Process identification method for multi-variety small-batch machining process

Abstract

A process identification method for a multi-variety and small-batch machining process comprises the steps of performing multiple groups of offline machining experiments, collecting power signals in the machining process, constructing a variable-point template in combination with process parameters, dividing the power signals into process segments by using a variable-point detection method, converting the process segments into a recursion plot by using recursion analysis, and identifying the process segments according to the recursion plot. Constructing a similarity calculation model for the recurrence plot by using a transfer learning method; and dividing a real-time power signal of the machining process into process segments, converting the process segments into a recurrence plot, inputting the recurrence plot into the similarity calculation model to obtain the similarity of each process, constructing a process identification mechanism, and inputting the similarity of each process into the process identification mechanism to determine the type of the process. According to the method and the system, accurate identification of multi-variety and small-batch machining procedures is realized, the convenience and the economical efficiency of identification are improved, automation of procedure division is realized, and the identification precision is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring multi-variety small batch mechanical processing, and in particular to a process identification method for multi-variety small batch mechanical processing. Background Art

[0002] With the increasingly fierce competition among enterprises and the personalized trend of customer orders, the rigid production model based on small varieties and large batches is gradually unable to meet the market demand, and the flexible machining model based on large varieties and small batches is becoming more and more popular in the market.

[0003] For manufacturing companies, once an abnormality occurs in the machining process, it will not only affect the final quality of the product but also lead to the consumption of raw materials and increase the company's production costs. Therefore, it is necessary to conduct real-time status monitoring of the machining process.

[0004] Current state monitoring research mostly takes the workpiece as the smallest research unit, and requires waiting for all processing steps of the workpiece to be completed before monitoring can be carried out. This lag makes it impossible to achieve real-time monitoring of the processing process. Only by further refining state monitoring to the process level can this lag be alleviated.

[0005] Existing research on process mostly focuses on identifying key processes in the processing. There are relatively few literatures that study whether the processing status of the process is abnormal, and most of them rely on installing multiple built-in sensors to collect processing status characteristics. However, these invasive devices are not only difficult to install but also greatly increase the cost. With the continuous development of science and technology, relatively convenient and low-cost non-invasive sensors will gradually become more in line with the needs of corporate benefits. Summary of the Invention

[0006] In order to solve the above-mentioned problems in the prior art, the present invention proposes a process identification method for multi-variety small-batch mechanical processing. The power signal collected by a non-invasive sensor is used as the data source, and the automatic change point detection technology is used to divide the process. The transfer learning method only requires a small amount of data to achieve accurate identification of the processing process.

[0007] The present invention is achieved through the following technical solutions.

[0008] The process identification method of a multi-variety small batch machining process according to the present invention comprises the following steps:

[0009] Step (1): performing multiple sets of machining experiments, collecting and recording power signals during the machining process, wherein the power signal is the input power signal of the machine tool's main power supply;

[0010] Step (2): using the Bayesian online change point detection algorithm to automatically identify the breakpoints of the power signal and obtain the first processing step power segment of the power signal;

[0011] Step (3): constructing a change point template based on the process parameter information on the process card, wherein the change point template includes the relative processing time of each processing step, that is, the time difference between the processing start point and the processing end point;

[0012] Step (4): Add the change point template to the change point detection algorithm and introduce the change point screening mechanism to divide the complete power signal into various process segments. The parameters of the change point detection algorithm include the risk function hazard, the discrete index prior rate λ gap , prior mean μ, prior variance σ, the set values ​​are 0.01, 1, 1, and 2 respectively. The change point screening mechanism refers to removing the change points between processing steps and only retaining the change points between processing procedures;

[0013] Step (5): using the recursive analysis method to convert each process segment into a recursive graph;

[0014] Step (6): Using the transfer learning method, the recursive graph is input into the pre-trained neural network to obtain a similarity calculation model, wherein the similarity calculation model is used to calculate the probability that each process is similar to other processes. If the maximum value of the output similarity is similar, it means that the recognition result is not unique. If the maximum value of the output similarity is less than 0.5, it means that the recognition result does not match. If the maximum value of the output similarity is not similar and is greater than 0.5, it means that the recognition is successful.

