Intelligent sensing method of shield tool based on speed-temperature-wear

Abstract

The present invention relates to the technical field of drilling in soil or rock, and specifically discloses a method for intelligent perception of speed-temperature-wear of shield tools, including: micro-synchronous acquisition, coupled feature extraction, physical modeling fitting, multi-scale thermal anomaly separation, physical constraint neural network prediction, time series evolution prediction, dynamic threshold risk decision-making, adaptive sampling scheduling and online incremental calibration; the present invention dynamically adjusts the warning threshold based on the prediction error distribution, and can flexibly switch between different reliability requirements to avoid the problems of "premature warning" or "missed reporting". At the same time, linking the sensor sampling frequency with the warning level can automatically improve the data resolution in the high-risk stage, save computing and transmission resources in the low-risk stage, and realize the efficient use of computing power and bandwidth by the monitoring system; therefore, it can not only cope with the distribution drift caused by changes in geological conditions and tool updates, but also extend the effective life of the model, with almost no need for downtime maintenance.

Description

Technical Field

[0001] The present invention belongs to the technical field of soil or rock drilling, and particularly relates to a rotation speed-temperature-wear intelligent sensing method for shield cutters. Background Art

[0002] When drilling soil or rock, shield cutters are key components that directly contact the rock and soil and enable tunneling. Their wear directly impacts the efficiency, quality, and cost of shield construction. Accurate, real-time sensing of shield cutter wear is crucial for timely tool replacement, preventing operational failures caused by excessive cutter wear, and optimizing shield tunneling parameters.

[0003] At present, the sensing technology for shield tool wear has made certain progress, but there are still many shortcomings.

[0004] First, existing techniques for heat flux calculation fail to fully consider the relationship between tool surface topography and the friction-heat transfer process. The 3D topography of the tool surface significantly influences the distribution of frictional heat flux. However, conventional methods lack an effective means of integrating 3D surface topography data with the friction-heat transfer equations. This results in inaccurate instantaneous frictional heat flux calculations, hindering the accurate assessment of tool wear.

[0005] Secondly, when it comes to separating wear-induced abnormal signals, existing methods struggle to effectively isolate the high-frequency thermal anomaly components directly associated with local tool defects when processing heat flux distribution data. Abnormal signals generated during tool wear are often masked by low-frequency environmental disturbances (such as overall temperature drift) and structured noise (such as mechanical vibration). This results in a significant amount of noise in the subsequent neural network input data, reducing the overall monitoring signal-to-noise ratio and accuracy, making the judgment of tool wear prone to deviations.

[0006] Finally, existing models for wear rate prediction lack the ability to capture the long- and short-term dependencies of wear rate. Shield cutter wear rates are influenced by both transient, sudden factors and long-term, cumulative wear processes. However, traditional models struggle to account for both these factors simultaneously, resulting in poor prediction accuracy for both short-term response and long-term planning, making them incapable of providing reliable decision-making for shield construction.

[0007] In this regard, this application proposes a speed-temperature-wear intelligent sensing method for shield tools to solve the above problems. Summary of the Invention

[0008] The purpose of the present invention is to provide a speed-temperature-wear intelligent sensing method for shield tools to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] The intelligent sensing method of shield tool based on speed, temperature and wear includes:

[0011] A thermocouple array and a photoelectric encoder are arranged on the surface of the tool to synchronously collect temperature time series signals and speed time series signals to construct a preliminary correlation matrix; mutual information analysis is performed on the correlation matrix to extract temperature-speed coupling features to obtain a feature set;

[0012] The feature set is combined with the 3D topography scanning data of the tool surface through least square fitting of the friction-heat transfer equation to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model; the heat flux distribution model is subjected to wavelet packet decomposition to separate the high-frequency thermal anomaly components induced by wear and obtain the abnormal components;

[0013] Inputting the feature set and the abnormal component into a physical constraint neural network model to predict the instantaneous wear rate and obtain an initial wear rate;

[0014] Time series evolution and online calibration: The initial wear rate and the latest temperature-speed history sequence are input into the LSTM-Transformer model to predict the future wear process. Online incremental calibration is then performed based on the actual wear measurements after replacement to obtain the final optimized prediction result.

[0015] The prediction results and the latest collected temperature-speed history sequence are used as input to the LSTM-Transformer model to predict the future wear process and obtain a prediction sequence. The prediction sequence is then subjected to dynamic threshold determination based on P-value control to generate a tool risk warning status.

[0016] The early warning status is used to adjust the sampling frequency and alarm level of the thermocouple and encoder to obtain an optimized acquisition strategy; the actual measured value of tool wear after replacement is compared with the predicted sequence to update the weights of the neural network and timing model online to obtain an optimized model.

