Abrasive paper grinding site quality monitoring method based on edge AI

Abstract

The invention relates to the technical field of precise grinding intelligent monitoring, and discloses an abrasive paper grinding site quality monitoring method based on edge AI. The method comprises the steps that grinding force, acoustic emission and thermal imaging data are collected and fused, and quality feature vectors are extracted through time-frequency transformation; and continuously tracking the feature trajectory, and calculating the deviation degree and stability index of the feature trajectory. And when the indexes exceed the limit, grinding force non-stationarity, acoustic emission burst mode and thermal imaging gradient anomaly are analyzed in parallel, and a collaborative diagnosis anomaly mark is generated. And positioning an abnormal stage, backtracking a feature track, and constructing an abnormal event file containing an abnormal mode, feature evolution and process parameters. And based on the file, a process parameter compensation instruction is generated in real time, a monitoring threshold value is updated, then a compensation effect is evaluated, diagnostic logic is iteratively optimized, and closed-loop monitoring is formed. According to the method, the abnormality diagnosis accuracy and the process adaptive capability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for precision grinding, specifically to a method for on-site quality monitoring of sandpaper grinding based on edge AI. Background Technology

[0002] In sandpaper grinding, real-time monitoring of workpiece surface quality and sandpaper wear is crucial for ensuring process stability. Existing technologies typically rely on single sensor signals, such as monitoring only grinding force or acoustic emission signals, and setting fixed thresholds for out-of-limit alarms. Another common approach is to simultaneously acquire multiple signals, but then analyze these signals independently or perform simple weighted fusion, making judgments based on comprehensive indicators. However, these methods struggle to accurately distinguish between normal adjustments to process parameters, external interference, and genuine sandpaper passivation or workpiece surface defects when dealing with complex and variable on-site grinding conditions.

[0003] Existing technical solutions have shortcomings. Single-signal monitoring is susceptible to interference, has a high false alarm rate, and cannot fully reflect the complex evolution of grinding conditions. Simple parallel processing of multiple signals lacks in-depth analysis of the physical nature of different signals, and cannot accurately correlate the specific fault modes reflected in different signals when an anomaly occurs, resulting in ambiguous diagnostic conclusions and an inability to guide precise process adjustments. In addition, conventional systems typically only record instantaneous data at alarm points, lacking the recording and analysis of the complete state evolution process before and after the anomaly occurs. This makes each anomaly event an isolated piece of information, preventing the system from accumulating experience from historical anomalies. Consequently, the monitoring logic and thresholds become rigid and cannot adapt to process drift caused by tool wear, material batch changes, etc.

[0004] There is a need for a technology capable of deep collaborative analysis of multi-source heterogeneous sensor information to achieve highly reliable anomaly diagnosis. Simultaneously, there is an urgent need for monitoring systems to possess the ability to record, learn, and utilize the complete contextual information of each anomaly event, enabling adaptive optimization of diagnostic logic and monitoring parameters, and overcoming the limitations of traditional methods that rely on passive alarms and lack evolutionary mechanisms. Summary of the Invention

[0005] The purpose of this invention is to provide a sandpaper grinding on-site quality monitoring method based on edge AI to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for on-site quality monitoring of sandpaper grinding based on edge AI, the method comprising: The multi-dimensional sensor data generated by the sandpaper grinding equipment during operation is collected and fused. The multi-dimensional sensor data includes grinding force signal, acoustic emission signal and thermal imaging information. A joint time-frequency domain transformation is performed on the fused multi-dimensional sensor data to extract the initial quality feature vectors characterizing the sandpaper wear state and the workpiece surface texture. Based on the initial quality feature vector, the change trajectory of the quality feature is tracked during a continuous grinding time period, and the deviation and stability index of the feature trajectory are calculated. When the deviation and stability indices exceed the threshold for dynamic adjustment, the non-stationarity of the grinding force signal, the burst mode of the acoustic emission signal, and the gradient anomaly of the thermal imaging information are analyzed in parallel to generate collaborative diagnostic anomaly markers. Based on the collaborative diagnostic anomaly markers, the abnormal grinding stage is located, and the quality characteristic change trajectory corresponding to the abnormal grinding stage is traced back to construct a grinding quality anomaly event archive containing anomaly patterns, feature evolution, and process parameters. Based on the grinding quality anomaly event archive, compensation instructions for grinding process parameters are generated in real time, and monitoring thresholds in subsequent grinding processes are updated. After the compensation command is executed, a new round of multi-dimensional sensor data is collected to evaluate the compensation effect and iteratively optimize the feature extraction and diagnostic logic, forming a closed-loop process for online quality monitoring.

[0007] Preferably, the step of collecting and fusing multi-dimensional sensor data generated during the operation of the sandpaper grinding equipment includes: Force sensors, acoustic emission sensors, and infrared thermal imagers are deployed at preset positions on the sandpaper grinding equipment to synchronously collect the raw signal stream during the grinding process; The original signal stream is timestamped and the sampling rate is normalized to form a time-synchronized sensor data sequence. Spatial registration is performed on the time-synchronized sensor data sequence to map and associate the location of the acoustic emission signal source, the high-temperature area of ​​thermal imaging, and the physical location of the grinding contact area. The statistical and morphological features of the correlated data within the same time window are extracted, and these features are spliced ​​together in the order of the time window to form the multi-dimensional sensing data.

[0008] Preferably, the step of performing a joint time-frequency domain transformation on the fused multi-dimensional sensing data to extract the initial quality feature vector characterizing the sandpaper wear state and the workpiece surface texture includes: Empirical mode decomposition is performed on the grinding force signal in the multi-dimensional sensing data to obtain a series of intrinsic mode function components. The acoustic emission signal in the multi-dimensional sensing data is subjected to wavelet packet transform, decomposed into a set frequency band, and the energy proportion of each frequency band is calculated. The thermal imaging information in the multi-dimensional sensing data is partitioned, and the average temperature, temperature gradient, and temperature distribution entropy of each partition are calculated. The initial quality feature vector is constructed by combining and reducing the energy entropy of the intrinsic mode function components, the proportion of acoustic emission energy in each frequency band, and the thermal characteristic parameters of each partition.

