Intelligent centerless grinding machine monitoring and regulating system with heat-force-vibration cooperative suppression function

Abstract

The invention discloses a heat-force-vibration cooperative inhibition centerless grinding machine intelligent monitoring and regulation system, and relates to the technical field of monitoring regulation, a multi-source sensor is used for collecting heat, force and vibration signals of a centerless grinding machine in real time, a sliding time window is used for synchronously segmenting the signals, and multi-modal features are extracted; calculating a complementary enhancement coefficient and a time coupling coefficient based on the features, comparing the current fusion feature with historical training data distribution, and calculating a drift metric; the fusion effect is judged by integrating the indexes, if the fusion effect is qualified, real-time intelligent monitoring and regulation are conducted on the centerless grinding machine according to the prediction result of the heat-force-vibration collaborative model, and stable and efficient operation of the system is ensured; in this way, an evaluation mechanism of fusion quality can be directly judged without depending on the final prediction effect of the heat-force-vibration collaborative model, so that the fusion failure risk is accurately recognized, whether re-fusion is needed or not is guided, and the adaptive capacity and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring and control, and in particular to an intelligent monitoring and control system for a centerless grinder with heat-force-vibration coordinated suppression. Background Art

[0002] In the machining of high-precision parts, centerless grinders are widely used for outer diameter grinding of shafts and sleeves due to their clamping-free nature and suitability for high-volume continuous processing. However, the grinding process is inherently a complex multi-physics coupling process involving thermal effects (heat accumulation in the workpiece and grinding wheel), mechanical loads (dynamic fluctuations in grinding forces), and structural vibrations (micro-displacements and resonances between the grinding wheel and workpiece). The interaction and enhanced coupling of these factors can easily lead to quality issues such as machining burns, chatter marks, and dimensional drift.

[0003] To this end, in recent years, an intelligent monitoring and control technology for centerless grinders with collaborative suppression of heat, force, and vibration has been proposed. Its core goal is to synchronously collect three types of physical signals, namely heat, force, and vibration, through multi-source sensors, and to build a heat-force-vibration coupling relationship with the help of a multimodal data fusion model. Based on this, advance prediction and judgment of key states (such as burn precursors, excessive vibration, thermal deformation trends, etc.) can be achieved. According to the prediction results, the servo drive, hydraulic and cooling units can be linked to achieve real-time control to prevent quality problems and ensure processing stability and accuracy. At present, most actual deployments use deep learning or data-driven heat-force-vibration collaborative models for feature extraction and state prediction. The model fuses the three types of modal information, namely heat, force, and vibration, to jointly predict and judge the state of the centerless grinder and control the centerless grinder in advance.

[0004] Therefore, in the thermal-mechanical-vibration collaborative model, the fusion effect of the three types of modal information of thermal-mechanical-vibration directly determines the subsequent state prediction and control of the centerless grinder. Once the fusion effect of the three types of modal information of thermal-mechanical-vibration has problems, it may cause problems in the subsequent state prediction and control of the centerless grinder.

[0005] Existing methods usually rely solely on the model's prediction results (such as accuracy, false alarm rate, etc.) to indirectly judge the quality of thermal-mechanical-vibration data fusion. This approach has two shortcomings. First, a decline in model prediction performance does not necessarily mean that the fusion has failed. The deterioration of model results may be caused by a variety of factors, such as changes in working conditions, sensor drift, data quality fluctuations, sample distribution shifts, etc., and does not necessarily mean that there is a problem with the fusion quality itself. Second, the cost of retraining the model is high. Once the prediction results are considered "unsatisfactory", existing methods often directly choose to retrain the fusion model, but this usually requires a large amount of new data, manual labeling, and redeployment, resulting in high costs, long cycles, and low efficiency. Therefore, it is urgent to build an evaluation mechanism that does not rely on the final prediction results and can directly judge the quality of fusion. This can accurately identify the risk of fusion failure, guide whether re-fusion is necessary, and improve the adaptive ability and stability. Summary of the Invention

[0006] The purpose of the present invention is to solve the above-mentioned problems and provide an intelligent monitoring and control system for centerless grinders with coordinated thermal, mechanical and vibration suppression.

[0007] In a first aspect of the present invention, an intelligent monitoring and control system for a centerless grinder with thermal, mechanical and vibration coordinated suppression is proposed, the system comprising:

[0008] Data acquisition module: collects thermal signals, force signals and vibration signals in the centerless grinding process in real time through multi-source sensors;

[0009] Partitioning module: Synchronously segment the three types of signals using a sliding time window and extract the corresponding thermal-mechanical-vibration multimodal features within each window;

[0010] Complementary enhancement module: Based on the extracted thermal-mechanical-vibration multimodal features, the complementary enhancement coefficient of the multimodal features is calculated to evaluate the mutual information strength and redundancy between different modes;

[0011] Temporal Coupling Module: This module performs correlation analysis on time-aligned multimodal features and calculates the temporal coupling coefficient of multimodal features to measure the synchronization and coordinated change trends between different modalities.

