Machine tool condition monitoring and tool wear visual diagnosis system

Abstract

The present application relates to the field of machining monitoring, in particular to a machine tool state monitoring and tool wear visual diagnosis system based on differential geometry, comprising an industrial camera acquisition unit, a differential geometry vibration analysis module, a tool wear evaluation module, an intelligent decision module and a database module, the present application innovatively maps the image and video data collected by the industrial camera to the Riemann manifold representation space, extracts the differential geometry features of vibration; Through manifold representation construction, differential invariant feature extraction, multi-scale decomposition, manifold learning and dimension reduction, accurate identification of the vibration state is realized; Based on the manifold evolution model, the machine tool vibration state is predicted, combined with the tool wear evaluation, the remaining service life is predicted, a non-contact monitoring method is adopted, traditional physical sensors are not needed, and the system complexity is reduced; The vibration analysis based on differential geometry improves the accuracy and robustness of feature extraction.

Description

Technical Field

[0001] This invention relates to the field of machining monitoring, and in particular to a machine tool condition monitoring and tool wear visual diagnosis system based on differential geometry, which is suitable for real-time monitoring of machine tool condition and assessment and prediction of tool wear in the field of precision machining. Background Technology

[0002] In modern manufacturing, machine tools are the core equipment for processing metal parts, and their working condition and tool wear directly affect the machining accuracy and production efficiency. Traditional machine tool condition monitoring mainly relies on physical sensing devices such as accelerometers and current sensors. These sensors need to be directly installed on critical parts of the machine tool, which not only increases system complexity but is also susceptible to the influence of the machine tool's working environment, leading to inaccurate monitoring. At the same time, traditional tool wear monitoring usually adopts offline detection methods, requiring machine shutdown and tool disassembly for measurement, which greatly reduces production efficiency.

[0003] Existing machine tool vibration analysis techniques are mainly based on time-domain and frequency-domain analysis, such as Fast Fourier Transform and wavelet analysis. However, these methods struggle to capture the inherent geometric structure and topological characteristics of complex nonlinear vibrations in machine tools, resulting in limited predictive capabilities for machine tool conditions. Meanwhile, tool wear detection primarily relies on direct measurement and resistance methods, making online real-time monitoring difficult.

[0004] With the development of computer vision technology, image-based machine tool condition monitoring and tool wear diagnosis methods have gradually emerged. However, existing visual diagnostic techniques often only focus on the surface features of images and fail to delve into the geometric structural information contained in the image sequence, making it difficult to accurately predict machine tool condition and tool wear.

[0005] Therefore, there is an urgent need for a machine tool condition monitoring and tool wear diagnosis system that can combine vision technology and advanced mathematical theory to achieve non-contact, high-precision, and predictive monitoring of machine tool condition and tool wear. Summary of the Invention

[0006] The purpose of this invention is to provide a machine tool condition monitoring and tool wear visual diagnosis system based on differential geometry. This system collects image and video data of the machine tool during operation using an industrial camera, and uses differential geometry theory to map vibration information onto the Riemannian manifold space for analysis, so as to achieve accurate identification and prediction of machine tool vibration status. At the same time, it assesses the degree of tool wear and predicts the remaining service life, and finally generates maintenance suggestions to improve machine tool operating efficiency and machining quality.

[0007] This invention proposes a machine tool condition monitoring and tool wear visual diagnosis system, comprising:

[0008] Industrial camera acquisition unit, used to acquire image and video data of cutting tools and workpieces on machine tools;

[0009] The differential geometric vibration analysis module is connected to the industrial camera acquisition unit and is used to map the image and video data to the Riemannian manifold representation space, extract the differential geometric features of vibration, establish a manifold evolution model, and predict the vibration state of the machine tool.

[0010] The tool wear assessment module, connected to the differential geometry vibration analysis module, is used to assess the degree of tool wear and predict the remaining tool life based on the differential geometry characteristics and the manifold evolution model.

[0011] The intelligent decision-making module, connected to the differential geometry vibration analysis module and the tool wear assessment module, is used to generate tool replacement and machine tool maintenance suggestions based on the vibration state prediction and the remaining tool life.

[0012] The database module is connected to the differential geometry vibration analysis module, the tool wear assessment module, and the intelligent decision-making module, respectively, and is used to store historical vibration data, tool wear trend data, and maintenance suggestions.

[0013] Preferably, the differential geometric vibration analysis module includes:

[0014] The data preprocessing unit is used to stabilize the image and video data, extract the region of interest, and perform image enhancement.

[0015] The manifold representation construction unit, connected to the data preprocessing unit, is used to generate a frame difference map sequence, construct a Riemannian manifold representation, and establish a manifold coordinate system.

[0016] The differential invariant feature extraction unit, connected to the manifold characterization construction unit, is used to perform curvature analysis, extract geodesics, perform manifold spectrum analysis, and generate feature vectors describing the intrinsic geometric properties of vibration.

