CN117495211A - Industrial master machining workpiece quality prediction method based on self-adaptive period discovery - Google Patents
Industrial master machining workpiece quality prediction method based on self-adaptive period discovery Download PDFInfo
- Publication number
- CN117495211A CN117495211A CN202410004935.0A CN202410004935A CN117495211A CN 117495211 A CN117495211 A CN 117495211A CN 202410004935 A CN202410004935 A CN 202410004935A CN 117495211 A CN117495211 A CN 117495211A
- Authority
- CN
- China
- Prior art keywords
- workpiece
- processing data
- time sequence
- data
- dimensional
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 61
- 238000003754 machining Methods 0.000 title claims abstract description 38
- 238000012545 processing Methods 0.000 claims abstract description 169
- 230000003746 surface roughness Effects 0.000 claims abstract description 81
- 238000012549 training Methods 0.000 claims abstract description 23
- 230000015654 memory Effects 0.000 claims abstract description 21
- 238000004140 cleaning Methods 0.000 claims abstract description 20
- 230000008569 process Effects 0.000 claims abstract description 14
- 238000012795 verification Methods 0.000 claims abstract description 7
- 238000012360 testing method Methods 0.000 claims abstract description 6
- 230000001133 acceleration Effects 0.000 claims description 33
- 230000008859 change Effects 0.000 claims description 32
- 230000002776 aggregation Effects 0.000 claims description 17
- 238000004220 aggregation Methods 0.000 claims description 17
- 238000006243 chemical reaction Methods 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 10
- 230000007246 mechanism Effects 0.000 claims description 10
- 238000005520 cutting process Methods 0.000 claims description 9
- 230000003044 adaptive effect Effects 0.000 claims description 8
- 238000012935 Averaging Methods 0.000 claims description 6
- 238000000227 grinding Methods 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 4
- 239000000463 material Substances 0.000 claims description 4
- 238000004439 roughness measurement Methods 0.000 claims description 4
- 230000004931 aggregating effect Effects 0.000 claims description 3
- 230000010354 integration Effects 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 239000013589 supplement Substances 0.000 claims description 3
- 230000000737 periodic effect Effects 0.000 abstract description 8
- 230000002159 abnormal effect Effects 0.000 abstract description 3
- 230000009466 transformation Effects 0.000 abstract description 2
- 230000000875 corresponding effect Effects 0.000 description 18
- 238000013480 data collection Methods 0.000 description 5
- 238000003801 milling Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000007774 longterm Effects 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 239000000047 product Substances 0.000 description 3
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 2
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000005299 abrasion Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000010921 in-depth analysis Methods 0.000 description 1
- 238000012994 industrial processing Methods 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06395—Quality analysis or management
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/04—Manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Strategic Management (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Economics (AREA)
- Marketing (AREA)
- Educational Administration (AREA)
- General Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Health & Medical Sciences (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Entrepreneurship & Innovation (AREA)
- Biophysics (AREA)
- Game Theory and Decision Science (AREA)
- Quality & Reliability (AREA)
- Manufacturing & Machinery (AREA)
- Primary Health Care (AREA)
- Biomedical Technology (AREA)
- Operations Research (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Numerical Control (AREA)
Abstract
The invention discloses a quality prediction method of an industrial master machining workpiece based on self-adaptive period discovery, and relates to the fields of time sequence prediction and industrial master machining data. Acquiring a sample data set; performing data cleaning on the sample data set; performing SMOTE data enhancement; randomly dividing a sample data set into a training set, a verification set and a test set according to a set proportion; constructing a surface roughness prediction model of the machined workpiece; training a surface roughness prediction model of the machined workpiece by using a training set; and obtaining a predicted value of the surface roughness of the machined workpiece by using the trained surface roughness prediction model of the machined workpiece. The invention solves the problems of less abnormal data, difficult learning of periodic time characteristics in the processing process and the like; the periodic information of the self-adaptive learning data is subjected to fast Fourier transformation, and the dependency relationship between the learning data of the multi-scale convolution module and the long-short-time memory module is used for obtaining a more accurate predicted value of the surface roughness of the processed workpiece.
Description
Technical Field
The invention relates to the field of time sequence prediction and industrial master machining data, in particular to a method for predicting quality of an industrial master machining workpiece based on self-adaptive period discovery.
Background
In the industrial mother machining process, the cutter is subjected to higher milling force and milling temperature due to impact in the process and friction between the cutter, chips and the surface of a workpiece, so that the condition that the cutter is worn is unavoidable. The cutter can be worn or damaged continuously in the machining process, and the tiny state change of each cutting edge in the cutting process can influence the quality of products. If the cutter reaches the grinding standard, the cutter cannot be replaced in time, so that the surface roughness of the workpiece is increased, the size precision is reduced, the milling force is increased, the milling temperature is increased, even the machine tool is in fault, the workpiece cannot be continuously and normally processed, and the product is scrapped in batches, so that the manpower, material resources and time are greatly consumed. It can be seen that the replacement of the tool at a proper time can prevent product defects caused by abrasion and breakage. Therefore, it is important to predict the current surface roughness of the tool from each piece of processing data at the time of processing.
In recent years, with the increasing level of industrial fields, the level of management of industrial equipment is also increasing. Existing methods for predicting the surface roughness of a machined workpiece implement predictive maintenance by periodically collecting, recording, analyzing sensor data, setting "alarm load values" based on machining experience, or building classification models, regression models, etc. based on machine learning.
However, the existing method for predicting the surface roughness of a machined workpiece has the following problems:
1) Because of the complexity of the process, the data cleaning and data marking are required to be performed manually, and great cost is consumed, so that the situation of insufficient data with labels often occurs.
2) The process of processing the workpiece has a certain periodicity, however, the existing method only depends on the association learning between discrete time points, so that correct periodic information is difficult to learn, and the processing process cannot be well modeled, so that reliable time sequence dependence is dug out, the model is misjudged, and an incorrect prediction result is generated.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides the industrial master machining workpiece quality prediction method based on self-adaptive period discovery, solves the problem of high acquisition cost of a data set through data enhancement, learns the periodic association of machining data through frequency domain transformation, and improves the prediction precision of the surface roughness of a machined workpiece.
The industrial master machining workpiece quality prediction method based on the self-adaptive period discovery comprises the following steps:
step 1: installing a power sensor and a vibration sensor on a numerical control machine tool, collecting cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration according to set frequencies in the cutting process to obtain a plurality of workpiece processing data samples, and performing manual roughness measurement on a set number of workpieces at intervals to serve as labels of the workpiece processing data samples, so as to obtain a sample data set consisting of the workpiece processing data samples; each workpiece processing data sample is a time sequence and comprises processing data acquired in a period of time for processing one workpiece and corresponding surface roughness of the processed workpiece, wherein the processing data comprises cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration; the x direction is the horizontal direction of the machine tool, namely the direction of a workpiece during transverse processing on the machine tool; the y direction is the vertical direction of the machine tool, namely the up-and-down movement direction of the machine tool; the z-direction is the feed direction of the material of the work piece being processed.
