Machine learning based diamond milling tool in-machine wear state prediction method and system

Abstract

This invention discloses a machine learning-based method and system for predicting the in-machine wear state of diamond end mills, belonging to the field of wear prediction technology. It includes: a baseline signal template acquisition module, a characterization data acquisition module, a defect feature extraction module, an image feature extraction module, and a wear state output module. This invention generates visual quality confidence scores through image quality assessment, automatically reducing the influence of low-quality images and making the model more reliant on residual signals. It generates synthetic residual signals by superimposing simulated defect signals on the baseline signal template, increasing training samples. A lightweight gradient boosting decision tree model is used to reduce computational requirements, enabling real-time monitoring on low-computing-power devices. This solves the problem in existing technologies where the need for large amounts of training data, high computational power, and clean imaging environments directly conflicts with the small-batch, low-computing-power, and high-oil-mist conditions of woodworking tool manufacturing workshops, leading to difficulties in implementation and high costs.

Description

Technical Field

[0001] This invention relates to the field of wear prediction technology, and in particular to a method and system for predicting the on-machine wear state of diamond end mills based on machine learning. Background Technology

[0002] With the widespread adoption of five-axis grinding machines in woodworking tool manufacturing workshops for on-machine cutting edge shaping and finishing of diamond end mills, the need for real-time on-site identification of grinding defects is becoming increasingly urgent.

[0003] Currently, the mainstream solution involves fixing an industrial camera on the top of the machine tool cabin and simultaneously acquiring the spindle's acoustic emission and current signals. The wear level is then output using a CNN-BiLSTM network, and this solution has been demonstrated and is currently in operation at large tool manufacturing companies.

[0004] In reality, the compact workshop cabin caused coolant atomization that obscured the camera, and small-batch model changes could not accumulate enough training samples. The edge computing power was mostly Jetson Nano level and there was no one to maintain it during the night shift, which led to frequent false alarms in deep learning models and a longer deployment cycle. Ultimately, the site was forced to revert to a conservative strategy of scrapping a fixed number of parts and conducting low-frequency manual spot checks.

[0005] In summary, existing technologies rely on massive training samples, high-performance servers, and clear imaging environments free of oil mist. However, in the reality of woodworking tool manufacturing workshops, tool types are changed multiple times a day, computing power is limited, and coolant mist is prevalent, making implementation difficult and costly. To address these issues, this invention proposes a solution. Summary of the Invention

[0006] This application provides a machine learning-based method and system for predicting the on-machine wear state of diamond end mills, which at least solves the problem that the need for a large amount of training data, high computing power and clean imaging environment in related technologies directly conflicts with the small-batch, low computing power and high oil mist conditions of woodworking tool manufacturing workshops, resulting in difficulties in implementation and high costs.

[0007] This application provides a machine learning-based method for predicting the on-machine wear state of diamond end mills, including:

[0008] Obtain the baseline signal template corresponding to the diamond end mill and a specific grinding wheel specification under normal, defect-free grinding conditions;

[0009] During the grinding of diamond end mills, sensor data is collected in real time, and image data of the diamond end mill cutting edge is also collected.

[0010] The real-time acquired sensor data is differentially processed with the baseline signal template to generate a residual signal that can characterize grinding anomalies, and the defect features of the residual signal are extracted.

[0011] The image data is subjected to quality assessment, visual quality confidence scores are generated, and image features are extracted.

[0012] The defect features and the image features are input into a preset machine learning-based fusion prediction model, and the input image features are weighted using visual quality confidence. The fusion prediction model outputs the wear state that characterizes the grinding defects.

[0013] Furthermore, the baseline signal template is obtained as follows:

[0014] For each type of grinding wheel, select multiple new diamond end mills that have been confirmed to have acceptable grinding quality for trial grinding;

[0015] During the trial grinding process, sensor data, including acoustic emission signals and spindle current signals, were collected. The signals from multiple qualified samples were aligned and averaged to construct a baseline signal template S representing the standard diamond end mill grinding process under this wheel specification. base,k (t), where k is the grinding wheel specification and t is time.

[0016] Furthermore, the residual signal generation step includes:

[0017] The real-time collected sensor data S real (t) and the baseline signal template S invoked according to the current grinding wheel specification identifier k. base,k Subtracting (t) yields the residual signal S. res (t):

[0018] S res (t)=S real (t)-S base,k (t).