[0017] Preferably, performing mutual information analysis on the correlation matrix to extract temperature-speed coupling features to obtain a feature set includes: calculating the mutual information between temperature and speed by a mutual information evaluation method to identify the most effective coupling amount for wear prediction from the correlation matrix, and selecting several features with the largest amount of information and the lowest redundancy based on the maximum correlation and minimum redundancy criterion;

[0018] The formula of the mutual information evaluation method is:

[0019] ;

[0020] Where T: tool surface temperature random variable, value set T={t1,…,tM};

[0021] ω: tool speed random variable, value set Ω={ω1,…,ωN};

[0022] p(t,ω): joint probability density, representing the frequency of observations at temperature t and speed ω;

[0023] p(t), p(ω): marginal probability density, which are the observation frequencies of temperature or speed, respectively.

[0024] Preferably, the feature set and the tool surface 3D topography scanning data are fitted with the least squares of the friction-heat transfer equation to calculate the instantaneous friction heat flux density, and the friction-heat transfer equation in the heat flux distribution model is obtained as follows:

[0025] ;

[0026] Where, q: instantaneous frictional heat flux density, that is, the frictional heat flux generated per unit area, unit W / m²;

[0027] h: surface heat transfer coefficient, reflecting the conduction and convection capacity between the tool and the environment, unit W / (m²·K);

[0028] Ts: Real-time temperature of tool surface, measured by thermocouple array, unit: °C;

[0029] Ta: ambient temperature, which can be the temperature of surrounding rock cuttings or lining, unit: °C;

[0030] μ: friction coefficient, related to tool material and rock type;

[0031] p: tool-rock contact pressure, obtained by force sensor or robotic arm feedback, unit: Pa;

[0032] v: Tool linear velocity, equal to the rotational speed ω multiplied by the radius, unit: m / s.

[0033] Preferably, the feature set and the abnormal component are input into a physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate, including adding constraints to the loss function and adopting a physical constraint loss function to ensure that the data-driven model does not violate the friction-heat transfer mechanism;

[0034] The formula of the physical constraint loss function is:

[0035] ;

[0036] Where, N: total number of training samples;

[0037] ,yi: wear rate predicted and measured by the neural network of the i-th sample, unit: mm / h;

[0038] ΔEi pred : the thermal energy change of the i-th sample calculated from the predicted wear amount, unit J;

[0039] ΔEi phys : the physical thermal energy change of the i-th sample calculated according to the friction-heat transfer equation, unit J;

[0040] λ: Constraint weight, used to balance fitting accuracy and physical consistency (typical value is 0.1–1.0).

[0041] Preferably, the wavelet packet decomposition of the heat flux distribution model is performed to separate the high-frequency thermal anomaly components induced by wear to obtain the abnormal components, including: using the db4 wavelet basis to decompose the heat flux distribution model into 4 levels to extract the high-frequency thermal anomaly components in the 3rd and 4th level sub-bands.

[0042] Preferably, the LSTM-Transformer model includes a bidirectional LSTM encoder and a multi-head self-attention decoder, wherein the number of attention heads is not less than 8, which is used to capture the long-term and short-term dependencies of the historical sequence;

[0043] The online incremental calibration adopts a sample replay strategy, in which the single weight update ratio does not exceed 5%.

[0044] Preferably, the feature set and the abnormal component are input into a physical constraint neural network model to predict the instantaneous wear rate to obtain the initial wear rate. The neural network structure is a three-layer fully connected network, the number of input layer nodes is equal to the sum of the feature set and the abnormal component dimensions, the middle layer includes energy conservation constraints, and the output layer is a single-node instant wear rate prediction.

[0045] Preferably, the dynamic threshold judgment based on P-value control of the prediction sequence is used to generate a tool risk warning state. The P-value control calculates the distribution parameters of the prediction error based on the sliding window method and dynamically adjusts the warning threshold to a 95% confidence level.

[0046] Preferably, the warning status is used to adjust the sampling frequency and alarm level of the thermocouple and encoder to obtain an optimized acquisition strategy, in which the sampling frequency is divided into three levels according to the warning status: high risk ≥20 kHz, medium risk 10-20 kHz and low risk 5-10 kHz.

[0047] Preferably, the actual measured value of tool wear after replacement is compared with the predicted sequence to update the weights of the neural network and timing model online to obtain the optimized model. The model update adopts an incremental learning strategy based on sample replay, and only 5% of the weight is updated each time calibration is performed.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The present invention extracts the coupling characteristics between temperature and speed through mutual information analysis. This method can capture implicit wear signals that are difficult to reflect with a single sensing dimension, significantly improving the sensitivity to the micro-wear stage of the tool. Compared with methods that rely solely on temperature or vibration, the multi-source coupling characteristics can reveal wear trends at an early stage, thereby providing early warning and reducing the "blind zone" time; the friction-heat transfer physics equation is integrated into the heat flux density fitting and neural network loss function, which not only closely combines empirical data drive with theoretical mechanisms, but also suppresses overfitting and physical violations through energy conservation constraints during model training. This not only enhances the robustness of the model under unknown working conditions, but also improves the interpretability of the results, making it easier for on-site engineers to perform fault diagnosis based on physical mechanisms.