[0009] Preferably, the step of tracking the change trajectory of quality features within a continuous grinding time period based on the initial quality feature vector, and calculating the deviation and stability index of the feature trajectory, includes: A sliding analysis time window is set, and within the analysis time window, the initial quality feature vectors are arranged in chronological order to form a feature trajectory sequence; Calculate the Euclidean distance between adjacent feature vectors in the feature trajectory sequence to form a distance variation curve; The distance change curve is piecewise linearly fitted, and the average slope of each line segment is calculated as the smoothness index of the trajectory. Using the mean of the feature vectors at the beginning of the analysis time window as a reference, the Mahalanobis distance between all subsequent feature vectors within the window and the reference is calculated, and the statistical maximum value of the Mahalanobis distance is defined as the feature trajectory deviation. The stability index is calculated by weighting the smoothness index and the deviation of the feature trajectory.

[0010] Preferably, the step of parallel analysis of the non-stationarity of the grinding force signal, the burst mode of the acoustic emission signal, and the gradient anomaly of the thermal imaging information when the deviation and stability indicators exceed the dynamically adjusted threshold includes: Based on historical grinding quality data, the warning thresholds and alarm thresholds of the deviation and stability indicators are dynamically updated; When the indicator exceeds the warning threshold, the parallel analysis thread is started: the first thread calculates the higher-order statistics of the grinding force signal for the corresponding time period and compares them with the baseline statistics under normal working conditions to quantify its non-stationarity index. The second thread detects abrupt changes in the signal envelope for acoustic emission signals within the corresponding time period, counts the frequency and energy of sudden events per unit time, and quantifies the intensity of sudden modes. The third thread calculates the consistency of the thermal gradient direction of the entire grinding area based on the thermal imaging information of the corresponding time period, and compares it with the standard texture direction to quantify the gradient anomaly coefficient.

[0011] Preferably, the step of generating collaborative diagnostic anomaly markers includes: Receive the nonstationarity index, burst mode intensity, and gradient anomaly coefficient; A three-dimensional decision space is established, with the three dimensions corresponding to the nonstationarity index, the intensity of the sudden mode, and the gradient anomaly coefficient, respectively. In the three-dimensional decision space, different abnormality types are divided into regions based on prior knowledge, including abrasive grain shedding regions, chip adhesion regions, and workpiece material inhomogeneity regions. The coordinate points formed by the currently calculated index, intensity, and coefficient are mapped to the three-dimensional decision space. Based on the anomaly type region to which they belong, corresponding collaborative diagnostic anomaly labels are generated. The collaborative diagnostic anomaly labels include anomaly type and confidence level.

[0012] Preferably, the step of locating the abnormal grinding stage, tracing back the quality characteristic change trajectory corresponding to the abnormal grinding stage, and constructing a grinding quality abnormality event file containing the abnormal mode, characteristic evolution, and process parameters includes: Based on the time point corresponding to the collaborative diagnostic anomaly marker, a time buffer is extended forward and backward, and the time buffer is defined as the abnormal grinding stage; Retrieve complete multi-dimensional sensor data, initial quality feature vector, and feature change trajectory during the abnormal grinding stage from the storage unit; The collaborative diagnostic anomaly markers, the retrieved anomaly stage data, and the process parameters recorded by the sandpaper grinding equipment during the anomaly grinding stage, including linear speed, feed rate, and grinding depth, are associated and encapsulated. Add timestamps, device identifiers, and sandpaper batch numbers to the associated encapsulated data packets to form a structured archive of grinding quality anomalies.

[0013] Preferably, the step of generating compensation instructions for grinding process parameters in real time based on the grinding quality anomaly event file and updating the monitoring thresholds in subsequent grinding processes includes: The grinding quality abnormality event file is analyzed, and the corresponding process parameter adjustment strategy is matched from the preset compensation strategy library according to the abnormality type recorded therein. The process parameter adjustment strategy includes suggestions on the adjustment amount for the linear speed, feed rate, or grinding depth of the grinding equipment. Based on the proposed adjustment amount, a specific, device-executable compensation instruction is generated; Simultaneously, based on the severity of the abnormality recorded in the grinding quality abnormality event file, the warning threshold and alarm threshold of the deviation and stability indicators are tightened proportionally to complete the threshold update.

[0014] Preferably, the step of continuing to collect a new round of multi-dimensional sensing data after the compensation instruction is executed, evaluating the compensation effect, and iteratively optimizing the feature extraction and diagnostic logic includes: Within a preset evaluation time after the compensation instruction is executed, a new round of multi-dimensional sensor data is collected. Based on the re-collected data, calculate the initial quality feature vector, feature trajectory deviation, and stability index for the new round; The new round of indicators is compared with the updated thresholds. If the indicators return to the normal range, the compensation is deemed effective. The successful compensation instruction is then associated with the corresponding abnormal event file and stored as a positive sample in the strategy library. If the indicators do not return to normal, the compensation is deemed invalid, a secondary compensation strategy is initiated, and the event is recorded to optimize the regional division or feature extraction parameters of the three-dimensional decision space, thereby achieving iterative optimization of the diagnostic logic.

[0015] Preferably, the closed-loop process for forming online quality monitoring is specifically embodied in the following ways: The online monitoring process, the initial feature construction of grinding quality, the dynamic quality evolution analysis, the collaborative diagnosis of multi-source anomalies, the construction of grinding quality anomaly event archives, the adaptive decision-making and evaluation iterative optimization steps are all run in a loop on the edge AI computing device. The output of the evaluation and optimization steps in each loop serves as the input for feature extraction, dynamic threshold adjustment, and diagnostic decisions in the next loop, enabling the monitoring system to have self-learning and self-adaptive capabilities. The closed-loop process continues to run until the sandpaper grinding task is completed.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By analyzing the non-stationarity of grinding force signals, the sudden modes of acoustic emission signals, and gradient anomalies in thermal imaging information in parallel when comprehensive indicators exceed limits, and then fusing these analyses to generate collaborative diagnostic anomaly markers, cross-modal anomaly correlation determination is achieved. This technical solution overcomes the shortcomings of single signals being susceptible to interference and the ambiguity of simple superposition of multiple signals. It can deeply analyze the specific performance of different physical signals during anomalies, logically correlate the slow degradation trend of force signals, the instantaneous impact events of acoustic emission signals, and the local distortion of thermal distribution, thereby improving the accuracy of identifying real grinding anomalies and reducing false alarms caused by fluctuations in single signals or normal process adjustments.