[0019] Furthermore, the extraction of defect features from the residual signal includes the following steps:

[0020] For the residual signal S res (t) Perform time-frequency analysis and statistically analyze its energy integral, kurtosis and crest factor within the preset defect frequency band to form a defect feature vector.

[0021] Furthermore, the visual quality confidence level includes the following steps:

[0022] For each frame of the image, the Laplacian variance is first calculated on the black area to obtain the fogging blur index I of the blade edge image. blur ;

[0023] The entire cutting edge region is segmented using the Otsu's method, and the proportion of pixels judged as white fog in the cutting edge region is counted to obtain the occlusion rate R of the effective cutting edge region. occ ;

[0024] The blurring index I of the cutting edge image is determined by the following functional relationship. blur The obstruction rate R of the effective area of ​​the cutting edge occ Mapped to visual quality confidence level W vision :

[0025]

[0026] Where c1 and c2 are weighting coefficients for adjusting the effects of blur and occlusion, c0 is the decision bias, both of which are constants, and e is a natural constant.

[0027] Furthermore, the fusion prediction model is specifically a lightweight gradient boosting decision tree model;

[0028] Before inputting the defect features and image features into the fusion prediction model, the image features are first multiplied by the visual quality confidence level W. vision It is used for dynamic adjustment of image feature weights.

[0029] Furthermore, it also includes training samples for enhancing the fusion prediction model:

[0030] Baseline signal templates S of different specifications base,k On (t), a series of preset characteristic signal waveforms simulating defects are superimposed to artificially synthesize residual signals with defect labels;

[0031] The synthesized residual signal and its defect features, combined with real-collected defect samples, are used to form a training set for training the fusion prediction model.

[0032] This application provides a machine learning-based system for predicting the on-machine wear condition of diamond end mills, including:

[0033] Baseline signal template acquisition module: Acquires the baseline signal template corresponding to the diamond end mill and a specific grinding wheel specification under normal, defect-free grinding conditions;

[0034] Characterization data acquisition module: During the grinding of diamond end mills, sensor data containing at least acoustic emission signals or spindle current signals are acquired in real time, and image data of the diamond end mill cutting edge is also acquired.

[0035] Defect feature extraction module: Performs differential operation on the real-time acquired sensor data and the baseline signal template to generate a residual signal that can characterize grinding anomalies, and extracts the defect features of the residual signal.

[0036] Image feature extraction module: performs quality assessment on the image data, generates visual quality confidence scores, and extracts image features;

[0037] Wear state output module: Input the defect features and the image features into a preset machine learning-based fusion prediction model, and use visual quality confidence to weight the input image features. The fusion prediction model outputs the wear state that represents the grinding defect.

[0038] This application assesses the quality of each image frame, statistically analyzes the fogging blur index and occlusion rate of the cutting edge image, and generates a visual quality confidence score. When image quality decreases due to coolant fogging, the visual quality confidence score decreases accordingly. Furthermore, before inputting image features into the fusion prediction model, the visual quality confidence score is multiplied. This automatically weakens the influence of unreliable image quality on the fusion prediction model, making it more reliant on reliable residual signals for judgment, thus solving the problem of coolant fogging occluding the camera. Next, a series of preset feature signal waveforms simulating defects such as micro-blowouts or cutting edge blunting are superimposed on baseline signal templates of different specifications, artificially synthesizing residual signals with clear defect labels. The generated samples are closer to physical reality and are specifically designed to train the "residual signal" recognition model, solving the problem of insufficient training samples caused by small-batch model changes. Finally, a lightweight gradient boosting decision tree model is used as the fusion prediction model. Compared to complex deep learning models, fusion prediction models have lower computational requirements, making them suitable for running on Jetson Nano-level devices with limited computing power. They also have fast inference speeds, meeting the needs of real-time monitoring. The acquisition and updating of baseline signal templates, sample enhancement, model training, and inference can all be completed automatically during device operation, reducing reliance on manual maintenance. This solves the problem of existing technologies conflicting with the small-batch, low-computing-power, and high-oil-mist conditions of woodworking tool manufacturing workshops, making implementation difficult and costly. Attached Figure Description

[0039] Figure 1 Flowchart of a machine learning-based method for predicting the on-machine wear state of diamond end mills provided in this application embodiment;

[0040] Figure 2 This is a diagram illustrating the architecture of a machine learning-based diamond end mill in-machine wear state prediction system provided in an embodiment of this application. Detailed Implementation

[0041] This application provides a machine learning-based method and system for predicting the in-machine wear state of diamond end mills. This solves the problem that the existing technology requires a large amount of training data, high computing power, and a clean imaging environment, which directly conflicts with the small-batch, low-computing-power, and high-oil-mist conditions of woodworking tool manufacturing workshops, resulting in difficulties in implementation and high costs. It achieves high-precision, high-robustness, and low-cost prediction of the in-machine wear state of diamond end mills.