[0050] (2) The present invention uses wavelet packet multi-scale decomposition to effectively separate high-frequency thermal anomalies directly associated with local defects in the tool, so that the subsequent neural network input is concentrated on the most representative abnormal components. This processing has significant advantages in suppressing low-frequency environmental disturbances (such as overall temperature drift) and structural noise (such as mechanical vibration), thereby improving the signal-to-noise ratio and accuracy of overall monitoring; the prediction model is designed as a hybrid structure of LSTM and multi-head self-attention, which can simultaneously capture short-term fluctuations and long-term trends in wear rate. This time series modeling method can not only reflect instantaneous sudden wear but also take into account the cumulative attenuation law, so that the prediction results remain highly accurate in both short-term response and long-term planning.

[0051] (3) The present invention dynamically adjusts the warning threshold based on the distribution of prediction errors, and can flexibly switch between different reliability requirements to avoid the problems of "premature warning" or "missed warning". At the same time, by linking the sensor sampling frequency with the warning level, it can automatically improve the data resolution in the high-risk stage and save computing and transmission resources in the low-risk stage, so as to achieve efficient utilization of computing power and bandwidth by the monitoring system; by comparing the actual wear after replacement with the predicted value for incremental learning, this scheme can continuously self-calibrate during operation and correct model deviations in a timely manner. This mechanism can not only cope with the distribution drift caused by changes in geological conditions and tool updates, but also extend the effective life of the model, with almost no downtime for maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of the speed-temperature-wear intelligent sensing method for shield tools of the present invention. DETAILED DESCRIPTION

[0053] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] Example 1: Please refer to Figure 1 As shown in FIG, the speed-temperature-wear intelligent sensing method for shield tools includes:

[0055] S1: A thermocouple array and a photoelectric encoder are arranged on the surface of the tool to synchronously collect temperature time series signals and speed time series signals and construct a preliminary correlation matrix; mutual information analysis is performed on the correlation matrix to extract temperature-speed coupling features and obtain a feature set;

[0056] The mutual information analysis of the correlation matrix in step S1 is performed to extract temperature-speed coupling features, and the feature set obtained includes: calculating the mutual information between temperature and speed by a mutual information evaluation method to identify the most effective coupling amount for wear prediction from the correlation matrix, and selecting several features with the largest amount of information and the lowest redundancy based on the maximum correlation and minimum redundancy criterion;

[0057] The formula of the mutual information evaluation method is:

[0058] ;

[0059] Where T: tool surface temperature random variable, value set T={t1,…,tM};

[0060] ω: tool speed random variable, value set Ω={ω1,…,ωN};

[0061] p(t,ω): joint probability density, representing the frequency of observations at temperature t and speed ω;

[0062] p(t), p(ω): marginal probability density, the observation frequency of temperature or speed respectively;

[0063] By calculating I(T;ω), we can screen out the temperature-speed coupling features that are most relevant to temperature changes and have the least redundancy. These features are then used to construct feature sets, significantly improving the sensitivity and generalization capabilities of subsequent models to operating condition switching.

[0064] S2: The feature set is combined with the 3D topography scanning data of the tool surface through least square fitting of the friction-heat transfer equation to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model; the heat flux distribution model is subjected to wavelet packet decomposition to separate the high-frequency thermal anomaly components induced by wear and obtain the abnormal components;

[0065] The feature set and the tool surface 3D topography scanning data are fitted with the least squares of the friction-heat transfer equation to calculate the instantaneous friction heat flux density, and the formula of the friction-heat transfer equation in the heat flux distribution model is obtained as follows:

[0066] ;

[0067] Where, q: instantaneous frictional heat flux density, that is, the frictional heat flux generated per unit area, unit W / m²;

[0068] h: surface heat transfer coefficient, reflecting the conduction and convection capacity between the tool and the environment, unit W / (m²·K);

[0069] Ts: Real-time temperature of tool surface, measured by thermocouple array, unit: °C;

[0070] Ta: ambient temperature, which can be the temperature of surrounding rock cuttings or lining, unit: °C;

[0071] μ: friction coefficient, related to tool material and rock type;

[0072] p: tool-rock contact pressure, obtained by force sensor or robotic arm feedback, unit: Pa;

[0073] v: tool linear velocity, equal to the rotational speed ω multiplied by the radius, unit: m / s;

[0074] This equation can be used to convert coupling characteristics into a heat flux distribution model, achieving physical quantification of tool force and thermal field. The temperature-speed characteristics can be mapped into a heat flux distribution model to capture the changes in heat sources during tool friction, providing an accurate physical basis for high-frequency thermal anomaly separation and wear prediction.

