrent neural networks are prone to gradient vanishing or exploding problems when processing long sequence signals. Furthermore, they lack systematic signal standardization methods and data support systems, resulting in semantic ambiguity in the sample space and difficulty in guaranteeing recognition accuracy and real-time performance. This implementation method introduces a database to build a complete data support system, realizing effective linkage between historical and real-time data. At the same time, it achieves standardized processing of TBM cutterhead vibration signals through angle domain downsampling, giving each sample a clear and unified spatial semantics, effectively eliminating the influence of rotational speed changes. The one-dimensional convolutional neural network model can not only automatically extract key discriminative features from the signal, avoiding information loss from manual feature extraction, but also overcome the technical defects of recurrent neural networks. Combined with confidence assessment and feedback optimization mechanisms, a closed loop of "data storage - real-time processing - intelligent recognition - model optimization" is formed, realizing real-time and accurate identification of surrounding rock grades, and providing reliable technical support for TBM tunneling parameter optimization and construction safety control.

[0036] In another embodiment of the present invention, step two, "applying mean filtering to the tangential vibration component to extract its periodic gravity component, and fitting the gravity component using a sine function model to obtain the periodic information of the cutter head rotation," is further refined as follows: The core objective of periodic extraction is to stably separate and fit the periodic gravity component that is strictly synchronized with the rotation of the cutterhead from the tangential vibration signal disturbed by complex rock impacts, thereby accurately obtaining the rotation frequency (period) and initial phase. This process mainly relies on two key calculations: mean filtering and sinusoidal function model fitting.

[0037] First, mean filtering is used to perform preliminary smoothing of the tangential vibration component. The mean filtering algorithm originates from linear smoothing techniques in the field of digital signal processing. Its essence is to apply a finite-length sliding window to the discrete signal sequence, replacing the original value at the center point of the window with the arithmetic mean of all sampling points within the window, thereby filtering out high-frequency fluctuations and abnormal spikes while retaining the low-frequency trend components of the signal. In this application, this classic algorithm is applied to the tangential vibration signal sequence of the cutterhead. Let the original tangential vibration component be a discrete sequence, where the i-th data point is denoted as x. i For a sliding window of length n (where n is an integer estimated based on the fundamental frequency of the cutter head, such as a fraction of the rotation period), the data value corresponding to the center position of the filtered window is calculated by formula (1): (Formula 1, i ranges from 1 to n) This formula is a direct application of the standard mathematical definition of the arithmetic mean in this scenario. Its physical meaning is that by summing and averaging the n data points (all in acceleration g) within the window, the resulting value is still in g, maintaining consistent dimensions. This operation is equivalent to a low-pass filter, effectively suppressing high-frequency, large-amplitude fluctuations caused by rock impacts, thus highlighting the implicit, relatively slowly changing, and periodic gravitational acceleration component curve. This curve represents the preliminary estimated value of the periodic gravitational component.

[0038] Subsequently, the gravitational component sequence obtained by the above filtering is fitted with parameters using a sine function model. The sine function is a standard mathematical model in physics that describes uniform circular motion or simple harmonic motion. When the accelerometer rotates approximately uniformly with the cutter head, the projection component of the Earth's gravity on the tangential sensitive axis of the sensor exhibits a sinusoidal change. Therefore, this application uses time t (in seconds) as the independent variable and the estimated value of the filtered gravitational component at time t as the dependent variable to construct the fitting model as shown in formula (2): g(t) = A × sin(2πft + φ) + B (Formula 2) Where g(t) represents the fitted value of the gravity component at time t, in g; A represents the amplitude of the sine curve, in g, reflecting the maximum change in the gravity projection; f represents the frequency of the cutter head rotation, in Hz, a key parameter directly determining the rotation period T (T=1 / f, in s); φ represents the initial phase, in rad, indicating the initial rotation angle position of the cutter head at t=0; B represents the offset, in g, used to compensate for possible zero-point drift of the sensor or the DC component caused by the installation tilt angle. Both sides of the formula represent acceleration, with the same unit of g.

[0039] The fitting process involves finding an optimal set of parameters [A, f, φ, B] such that the overall error between the fitted curve g(t) calculated by formula (2) and the estimated sequence of actual gravity components obtained by formula (1) is minimized. This process is usually implemented using optimization algorithms such as the least squares method. In specific implementation, the filtered data sequence (timestamp t) is used to... i The algorithm takes the corresponding acceleration value as input and inputs it into the fitting algorithm. Starting with a set of initial parameter guesses, the algorithm iteratively adjusts the parameters and finally outputs the optimal parameter estimates that minimize the mean square error. Thus, we can robustly extract the frequency f of the cutterhead rotation (and consequently the period T = 1 / f) and the initial phase φ from the mixed vibration signal. These two parameters are the absolute reference for subsequently calculating the cutterhead phase angle at any given time and achieving accurate mapping from the time domain signal to the angle domain.

[0040] Formula (1) is the preprocessing step, and its output (the filtered gravity component sequence) is the input data for the fitting process of Formula (2). The fitting result of Formula (2) (parameters f and φ) is the periodic information that needs to be obtained in the end, and provides direct parameters for the phase angle calculation formula. The whole process constitutes a complete and operable derivation chain from the original signal to the periodic parameters: original tangential signal → (mean filtering through Formula 1) → periodic gravity component sequence → (least square fitting as input to Formula 2) → obtain the rotation frequency f and the initial phase φ → calculate the rotation period T = 1 / f.

[0041] In another embodiment of the present invention, in step three, the phase angle θ corresponding to each moment of the preprocessed signal is calculated using equation (3). m (t). The fundamental purpose of calculating the phase angle at each sampling moment is to provide a precise time-domain partitioning basis for the subsequent "extraction of vibration samples after one revolution of the cutter head". This calculation directly relies on the two core parameters, the cutter head rotation frequency f and the initial phase φ, obtained through sine function fitting.

[0042] Phase angle θ m (t) Origin and derivation of the calculation formula: The sinusoidal function model g(t) = A × sin(2πft + φ) + B has its core oscillation as sin(2πft + φ). Here, the expression (2πft + φ) is referred to as the "instantaneous phase" or "total phase" of the signal in physics and signal processing. It is a continuous quantity that grows linearly with time t, and its unit is radians (rad). Its linear growth rate is determined by the rotational angular frequency 2πf.