[0012] Feature distribution drift module: compares the multimodal fusion features in the current time window with the fusion feature distribution of historical training data, and calculates the feature distribution drift metric;

[0013] Monitoring and control module: The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the complementary enhancement coefficient, time coupling coefficient and characteristic distribution drift measurement. If the data fusion effect is qualified, the centerless grinder is monitored and controlled according to the results predicted by the thermal-mechanical-vibration collaborative model.

[0014] Optionally, the steps for calculating the complementary enhancement coefficient of the multimodal features are:

[0015] For the sliding window t, construct three types of feature vectors: thermal, force, and vibration: thermal feature vector T t ∈R dT ; Force eigenvector F t ∈R dF ; Vibration eigenvector V t ∈R dV ;

[0016] The mutual information between two modes is calculated using non-parametric estimation methods, which are: I TF =MI(T t ,F t );I TV =MI(T t ,V t );I FV =MI(F t ,V t );I TF , I TV and I FV They represent the mutual information between thermal features and force features, the mutual information between thermal features and vibration features, and the mutual information between force features and vibration features respectively;

[0017] Concatenate the three modal feature vectors into the fusion feature Z t =[T t ,F t ,V t ]∈R dT+dF+dV ; Calculate Z based on kernel density estimation t The entropy H(Z t ), the calculation formula is: Where K(·,·) is the kernel function (such as Gaussian kernel), and N is the total number of time windows;

[0018] Calculate the complementary enhancement coefficient TG of the multimodal features. The calculation formula is:

[0019] Optionally, the steps for calculating the time coupling coefficient of the multimodal feature are:

[0020] The real-time collected thermal, force, and vibration signals are synchronously segmented according to the sliding time window t∈{1, 2, ..., T}. The length of each time window is set to Δt, so that the modal feature sequences of the three types of signals can be obtained in the t-th window. The thermal signal feature sequence is: Force signal characteristic sequence: Vibration signal characteristic sequence: N is the total number of time windows;

[0021] Calculate the first-order difference series of each modal signal in the time series: thermal modal difference: Force modal difference: Vibration modal difference:

[0022] The differential sequences of the three modes are compared pairwise, and the dynamic time warping distance is used to measure the degree of temporal coupling. The three groups of DTW distances are: D HF =DTW(ΔH t , ΔF t );D HV =DTW(ΔH t , ΔV t );D FV =DTW(ΔFt,ΔV t );

[0023] Calculate the time coupling coefficient YH of the multimodal feature. The calculation formula is:

[0024] Optionally, the calculation steps of the feature distribution drift metric are:

[0025] In each sliding time window, the fused multimodal features are extracted from the thermal, force, and vibration signals as the current real-time feature set;

[0026] Extract fusion features from historical data saved during the model training phase as a reference feature set;

[0027] The basic statistics such as mean and covariance of the current feature set and the reference feature set are counted respectively to characterize their probability distribution;

[0028] The KL divergence is used to calculate the relative entropy difference between the current feature distribution and the reference feature distribution as a measure of feature distribution drift.

[0029] Optionally, whether the data fusion effect of the current thermal-mechanical-vibration synergistic model is qualified is judged based on the complementary enhancement coefficient, the time coupling coefficient, and the characteristic distribution drift metric. If the data fusion effect is qualified, the steps of monitoring and controlling the centerless grinder based on the results predicted by the thermal-mechanical-vibration synergistic model are as follows:

[0030] The fusion qualification coefficient is calculated based on the complementary enhancement coefficient, time coupling coefficient and feature distribution drift metric, and the fusion qualification coefficient is compared with the preset fusion qualification coefficient threshold. The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the comparison result.

[0031] Optionally, the step of calculating the fusion qualification coefficient according to the complementary enhancement coefficient, the time coupling coefficient and the characteristic distribution drift metric is:

[0032] The complementary enhancement coefficient, temporal coupling coefficient and feature distribution drift metric are de-dimensionalized and normalized so that they are all between [0, 1]. The normalized complementary enhancement coefficient, temporal coupling coefficient and feature distribution drift metric are weighted and summed to obtain the fusion qualification coefficient. The calculation formula is: Where gh is the fusion qualification coefficient, xc, xb, and xm are the complementary enhancement coefficient, temporal coupling coefficient, and feature distribution drift metric after normalization, respectively.