[0017] Preferably, the differential geometric vibration analysis module further includes:

[0018] The manifold multi-scale decomposition unit, connected to the differential invariant feature extraction unit, is used to perform multi-scale decomposition on the feature vector and establish a vibration characterization hierarchy from coarse to fine.

[0019] The manifold learning and dimensionality reduction unit, connected to the manifold multi-scale decomposition unit, is used to perform local structure-preserving mapping, global structure embedding and adaptive dimensionality reduction to generate a dimensionality-reduced vibration feature representation.

[0020] The vibration pattern classification unit, connected to the manifold learning and dimensionality reduction unit, is used to match the dimensionality-reduced vibration feature representation with a pre-stored vibration pattern library, identify vibration pattern categories, detect abnormal vibrations, and classify vibration states into three categories: normal, abnormal, and high-risk.

[0021] Preferably, the differential geometric vibration analysis module further includes:

[0022] The historical trajectory analysis unit, connected to the database module, is used to construct a temporal manifold from the historical vibration data and identify inflection points and singularities in the trajectory.

[0023] An evolution model construction unit, connected to the historical trajectory analysis unit, is used to extract manifold evolution laws from the time-series manifold, establish a geodesic flow model, and verify the accuracy of the evolution model;

[0024] The future state prediction unit is connected to the evolution model construction unit and is used to perform short-term trajectory prediction and medium-to-long-term trend prediction based on the geodesic flow model to determine the optimal maintenance time point.

[0025] Preferably, the industrial camera acquisition unit includes:

[0026] The main camera, mounted above the tool, is used to capture a top view of the tool.

[0027] An auxiliary camera, mounted on the side of the tool, is used to acquire a side view of the tool;

[0028] The camera control module, connected to the main camera and the auxiliary camera, is used to adjust image acquisition parameters, trigger image acquisition, and control image transmission.

[0029] Preferably, the tool wear assessment module includes:

[0030] A tool contour extraction unit is used to extract the tool contour from the image and video data;

[0031] A wear calculation unit, connected to the tool profile extraction unit, is used to calculate the wear amount by comparing the current tool profile with a standard profile.

[0032] The life prediction unit, connected to the wear calculation unit and the differential geometric vibration analysis module, is used to predict the remaining service life of the tool based on the wear amount and the vibration state.

[0033] Preferably, the intelligent decision-making module includes:

[0034] The risk assessment unit is used to assess the machine tool failure risk based on the vibration state prediction and the remaining tool life.

[0035] A maintenance planning unit, connected to the risk assessment unit, is used to generate an optimal maintenance time plan;

[0036] The parameter optimization unit, connected to the risk assessment unit, is used to adjust the processing parameters based on the vibration state prediction.

[0037] Preferably, the system further includes:

[0038] An adaptive learning module is connected to the differential geometry vibration analysis module, the tool wear assessment module, and the database module, respectively, and is used to optimize vibration mode recognition parameters based on historical data, adjust the vibration-wear mapping relationship, and update the manifold evolution model.

[0039] Preferably, the manifold multi-scale decomposition unit decomposes the feature vector into 3-5 levels using spherical wavelet transform, extracting local geometric features at each scale; the manifold learning and dimensionality reduction unit reduces the dimensionality of the feature vector to 10% to 30% of the original dimension; the vibration mode classification unit sets the normal operation threshold to a deviation of less than 0.2, the abnormal warning threshold to a deviation of 0.2-0.5, and the high-risk warning threshold to a deviation of greater than 0.5.

[0040] Preferably, the future state prediction unit performs short-term trajectory prediction for 24 hours, medium-term prediction for 7 days, and long-term prediction for 30 days; wherein, the short-term prediction is updated hourly, and the medium- and long-term predictions are updated daily.

[0041] The beneficial effects of this invention include:

[0042] 1. Non-contact monitoring: Industrial cameras are used instead of traditional physical sensors to achieve non-contact monitoring of machine tool status and tool wear, reducing system complexity and potential failure points, and improving the reliability of the monitoring system.

[0043] 2. Vibration Analysis Based on Differential Geometry: This paper innovatively introduces differential geometry theory for vibration analysis, treating video sequences as spatiotemporal manifolds. By utilizing the geometric properties of the manifold, such as curvature and geodesics, the intrinsic structure of vibration is captured, significantly improving the accuracy and robustness of vibration feature extraction.

[0044] 3. Multi-scale analysis and prediction: Through multi-scale decomposition and manifold learning of the manifold, adaptive identification of vibration modes at different scales is achieved, enabling the detection of minute abnormal vibration modes; at the same time, based on the manifold evolution model, accurate prediction of vibration state is achieved, allowing potential faults to be detected in advance.

[0045] 4. Tool life optimization: By correlating vibration characteristics with tool wear, the remaining tool life can be accurately predicted, avoiding premature tool replacement or overuse, and significantly improving tool utilization efficiency and machining quality.