[0043] However, to characterize the spatial position of the cutter head on the circumference (i.e., the "mechanical phase angle"), what is needed is not an infinitely growing absolute phase, but rather its mapping to a fixed circumferential range, between 0 and 2π radians, which perfectly aligns with the geometric definition of a circle as 360° (2π rad). Therefore, a "modulo 2π" operation needs to be performed on the aforementioned continuously growing total phase to obtain its relative position within each rotation cycle.

[0044] Formula (3) is precisely the realization of this mathematical idea: θ m (t)=(2πft+φ)-k×2π The direct source of this formula is the phase normalization principle in trigonometry. Its derivation is as follows: 1. Calculate the continuous total phase: First, based on the fitted f (Hz) and φ (rad), calculate the continuous total phase at time t (s): Φt otal (t) = 2πft + φ (unit: rad).

[0045] 2. Period normalization: To obtain the mechanical phase angle within the range [0, 2π), an integer number of complete 2π periods needs to be subtracted from the continuous total phase. The integer k is determined by the following formula: k = floor(Φ total (t) / (2π))=floor((2πft+φ) / (2π)).

[0046] The `floor()` function used here is the standard floor function in mathematics, which returns the largest integer not greater than the parameter. This operation ensures the "wrap-up" effect of the phase and is a common basic operation for unwrapping digital phases.

[0047] 3. Obtain the normalized phase angle: Substitute the calculated k into formula (3) to obtain the normalized mechanical phase angle at time t: θ m (t)=Φ total (t)-k×2π. This result strictly satisfies 0≤θ. m (t)<2π.

[0048] Physical dimensions and logical relationships: In formula (3), the 2πft term: the unit of frequency f is Hz (i.e., 1 / s), and the unit of time t is s. Therefore, the product ft is dimensionless. Multiplying by 2π (rad), the unit of the entire term is rad. The unit of the initial phase φ is also rad. k is a pure integer. The unit of the whole right side of the equation is rad, and the phase angle θ on the left side is the same. m The units (t) are completely consistent in rad, which conforms to the principle of dimensional consistency.

[0049] This formula and formula (2) form a direct parameter transfer and calculation extension relationship. The output parameters (f, φ) of formula (2) are directly used as the input parameters of formula (3). At the same time, the phase angle θ calculated by formula (3) m (t) is the core criterion for performing the operation of "taking the 0 phase as the starting point and cyclically extracting a phase interval of length 2π" in step three.

[0050] Workable implementation process: In practice, the operation process is as follows: 1. Obtain fitting parameters: From the above fitting results, read the accurately obtained cutter head rotation frequency f (unit: Hz) and initial phase φ (unit: rad).

[0051] 2. Traverse sampling times: For each sampling time point t (in seconds) in the preprocessed vibration signal, perform the following calculations.

[0052] 3. Calculate the continuous total phase: Φ total =2π×f×t+φ.

[0053] 4. Calculate the periodic integer k: k = floor(Φ) total / (2π)). This step can be achieved through programming or integer division (quotient) operations in digital circuits.

[0054] 5. Calculate and output the mechanical phase angle: θ m (t)=Φt otal –k×2π.

[0055] 6. Generate phase angle sequence: Repeat steps 2-5 for all sampling points of the signal to finally generate a mechanical phase angle sequence that strictly corresponds one-to-one with the timing of the original vibration signal and has a value range of [0, 2π) rad.

[0056] Through the above calculations, the system can assign a precise angular label to each data point in the vibration signal, reflecting the spatial orientation of the cutterhead corresponding to its timing. It is based on this continuous phase angle sequence that the algorithm can reliably locate θ. m The moment (t) = 0 rad (i.e., due north of the cutterhead) is used as the starting point to accurately extract a complete single-cycle vibration data segment with a phase span of exactly 2π rad, thus laying a precise spatial reference for subsequent angle domain conversion.

[0057] In another embodiment of the present invention, the establishment of the mapping relationship between vibration acceleration and cutter head rotation angle and the resampling in step four are explained in detail. In this invention, the core purpose of angle domain conversion and standardization is to eliminate the influence of cutter head rotation speed changes on data samples, and to convert "single-cycle time-domain vibration signals" of varying durations into "standardized angle domain samples" with fixed lengths and each data point strictly corresponding to a fixed angle position on the face of the tunnel. This provides the neural network model with input that has clear spatial semantics and unified dimensions. Specifically: 1. Establishing the mapping relationship between the time domain and the angle domain The mathematical essence of this mapping relationship is to establish a linear transformation from time coordinates to angular coordinates. Its physical basis is: assuming that the cutter head undergoes approximately uniform circular motion within a single rotation, then its rotation angle is proportional to time.

[0058] Forward mapping function (Formula 4): θ n (t)=[(t-t n,start ) / T n ]×2π. This formula is the formula for the angle of uniform circular motion (θ=ωt, where the angular velocity ω=2π / T). n This is a direct application of the concept. It defines the rotation angle θ of the cutterhead at any time-domain sampling time t (in seconds) during the nth rotation. n (t) (unit: rad). Where t n,start (Unit: s) is the calculated phase angle θ within that circle.m (t) The initial time when T first equals 0 rad; n (Unit: seconds) is the rotation period (T) obtained by converting the fitted frequency f. n =1 / f`). In the formula (t-t) n,start ) / T n The calculation calculates the proportion of the rotation period completed since the start of the circle (dimensionless), which, when multiplied by 2π rad, is converted into the angle rotated from the starting point. The entire formula has consistent dimensions: time divided by time yields a dimensionless number, which is then multiplied by radians to obtain the final radian value.

[0059] Inverse mapping function (Formula 5): t n (θ)=(θ×T n ) / 2π+t n,start This formula is the mathematical inverse of the forward mapping function (Formula 4). It solves the inverse problem of "to obtain vibration data when rotated to a specific angle θ, it is necessary to find the corresponding sampling time". This formula can be derived by performing an algebraic transformation on Formula 4: From θ=[(t-t start From ) / T]×2π, we can deduce t-t start =(θ×T) / 2π, thus obtaining t=(θ×T) / 2π+t start In terms of dimensions, θ (rad) multiplied by T (s) and then divided by 2π (rad) results in seconds. Adding the initial time (s) gives the final time (s), which has the same dimensions.

