Model-based drilling object material classification and identification method and system

Abstract

The invention belongs to the technical field of production control, and provides a model-based drilling object material classification and identification method and system, electronic equipment and a computer program product. The method comprises the steps of firstly extracting original signals including vibration, sound waves, current, temperature and the like, performing signal synchronization processing, and acquiring various sensor signals based on a unified clock source by adopting hardware-level synchronization; extracting a physical parameter corresponding to the original signal based on a preset physical model; the original signals and the physical parameters are input into a pre-trained hybrid neural network model, the model comprises a first branch neural network and a second branch neural network, the original signals and the physical parameters are processed respectively to obtain corresponding features, the features are weighted and fused based on attention weights, and a classification result is output through a classifier. When abnormal working conditions such as physical parameter overrun, sensor faults or extreme working conditions are detected, the attention weight is corrected according to the abnormal type. The hybrid neural network model is pre-trained by adopting a combined constraint training mode containing classification loss, physical consistency loss and attention sparsification loss. According to the invention, the accuracy and reliability of classification and identification of drilling object materials are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of production control, and in particular to a method, system, electronic equipment and computer program product for classifying and identifying drilling object materials based on a model. Background Art

[0002] In fields such as mechanical processing, geological exploration, and construction, the accurate identification and classification of drilled materials is crucial for optimizing processing parameters, equipment maintenance, and operational safety. Currently, the industry primarily relies on sensor signals (such as vibration, torque, and acoustic emission signals) to build feature models and implement drilled material classification through traditional machine learning algorithms. However, this technology has significant limitations: on the one hand, sensor signals are easily affected by environmental noise, equipment wear, and fluctuations in operating conditions, making it difficult for a single signal feature to fully characterize material properties, resulting in a significant decrease in classification accuracy as operating conditions change. On the other hand, when training samples are insufficient, traditional models cannot effectively extract material features, and their generalization ability and robustness are weak under complex operating conditions.

[0003] Therefore, how to achieve efficient and accurate classification and identification of drilling object materials is an urgent problem to be solved in this field. Summary of the Invention

[0004] To this end, the present invention provides a model-based classification and identification method, system, electronic device and computer program product for drilling object materials to at least partially solve the above technical problems.

[0005] The present invention provides a model-based classification and identification method for drilling object materials, including the following method steps: extracting original signals, wherein the original signals include at least vibration signals, sound wave signals, current signals, and temperature signals; extracting corresponding physical parameters from the original signals based on a preset physical model; inputting the original signals and the physical parameters into a pre-trained hybrid neural network model, and outputting classification results based on the hybrid neural network model, wherein the hybrid neural network model includes a first branch neural network and a second branch neural network, the first branch neural network processes the original signal to obtain original signal features, the second branch neural network is used to process the physical parameters to obtain physical parameter features, the original signal features and the physical parameter features are weightedly fused based on attention weights to obtain fused features, and the classification results are obtained through a classifier based on the fused features.

[0006] Optionally, the method also includes, in response to detecting an abnormal operating condition, the abnormal operating condition type including at least a physical parameter exceeding a limit, a sensor failure or an extreme operating condition, correcting the attention weight based on the abnormal operating condition type, specifically including: if the predicted value of the physical parameter deviates from the theoretical value by more than a certain threshold, increasing the physical feature weight; if the sensor is abnormal, removing the sensor signal and renormalizing the remaining sensor signal weight; if the drill bit is severely worn, resulting in an abnormal vibration signal, constraining the range of change of the attention weight.

[0007] Optionally, the method also includes pre-training the hybrid neural network model using joint constraint training, where the joint constraint training includes loss function constraints, and the loss function includes classification loss, physical consistency loss, and attention sparsification loss.

[0008] Optionally, the method further includes extracting the original signal and signal synchronization, that is, simultaneously acquiring multiple sensor signals, and the signal synchronization adopts hardware-level synchronization to provide a time reference for all sensors based on a unified clock source.

[0009] The present invention also provides a model-based classification and identification system for drilling object materials, comprising: a first extraction module for extracting original signals, wherein the original signals include at least vibration signals, acoustic wave signals, current signals, and temperature signals; a second extraction module for extracting corresponding physical parameters from the original signals based on a preset physical model; a classification output module for inputting the original signals and the physical parameters into a pre-trained hybrid neural network model, and outputting classification results based on the hybrid neural network model, wherein the hybrid neural network model includes a first branch neural network and a second branch neural network, the first branch neural network processes the original signal to obtain original signal features, the second branch neural network is used to process the physical parameters to obtain physical parameter features, the original signal features and the physical parameter features are weightedly fused based on attention weights to obtain fused features, and the classification results are obtained through a classifier based on the fused features.

