CN117540256A - Method for detecting nonstandard operation process of electron microscope based on instrument working current - Google Patents
Method for detecting nonstandard operation process of electron microscope based on instrument working current Download PDFInfo
- Publication number
- CN117540256A CN117540256A CN202311757933.0A CN202311757933A CN117540256A CN 117540256 A CN117540256 A CN 117540256A CN 202311757933 A CN202311757933 A CN 202311757933A CN 117540256 A CN117540256 A CN 117540256A
- Authority
- CN
- China
- Prior art keywords
- current
- sequence
- lstm
- operation process
- training
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 134
- 230000008569 process Effects 0.000 title claims abstract description 109
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 20
- 230000002159 abnormal effect Effects 0.000 claims abstract description 16
- 238000001514 detection method Methods 0.000 claims abstract description 14
- 238000012549 training Methods 0.000 claims description 35
- 239000013598 vector Substances 0.000 claims description 13
- 238000000605 extraction Methods 0.000 claims description 12
- 238000009826 distribution Methods 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 10
- 238000013527 convolutional neural network Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 7
- 238000010276 construction Methods 0.000 claims description 6
- 230000005856 abnormality Effects 0.000 claims description 5
- 238000011017 operating method Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 4
- 238000002372 labelling Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 2
- 238000012804 iterative process Methods 0.000 claims 1
- 230000001902 propagating effect Effects 0.000 claims 1
- 230000001788 irregular Effects 0.000 abstract description 3
- 238000007405 data analysis Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 14
- 238000010586 diagram Methods 0.000 description 9
- 230000000694 effects Effects 0.000 description 5
- 108010076504 Protein Sorting Signals Proteins 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 4
- 238000004088 simulation Methods 0.000 description 4
- 238000010894 electron beam technology Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000007635 classification algorithm Methods 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000007599 discharging Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 108090000623 proteins and genes Proteins 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000008685 targeting Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06F18/23—Clustering 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
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Analysing Materials By The Use Of Radiation (AREA)
Abstract
The invention discloses an irregular detection method for an electron microscope operation process based on instrument working current, and belongs to the field of data analysis. According to the invention, working current of an electron microscope is firstly obtained from a scientific instrument management system, forward connection is carried out according to the size of a window, abnormal scoring is carried out on a current interval through a VAE-LSTM algorithm structure, and current points with abnormal score values are recorded. And finally, carrying out operation process non-standard classification on the abnormal current through a CWT-LSTM algorithm structure, thereby obtaining a specific operation non-standard type and sending out a prompt. The invention can obtain the abnormal current more accurately and predict the abnormal current in an operation non-standard type.
Description
Technical Field
The invention belongs to the technical field of data analysis, and particularly relates to an irregular detection method for an electron microscope operation process based on instrument working current data.
Background
During the use of the instruments, each instrument has a corresponding operation process, and during the use of the instruments by the instrument users, the phenomenon of nonstandard operation processes can occur. An out-of-specification procedure refers to a user of the instrument not proceeding in a predetermined step or sequence while performing a particular task or operation. The importance of identifying and correcting non-normative phenomena in the instrument operation process is not neglected. First, the correct operation helps to extend the useful life of the instrument and reduce the frequency of maintenance and replacement. Secondly, correct operation of the instrument can also improve the experimental effect of the user. Monitoring the instrument operating current through the sensor can help us identify the phenomenon of non-normative operation of the instrument.
Electron microscopes play an important role in the fields of scientific research, industrial production, medical diagnosis, and the like. The working process of the electron microscope mainly comprises the steps of lens taking and placing, light focusing, lens placing, focusing and observing, and the processes can generate corresponding working current signals, wherein the current in the lens taking and placing process is relatively smaller; the focusing process is carried out by adjusting the current of the lens so as to focus and align, a certain current is needed to control the behavior of the electron beam, and larger current protrusions exist; the current regulation is not involved in the process of discharging the sheet, and the current is relatively small; the focusing process needs to focus and deeply adjust the electron beam, and the current is relatively large; the "look" process current is longer in duration and may require adjustment of different current parameters depending on the characteristics of the sample and the imaging effect desired.