[0060] 2. Angle domain signal generation based on mapping relationship Once the inverse mapping function is obtained, a discrete signal sequence in the angle domain can be generated.

[0061] Calculation of angular domain vibration value (Formula 6): a n (θ)=S n (t n (θ)). This formula represents a composition operation of a function. S n Let be the preprocessed time-domain vibration signal vector (in units g) of the nth cycle, which is a function with time as the independent variable. To obtain the vibration acceleration value an(θ) (in units g) at angle θ, firstly, through the inverse mapping function t n (θ) Calculate the precise time in the time domain corresponding to the angle θ, and then substitute this time into the time domain signal function S. n In this process, the signal amplitude at that moment is calculated through interpolation. This process ensures the accuracy of the angle domain data point a. n (θ) physically corresponds to the vibration response at a specific orientation (angle θ) of the tunnel face.

[0062] 3. Implementation of Standardized Resampling To obtain samples with fixed lengths and uniform angular intervals, resampling is required.

[0063] Operation flow (corresponding to step S43): First, a fixed value M (e.g., 3600) is preset, representing the number of points to be sampled within a full 2π radian radius. Then, within the radian range of [0, 2π), M target angle points are uniformly generated: θ1=0, θ2=Δθ, θ3=2Δθ,…,θ M =(M-1)Δθ, where Δθ=2π / M.

[0064] Interpolation calculation: For each target angle point θ j Perform the following operations: 1. Based on the inverse mapping function t n (θ), calculate θ j The corresponding time domain time t j =t n (θ j ).

[0065] 2. Due to t j It is very likely that the sampling time will not coincide with the original time-domain sampling time, and it is necessary to obtain it from the original time-domain signal sequence S. n Estimate t j The signal value at time t. This is achieved through an interpolation algorithm. This implementation uses an "interpolation method," and in practice, classic numerical analysis algorithms such as cubic spline interpolation are typically chosen. This algorithm utilizes t j Construct a smooth spline curve from multiple nearby original data points, and calculate the curve at time t. j The function value at point a is used as a n (θ j ).

[0066] Generate standardized samples: After completing the above calculations for all M target angle points, a vibration acceleration value sequence of length M is obtained: {a n (θ1), a n (θ2), ..., a n (θ M This sequence is a standardized angle domain sample with a fixed length and uniform angular intervals. Regardless of the actual rotational speed of the cutter head (period T)... n (size), all samples are uniformly set to the same angular resolution (Δθ) and sample length (M), and the j-th point in the sequence always represents the vibration state when the cutter head rotates to the (j-1)Δθ radian position.

[0067] This implementation method constitutes a crucial data processing stage that connects the preceding and following steps, with a reception period T. n and the start time t of a single lap n,startAs input, by establishing a rigorous time-angle mapping (Equations 4 and 5) and function composition calculation (Equation 6), and using standard numerical interpolation algorithms, the standardized angle domain samples required by the neural network model are finally output. This process completely shifts the signal analysis benchmark from a variable time axis to a fixed spatial angle axis, which is the core technical step of this method to achieve "spatial semantic unification".

[0068] In another embodiment of the present invention, the core task of the one-dimensional convolutional neural network model designed in this invention is to automatically learn and extract deep feature patterns that can effectively distinguish different surrounding rock grades from the generated standardized vibration signal sequences with clear spatial angular coordinates. This model abandons the recurrent neural network architecture that relies on manual feature engineering or is difficult to handle long sequences, and instead adopts a hierarchical feature extraction and decision pipeline composed of convolutional layers, pooling layers, and fully connected layers.

[0069] Model hierarchical structure and processing flow First, standardized angle-domain samples are used as the input model of a 3D tensor, whose dimension is represented by the number of channels multiplied by the sequence length. The first stage of the model involves preliminary feature extraction and compression. The samples are then fed into the first convolutional layer, which has sixty-four independent convolutional kernels, each with a width of seven data points, sliding along the angle sequence direction with a stride of one. Each convolutional kernel performs an operation derived from cross-correlation in digital signal processing, acting as an adaptive filter to scan the input sequence and extract waveform patterns within a local range (covering approximately 0.06 radians), such as specific rising edges or oscillating segments. This layer uses the hyperbolic tangent function as the activation function, a common non-linear activation function in neural networks that maps the convolution calculation result to between -1 and 1, helping to stabilize gradient flow. The output of the first convolutional layer is then fed into the first max-pooling layer for downsampling. This pooling layer uses a pooling window with a width of ten and a stride of ten, a direct application of the max-pooling algorithm commonly used in computer vision and signal processing, which selects the maximum value from ten consecutive feature values ​​within each window for output. This operation compresses the length of the feature sequence to one-tenth, achieving data dimensionality reduction and enhancing the invariance of features to small angular shifts, while highlighting the most salient local activity features.

[0070] Next, the model enters the second stage for deeper feature abstraction. The output of the first max-pooling layer is fed into the second convolutional layer. This layer contains thirty-two convolutional kernels, with a kernel width of seven and a stride of one. Building upon the primary features extracted in the previous layer, this layer further combines these features to form more complex and semantically rich high-level patterns, such as identifying impact feature combinations caused by specific rock mass structures. This layer also uses the hyperbolic tangent activation function. Subsequently, the second max-pooling layer further compresses the feature map with a window of five width and a stride of five, reducing the sequence length to one-fifth again, further aggregating and strengthening the most discriminative feature information.

[0071] After two stages of convolution and pooling, the resulting feature map still has a two-dimensional structure. To adapt to the subsequent fully connected classification layer, it needs to be converted into a one-dimensional feature vector. This function is achieved by a flattening layer, which does not perform any computation but simply rearranges all elements of the multi-dimensional feature map into a long vector in sequence, thus completing the data structure transformation.