[0033] Optionally, the steps of comparing the fusion qualification coefficient with a preset fusion qualification coefficient threshold and judging whether the data fusion effect of the current thermal-mechanical-vibration synergistic model is qualified according to the comparison result are:

[0034] The fusion qualification coefficient is compared with the preset fusion qualification coefficient threshold. If the fusion qualification coefficient is not less than the preset fusion qualification coefficient threshold, it means that the data fusion effect of the current thermal-mechanical-vibration collaborative model is qualified. The centerless grinder is then monitored and controlled according to the results predicted by the thermal-mechanical-vibration collaborative model.

[0035] If the fusion qualification coefficient is less than the preset fusion qualification coefficient threshold, it means that the data fusion effect of the current thermal-mechanical-vibration collaborative model is unqualified, and the monitoring and control operations will no longer rely on the prediction results of the current thermal-mechanical-vibration collaborative model. At this time, the thermal-mechanical-vibration collaborative model can be automatically updated until the fusion qualification coefficient is not less than the preset fusion qualification coefficient threshold.

[0036] Beneficial effects of the present invention:

[0037] The present invention proposes an intelligent monitoring and control system for a centerless grinder with thermal-mechanical-vibration collaborative suppression. The system collects the thermal, force and vibration signals of the centerless grinder in real time through multi-source sensors, and uses a sliding time window to synchronously segment the signals and extract multimodal features. Based on these features, the complementary enhancement coefficient is calculated to evaluate the information complementarity between modalities. The time coupling coefficient is calculated through correlation analysis to measure the modal synchronization. The feature distribution drift module compares the current fusion feature with the historical training data distribution and calculates the drift metric. The monitoring and control module judges the fusion effect based on the above indicators. If it is qualified, the centerless grinder is intelligently monitored and controlled in real time according to the prediction results of the thermal-mechanical-vibration collaborative model to ensure stable and efficient operation of the system. In this way, an evaluation mechanism for directly judging the quality of fusion can be established without relying on the final prediction effect of the thermal-mechanical-vibration collaborative model, so as to accurately identify the risk of fusion failure and guide whether re-fusion is needed, thereby improving the adaptability and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings.

[0039] Figure 1Framework diagram of the intelligent monitoring and control system for centerless grinders with coordinated thermal, mechanical and vibration suppression. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe 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.

[0041] The embodiment of the present invention provides an intelligent monitoring and control system for a centerless grinder with thermal, mechanical and vibration coordinated suppression. Figure 1 , Figure 1 A framework diagram of an intelligent monitoring and control system for a centerless grinder with thermal, mechanical, and vibration coordinated suppression provided by an embodiment of the present invention. The system includes the following steps:

[0042] Data acquisition module: collects thermal signals, force signals and vibration signals in the centerless grinding process in real time through multi-source sensors;

[0043] Partitioning module: Synchronously segment the three types of signals using a sliding time window and extract the corresponding thermal-mechanical-vibration multimodal features within each window;

[0044] Complementary enhancement module: Based on the extracted thermal-mechanical-vibration multimodal features, the complementary enhancement coefficient of the multimodal features is calculated to evaluate the mutual information strength and redundancy between different modes;

[0045] Temporal Coupling Module: This module performs correlation analysis on time-aligned multimodal features and calculates the temporal coupling coefficient of multimodal features to measure the synchronization and coordinated change trends between different modalities.

[0046] Feature distribution drift module: compares the multimodal fusion features in the current time window with the fusion feature distribution of historical training data, and calculates the feature distribution drift metric;

[0047] Monitoring and control module: The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the complementary enhancement coefficient, time coupling coefficient and characteristic distribution drift measurement. If the data fusion effect is qualified, the centerless grinder is monitored and controlled according to the results predicted by the thermal-mechanical-vibration collaborative model.

[0048] The intelligent monitoring and control system for centerless grinders with thermal-mechanical-vibration collaborative suppression provided by the embodiment of the present invention can, through the above-mentioned method, directly judge the evaluation mechanism of the fusion quality without relying on the final prediction effect of the thermal-mechanical-vibration collaborative model, so as to accurately identify the risk of fusion failure, guide whether re-fusion is needed, and improve the adaptability and stability.

[0049] In one embodiment, multi-source sensors are used to collect thermal, force, and vibration signals in real time during the centerless grinding process. The three types of signals are synchronously segmented using a sliding time window, and the corresponding thermal-force-vibration multimodal features are extracted within each window.

[0050] During the centerless grinding process, multiple high-sensitivity sensors are deployed to achieve comprehensive perception and intelligent analysis of the three physical quantities of heat, force, and vibration:

[0051] Thermal signal acquisition: Infrared temperature sensors or thermocouples are placed near the grinding wheel contact area to obtain real-time temperature changes on the workpiece surface and grinding wheel area during the grinding process. The sampling frequency is generally set at 50-100Hz to balance response speed and signal stability.