[0042] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0043] like Figure 1 As shown in the embodiments of this application, a machine learning-based method for predicting the on-machine wear state of diamond end mills is provided, including:

[0044] Step 1: Obtain the baseline signal template of the diamond end mill and the specific grinding wheel specification under normal, defect-free grinding conditions.

[0045] The baseline signal template is obtained as follows:

[0046] For each type of grinding wheel, select multiple new diamond end mills that have been confirmed to have acceptable grinding quality for trial grinding.

[0047] During the trial grinding process, sensor data, including acoustic emission signals and spindle current signals, were collected. The signals from multiple qualified samples were aligned and averaged to construct a baseline signal template S representing the standard diamond end mill grinding process under this wheel specification. base,k (t), where k is the grinding wheel specification and t is time.

[0048] In this embodiment, the baseline signal template is used to capture the "normal grinding state" signal. For each wheel size, 3-5 brand-new, unused diamond end mills are selected from the current batch as samples. These samples must be pre-ground to ensure that the cutting edge is free of burrs or chipping. The selected diamond end mills are mounted on the machine tool, and trial grinding is performed using the current wheel size according to the corresponding standard grinding parameters, including rotational speed, feed rate, and depth of cut. For each wheel size, continuous grinding is performed using the new diamond end mill under standard feed, rotational speed, and cooling conditions, and acoustic emission / current signals are collected for 30–60 seconds. The time-domain average is then stored as the template. Data collection is repeated when the grinding wheel is changed.

[0049] Step 2: During the grinding of the diamond end mill, real-time acquisition of sensor data containing at least acoustic emission signals or spindle current signals, and acquisition of image data of the diamond end mill cutting edge.

[0050] Step 3: Perform differential operation between the real-time acquired sensor data and the baseline signal template to generate a residual signal that can characterize grinding anomalies, and extract the defect features of the residual signal.

[0051] The steps for generating the residual signal include:

[0052] The real-time collected sensor data S real (t) and the baseline signal template S invoked according to the current grinding wheel specification identifier k. base,kSubtracting (t) yields the residual signal S. res (t):

[0053] S res (t)=S real (t)-S base,k (t).

[0054] Among them, S res (t) The amplitude is close to zero under ideal grinding conditions, but significant pulses or fluctuations will occur when grinding defects such as micro-chipping or overheating occur.

[0055] The extraction of defect features from the residual signal includes the following steps:

[0056] For the residual signal S res (t) Perform time-frequency analysis and statistically analyze its energy integral (vis-viva formula), kurtosis, and crest factor within the preset defect frequency band to form a defect feature vector.

[0057] Step 4: Perform quality assessment on the image data, generate visual quality confidence scores, and extract image features.

[0058] The visual quality confidence level includes the following steps:

[0059] For each frame of the image, the Laplacian variance is first calculated on the black area to obtain the blurring index I of the blade edge image. blur .

[0060] The entire cutting edge region is segmented using the Otsu algorithm (maximum inter-class variance method). The proportion of pixels identified as white fog in the cutting edge region is then counted to obtain the occlusion rate R of the effective cutting edge region. occ .

[0061] The blurring index I of the cutting edge image is determined by the following functional relationship. blur The obstruction rate R of the effective area of ​​the cutting edge occ Mapped to visual quality confidence level W vision :

[0062]

[0063] Among them, W vision The value of is between 0 and 1. The higher the value, the more reliable the image quality. c1 and c2 are the weight coefficients for adjusting the effects of blur and occlusion. c0 is the decision bias, which are constants obtained through calibration experiments on sample images. e is a natural constant.

