Fusion feature vector construction mode determination and construction method for sealing electronic equipment redundancy positioning and redundancy positioning method

Abstract

The invention discloses a fusion feature vector construction mode determination and construction method for sealing electronic equipment redundancy positioning and a redundancy positioning method, and belongs to the technical field of sealing electronic equipment redundancy detection. The objective of the invention is to solve the problem that the classification effect of sealing electronic equipment redundancy positioning is limited by the obtained feature vector due to the fact that the existing signal for sealing electronic equipment redundancy positioning is constructed by taking a sound sensor as a unit. According to the method, channels and features are respectively used as basic construction units, and are arranged according to a channel and feature numbering sequence or a random disordered row; randomly disrupting features and channel numbers of each basic construction unit, or randomly disrupting features and channel numbers of each basic construction unit and arranging features of feature vectors according to an incompletely identical rule to obtain a fusion feature vector; screening based on the positioning identification accuracy to obtain a fusion feature vector construction mode; and feature vectors are constructed based on a feature vector construction method and are used for redundant object positioning.

Description

Technical Field

[0001] The present invention belongs to the technical field of redundant detection of sealed electronic devices, and particularly relates to a method for determining a construction method of a fusion feature vector for positioning redundant substances in a sealed electronic device, a construction method of the fusion feature vector, and a method for positioning redundant substances in a sealed electronic device. Background Art

[0002] Sealed electronic devices are core components of aerospace equipment, and their reliability directly determines the safe operation of aerospace equipment. Redundant substances are substances such as metal debris, solder balls, and wire segments generated during the production of sealed electronic devices, which damage the stable physical state inside the sealed electronic devices and pose an important threat to the reliability of the sealed electronic devices. Therefore, it is crucial to detect redundant substances before the sealed electronic devices leave the factory for service. The particle impact noise detection (PIND) method is currently the internationally recognized redundant substance detection method, and its principle based on vibration and sound generation is briefly described as follows. A sound sensor is placed on the surface of the test piece and a mechanical excitation is applied to activate the possible redundant substances to make them in a random motion state of collision or sliding, generating a sound signal, which is called a redundant substance signal. By judging whether the sound sensor captures the redundant substance signal, it is judged whether there are redundant substances inside the test piece.

[0003] For sealed electronic devices, when redundant substances are detected, the device must be opened to search for and clean the redundant substances, and the device must be repackaged for redundant substance detection until no more redundant substances are detected. However, sealed electronic devices are relatively large in size, and the aimless redundant substance cleaning work causes relatively large resource consumption. Therefore, obtaining the location of the redundant substances has important reference significance for the redundant substance cleaning work of sealed electronic devices. Redundant substance positioning has received extensive attention. In machine learning, different categories of objects have attribute differences, which can be quantified as distribution differences of data with different labels. In this way, by establishing a data set to train a classifier, data classification can be achieved, and then classification of different categories of objects can be achieved. Similarly, within a spatial range, the sound signals generated at different positions have characteristic differences, and the above idea can be followed to establish a data set and train a classifier, thereby achieving sound source positioning. Some scholars have carried out relevant research.

[0004] For example, Xiong et al. extracted nine typical time-frequency domain features from the minute vibration signals generated in four regions of the full-field target, established a dataset to train multiple classical classifiers, and achieved sound source localization of the full-field target, thereby verifying the feasibility of the data-driven sound source localization scheme. Tang et al. extracted multiple classical time-domain features from the sound signals generated at different positions in a plate-like structure, established a dataset to train a BP neural network, and achieved localization of the impact source inside the plate-like structure. Ojha et al. processed the sound signals generated at different positions in a CFRP composite plate using wavelet scattering transform, extracted wavelet scattering transform features to establish a dataset and train a probabilistic machine learning model, and achieved localization of the impact source inside the CFRP composite plate. Liang et al. extracted multiple time-domain features from the sound signals generated at different positions on the surface and inside a rock, respectively determined the feature combinations applicable to planar localization and spatial localization through feature engineering, and thereby established datasets to train and obtain the most suitable classifiers, a support vector machine and a gradient boosting regression model, to achieve fracture source localization on the surface and inside the rock. Nguyen-Vu and Lee extracted multiple multi-domain features from the sound signals generated at different positions in an indoor space, determined multiple applicable multi-domain features through feature engineering, and thereby established a dataset and trained multiple classical classifiers to achieve sound source localization in the indoor space.

[0005] Broadly speaking, foreign objects are regarded as sound sources, and foreign object signals are sound signals. The above-mentioned machine learning-based sound source localization research provides an important reference for foreign object localization work. Based on this, in previous research, Sun Zhigang et al. extracted multiple time-domain, frequency-domain, and time-frequency domain features from foreign object signals generated in different enclosed spaces inside sealed electronic devices, determined applicable feature combinations through feature engineering to establish a dataset, and continued to carry out incomplete data processing research to obtain a high-quality dataset to train multiple classifiers, and finally determined the random forest with the best performance. It should be noted that the sealed electronic device has a large geometric volume and a complex internal structure. When performing foreign object detection, it is usually necessary to place sound sensors at different positions on the surface of the sealed electronic device respectively to ensure that the combined detection range of multiple sound sensors covers the entire internal space of the sealed electronic device. In this way, when there are foreign objects inside the sealed electronic device, during foreign object detection, within the same detection time, multiple sound sensors will simultaneously capture a segment of foreign object signal, obtaining multiple segments of foreign object signals. At this time, when further using machine learning to quantify foreign object signals, construct feature vectors, and predict the position of foreign objects, it will be found that machine learning faces multiple segments of foreign object signals simultaneously. Usually, through machine learning, a segment of foreign object signal can be quantified into a feature vector, and then a prediction result is given by the classifier. In this way, it seems that multiple segments of foreign object signals can be quantified into multiple feature vectors, and then multiple prediction results are given by the classifier, but this is incorrect. On the one hand, one foreign object detection requires one prediction result, not multiple prediction results. Even if multiple prediction results are processed subsequently to give one prediction result, the reliability of the prediction result is reduced. On the other hand, multiple segments of foreign object signals represent sound sensors at different positions respectively, and there is a high degree of correlation and complementarity between them, such as phase complementarity. Therefore, multiple segments of foreign object signals cannot be arbitrarily processed as individuals, but should be considered as a whole. Thus, it is necessary to consider the multiple feature vectors quantified from multiple segments of foreign object signals as a whole and construct a fused feature vector. Summary of the Invention

[0006] The present invention aims to solve the problem that the existing signals for foreign object localization in sealed electronic devices are constructed based on sound sensors, and the obtained feature vectors limit the classification effect of foreign object localization in sealed electronic devices, and lack interpretability.

[0007] A method for determining a construction method of a fused feature vector for foreign object localization in a sealed electronic device, the method comprising:

[0008] For the sound signals for foreign object localization in a sealed electronic device collected by M channels, N features are obtained based on the sound signals corresponding to each channel;

[0009] Using channels as basic building blocks, arrange M basic building blocks in the order of channel numbers or randomly shuffle them; and randomly shuffle the N features of the feature vector for each basic building block by feature numbers, or randomly shuffle the N features of the feature vector by feature numbers and arrange the N features of the M feature vectors according to different rules to obtain the fused feature vector;

[0010] Using features as basic building blocks, arrange N basic building blocks in the order of feature numbers or randomly shuffle them; and randomly shuffle the M features of the feature vector for each basic building block by channel numbers, or randomly shuffle the M features of the feature vector by channel numbers and arrange the M features of the N feature vectors according to different rules to obtain the fused feature vector;

[0011] For the fused feature vector obtained with channels as basic building blocks and the fused feature vector obtained with features as basic building blocks, use a classifier to perform positioning and identification respectively, and screen the construction methods of the fused feature vectors based on the positioning and identification accuracy rates corresponding to the fused feature vectors;

[0012] The obtained construction method of the fused feature vector is used as the construction method of the fused feature vector for positioning the redundant objects in the sealed electronic device.

[0013] A method for constructing a fused feature vector for positioning redundant objects in a sealed electronic device, including:

[0014] For the sound signals for positioning redundant objects in the sealed electronic device collected from M channels, obtain N features based on the sound signals corresponding to each channel;

[0015] Using features as basic building blocks, arrange the N basic building blocks according to feature numbers, and for each basic building block, randomly shuffle the M features of the feature vector by channel numbers to construct a fused feature vector; or, using features as basic building blocks, randomly shuffle the N basic building blocks by feature numbers, and for each basic building block, randomly shuffle the M features of the feature vector by channel numbers to construct a fused feature vector.