[0017] After locating anomalies, the system proactively traces the complete quality characteristic change trajectory corresponding to the anomaly stage and constructs a structured archive of grinding quality anomaly events, including anomaly patterns, feature evolution, and process parameters. This archive is then used to drive parameter compensation and system self-optimization. This approach transforms discrete alarm events into structured knowledge with context. The system not only performs immediate compensation but also records and archives the complete "feature-parameter-pattern" relationship for each anomaly event. This allows monitoring thresholds to be dynamically adjusted based on statistical analysis of historical anomaly data, and feature extraction algorithms to be specifically optimized for identified typical anomaly patterns. The entire process forms a closed loop from "perception-diagnosis-decision-learning," enabling the system to accumulate experience as production progresses, adaptively track the natural drift of process states, and continuously improve the sensitivity and accuracy of subsequent monitoring, achieving a fundamental shift from static monitoring to dynamic evolution. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the sandpaper grinding on-site quality monitoring method based on edge AI described in this invention. Figure 2 A flowchart for multi-dimensional sensor data acquisition and fusion; Figure 3 A flowchart for extracting the initial quality feature vector; Figure 4 A three-dimensional decision space distribution diagram of sandpaper grinding quality abnormalities; Figure 5 This is a diagram analyzing the effect of sandpaper grinding quality compensation. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides a method for on-site quality monitoring of sandpaper grinding based on edge AI. The method includes: collecting and fusing multi-dimensional sensor data generated by the sandpaper grinding equipment during operation, wherein the multi-dimensional sensor data includes grinding force signals, acoustic emission signals, and thermal imaging information; performing a time-frequency domain joint transformation on the fused multi-dimensional sensor data to extract initial quality feature vectors characterizing the sandpaper wear state and workpiece surface texture; tracking the change trajectory of quality features within a continuous grinding time period based on the initial quality feature vectors; calculating the deviation and stability index of the feature trajectory; and performing parallel analysis of the grinding process when the deviation and stability index exceed a dynamically adjusted threshold. The non-stationarity of the cutting force signal, the burst mode of the acoustic emission signal, and the gradient anomaly of the thermal imaging information are used to generate collaborative diagnostic anomaly markers. Based on these markers, abnormal grinding stages are located, and the trajectory of quality characteristic changes corresponding to these stages is traced back. A grinding quality anomaly event archive containing anomaly modes, feature evolution, and process parameters is constructed. Based on this archive, compensation instructions for grinding process parameters are generated in real time, and monitoring thresholds in subsequent grinding processes are updated. After the compensation instructions are executed, a new round of multi-dimensional sensor data is collected to evaluate the compensation effect and iteratively optimize feature extraction and diagnostic logic, forming a closed-loop process for online quality monitoring.

[0021] In one embodiment of the present invention, see [reference] Figure 2 Force sensors, acoustic emission sensors, and infrared thermal imagers are deployed at preset positions on the sandpaper grinding equipment to synchronously collect the raw signal stream during the grinding process. The raw signal stream is then timestamped and the sampling rate is normalized to form a time-synchronized sensor data sequence. Spatial registration is performed on the time-synchronized sensor data sequence to map and associate the physical locations of the acoustic emission signal source, the high-temperature area of ​​the thermal imaging, and the grinding contact area. The statistical and morphological features of the associated data within the same time window are extracted, and these features are spliced ​​together in the order of the time window to form multi-dimensional sensing data.

[0022] In practical implementation, the acquisition and fusion of multi-dimensional sensor data generated by the sandpaper grinding equipment during operation is executed with a deterministic process. Taking the surface grinding of an aluminum alloy workpiece as an example scenario, a piezoelectric force sensor is deployed on the grinding head fixture of the sandpaper grinding equipment to collect grinding force signals, an acoustic emission sensor with a resonant frequency of 1MHz is deployed on the side of the workpiece fixture to collect acoustic emission signals, and an infrared thermal imager is deployed diagonally above the grinding point to collect thermal imaging information. All three sensors begin acquiring data synchronously upon receiving a unified start command, forming a raw signal stream containing force, acoustic, and thermal information. These raw signal streams have different sampling frequencies and initial time bases: the force signal sampling frequency is 500Hz, the acoustic emission signal sampling frequency is 1MHz, and the infrared thermal imager frame rate is 30Hz. The synchronous acquisition results in three independent data sequences with microsecond-level differences in their time bases. In practice, the original signal stream is timestamped and the sampling rate is normalized. The force sensor, acoustic emission sensor and infrared thermal imager are all connected to the same high-precision clock source. Each data point is stamped with an absolute timestamp based on this clock source. The data sequences of all sensors are aligned on the time axis according to the absolute timestamp. After alignment, the high-frequency acoustic emission signal is downsampled and the low-frequency infrared thermal imaging sequence is interpolated to be unified to a preset common sampling frequency, forming a time-synchronized sensor data sequence.