[0010] Optionally, the system also includes a weight correction module for correcting the attention weight in response to the detection of abnormal working conditions, wherein the abnormal working condition types include at least physical parameter exceeding the limit, sensor failure or extreme working conditions, based on the abnormal working condition types, specifically including: if the predicted value of the physical parameter deviates from the theoretical value by more than a certain threshold, increasing the physical feature weight; if the sensor is abnormal, removing the sensor signal and renormalizing the remaining sensor signal weight; if the vibration signal is abnormal due to severe wear of the drill bit, constraining the range of attention weight variation.

[0011] Optionally, the system also includes a training module for pre-training the hybrid neural network model using a joint constraint training method, wherein the joint constraint training includes loss function constraints, and the loss function includes classification loss, physical consistency loss, and attention sparsification loss.

[0012] Optionally, the system further includes extracting the original signal and signal synchronization, that is, simultaneously acquiring multiple sensor signals, wherein the signal synchronization adopts hardware-level synchronization and provides a time reference for all sensors based on a unified clock source.

[0013] The present invention also provides an electronic device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements any of the aforementioned methods when executed by the processor.

[0014] The present invention also provides a computer program product, which includes a computer program that can be executed by a processor to implement any of the methods described above.

[0015] The beneficial effects of the present invention are: multi-source signal fusion and synchronization improve accuracy: extract multiple original signals such as vibration, sound waves, current, temperature, etc., and obtain them based on a unified clock source through hardware-level synchronization. At the same time, physical parameters are extracted in combination with a preset physical model, and the original signals and physical parameters are input into a hybrid neural network model. Multi-dimensional information fusion processing makes full use of the effective features in the signal, significantly improving the accuracy of classification and identification of drilling object materials.

[0016] Hybrid neural network and attention mechanism optimize recognition effect: The hybrid neural network model obtains features by processing the original signal and physical parameters through the first and second branch neural networks respectively. Based on the weighted fusion feature of attention weight, it can adaptively allocate weights of different information sources, highlight key features, enable the model to classify materials more accurately, and further optimize the recognition effect. Especially when the training samples are insufficient, the optimization of hybrid neural network and attention mechanism can greatly improve the recognition accuracy of the model.

[0017] Abnormal working condition correction ensures stability: For abnormal working conditions such as physical parameter exceeding the limit, sensor failure, and extreme working conditions, corresponding attention weight correction strategies are designed. For example, when physical parameters are abnormal, the weight of physical features is increased, and when sensors fail, the signal weight is renormalized. This ensures that the model can still run stably under complex working conditions, effectively reduces the interference of abnormal conditions on classification results, and improves the stability and reliability of the system.

[0018] Joint constrained training enhances generalization capabilities: A joint constrained training method that includes classification loss, physical consistency loss, and attention sparsification loss is used to pre-train the hybrid neural network model. This constrains the model training process from multiple angles, enabling the model to learn more representative features, improving the model's generalization capabilities and maintaining good classification and recognition performance in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is a schematic diagram of a model-based classification and identification method for drilling object materials disclosed in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of a model-based classification and identification system for drilling object materials disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following specific embodiments illustrate the implementation of this application. Those familiar with the art can easily understand the other advantages and functions of this application from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of this application, but not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0023] In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0024] like Figure 1 As shown, an embodiment of the present invention discloses a model-based classification and identification method for drilling object materials, including the following method steps: S10, extracting original signals, wherein the original signals at least include vibration signals, sound wave signals, current signals, and temperature signals.

[0025] In some embodiments, the original signal is a multimodal signal, including but not limited to vibration signals, acoustic signals, current signals, and temperature signals. The vibration signal is collected using an accelerometer. Based on the MEMS (micro-electromechanical systems) principle, the accelerometer can convert the mechanical vibrations generated by the drill bit during the drilling process into electrical signals. During the actual acquisition process, key parameters such as the drill bit's vibration frequency, amplitude, and acceleration are monitored. The vibration frequency reflects the number of vibrations per unit time, while the amplitude reflects the amplitude of the drill bit's vibrations. Furthermore, the vibration signal can reflect physical information such as material hardness and elastic modulus.

[0026] Acoustic signal acquisition can be achieved using a microphone. Using the principle of acoustic-to-electrical conversion, the microphone converts the noise generated during drilling into an electrical signal. During acquisition, the focus is on the spectral characteristics of the drilling noise. Spectral characteristics refer to the energy distribution of the noise signal across different frequency bands. By analyzing these spectral characteristics, it is possible to identify friction between the drill bit and the material, material fracture, and other factors. Common spectral analysis methods include the Fast Fourier Transform (FFT), which converts acoustic signals in the time domain into frequency domain signals, clearly demonstrating the intensity and distribution of each frequency component. Furthermore, acoustic signals can capture physical information such as material fracture / friction characteristics.