[0052] Force signal acquisition: Strain gauges or multi-dimensional force sensors are deployed in the centerless grinder's support mechanism or at the grinding wheel spindle loading area to monitor cutting forces, normal forces, and other factors generated during the grinding process, capturing force fluctuations caused by changes in machining conditions. The force signal sampling rate is typically no less than 1kHz to ensure responsiveness to short-term fluctuations.

[0053] Vibration signal acquisition: High-frequency accelerometers are installed on key components of the grinding machine, such as the spindle box, bracket, and support wheel seat, to capture micro-vibration data during the grinding process. To ensure the capture of high-frequency characteristics such as chatter and impact, the vibration signal sampling rate is generally 5–20 kHz, and a high-pass filter is used to remove low-frequency structure-borne noise.

[0054] After completing the acquisition of the above multimodal data, in order to ensure the alignment and linkage analysis of the three types of signals in the time domain, a sliding time window mechanism is introduced for synchronous processing:

[0055] Sliding time window design: Align all modal signals to a unified time reference and construct sliding time windows of fixed length (e.g., 1 second with a 50% overlap). The window setting should comprehensively consider the sampling rates and physical response delays of different signals to ensure that the features of different modalities within the same window are comparable and interconnected.

[0056] Synchronous segmentation and feature extraction: Within each sliding window, the three types of modal signals are preprocessed (such as denoising and normalization) and then features are extracted:

[0057] Extract temperature change trend indicators such as mean, slope, maximum temperature rise rate, and instantaneous volatility from thermal signals;

[0058] Calculate the mean, root mean square, frequency energy distribution, force mutation slope and other indicators of the force signal to reflect the grinding state and material removal behavior;

[0059] The time domain features (such as RMS, kurtosis, and kurtosis), frequency domain features (main frequency band energy, spectrum center of gravity), and time-frequency domain features (such as wavelet packet energy ratio) are extracted from the vibration signal to characterize the indicators of processing stability.

[0060] Multimodal feature fusion construction: Finally, the features from the three types of signals, thermal, force, and vibration, in each time window are spliced ​​in chronological order to form a set of multimodal collaborative feature vectors for subsequent fusion quality assessment, anomaly prediction, or model update mechanism.

[0061] It is important to note that this step ensures the temporal alignment and physical correlation of multiple modal information, providing a solid data foundation for subsequent complementary enhancement analysis, coupling analysis, and fusion degradation identification. The sliding window mechanism not only improves the temporal representation of features but also enhances the model's response to transient anomalies and trend changes, making it a core data preprocessing step for intelligent monitoring.

[0062] In one embodiment, based on the extracted thermal-mechanical-vibration multimodal features, a complementary enhancement coefficient of the multimodal features is calculated to evaluate the mutual information strength and redundancy between different modalities;

[0063] Specifically, the calculation steps of the complementary enhancement coefficient of multimodal features are:

[0064] For the sliding window t, construct three types of feature vectors: thermal, force, and vibration: thermal feature vector T t ∈R dT ; Force eigenvector F t ∈R dF ; Vibration eigenvector V t ∈R dV ;

[0065] The mutual information between two modes is calculated using non-parametric estimation methods (such as k-nearest neighbor estimation or kernel density estimation), which are: I TF =MI(T t ,F t );I TV =MI(T t ,V t );I fV =MI(E t ,V t );I TF , I TV and I FV They represent the mutual information between thermal features and force features, the mutual information between thermal features and vibration features, and the mutual information between force features and vibration features respectively;

[0066] Concatenate the three modal feature vectors into the fusion feature Z t =[Tt ,F t ,V t ]∈R dT+dF+dV ; Calculate Z based on kernel density estimation (such as Renyi entropy approximation) t The entropy H(Z t ), the calculation formula is: Where K(·,·) is the kernel function (such as Gaussian kernel), and N is the total number of time windows;

[0067] Calculate the complementary enhancement coefficient TG of the multimodal features. The calculation formula is:

[0068] It should be noted that the complementary enhancement coefficient (CEC) of modal features reflects the degree of effective fusion of complementary information between different modal features. Essentially, it measures whether the fusion of multimodal features yields greater information gain than the individual modalities acting alone. In its calculation, the CEC uses the pairwise mutual information between the three modalities (thermal, mechanical, and vibration) as the denominator, representing the original redundancy and shared information between the modalities. The overall information entropy after the fusion of the three modalities is then used as the numerator to measure the overall information contained in the fusion. A larger CEC indicates a significant increase in the information entropy of the fused feature space, indicating that the fusion process effectively integrates the unique and complementary features of multiple modalities rather than simply stacking signals in a repetitive or redundant manner. Conversely, a smaller CEC indicates strong redundancy between the three signals, resulting in no significant information gain after fusion, and may even lead to "modal conflict" or fusion failure. Especially in complex nonlinear dynamics such as centerless grinding machines, where the thermal, mechanical, and vibration signals have distinct mechanisms of origin, a model can only achieve high discriminative power if they exhibit sufficient complementarity and can synergistically describe the grinding state. Therefore, the higher the complementary enhancement coefficient value, the more effectively the fusion process captures the key differential information in different modalities, and the information richness and discrimination ability are enhanced, indicating that the current fusion effect is good and the model has strong monitoring and control capabilities within this window; on the contrary, if the complementary enhancement coefficient continues to be low, it may indicate that there is redundancy in the thermal-mechanical-vibration fusion features or the fusion fails, and the fusion strategy needs to be reviewed or the model structure needs to be updated.