Step 2: and performing data cleaning on the workpiece processing data samples in the sample data set.
The specific method for cleaning the data comprises the following steps: when the missing value exists in the data in the workpiece processing data sample, the missing value is complemented by using the average value of the data in the time span set by the missing value; and deleting the workpiece processing data sample when the surface roughness of the processed workpiece in the workpiece processing data sample is not in the set range.
Step 3: and performing SMOTE data enhancement on the sample data set after data cleaning, and performing characteristic expansion according to a cutter grinding mechanism to obtain a processed sample data set.
Step 3.1: and for each workpiece processing data sample with the surface roughness of the processed workpiece being greater than a set threshold value in the sample data set after data cleaning, finding out a workpiece processing data sample closest to the surface roughness of the processed workpiece.
Step 3.2: randomly selecting one workpiece processing data sample from the found a workpiece processing data samples, and calculating the position difference of the workpiece processing data sample with the surface roughness of the processed workpiece being larger than a set threshold value and the randomly selected workpiece processing data sample in a feature space.
The calculation method of the position difference in the feature space comprises the following steps:
（1）
（2）
Wherein X is a workpiece processing data sample of which the surface roughness of the processed workpiece is greater than a set threshold value; y is a randomly selected workpiece processing data sample in the a workpiece processing data samples;sample data X for workpiece processing at +.>Values in the individual dimensions; />Sample Y of workpiece processing data is at +.>Values in the individual dimensions; />The number of dimensions for the workpiece processing data samples, X, Y, all contain n dimensions.
Step 3.3: for the difference in position in the feature space obtained from each set of workpiece processing data samples, randomly multiply by one [0, 1]Number of the twoAnd adding the result into the corresponding workpiece processing data sample with the surface roughness of the processed workpiece being greater than the set threshold value, thereby generating a new workpiece processing data sample +.>；
（3）
Wherein,processing data samples for a new workpiece; />Is a random number between (0, 1).
Step 3.4: judging whether the number of the generated new workpiece processing data samples reaches a set threshold value, if so, executing the step 3.5; if not, returning to the step 3.1.
Step 3.5: separately computing each workpiece processing data samplePower to vibration ratios in three directions and adding the power to vibration ratios in three directions to the workpiece processing data sample.
The power vibration ratio calculation method in the three directions comprises the following steps:
（4）
step 3.6: and screening workpiece processing data samples with the surface roughness of the processed workpiece being smaller than a set threshold value in the sample data set, and calculating the average value of processing data of each time point in the workpiece processing data samples as a normal processing data trend.
Step 3.7: calculating the difference value between the processing data of each workpiece processing data sample in the sample data set and the normal processing data trend to obtain a distance flat value, and adding the distance flat value to the corresponding workpiece processing data sample to obtain a processed sample data set.
Step 4: and randomly dividing the processed sample data set into a training set, a verification set and a test set according to a set proportion.
Step 5: and constructing a surface roughness prediction model of the machined workpiece.
The surface roughness prediction model of the machined workpiece comprises a fast Fourier transform module, a two-dimensional tensor conversion module, a multi-scale convolution module, an information aggregation module, a long-short-time memory module and a fully-connected network.
The fast Fourier transform module is used for inputting a one-dimensional time sequence with the length of T and the channel dimension of CPerforming fast Fourier transform in a time dimension to obtain frequency domain signals of different periods, calculating amplitude values of the obtained different periods, averaging to obtain average amplitude of each period, namely strength of each period, taking k periods with the largest average amplitude as the most significant k periods, and sending the most significant k periods to a two-dimensional tensor conversion module, wherein the process is expressed as follows:
（5）
（6）
（7）
Wherein,is a one-dimensional time sequence; FFT (·) represents the fast Fourier transform; amp (·) represents the calculation of the amplitude value; />Representing averaging; />Is->The average amplitude of each cycle, i.e. the intensity of each cycle;representing taking the largest k values; />Frequency corresponding to k cycles representing maximum average amplitude +.>Is->A frequency corresponding to the individual periods; />For a length of k periods with maximum average amplitude, +.>Is->The length of the individual periods; />Is a one-dimensional time seriesLength.
The two-dimensional tensor conversion module is used for inputting a one-dimensional time sequence with the length of T and the channel dimension of CFolding reshape based on each of the k most significant periods, respectively, to obtain k two-dimensional tensors +.>And sent to a multi-scale convolution module.
Further, when the one-dimensional time series cannot be periodic-lengthWhen dividing, 0 is added at the end of the one-dimensional time sequence, so that the length of the one-dimensional time sequence can be increased by the corresponding period length +>And (5) integer division.
The method for folding the reshape comprises the following steps:
（8）
wherein,is based on->Folding the two-dimensional tensor obtained in each period; />Is a folding operation;to supplement 0 at the end of the one-dimensional time sequence, so that the one-dimensional time sequence length can be +. >And (5) integer division.
The multi-scale convolution modulusThe block comprises 4 different paths for receiving k two-dimensional tensors sent by the two-dimensional tensor conversion moduleExtracting 4 two-dimensional time sequence change characteristics with different scales through 4 different paths, connecting the 4 two-dimensional time sequence change characteristics with different scales according to depth, and finally merging the two-dimensional time sequence change characteristics into two-dimensional time sequence change characteristic informationAnd sending the information to an information aggregation module; .
The information aggregation module is used for receiving the two-dimensional time sequence change characteristic information sent by the multi-scale convolution moduleConverting it back into one-dimensional space to obtain the time sequence feature +.>Aggregating the time sequence features to obtain the representation of the time sequence change feature information in one dimension>And sends to the long and short time memory module LSTM.
Representation of the time-series variation characteristic information in one dimensionThe calculation method of (1) is as follows:
（9）
（10）
（11）
wherein,indicating that padding operation is supplemented with 0 for removal; />Representation->Amplitude of the corresponding period; />Representation->Amplitude value after Softmax function; />Representation->Switching to one-dimensional space removes the timing characteristics after padding 0,>representing a two-dimensional time sequence variation characteristic, +.>Representing time sequence change characteristic information in one dimension;the representation is converted into a one-dimensional space; / >Representation->A function.
The long-short time memory module is used for receiving the representation of the time sequence change characteristic information sent by the information aggregation module in one dimensionAnd integrating the output results to obtain output results and sending the output results to the fully-connected network.
The fully-connected network FC is used for receiving the output result sent by the long-short-time memory module, analyzing, capturing and reducing the data dimension, and finally obtaining the predicted value of the surface roughness of the processed workpiece.
