CN117009858A - Synchronous classification method for redundant detection signals of aerospace sealed electronic components - Google Patents

Abstract

The invention discloses a synchronous classification method for redundant detection signals of a space-flight sealed electronic component, which designs a self-learning clustering threshold value determining algorithm according to pulse data characteristics of components in the space-flight sealed electronic component after forced vibration, and carries out first judgment and classification on pulse data by adopting a spectral clustering algorithm and a self-design classification standard. And then, a localized characteristic space data set is established by extracting a plurality of characteristic parameters of a large amount of pulse data, and the pulse data is subjected to secondary judgment and classification by adopting a k-NN algorithm after parameter optimization. Finally, the advantages of the two classification methods are combined, and the synchronous classification method is designed. The method solves the problem of multi-classification of the redundant detection signals based on the PIND method, has certain universality and practicability, and has important reference value for quality detection of aerospace sealed electronic components.

Description

Synchronous classification method for redundant detection signals of aerospace sealed electronic components

Technical Field

The invention belongs to the technical field of aerospace science, relates to a method for classifying and identifying redundant detection signals, and in particular relates to a method for classifying redundant detection signals of an aerospace sealing electronic component based on PIND.

Background

The spaceflight sealed electronic component is an important component with electromagnetic interference resistance and high reliability, and is widely applied to equipment such as satellites, rockets, missiles and the like. The aerospace sealing electronic component has the characteristics of complex structure, small volume, multiple types and the like. In the field of aerospace, the reliability requirements for sealing electronic components are very strict, and one of the key links is redundant detection, which is an indispensable part in the conventional inspection of reliability. By redundancy is meant substances that may occur during the production, packaging, transportation, etc. of the device or component, irrespective of the device or component itself, typical redundancy found at present mainly includes chip silicon residue, rosin residue, residual solder residue, metal chips, ceramic or glass chips, solder wire residue, insulation coating of the wire, and the like, as well as the part of the component itself that peels off inside. These redundant substances are difficult to find in the non-working state of the components, and irregular random movement can occur in the working state of severe vibration such as weightlessness or overweight, etc., and possibly contact with the sealed electronic component members or switch contacts in a contact manner, so that the components malfunction or refusing action is caused, and further damage to the system part or whole is caused.

At present, in the field of aerospace in China, a widely applied nondestructive detection method for redundant substances is generally adopted, namely a particle collision noise detection method (Particle ImpactNoise Detection, PIND for short). The PIND method has the core concept that the active impact and vibration of an external vibrating table activate redundant particles in the measured object to collide with the measured object in the measured object, and then the piezoelectric ceramic sensor is used for collecting acoustic signals in the measured object. Such an acoustic signal is referred to as a PIND detection signal (hereinafter referred to as detection signal). In practical detection, we find that the interior of the tested component has not only the acoustic signal (hereinafter referred to as the redundant signal) generated by the collision of the redundant material, but also the vibration signal (hereinafter referred to as the component signal) of the internal component of the tested component. We divide the detection signal into 3 cases: (1) a redundancy signal (containing redundancy pulses); (2) component signals (containing component pulses); (3) A mixed signal (hereinafter referred to as a mixed signal) composed of the redundant signal and the component signal.

The preliminary judgment (whether a detection signal exceeding a threshold exists or not) of the detection result of the redundant substance based on the particle collision noise detection system is easy, but the difficulty of further judging the type of the signal is relatively high, and in actual engineering detection, a detection person with abundant experience is often required to repeatedly observe and analyze the detection waveform for many times. This process is time consuming and labor intensive and is susceptible to subjective awareness by the inspector, leading to erroneous decisions. In summary, how to adopt a certain method to judge the PIND detection result is a difficult problem faced by the detection of redundancy in the current aerospace technology field.

Disclosure of Invention

The invention provides a synchronous classification method for redundant detection signals of spaceflight sealed electronic components in order to solve the problem of multiple classification of redundant detection signals based on a PIND method.