[0064] In this embodiment, based on the existing fixed light source and industrial camera, a matte black calibration plate, 5mm × 5mm in size, is attached above the cutting edge within the camera's field of view, flush with the cutting edge. This black plate remains in the field of view at any grinding stage, serving as a "built-in grayscale standard." Three levels—"no fog," "medium fog," and "heavy fog"—are generated sequentially, with at least 30 frames captured for each level. The three levels are further divided according to the concentration of atomized particulate matter from the workshop coolant. The no-fog level is defined as [0.05, 0.1) mg / m³. 3 Moderate fog level: [0.1~0.3) mg / m³ 3 Severe fog level: [0.3~0.5) mg / m³ 3 Based on each frame of the image, the Laplacian variance is first calculated on the black areas to obtain the fogging blur index I of the blade edge image. blur Next, Otsu segmentation is performed on the entire cutting edge area, and the proportion of pixels judged as "white fog" in the cutting edge area is counted to obtain the occlusion rate R of the effective cutting edge area. occ . Will I blur With R occ Inputting the data into the corresponding functional expression yields the visual quality confidence score W. vision Among them, c1, c2, and c0 are obtained by linear inverse calculation of three points automatically generated by the black film calibration film itself under three known atomization concentrations; the staff only needs to put the black film into the atomization concentration three times, and the system will automatically calculate c1, c2, and c0 and write them back to the JSON configuration file.

[0065] Step 5: Input the defect features and the image features into a preset machine learning-based fusion prediction model, and use visual quality confidence to weight the input of image features. The fusion prediction model outputs the wear state characterizing the grinding defect.

[0066] The defect features and the image features are normalized to ensure that the feature values ​​are within the same range. The defect feature vector and the weighted image feature vector are then concatenated to form a fused feature vector. The fused feature vector is input into a preset machine learning-based fusion prediction model, and the input image features are weighted using visual quality confidence. The fusion prediction model outputs the wear state characterizing the grinding defect.

[0067] In this embodiment, each fused feature vector is assigned a corresponding wear state label, such as "normal," "minor collapse," or "passivation." The labeled fused feature vector dataset is divided into a training set and a test set, with a ratio of 7:3 or 8:2. After the training set is input into the model, the model automatically learns the mapping relationship between the fused feature vectors and the wear state labels. The test set is used to verify the model's performance; after being input into the model, it outputs the predicted wear state.

[0068] The fusion prediction model is specifically a lightweight gradient boosting decision tree model (LightGBM).

[0069] Before inputting the defect features and image features into the fusion prediction model, the image features are first multiplied by the visual quality confidence level W. vision It is used for dynamic adjustment of image feature weights.

[0070] Furthermore, it also includes training samples for enhancing the fusion prediction model:

[0071] Baseline signal templates S of different specifications base,k On (t), a series of preset characteristic signal waveforms simulating defects such as micro-bumps or edge blunting are superimposed to artificially synthesize residual signals with clear defect labels.

[0072] By using the synthesized residual signal and its defect features, combined with a small number of real-collected defect samples, a balanced training set is constructed to train the fusion prediction model, thereby improving the model's ability to identify rare defect samples.

[0073] In this embodiment, sample augmentation is organically combined with model training to effectively address the problem of insufficient training data and improve the model's ability to identify different grinding defects. Simulated data expands the training set, enabling the model to learn from more diverse data, thereby more accurately outputting the wear state of grinding defects in practical applications.

[0074] like Figure 2 As shown in the embodiments of this application, a machine learning-based on-machine wear state prediction system for diamond end mills is provided, including:

[0075] Baseline signal template acquisition module: Acquires the baseline signal template corresponding to the diamond end mill and a specific grinding wheel specification under normal, defect-free grinding conditions;

[0076] Characterization data acquisition module: During the grinding of diamond end mills, sensor data containing at least acoustic emission signals or spindle current signals are acquired in real time, and image data of the diamond end mill cutting edge is also acquired.

[0077] Defect feature extraction module: Performs differential operation on the real-time acquired sensor data and the baseline signal template to generate a residual signal that can characterize grinding anomalies, and extracts the defect features of the residual signal.

[0078] Image feature extraction module: performs quality assessment on the image data, generates visual quality confidence scores, and extracts image features;

[0079] Wear state output module: Input the defect features and the image features into a preset machine learning-based fusion prediction model, and use visual quality confidence to weight the input image features. The fusion prediction model outputs the wear state that represents the grinding defect.