The operation process non-standardization of the electron microscope is divided into two types of operation step deletion and operation sequence mismatching. First, the absence of the operation step means that the operator has a step missing when performing the operation, and the two steps of "focusing" and "focusing" are easily missing during the operation of the electron microscope. Secondly, the improper operation sequence means that the operator does not operate according to the process sequence when performing the operation of the electron microscope, and during the operation of the electron microscope, there may be two steps of "putting a sheet" and "focusing" in the case of improper sequence. Typically, a "focusing" operation, i.e. focusing and aligning the electron beam, should be performed first to ensure that a clear image is obtained. Then "slide", the sample is placed in an electron microscope for viewing. Such a sequence can ensure that a clear image is observed. The method identifies the problem that operation steps are missing and operation flows are not right when an operator uses the electron microscope through monitoring the working current of the instrument, and sends out a prompt.
For such problems, the data points are typically classified by a clustering/classification algorithm in machine learning or a time series algorithm in deep learning. For example, document [ Shi Haipeng ] improved k-means intelligent plant production anomaly identification model based on improved k-means intelligent plant production anomaly identification model in the form of dynamic time warping (Dynamic Time Warping, DTW) algorithm is established based on improved k-means intelligent plant production anomaly identification and application study [ D ]. Guizhou university, 2022.doi:10.27047/d.cnki.ggudu.2021.002367 ], but if the data type is unbalanced or the feature value is selected improperly, the clustering effect is poor. For example, the motor fault diagnosis study [ J ]. Bao Steel technology based on CNN current data morphology recognition [ Jingqun, li Jie, huang Dongming ] adopts CNN convolutional neural network to recognize abnormal morphology of each feature under fault, establishes an optimization model targeting false alarm rate and false alarm rate, optimizes each feature morphology combination of different fault types by genetic algorithm using abnormal morphology as gene segment, but this approach ignores time series characteristic of current data and cannot fully utilize time-frequency characteristic of current data. The two methods neglect the time sequence characteristics of the current data, cannot accurately locate the abnormal point of the current, and cannot fully extract the time-frequency domain information in the current data.
Disclosure of Invention
In view of the foregoing, the present invention provides a method for detecting non-normative operation of an electron microscope based on VAE and CWT-LSTM. The invention collects a large amount of current data corresponding to the normal operation process of the electron microscope, trains the current data through a time sequence abnormality detection model, and positions the current which is not standard in the operation process through abnormality scoring; the time-frequency domain information of the nonstandard current in the operation process is fully extracted through wavelet transformation, then the data set with the nonstandard process labels is trained through supervised learning, the nonstandard current in the operation process is classified, and a universal model is trained, so that the method has a high engineering value.
The invention comprises the following steps:
(1) For a large amount of electron microscope running current data only including normal operation process, a running current data set corresponding to the normal operation process is constructed, and the data set is divided into a training set and a testing set.
(2) And constructing a model based on a VAE-LSTM time sequence anomaly detection algorithm, and constructing an LSTM network in the encoder and decoder part to train the training set.
(3) And aiming at the current in the normal operation process, carrying out current reconstruction according to the missing of the light process, the missing of the focusing process and the improper sequence of the light process by the film discharge, constructing a data set which is not standard in the operation process of the electron microscope, carrying out state marking, and dividing the data set into a training set and a test set.
(4) And constructing a CWT-LSTM-based supervised algorithm model, and training the training set based on wavelet transform (CWT), a feature extraction module and a BiLSTM network.
(5) And (3) carrying out anomaly scoring on the current sequence by using the trained VAE-LSTM model, and positioning the current with nonstandard operation process.
(6) And (3) inputting the current which is not standardized in the operation process into a trained CWT-LSTM model, so that the specific operation non-standardized type corresponding to each moment of the current can be obtained, and further, the specific operation non-standardized current interval can be further positioned.
Further, the specific implementation manner of the data set construction and the data preprocessing in the step (1) is as follows: firstly, a plurality of groups of electron microscope current sequences are selected from a background database, and the current sequences are intercepted by a professional technician. The intercepted current sequences are all in a working state and have a correct operation process. Then converting the intercepted current sequence into current input vectors at all moments according to 7: the scale of 3 is divided into training and test sets.
Further, the method for constructing the VAE-LSTM time sequence abnormality detection algorithm model in the step (2) is as follows: firstly, the encoder uses an LSTM network to encode input current through the LSTM network, and maps output into corresponding hidden variable distribution parameters through a full connection layer.
Further, the decoder uses an LSTM network to input the hidden variable sequence output by the encoder into the LSTM network for decoding, and the output result is averaged to obtain a reconstruction result of the sample.