Step 6: and training the surface roughness prediction model of the machined workpiece by using the training set, wherein the training is stopped when the number of rounds reaches a limited number or the mean square error loss function setting round of the verification set is not reduced, so that the trained surface roughness prediction model of the machined workpiece is obtained.
Step 7: and (3) acquiring processing data, performing data cleaning and data enhancement on the processing data according to the methods from the step (2) to the step (3), and inputting the processed processing data into a processed workpiece surface roughness prediction model after training to obtain a predicted value of the processed workpiece surface roughness.
Step 7.1: and acquiring machining data according to a set frequency in the cutting process, wherein the machining data comprise cutter machining power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration.
Step 7.2: and (3) performing data cleaning and data enhancement on the processing data according to the methods from the step (2) to the step (3) to obtain the processed processing data.
Step 7.3: the processed processing data is input into a fast Fourier transform module, and the most remarkable k periods are found.
Step 7.4: inputting the processed processing data into a two-dimensional tensor conversion module, and respectively folding the reserve based on each of the most obvious k periods to obtain k two-dimensional tensors。
Further, when the one-dimensional time series cannot be periodic-lengthWhen dividing, 0 is added at the end of the one-dimensional time sequence, so that the length of the one-dimensional time sequence can be increased by the corresponding period length +>And (5) integer division.
Step 7.5: inputting k two-dimensional tensors into a multi-scale convolution module, extracting 4 two-dimensional time sequence change characteristics of different scales through 4 different paths, and connecting and combining the 4 two-dimensional time sequence change characteristics of different scales according to depth to obtain two-dimensional time sequence change characteristic information。
Step 7.6: inputting the two-dimensional time sequence variation characteristic information into an information aggregation module, and converting the information back into a one-dimensional space to obtain time sequence characteristicsThen the time sequence characteristics are aggregated to obtain the representation of the time sequence variation characteristic information in one dimension +. >。
Step 7.7: representation of time sequence variation characteristic information in one dimensionAnd after the long-short time memory module is input for integration, an output result is obtained.
Step 7.8: and inputting the output result into a fully-connected network for analysis, capturing and data dimension reduction, and finally obtaining the predicted value of the surface roughness of the processed workpiece.
Compared with the prior art, the invention has the beneficial effects that:
1) Aiming at the industrial master machining data set with less data and unbalance, a more effective data enhancement method is provided, and the problems that the data with a label are difficult to acquire, the abnormal data (the surface roughness of a machined workpiece is high) is less and the like are solved.
2) And establishing a deep learning model, and obtaining a more accurate predicted value of the surface roughness of the processed workpiece through the periodic information of the learning data self-adapting by the fast Fourier transform and the dependency relationship between the learning data of the multi-scale convolution module and the long-short-time memory module.
Drawings
FIG. 1 is a flow chart of a method for predicting quality of an industrial master machined workpiece based on adaptive period discovery in accordance with an embodiment of the present invention;
FIG. 2 is a flow chart of a process for constructing a model for predicting the surface roughness of a workpiece in accordance with an embodiment of the present invention;
FIG. 3 is a schematic view of a surface roughness prediction model of a workpiece under processing in an embodiment;
FIG. 4 is a schematic diagram of a periodic conversion of an original time series into a two-dimensional tensor in accordance with an embodiment of the present invention;
fig. 5 is a schematic diagram of a multi-scale convolution acceptance module in an embodiment of the present invention.
Fig. 6 is a schematic diagram of a long and short term memory LSTM module in an embodiment of the present invention.
Detailed Description
In order to show effectiveness and generality, the invention is illustrated by taking data acquired in the machining process of a cutter of a numerical control machine tool as an example. The method comprises the steps that a sensor is arranged on a numerical control machine, raw data are collected from three-way acceleration and vibration frequency of a cutter in the machining process of the numerical control machine, and after machining of a workpiece is completed each time, roughness of the machined workpiece is measured. And carrying out data enhancement and mechanism characteristic expansion by using the processing data obtained from the sensor, constructing a deep learning model to analyze the relationship between the period and the period through the periodicity of a self-adaptive learning time sequence of the fast Fourier transform, and carrying out good modeling on the dependency relationship between the processing data and the surface roughness of the processed workpiece so as to obtain a more accurate surface roughness prediction method of the processed workpiece.
The industrial master machined workpiece quality prediction method based on the adaptive period discovery, as shown in fig. 1, comprises the following specific steps.
Step 1: installing a power sensor and a vibration sensor on a numerical control machine tool, collecting cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration according to set frequencies in the cutting process to obtain a plurality of workpiece processing data samples, and performing manual roughness measurement on a set number of workpieces at intervals to serve as labels of the workpiece processing data samples, so as to obtain a sample data set consisting of the workpiece processing data samples; each workpiece processing data sample is a time sequence and comprises processing data acquired in a period of time for processing one workpiece and corresponding surface roughness of the processed workpiece, wherein the processing data comprises cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration; the x direction is the horizontal direction of the machine tool, namely the direction of a workpiece during transverse processing on the machine tool; the y direction is the vertical direction of the machine tool, namely the up-and-down movement direction of the machine tool; the z-direction is the feed direction of the material of the work piece being processed.
In this embodiment, 81 pieces of workpiece processing data samples are collected during the cutting process from 2022.5.1 to 2023.4.31, and the collection frequency of each piece of workpiece processing data sample is 10Hz. And during acquisition, five workpieces are subjected to manual roughness measurement at intervals, and the data are stored by using a basic csv file. The workpiece processing data samples comprise cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration, z-direction vibration acceleration and processed workpiece surface roughness after the workpiece is processed. Because of the higher cost of tagged data collection, less usable data is collected. A partial data sample is shown in the following table:
| SN | power of | Vibration in X direction | Vibration in Y direction | Vibration in Z direction | RaCylinder |
| S220628340459 | 196 | 0.154205 | 0.392639 | 0.416969 | 1.143 |
| S220628340459 | 201 | 0.178035 | 0.415315 | 0.388742 | 1.143 |
| S220628340459 | 200 | 0.160285 | 0.378492 | 0.38525 | 1.143 |
| S220628340459 | 198 | 0.158359 | 0.550503 | 0.350014 | 1.143 |
| S220628340459 | 195 | 0.153522 | 0.530785 | 0.343116 | 1.143 |
| S220628340459 | 199 | 0.142647 | 0.379423 | 0.332812 | 1.143 |
| S220628340459 | 208 | 0.111487 | 0.337335 | 0.359088 | 1.143 |
Step 2: and performing data cleaning on the workpiece processing data samples in the sample data set.
The specific method for cleaning the data comprises the following steps: when the missing value exists in the data in the workpiece processing data sample, the missing value is complemented by using the average value of the data in the time span set by the missing value; and deleting the workpiece processing data sample when the surface roughness of the processed workpiece in the workpiece processing data sample is not in the set range.
