Two-level adaptive multi-threshold pulse extraction method for redundant object detection

Through the two-stage adaptive multi-threshold pulse extraction method, the problem of insufficient adaptability of the pulse extraction algorithm in the prior art is solved, and more accurate pulse extraction and higher classification accuracy are achieved, which is suitable for redundant detection.

CN117912487BActive Publication Date: 2024-11-12HARBIN INST OF TECH
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202410074529.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-18
Publication Date
2024-11-12
Estimated Expiration
2044-01-18

AI Technical Summary

Technical Problem

In the existing redundant detection methods, the pulse extraction algorithm lacks adaptability, resulting in insufficient accuracy of the extracted pulse signal and inability to effectively identify the redundant signal and component signals.

Method used

A two-stage adaptive multi-threshold pulse extraction method is proposed. By calculating the average energy and average zero crossing rate of the entire sound signal, setting a 3σ distribution criterion based on normal distribution, processing frame by frame, and comparing signal characteristics accurately. Useful pulses to solve the problem of continuous multi-pulse.

Benefits of technology

It improves the accuracy and stability of the pulse signal, and enhances the classification accuracy of the classifier, including the classification accuracy of the excess signal and component signal, and the classification accuracy of the excess material of the specific material.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117912487B_ABST
    Figure CN117912487B_ABST
Patent Text Reader

Abstract

A two-stage adaptive multi-threshold pulse extraction method for detecting unwanted objects belongs to the technical field of monitoring unwanted objects in sealed electronic equipment. In order to solve the problem that the accuracy of the extracted pulse signal in the existing limited pulse extraction method for detecting unwanted objects needs to be further improved due to the lack of adaptability of the threshold. The present invention first calculates the average energy and average zero-crossing rate of all sampling points contained in the sound signal, and sets the peak energy threshold, the endpoint energy threshold and the endpoint zero-crossing rate threshold according to the 3σ distribution criterion; first quickly finds the highest point of the useful pulse through the peak energy threshold, first finds the starting point and the ending point of a round of useful pulses through the endpoint energy threshold, and then resets the endpoint energy threshold and the endpoint zero-crossing rate threshold according to the 3σ distribution criterion again. Two new thresholds, the starting point and the ending point of the second round of useful pulses, and finally determines the final starting point and the ending point according to the duration of the two rounds of pulses.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of monitoring excess matter in sealed electronic equipment, and in particular relates to a pulse extraction method for excess matter signals in sealed electronic equipment. Background Art

[0002] Sealed electronic components are an important part of aerospace electronic equipment. During their production process, particles such as solder particles, copper wires, and silicone wires may be encapsulated inside sealed electronic components. These particles that are introduced from the outside or generated internally and do not conform to the standard physical state specified for the product are called excess matter. Aerospace electronic equipment usually works in environments such as super-weightlessness and severe impact. The excess matter remaining inside sealed electronic components will be activated and in a state of random motion. It will collide with internal components to damage them, adhere to or get stuck on circuit boards to cause short circuits and open circuits, slide to generate static electricity and electromagnetic interference, etc. This will cause sealed electronic components to fail, causing aerospace electronic equipment to work abnormally, causing aerospace missions to fail, and even causing casualties. Some reports of major accidents caused by excess matter that have attracted global attention can be searched and referenced.

[0003] The particle impact noise detection (PIND) method is a widely used method for detecting unwanted objects. Its detection principle is as follows. By applying mechanical excitation to the DUT, the unwanted objects that may exist inside it are activated and random collision or sliding signals are generated, which are called unwanted object signals. The acoustic emission sensor arranged on the surface of the DUT captures the sound signal generated inside the DUT in real time, transmits it to the conditioning circuit for processing and outputs it to the oscilloscope. By observing whether the oscilloscope generates a waveform, it can be judged whether there are unwanted objects inside the DUT. However, usually, there are loose components such as bonding wires that are fixed in position but still have freedom inside the sealed electronic components. Under mechanical excitation, the loose components will generate a fixed period vibration signal with the sealed electronic components, which is called a component signal. Therefore, in the PIND method, the sound signal generated inside the DUT may be a unwanted object signal, a component signal, or even two signals. Therefore, accurately identifying unwanted object signals and component signals is the key to unwanted object detection. In addition, the degree of harm of unwanted objects of different materials to sealed electronic components is different. Excess metal materials are generally hard and conductive, and have a high degree of harm, but the impact of excess non-metal materials is limited. Some sealed electronic components in the aerospace field that contain excess non-metal materials can even be downgraded to the civilian field for continued use. At the same time, obtaining the material of the excess can trace the production link that produces the excess, which has important reference value for excess control work.

[0004] As mentioned above, the randomness of the redundant signal and the periodicity of the component signal are the most significant differences in signal characteristics between the two. There must be differences in signal characteristics between the redundant signals generated by redundant objects of different materials. The principle of signal recognition technology based on deep machine learning is to extract signal features from multiple domains to quantify the differences in signal characteristics between signals of different categories, construct a data set or a picture set, and train a suitable classifier to complete data or picture classification, thereby realizing signal recognition. This technology has been widely used and verified in the fields of acoustic emission detection, fault diagnosis, and speech signal processing. Similarly, deep machine learning methods can be used to carry out research on the recognition of redundant signals and component signals, as well as the recognition of redundant signals generated by redundant objects of different materials (referred to as redundant material recognition research). Many scholars have carried out some research work in the early stage.

[0005] In the identification of redundant signals and component signals, for example, Liang Xiaowen et al. (2020) extracted four signal features from the frequency domain to construct a data set, and the trained parameter-optimized decision tree achieved a classification accuracy of more than 80%. At the same time, they (2020) took into account the data imbalance problem in the data set and proposed a new LR-SMOTE algorithm to further improve the classification accuracy to more than 90%. Li Chaoran et al. (2020) extracted thirteen signal features from the time domain and frequency domain to construct a data set, and trained the parameter-optimized XGBoost to achieve a classification accuracy of about 90%. Li Xiang et al. (2020) extracted fourteen signal features from the time domain and frequency domain, and used the feature selection method combining the chi-square test and the tree model to select four excellent signal features to construct a data set. The classification accuracy of the trained hyperparameter-optimized multilayer perceptron was 87.1%. Xue Yongyue (2020) extracted ten signal features from the time domain and frequency domain, and on this basis, conducted a study on the combination transformation of signal features, and focused on analyzing the signal features that contribute greatly to the classification performance of the classifier. The results show that four signal features were selected to construct a data set, support vector machines and random forests were trained, and more than 95% classification accuracy was achieved by both. Gao Yajie et al. (2023) extracted ten signal features from the time domain and frequency domain to construct a data set, and the trained fusion machine learning (FML) method based on the Boosting algorithm achieved a classification accuracy of more than 95%. Sun Zhigang et al. (2024) extracted fifteen signal features from the time domain and frequency domain, selected eleven of the outstanding signal features to construct a data set, and trained a support vector machine based on a linear kernel function with parameter optimization to achieve a classification accuracy of 92.16%.

[0006] In the identification of the material of the waste, for example, for the waste of tin slag, epoxy and filament, Meng Si et al. (2020) extracted seven signal features from the time domain and frequency domain, constructed a data set containing three labels, and the classification accuracy of the support vector machine with optimized training parameters reached 98%. For the waste of metal (lead) and non-metal (plastic, plastic) materials, Zhao Runsen (2020) extracted 58 signal features from the frequency domain, wavelet transform and Hilbert Huang transform, respectively, constructed a data set containing two labels, and the trained BP neural network achieved a classification accuracy of 91.11%. For the redundant objects of metal (copper wire, solder particles, aluminum particles) and non-metal (hot melt adhesive particles, silicone particles), Gao Yajie et al. (2021) first extracted five signal features from the time domain and frequency domain, then converted the one-dimensional signal into a two-dimensional spectrogram, and extracted four new signal features from the spectrogram of the redundant object signal. On this basis, they constructed data sets containing two labels and five labels respectively, and the classification accuracy achieved by the trained parameter-optimized random forest was 90% and 78%, respectively. For the redundant objects of metal and non-metal materials, Lü Bingze et al. (2021) extracted fourteen signal features from the time domain and frequency domain, constructed a data set containing two labels, and trained a parameter-optimized gradient boosting decision tree (GBDT) to achieve a classification accuracy of 93.2%. For the waste materials of metal (copper wire, solder particles, aluminum particles) and non-metal (hot melt adhesive particles, PVC particles, silicone particles), Sun Zhigang et al. (2022) extracted fourteen signal features from the time domain and frequency domain, conducted feature optimization research on this basis, and focused on analyzing the signal features that contributed most to the classification performance of the classifier. The results showed that twelve signal features were selected to construct a dataset containing six labels, and a random forest was trained to achieve a classification accuracy of 63.6%. In a subsequent study (2024), they trained the parameter-optimized XGBoost and achieved a classification accuracy of 80.41%. For the waste materials of four materials, aluminum scraps, solder, plastic and rubber, Zhang Heng and Liu Haijiang (2023) extracted twenty-four energy-weighted signal features from the Mel frequency cepstral coefficients, constructed a dataset containing four labels, and trained an optimized probabilistic neural network to achieve a classification accuracy of more than 90%.

[0007] In machine learning, signals or audio cannot be directly input into the classifier for recognition, but need to be converted into interpretable data or feature vectors. Therefore, the basic idea of ​​existing research on the identification of redundant signals and component signals and redundant material identification (referred to as existing research) is to convert signals into data through feature extraction, thereby constructing a data set to train the classifier. In order to achieve higher classification accuracy, in the early stage, scholars focused their research on finding and training the latest suitable classifiers with good performance, and optimized parameters through various means to achieve the best classification performance. Recently, some scholars have begun to shift their attention to the data set for training classifiers and focus their research on the construction of high-quality data sets. They look for suitable signal features with good performance from multiple domains, and use feature optimization methods to retain special signal features to construct high-quality data sets. At present, the classification accuracy of redundant signals and component signals obtained by the classifier exceeds 90%, the classification accuracy of redundant objects of metal materials and non-metallic materials reaches 95%, and the classification accuracy of redundant objects of specific materials exceeds 80%. But they did not shift their attention to the signal for constructing the data set. In fact, high-quality signals determine high-quality data sets, and high-quality data sets determine high-performance classifiers. It can be seen that the signal is the source that affects the classification accuracy, but few studies have paid attention to it. As mentioned above, the purpose of feature extraction is to convert signals into data. Considering that most existing studies directly convert data based on signals, a few convert signals into images and then convert data. This may be because some information will be lost in the process of converting signals to images. Therefore, the focus is also on feature extraction based on signals.

[0008] After a long period of development, the PIND detection equipment developed based on the PIND method is an important tool for conducting redundant object detection research. Representative ones include the 4511 type equipment of SD Company in the United States and the DZJC-III type equipment of Yugao Electronic Technology Co., Ltd. in China. On the basis of meeting the hardware requirements of the PIND method, the PIND detection equipment has added peripherals such as acquisition cards and built-in computers to provide functions such as signal processing, synchronous acquisition, waveform display, and data storage. Therefore, the sound signals collected and saved by the PIND detection equipment are already of high intensity and high quality, and can be directly used for research. In fact, existing research does indeed directly extract features from these sound signals. The acquisition time of these sound signals is fixed, usually five or ten seconds. Because these sound signals are digital signals, the sound signals collected and saved by the PIND detection equipment are actually the amplitudes of the sampling points arranged in sequence. Taking the DZJC-III type equipment as an example, its sampling frequency is 500kHz and the acquisition time is five seconds. Then, the sound signals it collects and saves are arranged in sequence. The amplitude of the sampling points, the interval between two adjacent sampling points Sound signals are composed of useful pulses and noise pulses. Useful pulses refer to pulses that rise and then fall in the shape of a mountain peak, and the amplitudes of the sampling points they contain are all greater than 0. Noise pulses are used to connect useful pulses and are basically horizontal lines. The amplitudes of the sampling points they contain are partly 0 and partly slightly greater than 0, corresponding to the residual background noise in the signal. Feature extraction based on the signal is actually performed on the sampling points. For example, extracting signal features from the time domain is to calculate their energy, variance, etc. based on the amplitude of the sampling points. Extracting signal features from the frequency domain is to perform Fourier transform on the sampling points and then calculate their spectral centroid, frequency standard deviation, etc. The calculation of signal features in other domains is the same.