[0075] S3: Inputting the feature set and the abnormal component into a physical constraint neural network model to predict the instantaneous wear rate and obtain an initial wear rate;

[0076] The feature set and the abnormal component are input into a physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate, including adding constraints to the loss function and adopting a physical constraint loss function to ensure that the data-driven model does not violate the friction-heat transfer mechanism;

[0077] The formula of the physical constraint loss function is:

[0078] ;

[0079] Where, N: total number of training samples;

[0080] ,yi: wear rate predicted and measured by the neural network of the i-th sample, unit: mm / h;

[0081] ΔEi pred : the thermal energy change of the i-th sample calculated from the predicted wear amount, unit J;

[0082] ΔEi phys : the physical thermal energy change of the i-th sample calculated according to the friction-heat transfer equation, unit J;

[0083] λ: Constraint weight, used to balance fitting accuracy and physical consistency (typical value is 0.1–1.0).

[0084] The first term, LMSE, ensures the fitting accuracy of the predicted wear rate, while the second term, Lcons, forces the output to satisfy energy conservation, significantly improving the model's physical interpretability and robustness under unknown working conditions.

[0085] S4: Time series evolution and online calibration step: The initial wear rate and the latest temperature-speed history sequence are input into the LSTM-Transformer model to predict the future wear process, and the actual wear measurement value after replacement is combined with online incremental calibration to obtain the final optimized prediction result.

[0086] S5: The prediction results and the latest collected temperature-speed history sequence are used as input to the LSTM-Transformer model to predict the future wear process and obtain a prediction sequence. The prediction sequence is subjected to dynamic threshold determination based on P-value control to generate a tool risk warning status.

[0087] S6: The warning status is used to adjust the sampling frequency and alarm level of the thermocouple and encoder to obtain the optimized acquisition strategy; the actual measured value of tool wear after replacement is compared with the predicted sequence to update the weights of the neural network and time series model online to obtain the optimized model.

[0088] Specifically, the wavelet packet decomposition of the heat flux distribution model is performed to separate the high-frequency thermal anomaly components induced by wear to obtain the abnormal components, including: using the db4 wavelet basis to decompose the heat flux distribution model into 4 levels to extract the high-frequency thermal anomaly components in the 3rd and 4th level sub-bands.

[0089] Specifically, the LSTM-Transformer model includes a bidirectional LSTM encoder and a multi-head self-attention decoder, where the number of attention heads is no less than 8, which is used to capture the long-term and short-term dependencies of the historical sequence;

[0090] The online incremental calibration adopts a sample replay strategy, in which the single weight update ratio does not exceed 5%.

[0091] Specifically, the feature set and the abnormal component are input into a physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate. The neural network structure is a three-layer fully connected network. The number of nodes in the input layer is equal to the sum of the feature set and the abnormal component dimensions. The middle layer includes energy conservation constraints, and the output layer is a single-node instant wear rate prediction.

[0092] Specifically, the dynamic threshold judgment based on P-value control of the prediction sequence is used to generate a tool risk warning state. The P-value control calculates the distribution parameters of the prediction error based on the sliding window method and dynamically adjusts the warning threshold to a 95% confidence level.

[0093] Specifically, the early warning status is used to adjust the sampling frequency and alarm level of the thermocouple and encoder, and in the optimized acquisition strategy, the sampling frequency is divided into three levels according to the early warning status: high risk ≥20 kHz, medium risk 10-20 kHz, and low risk 5-10 kHz.

[0094] Specifically, the actual measured value of tool wear after replacement is compared with the predicted sequence, which is used to update the weights of the neural network and timing model online to obtain the optimized model. The model update adopts an incremental learning strategy based on sample replay, and only 5% of the weight is updated each time calibration is performed.

[0095] As can be seen above, extracting the coupling characteristics between temperature and speed through mutual information analysis can capture implicit wear signals that are difficult to detect using a single sensor dimension, significantly improving sensitivity to subtle stages of tool wear. Compared to methods that rely solely on temperature or vibration, multi-source coupling characteristics can reveal wear trends at an early stage, providing early warning and reducing "blind spot" time.

[0096] Integrating friction-heat transfer physics equations into heat flux fitting and neural network loss functions not only closely combines empirical data-driven approaches with theoretical mechanisms, but also suppresses overfitting and physical violations during model training through energy conservation constraints. This not only enhances the robustness of the model under unknown operating conditions but also improves the interpretability of the results, facilitating on-site engineers' fault diagnosis based on physical mechanisms.

[0097] Wavelet packet multi-scale decomposition effectively isolates high-frequency thermal anomalies directly associated with local tool defects, allowing subsequent neural network input to focus on the most representative anomaly components. This process significantly suppresses low-frequency environmental disturbances (such as overall temperature drift) and structured noise (such as mechanical vibration), thereby improving the signal-to-noise ratio and accuracy of overall monitoring.

[0098] Example 2: Tool wear monitoring in soft soil conditions:

[0099] 1. Working conditions and goals

[0100] The scheme's accuracy in predicting real-time wear rates was validated in a soft, high-moisture soil layer (average moisture content 25%), with no cohesion, a standard penetration resistance (N) of approximately 2–4, and a hardness of approximately 0.5 MPa. The goal was to achieve a prediction error of less than 5% within one hour.