In the present embodiment, the time span is set to 10, that is, 10 data are taken before and after the missing value to calculate the average value; and deleting the error samples with the surface roughness of the processed workpiece less than 0.5 and greater than 2.5.
Step 3: and performing SMOTE data enhancement on the sample data set after data cleaning, and performing characteristic expansion according to a cutter grinding mechanism to obtain a processed sample data set.
For a sample dataset, it is characterized by:
1) The sample dataset has few features: because the data collection needs to be carried out under the condition of not influencing normal processing, the normal production plan of an enterprise can be successfully completed, the positions where the sensors can be added for data collection are fewer, and the vibration sensors and the power sensors which are simpler in construction and simpler in operation are preferably selected for data collection.
2) The data containing the tag is small: because the roughness of the machined workpiece is required to be detected artificially after the workpiece is machined, the normal machining progress is inevitably affected, and the labor cost and the time consumption of measurement are increased. Therefore, the amount of tagged data that can be obtained tends to be small.
3) The data correlation is strong: because various indexes contained in the processed workpiece roughness data set are often associated with each other in the workpiece processing process, deep analysis of correlation among the indexes is needed, and effective mechanism characteristics are dug out according to experience provided by professionals so as to find possible dependency relationship among the data and improve the accuracy of the surface roughness predicted value of the processed workpiece.
4) There are inter-period dependencies and intra-period dependencies between data: the time sequence data in the industrial processing process is often mutually overlapped in different period processes, and the time sequence data in the period are displayed. Specifically, for a particular cycle, the data at each time point is highly correlated with both the time of day and the same phase of the other cycles. Wherein the periodic variation corresponds to a short term process and the periodic variation may reflect a long term trend over successive weeks. The in-depth analysis of the cycle dependence and the in-cycle dependence plays a vital role in improving the accuracy of the surface roughness prediction of the machined workpiece.
5) Imbalance of positive and negative samples: the processing efficiency of the workpiece is reduced and even the whole production line is scrapped due to the fact that the surface roughness of the processed workpiece is too high. Therefore, it is difficult to collect abnormal data having excessively high surface roughness at the time of data collection. Data enhancement is needed for data with partial surface roughness greater than 1.6 to improve the learning effect of the model.
By reasonably applying the characteristics, the method adopts SMOTE for data enhancement and adds features according to a cutter grinding mechanism, and comprises the following specific steps:
Step 3.1: and for each workpiece processing data sample with the surface roughness of the processed workpiece being greater than a set threshold value in the sample data set after data cleaning, finding out a workpiece processing data sample closest to the surface roughness of the processed workpiece.
In this embodiment, for a sample having a surface roughness greater than 1.6, a nearest neighbor samples are selected near the roughness value thereof. a is a preset super parameter for controlling the number of synthesized samples.
Step 3.2: randomly selecting one workpiece processing data sample from the found a workpiece processing data samples, and calculating the position difference of the workpiece processing data sample with the surface roughness of the processed workpiece being larger than a set threshold value and the randomly selected workpiece processing data sample in a feature space.
The calculation method of the position difference in the feature space comprises the following steps:
（1）
（2）
wherein X is a workpiece processing data sample of which the surface roughness of the processed workpiece is greater than a set threshold value; y is a randomly selected workpiece processing data sample in the a workpiece processing data samples;sample data X for workpiece processing at +.>Values in the individual dimensions; />Sample Y of workpiece processing data is at +.>Values in the individual dimensions; />The number of dimensions for the workpiece processing data samples, X, Y, all include n dimensions Degree.
In this embodiment, for each sample having a surface roughness of greater than 1.6, a sample is randomly selected among its a nearest neighbors to calculate the difference in position between the two in the feature space.
Step 3.3: for the difference in position in the feature space obtained from each set of workpiece processing data samples, randomly multiply by one [0, 1]Number of the twoAnd adding the result into the corresponding workpiece processing data sample with the surface roughness of the processed workpiece being greater than the set threshold value, thereby generating a new workpiece processing data sample +.>；
（3）
Wherein,processing data samples for a new workpiece; />Is a random number between (0, 1).
Step 3.4: judging whether the number of the generated new workpiece processing data samples reaches a set threshold value, if so, executing the step 3.5; if not, returning to the step 3.1.
Step 3.5: separately computing each workpiece processing data samplePower to vibration ratios in three directions and adding the power to vibration ratios in three directions to the workpiece processing data sample.
The new mechanism characteristic, namely the power vibration ratio, is added in the embodiment: the relative mechanism can obtain that the ratio of the power to the three-way acceleration of vibration has a larger correlation with the surface roughness of the processed workpiece, and the value is calculated and added into a sample data set;
（4）
Step 3.6: and screening workpiece processing data samples with the surface roughness of the processed workpiece being smaller than a set threshold value in the sample data set, and calculating the average value of processing data of each time point in the workpiece processing data samples as a normal processing data trend.
New mechanism features-normal processing data trend-are added in the embodiment: and calculating the average value of 254 corresponding time points according to the collected normal machining data (machining data with the roughness of the machined workpiece being less than 1.6), and obtaining the trend of the normal machining data.
Step 3.7: calculating the difference value between the processing data of each workpiece processing data sample in the sample data set and the normal processing data trend to obtain a distance flat value, and adding the distance flat value to the corresponding workpiece processing data sample to obtain a processed sample data set.
Step 4: and randomly dividing the processed sample data set into a training set, a verification set and a test set according to a set proportion.
Step 5: and constructing a surface roughness prediction model of the machined workpiece.
Based on the characteristics of the surface roughness data set of the industrial master machining workpiece, the traditional method is difficult to obtain a large amount of effective data in the machining process with low cost and high efficiency, and is difficult to construct a relatively accurate machining model from the existing data set, so that the surface roughness of the machining workpiece is difficult to accurately predict, the tool changing time is difficult to accurately and timely early warn, and the production plan of an enterprise is delayed.
According to the invention, an industrial master machining workpiece quality prediction method based on self-adaptive period discovery is adopted, data is converted into a two-dimensional tensor through the periodicity of a fast Fourier transform self-adaptive learning time sequence, a multi-scale convolution module is constructed, the dependence relationship between the period and the period is learned, and finally, a long-time memory module is constructed to obtain a predicted value of the surface roughness of a machined workpiece. Meanwhile, the time sequence relation change between the period and the period is considered, so that the accuracy of prediction is improved. The construction flow is as shown in FIG. 2:
the method provided by the application can be used for predicting the surface roughness of the machined workpiece, can also be used for predicting other data characteristics in the machining process of the industrial master machine, and has good universality. The model structure is shown in fig. 3, and the model is constructed as follows.
The surface roughness prediction model of the machined workpiece comprises a fast Fourier transform module, a two-dimensional tensor conversion module, a multi-scale convolution module, an information aggregation module, a long-short-time memory module and a fully-connected network.