The invention aims at realizing the following technical scheme:

a synchronous classification method for redundant detection signals of spaceflight sealed electronic components comprises the following steps:

step one, data processing

Constructing a redundancy detection platform, and performing data processing on the acquired pulse data to acquire the frequency domain characteristics, the time domain characteristics, the energy characteristics and the distribution characteristics of the pulse;

step two, signal classification design based on spectral clustering algorithm

Step two, taking the pulse frequency domain characteristic data obtained in the step one as an input item, performing pulse similarity calculation to obtain similarity coefficients of all pulses in the PIND detection signal, and obtaining a fuzzy similarity matrix, wherein a calculation formula of the similarity coefficients is as follows:

wherein: r is (r) _ij Representing the degree of similarity of pulse i to pulse j, F _i [k]Representing the spectral sequence of pulse i, F _j [k]Represents the spectral sequence of pulse j, N represents the data length of the short-time fourier transform, n=2 ⁿ (n is a positive integer);

step two, converting the fuzzy similar matrix into a fuzzy equivalent matrix through transfer closed-packet operation;

step two, determining a clustering threshold lambda through a self-learning clustering algorithm, wherein the method specifically comprises the following steps:

setting the number of components in a tested piece as N, loading a self-learning signal, carrying out pulse extraction on the self-learning signal, calculating a fuzzy equivalent matrix, sequencing all similar coefficients in the fuzzy equivalent matrix from large to small, and sequentially takingRecording a current clustering threshold lambda until the pulse number ratio in N classes is greater than 95%, and selecting the smallest clustering threshold lambda from all the recorded clustering thresholds lambda as a learning result after all the self-learning signals are completely learned;

step two, comparing through traversing the fuzzy equivalent matrix elements to obtain a clustering result M, wherein the specific steps are as follows:

in the fuzzy equivalence matrix, when r _ij At > λ, pulses i and j are grouped into one class at truncated level λ; when r is _ij When the lambda is less than or equal to lambda, the pulse i and the pulse j cannot be gathered into one type under the truncated level lambda; if (lambda) _k ) _max ≤(r _ij ) _min Wherein lambda is _k Representing the truncated level data set, then the full of the redundancy detection signalThe partial pulses are grouped into one class; if (lambda) _min ≥(r _ij ) _max All pulses of the redundancy detection signal are respectively of one type;

step two, judging and classifying the redundant signals, the component signals, the mixed signals and the oversized component signals according to the clustering result M and the pulse duty ratio P and the signal classification standard, wherein the specific classification steps are as follows: (1) if P is less than or equal to 0.5%, judging that the signal is an excessive signal; (2) If P is more than 0.5% and less than or equal to 5% and M is more than 3, judging that the signal is an excessive signal; (3) If P is more than or equal to 0.5% and less than or equal to 10% and M is more than or equal to 3, judging that the signal is a movable component signal; (4) If P is more than 5% and less than or equal to 10%, M is more than 3 and less than or equal to 15, judging that manual checking is needed; (5) If P is more than 5% and less than or equal to 10% and less than 15% and less than M, judging the mixed signal; (6) If the ratio is less than 10% < P, judging that the signal of the movable assembly is too large;

step three, signal classification design of k-NN algorithm based on parameter optimization

Step three, taking the pulse data obtained in the step one as an input item, and obtaining the processed data through short-time Fourier transform and pulse normalization processing;

step three, carrying out feature extraction on the data processed in the step three from four aspects of frequency domain features, time domain features, energy features and distribution features, selecting chi-square test and screening feature parameters based on a decision tree model, and obtaining a multi-dimensional feature set as a feature basis of k-NN algorithm classification, wherein the multi-dimensional feature set comprises the following 8 features: cepstrum coefficient difference, crest coefficient, area occupation ratio, zero crossing rate, energy density, variance, pulse area, and upper and lower pulse symmetry;

thirdly, adopting a cosine distance method as a space distance calculation method through a majority compliance method;

thirdly, calculating through testing accuracy under different k values, and determining the k value;

step three, training and testing pulse data through a KNN algorithm with optimized parameters, and obtaining classification results of different pulses;

and fourthly, comparing and selecting corresponding classification results according to the following rules to realize synchronous classification:

classifying the redundant signals and the oversized component signals by using the method of the second step;

for the component signals and the mixed signals, classification is performed by using the method of the third step.

Compared with the prior art, the invention has the following advantages:

according to the pulse data characteristics of the components in the space-flight sealed electronic component after forced vibration, a self-learning clustering threshold determining algorithm is designed, and the pulse data is judged and classified for the first time by adopting a spectral clustering algorithm and a self-design classification standard. And then, a localized characteristic space data set is established by extracting a plurality of characteristic parameters of a large amount of pulse data, and the pulse data is subjected to secondary judgment and classification by adopting a k-NN algorithm after parameter optimization. Finally, the advantages of the two classification methods are combined, and the synchronous classification method is designed. The method is a closed loop design fused by two methods, comprehensively considers the advantages of the two classification methods, gives consideration to the detection accuracy and the detection speed, and finally verifies the feasibility. The method solves the problem of multi-classification of the redundant detection signals based on the PIND method, has certain universality and practicability, and has important reference value for quality detection of aerospace sealed electronic components.