[0027] Current signal acquisition is achieved through Hall-effect current sensors, primarily used to detect motor current fluctuations, indirectly reflecting the drill bit's resistance torque. By connecting a high-precision current sensor in series with the motor's power supply circuit, changes in motor current can be monitored in real time. The current signal can characterize physical information such as the drilling resistance of different materials. When the drill bit encounters different material resistances during drilling, the motor outputs different torques to drive the drill bit, which causes corresponding fluctuations in the motor current. By accurately measuring and analyzing current fluctuations, a mathematical model can be established between current and drill bit resistance torque, enabling accurate assessment of the drill bit's load during the drilling process.

[0028] Temperature signals are collected using infrared sensors. Based on the principle of thermal radiation, infrared sensors can non-contactly monitor temperature changes at the interface between the drill bit and the material. During the drilling process, friction between the drill bit and the material generates significant heat, causing the interface temperature to rise. Furthermore, the temperature signal can be used to infer the thermal conductivity of different materials.

[0029] The aforementioned multimodal signal types, sensor selection, and corresponding physical meanings are detailed in Table 1 below.

[0030] Table 1 List of multimodal signals:

[0031] It is understandable that, according to actual needs, the multimodal signal is not limited to vibration signal, sound wave signal, current signal, and current signal.

[0032] Preferably, extracting raw signals also includes signal synchronization, i.e., simultaneously acquiring multiple sensor signals (e.g., vibration, acoustic waves, current, temperature) and ensuring their temporal alignment to accurately reflect the physical state of the drilling process. Signal synchronization can be achieved through hardware-level synchronization, for example, using a unified clock source (e.g., GPS / PTP protocol) to provide a time reference for all sensors. Synchronous pulses are distributed via an FPGA or dedicated synchronization chip. Each sensor triggers sampling on the rising edge of the pulse, with a time deviation of <1ms. Alternatively, software-level synchronization can be achieved, such as aligning the timestamps of asynchronously acquired signals. This can be achieved through existing dynamic time warping (DTW) or cross-correlation algorithms, which will not be detailed in this embodiment.

[0033] S20: extracting corresponding physical parameters from the original signal based on a preset physical model.

[0034] Based on the physical meaning of the above signal types, the correspondence between the original signal and the physical parameters can be established through the physical model. For example, vibration signal: the elastic modulus (Young's modulus) of the associated material is used to extract features through the vibration frequency attenuation formula. Current signal: the resistance characteristics are derived by combining the mathematical model of material hardness (Rockwell hardness) and current power consumption. Temperature signal: based on the heat conduction equation, the temperature gradient characteristics related to the thermal conductivity of the material are calculated. Acoustic wave signal: based on the parameters directly related to the material analyzed in the acoustic wave signal, relevant features are extracted.

[0035] Establish the mathematical relationship between sensor signals (vibration, sound waves, current, temperature) and material physical parameters (elastic modulus, hardness, thermal conductivity) and build a physical model.

[0036] As an example, consider vibration signal → elastic modulus (Young's modulus) feature extraction. The physical model can be: Drilling vibration frequency and material elastic modulus Related, approximate relationship: ;in, Indicates the main vibration frequency, which is the dominant vibration frequency generated by the interaction between the drill bit and the material during drilling. Elastic modulus: The ability of a material to resist elastic deformation, characterizing stiffness. The larger the value, the stiffer the material. The density of the material, the mass per unit volume of the material, affects the vibration inertia. According to the propagation speed of the elastic wave in the medium , vibration frequency and wave speed Directly proportional.

[0037] The feature extraction method can be: first, perform Fourier transform (FFT) on the vibration signal to extract the main frequency Secondly, combined with the known material density For example, steel , inverse elastic modulus characteristics .

[0038] The comparison of example materials is shown in Table 2 below.

[0039] Table 2 Example material elastic model and vibration frequency comparison:

[0040] If the dominant frequency of the vibration signal is 1100Hz, the model can initially identify it as steel rather than aluminum alloy. If low thermal conductivity (slow temperature rise) is also detected, it can further distinguish between stainless steel (low thermal conductivity) and carbon steel (higher thermal conductivity).

[0041] For example, the current signal → material hardness & drilling resistance characteristics, the physical model is: the relationship between motor current I and drilling resistance F: ;in, The drilling speed of the drill bit. Material hardness (such as Rockwell hardness HRC) affects drilling resistance. The greater the drilling resistance (harder materials) and the higher the speed, the higher the motor current required to maintain power output.

[0042] The feature extraction method is: calculate the dynamic fluctuation amplitude of the current signal , combined with drill parameters (diameter, speed), estimate the specific cutting force (Cutting force per unit area): ;in, is the motor efficiency constant, is the drill bit diameter. According to is the drill bit diameter. According to Infer the hardness range of the material, such as high Compatible with high hardness materials.