To uniformly represent the time sequence relation change between the period and the period, the periodicity of the time sequence is first adaptively learned.
The fast Fourier transform module is used for inputting a one-dimensional time sequence with the length of T and the channel dimension of C Performing fast Fourier transform in a time dimension to obtain frequency domain signals of different periods, calculating amplitude values of the obtained different periods, averaging to obtain average amplitude of each period, namely strength of each period, taking k periods with the largest average amplitude as the k most significant periods, and sending the k most significant periods to a two-dimensional tensor conversion module.
For a one-dimensional time sequence of length T and channel dimension CThe periodicity can be obtained by calculating the fast fourier transform in the time dimension as follows:
（5）
（6）
（7）
wherein,is a one-dimensional time sequence; FFT (·) represents the fast Fourier transform; amp (·) represents the calculation of the amplitude value; />Representing averaging; />Is->The average amplitude of each cycle, i.e. the intensity of each cycle;representing taking the largest k values; />Frequency corresponding to k cycles representing maximum average amplitude +.>Is->The frequency corresponding to the period is only considered +.>Frequencies within the range; />For a length of k periods with maximum average amplitude, +.>Is->The length of the individual periods; />Is the length of the one-dimensional time series.
In this embodiment, considering sparsity of the frequency domain, in order to avoid noise caused by meaningless frequency domain data, only k maximum amplitude values are selected, and the most significant k periods are obtained, where k is a super parameter.
The two-dimensional tensor conversion module is used for inputting a one-dimensional time sequence with the length of T and the channel dimension of CFolding reshape based on each of the k most significant periods, respectively, to obtain k two-dimensional tensors +.>And sent to a multi-scale convolution module.
Further, when the one-dimensional time series cannot be periodic-lengthWhen dividing, 0 is added at the end of the one-dimensional time sequence, so that the length of the one-dimensional time sequence can be increased by the corresponding period length +>And (5) integer division.
As shown in fig. 4, the method for folding the one-dimensional time sequence based on the selected period is as follows:
（8）
wherein,is based on->Folding the two-dimensional tensor obtained in each period; />Is folded intoStacking operation;to supplement 0 at the end of the one-dimensional time sequence, so that the one-dimensional time sequence length can be +.>And (5) integer division.
Each column and each row of data corresponds to an adjacent time and an adjacent period, respectively, and the adjacent time and period often contain similar time law and time dependency. The data passing through the modules is easy to capture information by two-dimensional convolution.
And then constructing a multi-scale convolution module, and learning the dependent relationship between the period and the period.
For two-dimensional tensors Because of its two-dimensional locality, we can extract information using 2D convolution.
From research experience and unified understanding, increasing model performance generally requires increasing model size, i.e., increasing width and depth, whereas simply increasing model size presents the following problems: (1) the large number of nodes and parameters results in the occurrence of overfitting; (2) oversized models result in more parameter budgets, resulting in wasted computing resources and extended processing time.
Therefore, the method selects a classical multi-scale convolution module. And providing a plurality of passages, extracting features of different scales at the same layer by using convolution kernels of different sizes, and combining the features in one module to form a multi-channel feature map. The sparse model is realized by the dense operation module through 4 different path processing nodes, the dense connection structure is replaced by the sparse link, the dense calculation advantage of hardware is fully utilized, training parameters are reduced, and the calculation resources and the data processing speed are saved.
As shown in fig. 5, the multi-scale convolution module includes 4 different paths for receiving k two-dimensional tensors sent by the two-dimensional tensor conversion module Extracting 4 two-dimensional time sequence variation characteristics with different scales through 4 different paths, connecting the 4 two-dimensional time sequence variation characteristics with different scales according to depth, for example, connecting 2 x 3 with 2 x 4 two-dimensional time sequence variation characteristics to obtain 2 x 7 output, and finally merging the two-dimensional time sequence variation characteristics into two-dimensional time sequence variation characteristic information%>And sent to the information aggregation module.
Wherein the first path is a convolution layer of 1*1; the second pass is a convolution layer of 1*1 and a convolution layer of 3*3 in order; the third path is a convolution layer of 1*1 and a convolution layer of 5*5 in sequence; the fourth pass is in turn the max-pooling layer of 3*3 and the convolutional layer of 1*1.
Because the deep network needs a small receptive field in the shallow layer to obtain local information, and the deep network needs a large receptive field to obtain long-term information, various needed methods are combined into a module, and the model is enabled to adaptively select the response rule. Introducing a 1*1-sized convolutional layer in a multi-scale convolutional module can reduce the number of parameters so as to reduce a certain amount of calculation.
The information aggregation module is used for receiving the two-dimensional time sequence change characteristic information sent by the multi-scale convolution moduleConverting it back into one-dimensional space for information aggregation to obtain timing characteristics +. >Aggregating the time sequence features to obtain the representation of the time sequence change feature information in one dimension>And sends to the long and short time memory module LSTM.
The extracted timing characteristics are converted into a one-dimensional space. Due to amplitude of vibrationThe relative importance of the corresponding frequency and period can be reflected, so that the relative importance can be used as the weight of the time sequence characteristic to aggregate the information;
（9）
（10）
（11）
wherein,indicating that padding operation is supplemented with 0 for removal; />Representation->Amplitude of the corresponding period; />Representation->The amplitude values after the Softmax function are aggregated as weights. />Representation->Switching to one-dimensional space removes the timing characteristics after padding 0,>representing a two-dimensional time sequence variation characteristic, +.>For time sequence variation characteristic signalsA representation of the information in one dimension; />The representation is converted into a one-dimensional space; />Representation->A function.
Through the above process, we have obtained time information aggregation during the period and in the period, and in order to further aggregate long-sequence time information in the whole processing process, the method constructs a long-time memory module (LSTM), as shown in fig. 6, to capture long-time dependencies in the sequence. The memory unit and the gating mechanism of the long-short time memory module can learn and memorize long-term dependency, solve the problems of gradient disappearance and gradient explosion in the traditional cyclic neural network and difficult capture of long-sequence time information, and are widely used in the problems of time sequence prediction and anomaly detection.
The long-short time memory module is used for receiving the representation of the time sequence change characteristic information sent by the information aggregation module in one dimensionAnd integrating the output results to obtain output results and sending the output results to the fully-connected network. />
Currently, time-dependent information of the processing data has been obtained, however, since there is a certain dependency between features of the processing data, modeling of the dependency between features is required.
The fully connected network (FC) is used for receiving the output result sent by the long-short-time memory module, analyzing, capturing and reducing the data dimension, and finally obtaining the predicted value of the surface roughness of the processed workpiece.