[0072] Finally, the classification module makes a decision on the one-dimensional feature vector. This module consists of two fully connected layers. The first fully connected layer contains 256 neurons, whose input is the flattened feature vector. This layer performs a linear transformation and nonlinear feature integration in a high-dimensional space, aiming to fuse all features learned from the global context of the vibration signal. A linear rectified unit is chosen as the activation function, which effectively introduces nonlinearity and alleviates the gradient vanishing problem, prompting the network to learn sparser and more efficient feature representations. The second fully connected layer has four neurons, directly corresponding to the number of rock grade classifications. The output of this layer is a four-dimensional vector, with each dimension corresponding to an unnormalized score for a rock grade. To transform these scores into an interpretable probability distribution, this layer uses a flexible maximum function as the activation function. This function is the standard normalization function in multi-class classification tasks, transforming the four input values ​​into a probability vector, where each component's value is between zero and one, and the sum of all components is one. The component with the highest probability is determined as the final rock grade output by the model.

[0073] The design logic of the entire model embodies a hierarchical feature learning approach, moving from local to global and from details to semantics. Shallow convolutional layers capture local waveform details in the signal, while deeper structures combine these details to form abstract patterns related to geological significance. Max pooling layers reduce computational complexity while improving the model's robustness. Fully connected layers ultimately synthesize global information to complete the classification decision. This one-dimensional convolutional neural network model directly receives standardized angle domain samples with fixed spatial semantics from the output of claim 4 as input. Its architecture is specifically optimized for such one-dimensional sequence signals, avoiding the gradient problem of recurrent neural networks when processing long sequences, and achieving end-to-end efficient and robust identification of surrounding rock information hidden in cutterhead vibration signals.

[0074] In another embodiment of the present invention, under complex geological and tunneling conditions, periodic extraction based solely on vibration signals may lead to inaccurate results due to strong, non-periodic rock impact interference. The enhancement mechanism introduced in this embodiment aims to add a quality monitoring and correction step to the core periodic extraction process, thereby improving the robustness and reliability of the entire system under harsh signal environments. Specifically: 1. Acquisition of preliminary periodic gravitational components First, the tangential component of the cutterhead vibration is processed using a mean filtering algorithm. This algorithm is a direct application of linear smoothing techniques in digital signal processing. By calculating local average values ​​through a sliding window, high-frequency impact noise is effectively suppressed, resulting in a relatively smooth curve that reflects the periodic changes in gravitational acceleration. This curve serves as a preliminary estimate of the gravitational component caused by the cutterhead rotation.

[0075] 2. Reliability assessment based on Hilbert transform After obtaining the preliminary estimate of the gravity component, this step introduces the Hilbert transform for signal quality analysis to determine its reliability as a periodic reference. The Hilbert transform is a classic algorithm in the fields of communication and signal processing, used to construct an analytic signal representation of a real signal.

[0076] Instantaneous amplitude and phase calculation: Applying a Hilbert transform to the preliminary gravity component estimation signal yields its corresponding analytic signal. Based on this analytic signal, the instantaneous amplitude and instantaneous phase of the original signal can be directly derived. The instantaneous amplitude describes the change of the signal envelope over time, and the instantaneous phase describes the change of the signal phase over time.

[0077] Evaluation Logic: For an ideal sinusoidal signal of gravity component generated by uniform rotation, its instantaneous amplitude should remain constant (stable), and its instantaneous phase should increase strictly linearly with time (good linearity). However, residual impact interference or small fluctuations in the rotational speed of the cutter head in actual signals can disrupt this ideal characteristic. Therefore, this invention evaluates its stability by calculating the variance (or fluctuation range) of the instantaneous amplitude and evaluates its linearity by calculating the fitting error between the instantaneous phase sequence and the ideal linear phase sequence. By comparing these two evaluation indicators with preset thresholds, the reliability of the currently initially extracted periodic gravity component can be quantitatively scored. If the evaluation result (such as the comprehensive score) is higher than the preset threshold, the signal quality is considered reliable, and it can directly proceed to the subsequent sinusoidal function model fitting step.

[0078] 3. Multi-source signal fusion and phase correction When the reliability assessment result is lower than the preset threshold, it indicates that the current vibration signal is subjected to too much interference, and it is difficult to guarantee the accuracy of period extraction by relying solely on the vibration signal itself. At this time, the system activates the multi-source signal correction mechanism.

[0079] Calibration reference signal source: The system calls the encoder pulse signal from the cutter head drive motor. This signal is a hardware digital signal that directly measures the angular displacement of the rotating shaft. Each pulse corresponds to a fixed mechanical rotation angle (e.g., 0.1 degrees), and it has extremely high timing accuracy and anti-vibration interference capability.

[0080] Phase resynchronization and correction: Using a high-precision encoder pulse sequence as the absolute time reference, the preliminary gravity component estimate obtained from the vibration signal is compared. Specifically, the pulse moments indicating "zero position" or specific angular positions in the encoder signal are identified and aligned with the phase information of the gravity component estimate from the vibration signal. The initial phase φ of the preliminary gravity component estimate is corrected by calculating the phase difference between the two on the time axis. Simultaneously, the instantaneous rotational speed directly calculated from the encoder signal can be used to verify or smooth the frequency f obtained by fitting the vibration signal. The signal obtained after this correction process is considered the final, reliable periodic gravity component, which is then used for accurate fitting of a sine function model.

[0081] This embodiment is one of the key designs that enables the present invention to operate stably in real and complex engineering environments, ensuring the accuracy of the most basic "time reference" information at the very beginning of the entire chain from the original sensor signal to the final identification of the surrounding rock grade.

[0082] In another embodiment of the present invention, the goal of angle domain resampling is not only to generate a uniform series of angle points, but more importantly, to preserve, to the greatest extent possible, the high-frequency impact characteristics of the original vibration signal generated by the intense interaction between the tool and the rock mass, which have important geological indicative significance, during the standardization process. Conventional single global interpolation methods (such as global cubic spline interpolation) often smooth or distort these local, steep amplitude abrupt changes when pursuing overall curve smoothness, leading to the loss of key information. The hybrid interpolation scheme proposed in this embodiment aims to solve this contradiction and achieve high-fidelity signal reconstruction with "overall smoothness and local shape preservation".