[0009] As mentioned above, the amplitude of the sampling points included in the noise pulse is partly 0, and the other part is a value slightly greater than 0. The signal features in the multi-domain calculated by them may be meaningless, and 0 or missing values ​​may appear. In addition, from the mechanism of the generation of the redundant signal and the component signal, a useful pulse is generated every time the redundant object collides or slides, or every time the loose component vibrates, and no useful pulses are generated in other non-moving states, but instead noise pulses are generated. In other words, the real redundant signal and component signal are actually only useful pulses. Therefore, it is only effective to extract features at the sampling points included in the useful pulse. At this time, a method is needed to extract useful pulses from the sound signal. In the field of speech recognition, it is called a pulse extraction algorithm, and the most commonly used is the three-threshold pulse extraction algorithm. It calculates the average energy of all the sampling points contained in the sound signal, and sets three thresholds of the start detection threshold, the peak detection threshold, and the end detection threshold on this basis. The peak detection threshold is used to quickly find the highest point of the useful pulse, and then the start detection threshold and the end detection threshold are used to find the start and end points of the useful pulse forward and backward respectively. However, there are no fixed rules for setting the start detection threshold, peak detection threshold, and end detection threshold. They are often based on manual experience and judgment, giving a relatively good result. This requires high professional skills and experience accumulation from the user, and the pulse extraction results are highly random. Most existing studies directly use the three-threshold pulse extraction algorithm to extract useful pulses from sound signals, and the settings of the three thresholds in different studies are also different, providing limited reference value. In addition, the three-threshold pulse extraction algorithm has a serious problem of continuous multiple pulses, and it will detect multiple continuous pulses as one pulse.

[0010] However, there are not many studies specifically targeting pulse extraction algorithms. For example, Gao Leizhen et al. (2019) proposed a new double-threshold pulse extraction algorithm. It calculates the average energy and average zero-crossing rate of the sampling points contained in the beginning of a noise pulse in the sound signal, and sets two thresholds, the endpoint energy threshold and the endpoint zero-crossing rate threshold, on this basis. The endpoint energy threshold is used to first find the starting point and end point of a round of useful pulses, and then the endpoint zero-crossing rate threshold is used to find the starting point and end point of the second round of useful pulses. The final starting point and end point of the useful pulse are determined based on the results of the two rounds of searches. However, the setting of the endpoint energy threshold and the endpoint zero-crossing rate threshold still relies on manual experience judgment, the problem of continuous multiple pulses is not considered, and the problems existing in the three-threshold pulse extraction algorithm have not been solved. Sun Zhigang et al. (2023) proposed a new two-stage double-threshold pulse extraction algorithm, which specifically includes two stages of processing. In the first stage, like the double-threshold pulse extraction algorithm, it determines the starting point and end point of the useful pulse by setting two thresholds, the endpoint energy threshold and the endpoint zero-crossing rate threshold. In the second stage, it resets two new thresholds, the endpoint energy threshold and the endpoint zero-crossing rate threshold, to determine the starting and ending points of the useful pulse again. According to the results of the two stages, the final starting and ending points of the useful pulse are given. This is designed to solve the problem of continuous multiple pulses, and the problem has been alleviated to a certain extent. However, the setting of the endpoint energy threshold and the endpoint zero-crossing rate threshold in the two stages still cannot get rid of manual experience judgment. They do not fundamentally solve the problems existing in the existing pulse extraction algorithms. Other studies on pulse extraction algorithms have not been retrieved in the database.

[0011] In summary, existing research has neglected the focus on sound signal processing, specifically, the focus on the useful pulse extraction algorithm between the sound signal output by the PIND detection device and the feature extraction of the constructed data set. In fact, the more accurate the useful pulse extraction is, the higher the quality of the constructed data set is, and the better the performance of the trained classifier is. It is the source of existing research and needs to be seriously considered. The three-threshold pulse extraction algorithm sets the threshold by referring to the average energy of all sampling points contained in the entire sound signal, and the double-threshold pulse extraction algorithm and the two-level double-threshold pulse extraction algorithm set the threshold by referring to the average energy and average zero-crossing rate of the sampling points contained in the beginning of a noise pulse in the sound signal. Their reference standards are not unified. In addition, the rules for setting the thresholds of the three pulse extraction algorithms are not fixed, and they rely on manual experience judgment, giving unstable and unreliable useful pulse extraction results. Moreover, the three-threshold pulse extraction algorithm and the double-threshold pulse extraction algorithm do not consider the problem of continuous multi-pulses. Although the two-level double-threshold pulse extraction algorithm considers it, the actual effect is not satisfactory. Finally, the three existing pulse extraction algorithms are directly used to extract useful pulses from sound signals. Their useful pulse extraction effects have not been seriously considered, and no relevant evaluation indicators have been used to give specific quantitative results of the pulse extraction effect. Gao Leizhen et al. (2019) and Sun Zhigang et al. (2023) only focused on the number of useful pulses, but this evaluation result is rather one-sided. Summary of the invention

[0012] The present invention aims to solve the problem that the accuracy of the extracted pulse signal needs to be further improved due to the lack of adaptability of the threshold in the existing limited pulse extraction method for redundant object detection.

[0013] A two-stage adaptive multi-threshold pulse extraction method for redundant object detection comprises the following steps:

[0014] Step 1: Calculate the average energy E of the entire sound signal a and the average zero-crossing rate Z a , and the standard deviation of energy σ E and the standard deviation of the zero-crossing rate σ Z ; According to the 3σ distribution criterion of normal distribution, set the peak detection threshold E p , endpoint energy threshold E be and endpoint zero crossing rate threshold Z be ;

[0015] Step 2: Frame the sound signal and calculate the energy and zero-crossing rate of each frame signal;

[0016] Step 3: Starting from the first frame signal, compare the energy of each frame signal with E p ; When the energy of a frame signal is greater than E pWhen the frame signal is recorded, the energy of each subsequent frame signal is compared with E p , until the energy of a frame signal is less than E p ; Find the frame signal with the largest energy from these frame signals, and obtain the time of the first sampling point, which is called the peak time t of the current useful pulse p ;

[0017] Step 4: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the energy of each frame signal is compared with E be Compare until the energy of a frame signal is found to be less than E in both directions. be ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the first candidate start time t of the current useful pulse b_f1 and the first candidate end time t e_f1 ;

[0018] Step 5: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the zero-crossing rate of each frame signal is calculated forward and backward respectively. be Compare until the zero-crossing rate of a frame signal is found to be less than Z in both directions. be ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the second candidate start time t of the current useful pulse b_f2 and the second candidate end time t e_f2 ;

[0019] Step 6: Compare the first candidate start time with the second candidate start time, and determine the larger one as the final start time t b_1 ; Compare the first candidate end time with the second candidate end time, and determine the smaller one as the final end time t e_1 ;

[0020] Step 7: Take the final end time t of the useful pulse extracted in step 6 e_1 The next frame signal is taken as a new starting point, and steps 3 to 6 are repeated to extract the second useful pulse; and so on, until the last frame signal of the sound signal is searched; and the extracted multiple useful pulses are saved;

[0021] Step 8: Calculate the average energy and average zero-crossing rate of each useful pulse, expressed as E i and Z i ; Calculate the standard deviation of the energy of each pulse σE_i and the standard deviation of the zero-crossing rate σ Z_i ; Where i represents the number of the useful pulse; Based on E i and σ E_i Determine the endpoint energy threshold E for each useful pulse be_i , and based on z i and σ Z_i Determine the endpoint zero-crossing rate threshold value of the useful pulse as Z be_i ;

[0022] Step 9: Take the peak time t of a useful pulse saved in step 7 p As the starting point, the energy of each frame signal is compared with the exclusive E be_i Compare until the energy of a frame signal is found to be less than E in both directions. be_i ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the third candidate start time t of the current useful pulse b_s1 and the third candidate end time t e_s1 ;

[0023] Step 10: Take the peak time t of a useful pulse saved in step 7 p As the starting point, the zero-crossing rate of each frame signal is compared with the exclusive Z be_i Compare until the zero-crossing rate of a frame signal is found to be less than Z in both directions. be_i ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the fourth candidate start time t of the current useful pulse b_s2 and the fourth candidate end time t e_s2 ;

[0024] Step 11: Compare the third candidate start time and the fourth candidate start time, and determine the larger one as the final start time t b_2 Similarly, compare the third candidate end time and the fourth candidate end time, and determine the smaller one as the final end time t e_2 ;

[0025] Step 12: starting from the second useful pulse saved in step 7, repeat the processing steps from step 9 to step 11; and so on, until all the useful pulses saved in step 7 are processed;

[0026] Step 13: Calculate the first duration = t e_1 -t b_1 and the second duration = t e_2-t b_2 ;

[0027] Calculate the ratio of the second duration to the first duration; if the ratio exceeds the duration ratio threshold, set the final start time t b_1 and the final ending time t e_1 The real starting time t of the current useful pulse is considered b and the actual end time t e ; If the ratio is lower than the duration ratio threshold, the final start time t b_1 and the final ending time t e_2 The real starting time t of the current useful pulse is considered b and the actual end time t e ;

[0028] Step 14: The useful pulses processed in step 13 are retained, and the useful pulse extraction of the entire sound signal is completed.

[0029] Beneficial Effects

[0030] The present invention shifts the focus of redundant object detection research to pulse extraction for the first time, and proposes a two-level adaptive multi-threshold pulse extraction method, which uses the average energy and average zero-crossing rate of all sampling points contained in the entire sound signal as a reference, and sets the threshold according to the 3σ distribution criterion of the normal distribution, which not only ensures its robustness and generalization; the two-level adaptive extraction method of the present invention can also accurately and stably extract useful pulses in sound signals.

[0031] According to the evaluation of the pulse extraction method of the present invention from multiple perspectives, the present invention can stably extract pulse signals, and the extracted pulses are more accurate and can reflect the signals more realistically, thereby effectively improving the classification accuracy of the classifier, including the classification accuracy of redundant signals and component signals and the classification accuracy of redundant objects of specific materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the three-threshold pulse extraction algorithm.

[0033] Figure 2 Schematic diagram of the dual-threshold pulse extraction algorithm.

[0034] Figure 3 Schematic diagram of the two-stage dual-threshold pulse extraction algorithm.

[0035] Figure 4 Schematic diagram of the continuous multi-pulse problem.

[0036] Figure 5 Schematic diagram of the two-stage adaptive multi-threshold pulse extraction algorithm.

[0037] Figure 6 This is a physical picture of a sealed relay.

[0038] Figure 7 A visual representation of the redundant signal.

[0039] Figure 8 This is the useful pulse extraction effect achieved by the three-threshold pulse extraction method on the redundant signal.

[0040] Fig. 9 This is the useful pulse extraction effect achieved by the double threshold pulse extraction method on the redundant signal.

[0041] Fig.10 This is the useful pulse extraction effect achieved by the two-stage double-threshold pulse extraction method on the redundant signal.

[0042] Fig.11 This is the useful pulse extraction effect achieved by the two-level adaptive multi-threshold pulse extraction method on the redundant signal.

[0043] Fig.12 The component classification accuracy achieved by the four classifiers on the test sets of the four component datasets.

[0044] Fig.13 The material classification accuracy of the four classifiers on the test sets of the four material datasets. DETAILED DESCRIPTION

[0045] This embodiment is a two-stage adaptive multi-threshold pulse extraction method for unwanted object detection. In order to fully illustrate the present invention, this embodiment also compares and illustrates the existing extraction method and the two-stage dual-threshold pulse extraction method.

[0046] A. Three-threshold pulse extraction method:

[0047] The three-threshold pulse extraction algorithm is one of the pulse extraction methods widely used in the field of speech signal processing. Figure 1 A schematic diagram of the three-threshold pulse extraction method is given. The specific process includes the following steps:

[0048] Step 1: Calculate the average energy of the entire sound signal, that is, the average of the sum of the squares of the amplitudes of all sampling points in the sound signal, expressed as E a , called the reference threshold. On this basis, the peak detection threshold E is set p , initial detection threshold E b and the end detection threshold E e These are the three thresholds mentioned in the algorithm name. In the prior art, the peak detection threshold is recommended to be set to E p=3Ea, the starting detection threshold and the ending detection threshold are recommended to be set to the same, that is, E b =E e =1.1E a .

[0049] Step 2: Perform the first frame processing on the sound signal, set the duration of each frame signal Δt1 = 100 μs, that is, each frame signal contains fifty sampling points. In this way, the sound signal is divided into multiple frame signals. Calculate the energy of each frame signal.