[0015] Step (7): Construct a process identification mechanism, wherein the process identification mechanism includes two parts: a non-circular structure and a cyclic structure, specifically including:

[0016] The non-cyclic structure is used to preliminarily identify the type of workpiece. The first processing step power segment of the real-time signal is converted into a recursive graph and input into the similarity calculation model. If the recognition is successful, the template change point of the workpiece type will be directly used in the Bayes change point detection algorithm for division. If the recognition result does not match, it means that a new type of workpiece has appeared. If the recognition result is not unique, the template containing the first processing step segment in the change point template is used to divide the entire real-time power signal into processes in turn. If the next process segment divided by one of the templates is successfully identified, it means that the template is the corresponding template. At this time, other templates can be removed to determine the type of workpiece to which the process segment belongs.

[0017] The loop structure is used to further determine the type of process. After the workpiece type is initially screened in the non-loop structure, if the workpiece type cannot be determined, the template change points that have been initially screened are used to divide the next process segment, and the templates that do not meet the requirements are gradually eliminated until the process segment is successfully identified. The corresponding template and workpiece type to which the process segment belongs are then determined. Then, the corresponding template change points are used in the Bayesian change point detection algorithm to gradually divide the process segments and input them into the similarity calculation model in sequence to determine the process type until the processing is completed.

[0018] Step (8): Input the real-time power signal into the process identification mechanism to determine the process type.

[0019] Furthermore, in step (3), the process parameters include the spindle speed n, feed rate f, cutting depth a of the CNC machine tool p .

[0020] Furthermore, in step (5), the parameters of the recursive analysis, including the embedding dimension m and the time delay τ, are all set to 1.

[0021] Furthermore, in step (6), the neural network used in the transfer learning method is ConvNeXt, and the migration method is to update only the last fully connected layer of the neural network.

[0022] The advantages of the present invention over the prior art are:

[0023] The present invention realizes the accurate identification of multi-variety and small-batch machining processes, and improves the convenience and economy of identification by using the machine tool processing power signal as the data source; the use of the change point detection method to divide the process reduces the interference of subjective factors and realizes the automation of process division; the use of the transfer learning method only requires a small amount of data to achieve extremely high recognition accuracy; and proposes a process recognition mechanism to solve the identification problem of similar or identical processes, further improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The flowchart of the process identification method of the multi-variety small batch machining process of the present invention.

[0025] Figure 2 This is the result of dividing the working steps in the embodiment of the present invention.

[0026] Figure 3 This is the process division result in the embodiment of the present invention.

[0027] Figure 4 This is a framework diagram of the process identification mechanism in the present invention. DETAILED DESCRIPTION

[0028] The present invention is described in detail below with reference to the following embodiments and accompanying drawings. However, it should be understood that the embodiments and accompanying drawings are merely illustrative of the present invention and do not limit the scope of protection of the present invention. All reasonable variations and combinations within the scope of the present invention fall within the scope of protection of the present invention.

[0029] like Figure 1 As shown, the present invention discloses various steps of a method for identifying a process step in a multi-variety small batch machining process. In this embodiment, the above steps are further described in combination with typical shaft workpieces and plate workpieces in machining, including:

[0030] In step S01, 50 sets of machining experiments are performed on three types of shaft workpieces and two types of plate workpieces. During the machining process, a power meter is connected to the main power supply of the machine tool to collect the input power of the machine tool during the machining process.

[0031] In step S02, the Bayesian online change point detection algorithm is used to automatically detect the change points of the power signal and divide it into multiple power segments. The first divided power segment is the power segment of the first processing step.

[0032] The Bayes online change point detection algorithm requires four parameters to be input in advance, namely hazard, λ gap , μ and σ, where hazard represents the risk function, λ gap represents the prior rate of the discrete index, μ represents the prior mean, σ represents the prior variance, the hazard of the present invention when using the Bayes online change point detection algorithm, λ gap , the setting values ​​of the four parameters μ and σ are 0.01, 1, 1 and 2 respectively.