Drawings

FIG. 1 is a general design diagram of a method for detecting and classifying redundancy;

FIG. 2 is a flow chart for determining a self-learning cluster threshold;

FIG. 3 is a flowchart of a spectral clustering algorithm;

FIG. 4 is a flow chart of a signal classification criteria design;

FIG. 5 is a graph of a typical redundancy signal, (a) redundancy signal, (b) monopulse waveform;

FIG. 6 is a graph of typical component signals, (a) component signals, (b) monopulse waveforms;

FIG. 7 is a graph of feature selection results, (a) chi-square test of feature parameters, (b) feature parameter importance ranking based on decision trees;

FIG. 8 shows the calculation accuracy for selecting different K values;

fig. 9 shows the time efficiency and accuracy of signal classification for different algorithms, 1: algorithm 1 calculates time, 2: algorithm 2 calculates time, 3: algorithm 1 calculates the accuracy, 4: calculating accuracy rate by algorithm 2, wherein: redundancy signal, ii: component signal, iii: mixed signal, IV, oversized component signal.

Detailed Description

The following description of the present invention is provided with reference to the accompanying drawings, but is not limited to the following description, and any modifications or equivalent substitutions of the present invention should be included in the scope of the present invention without departing from the spirit and scope of the present invention.

The invention provides a synchronous classification method for redundant detection signals of spaceflight sealed electronic components, which comprises the following steps:

1. overall design

Firstly, the research process of the method for detecting and classifying the redundant substances provided by the invention is as follows: firstly, constructing a redundancy detection platform through a DZJC-III type PIND detector, and collecting detection data. After acquisition of the pulse data, data processing is performed so as to acquire sufficient data features (frequency domain features, time domain features, energy features, distribution features).

Secondly, the calculation process of the method for detecting and classifying the redundant objects provided by the invention is as follows:

method 1: signal classification design based on spectral clustering algorithm

Firstly, taking frequency domain characteristic data of a pulse as an input item, calculating pulse similarity, and acquiring a fuzzy similarity matrix. Then, the fuzzy similar matrix is converted into a fuzzy equivalent matrix by passing a closed-packet operation. Secondly, according to a large amount of experimental experience, in a complete detection signal, the regular component pulse number is generally equal to or more than 95%, and the irregular pulse number is generally not more than 5%. With the accumulation of later data, the parameter may be adjusted, which is generally referred to herein only. Therefore, the invention designs a self-learning Xi Julei algorithm to determine the clustering threshold lambda, and obtains the clustering result M by traversing the fuzzy equivalent matrix elements for comparison. Secondly, according to the clustering result M and the pulse duty ratio P, a signal classification standard is designed according to experience, and the redundant signals, the component signals, the mixed signals and the oversized component signals are judged, identified and classified.

Method 2: first, the acquired pulse data is used as an input item, and the processed data is acquired through a series of processes such as short-time fourier transform and pulse normalization (data zero padding). And extracting the characteristics of the processed data from the angles of frequency domain, energy, time and distribution to obtain a multi-dimensional characteristic set. Secondly, a calculation method adopting a cosine distance method as a space distance is determined by a majority compliance method. And determining the magnitude of the k value through calculation of test accuracy under different k values. Finally, on the basis, training and testing pulse data through a KNN algorithm with optimized parameters, and obtaining classification results of different pulses.

Then, the two methods are combined to judge and recognize the detection data.

Finally, comparing and selecting the classification result corresponding to each method according to the rules in the invention, and realizing synchronous classification.

The general design of the method for detecting and classifying the residues is shown in fig. 1.

2. Method 1: signal classification design based on spectral clustering algorithm

The component signal is generated along with vibration of the vibration table when the redundant object is detected by the components in the space-flight sealed electronic component, and belongs to the normal signal category. The existing redundant detection system identifies the component signals according to the periodically occurring characteristics of the component signals. In actual detection, however, erroneous judgment often occurs because the periodicity of the component signals is not ideal. Aiming at the difference of the redundancy signal and the component signal in pulse amplitude, pulse energy and pulse length, a spectral clustering algorithm is introduced to cluster the component signal, and the accurate classification of the redundancy signal, the component signal, the mixed signal and the oversized component signal is realized by designing a signal classification standard.