[0023] It is understandable that although time-synchronized sensor data sequences are aligned in the time dimension, the physical events they reflect still need to be correlated in spatial location. Spatial registration is performed on the time-synchronized sensor data sequences to establish the mechanical coordinate system of the grinding equipment. The measurement center of the force sensor is defined as the origin. The acoustic emission sensor locates the acoustic emission events generated during the grinding process using the time difference of arrival method, and the calculated coordinates of the acoustic emission signal source are transformed into the mechanical coordinate system. The infrared thermal imager maps the pixel coordinates of each frame of its thermal imaging image to the three-dimensional mechanical coordinates of the workpiece surface using pre-calibrated extrinsic parameters, thereby mapping and associating the physical location of the high-temperature area in the thermal imaging with the grinding contact area. In some embodiments, the spatial registration operation enables the determination of the magnitude and direction of the grinding force, the precise location of the acoustic emission event, and the temperature field distribution on the surface of the grinding area within any identical time slice.

[0024] Optionally, after completing time synchronization and spatial registration, feature extraction is performed on the correlated data. Statistical and morphological features of the correlated data within the same time window are extracted. A fixed time window of 100 milliseconds is set. Within this time window, the mean, variance, and peak factor are extracted from the grinding force signal as statistical features; the ring count and root mean square value are extracted from the acoustic emission signal as statistical features; and the difference between the highest and average temperatures is extracted from the thermal imaging information as morphological features. These features extracted from the force, acoustic, and thermal data are concatenated in time window order. Assuming each time window extracts 3 features from the force signal, 2 features from the acoustic emission signal, and 1 feature from the thermal imaging information, each time window will output a feature vector containing 6 scalar values. Arranging these feature vectors in chronological order forms the multi-dimensional sensing data. In some embodiments, the multi-dimensional sensing data is a two-dimensional matrix, where the number of rows represents the number of time windows, and the number of columns represents the total number of features extracted and concatenated within each time window. Optionally, the order of feature concatenation is fixed, for example, force features first, acoustic features in the middle, and thermal features last, to ensure the consistency of the data structure. It can be understood that the multi-dimensional sensor data obtained through the above steps integrates time-aligned, spatially correlated, and physically meaningful force, acoustic, and thermal features, providing structured input for subsequent analysis. The formula for constructing the feature vector fusion can be expressed as:

[0025] in: This represents the fused feature vector. This represents the feature vector extracted from the grinding force signal. This represents the feature vector extracted from the acoustic emission signal. This represents a feature vector extracted from thermal imaging information. and It is a normalization coefficient used to balance the characteristics of different physical dimensions.

[0026] In one embodiment of the present invention, see [reference] Figure 3Empirical mode decomposition (EMD) is performed on the grinding force signal in the multidimensional sensing data to obtain a series of intrinsic mode function (IMF) components. Wavelet packet transform is performed on the acoustic emission signal in the multidimensional sensing data to decompose it into a set frequency band, and the energy proportion of each frequency band is calculated. The thermal imaging information in the multidimensional sensing data is partitioned, and the average temperature, temperature gradient, and temperature distribution entropy of each partition are calculated. The energy entropy of the IMF components, the acoustic emission energy proportion of each frequency band, and the thermal characteristic parameters of each partition are combined and dimension-reduced to construct an initial quality feature vector. A sliding analysis time window is set up. Initial quality feature vectors are arranged in chronological order within the analysis time window to form a feature trajectory sequence. The Euclidean distance between adjacent feature vectors in the feature trajectory sequence is calculated to form a distance change curve. Piecewise linear fitting is performed on the distance change curve, and the average slope of each line segment is calculated as the smoothness index of the trajectory. The mean of the feature vectors in the initial segment of the analysis time window is used as the reference benchmark. The Mahalanobis distance between all subsequent feature vectors in the window and the reference benchmark is calculated. The statistical maximum value of the Mahalanobis distance is defined as the feature trajectory deviation. The stability index is calculated by weighting the smoothness index and the feature trajectory deviation.

[0027] In practical implementation, a joint time-frequency domain transformation is performed on the fused multi-dimensional sensor data to extract the initial quality feature vector. Empirical mode decomposition (EMD) is performed on the grinding force signal from the multi-dimensional sensor data. A 200-millisecond segment of the grinding force signal is input into the EMD algorithm, which decomposes the signal into six intrinsic mode function components arranged from high frequency to low frequency and one residual component through an iterative sieving process. In practical implementation, wavelet packet transform is performed on the acoustic emission signal from the multi-dimensional sensor data. The db4 wavelet is selected as the basis function to perform three-level wavelet packet decomposition on the acoustic emission signal within the same time period, decomposing it into eight predetermined frequency bands. The energy of the reconstructed signal in each frequency band is calculated, and the energy percentage of each frequency band is obtained by dividing the energy value of each frequency band by the total energy value of the eight frequency bands.

[0028] In some embodiments, the thermal imaging information in the multi-dimensional sensing data is partitioned. For a frame of thermal imaging image from an infrared thermal imager, the image is divided into a 3x3 grid area centered on the grinding contact point. The average temperature of each partition is calculated by summing the temperature values ​​of all pixels within the partition and dividing by the total number of pixels. The temperature gradient of each partition is calculated by summing the absolute values ​​of the temperature differences between the partition center point and its four adjacent pixels (up, down, left, and right). The temperature distribution entropy of each partition is calculated based on the probability distribution of temperature values ​​within the partition. The energy entropy of the six intrinsic mode function components obtained from empirical mode decomposition, the proportion of acoustic emission energy in the eight frequency bands obtained from wavelet packet transform, and the average temperature, temperature gradient, and temperature distribution entropy of the nine partitions are combined to form an initial high-dimensional feature set containing six plus eight plus twenty-seven, totaling forty-one elements.

[0029] Optionally, principal component analysis (PCA) is performed on the initial high-dimensional feature set to reduce its dimensionality. The top ten principal components, whose cumulative contribution exceeds 95%, are retained and constitute the initial quality feature vectors. A sliding analysis time window is set, with a length that includes fifty consecutive initial quality feature vectors. These fifty initial quality feature vectors are arranged in chronological order within the analysis time window, forming a feature trajectory sequence. The Euclidean distance between adjacent feature vectors in the feature trajectory sequence is calculated. For the forty-ninth pair of adjacent vectors in the sequence, the square root of the sum of the squares of the differences between their corresponding ten principal component components is calculated to obtain the forty-ninth distance value. All forty-nine distance values ​​are arranged in chronological order to form a distance variation curve.