[0050] Step 3: Starting from the first frame signal, compare the energy of each frame signal with E p When the energy of a frame signal is greater than E p When , it means that the maximum frame signal of the current useful pulse is about to appear. Starting from this frame signal, continue to compare the energy of each subsequent frame signal with E p Compare until the energy of a frame signal is less than E p Find the frame signal with the largest energy from these frame signals and identify it as the maximum frame signal of the current useful pulse. The time of obtaining the first sampling point in the maximum frame signal is called the peak time t of the current useful pulse. p .

[0051] Step 4: Perform a second frame processing on the sound signal, set the duration of each frame signal to Δt2 = 50 μs, that is, each frame signal contains 25 sampling points, and recalculate the energy of each frame signal.

[0052] In fact, because the duration of the frame signal here is exactly half of the duration of the frame signal in step 2, it is equivalent to perfectly dividing a large interval into two small intervals. In this way, the first sampling point in the maximum frame signal in step 3 will still be the first sampling point in a new frame signal, that is, the peak time remains unchanged.

[0053] Step 5: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the energy of each frame signal is compared with E b Compare the energy of each frame signal with E e Compare until a frame signal energy is found to be less than E in the forward direction. b , the energy of a frame signal is found to be less than E in the backward direction e The previous frame signals of the two frame signals are respectively identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the start time t of the current useful pulse. b and the end time t e .

[0054] Step 6: Take the end time t of the useful pulse extracted in step 5 e The next frame signal is taken as a new starting point, and steps 2 to 5 are repeated to extract the second useful pulse. This process is repeated until the last frame signal of the sound signal is found, and the extraction of all useful pulses is completed.

[0055] In fact, steps 2 and 3 set a larger frame signal to quickly find the peak position of the useful pulse. Although this position may not be absolutely accurate, the difference is not large. This is considered a rough search process. Steps 4 and 5 set a smaller frame signal to accurately find the start and end positions of the useful pulse. This is considered a fine search process.

[0056] B. Dual threshold pulse extraction method:

[0057] Based on the three-threshold pulse extraction algorithm, Gao Leizhen et al. (2019) introduced the zero-crossing rate into the pulse extraction research. By comprehensively considering the characteristics of the energy and zero-crossing rate of the sound signal, appropriate endpoint energy thresholds and endpoint zero-crossing rate thresholds are set, and useful pulses in the sound signal are extracted based on these two thresholds. Figure 2 A schematic diagram of the dual threshold pulse extraction method is given. The specific process includes the following steps:

[0058] Step 1: Calculate the average energy and average zero-crossing rate of the beginning of a noise pulse in the sound signal, that is, the average sum of the squares of the amplitudes of the sampling points contained in this noise pulse, and the average zero-crossing rate of the amplitudes of the sampling points, expressed as E a and Z a , respectively referred to as the reference energy threshold and the reference zero-crossing rate threshold. On this basis, the endpoint energy threshold E is set be and endpoint zero crossing rate threshold z be In the prior art, the endpoint energy threshold is suggested to be set to E be =1.1E a , the endpoint zero crossing rate threshold is recommended to be set to Z be =2Z a .

[0059] Step 2: Process the sound signal by framing, set the duration of each frame signal to Δt=50 μs, that is, each frame signal contains 25 sampling points, and calculate the energy and zero-crossing rate of each frame signal.

[0060] Step 3: Starting from the first frame signal, compare the energy of each frame signal with E be When the energy of a frame signal is greater than E be When t is reached, the frame signal is identified as the candidate start frame signal of the current useful pulse. The time of obtaining the first sampling point in this frame signal is called the candidate start time - tb_c1 Starting from this frame signal, continue to compare the energy of each subsequent frame signal with E be , until the energy of a frame signal is less than E be , identify this frame signal as the candidate end frame signal of the current useful pulse. The time of obtaining the first sampling point in this frame signal is called the candidate end time -t e_c1 .

[0061] Step 4: Starting from the first frame signal, compare the zero-crossing rate of each frame signal with Z be When the zero-crossing rate of a frame signal is greater than Z be When t is 0, the frame signal is identified as the candidate start frame signal of the current useful pulse. The time of obtaining the first sampling point in this frame signal is called the candidate start time t2 b_c2 Starting from this frame signal, continue to compare the zero-crossing rate of each subsequent frame signal with Z be , until the zero-crossing rate of a frame signal is less than Z be , identify this frame signal as the candidate end frame signal of the current useful pulse. The time of obtaining the first sampling point in this frame signal is called the candidate end time t e_c2 .

[0062] Step 5: Compare candidate starting times 1 and 2, and identify the larger one as the final starting time t b Compare candidate end time 1 and 2, and identify the smaller one as the final end time t e .

[0063] Step 6: Take the end time t of the useful pulse extracted in step 5 e The next frame signal of the sound signal is taken as a new starting point, and steps 3 to 5 are repeated to extract the second useful pulse. This process is repeated until the last frame signal of the sound signal is found, and the extraction of all useful pulses is completed.

[0064] In fact, step three uses the energy of the frame signal to find the start and end times of the useful pulse in the first round, and step four uses the zero-crossing rate of the frame signal to find the start and end times of the useful pulse in the second round. If the combination of the start and end times found in the two rounds is regarded as an interval, the purpose of step five is actually to find the union of the two intervals. The interval corresponding to the union may be smaller than the original two intervals, but it must be the most core, which means that the sampling points contained in its corresponding time are most likely to be useful pulses.

[0065] C. Two-level dual-threshold pulse extraction method:

[0066] The three thresholds set in the three-threshold pulse extraction method are actually two thresholds, because the start detection threshold and the end detection threshold are the same. Therefore, in the study of Sun Zhigang et al. (2023), they called the start detection threshold and the end detection threshold the endpoint detection threshold, which is regarded as the first threshold, and retained the original peak detection threshold as the second threshold, thus proposing a two-stage dual-threshold pulse extraction method. The algorithm includes two stages of processing. Figure 3 A schematic diagram of a two-stage dual-threshold pulse extraction method is given, and its specific implementation steps are as follows:

[0067] Step 1: Calculate the average energy of the entire sound signal, that is, the average of the sum of the squares of the amplitudes of all sampling points in the sound signal, expressed as E a , called the reference threshold. On this basis, the peak detection threshold E is set p , endpoint detection threshold E be_1 In the prior art, the peak detection threshold is suggested to be set to E p =1.5E a , which is less than the value set in the three-threshold pulse extraction method, the endpoint detection threshold is recommended to be set to E be_1 =1.1E a .

[0068] Step 2: Perform the first frame processing on the sound signal, set the duration of each frame signal Δt1 = 100 μs, that is, each frame signal contains fifty sampling points. Calculate the energy of each frame signal.

[0069] Step 3: Starting from the first frame signal, compare the energy of each frame signal with E p When the energy of a frame signal is greater than E p When the frame signal is recorded, the energy of each subsequent frame signal is compared with E p , until the energy of a frame signal is less than E p Find the frame signal with the largest energy from these frame signals and obtain the time of the first sampling point, which is called the peak time t of the current useful pulse. p .

[0070] Step 4: Perform a second frame processing on the sound signal, set the duration of each frame signal to Δt2 = 50 μs, that is, each frame signal contains 25 sampling points, and recalculate the energy of each frame signal.

[0071] Step 5: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, compare the energy of each frame signal with E be_1 Until the energy of a frame signal is found to be less than E in both directions be_1The previous frame signals of the two frame signals are identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the candidate start time t of the current useful pulse. b_1 and the candidate end time t e_1 .

[0072] At this point, the processing of the first stage is finished, that is, the first-stage double-threshold pulse extraction method is finished. In fact, the processing process from step 1 to step 5 is the main processing process of the three-threshold pulse extraction method.

[0073] Step 6: Set a new endpoint detection threshold E be_2 =1.3E a . Continue with the peak time t of the useful pulse extracted in step 3 p As the starting point, compare the energy of each frame signal with E be_2 Until the energy of a frame signal is found to be less than E in both directions be_2 The previous frame signals of the two frame signals are identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the candidate start time t of the current useful pulse. b_2 and the candidate end time t e_2 .

[0074] Step 7: Compare the two candidate start times and identify the larger one as the final start time t b Similarly, compare the two candidate end times and identify the smaller one as the final end time t e .

[0075] Step 8: Take the end time t of the useful pulse extracted in step 7 e The next frame signal of the sound signal is taken as a new starting point, and steps 2 to 7 are repeated to extract the second useful pulse. This process is repeated until the last frame signal of the sound signal is found, and the extraction of all useful pulses is completed.

[0076] At this point, the second stage of processing is completed, that is, the two-stage double-threshold pulse extraction method is completed. In fact, the processing process of step six is ​​the main processing process of finding the starting point and ending point of the useful pulse in the three-threshold pulse extraction method, except that a larger endpoint detection threshold is replaced.

[0077] In fact, if Figure 3As shown, the processing in the second stage actually raises the entry threshold of the starting point and the ending point, and it seems that this effort is unnecessary. Perhaps the same purpose can be achieved by directly setting a larger endpoint detection threshold at the beginning, but this is not correct. Because the processing in the second stage is actually to solve the problem of continuous multiple pulses. The two-level dual-threshold pulse extraction method takes into account the problem of continuous multiple pulses for the first time, but the solution does not seem to be so smart, and it simply increases the endpoint detection threshold. The present invention provides a solution with reference value.

[0078] D. Two-level adaptive multi-threshold pulse extraction method of the present invention

[0079] In the three-threshold pulse extraction method, the average energy of the entire sound signal is used as the reference threshold, on which the peak detection threshold, the start detection threshold and the end detection threshold are set. They are the basis for judging useful pulses, but this basis is the artificial experience judgment given by scholars based on a large number of experiments, which has no theoretical basis and its reference value is not high. Moreover, in existing studies, the start detection threshold and the end detection threshold are actually the same. This makes the three-threshold pulse extraction method unworthy of its name, because only two thresholds actually work. In addition, the three-threshold pulse extraction method only extracts one round of useful pulses, and does not take into account the problem of continuous multiple pulses. These are the three shortcomings of the three-threshold pulse extraction method, which are also what the present invention needs to solve. However, although the setting of the three thresholds is based on artificial experience judgment, the calculation of the reference threshold is based on the average energy of the entire signal, which is relatively stable and representative, and can reflect the overall distribution level of the sound signal. Moreover, in the process of extracting useful pulses, the three-threshold pulse extraction method first finds the peak value of each useful pulse, and then finds their starting point and end point. It is worth noting that the peak detection threshold is much larger than the start detection threshold and the end detection threshold. Therefore, the amplitude of the useful pulses extracted is relatively large. Those pulses with relatively small amplitudes that may come from the background noise will be filtered out. In other words, the three-threshold pulse extraction method has a great advantage in finding the peak of useful pulses, which means that the useful pulses it extracts are more reliable.

[0080] In the two-stage dual-threshold pulse extraction method, the processing of the first stage is the same as that of the three-threshold pulse extraction method, and the processing of the second stage is actually a repetition of the processing of the first stage, except that the endpoint detection threshold, that is, the start detection threshold and the end detection threshold, are adjusted. However, it solves the two shortcomings of the three-threshold pulse extraction method very well, that is, it recognizes that only two thresholds are used, and uses two-stage processing to solve the problem of continuous multiple pulses. Unfortunately, in the two-stage dual-threshold pulse extraction method, the peak detection threshold and the endpoint detection threshold are also set based on artificial experience judgment, and this problem has not been improved. Therefore, in the present invention, the problem of setting thresholds based on artificial experience judgment needs to be taken seriously, and using two-stage processing to solve the problem of continuous multiple pulses is worthy of reference.