[0069] It is important to note that analyzing the complementary enhancement coefficient of modal features has the following benefits for determining the effectiveness of the fusion of the three modal information types: thermal, mechanical, and vibration, in a thermal-mechanical-vibration collaborative model: it directly quantifies the increase in information entropy after fusion. Compared to relying solely on model predictions, it can better reveal the true complementarity and information gain between multimodal features, avoiding attributing prediction errors to fusion failure. Secondly, it provides an immediate feedback mechanism for fusion quality, enabling early detection of potential risks before fusion performance deteriorates. This helps dynamically optimize fusion strategies or trigger adaptive re-fusion processes to improve robustness. Thirdly, this coefficient is model-independent and can be used to evaluate the pros and cons of different fusion schemes across models and tasks, providing a stable and highly interpretable fusion quality reference metric for subsequent decision-making modules. In summary, the complementary enhancement coefficient, as a fusion quality assessment tool, helps determine the reliability of fusion mechanisms at the underlying feature level, supporting the evolution of centerless grinding machine monitoring towards intelligent, self-diagnostic, and self-recovery capabilities.

[0070] In one embodiment, correlation analysis is performed on the time-aligned multimodal features to calculate the temporal coupling coefficient of the multimodal features to measure the synchronization and coordinated change trend between different modalities;

[0071] Among them, the calculation steps of the time coupling coefficient of the multimodal feature are:

[0072] The real-time collected thermal, force, and vibration signals are synchronously segmented according to the sliding time window t∈{1, 2, ..., T}. The length of each time window is set to Δt, so that the modal feature sequences of the three types of signals can be obtained in the t-th window. The thermal signal feature sequence is: Force signal characteristic sequence: Vibration signal characteristic sequence: N is the total number of time windows;

[0073] Calculate the first-order difference series of each modal signal in the time series: thermal modal difference: Force modal difference: Vibration modal difference:

[0074] The differential sequences of the three modes are compared pairwise, and the dynamic time warping distance is used to measure the degree of temporal coupling. The three groups of DTW distances are: D HF =DTW(ΔH t , ΔF t );D HV =DTW(ΔH t , ΔV t );D FV =DTW(ΔF t , ΔV t );

[0075] Calculate the time coupling coefficient YH of the multimodal feature. The calculation formula is:

[0076] It should be noted that the temporal coupling coefficient (TCC) of multimodal features is an indicator used to measure the temporal synchronization and synergistic trends of the three modal features—thermal, mechanical, and vibrational—in the temporal dimension. Essentially, it reflects the strength of temporal coupling between different modalities during dynamic processes. A larger TCC indicates more consistent variation patterns among the three modalities, meaning they exhibit high consistency and linkage within the same time window. This indicates strong coupling between the physical processes of the modalities, and the fused features are more representative of the actual grinding conditions. Conversely, a smaller TCC indicates divergent and inconsistent variation trends between the modalities. This may be due to interference with certain modalities, sensor distortion, or failure of the coupling mechanism, resulting in the fused features losing their global representativeness. Therefore, TCC can directly reflect the quality of fusion from the underlying temporal structure without relying on model output. It is particularly suitable for monitoring the stability and adaptability of thermal-mechanical-vibration model fusion during long-term operation, helping to determine whether the current fusion mechanism remains effective and whether model reconstruction or adjustment is necessary to maintain high accuracy and robustness. This indicator has the advantages of strong physical interpretability, real-time evaluation, and high model independence, and is a key structural reference parameter in fusion monitoring.

[0077] It is important to note that analyzing the temporal coupling coefficient of multimodal features is beneficial for determining the effectiveness of the fusion of the three modal information types within the thermal-mechanical-vibration synergistic model: it can reveal the synergistic response relationships between different modal signals during the grinding process from a temporal dynamic perspective, thereby enabling a structural diagnosis of fusion quality. Traditional fusion evaluation often relies on final prediction performance, making it difficult to trace the root cause of the problem. However, the temporal coupling coefficient, independent of model output, quantitatively analyzes the temporal consistency of the signals themselves, providing greater independence and foresight. When thermal, mechanical, and vibration signals exhibit a high degree of coupling within the same window, it indicates that the underlying physical mechanisms are interconnected. The fused features share a consistent dynamic structural foundation, helping to enhance the model's ability to perceive and interpret changing working conditions. Therefore, the temporal coupling coefficient not only monitors early signs of fusion failure in real time but also provides a basis for subsequent reconstruction of fusion paths or weight adjustments, significantly improving the adaptability and robustness of the thermal-mechanical-vibration synergistic model.