[0016] A method for positioning redundant objects in a sealed electronic device, uses the method for constructing a fused feature vector for positioning redundant objects in a sealed electronic device described above to construct a fused feature vector, and for the fused feature vector, uses a classifier to perform positioning and identification of redundant objects in the sealed electronic device to achieve positioning.

[0017] Beneficial effects: The present invention for the first time considers carrying out the research on the positioning of foreign objects in sealed electronic devices starting from the fused feature vectors, and proposes various methods for constructing fused feature vectors applicable to the positioning of foreign objects in sealed electronic devices. Experiments prove that the optimal feature change-ordered fused feature vector construction method and the random forest not only provide an interpretable basis for the research on the positioning of foreign objects in sealed electronic devices based on machine learning, but also have good classification effects for the positioning of foreign objects. Description of the Drawings

[0018] Figure 1 Schematic diagram of the fused feature vector construction method used in the existing research.

[0019] Figure 2 Schematic diagram of the channel number-ordered fused feature vector construction method.

[0020] Figure 3 Schematic diagram of the channel change-ordered fused feature vector construction method.

[0021] Figure 4 Schematic diagram of the channel number-unordered fused feature vector construction method.

[0022] Figure 5 Schematic diagram of the channel change-unordered fused feature vector construction method.

[0023] Figure 6 Schematic diagram of the feature basis-fused feature vector construction method.

[0024] Figure 7 Schematic diagram of the feature number-ordered fused feature vector construction method.

[0025] Figure 8 Schematic diagram of the feature change-ordered fused feature vector construction method.

[0026] Figure 9 Schematic diagram of the feature number-unordered fused feature vector construction method.

[0027] Figure 10 Schematic diagram of the feature change-unordered fused feature vector construction method.

[0028] Figure 11 Schematic diagram of the aerospace electronic unit in kind.

[0029] Figure 12 Spatial layout diagram of the foreign object automatic detection device and the sound sensor, where (a) is the foreign object automatic detection device and (b) is the spatial layout of the sound sensor.

[0030] Figure 13 F1-scores obtained by five classifiers on the channel-basic dataset.

[0031] Figure 14 The F1-scores obtained by five classifiers on the feature-basic dataset.

[0032] Figure 15 The average F1-score obtained by five classifiers on the channel number-ordered dataset.

[0033] Figure 16 The average F1-score obtained by five classifiers on the channel change-ordered dataset.

[0034] Figure 17 The average F1-score obtained by five classifiers on the channel number-unordered dataset.

[0035] Figure 18 The average F1-score obtained by five classifiers on the channel change-unordered dataset.

[0036] Figure 19 The average F1-score obtained by five classifiers on the feature number-ordered dataset.

[0037] Figure 20 The average F1-score obtained by five classifiers on the feature change-ordered dataset.

[0038] Figure 21 The average F1-score obtained by five classifiers on the feature number-unordered dataset.

[0039] Figure 22 The average F1-score obtained by five classifiers on the feature change-unordered dataset. Detailed implementation manners

[0040] In previous studies, the inventors numbered the sound sensors and features respectively. On this basis, for each piece of redundant signal, the numerical values of multi-domain features were calculated respectively, and according to the numbers of the features, the multiple calculated numerical values were arranged in a row to construct a feature vector. Multiple pieces of redundant signals constructed multiple feature vectors. For the multiple feature vectors, according to the numbers of the corresponding sound sensors, they were arranged in a row to construct a fused feature vector. After that, no relevant research has focused on the method for constructing the fused feature vector, and the above-mentioned processing process has been used as an established rule. In fact, the numbers of the sound sensors and features are both artificially set, and the designed method for constructing the fused feature vector lacks relatively theoretical basis and is regarded as a temporary strategy. Moreover, the numbering rules of the sound sensors and features are also adjustable. Whether different numbering rules have an impact on the method for constructing the fused feature vector, and whether the previously used method for constructing the fused feature vector is optimal in this case, all these need to be explored in depth. In addition, as an intermediate process between the redundant signal and the data set, the feature vector not only represents the interpretability of the input signal, but also affects the quality of the output data set and even the classification performance of the subsequent trained classifier. Therefore, the method for constructing the fused feature vector is crucial for the research on redundant object localization. Unfortunately, in previous studies, scholars including the inventors did not pay attention to this point. It is worth mentioning that in the research on the recognition of multi-channel signals in related fields, they all considered converting multi-channel signals into multiple pictures and using advanced neural networks for processing. However, compared with "extracting features from the signal", more information is lost in "converting the signal into an image first and then extracting features from the image". Obviously, in principle, constructing the fused feature vector has advantages.

[0041] Therefore, in the present invention, for the first time, the research on the positioning of foreign objects in sealed electronic devices is considered starting from the fused feature vector, and thus a method for constructing a fused feature vector is proposed. The sound sensors and features are regarded as reference objects, called channels and features, and channel numbers and feature numbers are set. On this basis, the method for constructing a fused feature vector used in existing research is referred to. On the one hand, taking the channel as the basic construction unit, the methods for constructing the fused feature vector considering only the channel number, only the feature number, and both the channel number and the feature number under the ordered rule and the disordered rule are respectively explored. On the other hand, taking the feature as the basic construction unit, the methods for constructing the fused feature vector considering only the channel number, only the feature number, and both the channel number and the feature number under the ordered rule and the disordered rule are respectively explored. Finally, taking the typical sealed electronic device in the aerospace field as the experimental object to conduct foreign object detection, a large number of representative data sets are respectively established with reference to the above-mentioned methods for constructing the fused feature vector, and a number of representative classifiers are respectively trained. The variation law of the classification performance of different types of classifiers on the data sets established by different methods for constructing the fused feature vector is analyzed, and the optimal method for constructing the fused feature vector suitable for the positioning of foreign objects in sealed electronic devices and the optimal classifier selection suggestions are given.

[0042] The present invention will be described in detail below in conjunction with the specific embodiments.

[0043] This embodiment proposes a method for determining a method for constructing a fused feature vector for positioning foreign objects in a sealed electronic device, a method for constructing a fused feature vector, and a method for positioning foreign objects in a sealed electronic device using the feature vector constructed by the method for constructing a fused feature vector. This embodiment first determines the optimal method for constructing a feature vector through a method for determining a method for constructing a fused feature vector for positioning foreign objects in a sealed electronic device, which is used to construct a fused feature vector for positioning foreign objects in a sealed electronic device; the constructed feature vector is finally used for positioning foreign objects in a sealed electronic device.

[0044] In the positioning of foreign objects in a sealed electronic device, it is assumed that M sound sensors need to be placed on the surface of the sealed electronic device to ensure that the joint detection range of the sound sensors covers all the internal spaces of the sealed electronic device. The M sound sensors are respectively numbered, that is: the 1st sound sensor, the 2nd sound sensor,..., the Mth sound sensor. The numbers are called sound sensor numbers. During the same detection time, the M sound sensors will simultaneously capture a segment of foreign object signal, a total of M segments of foreign object signals. For each segment of foreign object signal, it is assumed that the values of N features in different domains need to be calculated to realize the quantization of the foreign object signal into a feature vector, and N values are obtained. The N features are respectively numbered, that is: the 1st feature, the 2nd feature,..., the Nth feature. The numbers are called feature numbers. According to the numbers of the features, the corresponding N values are arranged in a row to construct a feature vector.

[0045] In the method for constructing a fusion feature vector used in existing research, first, M segments of foreign object signals collected by M sound sensors are respectively quantized into M feature vectors. Then, according to the numbers of the sound sensors, the corresponding M feature vectors are arranged in a row to construct an M×N type fusion feature vector. The principle is as Figure 1 shown.

[0046] It is worth noting that if the sound sensor is abbreviated as a channel, then the number of the sound sensor can be called the channel number. In this way, the 1st sound sensor can be called Channel 1, the 2nd sound sensor can be called Channel 2,..., and the Mth sound sensor can be called Channel M. Further, the foreign object signal collected by the 1st sound sensor can be called the foreign object signal of Channel 1, the foreign object signal collected by the 2nd sound sensor can be called the foreign object signal of Channel 2,..., and the foreign object signal collected by the Mth sound sensor can be called the foreign object signal of Channel M. At the same time, if Figure 1 the components included in the rectangular box in

[0047] are called the basic construction units of the fusion feature vector, then the method for constructing a fusion feature vector used in existing research is based on channels as the basic construction units, and each basic construction unit contains a feature vector constructed with a reference feature number. Considering that the internal values of these feature vectors are arranged according to the same rule (i.e., the same feature number), the fusion feature vector constructed thereby is said to satisfy the ordered rule. On the contrary, if the internal values of the feature vectors included in each basic construction unit are arranged according to rules that are not completely the same, the fusion feature vector constructed thereby is said to satisfy the disordered rule. At the same time, the method for constructing a fusion feature vector used in existing research is called a channel-based - fusion feature vector construction method with channels as the basic construction units.