2.1 self-learning Xi Julei threshold algorithm

For different types of tested pieces, the redundant signals are different, but the component signals in the same batch of tested pieces are certainly generated by a plurality of identical components, so that the component signals in the same batch of tested pieces have certain similarity. If a lot of tested pieces are detected, a large number of pure component signals are adopted for clustering training. And redundant detection is carried out on the rest tested pieces by adopting the trained clustering threshold value, so that the clustering threshold value can be flexibly adjusted according to the characteristics of PIND pulse signals of different tested pieces. Firstly, the clustering threshold value should be as high as possible to gather all the component signals into 1 class, so that the recognition rate of the system to the component signals is improved. Under the clustering threshold value, the redundant signals and the oversized component signals can be easily distinguished through the pulse length ratio, then the component signals are extracted through distinguishing the clustered category number, and the rest signals are mixed signals.

The number proportion of irregular pulses in a complete component signal in all pulses is generally not more than 5%, so the self-learning clustering threshold algorithm designed by the invention is as follows: the principle of jumping out of the condition is that all pulses are gathered into one type, and more than 95% of pulses are gathered into one type. And meanwhile, a plurality of components possibly exist in one tested piece at the same time, and the clustering result can have the category number corresponding to the number of the components. Considering this, in the clustering threshold algorithm, if the number of components in a signal is L, and the sum of the pulse numbers in the L classes is 95% or more, the self-learning is also ended.

On the other hand, for a given self-learning signal, its fuzzy equivalence matrix is fixed. If the number of pulses is N, the different similarity coefficients in the fuzzy similarity matrix are at mostAnd each. If the similarity coefficients are arranged from large to small, the similarity coefficients are sequentially r ₁ 、r ₂ ……r _k-1 、r _k ……/>If the clustering threshold is r _k-1 ≥λ＞r _k The similarity coefficient is greater than or equal to r _k-1 Is grouped into one class, and the similarity coefficient is smaller than r _k-1 Is self-organizing and lambda is in interval r _k-1 ,r _k ) The clustering result is not changed at any value. Therefore, the similarity coefficients in the fuzzy equivalent matrix can be directly ordered from big to small and sequentially takenAnd clustering respectively until a clustering result meeting the requirement is obtained.

Based on the design principle, the learning time of the self-learning algorithm can be greatly reduced, and the applicability of the learning result is improved. The flow of the self-learning algorithm designed by the invention is shown in figure 2. Firstly, setting the number of components in a tested piece as N, then loading a self-learning signal, carrying out pulse extraction on the self-learning signal, and calculating a fuzzy equivalent matrix. Sequencing all similar coefficients in the fuzzy equivalent matrix from big to small, and sequentially takingAnd recording the current clustering threshold lambda until the pulse number duty ratio in the N classes is greater than 95%, and selecting the smallest clustering threshold as a learning result from all the recorded clustering thresholds lambda after all the self-learning signals are completely learned.

2.2 Signal spectral clustering algorithm

After the pulses in the PIND detection signal are extracted by a two-threshold method, short-time Fourier transform is performed on the pulse sequence in order to obtain the frequency spectrum of the pulses. Since the pulse length is not fixed, each pulse needs to be pulse-normalized to meet the requirements. The pulse regulation method adopted by the invention is a zero filling method, namely zero is filled in the front and the back of the pulse data to be converted by equal amount, so that the length of the pulse data meets the requirement and the original information is not lost. After the spectrums of all the pulses are obtained through pulse normalization and short-time Fourier transformation, in order to express the similarity degree of two pulses, corresponding similarity evaluation is carried out on all the pulses, and the similarity coefficient is obtained. The calculation formula of the similarity coefficient is shown in (1).

Wherein: r is (r) _ij Representing the degree of similarity of pulse i to pulse j, F _i [k]Representing the spectral sequence of pulse i, F _j [k]Represents the spectral sequence of pulse j, N represents the data length of the short-time fourier transform, n=2 ⁿ (n is a positive integer).

The similarity coefficient of all pulses in the PIND detection signal can be calculated by the formula (1) and a fuzzy similarity matrix R is obtained _ij . Then, R is obtained by a flattening method through transfer closing operation _ij Is a transfer closure t (R) _ij ) I.e. fuzzy equivalence matrixBy the set clustering threshold lambda, if r _ij And (3) setting the pulse i and the pulse j to be in a class under the truncated level lambda. Let j _ij Indicating whether the pulses i and j are gathered into one type, when r _ij Not less than lambda, j _ij ＝1；r _ij < lambda, j _ij =0. Comparing all r _ij Obtaining a truncated matrix J _ij And obtaining a clustering result according to the truncated matrix analysis, as shown in fig. 3. The key parameter of pulse clustering is the clustering threshold lambda. In the fuzzy equivalence matrix, when r _ij At > λ, pulses i and j are grouped into one class at truncated level λ; when r is _ij When lambda is less than or equal to lambda, the pulse i and the pulse j cannot be gathered into one type under the truncated level lambda. When pulse clustering is completed, component signals can be well clustered into one type. If (lambda) _k ) _max ≤(r _ij ) _min Wherein lambda is _k Representing a truncated horizontal dataset, then all pulses of the redundancy detection signal are grouped into one class; if (lambda) _k ) _min ≥(r _ij ) _max All pulses of the unwanted detection signal are of one type each, which are two extreme cases. Therefore, too large or too small a lambda value has a large influence on the clustering result.