[0081] In the double threshold pulse extraction method, the average energy and average zero-crossing rate of the beginning of a noise pulse in the sound signal are used as the reference energy threshold and the reference zero-crossing rate threshold. On this basis, the endpoint energy threshold and the endpoint zero-crossing rate threshold are set. They are the basis for judging useful pulses, and this basis is also based on artificial experience judgment given by a large number of experiments. This is the same problem that the double threshold pulse extraction method has with the above two pulse extraction methods. At the same time, the problem of continuous multiple pulses is not considered in the double threshold pulse extraction method. In addition, the calculation of the reference energy threshold and the reference zero-crossing rate threshold is based on the beginning of a noise pulse in the sound signal, which is unstable and unrepresentative. Because they can only reflect the distribution level of this noise pulse, they can barely reflect the distribution level of all noise pulses in the sound signal, but cannot represent the overall distribution level of the sound signal. In addition, if accidental noise appears at the beginning of the sound signal, it will appear as an abrupt pulse with a large amplitude. The calculated reference energy threshold and reference zero-crossing rate threshold are relatively large, which does not conform to the overall distribution level of the sound signal, and seriously affects the extraction of useful pulses. These are the three shortcomings of the dual threshold pulse extraction method, among which the third one is different from the above two pulse extraction methods and needs to be considered extra. In the field of speech signal processing, the zero crossing rate refers to the number of times the amplitude of the sampling point in the frame signal changes from positive to negative or from negative to positive. For the frame signal of the main part of the useful pulse, that is, the frame signal of the middle part, the amplitude of the sampling points they contain is positive, and there will be no phenomenon of changing from positive to negative or from negative to positive, and the zero crossing rate is 0. Only those frame signals located at the edge of the useful pulse, that is, the frame signals at the two ends, the amplitude of the sampling points they contain will intermittently change from positive to negative or from negative to positive, and the zero crossing rate will be greater than 0. It can be seen that the zero crossing rate can play an important role in finding the starting point and end point of the useful pulse. In the dual threshold pulse extraction method, we can regard the process of extracting useful pulses using the endpoint energy threshold as rough, and the process of extracting useful pulses using the endpoint zero crossing rate threshold as fine. Therefore, the dual threshold pulse extraction method has great advantages in finding the endpoints of useful pulses.

[0082] As mentioned above, the three-threshold pulse extraction method and the double-threshold pulse extraction method do not take into account the problem of continuous multiple pulses. Figure 4 In the three-threshold pulse extraction method, after the peak detection threshold is used to find the maximum frame signal of a useful pulse, the end detection threshold E e The end frame signal of the useful pulse found may be the end frame signal of the next useful pulse, not the real end frame signal of the current useful pulse, as shown in the red box in the figure. The reason is that the signal attenuation between the two useful pulses does not reach the set end detection threshold. Similarly, in the dual threshold pulse extraction method, because the set endpoint energy threshold Ebe and the end detection threshold E e They play the same role and there is little difference between them. Therefore, when the endpoint energy threshold is used to extract useful pulses. It will also regard the two useful pulses in the red box as one useful pulse. In this way, when a comprehensive analysis is performed on the useful pulse extraction results obtained using the endpoint energy threshold and the endpoint zero-crossing rate threshold, if the endpoint zero-crossing rate threshold plays a dominant role, the problem of continuous multiple pulses may be avoided. However, when the endpoint energy threshold plays a dominant role, the problem of continuous multiple pulses is inevitable. Therefore, it is taken seriously in the processing of the present invention. It should be noted that for the purpose of visual display, Figure 4 What is shown is a continuous analog signal, not a discrete digital signal. In fact, it can be viewed as a curve connecting multiple dense sampling points.

[0083] In summary, in the present invention, the deficiencies of the three existing pulse extraction methods are fully and comprehensively considered, and their advantages are referred to, thereby proposing a two-stage adaptive multi-threshold pulse extraction method. Similarly, it includes two stages of processing, and multiple thresholds are set in each stage of processing to achieve different purposes. The purpose of the first stage of processing is to extract useful pulses with a high degree of credibility. Specifically, the average energy and average zero-crossing rate of the entire sound signal are calculated, and the peak detection threshold, the endpoint energy threshold and the endpoint zero-crossing rate threshold are set. Through the peak detection threshold, the maximum frame signal in the useful pulse is quickly found, and through the endpoint energy threshold and the endpoint zero-crossing rate threshold, the starting frame signal and the ending frame signal of the useful pulse are preliminarily determined. The purpose of the second stage of processing is to solve the possible continuous multi-pulse problem and ensure that the extracted useful pulses are independent individuals. Specifically, the processing results of the first stage are referenced, and the new endpoint energy threshold and the endpoint zero-crossing rate threshold are set. Through the endpoint energy threshold and the endpoint zero-crossing rate threshold, the starting frame signal and the ending frame signal of the useful pulse are determined again. The processing results of the two stages are comprehensively analyzed to determine the final starting frame signal and the ending frame signal of the useful pulse. It is worth noting that the setting of the peak detection threshold, endpoint energy threshold and endpoint zero-crossing rate threshold is no longer based on manual experience judgment, but refers to the 3σ principle of normal distribution.

[0084] According to the 3σ principle, the probability of the value being distributed within [μ-σ, μ+σ] is 68.27%, and the probability of the value being distributed within [μ-3σ, μ+3σ] is 99.73%. In other words, data with values ​​exceeding μ+3σ are considered outliers, that is, values ​​that are particularly large, while data with values ​​hovering around μ+σ are considered valid values. In other words, for a sound signal, when calculating its average energy E a , average zero-crossing rate Z a , the standard deviation of energy σ E and the standard deviation of the zero-crossing rate σZ Then, we can a +3σ E Find the maximum frame signal of the useful pulse in the frame signal, and when the energy hovers around E a +σ E The zero crossing rate hovers around Z a +σ Z The start frame signal and the end frame signal of the useful pulse are respectively searched in the nearby frame signals. Figure 5 A schematic diagram of a two-stage adaptive multi-threshold pulse extraction method is given, and its specific implementation steps are as follows:

[0085] Step 1: Calculate the average energy and average zero-crossing rate of the entire sound signal, expressed as E a and Z a , calculate the standard deviation of the energy and the standard deviation of the zero-crossing rate of the entire sound signal, expressed as σ E and σ Z On this basis, the peak detection threshold is set to E p =E a +3σ E , set the endpoint energy threshold to E be =E a +σ E , set the endpoint zero crossing rate threshold to Z be =Z a +σ Z .

[0086] Step 2: Frame the sound signal and set the duration of each frame signal to Δt = 50μs. In fact, it is to use Δt as a window sliding frame, and the starting position of the next frame signal is the end position of the previous frame signal. Calculate the energy and zero-crossing rate of each frame signal.

[0087] Each frame signal (50 μs) contains twenty-five sampling points.

[0088] Step 3: Starting from the first frame signal, compare the energy of each frame signal with E p When the energy of a frame signal is greater than E p When the frame signal is recorded, the energy of each subsequent frame signal is compared with E p , until the energy of a frame signal is less than E p Find the frame signal with the largest energy from these frame signals and obtain the time of the first sampling point, which is called the peak time t of the current useful pulse. p .

[0089] Step 4: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the energy of each frame signal is compared with Ebe Compare until the energy of a frame signal is found to be less than E in both directions. be The previous frame signals of the two frame signals are identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the first candidate start time t of the current useful pulse. b_f1 and the first candidate end time t e_f1 .

[0090] Step 5: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the zero-crossing rate of each frame signal is calculated forward and backward respectively. be Compare until the zero-crossing rate of a frame signal is found to be less than Z in both directions. be The previous frame signals of the two frame signals are identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the second candidate start time t of the current useful pulse. b_f2 and the second candidate end time t e_f2 .

[0091] Step 6: Compare the first candidate start time with the second candidate start time, and determine the larger one as the final start time t b_1 Similarly, compare the first candidate end time with the second candidate end time, and identify the smaller one as the final end time t e_1 .

[0092] Step 7: Take the final end time t of the useful pulse extracted in step 6 e_1 The next frame signal is taken as a new starting point, and steps 3 to 6 are repeated to extract the second useful pulse. This process is repeated until the last frame signal of the sound signal is found. The extracted multiple useful pulses are saved.

[0093] At this point, the first stage of processing is completed, that is, the first-level adaptive multi-threshold pulse extraction method is completed.

[0094] Step 8: Calculate the average energy and average zero-crossing rate of each useful pulse, expressed as E i and Z i Calculate the standard deviation of the energy and the standard deviation of the zero-crossing rate of each pulse, denoted as σ E_i and σ Z_i Where i represents the serial number of the useful pulse. On this basis, the endpoint energy threshold exclusive to each useful pulse is set to E be_i =E i +σ E_i , set the endpoint zero crossing rate threshold to Z be_i =Z i+σ Z_i .

[0095] Step 9: Take the peak time t of a useful pulse saved in step 7 p As the starting point, the energy of each frame signal is compared with the exclusive E be_i Compare until the energy of a frame signal is found to be less than E in both directions. be_i The previous frame signals of the two frame signals are identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the third candidate start time t of the current useful pulse. b_s1 and the third candidate end time t e_s1 .

[0096] Step 10: Take the peak time t of a useful pulse saved in step 7 p As the starting point, the zero-crossing rate of each frame signal is compared with the exclusive Z be_i Compare until the zero-crossing rate of a frame signal is found to be less than Z in both directions. be_i The previous frame signals of the two frame signals are identified as the start frame signal and the end frame signal of the current useful pulse. The time of obtaining the first sampling point in the two frame signals is called the fourth candidate start time t of the current useful pulse. b_s2 and the fourth candidate end time t e_s2 .

[0097] Step 11: Compare the third candidate start time and the fourth candidate start time, and determine the larger one as the final start time t b_2 Similarly, compare the third candidate end time and the fourth candidate end time, and determine the smaller one as the final end time t e_2 .

[0098] Step 12: Starting from the second useful pulse saved in step 7, repeat the processing steps from step 9 to step 11. And so on, until all the useful pulses saved in step 7 are processed.

[0099] Step 13: Calculate the first duration = t e_1 -t b_1 and the second duration = t e_2 -t b_2 , that is, the difference between the end time and the start time in step six and the difference between the end time and the start time in step eleven, to obtain the duration of the useful pulse before and after each processing.

[0100] Calculate the ratio of the second duration to the first duration. If the ratio exceeds 65%, set the final starting time t in step 6 to b_1and the final ending time t e_1 The real starting time t of the current useful pulse is considered b and the actual end time t e If the ratio is lower than 65%, the final starting time t in step 6 is b_1 and the final end time t in step 11 e_2 The real starting time t of the current useful pulse is considered b and the actual end time t e .

[0101] Step 14: The useful pulses processed in step 13 are retained, and the useful pulse extraction of the entire sound signal is completed.

[0102] At this point, the second stage of processing is completed, that is, the two-stage adaptive multi-threshold pulse extraction method is completed.

[0103] It can be seen from the above description that in the first stage of the processing, the present invention first uses the peak detection threshold to extract those useful pulses with a high degree of credibility, which effectively discards the pulses with a small amplitude in the sound signal, and then uses the endpoint energy threshold and the endpoint zero-crossing rate threshold to accurately find the starting point and the end point of the useful pulse, which effectively ensures that the main signal components in the useful pulse are extracted, and the components of the noise pulse are discarded as much as possible at both ends. Most importantly, the peak detection threshold, the endpoint energy threshold and the endpoint zero-crossing rate threshold are adaptively set with reference to the overall distribution characteristics of the sound signal and the 3σ principle of the normal distribution, with a credible mathematical theoretical basis, and get rid of manual experience judgment. In the second stage of the processing, the average energy, average zero-crossing rate, standard deviation of energy and standard deviation of zero-crossing rate of each useful pulse saved after the first stage of processing are first calculated, which actually reflects the overall distribution characteristics of each useful pulse. On this basis, the peak detection threshold, endpoint energy threshold and endpoint zero-crossing rate threshold exclusive to each useful pulse are set again with reference to the 3σ principle of the normal distribution. In this way, the starting point and the end point of each useful pulse saved after the first stage of processing are re-searched. If the useful pulse is indeed an independent individual, then referring to the 3σ principle of normal distribution, the duration of the useful pulse after the second stage of processing should be 68.27% of that before processing. If the useful pulse is a continuous multi-pulse, that is, a combination of multiple useful pulses, then the calculated average energy, average zero-crossing rate, standard deviation of energy and standard deviation of zero-crossing rate actually reflect their comprehensive distribution characteristics, and are no longer the overall distribution characteristics exclusive to a certain useful pulse. In this way, the useful pulse with a larger amplitude dominates the comprehensive distribution characteristics, resulting in the four calculated values ​​being larger, all greater than the four values ​​calculated based only on the useful pulse with a smaller amplitude. Therefore, the second stage of processing can find the useful pulse with a smaller amplitude in the continuous multi-pulse and discard it. In this way, the continuous multi-pulse after the second stage of processing not only discards the frame signal accounting for 31.73% of itself, but also discards the frame signal of the useful pulse with a smaller amplitude, which results in the ratio of the duration of the useful pulse before and after processing being much lower than 68.27%. This can effectively distinguish independent useful pulses from continuous multi-pulses, indicating that the judgment method adopted is feasible and has a credible mathematical theoretical basis. In fact, in the present invention, the ratio of the duration of the useful pulse before and after the processing is set to 65%, which reserves a certain flexibility and tolerates the occurrence of extreme situations.