[0078] In one embodiment, the multimodal fusion features in the current time window are compared with the fusion feature distribution of historical training data to calculate the feature distribution drift metric;

[0079] Among them, the calculation steps of the feature distribution drift metric are:

[0080] In each sliding time window, the fused multimodal features are extracted from the thermal, force, and vibration signals as the current real-time feature set;

[0081] Extract fusion features from the historical data saved during the model training phase as the reference feature set (i.e., the “feature distribution under the original working conditions”);

[0082] The basic statistics such as mean and covariance of the current feature set and the reference feature set are counted respectively to characterize their probability distribution;

[0083] The KL divergence is used to calculate the relative entropy difference between the current feature distribution and the reference feature distribution as a measure of feature distribution drift.

[0084] It should be noted that the feature distribution drift metric refers to the degree of difference in probability distribution between the currently collected fusion features and the reference features used during model training. Essentially, it serves as a measure of whether statistical patterns in the feature space have significantly changed. Methods such as the Kullback-Leibler divergence (KL divergence) are commonly used to measure the relative entropy difference between the two distributions. This metric assesses whether the multimodal fusion features under the current operating conditions remain consistent with those under the original training conditions, reflecting the applicability and reliability of the model fusion pipeline under new operating conditions. A small feature distribution drift metric indicates that the current fusion features are highly consistent with the historical reference distribution, indicating that the fusion relationship and expression patterns of the thermal, mechanical, and vibration modalities have not substantially shifted, the fusion mapping remains valid, the information representation fully preserves the original semantics, and the model can stably predict and control based on the fusion features. Conversely, a large drift metric indicates that the fusion features are statistically disconnected from the original data, weakening the fusion effect. Problems such as increased redundancy or information loss may occur, making it difficult for the collaborative model to accurately perceive the actual operating conditions. Therefore, a smaller feature distribution drift metric is a manifestation of good fusion quality and an important reference for judging the stability and generalization ability of the thermal-mechanical-vibration collaborative model.

[0085] In one implementation,

[0086] It should be noted that the feature distribution drift metric is crucial for determining the effectiveness of multimodal information fusion in thermal-mechanical-vibration collaborative models. Its advantage lies in providing an objective quantitative measure independent of model predictions, directly reflecting the degree of match between the current actual operating conditions and the training conditions in the fusion feature space. Compared to relying on output error for indirect judgment, the drift metric can more sensitively detect potential mismatches caused by operating condition changes, sensor errors, or signal anomalies during modal fusion earlier. This helps proactively intervene before model prediction accuracy degrades, enabling model updates or fusion reconstruction, thereby improving the sensitivity and response speed to changes in the grinding machine's operating state. Furthermore, drift detection can be used to establish a model self-calibration mechanism, enabling online adaptive adjustment, reducing the risk of false positives and false negatives, and improving robustness and stability. Therefore, the introduction of the feature distribution drift metric provides a fundamental basis for fusion quality assessment and is of great value in improving the generalization capability and long-term effectiveness of the thermal-mechanical-vibration collaborative model.

[0087] In one embodiment, the data fusion effect of the current thermal-mechanical-vibration synergistic model is determined to be qualified based on the complementary enhancement coefficient, the time coupling coefficient, and the characteristic distribution drift metric. If the data fusion effect is qualified, the steps of monitoring and controlling the centerless grinder based on the results predicted by the thermal-mechanical-vibration synergistic model are as follows:

[0088] The fusion qualification coefficient is calculated based on the complementary enhancement coefficient, time coupling coefficient and feature distribution drift metric, and the fusion qualification coefficient is compared with the preset fusion qualification coefficient threshold. The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the comparison result.

[0089] In one embodiment, the steps of calculating the fusion qualification coefficient according to the complementary enhancement coefficient, the temporal coupling coefficient and the characteristic distribution drift metric are as follows:

[0090] The complementary enhancement coefficient, temporal coupling coefficient and feature distribution drift metric are de-dimensionalized and normalized so that they are all between [0, 1]. The normalized complementary enhancement coefficient, temporal coupling coefficient and feature distribution drift metric are weighted and summed to obtain the fusion qualification coefficient. The calculation formula is: Where gh is the fusion qualification coefficient, xc, xb, and xm are the complementary enhancement coefficient, temporal coupling coefficient, and feature distribution drift metric after normalization, respectively.