2.3 Signal Classification Standard design

Since the redundancy signal pulses occur randomly with generally large occurrence intervals, the total length of the redundancy signal pulses has a small duty cycle in the PIND detection signal for a PIND detection signal containing redundancy pulses. According to the statistical analysis of PIND detection results in recent years, the invention discovers that the total time length of most redundant signal pulses in PIND detection signals with the total time length of 5s is less than 5%. In PIND detection signals containing component signal pulses, most component pulses occur periodically, with a total duration of 3% -20%. When the total time length of the component signal pulse is more than 10%, the component with larger activity space exists in the tested piece, the excessive object signal pulse is easily covered by the large-amplitude vibration, and the state is called an oversized component signal, which is a special condition of the component signal pulse. And for the case that the pulse length ratio is within the interval (0.5 percent, 10 percent), the pulse clustering result is combined to further distinguish the signal components contained in the pulse clustering result.

Because the component signal pulses have extremely high similarity in pulse length, pulse energy and pulse amplitude, the component signal pulses can be well clustered into one class, and the statistical analysis finds that the total class number of most component signal pulses after clustering is less than 3 classes. The redundant signal pulses have large differences in pulse length, pulse energy and pulse amplitude, and are difficult to gather into one class, so the total class number after clustering is usually more than 3 classes. In addition, the statistical analysis also shows that the clustering result of the mixed signal pulse should contain a major class and a plurality of minor classes, the total number of the classes is more than 15, the judgment standard is given by combining the pulse length duty factor P and the number M of the classes after pulse clustering, and the classification flow chart is shown in figure 4. The specific classification steps are as follows: (1) if P is less than or equal to 0.5%, judging that the signal is an excessive signal; (2) If P is more than 0.5% and less than or equal to 5% and M is more than 3, judging that the signal is an excessive signal; (3) If P is more than or equal to 0.5% and less than or equal to 10% and M is more than or equal to 3, judging that the signal is a movable component signal; (4) If P is more than 5% and less than or equal to 10%, M is more than 3 and less than or equal to 15, judging that manual checking is needed; (5) If P is more than 5% and less than or equal to 10% and less than 15% and less than M, judging the mixed signal; (6) If the ratio is less than 10% < P, judging that the signal of the movable assembly is too large;

3. method 2: signal classification design of k-NN algorithm based on parameter optimization

In order to more accurately identify the types of PIND detection signals, the characteristics of signal pulses of different types in the PIND detection signals are researched and analyzed, and a k-NN algorithm-based redundant detection signal classification algorithm is designed. The algorithm establishes a multidimensional feature space set by selecting features capable of representing the pulse characteristics of the redundant signals. And searching k training sample data which are closest to the characteristics of the test sample in the characteristic space set, and obtaining the signal characteristic type of the test sample according to the existing classification labels of the training sample. The redundant signal classification method based on the k-NN algorithm is mainly designed as follows:

step one: and respectively extracting the characteristics of the training sample and the test sample to generate a training sample characteristic set and a test sample characteristic set.

Step two: adding known class labels for the training samples, calculating feature distances between the feature sets of the test samples and the feature vectors in the feature sets of the training samples, and selecting k pieces of training sample data closest to the feature vectors of the test samples.

Step three: and analyzing the classification result of the test sample according to the selected class label of the training sample.

3.1 selection of Signal Properties and determination of feature distance

The unwanted material signal is generated by collision of the unwanted material particles which have been activated with the internal components of the sealed electronic component or the inner wall of the sealed housing. Fig. 5 is a schematic diagram of a typical unwanted signal pulse train and a single pulse waveform. The retentate signal is mainly represented as a random spike sequence. The single pulse has a unilateral oscillation attenuation trend, namely the initial amplitude of the pulse rises fast, and the pulse begins to decay rapidly after reaching a certain peak value. The component signal is a signal generated by forced vibration of the components in the sealed electronic component. Fig. 6 is a schematic diagram of a typical component signal pulse train and a single pulse waveform. The component signal is mainly represented by a periodic spike pulse sequence, which must have a certain starting process to activate the component, and when the external vibration disappears, it takes a certain time to return to the static state.