[0048] I. Basic construction unit - channel:

[0049] Explore the construction method of a fusion feature vector with channels as the basic building blocks. It consists of two parts. The first part is the construction method of a fusion feature vector that satisfies the ordered rule, and the second part is the construction method of a fusion feature vector that satisfies the unordered rule.

[0050] In the construction method of a fusion feature vector that satisfies the ordered rule, in the first step, arrange the M basic building blocks according to the Figure 1 channel numbers shown. On this basis, adjust the construction process of the feature vectors contained in each basic building block. Instead of arranging the N numerical values of the feature vectors according to the Figure 1 feature numbers shown, randomly shuffle the feature numbers to arrange the N numerical values of the feature vectors. In this way, the N numerical values of the M feature vectors are arranged according to the same rule to construct a fusion feature vector. This construction method of the fusion feature vector is called the channel number-ordered fusion feature vector construction method. According to the knowledge of permutations and combinations, there are N! permutation methods for N numerical values. In this way, the channel number-ordered fusion feature vector construction method contains N! specific cases. Figure 2 Shows three cases of numerical permutations among them.

[0051] In the construction method of a fusion feature vector that satisfies the ordered rule, in the second step, instead of arranging according to the Figure 1 channel numbers shown, randomly shuffle the channel numbers to arrange the M basic building blocks. On this basis, similar to the channel number-ordered fusion feature vector construction method, for the feature vectors contained in each basic building block, randomly shuffle the feature numbers to arrange the N numerical values of the feature vectors. In this way, the N numerical values of the M feature vectors are arranged according to the same rule to construct a fusion feature vector. This construction method of the fusion feature vector is called the channel variation-ordered fusion feature vector construction method. As mentioned before, there are N! permutation methods for N numerical values. Similarly, there are M! permutation methods for M channels. In this way, the channel variation-ordered fusion feature vector construction method contains M! × N! specific cases. Figure 3 Shows three cases of channel permutations among them.

[0052] In the construction method of a fusion feature vector that satisfies the unordered rule, in the first step, according to the Figure 1The channel numbers shown arrange the M basic building blocks. On this basis, the construction process of the feature vectors included in each basic building block is adjusted. Not only randomly shuffle the feature numbers to arrange the N values of the feature vectors, but also arrange the N values of the M feature vectors according to rules that are not completely the same (possibly partially the same, possibly completely different) to construct the fused feature vectors. This method of constructing fused feature vectors is called the channel number - disordered fused feature vector construction method. As mentioned before, there are N! permutation methods for N values, and M! permutation methods for M channels. In this way, the channel number - disordered fused feature vector construction method includes (N!) M - N! specific cases. Figure 4 Three cases of such permutations are shown.

[0053] In the method of constructing fused feature vectors that satisfies the disordered rule, in the second step, randomly shuffle the channel numbers to arrange the M basic building blocks. On this basis, similar to the channel number - disordered fused feature vector construction method, for the feature vectors included in each basic building block, not only randomly shuffle the feature numbers to arrange the N values of the feature vectors, but also arrange the N values of the M feature vectors according to rules that are not completely the same to construct the fused feature vectors. This method of constructing fused feature vectors is called the channel variation - disordered fused feature vector construction method. As mentioned before, there are N! permutation methods for N values, and M! permutation methods for M channels. In this way, the channel variation - disordered fused feature vector construction method includes M! × [(N!) M - N!] specific cases. Figure 5 Three cases of such permutations are shown.

[0054] II. Basic building block - feature:

[0055] Explore the method of constructing fused feature vectors with features as basic building blocks. It also includes two parts. The first part is the method of constructing fused feature vectors that satisfies the ordered rule, and the second part is the method of constructing fused feature vectors that satisfies the disordered rule. Imitating Figure 1 the channel - based fused feature vector construction method with channels as basic building blocks shown, a similar feature - based fused feature vector construction method with features as basic building blocks is given, as Figure 6 shown, and using this as a reference, explore the method of constructing fused feature vectors that satisfies the ordered rule and the disordered rule.

[0056] In the method of constructing fused feature vectors that satisfies the ordered rule, in the first step, arrange the N basic building blocks according to Figure 6 the feature numbers shown. On this basis, adjust the construction process of the feature vectors included in each basic building block. No longer according to Figure 6Instead of using the channel numbers shown, randomly shuffle the M values of the feature vectors for the channel numbers. In this way, the M values of the N feature vectors are arranged according to the same rule to construct a fused feature vector. This method of constructing a fused feature vector is called the feature number-ordered fused feature vector construction method. According to the knowledge of permutations and combinations, there are M! permutation methods for M values. Thus, the feature number-ordered fused feature vector construction method contains M! specific cases. Figure 7 Some of the permutation cases are shown.

[0057] In the method of constructing a fused feature vector that satisfies the ordered rule, in the second step, instead of following Figure 6 the feature numbers shown, randomly shuffle the feature numbers to permute the N basic building blocks. On this basis, similar to the feature number-ordered fused feature vector construction method, for the feature vectors contained in each basic building block, randomly shuffle the M values of the feature vectors for the channel numbers. In this way, the M values of the N feature vectors are arranged according to the same rule to construct a fused feature vector. This method of constructing a fused feature vector is called the feature variation-ordered fused feature vector construction method. As mentioned before, there are M! permutation methods for M values. Similarly, there are N! permutation methods for N features. Thus, the feature variation-ordered fused feature vector construction method contains N! × M! specific cases. Figure 8 Some of the permutation cases are shown.

[0058] In the method of constructing a fused feature vector that satisfies the unordered rule, in the first step, permute the N basic building blocks according to Figure 6 the feature numbers shown. On this basis, adjust the construction process of the feature vectors contained in each basic building block. Not only randomly shuffle the M values of the feature vectors for the channel numbers, but also arrange the M values of the N feature vectors according to not completely the same rules to construct a fused feature vector. This method of constructing a fused feature vector is called the feature number-unordered fused feature vector construction method. As mentioned before, there are M! permutation methods for M values and N! permutation methods for N features. Thus, the feature number-unordered fused feature vector construction method contains (M!) N - M! specific cases. Figure 9 Some representative cases are shown.

[0059] In the method for constructing a fusion feature vector that satisfies the disorder rule, in the second step, the N basic building units are arranged by randomly shuffling the feature numbers. On this basis, similar to the feature number-disorder fusion feature vector construction method, for the feature vectors contained in each basic building unit, not only are the M numerical values of the feature vector randomly shuffled by shuffling the channel numbers, but also the M numerical values of the N feature vectors are arranged according to rules that are not completely the same to construct a fusion feature vector. This method for constructing a fusion feature vector is called the feature variation-disorder fusion feature vector construction method. As mentioned above, there are M! permutation methods for the M numerical values, and N! permutation methods for the N features. Thus, the feature variation-disorder fusion feature vector construction method includes N! × [(M!) N - M!] specific cases. Figure 10 Shows a representative case among them.

[0060] Based on the above construction method, the following analysis is carried out to determine the optimal construction method:

[0061] 1. Obtain signals, extract features, and select to build a classifier:

[0062] 1.1. Sample selection:

[0063] The size of a typical sealed electronic device used in the aerospace system is usually less than 80 cm × 80 cm × 100 cm, and the smallest is even 20 cm × 20 cm × 16 cm. Therefore, in the present invention, an aerospace electronic unit with a size of 80 cm × 80 cm × 64 cm is used as a sample of the sealed electronic device, as Figure 11 shown. There are eight enclosed spaces inside the aerospace electronic unit where foreign objects may exist, numbered from 1 to 8.

[0064] According to the latest foreign object detection information of the aerospace electronic unit, the foreign objects are usually spherical solder balls of 2 mg to 12 mg, and the specific material is tin. Therefore, in the present invention, spherical solder balls of 2 mg are used as foreign object samples.