[0104] Example

[0105] A sample of sealed electronic components is produced to collect and save a large number of component signals and redundant signals generated by redundant objects of different materials. On this basis, the three existing pulse extraction methods and the two-level adaptive multi-threshold pulse extraction method proposed in the present invention are used to synchronously extract pulses to obtain four sets of multiple processed useful pulses. The useful pulse extraction effects of the four pulse extraction methods are comprehensively analyzed from three perspectives, including the number of pulses and distribution density from a mathematical perspective, the signal-to-noise ratio and normalized signal-to-noise ratio from a signal perspective, and two classification accuracies from a machine learning perspective, to fully verify the feasibility and superiority of the proposed two-level adaptive multi-threshold pulse extraction method.

[0106] 1 Evaluation Metrics

[0107] The three-threshold pulse extraction method is a classic and widely used pulse extraction method in the field of speech signal processing. The evaluation of its feasibility and practicality mostly comes from the observation of users, that is, observing the number, amplitude, distribution, etc. of useful pulses extracted on the visualized waveform. So far, no special study has given specific quantitative results. Similarly, in the study of Gao Leizhen et al. (2019), they only used the evaluation indicator of the number of pulses to compare the useful pulse extraction effects of the proposed double-threshold pulse extraction method and the three-threshold pulse extraction method. In fact, the reference value of this evaluation result is not high. Because if more noise pulses are extracted, the quality of all pulses will be reduced, which is not as good as extracting a slightly smaller number of useful pulses with high credibility. In the study of Sun Zhigang et al. (2023), they directly used the proposed two-stage double-threshold pulse extraction method for processing, and did not give the evaluation results of the useful pulse extraction effect. Therefore, in the present invention, the present invention considers quantitative evaluation of the useful pulse extraction effect of the pulse extraction method for the first time, and proposes a comprehensive analysis from the perspective of mathematics, signal and machine learning to give a comprehensive evaluation result.

[0108] From a mathematical point of view, the number of pulses and distribution density are mainly selected. The former is used to evaluate the number of useful pulses extracted from the sound signal by the pulse extraction method, and the latter is used to evaluate the root mean square error of the extracted useful pulses themselves and their distribution in the sound signal. The number of pulses is essentially the number of a certain object and is an existing evaluation indicator. Therefore, the present invention does not provide a detailed definition of it. Distribution density is a new evaluation indicator proposed by the present invention, which consists of two parts. The first part is to calculate the root mean square error of the useful pulse itself. Assume that a useful pulse contains N sampling points, and each sampling point has an amplitude. The amplitudes of the N sampling points are respectively represented as A i (i=1,2,…,N), μ is the mean value of the amplitude of N sampling points. The root mean square error RMSE of the useful pulse itself is calculated by the following formula:

[0109]

[0110] On this basis, the second part is to calculate the distribution of useful pulses in the sound signal. Assuming that M useful pulses are extracted from the sound signal, the root mean square error of each useful pulse is expressed as RMSE m (m=1,2,…,M), the distribution density DBDS of these useful pulses is calculated by the following formula:

[0111]

[0112] From the calculation process of the distribution density, it can be seen that the calculation results of the first part reflect the concentration of each useful pulse, that is, the proportion of the main signal components in the useful pulse. The higher the proportion of the main signal components in the useful pulse, the higher the concentration, and the smaller the calculated root mean square error RMSE. The calculation results of the second part reflect the average distribution level of the root mean square error of all useful pulses in the sound signal. Therefore, the smaller the distribution density, the higher the concentration of the useful pulses extracted from the sound signal, the more stable the average distribution level, and the better the useful pulse extraction effect of the corresponding pulse extraction method.

[0113] From the signal perspective, the present invention mainly selects the signal-to-noise ratio and the normalized signal-to-noise ratio. The former is used to evaluate the signal-to-noise ratio of the sound signal, and the latter is used to evaluate the signal-to-noise ratio of each useful pulse. In the field of speech signal processing, the definition of the signal-to-noise ratio is as follows. Assume that the noisy sound signal is f(t)=s(t1)+n(t2), where s(t1) and n(t2) represent the speech signal and noise, respectively. The signal-to-noise ratio SNR refers to the ratio of the speech signal to the noise in a sound signal, and its calculation process is as follows:

[0114]

[0115] Among them, P s Represents the average power of the speech signal, P n Represents the average power of the noise. Their calculation process is as follows:

[0116]

[0117]

[0118] Wherein, N represents the number of sampling points contained in the noisy audio signal, and M represents the number of sampling points contained in the speech signal.

[0119] The greater the signal-to-noise ratio of a sound signal, the greater the proportion of speech signals therein, and the better the quality. In the sound signal of the present invention, speech signals are all useful pulses, and noise is noise pulses. However, from the perspective of formula (3), the smaller the numerator, or the larger the denominator, the greater the calculated signal-to-noise ratio SNR. In this way, if a part of the noise pulses is mistakenly regarded as useful pulses, the numerator becomes larger while the denominator becomes smaller, and the signal-to-noise ratio also increases. This is undoubtedly incorrect. Therefore, in the present invention, the calculation process of the average power of the noise is adjusted, as shown in the following formula:

[0120]

[0121] Wherein, M represents the number of sampling points contained in the speech signal.

[0122] That is, for a sound signal, the pulse extraction method is used to extract the useful pulses therein, and the number of sampling points contained in all the useful pulses is calculated. In this way, the remaining sampling points in the sound signal are determined to be from noise pulses. Therefore, the present invention calculates the average amplitude of the remaining sampling points as the normalized amplitude of the noise pulses in this sound signal. On this basis, P s Calculated from the sampling points included in all useful pulses, P n It is obtained by accumulating the normalized amplitude with the same number of sampling points as the useful pulse. It can be seen that the calculation result of the proposed applicable improved signal-to-noise ratio is more fair and reliable.

[0123] From the description of the improved signal-to-noise ratio above, it can be seen that the useful pulse is regarded as a speech signal and the noise pulse is regarded as noise. From the original definition of the signal-to-noise ratio, this is not entirely correct, because the noise pulse is actually part of the sound signal. Therefore, referring to the traditional definition of the signal-to-noise ratio, a new evaluation index, the normalized signal-to-noise ratio, is proposed. Specifically, for a sound signal, the pulse extraction method is used to extract the useful pulses therein, and the number of sampling points M contained in all the useful pulses is calculated. Refer to formula (4) to calculate their average power P s . Next, the present invention adds Gaussian white noise to the sound signal, and calculates the average power of all sampling points contained in the sound signal, and the calculation process is as shown in formula (7). It should be noted that, at this time, the sound signal before adding Gaussian white noise is regarded as the speech signal s(t1), the Gaussian white noise is regarded as the noise n(t2), and the sound signal after adding Gaussian white noise is regarded as the noisy sound signal f(t).

[0124]

[0125] Wherein, N represents the number of sampling points contained in the noisy audio signal.

[0126] On this basis, referring to the number of sampling points M contained in the useful pulse, the ratio of the number of sampling points M contained in the sound signal is calculated to obtain the normalized average power of the noisy sound signal. The calculation process is as follows:

[0127]

[0128] Wherein, N represents the number of sampling points contained in the noisy sound signal, and M represents the number of sampling points contained in the useful pulse.

[0129] In this way, the normalized signal-to-noise ratio can be obtained by the average power P of the sampling points contained in all useful pulses. s and the normalized average power P of the noisy sound signal f_Nor The calculation process is as follows:

[0130]

[0131] It can be seen that Gaussian white noise is introduced in the definition process of the normalized signal-to-noise ratio, which is similar to the definition process of the traditional signal-to-noise ratio. Compared with the improved signal-to-noise ratio proposed above, the difference between equation (3) and equation (9) lies in the denominator. The denominator of the normalized signal-to-noise ratio is the average power of the noisy sound signal with the same proportion intercepted. The reason is the same as the normalized amplitude of the noise pulse with the same length intercepted in the improved signal-to-noise ratio. In the field of speech signal processing, the larger the signal-to-noise ratio, the higher the signal quality. Similarly, the larger the improved signal-to-noise ratio, or the larger the normalized signal-to-noise ratio, the higher the quality of the speech signal, that is, the higher the quality of the extracted useful pulse, and the better the useful pulse extraction effect of the corresponding pulse extraction method.

[0132] From the perspective of machine learning, two classification accuracies are mainly selected, specifically component classification accuracy and material classification accuracy. As described in the background technology, feature extraction is an important means to complete the conversion of sound signals into data sets, and feature extraction is performed on the basis of the sampling points contained in the useful pulses. Therefore, for sound signals, the present invention uses a pulse extraction method to extract useful pulses therein, calculates the characteristic values ​​of common signal features in the time domain and frequency domain on each useful pulse, and converts the useful pulses into characteristic vectors, thereby constructing a data set. The data set is divided into a training set and a test set in a ratio of 7:3, the classifier is trained on the training set, and the classification performance of the classifier is tested on the test set, that is, the classification accuracy obtained is counted. In this way, for the same sound signal, the three existing pulse extraction methods and the two-level adaptive multi-threshold pulse extraction method proposed by the present invention are used to synchronously extract pulses, and four sets containing multiple processed useful pulses are obtained. The characteristic values ​​of the same signal features are calculated on multiple useful pulses in each set, thereby constructing four data sets. The four data sets are divided into training sets and test sets in a ratio of 7:3, the same classifier is trained on the four training sets, the classification performance of the classifier is tested on the corresponding test sets, and the four classification accuracies obtained by them are counted. In this way, the classification performance of the classifier with higher classification accuracy is the best, which means that the quality of the training set used to train it is the highest, that is, the quality of the data set from which the training set comes is the highest. Furthermore, the quality of the multiple useful pulses used to construct this data set is the highest, and the corresponding pulse extraction method is the best.

[0133] Assume that the data set is D = {(x1, y1), (x2, y2), ..., (x n ,y n )}, where y i is the data x i The true label, R(x i ) is the predicted label given by the classifier R. The classification accuracy is expressed as the ratio of the number of data with correct label prediction to the total number of data, and its calculation process is as follows:

[0134]

[0135] Among them, i represents the serial number of the data in the data set, I is the indicator function, when R(x i )=y i When I(R(x i )=y i )=1.

[0136] As described in the background technology, the research that scholars in the field of redundant object detection are currently paying more attention to includes the research on the identification of redundant object signals and component signals, and the research on the identification of redundant object materials. For the former, a large number of redundant object signals and component signals can be collected, and a data set containing two labels can be constructed through pulse extraction, feature extraction and other processing. On this basis, a classifier is trained and the classification accuracy obtained is counted. This classification accuracy is called component classification accuracy, indicating that it is used to evaluate the classification performance obtained by the classifier on the data set constructed by the redundant object signal and the component signal. For the latter, a large number of redundant object signals generated by redundant objects of different materials can be collected, and a data set containing multiple labels can be constructed through pulse extraction, feature extraction and other processing. On this basis, a classifier is trained and the classification accuracy obtained is counted. This classification accuracy is called material classification accuracy, indicating that it is used to evaluate the classification performance obtained by the classifier on the data set constructed by the redundant object signals generated by redundant objects of different materials.

[0137] 2 Sound signal

[0138] The sealed relay is a kind of sealed electronic component and is also one of the most commonly used sealed electronic components in aerospace electronic equipment. Therefore, the present invention customizes a batch of sealed relays in a dedicated aerospace institute, including sealed relays that only contain loose components but no redundant objects, and sealed relays that only contain redundant objects but no loose components. Figure 6 The sealed relays used are given. The six sealed relays in the first row each contain a loose component but no excess components. The loose components are mainly bonding wires of different lengths or looseness. The six sealed relays in the second row each contain an excess of one material but no loose components. The six materials of excess are copper wire, solder particles, aluminum particles, hot melt adhesive particles, PVC particles and silicone particles.

[0139] With the help of DzJC-III equipment of China Yugao Electronic Technology Co., Ltd., the present invention conducts a large number of experiments on each sealed relay, respectively, and collects and saves one hundred sound signals, specifically, collects and saves six hundred component signals and six hundred superfluous object signals, wherein the six hundred superfluous object signals are composed of one hundred superfluous object signals corresponding to six kinds of materials, respectively.