[0101] 2. Data Acquisition

[0102] Temperature data: An eight-point array of Omega K-type thermocouples (model TC-08) with a 20 mm spacing was attached to the side of the tool's edge to collect the time series temperature, Ts,i. The data was transmitted to the edge computing unit via an NI USB-6210 acquisition card at a 1 kHz frequency.

[0103] Speed data: A Renishaw photoelectric encoder (resolution 2 000 PPR) is used, installed at the end of the tool holder, with a sampling frequency of 5 kHz, to output the speed ωi in real time.

[0104] 3D topography data: The tool surface was rotated and scanned using a Faro Focus 3D laser scanner with a point cloud accuracy of 0.1 mm for constructing a reference geometry model.

[0105] Ambient temperature: A SHT35 temperature and humidity sensor with a sampling frequency of 1 Hz is placed near the tool to record the ambient temperature Ta.

[0106] All sensors are aggregated to the NVIDIA Jetson AGX edge computing platform via industrial Ethernet and stored in real time on a local SSD.

[0107] 3. Data Processing

[0108] Construct the correlation matrix A

[0109] Synchronously collect N=10000 {Ts,i, ωi} pairs within 10 seconds to form a matrix

[0110] A=[(Ts,1,ω1),(Ts,2,ω2),…,(Ts,N,ωN)]

[0111] Mutual information feature extraction

[0112] A sliding window with a length of 500 points is used to calculate the mutual information I(T;ω) for each window segment.

[0113] Based on the mRMR algorithm, the 8-dimensional features with the highest mutual information value are selected to form the feature set

[0114] B={b1,b2,…,b8}.

[0115] Friction-heat transfer least squares fitting

[0116] Calculate the linear velocity vi=ωi / 60×(1.2 / 2)m / s.

[0117] Using the Fit Function

[0118] qi=1500 (Ts,i−Ta)+0.4×2×vi

[0119] The least squares fitting is performed on each sampling point to obtain the heat flow series {qi}.

[0120] Wavelet packet multi-scale decomposition

[0121] Perform db4, 4-level decomposition on {qi}, retaining the 3rd and 4th level high frequency coefficients to form abnormal components

[0122] D={d1,d2,…}.

[0123] Physically constrained neural network prediction

[0124] Network structure: input layer dimension 8+2 (main statistics of B and D); two hidden layers (number of nodes 64 and 32); output layer is single-node wear rate prediction.

[0125] The loss function is shown in the following formula:

[0126] ;

[0127] An energy conservation constraint term of λ=0.5 is introduced.

[0128] Time series evolution prediction and online calibration:

[0129] The instant rate sequence { } and the past 1,000 points {Ts,i,ωi} history are written into a bidirectional LSTM + multi-head attention decoder to predict the wear curve for the next 1 hour.

[0130] After every 1 h of monitoring, the cutter teeth were manually replaced and the average wear rate was measured. The network weight was updated by replaying the samples, and no more than 3% was allowed.

[0131] Instant wear rate based on neural network After prediction, the rate sequence and the synchronously collected historical temperature Ts,i and speed ωi are input into the trained LSTM-Transformer model:

[0132] Input structure:

[0133] Time series length: 1000 points (i.e., a 10 s history window).

[0134] Each point contains three dimensions: temperature, rotation speed, and instant wear rate.

[0135] Data normalization is dynamically updated using historical mean and standard deviation.

[0136] Model structure:

[0137] LSTM: bidirectional structure, 64 nodes in a single layer.

[0138] Transformer: uses a 4-head attention mechanism, with each head dimension of 16.

[0139] Output: The predicted sequence of tool wear rate per minute for the next 60 minutes {f1,f2,…,f60}.

[0140] Dynamic threshold setting method:

[0141] Based on the residual sequence of the model prediction error in the past 2 hours, a sliding window statistical distribution is constructed to calculate the confidence interval of the wear prediction at the current time point.

[0142] The confidence level α is set to 95%, and the dynamic threshold θt is calculated by the P value control function of the prediction error.

[0143] If any point in the current or future forecast sequence exceeds the threshold θt, a risk warning state G is generated and downstream scheduling is triggered.

[0144] For example:

[0145] In the 38th minute prediction, the predicted value for the 15th minute in the future is 0.052 mm / h, which exceeds the dynamic threshold of 0.048 mm / h (based on the 95% confidence interval of the previous two hours). The system immediately marks this point as "medium risk" and records the trigger time.

[0146] Sampling scheduling mechanism: Based on the warning status generated by the previous prediction result, the system automatically adjusts the sensor sampling strategy, as shown in Table 1 below:

[0147] Table 1

[0148]

[0149] The system controls the NI module and encoder sampling rate through the CAN bus to ensure that the data granularity during high-risk periods meets the model tracking requirements.