The component signal has the characteristic that the pulse sequence has periodicity in its entirety but locally has periodicity instability, for example, the pulse sequence occurrence moments as shown in fig. 1 have periodicity in its entirety but the local time interval at 4 is significantly smaller than the other time intervals.

The unwanted signal is an acoustic emission signal that broadly pertains to. Classification of acoustic emission signals requires investigation from the characterization of the signals. For signal pulse classification, the selected features need to be able to characterize the differences of the unwanted signal pulses from the component signal pulses, and the more pronounced the better. In view of the above requirements and by comprehensively considering the overall difference degree of the features, the invention mainly performs feature extraction from four aspects of frequency domain features, time domain features, energy features and distribution features: pulse area(s), pulse rise percentage (Tp), pulse bilateral symmetry (dczy), pulse up-down symmetry (dcsx), duration (Tl), energy density (MD), pulse duty cycle (ZB), crest factor (bf), zero crossing rate (zerorate), variance (var), area duty cycle (dp), spectral centroid (mainHz), cepstral coefficient (MSF), cepstral coefficient difference (MSFcha). I.e. a 14-dimensional feature data set is generated as the feature basis for classification by the k-NN algorithm.

The invention selects two methods, namely chi-square test and decision tree model-based method to screen characteristic parameters. The purpose of the chi-square test is to test the correlation of qualitative independent variables to qualitative dependent variables, i.e. the degree of correlation of classification features and classification targets. The larger the chi-square test value, the stronger the correlation. The feature selection method based on the decision tree model aims at calculating the importance degree of each feature to the model, and the larger the parameter value is, the higher the importance degree is represented. The test results are shown in FIG. 7.

The more pronounced the selected feature is able to characterize the difference of the unwanted signal pulses from the component signal pulses for a single signal pulse that needs to be classified. In view of the above requirements, the overall difference degree of the features is considered, and the first 5 features of the chi-square test method and the first 8 features of the decision tree model method are considered, and finally the cepstrum coefficient difference (MSFcha), the crest factor (bf), the area occupation ratio (dp), the zero crossing rate (zerorate), the energy density (MD), the variance (var), the pulse area(s) and the up-down pulse symmetry (dcsx) are selected, wherein the total 8 features are taken as the basis of classification, and the specific details are shown in table 1.

TABLE 1 characteristic formulas table

In consideration of the difference of pulse feature vectors of the comprehensive measurement PIND signals, the invention selects a common Euclidean distance algorithm and cosine distance algorithm, but the data needs to be normalized due to the large numerical difference among the feature data. I.e. by comparing each feature value with the maximum value in the feature data set, a new value is obtained which can replace the original feature. The new data characteristic value is between [0,1], and the calculation formula is as follows (2):

wherein x is _i Is the normalized value, t _i Values before normalization.

System comparison tests were performed for both distance algorithms. 1000 sets of redundant signals and component signals are taken as test samples respectively, k value is randomly selected to be 10, and a rule of minority compliance (namely a majority voting method) is taken as a judgment standard for calculation. As is apparent from the results shown in table 2, the accuracy of recognition of the redundant signal by the euclidean distance algorithm is 91.90%, the accuracy of recognition of the component signal is 89.70%, and the average accuracy is 90.80%; the accuracy of the cosine distance algorithm for identifying the redundant signals is 92.40%, the accuracy of identifying the component signals is 92.10%, and the average accuracy is 92.30%

Table 2 comparison of results of two distance algorithms

Therefore, although the two kinds of discrimination criteria can realize more accurate classification of the two kinds of signal feature vectors, the classification effect by taking the cosine distance algorithm as the discrimination criteria is better. In fact, the euclidean distance algorithm represents an absolute difference in value. The cosine distance algorithm reflects the relative difference in the direction, weakens the influence of a single feature on the whole, and is more suitable for classification of the invention. Therefore, the invention selects the basis method using cosine distance algorithm as the identification distance.

3.2k value determination

For the algorithm of classifying by the central clustering method, the key problem is to determine the k value so as to ensure that the k value can accurately represent the type of the test signal. If the given k value is smaller, the whole model is more complex, and the fitting is easy to happen in the estimation process; if the given k value is large, the model is simple as a whole, and a large amount of useful feature information in training sample features is easy to ignore. Thus, during the design process, the k value is typically tested specifically with specific sample data and evaluated for determination.