[0065] 1.2. Sound sensors:

[0066] Figure 12 Figure (a) in [reference] shows the foreign object detection system used, which can connect multiple sound sensors simultaneously through high-shielded transmission lines. The model of the sound sensor is the PXR04 resonant acoustic emission sensor, and the sensitive capture range is 80 cm. For the aerospace electronic unit, usually four sound sensors are placed at different positions on its surface to ensure that the combined detection range of the sound sensors covers all the internal space of the aerospace electronic unit. The spatial layout of the four sound sensors is pre-designed, as Figure 12As shown in (b). The four sound sensors are numbered from 1 to 4, referred to as Channel 1 to Channel 4, and thus the channel numbers are determined.

[0067] Place the debris sample into the enclosed space numbered 1 inside the aerospace electronics unit, re-seal the aerospace electronics unit, and fix it to the vibration table. Drive the vibration table to generate a mechanical excitation to activate the debris sample to be in a random motion state and generate debris signals. In each round of detection, the four sound sensors simultaneously capture a segment of debris signals to obtain a debris signal group, which contains the debris signals of Channel 1 to Channel 4. Through multiple rounds of detection, a large number of debris signal groups are obtained and labeled with "1". And so on, place the debris sample into the enclosed spaces numbered 2 to 8 respectively to obtain a large number of debris signal groups labeled with "2" to "8". After screening, for the debris signal groups of the eight labels, 2000 are retained respectively, a total of 16000 debris signal groups.

[0068] 1.3. Multi-domain features:

[0069] In previous studies, multiple features were extracted from the time domain, frequency domain, and time-frequency domain. Through feature selection, eleven features with good performance were determined. In the present invention, these eleven features are numbered, and thus the feature numbers are determined, as shown in Table 1. On this basis, the values of the eleven features are calculated on each segment of the debris signals in the 16000 debris signal groups.

[0070] Table 1 Eleven features

[0071]

[0072] Refer to respectively Figure 1 and Figure 6 shown channel-based - fusion feature vector construction method and feature-based - fusion feature vector construction method to construct 16000 fusion feature vectors respectively. According to the labels of the debris signal groups, add labels to the fusion feature vectors to obtain 16000 pieces of data respectively, and thus establish two basic data sets, respectively called channel - basic data set and feature - basic data set. Each basic data set contains 16000 pieces of data and eight labels, and each label contains 2000 pieces of data.

[0073] 1.4. Typical classifiers:

[0074] In machine learning, typical classifiers mainly include three types. The first type is the classifier based on mathematical theories, including k-nearest neighbor (kNN), support vector machine (SVM), Naive Bayes, etc. The second type is the classifier based on tree structures, including decision tree, random forest, XGBoost, etc. The third type is the classifier based on neural networks, including multi-layer perceptron, BP neural network, one-dimensional convolutional neural network (1D-CNN), etc. Considering that the currently popular neural networks are all used for image classification, in the present invention, the first two types of classifiers are mainly used, and kNN, SVM, random forest, XGBoost, and 1D-CNN are selected as typical classifiers.

[0075] The channel-basic dataset and the feature-basic dataset are divided into a training set and a test set according to a ratio of 7:3. The above five classifiers are trained on the training set, and their classification performances are evaluated on the test set. Table 2 shows the classification accuracies achieved by the five classifiers on the two basic datasets. Considering that both basic datasets contain eight labels, a radial basis kernel function is introduced for SVM to construct RBF-SVM.

[0076] Table 2 Classification accuracies achieved by the classifiers on the two basic datasets

[0077]

[0078] As can be seen from the table, on either basic dataset, random forest and XGBoost both achieve the most significant classification accuracies, which is consistent with the conclusions of the previous research conducted, showing the superiority of the classifiers based on tree structures for the two basic datasets. In addition, 1D-CNN achieves the lowest classification accuracy on both basic datasets, proving the correctness of using the third type of classifier as a supplement. In addition, the classification accuracies achieved by the five classifiers on the feature-basic dataset are slightly higher than those on the channel-basic dataset, which is worthy of attention.

[0079] At the same time, the F1-scores achieved by the five classifiers on the eight labels of the two basic datasets are statistically counted. This implementation does not list them in detail, see Figure 13 And Figure 14 . As can be seen from the figure, the five classifiers achieve similar F1-scores on the eight labels of the two basic datasets, and the difference from the classification accuracies they achieve is not significant. This shows that they have stable global and local classification performances on the two basic datasets, show the same preference for data with different labels, effectively illustrate the stability of the internal structures of the two basic datasets, and further illustrate the stability of the channel-based - fused feature vector construction method and the feature-based - fused feature vector construction method.

[0080] 2. Basic building unit - channel:

[0081] 2.1 Channel number - ordered fusion feature vector construction:

[0082] Based on the channel - basic dataset, referring to the aforementioned channel number - ordered fusion feature vector construction method, multiple datasets are constructed respectively, which are called multiple channel number - ordered datasets. At this time, N = 11, and a total of 11! = 39916800 channel number - ordered datasets can be constructed. Considering the computing resources and time consumption, 100000 channel number - ordered datasets are randomly selected for construction. Similarly, the 100000 channel number - ordered datasets are divided into a training set and a test set according to the ratio of 7:3. Five classifiers are trained on the training set, and their classification performances are evaluated on the test set. On this basis, for the classification accuracies obtained by the five classifiers on the 100000 channel number - ordered datasets respectively, on the one hand, their means are calculated to reflect the average level, which is called the average classification accuracy. On the other hand, their variances are calculated to reflect the degree of concentration, which is called the classification accuracy variance. Table 3 shows the average classification accuracy and classification accuracy variance obtained by the five classifiers on the channel number - ordered datasets.

[0083] Table 3 Average classification accuracy and classification accuracy variance obtained by the classifier on the channel number - ordered datasets

[0084]

[0085] By comparing the classification accuracies obtained by the five classifiers on the channel - basic dataset in Table 2 and the average classification accuracies obtained by them on the 100000 channel number - ordered datasets in Table 3, it can be found that the difference between the two is small and almost negligible. In addition, the classification accuracy variances obtained by the five classifiers on the 100000 channel number - ordered datasets are extremely small, indicating that the classification accuracies obtained by the five classifiers on the 100000 channel number - ordered datasets are relatively concentrated, all close to the average classification accuracy, with small volatility and difference. This shows that the classification performances of the five classifiers on the channel - basic dataset and the 100000 channel number - ordered datasets are similar, further indicating that the structures of the channel - basic dataset and the channel number - ordered dataset are similar.

[0086] From the perspective of the method for constructing the channel number-ordered fusion feature vector, both the method for constructing the channel number-ordered fusion feature vector and the method for constructing the channel basis-fusion feature vector use channels as the basic construction units and use the same channel numbers. Although the feature numbers they use are different, in the fusion feature vectors they construct, the arrangement structure of the feature vectors contained in each basic construction unit refers to the same feature number, that is, the feature vectors contained in each basic construction unit are arranged with 11 values by referring to the same feature number. It can be seen from this that in the fusion feature vector, whether the arrangement structure of the feature vectors contained in each basic construction unit refers to the same feature number is the core, and whether these methods use the same feature number is not the core. Therefore, generally speaking, the method for constructing the channel number-ordered fusion feature vector and the method for constructing the channel basis-fusion feature vector are approximately the same. This explains why the five classifiers achieve similar classification accuracies and average classification accuracies on the channel-basic dataset and the channel number-ordered dataset.

[0087] In addition, the average F1-scores obtained by the five classifiers on eight labels of 100,000 channel number-ordered datasets were statistically analyzed, as Figure 15 shown. It can also be seen from the figure that the five classifiers all achieved similar average F1-scores on eight labels of 100,000 channel number-ordered datasets, and the difference from the classification accuracies they achieved was not significant. This shows that they all have stable global and local classification performances on 100,000 channel number-ordered datasets, show the same preference for data with different labels, and effectively illustrate the stability of the method for constructing the channel number-ordered fusion feature vector.

[0088] 2.2. Construction of the channel change-ordered fusion feature vector:

[0089] Based on the channel-basic dataset, referring to the aforementioned method for constructing the channel change-ordered fusion feature vector, multiple datasets were constructed respectively, which are called multiple channel change-ordered datasets. At this time, M = 4, N = 11, and a total of 4! × 11! = 958,003,200 channel change-ordered datasets can be constructed. Similarly, 100,000 channel change-ordered datasets were randomly selected and constructed, and divided into a training set and a test set according to a ratio of 7:3. Five classifiers were trained on the training set, and their classification performances, including the average classification accuracy and the variance of the classification accuracy, were evaluated on the test set, as shown in Table 4.