[0030] In some embodiments, a piecewise linear fit is performed on the distance variation curve. The distance variation curve is evenly divided into five segments, and a least-squares linear fit is performed on the distance values ​​within each segment to obtain the slopes of the five line segments. The arithmetic mean of these five slopes is calculated as the smoothness index of the trajectory. The mean of the feature vectors of the first segment of the analysis time window is used as the reference benchmark. The first segment is defined as the first five initial quality feature vectors. The mean vector of these five vectors is calculated as the reference benchmark. The Mahalanobis distance between the subsequent forty-five feature vectors within the window and the reference benchmark is calculated. The statistical maximum value of the Mahalanobis distance is defined as the feature trajectory deviation.

[0031] In one embodiment of the present invention, the warning threshold and alarm threshold of deviation and stability indicators are dynamically updated based on historical grinding quality data. When the indicators exceed the warning threshold, a parallel analysis thread is started. The first thread calculates the higher-order statistics of the grinding force signal for the corresponding time period and compares them with the baseline statistics under normal working conditions to quantify the non-stationarity index. The second thread detects the abrupt change points of the signal envelope for the acoustic emission signal for the corresponding time period, counts the frequency and energy of the sudden events per unit time, and quantifies the intensity of the sudden mode. The third thread calculates the consistency of the thermal gradient direction of the entire grinding area based on the thermal imaging information for the corresponding time period and compares it with the standard texture direction to quantify the gradient anomaly coefficient. The non-stationarity index, the intensity of the sudden mode, and the gradient anomaly coefficient are received to establish a three-dimensional decision space. The three dimensions correspond to the non-stationarity index, the intensity of the sudden mode, and the gradient anomaly coefficient, respectively. Different anomaly type regions are divided in the three-dimensional decision space according to prior knowledge, including abrasive grain shedding regions, chip adhesion regions, and workpiece material inhomogeneity regions. The coordinate points formed by the currently calculated index, intensity, and coefficient are mapped to the three-dimensional decision space. Corresponding collaborative diagnostic anomaly labels are generated according to the anomaly type region to which they belong. The collaborative diagnostic anomaly labels include anomaly type and confidence level.

[0032] In practice, the warning and alarm thresholds for deviation and stability indicators are dynamically updated based on historical grinding quality data. This historical data comprises deviation and stability index values ​​calculated from the past one hundred normal grinding processes. The warning threshold is set at 90% of the average of these historical values, and the alarm threshold at 80% of the average. After grinding each new workpiece, the deviation and stability indicators for that grinding process are included in the historical dataset, and the average is recalculated to dynamically update the thresholds for the next round of monitoring. When the deviation and stability indicators exceed the warning threshold, a parallel analysis thread is initiated. The first thread calculates the higher-order statistics of the grinding force signal for the corresponding time period. These higher-order statistics include skewness and kurtosis, and are compared with the baseline statistics of skewness and kurtosis of the grinding force signal under normal operating conditions. The baseline statistics are derived from historical normal data, quantifying the non-stationarity index. The formula for calculating the non-stationarity index is:

[0033] in: Represents a non-stationary index. This represents the skewness of the current grinding force signal. The reference value representing the deviation of the grinding force signal under normal working conditions. The kurtosis of the current grinding force signal. The reference value representing the kurtosis of the grinding force signal under normal operating conditions. and These are the preset weighting coefficients for skewness difference and kurtosis difference, respectively.

[0034] In some embodiments, the second thread detects abrupt changes in the acoustic emission signal envelope for a corresponding time period. It calculates the local variance of the acoustic emission signal envelope using a sliding time window. A sudden change is identified when the local variance exceeds twice the variance of the previous window. The number of sudden changes per unit time is counted as the burst event frequency. The arithmetic mean of the signal energy within the corresponding time window for all sudden changes is calculated as the average burst event energy. The burst mode intensity is quantified, and it is represented by the product of the burst event frequency and the average burst event energy. The third thread calculates the consistency of the thermal gradient direction across the entire grinding area based on the thermal imaging information for the corresponding time period. In the thermal imaging image, the Sobel operator is used to calculate the temperature gradient direction and amplitude of each pixel. The gradient directions of all pixels with amplitudes greater than a threshold are counted, and the average direction of these direction vectors is calculated and compared with the standard texture direction of the workpiece surface. The standard texture direction is set by the process. The gradient anomaly coefficient is quantified, defined as the absolute value of the cosine of the angle between the average direction and the standard texture direction.

[0035] Optionally, after receiving the nonstationarity index, burst mode intensity, and gradient anomaly coefficient, a three-dimensional decision space is established. The three dimensions correspond to the nonstationarity index, burst mode intensity, and gradient anomaly coefficient, respectively. Different anomaly type regions are defined in the three-dimensional decision space based on prior knowledge derived from the analysis of historical anomaly events. For example, abrasive particle detachment regions are defined as regions with high nonstationarity index, medium burst mode intensity, and low gradient anomaly coefficient; chip adhesion regions are defined as regions with medium nonstationarity index, high burst mode intensity, and high gradient anomaly coefficient; and workpiece material inhomogeneity regions are defined as regions with low nonstationarity index, low burst mode intensity, and medium gradient anomaly coefficient. This can be understood as mapping the coordinate points formed by the currently calculated index, intensity, and coefficient to the three-dimensional decision space, where the coordinate points are represented as... ,in Represents a non-stationary index. Represents the intensity of the sudden mode. Represents the gradient anomaly coefficient, based on the coordinate point In the three-dimensional decision space, a corresponding collaborative diagnostic anomaly label is generated for the anomaly type region. The collaborative diagnostic anomaly label includes the anomaly type and confidence level. The confidence level is quantified by the reciprocal of the Euclidean distance from the coordinate point to the geometric center of the anomaly type region.