[0046] 5. Intelligent decision support: Based on vibration prediction and tool life assessment, the system automatically generates optimal maintenance time and parameter adjustment suggestions, reducing unplanned downtime and improving production efficiency. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0048] Figure 2 This is a schematic diagram of the differential geometric vibration analysis module in the system of the present invention;

[0049] Figure 3 This is a flowchart of the manifold characterization construction unit in the system of the present invention;

[0050] Figure 4 This is a schematic diagram of manifold multiscale decomposition in the system of the present invention;

[0051] Figure 5 This is a flowchart of the vibration state prediction based on manifold evolution in the system of the present invention;

[0052] Figure 6 This is a flowchart of the tool wear assessment module in the system of the present invention. Detailed Implementation

[0053] Please refer to Figures 1-6 The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 As shown, the machine tool condition monitoring and tool wear visual diagnosis system provided by the present invention mainly includes an industrial camera acquisition unit 1, a differential geometric vibration analysis module 2, a tool wear assessment module 3, an intelligent decision-making module 4, and a database module 5.

[0055] Industrial camera acquisition unit 1 is used to acquire image and video data of the cutting tools and workpieces on the machine tool. Differential geometry vibration analysis module 2 is connected to industrial camera acquisition unit 1 and is used to map the image and video data to the Riemannian manifold representation space, extract the differential geometric features of vibration, establish a manifold evolution model, and predict the vibration state of the machine tool. Tool wear assessment module 3 is connected to differential geometry vibration analysis module 2 and is used to assess the degree of tool wear and predict the remaining tool life based on differential geometric features and the manifold evolution model. Intelligent decision-making module 4 is connected to differential geometry vibration analysis module 2 and tool wear assessment module 3 and is used to generate tool replacement and machine tool maintenance suggestions based on vibration state prediction and remaining tool life. Database module 5 is connected to differential geometry vibration analysis module 2, tool wear assessment module 3, and intelligent decision-making module 4 respectively and is used to store historical vibration data, tool wear trend data, and maintenance suggestions.

[0056] Preferably, the system may also include an adaptive learning module 6, which is connected to the differential geometry vibration analysis module 2, the tool wear assessment module 3 and the database module 5 respectively, and is used to optimize vibration mode recognition parameters based on historical data, adjust the vibration-wear mapping relationship and update the manifold evolution model.

[0057] like Figure 2 As shown, the differential geometric vibration analysis module 2 is the core innovative part of this invention. It includes a data preprocessing unit 21, a manifold characterization construction unit 22, a differential invariant feature extraction unit 23, a manifold multi-scale decomposition unit 24, a manifold learning and dimensionality reduction unit 25, a vibration mode classification unit 26, a historical trajectory analysis unit 27, an evolution model construction unit 28, and a future state prediction unit 29.

[0058] The data preprocessing unit 21 is used to stabilize the images and video data acquired by the industrial camera, extract regions of interest (ROIs), and perform image enhancement. In one embodiment of the invention, the data preprocessing unit 21 first eliminates minor camera jitter using a region anchoring method. Specifically, 5-10 reference points are set in the static area of ​​the machine tool, the inter-frame displacement vectors of these reference points are calculated, and a global perspective transformation matrix is ​​established for image calibration. Next, the tool and workpiece areas are located through edge detection and morphological manipulation, and adaptive threshold segmentation is applied to extract precise ROIs, generating ROI mask images for subsequent analysis. Finally, the extracted ROI areas are subjected to image enhancement processing, preferably using contrast-limited adaptive histogram equalization (CLAHE) to enhance image details. In practice, the enhancement parameters can be adjusted for different workpiece materials. For example, typical parameters for metal materials are a grid size of 8×8 and a contrast limiting factor of 3.5; while for ceramic materials, the grid size can be set to 12×12 and the contrast limiting factor to 4.2. In addition, a bilateral filter is used to preserve edge details while suppressing noise. The filter parameters are typically set to σ=2.0 in the spatial domain and σ=30 in the value domain.

[0059] The manifold representation construction unit 22 is connected to the data preprocessing unit 21 and is used to generate frame difference map sequences, construct Riemannian manifold representations, and establish a manifold coordinate system. For example... Figure 3 As shown, the process of constructing a manifold representation includes the following steps:

[0060] First, the pixel-level differences between consecutive video frames are calculated, generating a sequence of frame difference maps. These frame difference maps highlight the areas of variation between adjacent frames, primarily caused by machine tool vibration. When calculating the frame differences, a pixel difference threshold of 5 (for an 8-bit grayscale image) is set, retaining only areas with differences greater than the threshold to filter out noise.

[0061] Secondly, each frame of the difference map is treated as a point in a high-dimensional space, and a Riemannian manifold representation is constructed. In this invention, a sliding window with a time window size of 30 frames is preferably used, and the sequence of difference maps within the window is organized into a manifold. For a pair of points i and j on the manifold, the Riemannian metric is defined as follows:

[0062] ,

[0063] in, Let i be the Riemann distance between points i and j. and These are the feature vectors of the corresponding frame difference maps. This is a weight matrix used to adjust the importance of different regions. In practical applications, for precision machine tools, a typical... The matrix assigns higher weights (typically 1.5-2.5 times) to pixels near the cutting area and lower weights (typically 0.5-0.8 times) to pixels far from the cutting area.