[0090] Table 4 Average classification accuracy and variance of classification accuracy obtained by the classifier on the channel change-ordered dataset

[0091]

[0092] By comparing the classification accuracies achieved by the five classifiers on the channel - basic dataset in Table 2 and their average classification accuracies on the 100,000 channel - change - ordered datasets in Table 4, it can be found that the difference between the two is negligible. At the same time, the variance of the classification accuracies achieved by the five classifiers on the 100,000 channel - change - ordered datasets is extremely small, indicating that the volatility and difference in the classification accuracies achieved by the five classifiers on the 100,000 channel - change - ordered datasets are small. This also shows that the classification performances of the five classifiers on the channel - basic dataset and the 100,000 channel - change - ordered datasets are similar, and the structures of the channel - basic dataset and the channel - change - ordered dataset are similar.

[0093] From the perspective of the construction method of the channel - change - ordered fusion feature vector, both the construction method of the channel - change - ordered fusion feature vector and the construction method of the channel - basic - fusion feature vector use channels as the basic construction units, but different channel numbers and feature numbers are used. They will be described separately. Although the channel numbers they use are different, the number of basic construction units they use and the dimension of the feature vectors contained in each construction unit are the same. Although the feature numbers they use are different, in the fusion feature vectors they construct, the arrangement structure of the feature vectors contained in each basic construction unit refers to the same feature number. It can be seen from this that in the fusion feature vector, whether the number of basic construction units and the dimension of the feature vectors contained in each construction unit are the same, and whether the arrangement structure of the feature vectors contained in each basic construction unit refers to the same feature number are the core, and whether these methods use the same channel number and feature number is not the core. Therefore, generally speaking, the construction method of the channel - change - ordered fusion feature vector and the construction method of the channel - basic - fusion feature vector are approximately the same. This explains the similar classification accuracies and average classification accuracies achieved by the five classifiers on the channel - basic dataset and the channel - change - ordered dataset.

[0094] In addition, the average F1 - scores achieved by the five classifiers on the eight labels of the 100,000 channel - change - ordered datasets were statistically calculated, as Figure 16 shown. It can also be seen from the figure that the five classifiers achieved similar average F1 - scores on the eight labels of the 100,000 channel - change - ordered datasets. This shows that they all have stable global and local classification performances on the 100,000 channel - change - ordered datasets, show the same preference for data with different labels, and effectively illustrate the stability of the construction method of the channel - change - ordered fusion feature vector.

[0095] 2.3. Construction of the channel - number - disordered fusion feature vector:

[0096] Based on the channel - basic dataset, referring to the aforementioned method for constructing the channel number - disordered fusion feature vectors, multiple datasets are constructed respectively, which are called multiple channel number - disordered datasets. At this time, M = 4, N = 11, and a total of (11!) 4 - 11! = 2.5387672e+30 channel number - disordered datasets can be constructed. Similarly, 100,000 channel number - disordered datasets are randomly selected and constructed, and they are divided into a training set and a test set according to a ratio of 7:3. Five classifiers are trained on the training set, and their classification performances are evaluated on the test set, including the average classification accuracy and the variance of the classification accuracy, as shown in Table 5.

[0097] Table 5 Average classification accuracy and variance of classification accuracy obtained by classifiers on the channel number - disordered dataset

[0098]

[0099] Comparing the classification accuracies obtained by the five classifiers on the channel - basic dataset in Table 2, their average classification accuracies obtained on 100,000 channel number - disordered datasets have decreased significantly. Especially for kNN, RBF - SVM, and 1D - CNN, the decline is relatively large, all exceeding 10% or even 20%. In contrast, the decline of Random Forest and XGBoost is relatively small, both not exceeding 10% or even 5%. Among them, Random Forest achieves the highest average classification accuracy, which is 85.45%. In addition, the variances of the classification accuracies obtained by the five classifiers on 100,000 channel number - disordered datasets vary greatly from each other. Among them, Random Forest and XGBoost also achieve relatively small classification accuracy errors, indicating that the classification accuracies they obtain on 100,000 channel number - disordered datasets are relatively concentrated, and the difference from the average classification accuracies they obtain is not large. The overall volatility and difference are within an acceptable range. In contrast, kNN, RBF - SVM, and 1D - CNN obtain relatively large classification accuracy errors, indicating that the classification accuracies they obtain on 100,000 channel number - disordered datasets are relatively dispersed and chaotic, with large volatility and difference. This shows that there are significant differences in the structures of the channel - basic dataset and the channel number - disordered dataset, which have different impacts on the classification performances of different types of classifiers. An explanation is given for this.

[0100] For kNN, the order of features in the fused feature vector, i.e., the feature numbers referred to by the basic building blocks, directly affects its distance metric. Messy and different feature numbers will cause a sharp decline in its classification performance. For RBF-SVM, although the RBF kernel function maps the feature space to a high-dimensional space, the order of features in the fused feature vector will cause great interference to the calculation of the kernel function. For 1D-CNN, the weights of each neuron in the fully connected layer correspond one-to-one with the positions of the input features. The change in the feature order will completely disrupt the weight assignment logic and require continuous re-learning of new parameters. In contrast, in random forest, each decision tree constructs a sample subset and a feature subset through random sampling. The feature order is offset by randomness during splitting, and the global importance assessment depends on the average results of multiple decision trees, diluting the perturbation of the local feature order. Moreover, the decision tree splits based on feature thresholds and is independent of the feature arrangement order, and it only focuses on the distribution relationship of feature values. XGBoost suppresses overfitting through L1 / L2 regularization terms and second-order Taylor expansion, reducing the dependence on feature order. Moreover, the gain-based feature selection mechanism of XGBoost itself can adaptively adjust the splitting path, reducing the influence of feature order. This explains why the decline in the classification performance of random forest and XGBoost is relatively small. In addition, because the double randomness of random forest (randomly constructing sample subsets + feature subsets) is stronger than the single-layer randomness of XGBoost (randomly constructing feature subsets), further suppressing its preference for feature order, therefore, the decline in the classification performance of random forest is slightly less than that of XGBoost.

[0101] From the perspective of the method for constructing the channel number-unordered fused feature vector, both the method for constructing the channel number-unordered fused feature vector and the method for constructing the channel-based fused feature vector use channels as the basic building blocks and use the same channel numbers. The difference is that not only do they use different feature numbers, but also in the fused feature vectors they construct, the feature numbers referred to by the arrangement structure of the feature vectors contained in each basic building block are not exactly the same. In step 2.1, it is concluded that whether the arrangement structure of the feature vectors contained in each basic building block in the fused feature vector refers to the same feature number is the core. Unfortunately, this core is discarded in the feature vector constructed by the method for constructing the channel number-unordered fused feature vector. This explains the large difference in the classification accuracy obtained by the five classifiers on the channel-based dataset and their average classification accuracy obtained on the channel number-unordered dataset.

[0102] In addition, the average F1-scores obtained by the five classifiers on eight labels of 100,000 channel number-unordered datasets were counted, as Figure 17As shown. It can also be seen from the figure that the average F1-scores obtained by kNN, RBF-SVM, and 1D-CNN on the eight labels of the 100,000 channel number-unordered dataset vary greatly from each other. The average F1-scores obtained by Random Forest and XGBoost on the eight labels of the 100,000 channel number-unordered dataset are relatively less different from each other, but the change is obvious compared with the average F1-scores they obtained in Step 2.1 and Step 2.2. This shows that kNN, RBF-SVM, and 1D-CNN have unstable global and local classification performance on the 100,000 channel number-unordered dataset, and Random Forest and XGBoost have non-absolutely stable global and local classification performance on the 100,000 channel number-unordered dataset. This effectively illustrates the instability of the channel number-unordered fusion feature vector construction method.

[0103] 2.4. Construction of Channel Change-Unordered Fusion Feature Vectors:

[0104] Based on the channel-basic dataset, referring to the aforementioned channel change-unordered fusion feature vector construction method, multiple datasets are constructed respectively, called multiple channel change-unordered datasets. At this time, M = 4, N = 11, and a total of 4! × [(11!) 4 - 11!] = 6.0930412e+31 channel change-unordered datasets can be constructed. Similarly, 100,000 channel change-unordered datasets are randomly selected and constructed, and divided into a training set and a test set according to a ratio of 7:3. Five classifiers are trained on the training set, and their classification performance is evaluated on the test set, including the average classification accuracy and the variance of the classification accuracy, as shown in Table 6.