[0140] 3 Evaluation of useful pulse extraction effect

[0141] A. Visual display:

[0142] The 1,200 sound signals collected and saved were processed using the three-threshold pulse extraction method, the double-threshold pulse extraction method, the two-level double-threshold pulse extraction method, and the two-level adaptive multi-threshold pulse extraction method, respectively, to obtain and save multiple useful pulse extraction results. Taking a randomly selected redundant signal as an example, the useful pulse extraction effects of the four pulse extraction methods are visualized. Figure 7 The waveform of this section of the redundant signal, as well as the calculated zero-crossing rate and energy are given. As can be seen from the figure, the zero-crossing rate and energy corresponding to the noise pulse are relatively stable, without large fluctuations. When a useful pulse appears, its corresponding zero-crossing rate shows obvious fluctuations of decreasing and increasing trends, which are manifested as light and dark peaks and contours, while its corresponding energy shows obvious fluctuations of increasing trends, which are manifested as continuous undulating peaks. This fully demonstrates the advantages of zero-crossing rate and energy in finding the starting and ending points of useful pulses.

[0143] Figure 8-Figure 11 The useful pulse extraction effects of the four pulse extraction methods on this section of the redundant signal are given respectively. In the figure, a red short vertical line and a blue short vertical line adjacent to each other are regarded as a combination, which represent the extraction of a useful pulse. It can be seen from the figure that the three-threshold pulse extraction method achieves a good useful pulse extraction effect on this section of the redundant signal. Most useful pulses are extracted, especially all useful pulses with a relatively large dispersion are extracted. Unfortunately, the useful pulses with a relatively large concentration between 2 seconds and 2.5 seconds cannot be extracted independently, that is, the problem of continuous multiple pulses cannot be solved. The same problem also occurs on several continuous useful pulses after 4.5 seconds. The double-threshold pulse extraction method achieves the worst useful pulse extraction effect on this section of the redundant signal, showing an overfitting phenomenon, that is, too many useful pulses are extracted, and even some tiny noise pulses are extracted. This is because the double-threshold pulse extraction method only detects the starting point and the end point of the useful pulse. Once a pulse shows a fluctuation trend, even if the amplitude is very small, the double-threshold pulse extraction method can capture it sensitively. This reflects the importance of finding the peak of the useful pulse first in the three-threshold pulse extraction method. In addition, the problem of continuous multiple pulses has not been solved. Compared with the double-threshold pulse extraction method, the two-stage double-threshold pulse extraction method has improved the useful pulse extraction effect on this section of redundant signal. Some continuous multiple pulses with longer lengths are divided, but the divided ones are still continuous multiple pulses with slightly smaller lengths, and the overall effect is still not good. Moreover, the problem of noise pulse extraction has not been solved. In addition, because the peak detection threshold is set to a small value, it does not inherit the advantages of the three-threshold pulse extraction method, but tends to the double-threshold pulse extraction method, retaining only the weak advantage of two-stage processing.

[0144] In contrast, the proposed two-level adaptive multi-threshold pulse extraction method achieves the best useful pulse extraction effect on this section of the redundant signal. Basically, all useful pulses are extracted and noise pulses are ignored. Most importantly, it shows a good processing effect on the continuous multi-pulse problem that the other three pulse extraction methods have not been able to solve. The useful pulses with a relatively high concentration between 2 seconds and 2.5 seconds in the redundant signal are extracted independently, and the same result is obtained for several consecutive useful pulses after 4.5 seconds. This preliminarily illustrates the feasibility and superiority of the two-level adaptive multi-threshold pulse extraction method proposed in the present invention. In the following description, the present invention will quantitatively demonstrate the useful pulse extraction effects of the four pulse extraction methods from three perspectives to further verify the proposed two-level adaptive multi-threshold pulse extraction method.

[0145] B. Evaluation from a mathematical perspective:

[0146] From a mathematical point of view, the number of pulses and distribution density are mainly selected to evaluate the useful pulse extraction effect of the four pulse extraction methods. The first is the number of pulses. For a section of the redundant signal shown in A, after statistics, the three-threshold pulse extraction method extracted a total of 85 useful pulses, the double-threshold pulse extraction method extracted a total of 172 useful pulses, the two-level double-threshold pulse extraction method extracted a total of 183 useful pulses, and the two-level adaptive multi-threshold pulse extraction method extracted a total of 108 useful pulses. This is consistent with the analysis in A. That is, the double-threshold pulse extraction method and the two-level double-threshold pulse extraction method directly detect the starting point and the end point of the pulse, lack a preprocessing process to judge the peak first, and are easily affected by noise pulses. Therefore, the number of useful pulses they extract is relatively large, and is significantly higher than the other two pulse extraction methods. In fact, although the two-level double-threshold pulse extraction method includes a preprocessing process to judge the peak first, it does not work because the peak detection threshold is too small, so it is regarded as directly detecting the starting point and the end point of the pulse. Moreover, because the two-stage dual-threshold pulse extraction method solves the problem of continuous multiple pulses to a certain extent, the number of useful pulses extracted by it is slightly more than that of the dual-threshold pulse extraction method. In contrast, the number of useful pulses extracted by the three-threshold pulse extraction method and the two-stage adaptive multi-threshold pulse extraction method is less. Figure 8-Figure 11 It can be seen that the reliability of these extracted useful pulses is higher. Similarly, because the two-level adaptive multi-threshold pulse extraction method effectively solves the problem of continuous multi-pulses, the number of useful pulses it extracts is slightly more than that of the three-threshold pulse extraction method.

[0147] Table 1 shows the number of useful pulses extracted from the other twenty sound signals by the four pulse extraction methods. These twenty sound signals are randomly extracted from the 1,200 sound signals collected and saved, including eight sound signals randomly extracted from the 600 component signals and two sound signals randomly extracted from the 100 redundant object signals corresponding to the six materials.

[0148] Table 1 The number of useful pulses extracted by four pulse extraction methods on twenty sound signals

[0149]

[0150]

[0151] As can be seen from the table, the number of useful pulses extracted by the double threshold pulse extraction method and the two-level double threshold pulse extraction method is still significantly higher than that of the three threshold pulse extraction method and the two-level adaptive multi-threshold pulse extraction method. The two-level double threshold pulse extraction method extracts the largest number of useful pulses, and the three-threshold extraction algorithm extracts the least number of useful pulses. This is consistent with the useful pulse extraction effect of the above redundant signal. In addition, according to statistics, for 1,200 sound signals, the three-threshold pulse extraction method extracts a total of 109,589 useful pulses, the double threshold pulse extraction method extracts a total of 218,895 useful pulses, the two-level double threshold pulse extraction method extracts a total of 232,028 useful pulses, and the two-level adaptive multi-threshold pulse extraction method extracts a total of 126,027 useful pulses. Therefore, in terms of the number of pulses, the two-level adaptive multi-threshold pulse extraction method does not have an advantage, only more than the three-threshold pulse extraction method, and significantly less than the double threshold pulse extraction method and the two-level double threshold pulse extraction method. But this reflects from the side that it selectively extracts useful pulses and ensures the reliability of the extracted useful pulses, rather than pursuing the extraction of more useful pulses. This is its advantage over the three-threshold pulse extraction method. In addition, the number of useful pulses it extracts is significantly greater than that of the three-threshold pulse extraction method, which reflects from the side its advantage in solving the problem of continuous multiple pulses.

[0152] The second is the distribution density. For the redundant signal shown above, after calculation, the distribution density of the useful pulses extracted by the three-threshold pulse extraction method is 4.6947, the distribution density of the useful pulses extracted by the double-threshold pulse extraction method is 7.6450, the distribution density of the useful pulses extracted by the two-stage double-threshold pulse extraction method is 7.6450, and the distribution density of the useful pulses extracted by the two-stage adaptive multi-threshold pulse extraction method is 2.4533. The above calculation results are all retained to four decimal places. The smaller the distribution density, the higher the concentration of the useful pulses extracted from the sound signal, the more stable the average distribution level, and the better the useful pulse extraction effect of the corresponding pulse extraction method. Therefore, the useful pulse extraction effect of the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention is the best, and the useful pulse extraction effects of the double-threshold pulse extraction method and the two-stage double-threshold pulse extraction method are close, which are significantly worse than the three-threshold pulse extraction method and the two-stage adaptive multi-threshold pulse extraction method. Similarly, the present invention calculates the distribution density of useful pulses extracted from the above twenty sound signals by four pulse extraction methods, as shown in Table 2.

[0153] Table 2 Distribution density of useful pulses extracted by four pulse extraction methods on twenty sound signals

[0154]

[0155]

[0156] It can be seen from the table that no matter which section of the sound signal is used, the distribution density of useful pulses extracted by the two-stage adaptive multi-threshold pulse extraction method is the smallest, and is significantly smaller than the other three pulse extraction methods. In addition, the distribution density of useful pulses extracted by the three-threshold pulse extraction method is the second smallest, and is significantly smaller than the double-threshold pulse extraction method and the two-stage double-threshold pulse extraction method. In contrast, the distribution density of useful pulses extracted by the double-threshold pulse extraction method and the two-stage double-threshold pulse extraction method is large and close to each other, more than twice that of the two-stage adaptive multi-threshold pulse extraction method, and nearly twice that of the three-threshold pulse extraction method. This shows that the two-stage adaptive multi-threshold pulse extraction method has the best useful pulse extraction effect. After calculation, for 1,200 sound signals, the average distribution density of useful pulses extracted by the three-threshold pulse extraction method is 4.6905, the average distribution density of useful pulses extracted by the double-threshold pulse extraction method is 7.6972, the average distribution density of useful pulses extracted by the two-level double-threshold pulse extraction method is 7.7306, and the average distribution density of useful pulses extracted by the two-level adaptive multi-threshold pulse extraction method is 2.4432. It should be noted that the average distribution density here is calculated in units of sound signals, not useful pulses. This effectively illustrates that the two-level adaptive multi-threshold pulse extraction method proposed in the present invention has the best useful pulse extraction effect, which is consistent with the evaluation result of the number of pulses. Its feasibility and superiority are effectively proven.

[0157] In terms of the number of pulses, the two-level adaptive multi-threshold pulse extraction method does not have an advantage, only more than the three-threshold pulse extraction method, and significantly less than the double-threshold pulse extraction method and the two-level double-threshold pulse extraction method. But this reflects from the side that it selectively extracts useful pulses to ensure the reliability of the extracted useful pulses, rather than pursuing the extraction of more useful pulses. This is its advantage over the three-threshold pulse extraction method. Moreover, the number of useful pulses it extracts is still significantly more than that of the three-threshold pulse extraction method, which reflects from the side its advantage in solving the problem of continuous multiple pulses. In fact, from the evaluation results of the number of pulses and distribution density, they seem to be a set of similar evaluation indicators. The better the useful pulse extraction effect, the fewer pulses and the smaller the distribution density. This can be seen from the comparison between the two-level adaptive multi-threshold pulse extraction method, the three-threshold pulse extraction method, the double-threshold pulse extraction method, and the two-level double-threshold pulse extraction method. But this is not entirely correct, because for the two-level adaptive multi-threshold pulse extraction method and the three-threshold pulse extraction method, the former has more pulses and a smaller distribution density. Therefore, the distribution density can be directly used to evaluate the useful pulse extraction effect of the pulse extraction method, and it is not ambiguous, but the number of pulses cannot. It does not mean that the more the number of useful pulses extracted, the better the useful pulse extraction effect of the corresponding pulse extraction method. Figure 8-Figure 11The visualization of shows that this is incorrect. Therefore, the number of pulses can only be used as an auxiliary evaluation indicator, and it cannot give an absolutely reliable and useful pulse extraction effect like the distribution density.

[0158] In summary, from a mathematical perspective, the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention achieves the best useful pulse extraction effect.

[0159] C. Evaluation of signal angle:

[0160] From the signal perspective, the signal-to-noise ratio and the normalized signal-to-noise ratio are mainly selected to evaluate the useful pulse extraction effects of the four pulse extraction methods. The first is the signal-to-noise ratio. For a section of the redundant signal shown in A, after calculation, the signal-to-noise ratio of the redundant signal after being processed by the three-threshold pulse extraction method is 9.2855, the signal-to-noise ratio of the redundant signal after being processed by the double-threshold pulse extraction method is 12.4346, the signal-to-noise ratio of the redundant signal after being processed by the two-stage double-threshold pulse extraction method is 12.6949, and the signal-to-noise ratio of the redundant signal after being processed by the two-stage adaptive multi-threshold pulse extraction method is 13.6371. The above calculation results are all retained to four decimal places. The larger the signal-to-noise ratio, the higher the quality of the speech signal, that is, the higher the quality of the extracted useful pulse, and the better the useful pulse extraction effect of the corresponding pulse extraction method. Therefore, the useful pulse extraction effect of the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention is the best, and the useful pulse extraction effect of the three-threshold pulse extraction method is the worst. Surprisingly, the useful pulse extraction effect of the double threshold pulse extraction method and the two-stage double threshold pulse extraction method is better, slightly worse than the two-stage adaptive multi-threshold pulse extraction method, but much better than the three-threshold pulse extraction method. The signal-to-noise ratio of the twenty sound signals processed by the four pulse extraction methods is calculated respectively, as shown in Table 3.