[0033] like Figure 2 As shown in the figure, since there is a power change between the first processing step and the second processing step, and this change can be accurately identified by the Bayes online change point detection algorithm, the first processing step can be automatically and completely divided.

[0034] In step S03, a variation point template is constructed based on the processing parameter information on the process card. This template records the relative processing time of each process and allows for error fluctuations at 20 time points before and after, with each time point separated by 50ms. Taking a shaft workpiece as an example, its processing parameter information is shown in Table 1:

[0035] Table 1 Processing parameter information table

[0036]

[0037] The change point template of the shaft workpiece is shown in Table 2:

[0038] Table 2 Change point template table

[0039]

[0040]

[0041] In step S04, a change point template is added to the Bayes online change point detection algorithm. The template only retains the change points between processes and removes the step change points between processes, thereby completely dividing the power segments of each process.

[0042] like Figure 3 As shown in the figure, all the processes are completely divided out, and the variation points between the steps are successfully removed.

[0043] Step S05: using a recursive analysis method to convert the power segments of each process into a recursive graph.

[0044] The recursive graph uses phase space reconstruction technology to increase the dimension of the signal, making the nonlinear dynamic system visual. Its key parameters include the embedding dimension m and the time delay τ, both of which are set to 1.

[0045] The recurrence graph includes threshold recurrence graph and no-threshold recurrence graph. To avoid subjective factors, the no-threshold recurrence graph is used to calculate the Euclidean distance between two points in the phase space. The closer the Euclidean distance, the darker the color in the recurrence graph.

[0046] Step S06: Use transfer learning to build a similarity calculation model. Pre-train the ConvNeXt network using the ImageNet 1K dataset, which contains 1.3 million training images and 50,000 validation images, all with a pixel size of 224 x 224. Save the neuron weights for each layer of the pre-trained model.

[0047] The data preprocessing of the recursive graph of each process segment is carried out, and the specific process includes:

[0048] Random cropping and scaling: By randomly cropping a region of the recurrence graph and scaling it to 224*224, we introduce local variations and enrich the model's input diversity.

[0049] Random horizontal flip: Randomly flip the recurrence graph horizontally to simulate the symmetry changes in the dataset and improve the robustness of the model.

[0050] Image tensorization: Convert the recurrent graph image to a PyTorch tensor and normalize the pixel values ​​from the range [0, 255] to the range [0, 1] to meet the input requirements of the model.

[0051] Standardization: The recurrence graph is normalized using the mean and standard deviation of the dataset to align the input data with the training distribution of the pre-trained model, thereby accelerating the convergence of the model.

[0052] Migrate the model, load the pre-trained model weights, and iterate over the keys of the weight dictionary, deleting the weights corresponding to the keys containing "head".

[0053] Input recursion graphs for model training. In this example, there are 5 types of workpieces, 46 types of process fragments and the first processing step fragment, and 2,300 recursion graphs. The last 20% of the samples are used as the test set, and the first 80% of the samples are divided into training set and validation set in a 4:1 ratio. The 1,472 recursion graphs in the training set are input into the model for training.

[0054] Only the fully connected layer parameters are updated to obtain the similarity calculation model.

[0055] In this example, the test set is used as the real-time power signal to verify the model effect.

[0056] Step S07: constructing a process identification mechanism.

[0057] like Figure 4 As shown in Figure 2, the process identification mechanism is mainly divided into two parts: non-cyclic structure and cyclic structure. The specific process includes:

[0058] Automatic work step segmentation: The real-time power signal is automatically segmented using Bayesian online change point detection to obtain the power signal of the first processing step segment;

[0059] Recursion graph conversion: convert the power signal of the first processing step segment into a recursion graph;

[0060] Similarity calculation: The power signal recursive graph of the first processing step segment is input into the similarity calculation model to calculate the similarity;

[0061] Non-cyclic structure: After calculating the similarity, if no two or more segments have close similarity, but the maximum similarity is greater than 0.5, it means that the segment recognition results match, and the subsequent segmentation will be directly performed using the template change point of the workpiece type.