Further, in the training process, for an input current sequence only containing correct operation process current, iteration times are set, parameters of an encoder and a decoder of a VAE-LSTM model are initialized, samples of the sequence are randomly sampled and traversed in the iteration process, a Loss function Loss is calculated after the samples pass through the encoder and the decoder, and model parameters are updated according to back propagation of the Loss function.
Further, the specific implementation manner of the data set construction and the data preprocessing in the step (3) is as follows: selecting a plurality of groups of current sequences of the electron microscope with normal operation processes from a background database, carrying out current reconstruction according to the missing of the light process, the missing of the focusing process and the improper sequence of the light process by the lens placement, converting the reconstructed current sequences into current input vectors at all moments, and carrying out state labeling on the current sequences according to the reconstruction types, wherein the reconstruction types comprise the missing of the light process, the missing of the focusing process, the improper sequence of the light process by the lens placement and other non-standard processes according to 7: the scale of 3 is divided into training and test sets.
Further, the method for constructing the CWT-LSTM supervised algorithm in the step (4) comprises the following steps: the signal is first time-frequency analyzed and processed using a continuous wavelet transform.
Furthermore, the feature extraction module performs feature extraction by using three parallel CNNs, and can enable convolution operation to capture features more comprehensively through the combination of convolution kernels with different sizes.
Further, the BiLSTM network obtains a final non-standard recognition result of the current operation process by enabling the two LSTMs to independently process the forward and backward input sequences respectively.
Further, in the training process, for an input reconstructed current sequence, model parameters are initialized, sequence samples are randomly sampled and traversed in the iteration process, a Loss function Loss is calculated after continuous wavelet transformation, a feature extraction module and a BiLSTM network are adopted, and the model parameters are updated according to the Loss function Loss in a counter-propagation mode.
Further, the procedure of abnormal scoring and positioning the current sequence in the step (5) is as follows: the input current sequence is first converted into current input vectors at various moments.
Further, the current sequence is scored using a VAE-LSTM model that has learned the normal operating procedure current sequence, locating an outlier score current point that is greater than the maximum normal operating procedure score.
Further, the specific process of obtaining the abnormality type corresponding to each abnormal current point in the step (6) is as follows: and (3) processing the abnormal fractional current sequence in the same way as training, and inputting the abnormal fractional current sequence into a trained CWT-LSTM model to obtain specific operation non-standard types of each current point of the instrument, thereby positioning a specific operation non-standard current interval.
The invention has the beneficial effects that:
1) The invention uses the VAE-LSTM model, trains through a large number of current sequences corresponding to normal operation processes, and positions out-of-specification current in the operation processes in a scoring mode.
2) In order to better extract the time-frequency domain information of the input sequence and simultaneously extract the forward current interval sequence information and the backward current interval sequence information when extracting the current interval sequence information, a CWT-LSTM model based on wavelet transformation and BiLSTM is adopted. After enough and complete data training, manual classification is not needed, and the method has certain engineering value.
3) The invention subdivides the irregular type of the electron microscope operation process into missing of the light process, missing of the focusing process and improper sequence of the light process of the film release, and has great reference value for operators.
4) The detection method for the non-standard operation process of the electron microscope based on the VAE and the CWT-LSTM has better effect than the identification effect based on the SVM or the BiLSTM.
Drawings
FIG. 1 is a schematic diagram of the overall process of the present invention.
Fig. 2 is a diagram showing the normal operation process and the corresponding current characteristics of the electron microscope.
FIG. 3 is a schematic diagram of the basic structure of the VAE-LSTM anomaly detection algorithm model of the present invention.
FIG. 4 is a schematic diagram of the basic structure of the model of the non-canonical classification algorithm in the CWT-LSTM operation of the present invention.
FIG. 5 is a schematic diagram of simulation results of an anomaly detection algorithm model according to the present invention.
FIG. 6 is a schematic diagram of simulation results of the overall algorithm model of the present invention.
Detailed Description
In order to more specifically describe the present invention, the following detailed description of the technical scheme of the present invention is given with reference to the accompanying drawings and examples.
Implementations of embodiments of the present application are based on a large amount of electron microscope current data collected by a current sensor, and the general detailed procedure is shown in fig. 1.
(1) Establishing VAE-LSTM algorithm data sets
1.1 selecting a Current sequence
And selecting a plurality of groups of current sequences in a working state from a background database, wherein the types of equipment are all electron microscopes.