The method comprises the steps of randomly dividing collected redundant signal pulses and component signal pulses into two groups of training sets and testing sets, wherein the ratio of the data quantity of the training sets to the data quantity of the testing sets is 2:1, then establishing a model for training detection, and finally selecting k values by taking classification accuracy as a basis. For a certain type of sealed relay, 400 redundant signals and 200 component signals are taken as training sets, the data results are shown in table 3, and the relation between the test accuracy and k value is shown in fig. 8. It is obvious that the effect is better when the k value is taken to be 20.

TABLE 3 test accuracy at different k values

4. Experimental verification and analysis

4.1 Signal classification experiment verification based on spectral clustering method

Cluster analysis is now performed on 3745 sets of known unwanted signal samples, 1695 sets of known component signal samples, 575 sets of known mixed signal samples, and 725 sets of known oversized component signal samples, summing up 6740 sets of data. The recognition results are shown in Table 4. Wherein, the redundant signal correctly identifies 3440 groups with the ratio of 91.86 percent, 85 groups need to be checked manually, and the ratio is 2.27 percent; component signals correctly identify 1420 groups, account for 83.76 percent, 245 groups need to be checked manually, and account for 14.45 percent; the mixed signal correctly identifies 290 groups, and the proportion is 50.43%; the 135 groups need to be manually checked, and the ratio is 23.47%. And the signals of the oversized assembly are correctly identified into 690 groups, the identification rate is 95.17%, 5 groups are required to be checked manually, and the occupied ratio is 0.69%. Therefore, the recognition accuracy of the spectral clustering algorithm to the redundant signals and the oversized component signals is higher, but the recognition accuracy of the mixed signals is lower, and meanwhile, a large number of situations needing to be checked manually exist in the recognition result.

Table 4 detection signal recognition results

The oversized component signal is an oversized signal generated by a sealing component containing the component during vibration. (2) In the invention, when the signal type is defined as manual checking, the signal type is tentatively defined as classification error. (3) ^△ The mixing signal, i.e. the unwanted signal, is mixed with the component signal.

4.2 Signal classification experiment verification of k-NN method based on parameter optimization

The results of signal classification by the k-NN algorithm using the same samples as those of 4.1 are shown in Table 5. The redundancy signal correctly identifies 3450 groups, accounting for 92.12%. The component signal correctly identifies 1526 groups, accounting for 90.03%. The mixed signal correctly identifies 466 group, accounting for 81.04%. Because the algorithm is complex in calculation process, the overall recognition speed is slower compared with the method 1. Because the pulse extraction in the oversized component signal is difficult, and the k-NN algorithm is considered to have strong dependence on the characteristics, the oversized component signal is not judged.

Table 5 detection signal recognition results

4.3 Signal Classification experiment verification based on synchronous algorithm

According to practical experience, the basis of selecting the method 1 and the method 2 is accuracy and calculation time, and the invention follows the following principles: (1) If the accuracy increases ar=n (0 < n < 5%), the calculation time needs to be reduced by ct=10n, at least by an order of magnitude, where the calculation time may be considered. (2) If the accuracy increases ar=n (5% +.n), the calculation time needs to be reduced by ct=50n, where the calculation time may be considered. Clearly, the accuracy is of higher priority. In summary, 4.1 and 4.2 were compared and analyzed as shown in FIG. 9. When the recognition result of the input signal is the redundant signal, AR/ct=0.0039 < 0.1, and method 1 is selected. When the input signal judgment result is the component signal, AR/CT=0.085 > 0.02, and selecting method 2. When the input signal judgment result is a mixed signal, AR/CT=0.56 > 0.02, and method 2 is selected. And when the input signal judgment result is an oversized component signal, selecting the method 1.

To sum up, table 6 shows the calculation results of the synchronous algorithm adopted in the present invention: the method 1 is used to classify the redundant signals and the oversized component signals, and the considered factors are calculation time. For the classification of component signals and mixed signals using method 2, the detection accuracy is considered as a factor. Further, the original signal data only needs to be extracted once, and the extracted data is input to the method 1 and the method 2 simultaneously. The method 1 only calculates the recognition effect of the redundant particles and the component signals, and the method 2 only calculates the recognition effect of the component signals and the mixed signals, so that compared with a single algorithm, the synchronous method comprehensively considers the calculation efficiency and the accuracy, and the advantages of the two methods are complemented.

Table 6 detection signal discrimination results of synchronous algorithm

Annotate (1) ^* 1 represents a self-learning signal spectrum clustering algorithm, and 2 represents a k-NN algorithm.