[0043] The comparison of example materials is shown in Table 3 below.

[0044] Table 3 Example material specific cutting force and hardness comparison:

[0045] If the current fluctuates violently and , the model prioritizes titanium alloy rather than ordinary steel, and combined with the temperature signal (titanium alloy heats up quickly when drilling), the classification confidence can be improved.

[0046] For example, the temperature signal → thermal conductivity feature extraction, the physical model is: the friction heat Q generated during drilling and the thermal conductivity of the material Relationship: .

[0047] Materials that conduct heat poorly, such as plastics, will heat up faster.

[0048] The feature extraction method is to monitor the temperature rise rate of the drill bit-material interface , according to the heat conduction equation, the thermal conductivity is deduced .

[0049] The comparison of example materials is shown in Table 4 below.

[0050] Table 4 shows the comparison between thermal conductivity and temperature rise rate of example materials:

[0051] If the temperature rise rate is >3°C / s, the model tends to identify it as plastic rather than metal. Combined with acoustic signals (high-frequency noise is less when drilling plastic), this can eliminate misidentification as metal.

[0052] For example, the acoustic waves generated during drilling are primarily due to the following physical processes: Friction sound: the friction between the drill bit and the material surface (high-frequency components). Fracture sound: the propagation of cracks within the material (burst pulse signals). Cavity resonance: the vibration of the cavity formed by drilling (specific frequency resonance peaks).

[0053] The correlation table of key physical quantities is shown in Table 5.

[0054] Table 5: Correspondence between sound waves and physical quantities:

[0055] The feature extraction method includes the following steps: Step 1: Signal preprocessing. Denoising: Wavelet threshold denoising (Daubechies 5 wavelet) is used to remove ambient noise (e.g., motor hum). Framing: The continuous acoustic signal is framed (20ms long, with 50% overlap) to capture transient events (e.g., cracking sounds). Step 2: Time-frequency domain feature extraction.

[0056] (1) Resonance frequency analysis (cavity resonance model). Physical model: The drilled cavity is approximately a Helmholtz resonator, and the resonance frequency and material elastic modulus and density Related: ;in, Represents the pore volume.

[0057] Feature extraction: Perform FFT spectrum analysis on the sound wave signal to extract the main resonance peak frequency Combined with the drill hole diameter (known), reverse calculation Ratio characteristics.

[0058] (2) Transient event detection (fracture acoustic model). Physical model: The acoustic emission signal released by material fracture conforms to the damped oscillation model: ;in, Indicates the instantaneous amplitude, the amplitude of the acoustic emission signal at time t, and quantifies the change of the intensity of the fracture sound over time. Indicates the initial amplitude, the maximum amplitude of the sound wave at the moment of fracture (t=0), which is usually proportional to the fracture energy of the material. Indicates the decay time constant, the amplitude decays to the initial value of The time required is related to the damping properties of the material. Indicates the fracture characteristic frequency, the main frequency of the acoustic emission signal oscillation (related to the crack size), represents the exponential decay term, which describes the envelope of the exponential decay of the amplitude over time and distinguishes the energy dissipation rate of different materials. Represents the oscillation term. The periodic oscillation component of the acoustic wave carries the frequency information of the dynamic crack expansion.

[0059] Feature extraction: locate the fracture sound event by short-time energy detection. Fit the attenuation envelope and extract and .

[0060] (3) Friction sound harmonic analysis (surface hardness model). Physical model: The degree of harmonic distortion of friction sound is related to the hardness of the material. The higher the hardness, the richer the high-frequency harmonics. Feature extraction: Calculate the harmonic distortion (THD) of the spectrum to quantify the proportion of harmonic components (nonlinear distortion) in the signal to the total signal, reflecting the purity of the sound wave signal. It is used to characterize the nonlinear characteristics of the material during friction or fracture (such as surface hardness and internal structural unevenness): ;in, Fundamental amplitude, indicating the main frequency component frequency of the signal The amplitude of The amplitude of the nth harmonic: the frequency is The component amplitudes of (n=2,3,4,5), Indicates the total harmonic energy: the root mean square (RMS) value of the 2nd to 5th harmonics, which represents the distortion intensity.

[0061] S30, input the original signal and the physical parameter into a pre-trained hybrid neural network model, and output a classification result based on the hybrid neural network model, wherein the hybrid neural network model includes a first branch neural network and a second branch neural network, the first branch neural network processes the original signal to obtain original signal features, the second branch neural network is used to process the physical parameters to obtain physical parameter features, the original signal features and the physical parameter features are weightedly fused based on the attention weight to obtain fused features, and the classification result is obtained through a classifier based on the fused features.