[0161] Table 3 Signal-to-noise ratio of twenty sound signals after processing by four pulse extraction methods

[0162]

[0163]

[0164] As can be seen from the table, the signal-to-noise ratio of the sound signal processed by the two-stage adaptive multi-threshold pulse extraction method is the largest, and the signal-to-noise ratio of the sound signal processed by the three-threshold pulse extraction method is the smallest, and the gap between the two is more obvious. In contrast, the signal-to-noise ratio of the sound signal processed by the double-threshold pulse extraction method is lower than that by the two-stage double-threshold pulse extraction method, but the gap is very small. Moreover, the signal-to-noise ratio of the sound signal processed by the double-threshold pulse extraction method and the two-stage double-threshold pulse extraction method is close to that of the two-stage adaptive multi-threshold pulse extraction method, and is also significantly greater than the three-threshold pulse extraction method. This shows that the two-stage adaptive multi-threshold pulse extraction method has the best useful pulse extraction effect. After calculation, for 1,200 sound signals, the average signal-to-noise ratio of the sound signal after being processed by the three-threshold pulse extraction method is 9.0879, the average signal-to-noise ratio of the sound signal after being processed by the double-threshold pulse extraction method is 12.5632, the average signal-to-noise ratio of the sound signal after being processed by the two-stage double-threshold pulse extraction method is 12.6857, and the average signal-to-noise ratio of the sound signal after being processed by the two-stage adaptive multi-threshold pulse extraction method is 13.2341. This effectively proves the feasibility and superiority of the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention. But this is different from the analysis results obtained from a mathematical point of view. From a mathematical point of view, the useful pulse extraction effects obtained by the two-stage adaptive multi-threshold pulse extraction method, the three-threshold pulse extraction method, the two-stage double-threshold pulse extraction method and the double-threshold pulse extraction method decrease in sequence. But here, the useful pulse extraction effects obtained by the two-stage adaptive multi-threshold pulse extraction method, the two-stage double-threshold pulse extraction method, the double-threshold pulse extraction method and the three-threshold pulse extraction method decrease in sequence. The difference lies in the useful pulse extraction effect achieved by the three existing pulse extraction methods.

[0165] From the definitions of equations (3) to (6), it can be seen that the signal-to-noise ratio used is actually an improved signal-to-noise ratio. The number of sampling points in the numerator part used for calculation is the same as the number of sampling points in the denominator part, and this number varies according to the sound signal and the pulse extraction method used. For the sound signal processed by the three-threshold pulse extraction method, those useful pulses with a relatively high degree of credibility are extracted, and some pulses with smaller amplitudes are discarded. In this way, although the average power P of the speech signal calculated from the extracted useful pulses is s However, due to the addition of pulses with smaller amplitudes, the normalized amplitude of the noise pulses will also increase, resulting in the average power of the noise P nThis explains why the average signal-to-noise ratio of the sound signal processed by the three-threshold pulse extraction method is relatively low. In contrast, for the sound signal processed by the double-threshold pulse extraction method and the two-stage double-threshold pulse extraction method, almost all pulses are extracted, including some noise pulses with small amplitudes. In this way, not only the average power P of the speech signal calculated from the extracted pulses is s is larger, and because there are almost no pulses, the normalized amplitude of the noise pulse will become smaller, resulting in the average power P of the noise n This explains why the average signal-to-noise ratio of the sound signal processed by the dual threshold pulse extraction method and the two-stage dual threshold pulse extraction method is relatively high. And it is conceivable that for the sound signal processed by the two-stage adaptive multi-threshold pulse extraction method, they have obtained the average power P of the speech signal. s The average power of the noise P n The optimal balance of , makes the calculated signal-to-noise ratio maximum. This again shows the advantages of the two-level adaptive multi-threshold pulse extraction method.

[0166] The second is the normalized signal-to-noise ratio. After calculation, the normalized signal-to-noise ratio of the sound signal processed by the three-threshold pulse extraction method is 0.6571 for the redundant signal shown above, the normalized signal-to-noise ratio of the sound signal processed by the double-threshold pulse extraction method is 0.8718, the normalized signal-to-noise ratio of the sound signal processed by the two-stage double-threshold pulse extraction method is 0.8836, and the normalized signal-to-noise ratio of the sound signal processed by the two-stage adaptive multi-threshold pulse extraction method is 1.1325. The larger the normalized signal-to-noise ratio, the higher the quality of the speech signal, that is, the higher the quality of the extracted useful pulse, and the better the useful pulse extraction effect of the corresponding pulse extraction method. Therefore, the useful pulse extraction effect of the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention is the best, and the useful pulse extraction effect of the three-threshold pulse extraction method is the worst. The useful pulse extraction effects of the dual threshold pulse extraction method and the two-stage dual threshold pulse extraction method are also good, which are worse than the two-stage adaptive multi-threshold pulse extraction method, but significantly better than the three-threshold pulse extraction method. This is consistent with the signal-to-noise ratio results of the sound signal after being processed by the four pulse extraction methods above. The normalized signal-to-noise ratios of the twenty sound signals processed by the four pulse extraction methods are calculated respectively, as shown in Table 4.

[0167] Table 4 Normalized signal-to-noise ratio of twenty sound signals after processing by four pulse extraction methods

[0168]

[0169] As can be seen from the table, the normalized signal-to-noise ratio of the sound signal processed by the two-stage adaptive multi-threshold pulse extraction method is the largest, and the normalized signal-to-noise ratio of the sound signal processed by the three-threshold pulse extraction method is the smallest. The normalized signal-to-noise ratio of the sound signal processed by the double-threshold pulse extraction method is lower than that by the two-stage double-threshold pulse extraction method. Moreover, the normalized signal-to-noise ratio of the sound signal processed by the double-threshold pulse extraction method and the two-stage double-threshold pulse extraction method is between the three-threshold pulse extraction method and the two-stage adaptive multi-threshold pulse extraction method, and there is a certain difference between the two. This is consistent with the signal-to-noise ratio results of the twenty sound signals processed by the four pulse extraction methods above. Similarly, this shows that the two-stage adaptive multi-threshold pulse extraction method has the best useful pulse extraction effect. After calculation, for 1,200 sound signals, the average normalized signal-to-noise ratio of the sound signal after being processed by the three-threshold pulse extraction method is 0.6809, the average normalized signal-to-noise ratio of the sound signal after being processed by the double-threshold pulse extraction method is 0.8676, the average normalized signal-to-noise ratio of the sound signal after being processed by the two-stage double-threshold pulse extraction method is 0.9007, and the average normalized signal-to-noise ratio of the sound signal after being processed by the two-stage adaptive multi-threshold pulse extraction method is 1.1326. This once again effectively proves the feasibility and superiority of the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention. At the same time, combined with the signal-to-noise ratio of the sound signal after being processed by the four pulse extraction methods, it can be comprehensively concluded that from the signal perspective, the useful pulse extraction effects obtained by the two-stage adaptive multi-threshold pulse extraction method, the two-stage double-threshold pulse extraction method, the double-threshold pulse extraction method and the three-threshold pulse extraction method decrease in turn, which is indeed different from the useful pulse extraction effect obtained from a mathematical perspective.

[0170] In summary, from the perspective of signal evaluation, the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention still achieves the best useful pulse extraction effect.

[0171] D. Evaluation from the perspective of machine learning:

[0172] From the perspective of machine learning, the component classification accuracy and material classification accuracy are mainly selected to evaluate the useful pulse extraction effect of the four pulse extraction methods. In this way, the two classification accuracies obtained by the classifier can indirectly reflect the useful pulse extraction effect of the pulse extraction method. Table 5 shows the fifteen signal features used in the present invention.

[0173] Table 5 Fifteen signal characteristics

[0174]

[0175] As mentioned above, for the collected and stored 1,200 sound signals, the three-threshold pulse extraction method, the double-threshold pulse extraction method, the two-level double-threshold pulse extraction method and the two-level adaptive multi-threshold pulse extraction method have been used to process them respectively, and multiple useful pulse extraction results have been obtained and saved. According to the corresponding pulse extraction method used, the present invention divides these useful pulses into four sets, namely, useful pulse set 1, useful pulse set 2, useful pulse set 3 and useful pulse set 4, which correspond to the three-threshold pulse extraction method, the double-threshold pulse extraction method, the two-level double-threshold pulse extraction method and the two-level adaptive multi-threshold pulse extraction method respectively. Taking the useful pulse set 1 as an example, the characteristic values ​​of the fifteen signal features shown in Table 5 are calculated on each useful pulse, and the fifteen characteristic values ​​calculated for each useful pulse are arranged in a row according to the order of the signal features in the table to construct a characteristic vector. In this way, the useful pulses in the useful pulse set 1 are all constructed into characteristic vectors. On this basis, for those feature vectors constructed by useful pulses contained in component signals, the label "0" representing the component signal is added, and for those feature vectors constructed by useful pulses contained in redundant signals, the label "1" representing the redundant signal is added. It should be noted that here the sound signals from the useful pulse set one are only divided into component signals and redundant signals in a broad sense, without considering the differences between the component signals and the redundant signals. In this way, the feature vectors constructed by the useful pulses in the useful pulse set one are all labeled, and a large amount of labeled data is obtained, thereby constructing component data set one. That is, component data set one contains two labels. Through the same processing process, component data set two, component data set three and component data set four are constructed. Table 6 gives their detailed descriptions. As can be seen from the table, the total data volume of each component data set is the same as the number of useful pulses extracted from one thousand two hundred sound signals by the corresponding pulse extraction method.

[0176] Table 6 Detailed description of the four component datasets

[0177]

[0178] In addition, the useful pulses of the source redundant signals in the four useful pulse sets are copied to form four new sets, namely, the new useful pulse set 1, the new useful pulse set 2, the new useful pulse set 3 and the new useful pulse set 4, which correspond to the three-threshold pulse extraction method, the double-threshold pulse extraction method, the two-level double-threshold pulse extraction method and the two-level adaptive multi-threshold pulse extraction method, respectively. Similarly, taking the new useful pulse set 1 as an example, the characteristic value is calculated on each useful pulse, and all the useful pulses are constructed into a characteristic vector. On this basis, for the characteristic vectors constructed by the useful pulses contained in the redundant signals generated by the copper wire, the label "1" representing the copper wire is added. For the characteristic vectors constructed by the useful pulses contained in the redundant signals generated by the solder particles, the label "2" representing the solder particles is added. For the characteristic vectors constructed by the useful pulses contained in the redundant signals generated by the aluminum particles, the label "3" representing the aluminum particles is added. For the characteristic vectors constructed by the useful pulses contained in the redundant signals generated by the hot melt adhesive particles, the label "4" representing the hot melt adhesive particles is added. For the feature vectors constructed from the useful pulses contained in the superfluous signals generated by PVC particles, the label "5" representing the PVC particles is added. For the feature vectors constructed from the useful pulses contained in the superfluous signals generated by silica gel particles, the label "6" representing the silica gel particles is added. In this way, the feature vectors constructed from the useful pulses in the new useful pulse set one are all labeled, and a large amount of labeled data is obtained, thereby constructing the material data set one. That is, the material data set one contains six labels. Through the same processing process, material data set two, material data set three and material data set four are constructed. Table 7 gives their detailed descriptions. It can be seen from the table that the total amount of data for each material data set is the same as the amount of data labeled "1" in Table 6.

[0179] Table 7 Detailed description of the four material datasets

[0180]

[0181] Most classifiers in machine learning are based on tree theory or mathematical theory. Commonly used classifiers based on tree theory include decision trees, random forests, XGBoost, etc. Commonly used classifiers based on mathematical theory include naive Bayes, k nearest neighbor, support vector machine, etc. In order to comprehensively evaluate the quality of the four component data sets and the four material data sets, and then evaluate the useful pulse extraction effect of the four pulse extraction methods, two representative classifiers based on tree theory, random forest and XGBoost, and two classifiers based on mathematical theory, k nearest neighbor and support vector machine, were selected. Therefore, the four component data sets were divided into four training sets and four test sets in a ratio of 7:3, and four classifiers were trained on the four training sets respectively. Their classification performance was tested on the corresponding test sets, and the component classification accuracy was statistically obtained. Fig.12 As shown. It should be noted that the four classifiers trained here are all configured with default parameters and no parameter optimization is performed. In fact, training high-performance classifiers is not the research focus of this invention. They are only tools for evaluating the effect of useful pulse extraction, as long as they have basic classification performance. In addition, because the four component data sets only contain two labels, the classic support vector machine is sufficient and there is no need to introduce a kernel function.