[0064] Finally, an intrinsic coordinate system is established on the constructed manifold, and the principal directions and principal curvatures are calculated to map the vibration intensity onto the geometric properties of the manifold. The intrinsic coordinate system is typically established by performing local tangent space analysis on points on the manifold, laying the foundation for subsequent geometric feature extraction.

[0065] The differential invariant feature extraction unit 23 is connected to the manifold characterization construction unit 22, and is used to perform curvature analysis, extract geodesics, perform manifold spectrum analysis, and generate feature vectors describing the intrinsic geometric characteristics of vibration. This is one of the core innovations of this invention: extracting the intrinsic geometric characteristics of vibration through differential geometry theory to achieve accurate characterization of the vibration state.

[0066] First, curvature analysis is performed to calculate the Gaussian curvature and mean curvature of the vibrating manifold. In differential geometry, Gaussian curvature K and mean curvature H are important features describing the local shape of a surface. For parameterized surfaces... The Gaussian curvature and mean curvature can be calculated using the following formulas:

[0067] ,

[0068] in, For Gaussian curvature, , , These are the coefficients of the second basic form. , , is the coefficient of the first fundamental form. The first fundamental form describes the metric properties of the surface, while the second fundamental form describes the curvature of the surface.

[0069] ,

[0070] in, The mean curvature is the standard value; other symbols have the same meaning as above. Gaussian curvature reflects the ellipticity or hyperbolicity of a surface, while mean curvature reflects its convexity or concavity.

[0071] In this system, a 16×16×8 curvature map is constructed, that is, 16×16 points are uniformly sampled on the manifold, and 8 different curvature features are calculated for each point. For machine tools that are operating normally, the standard deviation of Gaussian curvature is usually less than 0.2; while for machine tools with abnormal vibrations, the standard deviation of Gaussian curvature is usually greater than 0.5.

[0072] Secondly, geodesic features are extracted. Geodesics are the shortest paths between two points on a manifold, reflecting the manifold's inherent geometric structure. This system calculates the geodesic distances between key points on the manifold and analyzes the curvature and distribution of geodesics. The geodesic distances are calculated using the FastMarching Method, which obtains the distances by solving the Eikonal equation:

[0073] ,

[0074] in, For the magnitude of the gradient, From the starting point to the point Shortest path time, For the velocity field, The computational domain is defined as follows: In this system, the velocity field... It is related to local curvature; the speed is slower in regions with greater curvature.

[0075] Finally, manifold spectral analysis is performed, constructing the Laplace-Beltramian operator to calculate the spectral characteristics of the manifold. The Laplace-Beltramian operator is a generalization of the Laplace operator on a manifold, defined as:

[0076] ,

[0077] in, The Laplace-Beltrami operator acts on the function As a result, For function gradient, This is the divergence operator. It is obtained by solving the eigenvalue problem:

[0078] ,

[0079] in, For the first One characteristic function, These are the corresponding eigenvalues. The system extracts the top 10 most significant eigenvalues ​​and their corresponding eigenvectors as the spectral features of the vibration. For normally operating machine tools, the eigenvalue sequence typically exhibits a smooth decay; however, for machine tools with abnormalities, abnormal jumps will appear in the eigenvalue sequence.

[0080] Based on the above analysis, the differential invariant feature extraction unit 23 finally generates a feature vector describing the intrinsic geometric characteristics of vibration, which includes curvature features, geodesic features, and spectral features.

[0081] The manifold multi-scale decomposition unit 24 is connected to the differential invariant feature extraction unit 23, and is used to perform multi-scale decomposition of the feature vectors to establish a vibration representation hierarchy from coarse to fine. For example... Figure 4As shown, the multi-scale decomposition process includes:

[0082] First, the spherical wavelet transform is applied to perform multi-scale decomposition of the manifold. The spherical wavelet transform is a generalization of the traditional wavelet transform to spheres or manifolds, and can be expressed as:

[0083] ,

[0084] in, The result of wavelet transform, manifold The function on, Centered on Scale is wavelet basis functions, Let be the volume element on the manifold, and let the integral span the entire manifold. In this system, a four-level decomposition is preferred, which decomposes the vibration characteristics into four submanifolds of different scales. Each submanifold reflects the vibration information within a specific frequency range. For precision machine tools, the frequency ranges corresponding to each scale are typically: Scale 1 (0-50Hz), Scale 2 (50-200Hz), Scale 3 (200-500Hz), and Scale 4 (>500Hz).

[0085] Secondly, local geometric features are extracted at each scale to construct a multi-resolution vibration feature pyramid. For each scale's submanifold, local curvature, geodesic flow, and other geometric features are calculated to form a multi-scale feature representation. In practice, the importance of each scale usually varies, and the weights of each scale can be adjusted for different types of machine tools and machining processes. For example, for high-speed milling, the weights of scales 3 and 4 (corresponding to high-frequency vibration) are usually set higher (e.g., 0.3 and 0.4), while for low-speed turning, the weights of scales 1 and 2 (corresponding to low-frequency vibration) are higher (e.g., 0.4 and 0.3).