1.2 Current sequence division
The running current sequences in the training set are screened by professional technicians, so that each current interval sequence is ensured to contain the complete normal operation process of the electron microscope, and the operation process and the corresponding waveform characteristics of the electron microscope are shown in figure 2.
Assuming that the current data of one electron microscope is S= {0.04,0.04,0.05,0.04,0.05,0.04,0.04,1.34,1.37,1.33,1.35, …,1.31,0.34,0.34,0.35,0.36,0.35}, then the running current sequence is connected forward according to the window size, and converted into current input vectors at all times, namely, the current value i at any time t in the running current sequence t Taking i according to the preset window size w t And the current values of the previous w moments constitute a current input vector i at moment t t-w ，i t-w+1 ，...，i t-1 ，i t ]And converting into a Tensor data type to generate a training current set X corresponding to the moment t t . According to the method, the current at each moment is processed to obtain a training set sequence Seq, 70% of data volume is randomly selected as a training set, and 30% of data volume is selected as a test set.
X t ＝Tensor({i t-w ，i t-w+1 ，...，i t-1 ，i t })
Seq＝{X 0 ，X 1 ，X 2 ，...，X t ,X t+1 ，X n })
(2) Construction of VAE-LSTM algorithm model
2.1 LSTM based Encoder Structure
As shown in fig. 3, the LSTM network is first used to encode the timing information and learn the distribution of the input data. Mapping the coding result into hidden variable distribution parameter mu through the full connection layer i Sum sigma i The calculation formula is shown as follows:
μ i ，σ i ＝Enc(h i-1 ，s i-1 ，x i )
wherein mu i Sum sigma i Respectively represent input x i Corresponding hidden variable distribution mean and standard deviation, enc represents the parameters of the encoder, h i-1 Sum s i-1 Representing the hidden state vector and the cell state of the LSTM network, respectively.
By standard normal distributionThe middle sampling is carried out, and meanwhile, proper random noise xi is added, so that the input x can be obtained i Corresponding hidden variable z i In this way, the problem of reverse gradient fracture in neural network training can be avoided, and the calculation formula is shown as follows:
z i ＝μ i +σ i ⊙ξ
wherein, the product of Hadamard, ζ to N (0,I), I is the identity matrix.
Let the sampling times be K, sample x i The hidden variable sequence obtained after the coding operation isWherein z is i k Representing the kth time to sample x i Corresponding hidden variables are obtained when sampling is carried out on the hidden variable distribution.
2.2 LSTM based Decoder Structure
Sample x i Hidden variable sequence obtained after coding operationInputting the sample into an LSTM network for decoding, and obtaining a sample x after averaging i Reconstruction result of->The calculation formula is shown as follows:
wherein y is k Representing hidden variable Z i k Decoding output result h i-1 Sum s i-1 The hidden state vector and the cell state of the LSTM network are represented, respectively, and Dec represents the decoder parameters.
Will input a current signalThe reconstruction results of each sample in the sequence are arranged in sequence to obtain a final reconstruction current signal sequence
2.3 loss function
In the training process, the loss function has two aspects to be considered simultaneously. On the one hand, the hidden variable distribution of all samples in the current signal sequence is approximate to the standard normal distribution, and on the other hand, the reconstructed current signal sequence is similar to the input current signal sequence as much as possible. The Loss function Loss is thus composed of regularization Loss and mean square error Loss, as shown in the following equation:
2.4 training procedure
And inputting a current signal data set in a normal operation process, setting the iteration number as k, and initializing encoder parameters and decoder parameters of the VAE-LSTM model. M sequence samples are randomly sampled from the dataset when the number of iterations is less than k. Traversing M sequence samples from distribution N to (mu, sigma) by an encoder 2 ) A series of hidden variables z are sampled, decoding results are obtained by the decoder, and encoder and decoder parameters are updated according to a Loss function counter-propagating Loss.
(3) Establishing CWT-LSTM model data set
3.1 Current sequence selection and Current reconstruction
And selecting a plurality of groups of current sequences corresponding to the electron microscope equipment with normal operation processes from a background database, and carrying out current reconstruction according to the type of missing the focusing process, the type of improper sequence of the focusing process and other non-standard process types.