Step 6: and training the surface roughness prediction model of the machined workpiece by using the training set, wherein the training is stopped when the number of rounds reaches a limited number or the mean square error loss function setting round of the verification set is not reduced, so that the trained surface roughness prediction model of the machined workpiece is obtained.
The method is characterized by using the three-way acceleration, the power vibration ratio and the difference value between the normal trend in the industrial mother machining process as characteristics, the surface roughness of the machined workpiece after machining the workpiece is a target value, inputting a training set for training, and using a mean square error loss function to minimize the difference between the predicted value and the real roughness.
Step 7: and (3) acquiring processing data, performing data cleaning and data enhancement on the processing data according to the methods from the step (2) to the step (3), and inputting the processed processing data into a processed workpiece surface roughness prediction model after training to obtain a predicted value of the processed workpiece surface roughness.
Step 7.1: and acquiring machining data according to a set frequency in the cutting process, wherein the machining data comprise cutter machining power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration.
Step 7.2: and (3) performing data cleaning and data enhancement on the processing data according to the methods from the step (2) to the step (3) to obtain the processed processing data.
Step 7.3: the processed processing data is input into a fast Fourier transform module, and the most remarkable k periods are found.
Step 7.4: inputting the processed processing data into a two-dimensional tensor conversion module, and respectively folding the reserve based on each of the most obvious k periods to obtain k two-dimensional tensors。
Further, when the one-dimensional time series cannot be periodic-lengthWhen dividing, 0 is added at the end of the one-dimensional time sequence, so that the length of the one-dimensional time sequence can be increased by the corresponding period length +>And (5) integer division.
Step 7.5: inputting k two-dimensional tensors into a multi-scale convolution moduleThe block extracts 4 two-dimensional time sequence change characteristics with different scales through 4 different paths, and connects and merges the 4 two-dimensional time sequence change characteristics with different scales into two-dimensional time sequence change characteristic information according to depth。
Step 7.6: inputting the two-dimensional time sequence variation characteristic information into an information aggregation module, and converting the information back into a one-dimensional space to obtain time sequence characteristicsThen the time sequence characteristics are aggregated to obtain the representation of the time sequence variation characteristic information in one dimension +.>。
Step 7.7: representation of time sequence variation characteristic information in one dimensionAnd after the long-short time memory module is input for integration, an output result is obtained.
Step 7.8: and inputting the output result into a fully-connected network for analysis, capturing and data dimension reduction, and finally obtaining the predicted value of the surface roughness of the processed workpiece.
After model training is completed, the collected processing data can be subjected to data enhancement and processing workpiece surface roughness prediction, and the method can be used for cutter changing early warning and can improve the processing efficiency of enterprises.
Through multiple experiments, the effect achieved by the method in the test set is compared with various time sequence prediction methods, and the results are shown in the following table:
| The application | GBDT | Xgboost | lightGBM | GRU | LSTM | |
| MAE | 0.093 | 0.174 | 0.158 | 0.166 | 0.194 | 0.373 |
| MAPE | 9.42% | 18.5% | 14.6% | 16.0% | 21.9% | 34.2 % |
The method can achieve better effect than other comparison methods by adopting the actual surface roughness of the processed workpiece and the MAE (average absolute error) and MAPE (average absolute error) of the predicted roughness as model evaluation indexes. Through ten-fold cross validation, the data set was divided into ten parts, 9 of which were alternately used as training data, and 1 as test data.
Claims (7)
1. The industrial master machining workpiece quality prediction method based on the self-adaptive period discovery is characterized by comprising the following steps of:
step 1: installing a power sensor and a vibration sensor on a numerical control machine tool, collecting cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration according to set frequencies in the cutting process to obtain a plurality of workpiece processing data samples, and performing manual roughness measurement on a set number of workpieces at intervals to serve as labels of the workpiece processing data samples, so as to obtain a sample data set consisting of the workpiece processing data samples; each workpiece processing data sample is a time sequence and comprises processing data acquired in a period of time for processing one workpiece and corresponding surface roughness of the processed workpiece, wherein the processing data comprises cutter processing power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration; the x direction is the horizontal direction of the machine tool, namely the direction of a workpiece during transverse processing on the machine tool; the y direction is the vertical direction of the machine tool, namely the up-and-down movement direction of the machine tool; the z direction is the feeding direction of the material of the processed workpiece;
Step 2: performing data cleaning on workpiece processing data samples in the sample data set;
step 3: performing SMOTE data enhancement on the sample data set after data cleaning, and performing characteristic expansion according to a cutter grinding mechanism to obtain a processed sample data set;
step 4: randomly dividing the processed sample data set into a training set, a verification set and a test set according to a set proportion;
step 5: constructing a surface roughness prediction model of the machined workpiece;
step 6: training the surface roughness prediction model of the machined workpiece by using a training set, wherein the training is stopped when the number of rounds reaches a limited number or the mean square error loss function setting round of a verification set is not reduced, so that the trained surface roughness prediction model of the machined workpiece is obtained;
step 7: and (3) acquiring processing data, performing data cleaning and data enhancement on the processing data according to the methods from the step (2) to the step (3), and inputting the processed processing data into a processed workpiece surface roughness prediction model after training to obtain a predicted value of the processed workpiece surface roughness.
2. The method for predicting quality of an industrial master machined workpiece based on adaptive cycle discovery according to claim 1, wherein the specific method for cleaning data in step 2 is as follows: when the missing value exists in the data in the workpiece processing data sample, the missing value is complemented by using the average value of the data in the time span set by the missing value; and deleting the workpiece processing data sample when the surface roughness of the processed workpiece in the workpiece processing data sample is not in the set range.
3. The method for predicting quality of an industrial master machined workpiece based on adaptive period discovery as set forth in claim 1, wherein the step 3 specifically includes:
step 3.1: for each workpiece processing data sample with the surface roughness of the processed workpiece being greater than a set threshold value in the sample data set after data cleaning, finding out a workpiece processing data samples closest to the surface roughness of the processed workpiece;
step 3.2: randomly selecting one workpiece processing data sample from the found a workpiece processing data samples, and calculating the position difference of the workpiece processing data sample with the surface roughness of the processed workpiece being greater than a set threshold value and the randomly selected workpiece processing data sample in a feature space;
step 3.3: for the difference in position in the feature space obtained from each set of workpiece processing data samples, randomly multiply by one [0, 1]Number of the twoAnd adding the result into the corresponding workpiece processing data sample with the surface roughness of the processed workpiece being greater than the set threshold value, thereby generating a new workpiece processing data sample +.>；
（3）
Wherein,processing data samples for a new workpiece; />Is [0, 1]Random numbers in between;
step 3.4: judging whether the number of the generated new workpiece processing data samples reaches a set threshold value, if so, executing the step 3.5; if not, returning to the step 3.1;
Step 3.5: separately computing each workpiece processing data sampleThe power vibration ratios in the three directions are added into the workpiece processing data sample;
the power vibration ratio calculation method in the three directions comprises the following steps:
（4）
step 3.6: screening workpiece processing data samples with the surface roughness of the processed workpiece being smaller than a set threshold value in the sample data set, and calculating the average value of processing data of each time point in the workpiece processing data samples as a normal processing data trend;
step 3.7: calculating the difference value between the processing data of each workpiece processing data sample in the sample data set and the normal processing data trend to obtain a distance flat value, and adding the distance flat value to the corresponding workpiece processing data sample to obtain a processed sample data set.