[0091] It should be noted that, in general, the weight values ​​of the normalized complementary enhancement coefficient, the temporal coupling coefficient, and the characteristic distribution drift metric are equal in weighted summation, and the sum of the weight values ​​is 1.

[0092] In one embodiment, the steps of comparing the fusion qualification coefficient with a preset fusion qualification coefficient threshold and judging whether the data fusion effect of the current thermal-mechanical-vibration synergistic model is qualified according to the comparison result are as follows:

[0093] The fusion qualification coefficient is compared with the preset fusion qualification coefficient threshold. If the fusion qualification coefficient is not less than the preset fusion qualification coefficient threshold, it means that the data fusion effect of the current thermal-mechanical-vibration collaborative model is qualified. The centerless grinder is then monitored and controlled according to the results predicted by the thermal-mechanical-vibration collaborative model.

[0094] If the fusion qualification coefficient is less than the preset fusion qualification coefficient threshold, it means that the data fusion effect of the current thermal-mechanical-vibration collaborative model is unqualified, and the monitoring and control operations will no longer rely on the prediction results of the current thermal-mechanical-vibration collaborative model. At this time, the thermal-mechanical-vibration collaborative model can be automatically updated until the fusion qualification coefficient is not less than the preset fusion qualification coefficient threshold.

[0095] It should be noted that comparing the fusion qualification coefficient with the preset fusion qualification coefficient threshold is a key step in judging the fusion quality of the current thermal-mechanical-vibration collaborative model. This method achieves quantitative judgment of the model fusion status by setting a reasonable fusion effect benchmark value (i.e., the fusion qualification coefficient threshold). If the currently calculated fusion qualification coefficient (obtained by a comprehensive evaluation of the complementary enhancement coefficient, the time coupling coefficient, and the characteristic distribution drift metric) is not less than the threshold, it indicates that the multimodal information has good complementarity and temporal synergy, and the current working conditions are basically consistent with the training data distribution. The data fusion effect is stable and reliable, and the prediction results of the thermal-mechanical-vibration collaborative model can continue to be trusted, and the centerless grinder can be effectively monitored and controlled accordingly. On the contrary, if the fusion qualification coefficient is lower than the set threshold, it indicates that the modal fusion may have problems such as information loss, asynchrony, mismatch, or distribution drift, and the fusion quality is no longer reliable. At this point, the prediction results of the current model will no longer be relied upon to prevent serious consequences such as misjudgments, false alarms, and incorrect regulation. At the same time, the model fusion update process will be automatically triggered to perform incremental corrections, modal reconstruction, or feature screening on the fusion structure, and continuously monitor the new fusion effect until the fusion qualification coefficient rises above the set threshold. Through this mechanism, the risks caused by blindly relying on predictive performance can be effectively avoided, while avoiding the cost burden of frequent and full retraining. It achieves intelligent regulation with fusion quality perception, adaptive update rhythm, and controllable risks, effectively enhancing the robustness, economy, and self-healing ability of thermal-mechanical-vibration monitoring of centerless grinders in long-term operation.

[0096] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. The intelligent monitoring and control system for centerless grinders with coordinated suppression of heat, force and vibration is characterized by: The system comprises: Data acquisition module: collects thermal signals, force signals and vibration signals in the centerless grinding process in real time through multi-source sensors; Partitioning module: Synchronously segment the three types of signals using a sliding time window and extract the corresponding thermal-mechanical-vibration multimodal features within each window; Complementary enhancement module: Based on the extracted thermal-mechanical-vibration multimodal features, the complementary enhancement coefficient of the multimodal features is calculated to evaluate the mutual information strength and redundancy between different modes; Temporal Coupling Module: This module performs correlation analysis on time-aligned multimodal features and calculates the temporal coupling coefficient of multimodal features to measure the synchronization and coordinated change trends between different modalities. Feature distribution drift module: compares the multimodal fusion features in the current time window with the fusion feature distribution of historical training data, and calculates the feature distribution drift metric; Monitoring and control module: The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the complementary enhancement coefficient, time coupling coefficient and characteristic distribution drift measurement. If the data fusion effect is qualified, the centerless grinder is monitored and controlled according to the results predicted by the thermal-mechanical-vibration collaborative model.