Assuming that current data corresponding to a normal slice-setting and light-setting process of an electron microscope is D= {0.11,0.51,2.71,2.59,2.12,2.13,2.15,2.09,1.52,0.91,0.52,0.12,0.11,0.09,0.12,0.13}, reconstructing the current data according to a type of improper slice-setting and light-setting process sequence to obtain reconstructed data S= {0.12,0.11,0.09,0.12,0.13,0.11,0.51,2.71,2.59,2.12,2.13,2.15,2.09,1.52,0.91,0.52}, and corresponding to a label set C= {2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2}, wherein label 2 represents a state of improper slice-setting and light-setting process sequence, in addition, label 1 represents a focusing process missing state, label 0 represents a light-setting process missing state, and label 3 represents other non-standard process states.
3.2 forward connection
Assuming that the reconstructed data of a certain electron microscope is S, the initial current data is connected forward according to the window size. Generating a training current set corresponding to the moment t
S t ＝{i t-w ，i t-w+1 ，...，i t-1 ，i t }
(4) Constructing CWT-LSTM algorithm model
4.1 wavelet transform (CWT) extraction
As shown in fig. 4, the signal is time-frequency analyzed and processed using a continuous wavelet transform as shown in the following equation:
where x (t) is the input current signal, a is the scale, and τ is the offset.
The n input windows S obtained above are transformed by wavelet i And sequentially calculating a wavelet time-frequency diagram of each window, wherein the wavelet time-frequency diagram is shown in the following formula:
Y i ＝CWT(S i )
the wavelet time-frequency diagram of the continuous segments is rearranged into a sequence as shown in the following formula:
Y cwt ＝[Y 1 ，Y 2 ，...，Y n ]
and then to this time status label L t Mapping to finally generate a training set X corresponding to the moment t t The method comprises the steps of carrying out a first treatment on the surface of the The current at each moment is processed by the method to obtain the training setSequence Seq, randomly selecting 70% of data volume as training set and 30% of data volume as test set.
X t ＝[Tensor(Y t )，L t ]
Seq＝{X 0 ，X 1 ，X 2 ，...，X n-2 ，X n-1 ，X n })
4.2 feature extraction Module
Three parallel CNNs are used for feature extraction, and the convolution kernel sizes are set to 3, 5 and 7 in sequence. Each branch of the three parallel CNNs consists of four Conv and one-dimensional adaptive pooling layer. Each Conv contains one-dimensional convolution operation (Conv 1 d), batch Normalization (BN), linear rectification function (ReLU) and one-dimensional maximum pooling (Maxpool 1 d). By combining convolution kernels of different sizes, the convolution operation can capture features more comprehensively.
4.3 BiLSTM network
In the current state identification process, the machine-time state type corresponding to the current in the current section needs to consider both forward change and backward change. Therefore, in this embodiment, a BiLSTM network is used, and the following formula is used to process the forward and backward input sequences separately by two LSTMs, and combine the feature vectors obtained in the two directions to obtain the final current abnormal state recognition result:
wherein:for the hidden state of the forward LSTM cell at time t +.>Is the hidden layer state of the backward LSTM unit at the time t, x t For the time-frequency characteristic vector extracted by the characteristic extraction module at the time t,/time-frequency characteristic vector>Is the hidden layer state of the forward LSTM unit at the time t-1,,, the>Is the hidden layer state of the backward LSTM unit at the time t+1, LSTM + () And LSTM - () Representing the calculation functions of the forward LSTM cells and the reverse LSTM cells, respectively.
Further using Softmax activation to obtain a current classification status result, the following formula:
the loss function adopts Cross-Entropy Cross Entropy function aiming at multiple classifications to obtain a prediction resultAnd the true result y.
And (5) continuously and iteratively updating model parameters by using an optimizer through a gradient descent method according to the Loss function Loss until the Loss function Loss converges, and finishing training.
(5) Operation process non-canonical current point detection
5.1 forward connecting the input current sequence according to the window size during training.
5.2 scoring input Current
For the input intra-window data x= { X 1 ,x 2 ,,……,x n By learning the VAE-LSTM model of the current sequence of normal operation Calculating anomaly score of input windowThe current points at which an anomaly score greater than the maximum normal operation process score occurs are marked as anomaly current points, and the simulation result is shown in fig. 5.
(6) Obtaining the type of non-specification of specific operation process
And 6.1, forward connecting the non-standard current points of each operation process according to the window size during training, and inputting the forward connecting points into a trained CWT-LSTM model to obtain the non-standard type of the operation process corresponding to the non-standard current points of each operation process.