4. The method for predicting quality of an industrial master machined workpiece based on adaptive period discovery as set forth in claim 3, wherein the method for calculating the position difference in the feature space in step 3.2 is as follows:
（1）
（2）
wherein X is a workpiece processing data sample of which the surface roughness of the processed workpiece is greater than a set threshold value; y is a randomly selected workpiece processing data sample in the a workpiece processing data samples; Sample data X for workpiece processing at +.>Values in the individual dimensions; />Sample Y of workpiece processing data is at +.>Values in the individual dimensions; />The number of dimensions for the workpiece processing data samples, X, Y, all contain n dimensions.
5. The method for predicting the quality of an industrial master machined workpiece based on adaptive period discovery according to claim 1, wherein the machined workpiece surface roughness prediction model in step 5 comprises a fast fourier transform module, a two-dimensional tensor transform module, a multi-scale convolution module, an information aggregation module, a long-short-term memory module and a fully connected network;
the fast Fourier transform module is used for inputting a one-dimensional time sequence with the length of T and the channel dimension of CPerforming fast Fourier transform in a time dimension to obtain frequency domain signals of different periods, calculating amplitude values of the obtained different periods, averaging to obtain average amplitude of each period, namely strength of each period, taking k periods with the largest average amplitude as the most significant k periods, and sending the most significant k periods to a two-dimensional tensor conversion module, wherein the process is expressed as follows:
（5）
（6）
（7）
wherein,is a one-dimensional time sequence; FFT (·) represents the fast Fourier transform; amp (·) represents the calculation of the amplitude value; / >Representing averaging; />Is->The average amplitude of each cycle, i.e. the intensity of each cycle; />Representing taking the largest k values; />Frequency corresponding to k cycles representing maximum average amplitude +.>Is->A frequency corresponding to the individual periods; />For a length of k periods with maximum average amplitude, +.>Is->The length of the individual periods; />Is the length of a one-dimensional time sequence;
the two-dimensional tensor conversion module is used for inputting a one-dimensional time sequence with the length of T and the channel dimension of CFolding reshape based on each of the k most significant periods, respectively, to obtain k two-dimensional tensors +.>And sending the result to a multi-scale convolution module;
the method for folding the reshape comprises the following steps:
（8）
wherein,is based on->Folding the two-dimensional tensor obtained in each period; />Is a folding operation;to supplement 0 at the end of the one-dimensional time sequence, so that the one-dimensional time sequence length can be +.>Removing;
the multi-scale convolution module comprises 4 different paths for receiving k two-dimensional tensors sent by the two-dimensional tensor conversion moduleExtracting 4 two-dimensional time sequence change characteristics with different scales through 4 different paths, connecting the 4 two-dimensional time sequence change characteristics with different scales according to depth, and finally merging the two-dimensional time sequence change characteristics into two-dimensional time sequence change characteristic information +. >And sending the information to an information aggregation module;
the information aggregation module is used for receiving the two-dimensional time sequence change characteristic information sent by the multi-scale convolution moduleConverting it back into one-dimensional space to obtain the time sequence feature +.>Aggregating the time sequence features to obtain the representation of the time sequence change feature information in one dimension>Transmitting to a long and short time memory module LSTM;
the timing sequenceRepresentation of change characteristic information in one dimensionThe calculation method of (1) is as follows:
（9）
（10）
（11）
wherein,indicating that padding operation is supplemented with 0 for removal; />Representation->Amplitude of the corresponding period; />Representation->Amplitude value after Softmax function; />Representation->Switching to one-dimensional space removes the timing characteristics after padding 0,>representing twoDimensional time sequence variation feature->Representing time sequence change characteristic information in one dimension;the representation is converted into a one-dimensional space; />Representation->A function;
the long-short time memory module is used for receiving the representation of the time sequence change characteristic information sent by the information aggregation module in one dimensionIntegrating the output results to obtain output results and sending the output results to a fully-connected network;
the fully-connected network FC is used for receiving the output result sent by the long-short-time memory module, analyzing, capturing and reducing the data dimension, and finally obtaining the predicted value of the surface roughness of the processed workpiece.
6. The method for predicting quality of industrial master machined workpiece based on adaptive period discovery as recited in claim 5, wherein the one-dimensional time series cannot be cycled long when folding reshape is performedWhen dividing, 0 is added at the end of the one-dimensional time sequence, so that the length of the one-dimensional time sequence can be increased by the corresponding period length +>And (5) integer division.