[0105] Table 6 Average Classification Accuracy and Variance of Classification Accuracy Obtained by Classifiers on Channel Change-Unordered Datasets

[0106]

[0107] Comparing the classification accuracies achieved by the five classifiers on the channel - basic dataset in Table 2, their average classification accuracies on the 100,000 channel - varying - unordered datasets also decreased significantly, being similar to their average classification accuracies on the 100,000 channel - numbered - unordered datasets. More specifically, the decreases of kNN, RBF - SVM, and 1D - CNN all exceeded 10%, even 20%, while the decreases of Random Forest and XGBoost did not exceed 10%, even 5%. Moreover, Random Forest achieved the highest average classification accuracy of 85.33%. In addition, the classification accuracy errors of Random Forest and XGBoost were both relatively small, indicating that the classification accuracies they achieved on the 100,000 channel - varying - unordered datasets were relatively concentrated, with a small difference from their average classification accuracies, and the overall volatility and difference were within an acceptable range. The classification accuracy errors of kNN, RBF - SVM, and 1D - CNN were all relatively large, indicating that the classification accuracies they achieved on the 100,000 channel - varying - unordered datasets were relatively scattered and chaotic, with large volatility and difference. This shows that there are significant differences in the structures of the channel - basic dataset and the channel - varying - unordered dataset, which have different impacts on the classification performances of different types of classifiers. For the explanation of this part, refer to Step 2.3.

[0108] From the perspective of the construction method of the channel - varying - unordered fusion feature vector, both the construction method of the channel - varying - unordered fusion feature vector and the construction method of the channel - basic - fusion feature vector use channels as the basic construction units, but different channel numbers and feature numbers are used. They will be explained separately. For the channel numbers, although they are different, the number of basic construction units they use and the dimension of the feature vector contained in each construction unit are the same. However, for the feature numbers, they are not only different, but also the arrangement structure of the feature vectors contained in each basic construction unit in the fusion feature vectors they construct refers to different feature numbers. In Step 2.2, it was concluded that whether the number of basic construction units and the dimension of the feature vector contained in each construction unit in the fusion feature vector are the same, and whether the arrangement structure of the feature vectors contained in each basic construction unit refers to the same feature number are the cores. From this perspective, in the feature vector constructed by the construction method of the channel - varying - unordered fusion feature vector, the first two cores are retained, and the third core is discarded. This explains why there is a large difference in the classification accuracies achieved by the five classifiers on the channel - basic dataset and the average classification accuracies on the channel - varying - unordered dataset, and there is not much difference between the average classification accuracies on the channel - numbered - unordered dataset and the average classification accuracies on the channel - varying - unordered dataset.

[0109] In addition, the average F1-scores obtained by five classifiers on eight labels of 100,000 channel change-disorder datasets were counted, as Figure 18 shown. It can also be seen from the figure that the average F1-scores obtained by kNN, RBF-SVM, and 1D-CNN on eight labels of 100,000 channel change-disorder datasets vary greatly from each other. The average F1-scores obtained by the random forest and XGBoost on eight labels of 100,000 channel change-disorder datasets vary relatively little from each other, but the change is obvious compared with the average F1-scores they obtained in Step 2.1 and Step 2.2. This shows again that kNN, RBF-SVM, and 1D-CNN have unstable global and local classification performances on 100,000 channel change-disorder datasets, and the random forest and XGBoost have non-absolutely stable global and local classification performances on 100,000 channel change-disorder datasets. This effectively illustrates the instability of the channel change-disorder fusion feature vector construction method.

[0110] Based on the research content from Step 2.1 to Step 2.4, it is concluded that among the fusion feature vector construction methods with channels as the basic construction units, the channel number-ordered fusion feature vector construction method and the channel change-ordered fusion feature vector construction method are the optimal ones. That is, in the case of using channels as the basic construction units, the fusion feature vector construction method that satisfies the ordered rule is the optimal one. Considering that the channel number-ordered fusion feature vector construction method essentially belongs to the channel change-ordered fusion feature vector construction method, therefore, it is finally determined that the channel change-ordered fusion feature vector construction method is the optimal one. In addition, it is also concluded that the classification performance of the classifier based on the tree structure is the most prominent and stable, especially the random forest performs the best.

[0111] 3. Basic construction unit - feature:

[0112] 3.1 Feature number-ordered fusion feature vector construction:

[0113] Based on the feature-basic dataset, referring to the aforementioned feature number-ordered fusion feature vector construction method, multiple datasets were constructed respectively, called multiple feature number-ordered datasets. At this time, N = 4, and a total of 4! = 24 feature number-ordered datasets could be constructed. The 24 feature number-ordered datasets were divided into a training set and a test set according to a ratio of 7:3, five classifiers were trained, and their average classification accuracy and classification accuracy variance were calculated, as shown in Table 7.

[0114] Table 7 Average classification accuracy and classification accuracy variance obtained by classifiers on feature number-ordered datasets

[0115]

[0116] By comparing the classification accuracies achieved by the five classifiers on the feature-basic dataset in Table 2 and their average classification accuracies on the 24 feature-number-ordered datasets in Table 7, it can be found that the difference between the two is small and can be ignored, which is comparable to their performance in Step 2.1. In addition, the variances of the classification accuracies achieved by the five classifiers are small, indicating that the fluctuations and differences in the classification accuracies they achieved are small. The above conclusions can illustrate that the structures of the feature-basic dataset and the feature-number-ordered dataset are similar.

[0117] From the perspective of the construction method of the feature-number-ordered fusion feature vector, both the construction method of the feature-number-ordered fusion feature vector and the construction method of the feature-based fusion feature vector use features as the basic construction units and use the same feature numbers. Although the channel numbers they use are different, in the fusion feature vectors they construct, the arrangement structure of the feature vectors contained in each basic construction unit refers to the same channel number. Thus, whether the arrangement structure of the feature vectors contained in each basic construction unit refers to the same channel number is the core. Therefore, generally speaking, the construction method of the channel-number-ordered fusion feature vector is approximately the same as the construction method of the feature-based fusion feature vector.

[0118] It is worth noting that the average classification accuracies achieved by the five classifiers on the 24 feature-number-ordered datasets are slightly higher than those they achieved on the 100,000 channel-number-ordered datasets. Combining the fact that the classification accuracies achieved by the five classifiers on the feature-basic dataset are slightly higher than those they achieved on the channel-basic dataset, it can be preliminarily concluded that using features as the basic construction unit has advantages.

[0119] In addition, the average F1-scores achieved by the five classifiers on the eight labels of the 24 feature-number-ordered datasets were statistically analyzed, as Figure 19 shown. It can also be seen from the figure that the five classifiers achieved similar average F1-scores on the eight labels of the 24 feature-number-ordered datasets. This shows that they all have stable global and local classification performances on the 24 feature-number-ordered datasets, effectively illustrating the stability of the construction method of the feature-number-ordered fusion feature vector.

[0120] 3.2. Construction of the feature-variation-ordered fusion feature vector:

[0121] Based on the channel - basic dataset, referring to the aforementioned method for constructing the feature - change - ordered fusion feature vector, multiple datasets are constructed respectively, which are called multiple feature - change - ordered datasets. At this time, M = 4 and N = 11, and a total of 11!×4! = 958003200 feature - change - ordered datasets can be constructed. Similarly, 100000 feature - change - ordered datasets are randomly selected and constructed, five classifiers are trained, and the average classification accuracy and the variance of classification accuracy obtained by them are calculated, as shown in Table 8.

[0122] Table 8 Average classification accuracy and variance of classification accuracy obtained by the classifier on the feature - change - ordered dataset

[0123]

[0124] Comparing the classification accuracy of the five classifiers on the feature - basic dataset in Table 2 and the average classification accuracy they obtained on 100000 feature - change - ordered datasets in Table 8, it can be found that the difference between the two is relatively small and can be ignored, which is equivalent to their performance in step 2.2. In addition, the variance of the classification accuracy obtained by the five classifiers is relatively small, indicating that the volatility and difference of the classification accuracy they obtained are relatively small. The above conclusions can illustrate that the structures of the feature - basic dataset and the feature - change - ordered dataset are similar.