[0083] 1. Global preliminary resampling based on cubic spline interpolation First, for the vibration signal with an established time-angle mapping relationship in the nth cycle, in order to obtain the time-angle mapping relationship at M uniform target angle points {θ1, θ2, ..., θ...} M The acceleration values ​​on the curve are initially obtained using cubic spline interpolation. Cubic spline interpolation is a classic algorithm in numerical analysis used for function approximation and data interpolation. Its core idea is to construct a cubic polynomial curve between adjacent original data points (angle-acceleration pairs), and to force that not only the function values ​​at all connection points (nodes) are continuous, but also the first and second derivatives are continuous, thus ensuring the smoothness (second-order continuous differentiability) of the entire fitted curve.

[0084] Operation process: The original image, at a non-uniform angle θ... orig The corresponding vibration acceleration value a n (θ orig ( ) are known data points. Using these data points as nodes, construct a cubic spline interpolation function S that passes through all nodes. spline (θ). Then, apply this function to the uniform target angle point θ. j By evaluating at (j=1 to M), we can obtain the preliminary resampled sequence {a}. npre (θ1), a npre (θ2), ..., a npre (θM)}. This step ensures a smooth transition of the signal across the global angular range, effectively reconstructing the low- and mid-frequency components of the signal.

[0085] 2. Identification of impact feature angle sub-intervals After initial resampling, it is necessary to identify local intervals that may contain key impact characteristics. The impact of a cutting tool on rock mass typically manifests in vibration signals as drastic changes (sharp rises or falls) in acceleration amplitude within a very short angular span.

[0086] Recognition logic: For the sequence a obtained from the initial calculation... npre(θ) is analyzed, and its approximate first-order difference along the angle axis is calculated (or other edge detection algorithms are used), to find angle points whose absolute difference value exceeds a preset threshold. These points are used as boundaries to locate one or more angle sub-intervals [θ]. start θ end Within these sub-intervals, the amplitude of the original signal changes at a high rate, and the preliminary global spline interpolation results may be overly smoothed and fail to accurately reflect the true steepness of the impact waveform. The identification algorithm essentially detects the high-frequency components or abrupt changes in the signal.

[0087] 3. Local feature reconstruction based on piecewise cubic Hermitian interpolation For each identified impact feature sub-interval, the initial global spline interpolation results within that interval are discarded, and instead, piecewise cubic Hermitian interpolation polynomials are used for local recalculation.

[0088] Algorithm Principle: Piecewise cubic Hermite interpolation is another classic conformal interpolation algorithm. Unlike cubic splines, which require continuous second derivatives, Hermite interpolation independently constructs a cubic polynomial for each subinterval. This polynomial not only passes through the function values ​​(i.e., acceleration values) at the two endpoints of the interval but also through the derivative values ​​specified at the two endpoints (i.e., the first derivative of acceleration with respect to angle, which reflects the trend of change). This allows it to fit the data more flexibly within local intervals, making it particularly suitable for approximating functions with discontinuous or rapidly changing first derivatives.

[0089] Specific implementation: Obtaining local raw data: for the impact sub-interval [θ] start θ end From the original, high-resolution time-angle mapping data, extract all original data points (θ) falling within this angular range. origk a n (θ origk )).

[0090] Estimate the boundary derivative: Using these local raw data, estimate the starting point θ of the subinterval through numerical differentiation methods (such as central difference). start and endpoint θ end The value of the first derivative at that point.

[0091] Construct and calculate: Using the function values ​​(acceleration) and derivative values ​​at the two endpoints of the interval as constraints, construct a unique cubic Hermite interpolation polynomial H on this subinterval. hermite (θ). Then recalculate the polynomial in [θ] using this polynomial. start θ end All uniform target angle points θ within the range j The acceleration value a on nhermite (θ j ).

[0092] Replacement operation: using a nhermite (θ j Replace the initial sequence a npre The values ​​of the points corresponding to the same angle in (θ).

[0093] 4. Generate optimized standardized samples After completing the local recalculation and replacement of all identified impact sub-intervals, a complete sequence of vibration acceleration values ​​of length M is finally obtained {a nfinal (θ1), a nfinal (θ2), ..., a nfinal (θ M This sequence is the defined, optimized, standardized angle domain sample.

[0094] The optimized interpolation strategy in this implementation combines "global smooth interpolation (cubic spline) to ensure overall continuity" with "local conformal interpolation (Hermitian) to preserve impact features," effectively solving the problem pointed out in the background technology that "conventional sampling methods easily destroy the waveform integrity of key impact events or introduce distortion." This ensures that the signal input to the one-dimensional convolutional neural network model has a standardized format while preserving high-frequency impact details reflecting the rock mass's fracture state and mechanical properties to the greatest extent. This improvement in signal fidelity directly enhances the neural network model's ability to learn discriminative features, which is a crucial guarantee for ultimately achieving high-precision surrounding rock grade identification. This step, together with basic mapping and model architecture, constitutes a complete and refined data preprocessing chain from the raw signal to the high-quality model input.

[0095] In another embodiment of the present invention, the key to training an intelligent recognition model with strong generalization ability lies not only in having a sufficient quantity of training data, but also in covering the reasonable variations that vibration signals may exhibit under real geological conditions. This embodiment innovatively introduces a data augmentation method based on physical mechanisms, building upon the conventional methods of collecting historical data and constructing a training set. This expands the data diversity, enabling the model to learn and adapt to the intrinsic relationship between geological parameters and vibration signals, rather than merely memorizing limited historical sample patterns. Specifically: 1. Construction of the basic training dataset Model training begins with a foundational sample library comprised of historical tunneling data. First, raw triaxial vibration data of the cutterhead corresponding to different rock grade sections recorded in past TBM projects, along with synchronously recorded cutterhead operating parameters and rock grade labels confirmed by on-site geological engineers, are collected. Then, following the identification methodology (including signal preprocessing, period extraction, single-cycle truncation, and angle domain standardization), these historical raw vibration data are processed into standardized angle domain samples. Each sample is a sequence of fixed length M, associated with a specific rock grade label. These processed "sample-label" pairs constitute the initial standardized angle domain sample training set. This step ensures that the training data is completely consistent with the data format the model will process during future online identification.