[0036] See Figure 4This is a 3D decision space distribution map of sandpaper grinding quality anomalies, showcasing the clustering features of different anomaly types across three dimensions: "non-stationarity index," "burst mode intensity," and "gradient anomaly coefficient." This map is a core diagnostic tool for edge AI grinding quality monitoring. Newly acquired signal features can be mapped to this 3D space for rapid matching of their respective anomaly types; the clear clustering boundaries of different anomaly types improve the accuracy and confidence of anomaly diagnosis; further optimization of clustering regions through accumulating more samples enhances the system's self-learning capabilities.

[0037] In one embodiment of the present invention, a time buffer is extended forward and backward from the time point corresponding to the collaborative diagnostic anomaly marker, defining the time buffer as an abnormal grinding stage. Complete multi-dimensional sensor data, initial quality feature vectors, and feature change trajectories within the abnormal grinding stage are retrieved from the storage unit. The collaborative diagnostic anomaly marker, the retrieved abnormal stage data, and the process parameters recorded by the sandpaper grinding equipment during the abnormal grinding stage, including linear speed, feed rate, and grinding depth, are associated and encapsulated. A timestamp, equipment identifier, and sandpaper batch number are added to the associated encapsulated data packet to form a structured grinding quality anomaly event file. The grinding quality anomaly event file is parsed, and a corresponding process parameter adjustment strategy is matched from a preset compensation strategy library based on the anomaly type recorded therein. The process parameter adjustment strategy includes adjustment suggestions for the linear speed, feed rate, or grinding depth of the grinding equipment. Specific equipment-executable compensation instructions are generated based on the adjustment suggestions. Simultaneously, the warning thresholds and alarm thresholds of deviation and stability indicators are tightened proportionally according to the anomaly severity recorded in the grinding quality anomaly event file, completing the threshold update.

[0038] In practical implementation, a time buffer is extended forward and backward based on the time point corresponding to the collaborative diagnostic anomaly marker. The time point of the collaborative diagnostic anomaly marker is t1, the time extension is Δt1 forward, and the time extension is Δt2 backward, defining the extended time interval [t1-Δt1, t1+Δt2] as the abnormal grinding stage. Complete multi-dimensional sensor data, initial quality feature vectors, and feature change trajectories within the abnormal grinding stage are retrieved from the storage unit. The storage unit stores the original data and processing results indexed by time during the grinding process. All relevant data within the time interval [t1-Δt1, t1+Δt2] are retrieved and extracted using timestamps. The collaborative diagnostic anomaly marker, the retrieved abnormal stage data, and the process parameters recorded by the sandpaper grinding equipment during the abnormal grinding stage are associated and encapsulated. The process parameters include linear speed, feed rate, and grinding depth, which are obtained from the CNC system log of the grinding equipment, aligned by time.

[0039] In some embodiments, a timestamp, device identifier, and sandpaper batch number are added to the associated encapsulated data packets. The timestamp records the start and end times of the abnormal grinding stage, the device identifier is used to distinguish different grinding equipment, and the sandpaper batch number identifies the production batch to which the currently used sandpaper belongs, forming a structured grinding quality abnormality event file. The grinding quality abnormality event file is parsed, and the corresponding process parameter adjustment strategy is matched from a preset compensation strategy library based on the abnormality type recorded in the grinding quality abnormality event file. The compensation strategy library is a preset mapping table, see Table 1.

[0040] Table 1: Mapping Table of Anomaly Types and Process Parameter Adjustment Strategies

[0041] The process parameter adjustment strategy includes suggested adjustments to the grinding equipment's linear speed, feed rate, or grinding depth. Based on these suggestions, specific, executable compensation instructions are generated. These instructions are issued in G-code or specific protocol commands recognizable by the CNC system, and contain the adjusted target parameter values. Simultaneously, the warning and alarm thresholds for deviation and stability indicators are tightened proportionally based on the severity of anomalies recorded in the grinding quality anomaly event file. The anomaly severity is quantified by the confidence level in the collaborative diagnostic anomaly markers, and the threshold update formula is:

[0042] in: This represents the updated threshold. This represents the original threshold before the update. Represents the confidence level in collaborative diagnostic anomaly markers. It is a proportional coefficient used to control the degree of tightening and to update the threshold.

[0043] It is understandable that the grinding quality anomaly event archive, as a structured data package, not only records the raw data and diagnostic results of the anomaly events, but also serves as the direct basis for triggering process parameter adjustments and monitoring parameter adaptation. Optionally, the adjustment suggestions in the compensation strategy library are generated based on the process knowledge base and historical successful compensation cases, and are continuously optimized during system operation. In some embodiments, the threshold update operation takes effect immediately, and subsequent analysis windows will immediately adopt the updated threshold. Perform anomaly detection.

[0044] In one embodiment of the present invention, a new round of multi-dimensional sensor data is re-acquired within a preset evaluation time after the compensation command is executed. Based on the re-acquired data, a new round of initial quality feature vector, feature trajectory deviation, and stability index are calculated. The new round of indexes are compared with the updated thresholds. If the indexes return to the normal range, the compensation is deemed effective, and the successful compensation command is associated with the corresponding abnormal event file as a positive sample and stored in the strategy library. If the indexes do not return to normal, the compensation is deemed ineffective, a secondary compensation strategy is initiated, and the event is recorded to optimize the region division or feature extraction parameters of the three-dimensional decision space, thereby achieving iterative optimization of the diagnostic logic. The online monitoring process, initial feature construction of grinding quality, dynamic quality evolution analysis, multi-source anomaly collaborative diagnosis, construction of grinding quality abnormal event files, adaptive decision-making, and evaluation iterative optimization steps are cyclically run on the edge AI computing device. The output of each cycle's evaluation iterative optimization step serves as the input for feature extraction, dynamic threshold adjustment, and diagnostic decision-making in the next cycle, enabling the monitoring system to have self-learning and adaptive capabilities. The closed-loop process continues to run until the sandpaper grinding task is completed.