6.2 as shown in the simulation result of fig. 6, the operation process non-canonical sequence can be obtained.
The evaluation indexes of the method comprise Precision (PR), recall (RE) and F1 score (F1-score, F1), and the total Accuracy (ACC) is calculated according to the following formula:
where TP represents positive samples and FP represents negative samples and TN represents negative samples, and FN represents positive samples and FN represents negative samples. PR is used for evaluating the proportion of the positive class predicted by the classifier to be the actual positive class, RE is used for evaluating the proportion of the positive class sample predicted correctly to the positive class sample in the data set, F1 is the harmonic mean of PR and RE, and ACC is used for evaluating the proportion of the correct sample number of the model to the total number of the data set samples.
Table 1 gives the comparison of 3 different classification models with the model presented herein for the non-canonical detection of the operation of electron microscope current sequences. Compared with other models, the model has the best overall performance and reaches the highest overall evaluation index. The model utilizes anomaly detection to accurately locate the current which is not normalized in the operation process, further effectively captures time-frequency characteristics by using continuous wavelet transformation, and utilizes forward and backward characteristics by using bidirectional LSTM to better classify the type of the non-normalized operation process, thereby locating the non-normalized operation process.
TABLE 1
The embodiments described above are described in order to facilitate the understanding and application of the present invention to those skilled in the art, and it will be apparent to those skilled in the art that various modifications may be made to the embodiments described above and that the general principles described herein may be applied to other embodiments without the need for inventive faculty. Therefore, the present invention is not limited to the above-described embodiments, and those skilled in the art, based on the present disclosure, should make improvements and modifications within the scope of the present invention.
Claims (8)
1. The method for detecting the non-standard operation process of the electron microscope based on the instrument working current is characterized by comprising the following steps of:
(1) Aiming at a large number of electron microscope operation current data only including normal operation process, constructing an operation current data set corresponding to the normal operation process, and dividing the data set into a first training set and a first test set;
(2) Constructing a model based on a VAE-LSTM time sequence anomaly detection algorithm, constructing an LSTM network at the encoder and decoder part, and training a first training set;
(3) Aiming at normal operation process current, current reconstruction is carried out according to the missing of the light process, the missing of the focusing process and the improper sequence of the light process of the light placing, a data set which is not standard in the operation process of the electron microscope is constructed, state labeling is carried out, and the data set is divided into a second training set and a second test set;
(4) Constructing a CWT-LSTM-based supervised algorithm model, and training a second training set based on wavelet transformation, a feature extraction module and a BiLSTM network;
(5) Performing abnormal scoring on the current sequence by using a trained VAE-LSTM model, and positioning the current with nonstandard operation process;
(6) And (3) inputting the current which is not standardized in the operation process into a trained CWT-LSTM supervised algorithm model, so that the specific operation non-standardized type corresponding to each moment of the current can be obtained, and further, the specific operation non-standardized current interval is further positioned.
2. The method of claim 1, wherein: the specific implementation mode of the data set construction and the data preprocessing in the step (1) is as follows:
firstly, selecting a plurality of groups of current sequences of the electron microscope from a background database, and intercepting the current sequences by a professional technician; the intercepted current sequences are all in a working state and have a correct operation process;
then converting the intercepted current sequence into current input vectors at all moments according to 7: the scale of 3 is divided into training and test sets.
3. The method of claim 1, wherein: the method for constructing the non-supervision model based on the VAE-LSTM time sequence abnormality detection algorithm in the step (2) is as follows: firstly, the encoder uses an LSTM network to encode input current through the LSTM network, and maps output into corresponding hidden variable distribution parameters through a full connection layer;
and the decoder uses an LSTM network to input the hidden variable sequence output by the encoder into the LSTM network for decoding, and averages the output result to obtain a reconstruction result of the sample.
4. A detection method according to claim 3, wherein: the training process in the step (2) sets iteration times for an input current sequence only containing correct operation process current, and initializes encoder and decoder parameters of the VAE-LSTM model; in the iterative process, sequence samples are randomly sampled and traversed, a Loss function Loss is calculated after the sequence samples pass through the encoder and decoder structures, and model parameters are updated according to the back propagation of the Loss function.