Claims (5)

1. A synchronous classification method for redundant detection signals of spaceflight sealed electronic components is characterized by comprising the following steps:

step one, data processing

Constructing a redundancy detection platform, and performing data processing on the acquired pulse data to acquire the frequency domain characteristics, the time domain characteristics, the energy characteristics and the distribution characteristics of the pulse;

step two, signal classification design based on spectral clustering algorithm

Step two, taking the pulse frequency domain characteristic data obtained in the step one as an input item, performing pulse similarity calculation to obtain similarity coefficients of all pulses in the PIND detection signal, and obtaining a fuzzy similarity matrix;

step two, converting the fuzzy similar matrix into a fuzzy equivalent matrix through transfer closed-packet operation;

step two, determining a clustering threshold lambda through a self-learning clustering algorithm;

step two, comparing through traversing the fuzzy equivalent matrix elements to obtain a clustering result M;

step five, judging and classifying the redundant signals, the component signals, the mixed signals and the oversized component signals according to the clustering result M and the pulse duty ratio P and the signal classification standard;

step three, signal classification design of k-NN algorithm based on parameter optimization

Step three, taking the pulse data obtained in the step one as an input item, and obtaining the processed data through short-time Fourier transform and pulse normalization processing;

step three, carrying out feature extraction on the data processed in the step three from four aspects of frequency domain features, time domain features, energy features and distribution features, selecting chi-square test and screening feature parameters based on a decision tree model, and obtaining a multi-dimensional feature set as a feature basis of k-NN algorithm classification, wherein the multi-dimensional feature set comprises the following 8 features: cepstrum coefficient difference, crest coefficient, area occupation ratio, zero crossing rate, energy density, variance, pulse area, and upper and lower pulse symmetry;

thirdly, adopting a cosine distance method as a space distance calculation method through a majority compliance method;

thirdly, calculating through testing accuracy under different k values, and determining the k value;

step three, training and testing pulse data through a KNN algorithm with optimized parameters, and obtaining classification results of different pulses;

and fourthly, comparing and selecting corresponding classification results according to the following rules to realize synchronous classification:

classifying the redundant signals and the oversized component signals by using the method of the second step;

for the component signals and the mixed signals, classification is performed by using the method of the third step.

2. The method for synchronously classifying the redundant detection signals of the aerospace-sealed electronic components according to claim 1, wherein in the step two, a calculation formula of the similarity coefficient is as follows:

wherein: r is (r) _ij Representing the degree of similarity of pulse i to pulse j, F _i [k]Representing the spectral sequence of pulse i, F _j [k]Represents the spectral sequence of pulse j, N represents the data length of the short-time fourier transform, n=2 ⁿ N is a positive integer.

3. The method for synchronously classifying the redundant detection signals of the aerospace sealed electronic components according to claim 1, wherein the specific steps of the second step are as follows:

setting the number of components in a tested piece as N, loading a self-learning signal, carrying out pulse extraction on the self-learning signal, calculating a fuzzy equivalent matrix, sequencing all similar coefficients in the fuzzy equivalent matrix from large to small, and sequentially takingAnd recording the current clustering threshold lambda until the pulse number duty ratio in the N classes is greater than 95%, and selecting the smallest clustering threshold as a learning result from all the recorded clustering thresholds lambda after all the self-learning signals are completely learned.

4. The method for synchronously classifying the redundant detection signals of the aerospace sealed electronic components according to claim 1, wherein the specific steps of the second step are as follows:

in the fuzzy equivalence matrix, when r _ij At > λ, pulses i and j are grouped into one class at truncated level λ; when r is _ij When the lambda is less than or equal to lambda, the pulse i and the pulse j cannot be gathered into one type under the truncated level lambda; if (lambda) _k ) _max ≤(r _ij ) _min Wherein lambda is _k Representing a truncated horizontal dataset, then all pulses of the redundancy detection signal are grouped into one class; if (lambda) _min ≥(r _ij ) _max All pulses of the redundancy detection signal are each of one type.

5. The method for synchronously classifying the redundant detection signals of the aerospace-sealed electronic components according to claim 1, wherein the specific classification steps in the second five steps are as follows: (1) if P is less than or equal to 0.5%, judging that the signal is an excessive signal; (2) If P is more than 0.5% and less than or equal to 5% and M is more than 3, judging that the signal is an excessive signal; (3) If P is more than or equal to 0.5% and less than or equal to 10% and M is more than or equal to 3, judging that the signal is a movable component signal; (4) If P is more than 5% and less than or equal to 10%, M is more than 3 and less than or equal to 15, judging that manual checking is needed; (5) If P is more than 5% and less than or equal to 10% and less than 15% and less than M, judging the mixed signal; (6) If 10% < P, it is determined that the movable element signal is too large.