[0086] Through multi-scale decomposition, the system can simultaneously analyze vibration characteristics in different frequency ranges, improving the accuracy and robustness of vibration mode recognition.

[0087] The manifold learning and dimensionality reduction unit 25 is connected to the manifold multiscale decomposition unit 24 to perform local structure-preserving mapping, global structure embedding and adaptive dimensionality reduction to generate a dimensionality-reduced vibration feature representation.

[0088] First, a local distance-preserving mapping is constructed to optimize the objective function of preserving local geometry. The goal of the local distance-preserving mapping is to maintain the local distance relationships between points in the original high-dimensional space within the reduced-dimensional space. Specifically, for each point on the manifold... Define its k-nearest neighbor set The objective function for preserving local distance is:

[0089] ,

[0090] in, The objective function for preserving the local structure. The total number of data points. For point The nearest neighbor set, For point and points The weights between them and For point and Mapping in low-dimensional space, The distance is Euclidean. In this system, the number of nearest neighbors is adaptively set according to the vibration complexity. For simple vibration modes It is usually set to 5-8, but for complex vibration modes It can be set to 10-15.

[0091] Secondly, global alignment constraints are constructed to optimize the global topology preservation objective function. The goal of the global alignment constraints is to preserve the global topological structure of the original high-dimensional space in the reduced-dimensional space. The global topology preservation objective function can be expressed as:

[0092] ,

[0093] in, The objective function for preserving the global structure. The global weight reflects the point and Global relationships in the original space.

[0094] The overall objective function is a weighted sum of the local and global objective functions:

[0095] ,

[0096] in, Let be the overall objective function. This is a balancing factor with a value range of [0,1]. It is useful for scenarios requiring the preservation of fine local structures (such as high-precision machining). Typically, a larger value is set (e.g., 0.7); however, for scenarios where the global structure needs to be maintained (e.g., rough machining), a larger value is preferred. It can be set to a smaller value (e.g., 0.3).

[0097] Finally, adaptive dimensionality reduction is performed to determine the optimal embedding dimension. Kernel tricks are then applied to handle nonlinear structures, mapping high-dimensional vibration features to a low-dimensional feature space. In this system, the reduced dimension is typically set to 10%–30% of the original dimension, depending on the complexity of the vibration.

[0098] The vibration mode classification unit 26 is connected to the manifold learning and dimensionality reduction unit 25. It is used to match the dimensionality-reduced vibration feature representation with the pre-stored vibration mode library, identify the vibration mode category, detect abnormal vibration, and classify the vibration state into three categories: normal, abnormal, and high-risk.

[0099] First, the current vibration characteristics are matched against a pre-stored vibration pattern library. The matching process uses cosine similarity to calculate the similarity between the current vibration and each pattern in the library:

[0100] ,

[0101] in, For vectors and Cosine similarity between them This is the current vibration feature vector. The feature vectors in the pattern library, and They are vectors and The One portion, For vector dimensions, Represents the dot product of vectors. and These represent the magnitudes of the vectors.

[0102] Secondly, detect abnormal vibrations. Calculate the deviation of the current vibration mode from the normal vibration mode:

[0103] ,

[0104] Where deviation is the degree of deviation, and normal is the set of normal vibration modes. Represents the current vibration eigenvector Similarity to the one most similar to all normal vibration modes.

[0105] Finally, the vibration status is classified into three categories: normal, abnormal, and high-risk, based on the deviation. The system sets the normal operation threshold to a deviation of less than 0.2, the abnormal warning threshold to a deviation between 0.2 and 0.5, and the high-risk warning threshold to a deviation greater than 0.5. These thresholds are determined based on extensive experimental data and can be fine-tuned according to the specific machine tool type and machining process.

[0106] The historical trajectory analysis unit 27 is connected to the database module 5 and is used to construct a time-series manifold from historical vibration data and identify inflection points and singularities in the trajectory.

[0107] First, the historical vibration data is organized into a time-series manifold. A time-series manifold is a sequence of manifolds arranged in chronological order, reflecting the evolution of the system state over time. When constructing the time-series manifold, an appropriate time step is set, typically 5-10 minutes.

[0108] Secondly, key points, including inflection points and singularities, are identified on the temporal manifold. Inflection points are points where the direction of the manifold trajectory changes significantly, while singularities are points where the local structure of the manifold is abnormal. These key points typically correspond to changes or anomalies in the machine tool's state.

[0109] Inflection points are identified using a curvature thresholding method. The curvature of each point on the manifold trajectory is calculated as follows:

[0110] ,

[0111] in, For time curvature at that point and These are the first and second derivatives of the trajectory, respectively. Represents the cross product. This represents the magnitude of the vector. When the curvature exceeds a threshold (usually set to 2-3 times the average curvature), it is considered an inflection point.