[0062] If there are no two or more segments with close similarity, but the maximum similarity is less than 0.5, it means that the segment recognition result does not match, that is, there is no matching image in the trained images, which indicates that a new type of artifact has appeared.

[0063] If there are two or more segments with close similarity, it means that the segment recognition result is not unique, that is, the first step segment of different workpieces is the same as the predicted first step segment. In this case, the template containing the first processing step segment will be used to divide the entire real-time power signal into process segments. If the next process segment divided by a certain template is successfully recognized, it means that the template is the corresponding template. Remove other templates and determine the type of workpiece to which the step segment belongs.

[0064] For example, after the first processing step fragment x is input into the transfer learning model to calculate the similarity, it is found that the similarity between the fragment x and the first processing step fragments of workpiece 1 and workpiece 2 is 0.5, and the similarity between the fragment x and the first processing step fragments of other workpieces such as workpiece 3 and workpiece 4 is 0. This means that the fragment x is the first processing step fragment of workpiece 1 or workpiece 2. At this time, the two template change points of workpiece 1 and workpiece 2 are used to divide the original real-time signal into processes respectively. If the next process divided by the template change point of workpiece 1 does not match, but the next process divided by the template change point of workpiece 2 matches, it means that the type of the real-time signal is workpiece 2, and the template change point of workpiece 2 is used for process division in the future;

[0065] Loop structure: The workpiece type screening method based on the first processing step can quickly help lock the workpiece type. However, there is still a possibility that after locking the possibility of two workpieces, the next or even the next process segment divided by the change points of the two workpiece templates will still be the same. In this case, a loop structure can be used to solve this problem.

[0066] Take the first process segment as the starting point. After the workpiece type is initially screened based on the first processing step segment, if the workpiece type cannot be determined, continue to use the template change points that have been initially screened to divide the next process segment until the process segment is successfully identified. Remove other templates, determine the type of the process segment, and continue to input the next process segment until the processing is completed.

[0067] After determining the workpiece type, it is possible that a certain process in the middle is the same as the process of other workpieces, or the same as a subsequent process of the workpiece itself. For the same process, the transfer learning model cannot be used to calculate the similarity for identification, so the previous process segment is used to assist in identification.

[0068] Since each matching and unique process is recorded, the same process can be dynamically filtered using prior knowledge, that is, the previous process fragment is used to assist in identification.

[0069] For example, processes 1, 2, 4, and 5 of a workpiece are all matched and unique, but processes 3 and 6 are similar to processes of another workpiece. In this case, they cannot be distinguished by calculating similarity. In this case, they are distinguished by the previous process segment. If the previous process is process 2, then this process is process 3 and not other processes.

[0070] Step S08: Use the test set to verify the model effect. The final process recognition results are as follows:

[0071] Workpiece 1, all processes were identified successfully;

[0072] For workpiece 2, the end-facing process recognition results of the 41st and 42nd samples were unique and were directly misidentified as the end-facing process of workpiece 1. The remaining processes were all successfully recognized.

[0073] For workpiece 3, the first idle segments of the 41st, 44th, and 47th samples were uniquely identified and directly misidentified as the first processing step of workpiece 2. The second idle segment of the 47th sample was uniquely identified and directly misidentified as the first processing step of workpiece 2. The remaining processes were successfully identified.

[0074] For workpiece 4, the second unloaded segments of the 47th and 50th samples were uniquely identified and were directly misidentified as the fourth unloaded segments of workpiece 1. The remaining processes were all successfully identified.

[0075] For workpiece 5, all processes were identified successfully.

[0076] The final process recognition results showed that out of a total of 460 test set samples, only 8 samples were misidentified, and the remaining 452 samples were correctly identified, with an accuracy rate of 98.3%.