[0080] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0082] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0083] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0085] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for predicting the on-machine wear state of diamond end mills, characterized in that, Includes the following steps: Obtain the baseline signal template corresponding to the diamond end mill and a specific grinding wheel specification under normal, defect-free grinding conditions; During the grinding of diamond end mills, sensor data is collected in real time, and image data of the diamond end mill cutting edge is also collected. The real-time acquired sensor data is differentially processed with the baseline signal template to generate a residual signal that can characterize grinding anomalies, and the defect features of the residual signal are extracted. The image data is subjected to quality assessment, visual quality confidence scores are generated, and image features are extracted. The visual quality confidence level includes the following steps: For each frame of the image, the Laplacian variance is first calculated on the black area to obtain the fogging blur index of the blade edge image. ; The entire cutting edge region is segmented using the Otsu's method, and the proportion of pixels judged as white fog in the cutting edge region is counted to obtain the occlusion rate of the effective cutting edge region. ; The blur index of the cutting edge image is determined by the following functional relationship. Obstruction rate of the effective area of ​​the cutting edge Mapped to visual quality confidence : ； in, To adjust the weighting coefficients for the effects of blur and occlusion, The decision bias is a constant, and e is the natural constant. The defect features and the image features are input into a preset machine learning-based fusion prediction model, and the input image features are weighted using visual quality confidence. The fusion prediction model outputs the wear state characterizing the grinding defect. The fusion prediction model is specifically a lightweight gradient boosting decision tree model; Before inputting defect features and image features into the fusion prediction model, the image features are first multiplied by the visual quality confidence level. It is used for dynamic adjustment of image feature weights.

2. The machine learning-based method for predicting the on-machine wear state of diamond end mills as described in claim 1, characterized in that, The baseline signal template is obtained as follows: For each type of grinding wheel, select multiple new diamond end mills that have been confirmed to have acceptable grinding quality for trial grinding; During the trial grinding process, sensor data, including acoustic emission signals and spindle current signals, were collected. The signals from multiple qualified samples were aligned and averaged to construct a baseline signal template representing the standard grinding process of a diamond end mill under this wheel specification. ,in For marking the specifications of grinding wheels, For time.

3. The machine learning-based method for predicting the on-machine wear state of diamond end mills as described in claim 1, characterized in that, The steps for generating the residual signal include: Real-time collected sensor data According to the current grinding wheel specification markings Baseline signal template invoked The residual signal is obtained by subtraction. : 。 4. The machine learning-based method for predicting the on-machine wear state of diamond end mills as described in claim 1, characterized in that, The extraction of defect features from the residual signal includes the following steps: For the residual signal Time-frequency analysis is performed to statistically analyze the energy integral, kurtosis, and crest factor within the preset defect frequency band, thus constructing a defect feature vector.

5. The machine learning-based method for predicting the on-machine wear state of diamond end mills as described in claim 1, characterized in that, It also includes training samples for the enhanced fusion prediction model: Baseline signal templates of different specifications On top of this, a series of preset, simulated defect characteristic signal waveforms are superimposed to artificially synthesize residual signals with defect labels; The synthesized residual signal and its defect features, combined with real-collected defect samples, are used to form a training set for training the fusion prediction model.

6. A machine learning-based system for predicting the on-machine wear state of diamond end mills, used to implement the machine learning-based method for predicting the on-machine wear state of diamond end mills as described in any one of claims 1-5, characterized in that, include: Baseline signal template acquisition module: Acquires the baseline signal template corresponding to the diamond end mill and a specific grinding wheel specification under normal, defect-free grinding conditions; Characterization data acquisition module: During the grinding of diamond end mills, sensor data containing at least acoustic emission signals or spindle current signals are acquired in real time, and image data of the diamond end mill cutting edge is also acquired. Defect feature extraction module: Performs differential operation on the real-time acquired sensor data and the baseline signal template to generate a residual signal that can characterize grinding anomalies, and extracts the defect features of the residual signal. Image feature extraction module: performs quality assessment on the image data, generates visual quality confidence scores, and extracts image features; The visual quality confidence level includes the following steps: For each frame of the image, the Laplacian variance is first calculated on the black area to obtain the fogging blur index of the blade edge image. ; The entire cutting edge region is segmented using the Otsu's method, and the proportion of pixels judged as white fog in the cutting edge region is counted to obtain the occlusion rate of the effective cutting edge region. ; The blur index of the cutting edge image is determined by the following functional relationship. Obstruction rate of the effective area of ​​the cutting edge Mapped to visual quality confidence : ； in, To adjust the weighting coefficients for the effects of blur and occlusion, The decision bias is a constant, and e is the natural constant. Wear state output module: Input the defect features and the image features into a preset machine learning-based fusion prediction model, and use visual quality confidence to weight the input image features. The fusion prediction model outputs the wear state characterizing the grinding defects. The fusion prediction model is specifically a lightweight gradient boosting decision tree model; Before inputting defect features and image features into the fusion prediction model, the image features are first multiplied by the visual quality confidence level. It is used for dynamic adjustment of image feature weights.

Images