[0062] In one embodiment, the hybrid neural network model uses a two-branch hybrid constraint network, combining data-driven features with physical law features, dynamically integrating them through an attention mechanism, and ultimately outputting material classification results. The model's core modules include: 1. Input layer: Synchronously acquires multimodal signals. The original signals include at least one of vibration, acoustic, current, and temperature signals.

[0063] Second, the first neural network branch is responsible for extracting deep statistical features from raw sensor signals. In this embodiment, this network combines the local feature capture capabilities of CNN and the time series modeling capabilities of LSTM. For example, the specific architecture of this network includes: 1. Input Data Processing: Input signal: multimodal time series data (vibration, current, temperature, etc.) with a uniform sampling length.

[0064] Data preprocessing: Normalization: Z-score normalization was performed on each signal channel.

[0065] Segmentation: The vibration signal is divided into frames, and the sound wave signal retains the original waveform.

[0066] 2. CNN module: spatial / frequency domain feature extraction: (1) Vibration signal processing (1D-CNN): Input: vibration acceleration signal (3-axis, length 10,000 points, 10kHz sampling).

[0067] The network structure is as follows: Table 6 Network structure table:

[0068] Among them, a wide convolution kernel (Kernel=5) is used to capture the main frequency of low-frequency vibration, and GAP replaces the fully connected layer to reduce the number of parameters and prevent overfitting.

[0069] (2) Acoustic signal processing (2D-CNN): Input: acoustic spectrum (STFT transform, 128 Mel frequency band).

[0070] The network structure is as follows: Table 7 Network structure table:

[0071] 3. LSTM module: Time series dynamic modeling: (1) Current / temperature signal processing: Input: current and temperature signals (length 1000 points, 1kHz sampling), spliced ​​into 2D time series data.

[0072] The network structure is as follows: Table 8 Network structure table:

[0073] In this embodiment, a bidirectional LSTM is adopted to enhance the utilization of historical / future information, and Dropout=0.3 is adopted to prevent overfitting of small samples.

[0074] (2) Multimodal time series fusion: Input: vibration (64-dimensional) + sound wave (32-dimensional) features output by CNN, spliced ​​with the 32-dimensional time series features of LSTM.

[0075] Fusion layer: total dimension: 64+32+32=128 dimensions.

[0076] 4. Output: Feature output: 128-dimensional data-driven feature vector .

[0077] 3. Second branch neural network: Differentiable physical calculation layer, which directly maps the original sensor signal into physical parameters (such as elastic modulus, hardness, thermal conductivity). This network strictly follows the laws of physics to ensure that the model output conforms to the intrinsic properties of the material. The network structure includes: (1) Elastic modulus calculation unit: Physical model: Relationship between vibration frequency and material stiffness: ;in, is the material density, which can be obtained by preset or online estimation, The calibration coefficient can be determined through experiments, which will not be described in detail here. The calculation unit can be used to convert the main vibration frequency into a stiffness parameter to distinguish between metal (high E) and non-metal (low E).

[0078] (2) Specific cutting force calculation unit: Physical model: Current-resistance-hardness relationship: Where d is the drill diameter (known parameter) and k is the motor efficiency constant (calibrated value). Used to quantify the hardness of materials, such as titanium alloys. , aluminum .

[0079] (3) Thermal conductivity calculation unit: Physical principle: Heat conduction equation: Where Q represents the input heat flux (estimated by the current power) and A is the drill-material contact area, identifying low thermal conductivity materials (such as plastic) versus high thermal conductivity materials (aluminum).

[0080] (4) Acoustic attenuation unit: Physical principle: Damped oscillation model: ;in, Represents the envelope of the acoustic wave signal, which is used to reflect the internal damping characteristics of the material (such as cast iron ,rubber ).

[0081] The network structure is based on the key physical quantities of the multimodal sensor signal and outputs the normalized physical driving feature vector: .

[0082] Each parameter is normalized to the theoretical range, e.g. .

[0083] It should be noted that the calculation formulas of all the above calculation units are designed as explicit differentiable functions to ensure that the gradient can be back-propagated. For example Can be calculated automatically.

[0084] IV. Classification output: 1. Feature splicing: For example, 128-dimensional data drives the feature vector and 64-dimensional physical driver feature vector The two types of heterogeneous feature vectors are merged into a unified high-dimensional representation. Specifically, the following steps are performed: dimensional alignment: ensuring that the number of samples (batch size) of the two vectors is consistent; normalization: scaling the physical features to the theoretical range to avoid dimensional differences; and concatenation: directly concatenating the two vectors while maintaining the feature order to form a 192-dimensional concatenated feature vector. .