2. The intelligent monitoring and control system for centerless grinding machines with coordinated thermal-mechanical-vibration suppression according to claim 1 is characterized in that: The calculation steps of the complementary enhancement coefficient of multimodal features are: For the sliding window t, construct three types of feature vectors: thermal, force, and vibration: thermal feature vector T t ∈R dT ; Force eigenvector F t ∈R dF ; Vibration eigenvector V t ∈R dV ; The mutual information between two modes is calculated using non-parametric estimation methods, which are: I TF =MI(T t ,F t );I TV =MI(T t ,V t );I FV =MI(F t ,V t );I TF , I TV and I FV They represent the mutual information between thermal features and force features, the mutual information between thermal features and vibration features, and the mutual information between force features and vibration features respectively; Concatenate the three modal feature vectors into the fusion feature Z t =[t t ,F t ,V t ]∈R dT+dF+dV ; Calculate Z based on kernel density estimation t The entropy H(Z t ), the calculation formula is: Where K(·,·) is the kernel function and N is the total number of time windows; Calculate the complementary enhancement coefficient TG of the multimodal features. The calculation formula is:

3. The intelligent monitoring and control system for centerless grinding machines with coordinated thermal, mechanical and vibration suppression according to claim 1 is characterized in that: The calculation steps of the time coupling coefficient of multimodal features are: The real-time collected thermal, force, and vibration signals are synchronously segmented according to the sliding time window t∈{1, 2, ..., T}. The length of each time window is set to Δt, so that the modal feature sequences of the three types of signals can be obtained in the t-th window. The thermal signal feature sequence is: Force signal characteristic sequence: Vibration signal characteristic sequence: N is the total number of time windows; Calculate the first-order difference series of each modal signal in the time series: thermal modal difference: Force modal difference: Vibration modal difference: The differential sequences of the three modes are compared pairwise, and the dynamic time warping distance is used to measure the degree of temporal coupling. The three groups of DTW distances are: D HF =DTW(ΔH t , ΔF t );D HV =DTW(ΔH t , ΔV t );D FV =DTW(ΔF t , ΔV t ); Calculate the time coupling coefficient YH of the multimodal feature. The calculation formula is:

4. The intelligent monitoring and control system for centerless grinding machines with coordinated thermal, mechanical and vibration suppression according to claim 1 is characterized in that: The calculation steps of the feature distribution drift metric are: In each sliding time window, the fused multimodal features are extracted from the thermal, force, and vibration signals as the current real-time feature set; Extract fusion features from historical data saved during the model training phase as a reference feature set; The basic statistics of mean and covariance of the current feature set and the reference feature set are respectively counted to characterize their probability distribution; The KL divergence is used to calculate the relative entropy difference between the current feature distribution and the reference feature distribution as a measure of feature distribution drift.

5. The intelligent monitoring and control system for centerless grinding machines with coordinated thermal-mechanical-vibration suppression according to claim 1 is characterized in that: The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the complementary enhancement coefficient, time coupling coefficient and characteristic distribution drift metric. If the data fusion effect is qualified, the steps for monitoring and controlling the centerless grinder based on the results predicted by the thermal-mechanical-vibration collaborative model are as follows: The fusion qualification coefficient is calculated based on the complementary enhancement coefficient, time coupling coefficient and feature distribution drift metric, and the fusion qualification coefficient is compared with the preset fusion qualification coefficient threshold. The data fusion effect of the current thermal-mechanical-vibration collaborative model is judged to be qualified based on the comparison result.

6. The intelligent monitoring and control system for centerless grinding machines with coordinated thermal-mechanical-vibration suppression according to claim 5 is characterized in that: The steps for calculating the fusion qualification coefficient based on the complementary enhancement coefficient, the temporal coupling coefficient and the feature distribution drift metric are as follows: The complementary enhancement coefficient, temporal coupling coefficient and feature distribution drift metric are de-dimensionalized and normalized so that they are all between [0, 1]. The normalized complementary enhancement coefficient, temporal coupling coefficient and feature distribution drift metric are weighted and summed to obtain the fusion qualification coefficient. The calculation formula is: Where gh is the fusion qualification coefficient, xc, xb, and xm are the complementary enhancement coefficient, temporal coupling coefficient, and feature distribution drift metric after normalization, respectively.

7. The intelligent monitoring and control system for centerless grinding machines with coordinated thermal, mechanical and vibration suppression according to claim 1 is characterized in that: The steps of comparing the fusion qualification coefficient with the preset fusion qualification coefficient threshold and judging whether the data fusion effect of the current thermal-mechanical-vibration collaborative model is qualified according to the comparison result are as follows: The fusion qualification coefficient is compared with the preset fusion qualification coefficient threshold. If the fusion qualification coefficient is not less than the preset fusion qualification coefficient threshold, it means that the data fusion effect of the current thermal-mechanical-vibration collaborative model is qualified. The centerless grinder is then monitored and controlled according to the results predicted by the thermal-mechanical-vibration collaborative model. If the fusion qualification coefficient is less than the preset fusion qualification coefficient threshold, it means that the data fusion effect of the current thermal-mechanical-vibration collaborative model is unqualified, and the monitoring and control operations will no longer rely on the prediction results of the current thermal-mechanical-vibration collaborative model. At this time, the thermal-mechanical-vibration collaborative model can be automatically updated until the fusion qualification coefficient is not less than the preset fusion qualification coefficient threshold.

Images