[0182] from Fig.12 It can be clearly seen that no matter which classifier, it achieves the highest component classification accuracy on component data set four, and the second highest component classification accuracy on component data set one, and the gap between the two is small. In contrast, the classifiers achieve lower component classification accuracy on component data sets two and three, which is quite different from the component classification accuracy achieved on component data sets one and four. Referring to Table 6, it can be seen that component data set four corresponds to the proposed two-level adaptive multi-threshold pulse extraction method. This effectively illustrates the feasibility and superiority of the two-level adaptive multi-threshold pulse extraction method. This is consistent with the analysis results obtained from a mathematical point of view. After calculation, the average component classification accuracy achieved by the four classifiers on the four component data sets is 96.05%, 82.41%, 82.31% and 96.57% respectively. Compared with the existing redundant signal and component signal recognition research conducted by scholars in the background technology, the individual component classification accuracy and average component classification accuracy achieved by the four classifiers on component data sets one and four are higher than the highest classification accuracy achieved in existing studies. This effectively illustrates the necessity of carrying out the present invention.

[0183] Similarly, the four material data sets are divided into four training sets and four test sets in a ratio of 7:3. Four classifiers are trained on the four training sets respectively, and their classification performance is tested on the corresponding test sets. The material classification accuracy is statistically obtained. Fig.13 As shown. It should be noted that the four classifiers trained here are also configured with default parameters. However, because the four material datasets all contain six labels, the radial basis function is introduced on the basis of the classic support vector machine to ensure that it has good and stable classification performance when dealing with multi-classification problems.

[0184] from Fig.13It can also be clearly seen that no matter which classifier, it achieves the highest material classification accuracy on material dataset four, the second highest material classification accuracy on material dataset one, and lower material classification accuracy on material dataset two and material dataset three. Moreover, the gap between the material classification accuracy achieved by the classifier on material dataset two and material dataset three and the material classification accuracy achieved on material dataset one and material dataset four is further widened. Referring to Table 7, it can be seen that material dataset four corresponds to the proposed two-level adaptive multi-threshold pulse extraction method. This once again effectively illustrates the feasibility and superiority of the two-level adaptive multi-threshold pulse extraction method. After calculation, the average material classification accuracy achieved by the four classifiers on the four material datasets is 86.67%, 69.01%, 68.64% and 87.56%, respectively. Refer to the existing redundant material recognition research conducted by scholars in the background technology, especially the recognition research on specific materials. For the three-classification problem, Meng Si et al. (2020) achieved a classification accuracy of 98%. For the five-class classification problem, Gao Yajie et al. (2021) achieved a classification accuracy of 78%. For the six-class classification problem, Sun Zhigang et al. (2022) trained a random forest with a classification accuracy of 63.60%, and trained a parameter-optimized XGBoost with a classification accuracy of 80.41%.

[0185] Specifically, although it seems that the three-classification problem is only half of the six-classification problem, its complexity is much less than half of the six-classification problem. They are not a simple linear superposition relationship. Therefore, it is unfair and meaningless to compare with the classification accuracy achieved by Meng Si et al. (2020). The five-classification problem is close to the six-classification problem, but its complexity is still less than the six-classification problem. But despite this, the individual material classification accuracy and average material classification accuracy achieved by the four classifiers on material dataset one and material dataset four are higher than those achieved by Gao Yajie et al. (2021). The closest to the present invention is the research of Sun Zhigang et al. (2022), who also trained random forests and parameter-optimized XGBoost. The classification accuracy achieved by the random forest they trained was 63.60%, which is lower than the individual material classification accuracy and average material classification accuracy achieved by the four classifiers trained on the four material datasets. The classification accuracy achieved by the parameter-optimized XGBoost they trained was 80.41%, which is lower than the individual material classification accuracy and average material classification accuracy achieved by the four classifiers trained on material dataset one and material dataset four. Not to mention that the training is only for XGBoost with default parameter configuration. The above analysis can fully illustrate the necessity of carrying out the present invention, as well as the feasibility and superiority of the two-level adaptive multi-threshold pulse extraction method proposed in the present invention.

[0186] In summary, from the perspective of machine learning, the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention still achieves the best useful pulse extraction effect.

[0187] Based on the evaluation results of the useful pulse extraction effects of four pulse extraction methods obtained from the mathematical perspective, signal perspective and machine learning perspective, the present invention gives the following conclusion: the two-stage adaptive multi-threshold pulse extraction method proposed in the present invention achieves the best useful pulse extraction effect.

[0188] The component classification accuracy of the trained XGBoost is 97.60%, and the material classification accuracy is 88.95%. Perhaps after parameter optimization, the component classification accuracy and material classification accuracy will be improved, but they are only classification accuracy at the data level, not recognition accuracy in engineering application scenarios. This problem also exists in the existing research on the recognition of redundant signals and component signals, and the research on the recognition of redundant materials. The classification accuracy of redundant signals and component signals obtained by the classifiers trained by scholars, the classification accuracy of redundant objects of metal materials and non-metallic materials, and the classification accuracy of redundant objects of specific materials are essentially classification accuracy at the data level. There is a lack of a majority voting process between classification accuracy and recognition accuracy.

[0189] As mentioned above, the two-level adaptive multi-threshold pulse extraction method can extract 108 useful pulses from this sound signal. In this way, by calculating the characteristic values ​​of fifteen signal features on each useful pulse, 108 useful pulses are constructed into 108 feature vectors, and 108 unlabeled data are obtained. When the classifier trained on the component dataset, such as XGBoost, is applied to predict the labels of these 108 data, 108 predicted labels will be obtained. Most of these predicted labels are "1", and a small number are "0". In this way, in the majority voting process, the label "1" with a voting advantage will win, and it is determined as the common label of these 108 data. Referring to Table 6, it can be seen that the label corresponds to the redundant signal, which is consistent with the actual situation. This completes the classification from the data level to the recognition in the engineering application scenario. In this way, the recognition accuracy can be defined as the proportion of the number of correctly recognized sound signals in multiple sound signals. Similarly, when the classifier trained on the material dataset is applied to predict the labels of these 108 data, 108 predicted labels will also be obtained. In the majority voting process, a label from "1" to "6" will also have a voting advantage and win. Refer to Table 7, the corresponding material is the redundant material recognition result of this sound signal. This is worthy of in-depth research by scholars in related fields, and it can be used to promote the application of existing methods or algorithms in engineering application scenarios.

[0190] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection, characterized in that: The following steps are involved: Step 1: Calculate the average energy E of the entire sound signal a and the average zero-crossing rate Z a , and the standard deviation of energy σ E and the standard deviation of the zero-crossing rate σ Z ; According to the 3σ distribution criterion of normal distribution, set the peak detection threshold E p , endpoint energy threshold E be and endpoint zero crossing rate threshold Z be ; Step 2: Frame the sound signal and calculate the energy and zero-crossing rate of each frame signal; Step 3: Starting from the first frame signal, compare the energy of each frame signal with E p ; When the energy of a frame signal is greater than E p When the frame signal is recorded, the energy of each subsequent frame signal is compared with E p , until the energy of a frame signal is less than E p ; Find the frame signal with the largest energy from these frame signals, and obtain the time of the first sampling point, which is called the peak time t of the current useful pulse. p ; Step 4: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the energy of each frame signal is compared with E be Compare until the energy of a frame signal is found to be less than E in both directions. be ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the first candidate start time t of the current useful pulse b_f1 and the first candidate end time t e_f1 ; Step 5: Take the peak time t of the useful pulse extracted in step 3 p As the starting point, the zero-crossing rate of each frame signal is calculated forward and backward respectively. be Compare until the zero-crossing rate of a frame signal is found to be less than Z in both directions. be ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the second candidate start time t of the current useful pulse b_f2 and the second candidate end time t e_f2 ; Step 6: Compare the first candidate start time with the second candidate start time, and determine the larger one as the final start time t b_1 ; Compare the first candidate end time with the second candidate end time, and determine the smaller one as the final end time t e_1 ; Step 7: Take the final end time t of the useful pulse extracted in step 6 e_1 The next frame signal is taken as a new starting point, and steps 3 to 6 are repeated to extract the second useful pulse; and so on, until the last frame signal of the sound signal is searched; and the extracted multiple useful pulses are saved; Step 8: Calculate the average energy and average zero-crossing rate of each useful pulse, expressed as E i and Z i ; Calculate the standard deviation of the energy of each pulse σ E_i and the standard deviation of the zero-crossing rate σ Z_i ; Where i represents the number of the useful pulse; Based on E i and σ E_i Determine the endpoint energy threshold E for each useful pulse be_i , and based on Z i and σ Z_i Determine the endpoint zero-crossing rate threshold value of the useful pulse as Z be_i ; Step 9: Take the peak time t of a useful pulse saved in step 7 p As the starting point, the energy of each frame signal is compared with the exclusive E be_i Compare until the energy of a frame signal is found to be less than E in both directions. be_i ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the third candidate start time t of the current useful pulse b_s1 and the third candidate end time t e_s1 ; Step 10: Take the peak time t of a useful pulse saved in step 7 p As the starting point, the zero-crossing rate of each frame signal is compared with the exclusive Z be_i Compare until the zero-crossing rate of a frame signal is found to be less than Z in both directions. be_i ; Respectively identify the previous frame signals of the two frame signals as the start frame signal and the end frame signal of the current useful pulse; obtain the time of the first sampling point in the two frame signals, which is called the fourth candidate start time t of the current useful pulse b_s2 and the fourth candidate end time t e_s2 ; Step 11: Compare the third candidate start time and the fourth candidate start time, and determine the larger one as the final start time t b_2 Similarly, compare the third candidate end time and the fourth candidate end time, and determine the smaller one as the final end time t e_2 ; Step 12: starting from the second useful pulse saved in step 7, repeat the processing steps from step 9 to step 11; And so on, until all the useful pulses saved in step 7 are processed; Step 13: Calculate the first duration = t e_1 -t b_1 and the second duration = t e_2 -t b_2 ; Calculate the ratio of the second duration to the first duration; if the ratio exceeds the duration ratio threshold, set the final start time t b_1 and the final ending time t e_1 The real starting time t of the current useful pulse is considered b and the actual end time t e ; If the ratio is lower than the duration ratio threshold, the final start time t b_1 and the final ending time t e_2 The real starting time t of the current useful pulse is considered b and the actual end time t e ; Step 14: The useful pulses processed in step 13 are retained, and the useful pulse extraction of the entire sound signal is completed.

2. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 1, characterized in that: The average energy E a and the average zero-crossing rate Z a It is obtained by calculating the average energy and average zero-crossing rate of all sampling points contained in the sound signal.

3. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 2, characterized in that: The peak detection threshold E in step 1 p =E a +3σ E .

4. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 3, characterized in that: The endpoint energy threshold E in step 1 be =E a +σ E .

5. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 4, characterized in that: The endpoint zero-crossing rate threshold Z in step 1 be =z a +σ Z .

6. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 5, characterized in that: The endpoint energy threshold E of the useful pulse described in step 8 be_i =E i +σ E_i .

7. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 6, characterized in that: The endpoint zero-crossing rate threshold Z of the useful pulse described in step 8 is be_i =Z i +σ Z_i .

8. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 7, characterized in that: When performing frame processing on the sound signal in step 2, the duration of each frame signal is set to Δt=50 μs.

9. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to claim 8, characterized in that: When sampling the sound signal, the interval between sampling points is 2μs.

10. A two-stage adaptive multi-threshold pulse extraction method for redundant object detection according to any one of claims 1 to 9, characterized in that: The duration percentage threshold is 65%.

Citation Information

Patent Citations

  • Method for identifying redundancies and assembly of satellite-borne electronic device

    CN102830421A

  • Adaptive threshold setting voice endpoint detection method and device thereof, and readable storage medium

    CN108847218A

  • Method for identifying detection result of particle collision noise detection system

    CN115931114A

  • Sealed electronic equipment redundancy positioning definite reliability analysis method

    CN116956745A

  • Voice signal detecting system and method

    KR1020070045846A