[0150] Measured feedback and incremental learning:

[0151] After each manual maintenance or replacement of the cutter head, measure and record the actual wear amount.

[0152] Take out the historical forecast sequence {ft} corresponding to this time period and calculate the overall error.

[0153] Adopting sample replay incremental learning mechanism:

[0154] 100 input–output pairs are sampled from the time period with the maximum prediction error.

[0155] Fine-tune the existing neural network model and use a micro learning rate of 1×10 -5 , only the terminal weights are updated.

[0156] The weight update amplitude is controlled to be less than 3% of the overall model parameters to prevent "catastrophic forgetting";

[0157] 4. Performance comparison is shown in Table 2 below:

[0158] Table 2

[0159]

[0160] From the above, we can see that under the conditions of soft soil and high water content, this scheme reduces the average error by about 80% compared with the traditional single temperature method, significantly improving the early warning accuracy.

[0161] As can be seen from the above, designing the prediction model as a hybrid structure of LSTM and multi-head self-attention can simultaneously capture both short-term fluctuations and long-term trends in the wear rate. This time series modeling approach can reflect both instantaneous bursts of wear and account for the law of cumulative attenuation, ensuring highly accurate prediction results for both short-term response and long-term planning.

[0162] Dynamically adjusting the warning threshold based on the distribution of prediction errors allows for flexible switching between different reliability requirements, avoiding issues like premature warnings or missed warnings. Furthermore, linking the sensor sampling frequency with the warning level automatically improves data resolution during high-risk phases and conserves computing and transmission resources during low-risk phases, enabling efficient utilization of computing power and bandwidth by the monitoring system.

[0163] Example 3: Tool wear monitoring in hard rock conditions:

[0164] 1. Working conditions and goals

[0165] Working in hard sandstone (standard penetration resistance N value of approximately 50, hardness of 5 MPa), in order to cope with higher friction resistance and thermal loads, the robustness of this solution in extremely harsh environments was verified, with a target prediction error of <5%.

[0166] 2. Data Acquisition

[0167] Thermocouple array: OMEGA 12-point K-type, point spacing 15 mm; sampling frequency 2 kHz, transmitted via Advantech ADAM-4017 module.

[0168] Photoelectric encoder: Renishaw RGH24 series, 5 000 PPR, sampling frequency 10 kHz, combined with robot arm rotation angle compensation.

[0169] 3D laser scanning: Leica BLK360, accuracy 0.05 mm. After scanning, the point cloud was subjected to PCL (Point Cloud Library) noise reduction processing.

[0170] Environmental data collection: A PT100 platinum resistance meter is used to measure the ambient temperature Ta inside the silo, and a self-designed soil hardness sensor probe is used to obtain real-time hardness values.

[0171] 3. Data Processing

[0172] Time synchronization and correlation matrix

[0173] All device timestamps are aligned using the Precision Time Protocol (PTP).

[0174] 10,000 sets of data are collected synchronously within 5 s to form the correlation matrix A.

[0175] Coupling feature screening

[0176] The mutual information is calculated using a 200-point sliding window, and the most sensitive 10-dimensional features are selected to form the feature set.

[0177] Enhanced friction-heat transfer fitting

[0178] The formula is adjusted to:

[0179] qi=1800 (Ts,i−Ta)+0.6×4×vi+α Hi

[0180] Where Hi is the geological hardness value and α=50 W / (m²·MPa) is the empirical coupling coefficient.

[0181] The least squares fit also estimates α.

[0182] Wavelet packet decomposition and noise suppression:

[0183] In addition to db4 and 4-level decomposition, wavelet denoising (combined soft and hard thresholding) is first performed on {qi}, and then high-frequency anomalies are extracted.

[0184] Depth prediction and calibration

[0185] The physical constraint network is the same as that in Example 1, but λ is adjusted to 0.7 and the incremental learning weight update rate is 5%.

[0186] LSTM-Transformer is deployed with two layers of 128-unit LSTM, 12-head attention, and Dropout 0.2 to prevent overfitting.

[0187] After completing the instant wear prediction with geological hardness correction, a more complex multivariate prediction model is constructed using the structurally enhanced LSTM-Transformer:

[0188] Input structure:

[0189] The historical window has 2000 points, and a single point contains four dimensions: temperature, rotation speed, wear rate, and rock hardness.

[0190] All inputs are normalized and hardness is taken as a relative value (normalized to 0~1).

[0191] Model structure:

[0192] LSTM: bidirectional two-layer, 128 units.

[0193] Transformer: 12 attention heads, hidden layer dimension 64.

[0194] Output: Forecast sequence for the next 90 minutes, 1 point every 5 minutes, a total of 18 points of wear rate sequence output.

[0195] Dynamic threshold algorithm:

[0196] Add a multivariate residual control mechanism: consider the covariance of three pairs of errors: temperature-wear, speed-wear, and hardness-wear.