[0045] In practice, a new round of multi-dimensional sensor data is re-acquired within a preset evaluation time after the compensation command is executed. The preset evaluation time is set to the grinding process lasting thirty seconds after the compensation command is issued. During these thirty seconds, the force sensor, acoustic emission sensor, and infrared thermal imager re-synchronize, align timestamps, normalize sampling rates, spatially register, and extract features to generate a new round of multi-dimensional sensor data in a thirty-second time window sequence. Based on the re-acquired multi-dimensional sensor data, the initial quality feature vector, feature trajectory deviation, and stability index of the new round are calculated. Empirical mode decomposition, wavelet packet transform, thermal imaging partitioning, and feature dimensionality reduction are performed on the new sensor data to obtain the new initial quality feature vector. The feature change trajectory is tracked within the sliding analysis time window to calculate the new feature trajectory deviation and the new stability index.

[0046] In some embodiments, the new round of feature trajectory deviation and stability indicators are compared with the updated warning threshold and alarm threshold. The updated threshold is derived from the tightened values ​​of the anomaly severity. If both the new round of feature trajectory deviation and stability indicators are lower than the updated warning threshold, the compensation is deemed effective. The compensation instruction generated and successfully executed is then associated with the grinding quality anomaly event file that triggered this compensation and stored as a new positive sample entry in the compensation strategy library. If either the new round of feature trajectory deviation or stability indicator is still higher than the updated warning threshold, the compensation is deemed ineffective, and a secondary compensation strategy is initiated. The secondary compensation strategy is a backup strategy that increases the adjustment range or changes the adjustment parameters based on the initial compensation strategy. The failed compensation event is recorded to optimize the region division or feature extraction parameters of the three-dimensional decision space.

[0047] The online monitoring process, initial feature construction for grinding quality, dynamic quality evolution analysis, multi-source anomaly collaborative diagnosis, construction of grinding quality anomaly event archives, adaptive decision-making, and evaluation iterative optimization steps are all run cyclically on an edge AI computing device, which is an embedded computing unit deployed on-site at the grinding equipment. The output of each iterative optimization step serves as the input for feature extraction, dynamic threshold adjustment, and diagnostic decisions in the next cycle. For example, the optimized 3D decision space region division is used for the next anomaly type discrimination, and successfully stored compensation cases are used to enrich the basis for generating subsequent compensation instructions, enabling the monitoring system to have self-learning and adaptive capabilities. The closed-loop process continues until the sandpaper grinding task ends. The end-of-grind instruction is issued by the equipment's CNC system. Upon receiving this instruction, the edge AI computing device terminates the data acquisition and analysis cycle and saves a complete monitoring log.

[0048] See Figure 5 This is a graph analyzing the effect of sandpaper grinding quality compensation, which primarily shows the changes in "feature trajectory deviation" under different compensation types. The compensation measures are significantly effective in reducing feature trajectory deviation, and the higher the compensation level, the greater the reduction. This indicates that a multi-level compensation strategy can effectively repair grinding quality anomalies. Based on this result, the compensation process can be optimized: when initial compensation is ineffective, quickly upgrade to secondary / tertiary compensation to improve the efficiency and thoroughness of anomaly repair.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for on-site quality monitoring of sandpaper grinding based on edge AI, characterized in that, The method includes: The multi-dimensional sensor data generated by the sandpaper grinding equipment during operation is collected and fused. The multi-dimensional sensor data includes grinding force signal, acoustic emission signal and thermal imaging information. A joint time-frequency domain transformation is performed on the fused multi-dimensional sensor data to extract the initial quality feature vectors characterizing the sandpaper wear state and the workpiece surface texture. Based on the initial quality feature vector, the change trajectory of the quality feature is tracked during a continuous grinding time period, and the deviation and stability index of the feature trajectory are calculated. When the deviation and stability indices exceed the threshold for dynamic adjustment, the non-stationarity of the grinding force signal, the burst mode of the acoustic emission signal, and the gradient anomaly of the thermal imaging information are analyzed in parallel to generate collaborative diagnostic anomaly markers. Based on the collaborative diagnostic anomaly markers, the abnormal grinding stage is located, and the quality characteristic change trajectory corresponding to the abnormal grinding stage is traced back to construct a grinding quality anomaly event archive containing anomaly patterns, feature evolution, and process parameters. Based on the grinding quality anomaly event archive, compensation instructions for grinding process parameters are generated in real time, and monitoring thresholds in subsequent grinding processes are updated. After the compensation command is executed, a new round of multi-dimensional sensor data is collected to evaluate the compensation effect and iteratively optimize the feature extraction and diagnostic logic, forming a closed-loop process for online quality monitoring.

2. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 1, characterized in that, The steps for collecting and fusing multi-dimensional sensor data generated during the operation of the sandpaper grinding equipment include: Force sensors, acoustic emission sensors, and infrared thermal imagers are deployed at preset positions on the sandpaper grinding equipment to synchronously collect the raw signal stream during the grinding process; The original signal stream is timestamped and the sampling rate is normalized to form a time-synchronized sensor data sequence. Spatial registration is performed on the time-synchronized sensor data sequence to map and associate the location of the acoustic emission signal source, the high-temperature area of ​​thermal imaging, and the physical location of the grinding contact area. The statistical and morphological features of the correlated data within the same time window are extracted, and these features are spliced ​​together in the order of the time window to form the multi-dimensional sensing data.

3. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 2, characterized in that, The step of performing a joint time-frequency domain transformation on the fused multi-dimensional sensing data to extract the initial quality feature vectors characterizing the sandpaper wear state and workpiece surface texture includes: Empirical mode decomposition is performed on the grinding force signal in the multi-dimensional sensing data to obtain a series of intrinsic mode function components. The acoustic emission signal in the multi-dimensional sensing data is subjected to wavelet packet transform, decomposed into a set frequency band, and the energy proportion of each frequency band is calculated. The thermal imaging information in the multi-dimensional sensing data is partitioned, and the average temperature, temperature gradient, and temperature distribution entropy of each partition are calculated. The initial quality feature vector is constructed by combining and reducing the energy entropy of the intrinsic mode function components, the proportion of acoustic emission energy in each frequency band, and the thermal characteristic parameters of each partition.

4. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 3, characterized in that, The step of tracking the change trajectory of quality features within a continuous grinding time period based on the initial quality feature vector, and calculating the deviation and stability index of the feature trajectory, includes: A sliding analysis time window is set, and within the analysis time window, the initial quality feature vectors are arranged in chronological order to form a feature trajectory sequence; Calculate the Euclidean distance between adjacent feature vectors in the feature trajectory sequence to form a distance variation curve; The distance change curve is piecewise linearly fitted, and the average slope of each line segment is calculated as the smoothness index of the trajectory. Using the mean of the feature vectors at the beginning of the analysis time window as a reference, the Mahalanobis distance between all subsequent feature vectors within the window and the reference is calculated, and the statistical maximum value of the Mahalanobis distance is defined as the feature trajectory deviation. The stability index is calculated by weighting the smoothness index and the deviation of the feature trajectory.

5. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 4, characterized in that, The step of parallel analysis of the non-stationarity of the grinding force signal, the burst mode of the acoustic emission signal, and the gradient anomaly of the thermal imaging information when the deviation and stability indices exceed the dynamically adjusted threshold includes: Based on historical grinding quality data, the warning thresholds and alarm thresholds of the deviation and stability indicators are dynamically updated; When the indicator exceeds the warning threshold, the parallel analysis thread is started: the first thread calculates the higher-order statistics of the grinding force signal for the corresponding time period and compares them with the baseline statistics under normal working conditions to quantify its non-stationarity index. The second thread detects abrupt changes in the signal envelope for acoustic emission signals within the corresponding time period, counts the frequency and energy of sudden events per unit time, and quantifies the intensity of sudden modes. The third thread calculates the consistency of the thermal gradient direction of the entire grinding area based on the thermal imaging information of the corresponding time period, and compares it with the standard texture direction to quantify the gradient anomaly coefficient.

6. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 5, characterized in that, The step of generating collaborative diagnostic anomaly markers includes: Receive the nonstationarity index, burst mode intensity, and gradient anomaly coefficient; A three-dimensional decision space is established, with the three dimensions corresponding to the nonstationarity index, the intensity of the sudden mode, and the gradient anomaly coefficient, respectively. In the three-dimensional decision space, different abnormality types are divided into regions based on prior knowledge, including abrasive grain shedding regions, chip adhesion regions, and workpiece material inhomogeneity regions. The coordinate points formed by the currently calculated index, intensity, and coefficient are mapped to the three-dimensional decision space. Based on the anomaly type region to which they belong, corresponding collaborative diagnostic anomaly labels are generated. The collaborative diagnostic anomaly labels include anomaly type and confidence level.

7. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 6, characterized in that, The steps of locating abnormal grinding stages, tracing back the quality characteristic change trajectory corresponding to the abnormal grinding stages, and constructing a grinding quality abnormality event archive containing abnormal patterns, characteristic evolution, and process parameters include: Based on the time point corresponding to the collaborative diagnostic anomaly marker, a time buffer is extended forward and backward, and the time buffer is defined as the abnormal grinding stage; Retrieve complete multi-dimensional sensor data, initial quality feature vector, and feature change trajectory during the abnormal grinding stage from the storage unit; The collaborative diagnostic anomaly markers, the retrieved anomaly stage data, and the process parameters recorded by the sandpaper grinding equipment during the anomaly grinding stage, including linear speed, feed rate, and grinding depth, are associated and encapsulated. Add timestamps, device identifiers, and sandpaper batch numbers to the associated encapsulated data packets to form a structured archive of grinding quality anomalies.

8. The method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 7, characterized in that, The step of generating compensation instructions for grinding process parameters in real time based on the grinding quality anomaly event file, and updating the monitoring thresholds in subsequent grinding processes includes: The grinding quality abnormality event file is analyzed, and the corresponding process parameter adjustment strategy is matched from the preset compensation strategy library according to the abnormality type recorded therein. The process parameter adjustment strategy includes suggestions on the adjustment amount for the linear speed, feed rate, or grinding depth of the grinding equipment. Based on the proposed adjustment amount, a specific, device-executable compensation instruction is generated; Simultaneously, based on the severity of the abnormality recorded in the grinding quality abnormality event file, the warning threshold and alarm threshold of the deviation and stability indicators are tightened proportionally to complete the threshold update.

9. A method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 8, characterized in that, The steps of continuing to collect a new round of multi-dimensional sensor data after the compensation instruction is executed, evaluating the compensation effect, and iteratively optimizing the feature extraction and diagnostic logic include: Within a preset evaluation time after the compensation instruction is executed, a new round of multi-dimensional sensor data is collected. Based on the re-collected data, calculate the initial quality feature vector, feature trajectory deviation, and stability index for the new round; The new round of indicators is compared with the updated thresholds. If the indicators return to the normal range, the compensation is deemed effective. The successful compensation instruction is then associated with the corresponding abnormal event file and stored as a positive sample in the strategy library. If the indicators do not return to normal, the compensation is deemed invalid, a secondary compensation strategy is initiated, and the event is recorded to optimize the regional division or feature extraction parameters of the three-dimensional decision space, thereby achieving iterative optimization of the diagnostic logic.

10. A method for on-site quality monitoring of sandpaper grinding based on edge AI according to claim 9, characterized in that, The closed-loop process for forming online quality monitoring is specifically manifested as follows: The online monitoring process, the initial feature construction of grinding quality, the dynamic quality evolution analysis, the collaborative diagnosis of multi-source anomalies, the construction of grinding quality anomaly event archives, the adaptive decision-making and evaluation iterative optimization steps are all run in a loop on the edge AI computing device. The output of the evaluation and optimization steps in each loop serves as the input for feature extraction, dynamic threshold adjustment, and diagnostic decisions in the next loop, enabling the monitoring system to have self-learning and self-adaptive capabilities. The closed-loop process continues to run until the sandpaper grinding task is completed.

Images