[0085] 2. Attention weight calculation: Based on a single-layer fully connected network combined with softmax calculation weights and output classification, for example, attention weight The calculation process is as follows: .

[0086] Expand into step-by-step calculations: Linear transformation: ;in represents the data-driven feature weights, represents the physical driving feature weight, , an unnormalized logical value, reflecting the contribution of the feature to the current task, where represents the weight matrix ( is the coefficient of the i-th weight to the j-th splicing feature), Represents the bias term , Represents the concatenated feature vector ( arrive ).

[0087] Softmax normalization: .

[0088] make sure .

[0089] Weighted fusion: Final feature = data feature × data weight + physical feature × physical weight.

[0090] 3. Result output: The fully connected layer outputs the classification results, and the physical parameters are verified. If the predicted value is within the theoretical range, the result is credible and the final classification is output.

[0091] For example, the material classification process is implemented according to the model of this embodiment, as shown below: 1. Input data preparation: original sensor signal input: vibration acceleration signal: main frequency 1050Hz, amplitude 3.5g (sampling rate 10kHz).

[0092] Motor current signal: average value 4.2A, fluctuation amplitude ±0.6A (sampling rate 1kHz).

[0093] Temperature signal: The drill contact surface temperature rise rate is 2.4°C / s (sampling rate is 100Hz).

[0094] Sound wave signal: total harmonic distortion (THD) = 12%, main frequency 8kHz (sampling rate 44.1kHz).

[0095] Physical parameter input: Material density ρ = 4500 kg / m³ (default).

[0096] Drill bit diameter d=8mm.

[0097] Drilling speed v=1.2m / s.

[0098] The motor efficiency constant k=0.8.

[0099] 2. Model processing flow: (1) First branch network: data-driven feature extraction.

[0100] Vibration signal processing: 1D-CNN extracts frequency domain features: main frequency energy 0.85, second harmonic energy 0.12.

[0101] Output feature vector: [0.85, 0.12, ...] (64 dimensions).

[0102] Current / temperature signal processing: LSTM identifies current fluctuation patterns (sawtooth waveforms).

[0103] Output feature vector: [0.62, 0.33, ...] (32 dimensions).

[0104] Acoustic signal processing: 2D-CNN analysis of Mel spectrum.

[0105] Output feature vector: [0.71, 0.08, ...] (32 dimensions).

[0106] Merged output: F_data = [vibration characteristics; current and temperature characteristics; acoustic wave characteristics]∈R^128.

[0107] (2) The second branch network: physical parameter feature calculation.

[0108] Calculation of elastic modulus: E = 4500 × (1.1 × 1050) ^ 2 ≈ 114 GPa → normalized value 0.57.

[0109] Specific cutting force calculation: K_c=(4.2×0.8) / (1.2×0.008^2)≈2187N / mm² → normalized value 0.88.

[0110] Thermal conductivity calculation: λ = 50 / (5 × 10^-5 × 2.4) ≈ 7 W / m·K → normalized value 0.35.

[0111] Merged output: F_physics=[0.57,0.88,0.35,...]∈R^64.

[0112] (3) Attention fusion.

[0113] Feature concatenation: F_concat=[F_data;F_physics]∈R^192.

[0114] Weight calculation: K_c≈2200N / mm² and λ≈7W / m·K both match the characteristics of titanium alloy.

[0115] Automatically assign weights: a=[0.25,0.75] (data features: physical features).

[0116] Weighted fusion: F_final=0.25×F_data+0.75×F_physics.

[0117] (4) Classification decision.

[0118] The fully connected layer outputs logits: [titanium alloy: 3.2, stainless steel: 1.1, aluminum alloy: 0.5].

[0119] Softmax probability: [titanium alloy: 0.91, stainless steel: 0.06, aluminum alloy: 0.03].

[0120] Physical verification: λ = 7W / m·K is within the theoretical range (6-8) for titanium alloys → Verification passed.

[0121] 3. Final output: Classification result: Prediction category: Titanium alloy Ti-6Al-4V.

[0122] Confidence level: 91%.

[0123] Key parameters: Specific cutting force: 2187N / mm².

[0124] Thermal conductivity: 7W / m·K.

[0125] Feature weighting: Data 25%, Physical 75%.

[0126] Preferably, in response to detecting an abnormal operating condition, the abnormal operating condition type includes at least physical parameter exceeding the limit, sensor failure or extreme operating condition, and the attention weight is corrected based on the abnormal operating condition type, specifically including: if the predicted value of the physical parameter deviates from the theoretical value by more than a certain threshold, the physical feature weight is increased; if the sensor is abnormal, the sensor signal is eliminated and the remaining sensor signal weight is renormalized; if the drill bit is severely worn, resulting in an abnormal vibration signal, the range of attention weight variation is constrained.