[0112] Singularity identification employs local density analysis. The local density at each point on the manifold is calculated as follows:

[0113] ,

[0114] in, For point Local density, For point and The distance between them For kernel parameters, The total number of data points. This represents an exponential function. A point is considered a singular when its local density deviates significantly from the surrounding region (usually set to deviate by 2 standard deviations).

[0115] By analyzing key points in historical trajectories, we can discover patterns in vibration mode transitions, providing a basis for constructing evolutionary models.

[0116] The evolution model construction unit 28 is connected to the historical trajectory analysis unit 27, and is used to extract the manifold evolution law from the time series manifold, establish a geodesic flow model, and verify the accuracy of the evolution model.

[0117] First, we analyze the temporal laws governing manifold evolution and establish dynamic equations describing manifold deformation. In differential geometry, the evolution of a manifold can be described by vector fields on the manifold. Given an initial manifold... sum vector field The evolution of a manifold over time can be represented as:

[0118] ,

[0119] in, Representing a manifold Regarding time The partial derivatives, In manifold A vector field on the vector field. In this system, the vector field... Learned from historical data.

[0120] Secondly, a geodesic flow model is established. Geodesic flow is the flow along geodesics on a manifold, preserving the manifold's inherent geometric structure. The geodesic flow equation can be expressed as:

[0121] ,

[0122] in, For curves The second covariant derivative, The parameter is denoted by . This equation describes the properties of geodesics: particles moving along geodesics are not subject to any external forces.

[0123] Finally, the accuracy of the evolutionary model is validated. The model is trained using a subset of historical data (typically 80%), and the remaining subset (typically 20%) is used for validation. Model accuracy is measured by prediction error.

[0124] ,

[0125] in, The average prediction error, To verify the number of data points, For model prediction time manifold, For the manifold actually observed, This indicates a suitable distance metric. In practical applications, a prediction error of less than 15% is required for the model to be considered effective.

[0126] The future state prediction unit 29 is connected to the evolution model construction unit 28, and is used to perform short-term trajectory prediction, medium- and long-term trend prediction, and determine the optimal maintenance time point based on the geodesic flow model. For example... Figure 5 As shown, the state prediction process includes:

[0127] First, short-term trajectory prediction is performed to forecast the evolution of the vibration state over the next 24 hours. Short-term predictions are primarily based on the current vibration state and recent evolutionary trends, and their accuracy is typically high (error is usually less than 10%). Short-term predictions are updated hourly to ensure timeliness.

[0128] Secondly, we perform medium- to long-term trend forecasting, predicting the fluctuation trend for the next 1-30 days. This medium- to long-term forecast considers not only the current state but also historical evolution patterns and external factors (such as processing task plans). The medium- to long-term forecast is updated daily.

[0129] Finally, the optimal maintenance time is determined based on the prediction results. The optimal maintenance time needs to balance risk and cost, and can be optimized using the following objective function:

[0130] ,

[0131] in, For optimal maintenance time, This means that the expression takes the minimum value. value, For time The risk function is related to the predicted vibration state. For time The maintenance cost function is related to the production plan and the difficulty of maintenance. In practical applications, the risk function is usually proportional to the predicted vibration deviation, while the cost function is proportional to the degree of interruption in the production plan.

[0132] The industrial camera acquisition unit 1 is the data input terminal of this system, including the main camera, auxiliary camera and camera control module.

[0133] The main camera is mounted above the cutting tool to capture a top view of the tool. The main camera is typically a high-resolution industrial camera with a resolution of at least 1920×1080 pixels and a frame rate of at least 100fps to ensure the capture of high-frequency vibration information. The preferred mounting position for the main camera is 50-80cm above the cutting tool, with a top-down angle of 15-30 degrees.

[0134] An auxiliary camera is mounted on the side of the cutting tool to acquire a side view of the tool. The auxiliary camera's parameters are similar to the main camera, but it is mounted 30-50cm from the side of the tool at a horizontal angle of 60-75 degrees. Through the coordinated operation of the main and auxiliary cameras, the system can acquire three-dimensional information about the cutting tool.

[0135] The camera control module connects to the main camera and auxiliary camera, and is used to adjust image acquisition parameters, trigger image acquisition, and control image transmission. The camera control module can dynamically adjust parameters such as exposure time, gain, and frame rate according to different stages of the processing.

[0136] like Figure 6 As shown, the tool wear assessment module 3 is a key component of this system, including a tool profile extraction unit 31, a wear calculation unit 32, and a life prediction unit 33.

[0137] The tool contour extraction unit 31 is used to extract the tool contour from image and video data. In embodiments of the present invention, contour extraction employs a deep learning-based semantic segmentation method to accurately segment the tool region in the image. The segmentation model typically uses network structures such as U-Net or DeepLab, trained on a large amount of labeled data.