[0077] Through the method described in this embodiment, the process identification of multi-variety small-batch mechanical processing is realized, and the real-time monitoring of the production process is improved. The present invention uses the total power input power of the machine tool as the data source. This acquisition method is not only convenient to install but also low-cost. At the same time, the present invention uses change point detection technology to divide the process in real time, reducing the interference of subjective factors on the division results. In order to solve the problem of insufficient data volume of multi-variety small-batch mechanical processing processes, the present invention uses transfer learning methods to obtain a high-precision similarity calculation model with only a small amount of experiments. Finally, the present invention proposes a process identification mechanism for screening similar process fragments in the processing process, which further improves the accuracy of process identification.

[0078] The above applications are only some embodiments of the present application. For those skilled in the art, without departing from the inventive concept of the present application, several modifications and improvements can be made, which all fall within the scope of protection of the present application.

Claims

1. A process identification method for a multi-variety small batch machining process, characterized by The steps include: Step (1): Conduct multiple sets of machining experiments and collect and record the input power signal of the machine tool's total power supply during the machining process; Step (2): using the Bayesian online change point detection algorithm to automatically identify the breakpoints of the power signal and obtain the first processing step power segment of the power signal; Step (3): constructing a change point template based on the process parameter information on the process card, wherein the change point template includes the relative processing time of each processing step; Step (4): Add the change point template to the change point detection algorithm and introduce the change point screening mechanism to divide the complete power signal into various process segments. The parameters of the change point detection algorithm include the risk function hazard, the discrete index prior rate λ gap , prior mean μ, prior variance σ, the set values ​​are 0.01, 1, 1, and 2 respectively. The change point screening mechanism refers to removing the change points between processing steps and only retaining the change points between processing procedures; Step (5): using the recursive analysis method to convert each process segment into a recursive graph; Step (6): Using the transfer learning method, the recursive graph is input into the pre-trained neural network to obtain a similarity calculation model, wherein the similarity calculation model is used to calculate the probability that each process is similar to other processes. If the maximum value of the output similarity is similar, it means that the recognition result is not unique. If the maximum value of the output similarity is less than 0.5, it means that the recognition result does not match. If the maximum value of the output similarity is not similar and is greater than 0.5, it means that the recognition is successful. Step (7): Construct a process identification mechanism, wherein the process identification mechanism includes two parts: a non-cyclic structure and a cyclic structure, specifically including: The non-cyclic structure is used to preliminarily identify the type of workpiece. The first processing step power segment of the real-time signal is converted into a recursive graph and input into the similarity calculation model. If the recognition is successful, the template change point of the workpiece type will be directly used in the Bayes change point detection algorithm for division. If the recognition result does not match, it means that a new type of workpiece has appeared. If the recognition result is not unique, the template containing the first processing step segment in the change point template is used to divide the entire real-time power signal into processes in turn. If the next process segment divided by one of the templates is successfully identified, it means that the template is the corresponding template. At this time, other templates can be removed to determine the type of workpiece to which the process segment belongs. The loop structure is used to further determine the type of process. After the workpiece type is initially screened in the non-loop structure, if the workpiece type cannot be determined, the template change points that have been initially screened are used to divide the next process segment, and the templates that do not meet the requirements are gradually eliminated until the process segment is successfully identified. The corresponding template and workpiece type to which the process segment belongs are then determined. Then, the corresponding template change points are used in the Bayesian change point detection algorithm to gradually divide the process segments and input them into the similarity calculation model in sequence to determine the process type until the processing is completed. Step (8): Input the real-time power signal into the process identification mechanism to determine the process type.

2. The process identification method for a multi-variety small batch machining process according to claim 1 is characterized in that In step (3), the process parameters include the spindle speed n, feed rate f, cutting depth a of the CNC machine tool p .

3. The process identification method for a multi-variety small batch machining process according to claim 1 is characterized in that In step (5), the parameters of the recursive analysis include the embedding dimension m and the time delay τ, both of which are set to 1.

4. The process identification method for a multi-variety small batch machining process according to claim 1 is characterized in that In step (6), the neural network used in the transfer learning method is ConvNeXt, and the migration method is to update only the last fully connected layer of the neural network.

Images