[0127] In some cases, the attention weights automatically calculated by the model may become invalid due to signal noise, physical parameter excursions, sensor failure, or extreme operating conditions (e.g., severe drill bit wear). To address this issue, this embodiment introduces a weight correction strategy to ensure that the model still properly assigns feature weights under abnormal circumstances. For example, the triggering conditions for weight correction are shown in the following table.

[0128] Table 9 Weight correction comparison table:

[0129] Exemplarily, the weight correction process for different anomaly types includes: (1) weight constraints based on physical laws.

[0130] Problem: The model may over-rely on data features (such as vibration signals) and ignore physical parameters (such as thermal conductivity).

[0131] Correction method: If the physical parameters Deviation from theoretical value Exceeding the threshold Force physical feature weights to be higher: ;in, Physical parameter weights automatically calculated for the model, is the correction coefficient, is the empirical value, and the threshold For experience value.

[0132] (2) Weight redistribution under sensor failure.

[0133] Problem: The temperature sensor failed, resulting in abnormal temperature rise rate data.

[0134] Correction method: Fault detection: If the temperature signal standard deviation is less than the preset threshold (for example, the change is less than 1°C within 10 minutes), it is judged as a failure.

[0135] Weight reset: remove the temperature signal and renormalize the remaining weights: ;in, 、 is the original weight, satisfying .

[0136] is the sum of the effective weights, and the sum of the remaining weights after removing the weights corresponding to the failure signals.

[0137] 、 is the new weight after reset, still satisfying .

[0138] For example, if a temperature sensor fails, its associated physical feature weights need to be partially eliminated.

[0139] Original weight: a=[0.6,0.4] (data feature weight: 0.6, physical feature weight: 0.4).

[0140] Failure impact: The temperature signal contributes 30% of the physical weight → failure weight = 0.4 × 0.3 = 0.12.

[0141] The effective weight sum is: .

[0142] Weight reset: .

[0143] The weight of the data features is increased (0.6→0.68) to compensate for the partial failure of the physical features. The weight of the physical features is reduced but the valid parts are retained.

[0144] (3) Correction of attention weight constraints.

[0145] Problem: Drill bit wear causes abnormal vibration signals, misleading the model.

[0146] Correction method: constrain the range of attention weight changes during training: ,in, Represents the attention weight of the data-driven feature. The initial value is calculated by the network. The Clip function: forces the weight to be limited to the range of [0.2, 0.8] to ensure that the model always takes into account both data and physical features and avoids extreme dependence.

[0147] Preferably, the hybrid neural network model is pre-trained using a joint constraint training method, wherein the joint constraint training includes loss function constraints, and the loss function includes classification loss, physical consistency loss, and attention sparsification loss.

[0148] Exemplarily, the specific training process includes: Step 1: data preprocessing.

[0149] Synchronize multimodal signals (vibration, current, temperature, sound waves).

[0150] Mark the material category (e.g., "304 stainless steel," "C30 concrete").

[0151] Step 2: Network forward propagation (taking vibration signals and temperature signals as examples).

[0152] Data branch: input vibration signal → CNN → extract frequency domain features .

[0153] Physics branch: Input temperature signal → FCN → Predict thermal conductivity .

[0154] Fusion layer: and Splicing, input classifier.

[0155] Step 3: Loss calculation: .

[0156] The parameters are explained in the following table: Table 10 Parameter Explanation Table:

[0157] Step 4: Backpropagation and optimization.

[0158] Use gradient descent (e.g., Adam) to simultaneously optimize the classification loss, physical consistency loss, and attention sparsification loss.

[0159] The advantage of hybrid constraint training is that it can overcome the problems of overfitting and poor generalization of pure data-driven models based on small sample training. By jointly training the constraint model, the physical constraints reduce the dependence on the amount of data. When the training data is insufficient, the prior knowledge provided by the physical model can improve the classification accuracy, reduce the dependence on large amounts of training data, and improve the efficiency of model training. At the same time, through physical consistency verification, abnormal signals (such as current signal distortion caused by drill bit wear) can be identified to avoid misclassification and improve the anti-interference ability and robustness of the model classification.

[0160] Corresponding to the above embodiment, a model-based classification and identification method for drilling object materials is provided. Figure 2 A structural block diagram of a model-based drilling object material classification and identification system provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0161] See also Figure 2 As shown, a model-based drilling object material classification and identification system 200 provided in an embodiment of the present application includes: a first extraction module for extracting original signals, wherein the original signals include at least vibration signals, sound wave signals, current signals, and temperature signals.

[0162] The second extraction module is used to extract corresponding physical parameters from the original signal based on a preset physical model.