[0138] Wear calculation unit 32 is connected to tool profile extraction unit 31 and is used to calculate wear amount by comparing the current tool profile with a standard profile. Wear amount calculation is based on multiple indicators, such as profile area reduction, edge roughness increase, and angle change. The Comprehensive Wear Index (CMI) is calculated using the following formula:

[0139] ,

[0140] in, The comprehensive wear index, and These are the areas of the standard contour and the current contour, respectively. and These are the edge roughness values ​​for the standard profile and the current profile, respectively. and These are the feature angles of the standard contour and the current contour, respectively. , and This is a weighting factor, adjusted according to tool type and machining process. For turning tools, a typical weighting setting is... , , For milling tools, the typical weighting is set to... , , .

[0141] The tool life prediction unit 33 is connected to the wear calculation unit 32 and the differential geometric vibration analysis module 2, and is used to predict the remaining tool life based on wear amount and vibration state. In this system, tool life prediction adopts a data-driven method, taking wear amount, vibration characteristics, and machining parameters as inputs to predict the remaining tool life. Specifically, the remaining tool life... It can be estimated using the following model:

[0142] ,

[0143] in, The predicted remaining service life (usually expressed in hours or number of processed parts). The comprehensive wear index, For vibration characteristic vectors, For the processing parameter vector, For prediction models, random forests or neural networks are typically used.

[0144] The intelligent decision-making module 4 is the decision output end of this system, including the risk assessment unit 41, the maintenance planning unit 42, and the parameter optimization unit 43.

[0145] Risk assessment unit 41 is used to assess machine tool failure risk based on vibration state prediction and remaining tool life. The risk assessment employs a multi-factor comprehensive analysis method, comprehensively considering factors such as vibration state, tool life, and historical failure data to generate a risk index. The risk index is typically divided into 5 levels, from level 1 (low risk) to level 5 (extremely high risk).

[0146] The maintenance planning unit 42 is connected to the risk assessment unit 41 and is used to generate an optimal maintenance schedule. Maintenance planning needs to comprehensively consider risk levels, production plans, and maintenance costs to find the best balance. For example, for a risk level of 4-5, the system will suggest maintenance during the nearest production break; for a risk level of 2-3, the system will suggest maintenance on weekends or during planned downtime; and for a risk level of 1, the system will suggest checking during the next routine maintenance.

[0147] The parameter optimization unit 43 is connected to the risk assessment unit 41 and is used to adjust machining parameters based on vibration state prediction. Parameter optimization adopts a model predictive control method to predict the vibration and wear state under different parameter settings and select the optimal parameter combination. The optimized parameters typically include cutting speed, feed rate, and depth of cut.

[0148] Database module 5 is used to store historical vibration data, tool wear trend data, and maintenance recommendations. This module contains several sub-databases, such as a vibration characteristics database, a tool wear database, and a failure case database.

[0149] The adaptive learning module 6 is connected to the differential geometry vibration analysis module 2, the tool wear assessment module 3, and the database module 5, respectively. It is used to optimize vibration mode recognition parameters based on historical data, adjust the vibration-wear mapping relationship, and update the manifold evolution model. Adaptive learning enables the system to continuously learn and optimize, adapting to different machine tools, cutting tools, and machining processes.

[0150] During adaptive learning, the system periodically (typically every 200 hours) performs a complete model update, including retraining the vibration mode classification model, updating the vibration-wear mapping, and optimizing the manifold evolution prediction model. In addition, the system performs incremental learning every 24 hours, fine-tuning the model parameters based on newly collected data.

[0151] In a preferred embodiment of the present invention, the manifold multi-scale decomposition unit 24 performs a four-level decomposition of the feature vector using spherical wavelet transform, extracting local geometric features at each scale; the manifold learning and dimensionality reduction unit 25 reduces the dimensionality of the feature vector to 15% to 25% of the original dimension; the vibration mode classification unit 26 sets the normal operation threshold to a deviation of less than 0.2, the abnormal warning threshold to a deviation of 0.2-0.5, and the high-risk warning threshold to a deviation of greater than 0.5.

[0152] In another preferred embodiment, the future state prediction unit 29 performs short-term trajectory predictions for 24 hours, medium-term predictions for 7 days, and long-term predictions for 30 days; the short-term predictions are updated hourly, while the medium- and long-term predictions are updated daily. Short-term predictions are primarily used to guide daily operations and immediate maintenance decisions, while medium- and long-term predictions are used for maintenance planning and resource scheduling.

[0153] This invention innovatively applies differential geometry theory to machine tool condition monitoring and tool wear diagnosis, and achieves non-contact, high-precision monitoring through vision technology, providing technical support for intelligent manufacturing.

[0154] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A machine tool condition monitoring and tool wear visual diagnosis system, characterized in that, include: Industrial camera acquisition unit, used to acquire images and video data of cutting tools and workpieces on machine tools; The differential geometric vibration analysis module is connected to the industrial camera acquisition unit and is used to map the image and video data to the Riemannian manifold representation space, extract the differential geometric features of vibration, establish a manifold evolution model, and predict the vibration state of the machine tool. The tool wear assessment module, connected to the differential geometry vibration analysis module, is used to assess the degree of tool wear and predict the remaining tool life based on the differential geometry characteristics and the manifold evolution model. The intelligent decision-making module, connected to the differential geometry vibration analysis module and the tool wear assessment module, is used to generate tool replacement and machine tool maintenance suggestions based on vibration state prediction and the remaining tool life. The database module is connected to the differential geometry vibration analysis module, the tool wear assessment module, and the intelligent decision-making module, respectively, and is used to store historical vibration data, tool wear trend data, and maintenance suggestions. The differential geometric vibration analysis module includes: The data preprocessing unit is used to stabilize the image and video data, extract the region of interest, and perform image enhancement. The manifold representation construction unit, connected to the data preprocessing unit, is used to generate a frame difference map sequence, construct a Riemannian manifold representation, and establish a manifold coordinate system. The differential invariant feature extraction unit is connected to the manifold characterization construction unit and is used to perform curvature analysis, extract geodesics, perform manifold spectrum analysis, and generate feature vectors describing the intrinsic geometric properties of vibration. The differential geometric vibration analysis module also includes: The manifold multi-scale decomposition unit, connected to the differential invariant feature extraction unit, is used to perform multi-scale decomposition on the feature vector and establish a vibration characterization hierarchy from coarse to fine. The manifold learning and dimensionality reduction unit, connected to the manifold multi-scale decomposition unit, is used to perform local structure-preserving mapping, global structure embedding and adaptive dimensionality reduction to generate a dimensionality-reduced vibration feature representation. The vibration pattern classification unit, connected to the manifold learning and dimensionality reduction unit, is used to match the dimensionality-reduced vibration feature representation with the pre-stored vibration pattern library, identify vibration pattern categories, detect abnormal vibrations, and classify vibration states into three categories: normal, abnormal, and high-risk. The differential geometric vibration analysis module also includes: The historical trajectory analysis unit, connected to the database module, is used to construct a temporal manifold from the historical vibration data and identify inflection points and singularities in the trajectory. An evolution model construction unit, connected to the historical trajectory analysis unit, is used to extract manifold evolution laws from the time-series manifold, establish a geodesic flow model, and verify the accuracy of the evolution model; The future state prediction unit is connected to the evolution model construction unit and is used to perform short-term trajectory prediction and medium-to-long-term trend prediction based on the geodesic flow model to determine the optimal maintenance time point.

2. The machine tool condition monitoring and tool wear visual diagnosis system according to claim 1, characterized in that, The industrial camera acquisition unit includes: The main camera, mounted above the tool, is used to capture a top view of the tool. An auxiliary camera, mounted on the side of the tool, is used to acquire a side view of the tool; The camera control module, connected to the main camera and the auxiliary camera, is used to adjust image acquisition parameters, trigger image acquisition, and control image transmission.

3. The machine tool condition monitoring and tool wear visual diagnosis system according to claim 1, characterized in that, The tool wear assessment module includes: A tool contour extraction unit is used to extract the tool contour from the image and video data; A wear calculation unit, connected to the tool profile extraction unit, is used to calculate the wear amount by comparing the current tool profile with a standard profile. The life prediction unit, connected to the wear calculation unit and the differential geometric vibration analysis module, is used to predict the remaining service life of the tool based on the wear amount and vibration state.

4. The machine tool condition monitoring and tool wear visual diagnosis system according to claim 1, characterized in that, The intelligent decision-making module includes: The risk assessment unit is used to assess the machine tool failure risk based on vibration state prediction and the remaining service life of the cutting tool; A maintenance planning unit, connected to the risk assessment unit, is used to generate an optimal maintenance time plan; The parameter optimization unit, connected to the risk assessment unit, is used to adjust the processing parameters based on vibration state prediction.

5. The machine tool condition monitoring and tool wear visual diagnosis system according to claim 1, characterized in that, The system also includes: An adaptive learning module is connected to the differential geometry vibration analysis module, the tool wear assessment module, and the database module, respectively, and is used to optimize vibration mode recognition parameters based on historical data, adjust the vibration-wear mapping relationship, and update the manifold evolution model.

6. The machine tool condition monitoring and tool wear visual diagnosis system according to claim 1, characterized in that, The manifold multi-scale decomposition unit decomposes the feature vector into 3-5 levels using spherical wavelet transform, extracting local geometric features at each scale; the manifold learning and dimensionality reduction unit reduces the dimensionality of the feature vector to 10%-30% of the original dimension; the vibration mode classification unit sets the normal operation threshold to a deviation of less than 0.2, the abnormal warning threshold to a deviation of 0.2-0.5, and the high-risk warning threshold to a deviation of greater than 0.

5.

7. The machine tool condition monitoring and tool wear visual diagnosis system according to claim 1, characterized in that, The future state prediction unit performs short-term trajectory predictions for 24 hours, medium-term predictions for 7 days, and long-term predictions for 30 days; short-term predictions are updated hourly, while medium- and long-term predictions are updated daily.

Images