[0096] 2. Data augmentation processing based on geomechanical principles To enable the model to cope with unseen geological conditions, this step performs data augmentation on the aforementioned basic training set. Unlike traditional computer vision augmentations such as random cropping and rotation, the augmentation strategy of this invention is closely based on the principles of geomechanics, aiming to simulate the physical effects of different rock strengths and rock mass structures on vibration signals.

[0097] Amplitude modulation related to rock strength: The uniaxial compressive strength of rock is a key parameter affecting tool cutting force and impact vibration intensity. Higher rock strength typically leads to more severe tool vibration. Based on this physical law, amplitude modulation is applied to the training sample signal. Specifically, based on the typical rock strength range associated with the original surrounding rock grade label of the sample, a strength scaling factor α is randomly selected within this range according to a certain distribution law. s (For example, based on the typical uniaxial compressive strength range corresponding to the original surrounding rock grade of the sample, the strength scaling factor α is determined according to a preset mapping relationship.) s A baseline value is established, and random fluctuations are made within a reasonable range (e.g., ±20%) around this baseline value to simulate strength differences in similar rock masses. Then, each acceleration value in the entire normalized angular domain sample sequence is multiplied by α. s :a enhanced (θ)=α s ×a original (θ). This operation is derived from amplitude scaling in signal processing, which keeps the physical dimensions unchanged (g dimensionless factor = g), but simulates the change in the overall vibration energy level caused by subtle differences in intensity in the same type of rock mass.

[0098] Phase modulation related to rock mass structure: Joints, fissures, and other structural surfaces in the rock mass can cause abnormal vibration events due to local instability of the rock mass when the cutterhead rotates to a specific orientation. This manifests as additional impacts or disturbances near a specific phase angle in the signal. To simulate this effect, phase modulation is performed. First, one or more angle intervals Δφ are randomly selected within the range of [0, 2π) radians to represent the assumed dominant orientation of the structural surface. Then, using a mapping relationship, equivalent local time stretching or compression is performed in the time domain. However, a more direct method is to cyclically shift the original signal sequence in the angle domain or superimpose an attenuated transient waveform template simulating the structural surface response onto its specific phase interval. For example, within the selected Δφ interval, the original signal a can be... original (θ ) A time-angle mapped wavelet function is added to simulate the additional vibrational component caused by the presence of structural surfaces. This operation must ensure that the added perturbation conforms to geological logic and does not disrupt the overall continuity of the signal.

[0099] Superimposed noise matching the spectral characteristics of rock mass fracture: The vibration noise generated by the actual rock mass fracture process is not white noise; its energy has a specific distribution in the frequency domain. This step generates a random noise sequence with corresponding spectral characteristics based on the typical fracture characteristics of the target surrounding rock grade (e.g., hard rock is mainly characterized by high-frequency brittle fracture, while soft rock may contain more low-frequency plastic deformation). First, the characteristic frequency band corresponding to the target surrounding rock grade is determined by consulting rock mechanics literature or analyzing actual fracture signals. Then, colored noise is generated using filtering methods in digital signal processing: that is, a Gaussian white noise sequence is first generated, and then it is passed through a bandpass filter designed according to the target spectral characteristics (such as center frequency and bandwidth) to obtain a noise sequence n(θ) with spectral characteristics matched. Finally, this noise sequence is superimposed on the original angular domain sample at a certain signal-to-noise ratio: a enhanced (θ)=a original (θ)+β×n(θ), where β is a coefficient controlling the noise intensity. The unit of the superimposed noise remains g, maintaining the same dimensions.

[0100] 3. Model Training and Acquisition The enhanced samples generated after the above enhancement processing (multiple enhancements can be combined) are merged with the original base samples to form a larger and more diverse enhanced training set for final training. This enhanced training set is then used to train the one-dimensional convolutional neural network model. The training process adopts the standard supervised learning workflow in deep learning: using the enhanced samples as input and their corresponding rock grade labels as the supervision target, all weight parameters of the model are iteratively adjusted through backpropagation algorithm and optimizer (such as Adam) to minimize the cross-entropy loss function between the predicted output and the true label. After sufficient training and validation, the model parameters converge and stabilize, and the accuracy meets the requirements, thus obtaining the "pre-trained one-dimensional convolutional neural network model".

[0101] In this embodiment, by introducing directional data augmentation based on geomechanical principles, the physical interpretability and engineering generalization ability of the model training are greatly improved. This allows the model to not only see "what" the vibration patterns are during the learning process, but also indirectly learn the physical laws of how geological parameters (intensity, structure) "affect" the vibration signals through augmented data. This enables the finally trained model to make more reasonable inferences based on its learned physical correlations when faced with new combinations of geological conditions not covered by the training data, significantly improving the adaptability and reliability of the recognition system in complex and variable engineering environments. This training method is one of the core supporting components ensuring the successful application of the intelligent recognition method of this invention.

[0102] In another embodiment of the present invention, after the trained model is deployed in a real engineering environment, the model's single prediction result is uncertain due to complex and variable geological conditions that may exceed the distribution of the training data. This embodiment aims to add a closed loop of self-monitoring, decision support, and continuous learning to the entire intelligent recognition system, enabling it to evolve from a static prediction tool into an intelligent system capable of accumulating experience and adapting to new environments, thereby enhancing its engineering practicality and long-term reliability. Specifically: 1. Quantitative evaluation of confidence in real-time identification results During the online recognition phase, whenever a new standardized angle domain sample is input into the trained one-dimensional convolutional neural network model, the model's final fully connected layer (Softmax activation) outputs a probability vector, denoted as P=[p II p Ⅲ p IV p V This vector is the direct input for this step.

[0103] Confidence score calculation: This probability vector P is the model's confidence distribution regarding the current sample belonging to one of four predefined surrounding rock classes (e.g., Class II, III, IV, V). Maximum probability value p. max=max(P) intuitively reflects the model's confidence in its preferred class. Therefore, this invention will use p max This is directly defined as the confidence score of the surrounding rock identification results. This score is a scalar between 0 and 1, with a higher value indicating a more certain model prediction.

[0104] 2. Auxiliary processing and suggestions for low confidence results Set a predefined confidence threshold C th (For example, 0.7 or 0.8). This threshold serves as the criterion for triggering subsequent auxiliary processes.