[0125] From the perspective of the method for constructing the feature - change - ordered fusion feature vector, both the method for constructing the feature - change - ordered fusion feature vector and the method for constructing the feature - based - fusion feature vector use features as the basic construction units, but different feature numbers and channel numbers are used. They are described separately. Although the feature numbers they use are different, the number of basic construction units they use and the dimension of the feature vector contained in each construction unit are the same. Although the channel numbers they use are different, in the fusion feature vectors they construct, the arrangement structure of the feature vectors contained in each basic construction unit refers to the same channel number. Thus, whether the number of basic construction units and the dimension of the feature vector contained in each construction unit are the same, and whether the arrangement structure of the feature vectors contained in each basic construction unit refers to the same channel number are the core. Therefore, generally speaking, the method for constructing the feature - change - ordered fusion feature vector and the method for constructing the feature - based - fusion feature vector are also approximately the same.

[0126] It is worth noting that the average classification accuracy of the five classifiers obtained on 100000 feature - change - ordered datasets is slightly higher than that obtained on 100000 channel - change - ordered datasets, which once again shows the advantage of using features as the basic construction unit.

[0127] In addition, the average F1-scores obtained by five classifiers on eight labels of 100,000 feature-variation-ordered datasets were counted, as Figure 20 shown. It can also be seen from the figure that the five classifiers all achieved similar average F1-scores on eight labels of 100,000 feature-variation-ordered datasets. This shows that they all have stable global and local classification performances on 100,000 feature-variation-ordered datasets, effectively demonstrating the stability of the feature-variation-ordered fusion feature vector construction method.

[0128] 3.3. Construction of feature-number-unordered fusion feature vectors:

[0129] Based on the channel-basic dataset, referring to the aforementioned feature-number-unordered fusion feature vector construction method, multiple datasets were constructed respectively, called multiple feature-number-unordered datasets. At this time, M = 4, N = 11, and a total of (4!) 11 - 4! = 1.5216811e+15 feature-number-unordered datasets could be constructed. Similarly, 100,000 feature-number-unordered datasets were randomly selected and constructed, five classifiers were trained, and the average classification accuracy and classification accuracy variance they obtained were calculated, as shown in Table 9.

[0130] Table 9 Average classification accuracy and classification accuracy variance obtained by classifiers on feature-number-unordered datasets

[0131]

[0132] Comparing the classification accuracies obtained by the five classifiers on the feature-basic dataset in Table 2, their average classification accuracies on 100,000 feature-number-unordered datasets decreased significantly, which was comparable to their performance in Step 2.3. Among them, the decline rates of kNN, RBF-SVM, and 1D-CNN all exceeded 10%, even 20%, and the decline rates of random forest and XGBoost did not exceed 5%. Moreover, the random forest achieved the highest average classification accuracy, which was 85.71%. In addition, the classification accuracy errors obtained by the random forest and XGBoost were both small, indicating that the classification accuracies they obtained on 100,000 feature-number-unordered datasets were relatively concentrated, and the difference from their average classification accuracies was not large. The overall volatility and difference were within an acceptable range. The classification accuracy errors obtained by kNN, RBF-SVM, and 1D-CNN were all large, indicating that the classification accuracies they obtained on 100,000 feature-number-unordered datasets were relatively scattered and chaotic, with large volatility and difference. This shows that there are significant differences in the structures of the channel-basic dataset and the feature-number-unordered dataset, which have different impacts on the classification performances of different types of classifiers. For the explanation of this part, please refer to Step 2.3.

[0133] From the perspective of the construction method of feature number-unordered fusion feature vectors, both the construction method of feature number-unordered fusion feature vectors and the construction method of feature basis-fusion feature vectors use features as the basic construction units and use the same feature numbers. The difference is that not only do they use different channel numbers, but also in the fusion feature vectors they construct, the arrangement structure of the feature vectors contained in each basic construction unit refers to different channel numbers. In step 3.1, it is obtained that whether the arrangement structure of the feature vectors contained in each basic construction unit in the fusion feature vector refers to the same channel number is the core. Unfortunately, in the feature vectors constructed by the construction method of feature number-unordered fusion feature vectors, this core is discarded. This explains the large difference in the classification accuracy obtained by the five classifiers on the feature-basic dataset and their average classification accuracy on the feature number-unordered dataset.

[0134] It is worth noting that the average classification accuracy obtained by the five classifiers on 100,000 feature number-unordered datasets is slightly higher than that obtained on 100,000 channel number-unordered datasets, which once again shows the advantage of using features as the basic construction unit.

[0135] In addition, the average F1-scores obtained by the five classifiers on eight labels of 100,000 feature number-unordered datasets are statistically calculated, as Figure 21 shown. It can also be seen from the figure that the average F1-scores obtained by kNN, RBF-SVM, and 1D-CNN on eight labels of 100,000 feature number-unordered datasets vary greatly from each other. The average F1-scores obtained by Random Forest and XGBoost on eight labels of 100,000 feature number-unordered datasets vary relatively little from each other, but the change is obvious compared with their average F1-scores obtained in step 3.1 and step 3.2. This shows that kNN, RBF-SVM, and 1D-CNN have unstable global and local classification performances on 100,000 feature number-unordered datasets, and Random Forest and XGBoost have non-absolutely stable global and local classification performances on 100,000 feature number-unordered datasets. This effectively illustrates the instability of the construction method of feature number-unordered fusion feature vectors.

[0136] 3.4. Construction of Feature Variation-Unordered Fusion Feature Vectors:

[0137] Based on the feature-basic dataset, referring to the aforementioned construction method of feature variation-unordered fusion feature vectors, multiple datasets are constructed respectively, which are called multiple feature variation-unordered datasets. At this time, M = 4, N = 11, and a total of 11! × [(4!)11 -4!] = 6.0740642e+22 feature variation - unordered datasets. Similarly, randomly select 100,000 feature variation - unordered datasets from them, train five classifiers, and calculate the average classification accuracy and the variance of classification accuracy they achieved, as shown in Table 10.

[0138] Table 10 Average classification accuracy and variance of classification accuracy achieved by classifiers on feature variation - unordered datasets

[0139]

[0140] Comparing the classification accuracies achieved by the five classifiers on the feature - basic datasets in Table 2, their average classification accuracies achieved on 100,000 feature variation - unordered datasets also decreased significantly, were similar to their average classification accuracies achieved on 100,000 feature number - unordered datasets, and were comparable to their performances in Step 2.4. Among them, the decline rates of kNN, RBF - SVM, and 1D - CNN all exceeded 10%, even 20%, the decline rate of the random forest did not exceed 5%, the decline rate of XGBoost was 5.04% (slightly exceeding 5%), and the random forest achieved the highest average classification accuracy of 85.32%. In addition, the classification accuracy errors achieved by the random forest and XGBoost were both small, indicating that the classification accuracies they achieved on 100,000 feature variation - unordered datasets were relatively concentrated, the difference from their average classification accuracies was not large, and the overall volatility and difference were within an acceptable range. The classification accuracy errors achieved by kNN, RBF - SVM, and 1D - CNN were all large, indicating that the classification accuracies they achieved on 100,000 feature variation - unordered datasets were relatively scattered and chaotic, with large volatility and difference. This shows that there are significant differences in the structures of the channel - basic datasets and the feature variation - unordered datasets, which have different impacts on the classification performances of different types of classifiers. For the explanation of this part, refer to Step 2.3.

[0141] From the perspective of the method for constructing the feature change-disorder fusion feature vector, both the method for constructing the feature change-disorder fusion feature vector and the feature basis-fusion feature vector construction method use channels as the basic construction units, but use different feature numbers and channel numbers. They will be described separately. For the feature numbers, although they are different, the number of basic construction units they use and the dimension of the feature vector contained in each construction unit are the same. However, for the channel numbers, they are not only different, but also the channel numbers referred to by the arrangement structure of the feature vectors contained in each basic construction unit in the fusion feature vectors they construct are not the same. In step 3.2, it is concluded that whether the number of basic construction units in the fusion feature vector and the dimension of the feature vector contained in each construction unit are the same, and whether the arrangement structure of the feature vectors contained in each basic construction unit refers to the same channel number are the core. From this, it can be seen that in the feature vector constructed by the method for constructing the feature change-disorder fusion feature vector, the first two cores are retained and the third core is discarded. This explains the large difference in the classification accuracy obtained by the five classifiers on the channel-basic dataset and the average classification accuracy obtained on the feature change-disorder dataset, and the small difference in the average classification accuracy obtained on the feature number-disorder dataset and the average classification accuracy obtained on the feature change-disorder dataset.