5. The method of claim 1, wherein: the specific implementation mode of the data set construction and the data preprocessing in the step (3) is as follows:
selecting a plurality of groups of current sequences of the electron microscope with normal operation processes from a background database, carrying out current reconstruction according to the missing of the light process, the missing of the focusing process and the improper sequence of the light process by the lens placement, converting the reconstructed current sequences into current input vectors at all moments, and carrying out state labeling on the current sequences according to the reconstruction types, wherein the reconstruction types comprise the missing of the light process, the missing of the focusing process, the improper sequence of the light process by the lens placement and other non-standard processes according to 7: the scale of 3 is divided into training and test sets.
6. The method of claim 1, wherein: the construction method of the step (4) is based on a CWT-LSTM supervised algorithm and comprises the following steps: performing time-frequency analysis and processing on the signal by using continuous wavelet transformation;
the feature extraction module performs feature extraction by using three parallel CNNs;
the BiLSTM network obtains a final non-standard recognition result in the current operation process by enabling the two LSTMs to independently process the forward input sequence and the backward input sequence respectively.
7. The method of detecting according to claim 6, wherein: in the training process of the step (4), for an input reconstruction current sequence, initializing model parameters, randomly sampling sequence samples in the iteration process and traversing, calculating a Loss function Loss after passing through continuous wavelet transformation, a feature extraction module and a BiLSTM network, and reversely propagating and updating the model parameters according to the Loss function Loss.
8. The method of claim 1, wherein: the abnormal scoring and positioning operation of the current sequence in the step (5) is performed by the following specific processes:
converting the input current sequence into current input vectors at all moments;
scoring the current sequence using a VAE-LSTM model that has learned the normal operating procedure current sequence, locating an abnormal score current point that is greater than a maximum normal operating procedure score.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311757933.0A CN117540256A (en) | 2023-12-20 | 2023-12-20 | Method for detecting nonstandard operation process of electron microscope based on instrument working current |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311757933.0A CN117540256A (en) | 2023-12-20 | 2023-12-20 | Method for detecting nonstandard operation process of electron microscope based on instrument working current |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117540256A true CN117540256A (en) | 2024-02-09 |
Family
ID=89782389
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311757933.0A Pending CN117540256A (en) | 2023-12-20 | 2023-12-20 | Method for detecting nonstandard operation process of electron microscope based on instrument working current |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117540256A (en) |
-
2023
- 2023-12-20 CN CN202311757933.0A patent/CN117540256A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110674752B (en) | Tool wear state identification and prediction method based on hidden Markov model | |
| CN103968937B (en) | A kind of distribution switch mechanical state diagnostic method based on EMD Sample Entropies and FCM | |
| CN112529234A (en) | Surface water quality prediction method based on deep learning | |
| CN111898690B (en) | Power transformer fault classification method and system | |
| CN112904810B (en) | Process industry nonlinear process monitoring method based on effective feature selection | |
| CN112347910B (en) | Signal fingerprint identification method based on multi-mode deep learning | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN116070527B (en) | Milling cutter residual life prediction method based on degradation model | |
| CN111783336B (en) | Uncertain structure frequency response dynamic model correction method based on deep learning theory | |
| CN110659682A (en) | Data classification method based on MCWD-KSMOTE-AdaBoost-DenseNet algorithm | |
| CN110995461B (en) | Network fault diagnosis method | |
| CN113837046A (en) | Small sample remote sensing image scene classification method based on iterative feature distribution learning | |
| CN112990593A (en) | Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm | |
| CN117540256A (en) | Method for detecting nonstandard operation process of electron microscope based on instrument working current | |
| CN110347579B (en) | Deep learning test case selection method based on neuron output behavior pattern | |
| CN115628910B (en) | Embedded bearing fault diagnosis device and equipment based on transfer learning | |
| CN116738204A (en) | Power load data anomaly detection method based on generation of countermeasure transformers | |
| CN115407753B (en) | Industrial fault diagnosis method for multi-variable weighting integrated learning | |
| CN107341503B (en) | Identification method for multi-source energy efficiency state in cutting process | |
| CN113835964B (en) | Cloud data center server energy consumption prediction method based on small sample learning | |
| CN118035923B (en) | Power grid wave recording abnormal signal identification method | |
| CN114660931A (en) | Method and system for diagnosing and identifying petrochemical process fault | |
| CN112001530A (en) | Predictive maintenance method and system for transformer oil chromatography online monitoring device | |
| CN113792610B (en) | Health assessment method and device for harmonic reducer | |
| CN117210403B (en) | Preparation method and application of stem cell and immune cell composition for resisting aging |
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 |