[0197] Construct a joint P-value model (multivariate statistics):

[0198] ;

[0199] Where Pt: the multivariate joint P value at the current time point t, used to determine whether the current wear prediction significantly deviates from the normal distribution;

[0200] : The cumulative distribution function (CDF) of the chi-square distribution with k degrees of freedom is used to calculate the position of the observation in the distribution;

[0201] k: variable dimension, which is 3 in this embodiment (corresponding to the coupling residual vector between temperature, speed, hardness and wear);

[0202] rt∈Rk: The residual vector at time point t, that is, the deviation combination between the model prediction value and the measured value:

[0203] Σ∈Rk×k: residual covariance matrix, reflecting the joint change structure between variables;

[0204] : The square of the Mahalanobis distance, which measures the degree of deviation between the current residual and the historical distribution;

[0205] Dynamically adjust the warning level (such as high, medium, and low) and automatically recommend manual review based on geological conditions;

[0206] Edge device scheduling: Use the Jetson Xavier module to implement local model execution and policy switching, ensuring that the system can still operate stably in complex and high-vibration environments.

[0207] When the prediction is "high risk", the system automatically increases the thermocouple to 5 kHz and the encoder to 10 kHz, and starts the backup acquisition channel for redundancy.

[0208] Model self-update mechanism: After each round of maintenance, the measured wear amount is uploaded as a label and added to the "rolling training cache".

[0209] Training using online pruning optimization + fine-tuning:

[0210] Leverage “layer-wise learning rates”: Freeze the underlying feature extraction layers and only fine-tune the intermediate GRU units.

[0211] Dropout (0.2) is added to prevent overfitting.

[0212] After each online update, the model's error change on the validation set is automatically evaluated to ensure that incremental updates do not introduce performance degradation.

[0213] 4. Performance comparison is shown in Table 3 below:

[0214] Table 3

[0215]

[0216] As can be seen from the above, under hard rock high-pressure friction conditions, the accuracy of this scheme is further improved after adding geological hardness coupling and noise suppression, and the prediction error is maintained at around 2.5%, meeting the real-time warning needs on site.

[0217] As can be seen from the above, through the comparative tests under the above two working conditions, it can be seen that this scheme can maintain extremely high prediction accuracy under different geological conditions, far exceeding the traditional single temperature monitoring technology, and fully demonstrates the innovative value and engineering applicability of the "speed-temperature-wear" coupling and physical constraint algorithm.

[0218] As shown above, by performing incremental learning by comparing measured wear after tool changes with predicted values, this solution can continuously self-calibrate during operation and promptly correct model deviations. This mechanism not only addresses distribution drift caused by changing geological conditions and tool upgrades, but also extends the effective life of the model, virtually eliminating the need for downtime for maintenance.

[0219] This solution integrates collaborative optimization across multiple layers, including multi-physics acquisition, physical modeling, signal processing, deep learning, and online calibration, to form a closed-loop perception-prediction-feedback system. This system maintains high stability and accuracy in complex and variable formations and operating conditions, and boasts excellent scalability and portability. It can be extended to other multi-physics sensing monitoring scenarios (such as bearing wear and motor failure), offering broad application prospects and economic benefits.

[0220] In the description of this specification, the reference terms "one embodiment", "some embodiments", "examples", "specific examples" or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0221] In the drawings of the embodiments disclosed in the present invention, only the structures related to the embodiments disclosed in the present invention are involved. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of the present invention can be combined with each other.

[0222] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The intelligent sensing method of shield tool based on speed, temperature and wear is characterized by: The following steps are involved: A thermocouple array and a photoelectric encoder are placed on the surface of the tool to synchronously collect temperature and speed timing signals and construct a preliminary correlation matrix. Performing mutual information analysis on the correlation matrix to extract temperature-speed coupling features and obtain a feature set; The feature set is combined with the 3D topography scanning data of the tool surface through least square fitting of the friction-heat transfer equation to calculate the instantaneous friction heat flux density and obtain a heat flux distribution model; the heat flux distribution model is subjected to wavelet packet decomposition to separate the high-frequency thermal anomaly components induced by wear and obtain the abnormal components; Inputting the feature set and the abnormal component into a physical constraint neural network model to predict the instantaneous wear rate and obtain an initial wear rate; The initial wear rate and the latest temperature-speed history sequence are input into the LSTM-Transformer model to predict the future wear process. The model then performs online incremental calibration based on the actual wear measured after replacement to obtain the final optimized prediction result. The prediction results and the latest collected temperature-speed history sequence are used as input to the LSTM-Transformer model to predict the future wear process and obtain a prediction sequence. The prediction sequence is then subjected to dynamic threshold determination based on P-value control to generate a tool risk warning status. The early warning status is used to adjust the sampling frequency and alarm level of the thermocouple and encoder to obtain an optimized acquisition strategy; the actual measured value of tool wear after replacement is compared with the predicted sequence to update the weights of the neural network and timing model online to obtain an optimized model.

2. The speed-temperature-wear intelligent sensing method for shield tools according to claim 1 is characterized in that: The mutual information analysis of the correlation matrix is performed to extract temperature-speed coupling features to obtain a feature set, including: calculating the mutual information between temperature and speed by a mutual information evaluation method to identify the most effective coupling amount for wear prediction from the correlation matrix, and selecting several features with the largest amount of information and the lowest redundancy based on the maximum correlation and minimum redundancy criterion; The formula of the mutual information evaluation method is: ； Where T: tool surface temperature random variable, value set T={t1,…,tM}; ω: tool speed random variable, value set Ω={ω1,…,ωN}; p(t,ω): joint probability density, representing the frequency of observations at temperature t and speed ω; p(t), p(ω): marginal probability density, which are the observation frequencies of temperature or speed, respectively.

3. The intelligent sensing method for shield tool based on speed, temperature and wear according to claim 1 is characterized in that: The feature set and the tool surface 3D topography scanning data are fitted with the least squares of the friction-heat transfer equation to calculate the instantaneous friction heat flux density, and the formula of the friction-heat transfer equation in the heat flux distribution model is obtained as follows: ； Where, q: instantaneous frictional heat flux density, that is, the frictional heat flux generated per unit area, unit W / m²; h: surface heat transfer coefficient, reflecting the conduction and convection capacity between the tool and the environment, unit W / (m²·K); Ts: Real-time temperature of tool surface, measured by thermocouple array, unit: °C; Ta: ambient temperature, which can be the temperature of surrounding rock cuttings or lining, unit: °C; μ: friction coefficient, related to tool material and rock type; p: tool-rock contact pressure, obtained by force sensor or robotic arm feedback, unit: Pa; v: Tool linear velocity, equal to the rotational speed ω multiplied by the radius, unit: m / s.

4. The intelligent sensing method for shield tool based on speed, temperature and wear according to claim 1 is characterized in that: The feature set and the abnormal component are input into a physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate, including adding constraints to the loss function and adopting a physical constraint loss function to ensure that the data-driven model does not violate the friction-heat transfer mechanism; The formula of the physical constraint loss function is: ； Where, N: total number of training samples; ,yi: wear rate predicted and measured by the neural network of the i-th sample, unit: mm / h; ΔEi pred : the thermal energy change of the i-th sample calculated from the predicted wear amount, unit J; ΔEi phys : the physical thermal energy change of the i-th sample calculated according to the friction-heat transfer equation, unit J; λ: Constraint weight, used to balance fitting accuracy and physical consistency.

5. The speed-temperature-wear intelligent sensing method for shield tools according to claim 1 is characterized in that: The heat flux distribution model is subjected to wavelet packet decomposition to separate the high-frequency thermal anomaly components induced by wear to obtain the abnormal components, including: using the db4 wavelet basis to decompose the heat flux distribution model into 4 levels to extract the high-frequency thermal anomaly components in the 3rd and 4th level sub-bands.

6. The speed-temperature-wear intelligent sensing method for shield tools according to claim 1 is characterized in that: The LSTM-Transformer model includes a bidirectional LSTM encoder and a multi-head self-attention decoder, where the number of attention heads is no less than 8, which is used to capture the long-term and short-term dependencies of the historical sequence; The online incremental calibration adopts a sample replay strategy, in which the single weight update ratio does not exceed 5%.

7. The intelligent sensing method for shield tool based on speed, temperature and wear according to claim 1 is characterized in that: The feature set and the abnormal component are input into a physical constraint neural network model to predict the instantaneous wear rate and obtain the initial wear rate. The neural network structure is a three-layer fully connected network. The number of nodes in the input layer is equal to the sum of the feature set and the abnormal component dimensions. The middle layer includes energy conservation constraints. The output layer is a single-node instant wear rate prediction.

8. The intelligent sensing method for shield tool based on speed, temperature and wear according to claim 1 is characterized in that: The dynamic threshold determination based on P-value control of the prediction sequence is used to generate a tool risk warning state. The P-value control calculates the distribution parameters of the prediction error based on the sliding window method and dynamically adjusts the warning threshold to a 95% confidence level.

9. The intelligent sensing method for shield tool based on speed, temperature and wear according to claim 1, characterized in that: The early warning status is used to adjust the sampling frequency and alarm level of the thermocouple and encoder, and an optimized acquisition strategy is obtained. The sampling frequency is divided into three levels according to the early warning status: high risk ≥20 kHz, medium risk 10-20 kHz, and low risk 5-10 kHz.

10. The intelligent sensing method for shield tool based on speed, temperature and wear according to claim 1, characterized in that: The actual measured value of tool wear after replacement is compared with the predicted sequence F, which is used to update the weights of the neural network and the timing model online. In the optimized model, the model update adopts an incremental learning strategy based on sample replay, and only 5% of the weight is updated each time calibration is performed.

Images