[0163] A classification output module is used to input the original signal and the physical parameters into a pre-trained hybrid neural network model, and output a classification result based on the hybrid neural network model, wherein the hybrid neural network model includes a first branch neural network and a second branch neural network, the first branch neural network processes the original signal to obtain original signal features, the second branch neural network is used to process the physical parameters to obtain physical parameter features, the original signal features and the physical parameter features are weightedly fused based on the attention weight to obtain fused features, and the classification result is obtained through a classifier based on the fused features.

[0164] An embodiment of the present invention further provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements any of the aforementioned methods when executed by the processor.

[0165] An embodiment of the present invention further provides a computer program product, which includes a computer program that can be executed by a processor to implement any of the methods described above.

[0166] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0167] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A model-based method for classifying and identifying drilling object materials, characterized in that: The method includes the following steps: extracting the original signal, wherein the original signal includes at least a vibration signal, an acoustic wave signal, a current signal, and a temperature signal; extracting corresponding physical parameters from the original signal based on a preset physical model; inputting the original signal and the physical parameters into a pre-trained hybrid neural network model, and outputting a classification result based on the hybrid neural network model, wherein the hybrid neural network model includes a first branch neural network and a second branch neural network, the first branch neural network processes the original signal to obtain original signal features, the second branch neural network is used to process the physical parameters to obtain physical parameter features, the original signal features and the physical parameter features are weightedly fused based on attention weights to obtain fused features, and the classification results are obtained through a classifier based on the fused features.

2. The method for classifying and identifying drilling object materials based on a model according to claim 1, characterized in that: In response to detecting an abnormal operating condition, the abnormal operating condition type includes at least physical parameter exceeding the limit, sensor failure or extreme operating condition, and the attention weight is corrected based on the abnormal operating condition type. Specifically, if the predicted value of the physical parameter deviates from the theoretical value by more than a certain threshold, the physical feature weight is increased; if the sensor is abnormal, the sensor signal is eliminated and the remaining sensor signal weight is renormalized; if the drill bit is severely worn, resulting in an abnormal vibration signal, the range of attention weight change is constrained.

3. The method for classifying and identifying drilling object materials based on a model according to claim 2, characterized in that: The hybrid neural network model is pre-trained using a joint constraint training method, wherein the joint constraint training includes loss function constraints, and the loss function includes classification loss, physical consistency loss, and attention sparsification loss.

4. The method for classifying and identifying drilling object materials based on a model according to claim 1, characterized in that: Extracting the original signal also includes signal synchronization, that is, simultaneously acquiring multiple sensor signals. The signal synchronization adopts hardware-level synchronization and provides a time reference for all sensors based on a unified clock source.

5. A model-based classification and identification system for drilling object materials, characterized in that: include: A first extraction module is used to extract original signals, wherein the original signals include at least vibration signals, sound wave signals, current signals, and temperature signals; A second extraction module is used to extract corresponding physical parameters from the original signal based on a preset physical model; A classification output module is used to input the original signal and the physical parameters into a pre-trained hybrid neural network model, and output a classification result based on the hybrid neural network model, wherein the hybrid neural network model includes a first branch neural network and a second branch neural network, the first branch neural network processes the original signal to obtain original signal features, the second branch neural network is used to process the physical parameters to obtain physical parameter features, the original signal features and the physical parameter features are weightedly fused based on the attention weight to obtain fused features, and the classification result is obtained through a classifier based on the fused features.

6. The model-based drilling object material classification and identification system according to claim 5, characterized in that: include: A weight correction module is used to respond to the detection of abnormal working conditions, where the abnormal working condition types include at least physical parameter exceeding the limit, sensor failure or extreme working conditions, and correct the attention weight based on the abnormal working condition type. Specifically, if the predicted value of the physical parameter deviates from the theoretical value by more than a certain threshold, the physical feature weight is increased; if the sensor is abnormal, the sensor signal is eliminated and the remaining sensor signal weight is renormalized; if the vibration signal is abnormal due to severe wear of the drill bit, the range of attention weight change is constrained.

7. The model-based drilling object material classification and identification system according to claim 5, characterized in that: include: A training module is used to pre-train the hybrid neural network model using a joint constraint training method, wherein the joint constraint training includes loss function constraints, and the loss function includes classification loss, physical consistency loss, and attention sparsification loss.

8. The model-based drilling object material classification and identification system according to claim 5, characterized in that: include: Extracting the original signal also includes signal synchronization, that is, simultaneously acquiring multiple sensor signals. The signal synchronization adopts hardware-level synchronization and provides a time reference for all sensors based on a unified clock source.

9. An electronic device, characterized in that: The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements the method according to any one of claims 1 to 4 when executed by the processor.

10. A computer program product, characterized in that: The computer program product comprises a computer program that can be executed by a processor to implement the method according to any one of claims 1 to 4.

Images