7. The method for predicting quality of an industrial master machined workpiece based on adaptive period discovery as set forth in claim 1, wherein the step 7 specifically includes:
step 7.1: acquiring machining data according to a set frequency in a cutting process, wherein the machining data comprise cutter machining power, x-direction vibration acceleration, y-direction vibration acceleration and z-direction vibration acceleration;
step 7.2: performing data cleaning and data enhancement on the processing data according to the methods from the step 2 to the step 3 to obtain processed processing data;
step 7.3: inputting the processed processing data into a fast Fourier transform module, and finding the most remarkable k periods;
step 7.4: inputting the processed processing data into a two-dimensional tensor conversion module, and respectively folding the reserve based on each of the most obvious k periods to obtain k two-dimensional tensors ；
Further, when the one-dimensional time series cannot be periodic-lengthWhen dividing, 0 is added at the end of the one-dimensional time sequence, so that the length of the one-dimensional time sequence can be increased by the corresponding period length +>Removing;
step 7.5: inputting k two-dimensional tensors into a multi-scale convolution module, extracting 4 two-dimensional time sequence change characteristics of different scales through 4 different paths, and connecting and combining the 4 two-dimensional time sequence change characteristics of different scales according to depth to obtain two-dimensional time sequence change characteristic information；
Step 7.6: inputting the two-dimensional time sequence variation characteristic information into an information aggregation module, and converting the information back into a one-dimensional space to obtain time sequence characteristicsThen the time sequence characteristics are aggregated to obtain the representation of the time sequence variation characteristic information in one dimension +.>；
Step 7.7: representation of time sequence variation characteristic information in one dimensionAfter the long-short time memory module is input for integration, an output result is obtained;
step 7.8: and inputting the output result into a fully-connected network for analysis, capturing and data dimension reduction, and finally obtaining the predicted value of the surface roughness of the processed workpiece.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410004935.0A CN117495211B (en) | 2024-01-03 | 2024-01-03 | Industrial master machining workpiece quality prediction method based on self-adaptive period discovery |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410004935.0A CN117495211B (en) | 2024-01-03 | 2024-01-03 | Industrial master machining workpiece quality prediction method based on self-adaptive period discovery |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117495211A true CN117495211A (en) | 2024-02-02 |
| CN117495211B CN117495211B (en) | 2024-03-19 |
Family
ID=89683442
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410004935.0A Active CN117495211B (en) | 2024-01-03 | 2024-01-03 | Industrial master machining workpiece quality prediction method based on self-adaptive period discovery |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117495211B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3421988A1 (en) * | 2017-06-29 | 2019-01-02 | Domesens srl | Device and a method for the continuous measurement of one or more tree staticity |
| JP2019111969A (en) * | 2017-12-25 | 2019-07-11 | トヨタ自動車株式会社 | Hybrid automobile |
| CN112069966A (en) * | 2020-08-31 | 2020-12-11 | 中国计量大学 | Surface roughness prediction method based on improved LSTM network |
| US10950352B1 (en) * | 2020-07-17 | 2021-03-16 | Prince Mohammad Bin Fahd University | System, computer-readable storage medium and method of deep learning of texture in short time series |
| CN114869272A (en) * | 2021-09-03 | 2022-08-09 | 中国人民解放军总医院 | Posture tremor detection model, posture tremor detection algorithm, and posture tremor detection apparatus |
| CN116304861A (en) * | 2023-02-09 | 2023-06-23 | 河北工业大学 | Time-frequency characteristic fusion fault diagnosis method based on self-attention |
| CN116578832A (en) * | 2023-07-13 | 2023-08-11 | 兰州理工大学 | Device for predicting surface roughness of machined part |
| CN116787225A (en) * | 2023-04-10 | 2023-09-22 | 杭州电子科技大学 | Rotary cutter abrasion monitoring method based on sound vibration signal characteristic fusion graph |
-
2024
- 2024-01-03 CN CN202410004935.0A patent/CN117495211B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3421988A1 (en) * | 2017-06-29 | 2019-01-02 | Domesens srl | Device and a method for the continuous measurement of one or more tree staticity |
| JP2019111969A (en) * | 2017-12-25 | 2019-07-11 | トヨタ自動車株式会社 | Hybrid automobile |
| US10950352B1 (en) * | 2020-07-17 | 2021-03-16 | Prince Mohammad Bin Fahd University | System, computer-readable storage medium and method of deep learning of texture in short time series |
| CN112069966A (en) * | 2020-08-31 | 2020-12-11 | 中国计量大学 | Surface roughness prediction method based on improved LSTM network |
| CN114869272A (en) * | 2021-09-03 | 2022-08-09 | 中国人民解放军总医院 | Posture tremor detection model, posture tremor detection algorithm, and posture tremor detection apparatus |
| CN116304861A (en) * | 2023-02-09 | 2023-06-23 | 河北工业大学 | Time-frequency characteristic fusion fault diagnosis method based on self-attention |
| CN116787225A (en) * | 2023-04-10 | 2023-09-22 | 杭州电子科技大学 | Rotary cutter abrasion monitoring method based on sound vibration signal characteristic fusion graph |
| CN116578832A (en) * | 2023-07-13 | 2023-08-11 | 兰州理工大学 | Device for predicting surface roughness of machined part |
Non-Patent Citations (3)
| Title |
|---|
| HU, M等: "Bearing fault diagnosis method based on data augmentation and MCNN-LSTM", 《 2023 IEEE 6TH INFORMATION TECHNOLOGY, NETWORKING, ELECTRONIC AND AUTOMATION CONTROL CONFERENCE (ITNEC)》, vol. 6, 22 April 2023 (2023-04-22) * |
| 厉大维;沈明瑞;张贺清;王书利;陈锋;刘阔;王永青;: "基于长短时记忆网络的深孔镗削刀具状态监测", 现代制造工程, no. 08, 18 August 2020 (2020-08-18) * |
| 潘飞: "基于多尺度特征的 Attention-MCNN-LSTM 模型的构建和实证研究", 《中国优秀硕士学位论文全文数据库 信息科技；经济管理与科学辑》, no. 2022, 15 June 2022 (2022-06-15) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117495211B (en) | 2024-03-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110361176B (en) | Intelligent fault diagnosis method based on multitask feature sharing neural network | |
| Madhusudana et al. | Condition monitoring of face milling tool using K-star algorithm and histogram features of vibration signal | |
| Wong et al. | Technical data-driven tool condition monitoring challenges for CNC milling: a review | |
| CN103465107B (en) | A kind of Tool Wear Monitoring method | |
| CN109947086B (en) | Mechanical fault migration diagnosis method and system based on counterstudy | |
| CN108942409B (en) | Modeling and monitoring method for tool wear loss based on residual convolutional neural network | |
| Wang et al. | CHMM for tool condition monitoring and remaining useful life prediction | |
| Li et al. | Similarity-measured isolation forest: Anomaly detection method for machine monitoring data | |
| CN110303380A (en) | A kind of cutting tool for CNC machine method for predicting residual useful life | |
| CN112613646A (en) | Equipment state prediction method and system based on multi-dimensional data fusion | |
| CN107553219A (en) | A kind of Tool Wear Monitoring method based on multiple types sensor composite signal | |
| CN114619292B (en) | Milling cutter wear monitoring method based on fusion of wavelet denoising and attention mechanism with GRU network | |
| CN111366123A (en) | Part surface roughness and cutter wear prediction method based on multi-task learning | |
| CN112766182B (en) | Broaching tool wear state identification method | |
| Dong et al. | Multiscale feature extraction from the perspective of graph for hob fault diagnosis using spectral graph wavelet transform combined with improved random forest | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN112192318B (en) | Machining tool state monitoring method and system | |
| CN112372371B (en) | Method for evaluating abrasion state of numerical control machine tool cutter | |
| CN116520772A (en) | Sheet part numerical control machining quality prediction method based on multitasking transfer learning | |
| CN114491828A (en) | Compound fault diagnosis method based on multi-gating hybrid expert multi-task learning | |
| CN114354184A (en) | Deep learning-based method and device for establishing health early warning model of main shaft of large-scale rotating equipment | |
| CN115221973A (en) | Aviation bearing fault diagnosis method based on enhanced weighted heterogeneous ensemble learning | |
| CN112801139B (en) | Intelligent cutter wear state identification method based on heterogeneous domain self-adaptive transfer learning | |
| Xu et al. | Total process of fault diagnosis for wind turbine gearbox, from the perspective of combination with feature extraction and machine learning: A review | |
| CN117495211B (en) | Industrial master machining workpiece quality prediction method based on self-adaptive period discovery |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |