Intelligent detection method and system for support turning

Abstract

The present application relates to the technical field of image processing, in particular to a kind of intelligent detection method and system for support turning processing.The method includes: real-time acquisition sensor signal in the turning processing process of support;Sensor topological graph is constructed;Extract the real-time time series process feature reflecting tool-workpiece-machine tool system dynamic response in sensor topological graph;The real-time time series process feature is input into trained multi-task prediction model, and the predicted deviation value of multiple geometric dimensions of support and predicted surface defect mask chart are output;If all predicted deviation values are within the corresponding tolerance threshold, and the ratio of the defect area of predicted surface defect mask chart and the total area of support is less than the set area threshold, then determine that the expected quality state of the support is qualified.The scheme of the present application can comprehensively and real-time determine the expected quality state of support product.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to an intelligent detection method and system for support turning. BACKGROUND

[0002] As a kind of precision implantable medical device, the geometric dimension precision and surface integrity of the stent are crucial to the clinical performance and service life of the product. In the production and manufacturing process of the stent, turning is the core process to ensure the key dimensions and surface quality of the stent.

[0003] The traditional stent quality control method mainly relies on offline detection after processing, that is, through three-coordinate measuring machine, industrial CT, optical profiler or high-power microscope and other equipment to carry out sampling inspection or full inspection on the finished product. This post-detection mode has significant lag, which cannot find and warn the quality abnormality in real time during the processing, resulting in that once the waste product is produced, it cannot be recovered, which not only causes waste of materials and working hours, but also significantly increases the production cost and reduces the overall manufacturing efficiency. In addition, the offline detection obtains the final quality result, which is difficult to establish a direct and accurate causal relationship with the dynamic changes in the processing, so it cannot provide effective guidance for real-time optimization and adjustment of process parameters, which limits the intelligent and automatic level of the stent manufacturing process.

[0004] In order to overcome the disadvantages of offline detection, the academia and industry have begun to study the online monitoring technology of processing based on sensor signals. These technologies usually deploy force, vibration, acoustic emission, power or temperature sensors on the machine tool, trying to indirectly judge the processing state by analyzing the real-time collected signals. However, the existing online monitoring technology still has many limitations, which are as follows:

[0005] In the aspect of information fusion, most methods analyze the sensor signals from different physical sources independently or simply weighted fusion, ignoring the complex physical coupling relationship and dynamic interaction between the components inside the tool-workpiece-machine tool system, resulting in that the extracted features cannot comprehensively and accurately represent the overall dynamic response of the system.

[0006] In the modeling of features and quality, the existing technology mostly uses traditional signal processing methods combined with shallow machine learning models, which is difficult to mine deep features from high-dimensional, nonlinear sensor signals that are strongly related to the final part geometric dimension deviation and microscopic surface defects. Moreover, most prediction models are only for a single quality indicator (such as only predicting surface roughness or only predicting tool wear), which cannot simultaneously perform parallel and comprehensive prediction and evaluation on the two core quality elements of dimensional accuracy and surface defects, so it is difficult to realize the comprehensive and real-time determination of the expected quality state of the stent product. SUMMARY

[0007] The purpose of the present application is to provide an intelligent detection method and system for support turning machining, to solve the problem that it is difficult to realize comprehensive and real-time judgment of the expected quality state of the support product in the prior art. To this end, the present application provides a scheme in the following two aspects.

[0008] In the first aspect, the present application provides an intelligent detection method for support turning machining, comprising:

[0009] Real-time acquisition of sensor signals of the support in the turning machining process, the sensor signals including three-axis vibration signals, acoustic emission signals and spindle load signals;

[0010] Constructing a sensor topology graph, the nodes of the sensor topology graph being any one of the in-situ sensor signals, and the connection relationship between the nodes being the coupling strength between the corresponding two signals;

[0011] Extracting real-time time series process features reflecting the dynamic response of the tool-workpiece-machine tool system in the sensor topology graph;

[0012] Inputting the real-time time series process features into a trained multi-task prediction model to output predicted deviation values of multiple geometric dimensions of the support and a predicted surface defect mask map;

[0013] If all the predicted deviation values are within the corresponding tolerance threshold, and the ratio of the defect area of the predicted surface defect mask map to the total area of the support is less than a set area threshold, it is determined that the expected quality state of the support is qualified.

[0014] In the above scheme, the turning machine sensor signals of the support during the machining process are graphically structured, and real-time time series process features reflecting the dynamic response of the tool-workpiece-machine tool system are extracted therefrom, and a multi-task prediction model is used to predict the real-time time series process features to obtain predicted deviation values of multiple geometric dimensions of the support and a predicted surface defect mask map, so as to realize the evaluation of whether the expected quality state of the support is qualified. That is, the scheme of the present application can comprehensively and real-time evaluate the expected quality state of the support.

[0015] Optionally, the multiple geometric dimensions include: the maximum outer diameter dimension of the support body, the inner diameter dimension of the support center through hole, and the flatness of the support mounting flange end face.

[0016] The key geometric dimensions of the support are selected, which provides data support for subsequent detection of the support.

[0017] Optionally, the multi-task prediction model comprises a decoder and an encoder, the decoder is a long short-term memory network model, the encoder comprises two branches, a first branch is composed of multiple connection layers to output a plurality of geometric size prediction deviation values, and a loss function is a mean square error loss function; a second branch adopts a U-Net network structure, and a loss function is a combination of a Dice loss and a cross-entropy loss.

[0018] Optionally, the training process of the multi-task prediction model is as follows:

[0019] A training set is obtained, the training set comprises time sequence process features of sensor signals of a plurality of stent samples in a processing process and corresponding labels, the labels comprise deviation values of a plurality of geometric sizes and a surface defect map;

[0020] The multi-task prediction model is trained by using the training set, and the losses of the first branch and the second branch are calculated respectively, the parameters of the network prediction model are adjusted by using a gradient descent algorithm, and the training is stopped when the loss value between the prediction values output by the two branches and the labels is less than a threshold value or the number of training reaches a set number of times, and a trained network prediction model is obtained.

[0021] Through the training of the multi-task prediction model, the robustness of the multi-task prediction model can be improved.

[0022] Optionally, the deviation values of the plurality of geometric sizes and the surface defect map are obtained as follows:

[0023] Three-dimensional point cloud data of the stent sample is obtained, and the three-dimensional point cloud data is aligned and registered with a standard CAD model of the stent sample to calculate deviation values of geometric sizes;

[0024] The point cloud whose deviation value of the geometric size is greater than a set deviation threshold value is marked as a surface defect to obtain a surface defect map of the three-dimensional point cloud data.

[0025] The stent sample is labeled to provide a supervision signal for the training of the multi-task prediction model and drive the update of model parameters.

[0026] Optionally, the real-time time sequence process feature is obtained by using a spatio-temporal graph convolution network model.

[0027] Optionally, the sensor topology graph is obtained by processing the sensor signals by using a sensor topology graph neural network.

[0028] Optionally, the connection relationship between the nodes comprises a spatial edge, a functional edge and a dynamic edge; the spatial edge is the inverse of the Euclidean distance between sensors in spatial layout; the functional edge is the causality or correlation strength between the time series of the signals of the sensors corresponding to different two nodes analyzed by Granger causality test or mutual information algorithm; and the dynamic edge is based on the signal similarity between the time series of the signals of the sensors corresponding to two nodes.

[0029] The connection relationship can quantize the strength of the relationship between the nodes, thereby driving more fine graph algorithm behaviors.

[0030] Optionally, the step of acquiring the sensor signals comprises: installing at least one vibration sensor on a machine tool tool holder to collect multi-axis vibration signals; installing an acoustic emission sensor on a workpiece clamp or a tool holder close to a cutting area to collect acoustic emission signals; and acquiring a spindle load signal from a machine tool control system, the spindle load signal being a current signal.

[0031] In a second aspect, an intelligent detection system for support turning machining comprises:

[0032] a processor;

[0033] a memory storing computer instructions for intelligent detection of support turning machining, when the computer instructions are executed by the processor, the system executes the above-mentioned intelligent detection method for support turning machining.

[0034] The present application has the following advantages:

[0035] The present application uses a sensor topological graph neural network to explicitly model the coupling relationship between different sensors as a graph structure, which can extract real-time time series process features from multi-source sensor signals that can more comprehensively and accurately reflect the overall state of the tool-workpiece-machine system, overcoming the defects of traditional methods that simply fuse signals and lose key associated information. Further, by constructing a multi-task prediction model that associates sensor signals with finished support three-dimensional point cloud data, the problem of being difficult to synchronously and comprehensively evaluate multi-dimensional quality in the prior art is solved. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 A step flowchart of an intelligent detection method for support turning machining in the embodiment is schematically shown;

[0037] Figure 2 A structural block diagram of an intelligent detection system for support turning machining in the embodiment is schematically shown. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0039] As shown in the figure, the intelligent detection method for support turning machining in the embodiment includes the following steps: Figure 1

[0040] Step S1, real-time acquisition of sensor signals of vibration, acoustic emission and spindle load of the support in the turning machining process.

[0041] Specifically, the sensor signals include X, Y and Z axis vibration signals collected by at least one three-direction vibration sensor installed on the machine tool turret; high-frequency stress wave signals collected by an acoustic emission sensor installed on the workpiece clamp or the turret near the cutting area; and current signals (i.e., spindle load signals) reflecting the change of spindle load collected by a current sensor integrated on the machine tool spindle drive motor.

[0042] In the embodiment, a high-speed data acquisition card, such as a PXI series acquisition card of NI Company, is used to synchronously collect the vibration and load signals at a sampling frequency of not less than 50 kHz, and collect the acoustic emission signals at a sampling frequency of not less than 2 MHz, to form continuous, multi-channel time sequence data streams, and transmit them to an edge computing unit or an industrial computer in real time for subsequent processing.

[0043] Among them, the sampling frequency of the three-axis accelerometer is usually set to 25.6 kHz to cover the main dynamic information in the machining process, such as chatter characteristic frequency; the signal frequency range collected by the acoustic emission sensor can reach 100 to 400 kHz, which can accurately capture the instantaneous signals of tool micro-chip or chip morphology transition. The spindle motor load data (current signal) is read from the servo driver of the machine tool numerical control system at a frequency of 100 times per second, which can reflect the macroscopic change trend of cutting force, such as the gradual wear of the tool, which will show a slow upward trend in spindle load.

[0044] Step S2, constructing a sensor topology graph, the nodes of the sensor topology graph being any one of the in-situ sensor signals, and the connection relationship between the nodes being the physical coupling strength between the corresponding two signals; extracting real-time time sequence process features reflecting the dynamic response of the tool-workpiece-machine tool system in the sensor topology graph.

[0045] Specifically, a sensor topology graph is constructed, in which each node represents a sensor signal, such as X-axis vibration signal, Y-axis vibration signal, acoustic emission channel, etc.; and the connection relationship between the nodes represents the coupling strength between different signal sources.

[0046] ​Exemplarily, the multi-source sensor system is abstracted as an information-physical graph. Each node in the graph represents an independent signal stream. For example, in a system, there can be five nodes representing X-axis vibration, Y-axis vibration, Z-axis vibration, acoustic emission signal, and main shaft load signal. The feature of each node at any time is the original signal data or the extracted feature vector of the corresponding sensor signal channel in the corresponding time window.

[0047] In an embodiment, the sensor signals are processed using a sensor topology graph neural network to construct a graph structure.

[0048] The weights of the connection edges between the nodes can be obtained empirically. For example, since the three vibration sensor nodes are physically closely coupled, the edge weights between the three vibration sensor nodes are set to be relatively high, such as 0.9. The edge weights between the vibration sensor nodes and the main shaft motor load node are set according to the force transmission path in the machine tool structure. The longer the path and the more complex the connection, the lower the weight, such as 0.4.

[0049] As other implementations, Granger causality test or mutual information algorithm can also be used to analyze the causality or correlation strength between different sensor signal time series, and the calculated quantitative values are used as the weights of the edges in the graph.

[0050] The above-mentioned use of the sensor topology graph neural network in sensor information aggregation can more focusedly integrate sensor information with stronger physical correlation, thereby more deeply understanding the complex evolution process of the machining state.

[0051] The real-time time series process features are input into a spatio-temporal graph convolution network (STGCN) model through the graph structure.

[0052] The spatio-temporal graph convolution network model (STGCN) propagates and aggregates information in the sensor spatial dimension through graph convolution operation, and captures the time evolution law of each sensor signal itself through time convolution operation, to output low-dimensional, high-information-density time series process features that integrate the physical coupling relationship of the system.

[0053] In step S3, the real-time time series process features are input into a pre-trained multi-task prediction model to output the predicted deviation values of the multiple geometric dimensions of the support and the predicted surface defect mask map.

[0054] The multi-task prediction model adopts an encoder-decoder architecture. The encoder part adopts a long short-term memory network model (LSTM) to encode the extracted time series process features into a fixed-length context vector, which contains the comprehensive state information of the entire machining process.

[0055] The decoder comprises two branches, the first branch is a size deviation regression head composed of multiple fully connected layers, and the output thereof is three or more continuous values respectively corresponding to the size deviations of the inner diameter, the outer diameter and the rib width of the support, and the first branch adopts a mean square error loss function.

[0056] The second branch is a surface defect segmentation head, the structure of which adopts a U-Net network structure, and the loss function thereof adopts a combination of a Dice loss and a cross-entropy loss; the context vector is decoded and up-sampled, and a defect mask image consistent in size with a two-dimensional diagram developed on the surface of the support is output, and the numerical value of each pixel point in the image represents the probability of the existence of a defect at the position, and the category of the defect, such as a scratch, a burr or a pit, is also output.

[0057] The pre-training process of the multi-task prediction model is as follows:

[0058] First, a training set is obtained.

[0059] The training set comprises historical sensor signals of multiple support samples and corresponding labels.

[0060] The historical sensor signals are running state data of a lathe in a machining process, and the labels are deviation values of geometric dimensions of the support and a surface defect map of the support.

[0061] The multiple support samples can comprise 200 support samples in one sample batch.

[0062] The geometric dimensions of the support in the above include an outer diameter dimension of a support body, an inner diameter dimension of a support center through hole and a flatness of a support mounting flange end face. The maximum outer diameter dimension determines whether the support can be successfully installed into a matched housing or cavity; the support center through hole is used for mounting a rotating shaft or a fixing bolt, and the inner diameter dimension precision directly affects the stability and dynamic balance of the connection; the flatness of the mounting flange end face is the basis for ensuring that the support closely fits with other components and is uniformly stressed, and if the flatness prediction value exceeds the tolerance requirement of 0.03 millimeters, stress concentration and small gaps will be generated after assembly, which may cause vibration and fatigue fracture under long-term operation.

[0063] The defect results include defect categories, wherein the defect categories include surface scratches caused by chip winding, machining surface chatter of the support caused by tool wear, or surface burrs caused by the shedding of built-up edge. Generally, the machining surface chatter caused by tool wear is manifested as the formation of regular ripples on the workpiece surface, which is caused by the increased tool relief wear, leading to a decrease in cutting system rigidity and high-frequency chatter. For example, when the vibration sensor detects that the main vibration frequency deviates from the normal 200 Hz and stabilizes at 650 Hz, it often indicates that the surface has generated chatter, and the roughness value may deteriorate from Ra0.8 microns to Ra3.2 microns. The surface burrs caused by the shedding of built-up edge are small welds formed by chip material at high temperature and high pressure near the tool tip. The welds are unstable and periodically shed, and part of them are embedded in the newly machined surface to form hard points or burrs, which are usually associated with small fluctuations in the spindle load signal.

[0064] The deviation value of the geometric size of the support is obtained by using a three-dimensional measuring device (such as a high-precision laser three-dimensional scanner) to obtain the three-dimensional point cloud data of each support sample; aligning and registering the three-dimensional point cloud data with the standard CAD model of the support, and automatically calculating the absolute value of the difference between the actual value and the design value of the geometric size by software as the deviation value of the corresponding geometric size.

[0065] Exemplarily, after the three-dimensional point cloud data is aligned and registered with the standard CAD model of the support, the deviation values of the geometric sizes can be obtained, such as an outer diameter deviation of 0.08 mm, a flatness deviation of 0.04 mm, and other specific numerical labels.

[0066] The surface defect map of the support is obtained by marking the point cloud with a deviation value greater than the set deviation threshold as a surface defect, so as to obtain a binary defect map of the support sample, and the pixel value of the marked surface defect is 1 and the pixel value of the non-defect is 0.

[0067] The set deviation threshold value can be 20 microns. Of course, as other embodiments, it can also be determined according to the actual situation.

[0068] It should be noted that when obtaining the training set, different cutting parameter combinations need to be introduced in the machining process of the support sample to cover various quality states from qualified to unqualified. For each support sample, the corresponding multi-channel sensor data from the blank clamping to the finished unloading process are recorded completely and uniquely bound to the corresponding support sample.

[0069] After obtaining the trained multi-task prediction model, the real-time time series process features are input into the trained multi-task prediction model to obtain the predicted deviation value and the predicted surface defect mask map.

[0070] Step S4, according to the predicted deviation values of the plurality of geometric dimensions of the stent and the predicted surface defect mask map, the expected quality state of the stent is determined in real time.

[0071] Specifically, the evaluation process of the expected quality state of the stent in the embodiment is as follows:

[0072] If the predicted deviation values of all geometric dimensions are within the corresponding tolerance threshold, and the predicted defects do not meet the preset defect determination criteria, it is determined that the expected quality state of the stent is qualified; otherwise, if the predicted deviation value of any geometric dimension exceeds the tolerance threshold or the predicted defect meets the preset defect determination criteria, it is determined to be unqualified, and a warning information can be given according to the specific predicted deviation value or defect result.

[0073] The preset defect determination criteria in the above include that the ratio of the defect area of the predicted surface defect mask map to the total area of the stent is greater than or equal to a preset area threshold. The total area of the stent is the sum of the pixel points in the image of the stent taken.

[0074] The preset area threshold has different values according to different defects, for example, the preset threshold corresponding to surface scratches and burrs is 0.2%.

[0075] The tolerance threshold includes an inner diameter tolerance threshold, an outer diameter tolerance threshold, and a rib width tolerance threshold, and the specific values can be 5 microns, 5 microns, and 3 microns, respectively.

[0076] The above-mentioned quality determination logic through the establishment of a double check can accurately evaluate the quality of the stent.

[0077] The scheme of the present application can predict the deviation values of the geometric dimensions of the stent and the surface defects, which are two core quality indicators, in parallel from the machining data of the machining process, and thus the expected quality of the stent can be evaluated in time and effectively.

[0078] The present application also provides an intelligent detection system for stent turning machining. Figure 2 As shown in the figure, the system comprises a processor and a memory, and the memory stores computer program instructions, which realize the intelligent detection method for stent turning machining according to the present application when executed by the processor.

[0079] The system also includes a communication bus and a communication interface and other components familiar to those skilled in the art, and their settings and functions are known in the art, so they will not be described here.

[0080] In this description, the term "application" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "memory" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "computer-readable medium" means any tangible medium that stores, communicates, or otherwise provides data that can be used by an instruction execution system, apparatus or device. The computer-readable medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as, for example, resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), and the like, or any other medium that can be used to store the desired information and that can be accessed by an application, module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto. Any application or module described in this description can be implemented by computer-readable / executable instructions stored or otherwise held by such computer-readable media.

[0081] In the description of the present description, the meaning of "a plurality of" is at least two, for example, two, three or more, and the like, unless otherwise explicitly specified.

[0082] Although the present description has shown and described a number of embodiments of the present application, it will be apparent to those skilled in the art that many modifications, variations, and alternatives to the embodiments described herein can be made in light of the teachings herein.

Claims

1. An intelligent detection method for stent turning, characterized in that: The following steps are involved: Acquiring sensor signals of the bracket during the turning process in real time, wherein the sensor signals include a three-axis vibration signal, an acoustic emission signal, and a spindle load signal; Constructing a sensor topology graph, wherein the nodes of the sensor topology graph are any of the in-situ sensor signals, and the connection relationship between the nodes is the coupling strength between the corresponding two signals; Extract the real-time time series process characteristics reflecting the dynamic response of the tool-workpiece-machine system from the sensor topology map; Inputting the real-time time series process features into a trained multi-task prediction model, and outputting predicted deviation values ​​of multiple geometric dimensions of the bracket and a predicted surface defect mask map; If all predicted deviation values ​​are within the corresponding tolerance threshold, and the ratio of the defect area of ​​the predicted surface defect mask image to the total area of ​​the bracket is less than the set area threshold, the expected quality status of the bracket is determined to be qualified.

2. The intelligent detection method for stent turning according to claim 1, characterized in that: The multiple geometric dimensions include: the maximum outer diameter of the bracket body, the inner diameter of the bracket center through hole, and the flatness of the end surface of the bracket mounting flange.

3. The intelligent detection method for stent turning according to claim 1, characterized in that: The multi-task prediction model includes a decoder and an encoder. The decoder is a long short-term memory network model. The encoder includes two branches. The first branch is composed of multiple connection layers to output prediction deviation values ​​of multiple geometric dimensions, and the loss function is a mean square error loss function; the second branch adopts a U-Net network structure, and the loss function is a combination of Dice loss and cross entropy loss.

4. The intelligent detection method for stent turning according to claim 3, characterized in that: The training process of the multi-task prediction model is as follows: Acquire a training set; the training set includes time series process characteristics of sensor signals of multiple stent samples during processing and corresponding labels; the labels include deviation values ​​of multiple geometric dimensions and surface defect maps; The multi-task prediction model is trained using the training set, and the losses of the first branch and the second branch are calculated respectively. The parameters of the network prediction model are adjusted using the gradient descent algorithm until the loss value between the predicted value and the label output by the two branches is less than the threshold or the number of training times reaches the set number. The training is stopped and the trained network prediction model is obtained.

5. The intelligent detection method for stent turning according to claim 4, characterized in that: The process of obtaining the deviation values ​​of the multiple geometric dimensions and the surface defect map is as follows: Acquire three-dimensional point cloud data of the stent sample; align and register the three-dimensional point cloud data with a standard CAD digital model of the stent sample to calculate deviation values ​​of various geometric dimensions; Point clouds whose geometric dimension deviation values ​​are greater than the set deviation threshold are marked as surface defects to obtain a surface defect map of the three-dimensional point cloud data.

6. The intelligent detection method for stent turning according to claim 1, characterized in that: The real-time time series process features are obtained using a spatiotemporal graph convolutional network model.

7. The intelligent detection method for stent turning according to claim 1, characterized in that: The sensor topology map is obtained by processing the sensor signal using a sensor topology map neural network.

8. The intelligent detection method for stent turning according to claim 7, characterized in that: The connection relationship between the nodes includes spatial edges, functional edges and dynamic edges; the spatial edge is the inverse of the Euclidean distance between sensors in the spatial layout; the functional edge is the causal or correlation strength between the sensor signal time series corresponding to two different nodes using Granger causality test or mutual information algorithm; the dynamic edge is based on the signal similarity between the sensor signal time series corresponding to two nodes.

9. The intelligent detection method for stent turning according to claim 1, characterized in that: The sensor signal acquisition step includes: installing at least one vibration sensor on the machine tool tool holder to collect multi-axial vibration signals; installing an acoustic emission sensor on a workpiece fixture or tool holder near the cutting area to collect acoustic emission signals; and obtaining a spindle load signal from the machine tool control system, wherein the spindle load signal is a current signal.

10. An intelligent detection system for stent turning, characterized in that: include: processor; A memory storing computer instructions for intelligent detection of bracket turning processing, wherein when the computer instructions are executed by the processor, the system executes an intelligent detection method for bracket turning processing according to any one of claims 1 to 9.

Images