[0105] Low confidence level determination and feature matching: After each identification, the system automatically compares the confidence score p. max With threshold C th If p max <C th If the current identification result is deemed insufficient in confidence and should not be relied upon entirely, the system initiates an auxiliary analysis process. First, it extracts a set of predefined statistical features from the standardized angular domain samples that are currently causing low confidence. These features include the root mean square value of the signal in the three axes, peak factor, kurtosis, and the energy proportion of a specific frequency band. These features are commonly used indicators in signal processing to characterize the signal amplitude distribution, impulse characteristics, and frequency content.

[0106] Similarity matching: Subsequently, the system accesses a historical high-confidence sample database. This database stores all past samples with confidence scores higher than C. th The system generates identification records, each containing its standardized angular domain sample, extracted statistical features, and the final rock class output by the model. The system performs similarity matching (e.g., calculating Euclidean distance or cosine similarity) between the statistical features of the current low-confidence sample and the features of all samples in the database. The matching results can be presented to field engineers, for example, "The current signal features are most similar to a high-confidence sample historically classified as 'Class IV rock'," providing valuable reference clues for their comprehensive manual judgment based on geological sketches, drilling parameters, etc.

[0107] 3. Incremental learning and model optimization based on on-site confirmation After referring to system prompts or through other means (such as geological logging), the on-site engineer will ultimately confirm the actual surrounding rock grade at the working face. This confirmation information is a valuable closed-loop feedback.

[0108] Construction of new training data pairs: The system uses the final confirmed rock grade on site as the gold standard label and binds it with the standardized angle domain sample that triggered this manual review to form a new, high-quality "sample-label" training data pair.

[0109] Data accumulation and incremental learning triggering: These new data pairs are not immediately used for training, but are stored in a temporary buffer or directly appended to the historical training dataset in the background. The system sets a periodic execution strategy (e.g., every N new data pairs accumulated, or every fixed tunneling mileage). When conditions are met, the incremental learning process of the model is triggered.

[0110] Incremental learning implementation: Incremental learning refers to fine-tuning the model parameters using new data on top of an existing pre-trained model, adapting it to new data distributions without forgetting existing knowledge. In practice, the currently deployed model is used as the initial state. The original training set is merged with the new data to form an expanded training set, and a new round of training is conducted with a smaller learning rate (or only some layers are trained). This process fine-tunes the model parameters, allowing it to better fit the vibration patterns corresponding to newly encountered geological scenes while maintaining its original recognition capabilities.

[0111] This invention also provides a TBM cutterhead vibration signal surrounding rock identification system based on angle domain downsampling. The system includes: a signal acquisition and preprocessing module for real-time acquisition and preprocessing of triaxial vibration acceleration signals at the cutterhead during TBM tunneling; a period extraction module for extracting tangential vibration components from the preprocessed signal and obtaining the period information of the cutterhead rotation through mean filtering and sine fitting; a phase truncation module for calculating the phase angle at each moment based on the period information and truncating vibration samples from one rotation of the cutterhead starting at phase 0; an angle domain standardization module for converting the vibration samples from the time domain to the angle domain and resampling them through interpolation into standardized angle domain samples with fixed lengths and uniform angle intervals; and a surrounding rock intelligent identification module for inputting the standardized angle domain samples into a pre-trained one-dimensional convolutional neural network model and outputting the surrounding rock grade corresponding to the current tunnel face.

[0112] The present invention also provides an electronic device. The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signals as described above. The electronic device may be an industrial computer, server, or dedicated embedded data processing device deployed in the TBM control room or field edge computing unit, connected to a vibration acceleration sensor at the cutterhead via a data acquisition interface, and achieving real-time, automatic identification of the surrounding rock grade by executing the method instructions.

[0113] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the aforementioned method for identifying surrounding rock based on angle-domain downsampling TBM cutterhead vibration signals. The computer-readable storage medium can be any tangible medium containing or storing the program, such as ROM, RAM, disk, optical disk, etc. The computer program exists in the form of instructions, which, when executed by a computing device (such as the aforementioned electronic device), cause the computing device to complete the entire process of signal processing, angle-domain conversion, feature extraction, and classification decision-making in the method, thereby transforming a general-purpose computing device into a specific machine dedicated to TBM surrounding rock identification.

[0114] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0115] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for identifying surrounding rock based on angle-domain downsampling TBM cutterhead vibration signals, characterized in that, Includes the following steps: Step 1: Real-time acquisition of triaxial vibration acceleration signals at the cutterhead during TBM tunneling, and preprocessing of the triaxial vibration acceleration signals to obtain preprocessed signals; Step 2: Extract the tangential vibration component from the preprocessed signal, perform mean filtering on the tangential vibration component to extract its periodic gravity component, and fit the gravity component using a sine function model to obtain the periodic information of the cutter head rotation. Step 3: Based on the periodic information of the cutter head rotation, calculate the phase angle corresponding to each moment of the preprocessed signal, and starting from the 0 phase, cyclically extract the vibration signal segments corresponding to the phase intervals of length 2π to obtain the vibration sample of one rotation of the cutter head. Step 4: Convert the vibration samples from the time domain to the angle domain and establish the mapping relationship between vibration acceleration and the rotation angle of the cutter head; resample the vibration samples into standardized angle domain samples with fixed length and uniform angle intervals through interpolation. Step 5: Input the standardized angle domain samples into the pre-trained one-dimensional convolutional neural network model and output the surrounding rock grade corresponding to the current working face.

2. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 1, characterized in that, Step two involves mean filtering of the tangential vibration component, achieved using equation (1). （1）； Where, represents the filtered data value. n For window length, x i Indicates the first in the window i One data point; The model of the sine function is shown in equation (2). g(t)=A×sin(2πft+φ)+B(2); Where g(t) is the gravitational component at time t; A is the amplitude in g; f is the frequency in Hz; t is the time in s; φ is the initial phase in rad; and B is the offset in g.

3. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 2, characterized in that, In step three, the phase angle θ of the preprocessed signal at each moment is calculated using equation (3). m (t); i m (t)=(2πft+φ)-k×2π(3) Where, θ m (t) represents the mechanical phase angle of the sensor relative to the true north direction of the cutter head at time t, in rad; f is the frequency, in Hz; t is the time, in s; φ is the initial phase, in rad; k is the value satisfying 0 ≤ θ m Integers whose t is less than 2π are obtained by k=floor((2πft+φ) / (2π)), where floor() represents the floor function.

4. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 1, characterized in that, Step four establishes the mapping relationship between vibration acceleration and the tool head rotation angle, and resampling is achieved through the following methods: S41. For the nth vibration signal, establish the mapping relationship between its time-domain acquisition time and the rotation angle of the cutter head, which is defined by the forward mapping function (4) and the inverse mapping function (5): θ n (t)=[(t-t n,start ) / T n ]×2π,t∈[t n,start ,t n,start +T n ](4); t n (θ)=(θ×T n ) / 2π+t n,start (5); Where, θ n (t) represents the rotation angle of the nth cutter head at time t, in rad, with a value range of [0, 2π). n,start T represents the starting time of the nth rotation of the cutter head, in seconds. n The nth rotation period of the cutter head, in seconds, is obtained by converting the cutter head rotation frequency f obtained from the fitting of Formula 2. The conversion relationship is T. n =1 / f; t n (θ) represents the time domain acquisition time corresponding to angle θ, in seconds; t represents the time domain acquisition time of the signal, in seconds; θ represents the rotation angle of the nth turn of the cutter head, in rad, with a value range of [0, 2π). S42. Based on the inverse mapping function, the vibration acceleration value a at angle θ is obtained through formula (6). n (θ); a n (θ)=S n (t n (i))(6); Among them, a n (θ) represents the vibration acceleration value of the cutter head at angle θ during the nth rotation, in g; S n is the preprocessed time-domain vibration signal vector of the nth cycle, in units of g; S43. Within a complete rotation cycle of 0 to 2π radians ([0, 2π)), a fixed number of angle points M are preset; based on the mapping relationship, the angle points {θ1, θ2, ..., θ3} are calculated using an interpolation method. M The corresponding vibration acceleration value sequence {a} on} n (θ1),a n (θ2),...,a n (θ M This sequence is a standardized angular domain sample with a fixed length and uniform angular spacing.

5. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 1, characterized in that, A one-dimensional convolutional neural network model performs the following processes in sequence: S51. Input the standardized angle domain samples into the first convolutional layer for feature extraction, and input the output of the first convolutional layer into the first max pooling layer for downsampling; the number of channels of the first convolutional layer is 64, the kernel size is 7, the stride is 1, the activation function is Tanh, and the kernel size of the first max pooling layer is 10, the stride is 10. S52. Input the output of the first max pooling layer into the second convolutional layer for feature extraction, and input the output of the second convolutional layer into the second max pooling layer for downsampling; the second convolutional layer has 32 channels, a kernel size of 7, a stride of 1, and an activation function of Tanh; the second max pooling layer has a kernel size of 10 and a stride of 10. S53. Input the output of the second max pooling layer into the flattening layer to convert it into a one-dimensional feature vector; S54. The one-dimensional feature vector is sequentially input into the fully connected layer of the classification module for decision-making. The classification module outputs the classification probability of the surrounding rock grade. The fully connected layer of the classification module includes a first fully connected layer and a second fully connected layer. The first fully connected layer has 256 neurons and the activation function is ReLU. The second fully connected layer has 4 neurons and the activation function is Softmax.

6. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 1, characterized in that, Step two involves mean filtering of the tangential vibration component to extract its periodic gravitational component, which is specifically achieved through the following steps: S21. Use mean filtering to process the tangential component of the cutterhead vibration and obtain the gravitational acceleration variation curve of the tangential component of the cutterhead vibration as a preliminary estimate of the periodic gravitational component. S22. Perform Hilbert transform on the preliminary estimate of the periodic gravity component, calculate its instantaneous amplitude and instantaneous phase, and evaluate the reliability based on the stationarity of the instantaneous amplitude and the linearity of the instantaneous phase. S23. When the confidence assessment result is lower than the preset threshold, the initial periodic gravity component estimate is resynchronized and corrected based on the pulse signal of the cutter head drive encoder, and the corrected signal is used as the final periodic gravity component. Otherwise, the initial periodic gravity component estimate is directly used as the final periodic gravity component for sine function model fitting.

7. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 4, characterized in that, Step S43 involves calculation using interpolation methods, specifically including: a) Using cubic spline interpolation, the vibration acceleration value sequence corresponding to M uniformly distributed angle points is initially calculated; b) Identify the angle sub-intervals in the preliminarily calculated vibration acceleration value sequence that correspond to the drastic changes in signal amplitude caused by tool impact; c) Within the angle sub-interval, a piecewise cubic Hermite interpolation polynomial algorithm is used to locally recalculate the vibration acceleration values, replacing the initial calculation results within the sub-interval, to obtain optimized standardized angle domain samples for model input.

8. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 1, characterized in that, One-dimensional convolutional neural network models are obtained in the following way: Historical tunneling data was acquired and processed to obtain a standardized angle domain sample training set; Data augmentation processing is performed on the training set, including: based on the geomechanical parameters associated with the surrounding rock grade corresponding to the sample, amplitude modulation related to rock strength and phase modulation related to rock mass structure are applied to the sample signal, and noise that conforms to the spectral characteristics of rock mass fracture is superimposed to generate augmented samples. The one-dimensional convolutional neural network model is trained using a training set of augmented samples to obtain a pre-trained model.

9. The surrounding rock identification method based on angle-domain downsampling TBM cutterhead vibration signal as described in claim 1, characterized in that, It also includes step six, confidence assessment and feedback optimization of the surrounding rock identification results, specifically including: S61. Record the surrounding rock grade classification probability vector output by the one-dimensional convolutional neural network model in step five, and take the maximum probability value as the confidence score of the identification result. S62. Set a confidence threshold. When the confidence score is lower than the confidence threshold, determine that the confidence of the current recognition result is insufficient. Extract the statistical features of the current standardized angle domain sample and perform similarity matching with the features of high-confidence samples in the historical database. S63. Bind the final confirmed surrounding rock grade on site with its corresponding standardized angle domain sample to form a new training data pair, and periodically add it to the training dataset to trigger incremental learning and parameter fine-tuning of the one-dimensional convolutional neural network model.

Images