[0142] It is worth noting that the average classification accuracy obtained by the five classifiers on 100,000 feature change-disorder datasets is slightly higher than that obtained on 100,000 channel change-disorder datasets, which once again shows the advantage of using features as the basic construction unit.

[0143] In addition, the average F1-scores obtained by the five classifiers on eight labels of 100,000 feature change-disorder datasets were statistically analyzed, as Figure 22 shown. It can also be seen from the figure that the average F1-scores obtained by kNN, RBF-SVM, and 1D-CNN on eight labels of 100,000 feature change-disorder datasets are quite different from each other. The average F1-scores obtained by the random forest and XGBoost on eight labels of 100,000 feature change-disorder datasets are relatively less different from each other, but the change is obvious compared with the average F1-scores they obtained in steps 3.1 and 3.2. This once again shows that kNN, RBF-SVM, and 1D-CNN have unstable global and local classification performances on 100,000 feature change-disorder datasets, and the random forest and XGBoost have non-absolutely stable global and local classification performances on 100,000 feature change-disorder datasets. This effectively illustrates the instability of the method for constructing the feature change-disorder fusion feature vector.

[0144] Based on the research contents of steps 3.1 to 3.4, it is also concluded that in the method for constructing a fused feature vector with features as the basic building blocks, the method for constructing a fused feature vector that satisfies the ordered rule is the optimal one. Specifically, the feature change-ordered fused feature vector construction method is the optimal one. In addition, it is also concluded that the classification performance of the classifier based on the tree structure is the most prominent and stable, and especially the random forest performs the best.

[0145] The final conclusion is given that for the positioning of foreign objects in sealed electronic devices, the method for constructing a fused feature vector with features as the basic building blocks and satisfying the ordered rule is the optimal one. Specifically, the feature change-ordered fused feature vector construction method is the optimal one. Moreover, the classification performance of the classifier based on the tree structure is the most prominent and stable, and the random forest is a recommended classifier. In previous studies, the channel-based fused feature vector construction method with channels as the basic building blocks was adopted, and the dataset was established and the random forest with the best performance was trained. The research of the present invention will further improve the quality of the dataset and the classification performance of the classifier, and further enhance the ability to locate foreign objects in sealed electronic devices.

[0146] Aiming at the key problem of multi-sound sensor signal fusion in the positioning of foreign objects in sealed electronic devices, the present invention systematically proposes a method for constructing a fused feature vector with channels or features as the basic building blocks, and verifies the decisive role of the ordered rule in the classification performance of the classifier through experiments. The main conclusions are as follows:

[0147] (1) The method for constructing a fused feature vector with features as the basic building blocks and satisfying the ordered rule is the optimal one. Under the same conditions, the method for constructing a fused feature vector with features as the basic building blocks is superior to the method for constructing a fused feature vector with channels as the basic building blocks, and the method for constructing a fused feature vector that satisfies the ordered rule is superior to the method for constructing a fused feature vector that satisfies the disordered rule. The feature change-ordered fused feature vector construction method performs the best, and the most prominent random forest achieves an average classification accuracy of 88.89%.

[0148] (2) The method for constructing a fused feature vector that satisfies the disordered rule leads to a significant degradation in the classification performance of the classifier. On the dataset constructed by the method for constructing a fused feature vector that satisfies the ordered rule, the average classification accuracies obtained by multiple classifiers all decrease significantly, and the variances of the classification accuracies all increase significantly. The average classification accuracy obtained by the random forest drops to 85.32%, and the variance of the classification accuracy increases to 1.0692%. The maximum decrease in the average classification accuracy obtained by other classifiers exceeds 20%, and the maximum increase in the variance of the classification accuracy exceeds 4%.

[0149] (3) Random forests demonstrate remarkable robustness and superiority among various methods for constructing fused feature vectors. Whether in methods for constructing fused feature vectors with channels or features as basic building blocks, satisfying the construction of fused feature vectors with or without, or in an unordered manner, the double randomness mechanism of random forests effectively suppresses the influence of feature permutation perturbations, achieving significant and stable average classification accuracy and variance of classification accuracy.

[0150] Finally, the feature vectors constructed through the method for constructing fused feature vectors and the corresponding classifiers are ultimately used for the positioning of foreign objects in sealed electronic devices.

[0151] The present invention effectively improves the ability to locate foreign objects in sealed electronic devices and reduces the equipment maintenance cost. In addition, the feature logic concerned in the present invention enhances the interpretability of the model, providing a theoretical reference for the intelligent detection of large and complex equipment.

[0152] The above examples of the present invention are only for illustrating in detail the calculation model and calculation process of the present invention, rather than limiting the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or variations can be made on the basis of the above description. It is impossible to enumerate all the implementation manners here. Any obvious changes or variations derived from the technical solutions of the present invention still fall within the protection scope of the present invention.

Claims

1. A method for determining a construction method of a fusion feature vector for locating redundant objects in an electronic device, characterized in that, The method includes: For the sound signals collected from M channels for locating the redundant objects in the sealed electronic device, N features are obtained based on the sound signals corresponding to each channel; Taking the channels as basic building units, the M basic building units are arranged in the order of channel numbers or randomly shuffled; and for each basic building unit, the N features of the feature vector are arranged by randomly shuffling the feature numbers, or the N features of the feature vector are arranged by randomly shuffling the feature numbers and the N features of the M feature vectors are arranged according to rules that are not completely the same, to obtain the fused feature vectors; Taking the features as basic building units, the N basic building units are arranged in the order of feature numbers or randomly shuffled; and for each basic building unit, the M features of the feature vector are arranged by randomly shuffling the channel numbers, or the M features of the feature vector are arranged by randomly shuffling the channel numbers and the M features of the N feature vectors are arranged according to rules that are not completely the same, to obtain the fused feature vectors; For the fused feature vectors obtained by taking the channels as basic building units and the fused feature vectors obtained by taking the features as basic building units, a classifier is respectively used for location identification, and the construction method of the fused feature vectors is selected based on the location identification accuracy rate corresponding to the fused feature vectors; The obtained construction method of the fused feature vectors is used as the construction method of the fused feature vectors for locating the redundant objects in the sealed electronic device.

2. The method for determining a construction method of a fused feature vector for positioning redundant objects in an electronic device according to claim 1, wherein The M channels are M sound sensors.

3. The determination method for constructing a fusion feature vector for positioning redundant objects in an electronic device according to claim 2, wherein The number of sound sensors serving as channels is at least set to 4.

4. A determination method for constructing a fusion feature vector for positioning redundant objects in an electronic device according to claim 1, characterized in that The N features obtained based on the sound signals corresponding to each channel include time domain features and frequency domain features; The time domain features include time delay, pulse rise time, pulse symmetry, pulse amplitude, pulse energy, root mean square voltage, and zero crossing rate; the frequency domain features include spectral centroid, spectral mean square deviation, root mean square probability, and frequency standard deviation.

5. A method for constructing a fusion feature vector for positioning redundant objects in an electronic device, characterized in that, It includes: For the sound signals collected from M channels for locating the redundant objects in the sealed electronic device, N features are obtained based on the sound signals corresponding to each channel; Taking the features as basic building units, the N basic building units are arranged according to the feature numbers, and for each basic building unit, the M features of the feature vector are arranged by randomly shuffling the channel numbers to construct the fused feature vectors; or taking the features as basic building units, the N basic building units are randomly shuffled according to the feature numbers, and for each basic building unit, the M features of the feature vector are arranged by randomly shuffling the channel numbers to construct the fused feature vectors.

6. A method for constructing a fusion feature vector for positioning redundant objects in an electronic device according to claim 5, characterized in that, The M channels are M sound sensors.

7. A method for constructing a fusion feature vector for positioning redundant objects in an electronic device according to claim 6, characterized in that The number of sound sensors serving as channels is at least set to 4.

8. A method for constructing a fusion feature vector for positioning redundant objects in an electronic device according to claim 5, characterized in that, The number of sound sensors serving as channels is at least set to 4.

9. A method for locating foreign objects in a sealed electronic device, characterized in that, Using the construction method of a fused feature vector for locating the redundant objects in the sealed electronic device according to any one of claims 5 to 8 to construct the fused feature vectors, and for the fused feature vectors, a classifier is used to perform location identification of the redundant objects in the sealed electronic device to achieve location.

10. A method for positioning foreign objects in a sealed electronic device according to claim 9, characterized in that, The classifier is a random forest model.

Images