A nondestructive detection and identification method of wood borers based on acoustic signals

Abstract

A nondestructive detection and identification method for wood borers based on acoustic signals, which performs pulse waveform detection on the collected wood audio signals, extracts the average energy characteristics and zero-crossing rate characteristics of the pulse waveform in 10 frequency bands, uses a linear discriminant analysis (LDA) algorithm to reduce the dimension, and inputs the reduced dimension features into a SVM classifier using a Gaussian kernel function to identify whether there is wood borer activity sound. The present invention can realize nondestructive detection of wood borers and can effectively improve the quarantine efficiency of wood borers.

Description

Technical Field

[0001] The invention belongs to the technical field of audio signal detection, and in particular relates to a non-destructive detection and identification method of wood borers based on acoustic signals. Technical Background

[0002] In recent years, my country's annual timber import volume has been rising steadily. Wood borers have invaded my country along with imported timber, threatening my country's ecological environment security at all times, and causing huge economic losses to my country's forestry, construction, furniture manufacturing and other industries every year.

[0003] At present, the main means of preventing wood borers at customs ports is manual sampling, which is not only time-consuming and labor-intensive, but also has a high rate of missed inspections and cannot quickly and non-destructively complete the inspection and quarantine of imported wood products.

[0004] Patent ZL202020565730.7 provides a device for collecting the sound of wood borers, which can obtain low-noise recording data of imported wood borer activity sounds by collecting vibration sounds and performing internal and external noise reduction processing. The patent provides data collection and preprocessing functions, but does not further provide a detection and recognition algorithm for wood borer activity sounds.

[0005] Support vector machine (SVM) is a binary classification model, and its basic model is a linear classifier with the largest interval defined in the feature space. SVM adapts to the needs of nonlinear classification by setting different kernel functions. The linear discriminant analysis (LDA) algorithm can achieve feature dimension reduction, thereby reducing the complexity of the algorithm. At present, insect sound detection is mostly carried out manually, which is inefficient, and few machine learning methods are applied to insect sound detection. Summary of the invention

[0006] The purpose of the present invention is to provide an insect sound detection algorithm for non-destructive detection of wood borers in view of the deficiencies of the prior art. The LDA dimensionality reduction method and the excellent performance of the SVM machine learning model are used to realize the detection of wood insect sounds, with a detection accuracy of more than 92%, while improving the quarantine efficiency of wood borers.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A method for nondestructive detection and identification of wood borers based on acoustic signals, the method comprising the following steps:

[0009] Step 1: Collect wood sound signals: Place the sound amplification device on the surface of the wood sample to collect wood sound signals;

[0010] Step 2: Detect and extract the effective pulse waveform in the wood sound signal:

[0011] Extract the envelope of wood sound signal, and then perform two smoothing filters on the extracted envelope;

[0012] According to the smoothed envelope, set the starting threshold and peak threshold of the effective pulse;

[0013] Starting from the first data point of the smoothed envelope, traverse the entire envelope and find the point where the envelope line changes from less than or equal to the starting threshold to greater than the starting threshold, which is taken as the pulse starting point;

[0014] Starting from the pulse starting point, find the point where the envelope changes from greater than the starting threshold to less than or equal to the starting threshold, which is taken as the end point of the pulse;

[0015] Find the maximum peak value of the envelope between the pulse start point and the pulse end point. If the value is greater than the set peak threshold of the valid pulse, then this pulse is considered to be a valid pulse waveform, otherwise it is discarded.

[0016] Extract all valid pulse waveform segments that meet the conditions;

[0017] Step 3: Perform data enhancement on the extracted effective pulse waveform segment;

[0018] Step 4: Extract the features of the effective pulse waveform segment: For the detected effective pulse waveform segment, extract the average energy features and zero-crossing rate features in 10 frequency bands respectively, then perform feature dimensionality reduction, and combine the signal features after dimensionality reduction into a feature vector;

[0019] Step 5: Use the SVM model for training and testing:

[0020] The extracted effective pulse waveform segments are labeled into two categories: "with insects" and "without insects";

[0021] Randomly divide the dataset with labeled information into training set and test set;

[0022] Input the feature vector of the training set data into the SVM model based on the Gaussian kernel function and perform model training;

[0023] The feature vector of the test set data is input into the trained model, and the test set is classified and tested to obtain the identification result of whether there are insects.

[0024] The smoothing filtering method described in step 2 comprises the following steps:

[0025] Step 21: Find the maximum value: traverse the data points of the input signal, find all the maximum points of the input signal, and record their positions p k and amplitude x k (k=1, 2, 3, ...).

[0026] Step 22: Average adjacent maxima: Calculate a new value for each maximum point. The new value at the kth maximum point where x0 takes the value of the first data point of the input signal.

[0027] Step 2 and 3: Linear interpolation: The value y of the pth data point of the filtered signal satisfies the formula:

[0028]

[0029] Among them, x′0=x0.

[0030] The calculation method of the starting threshold and the maximum threshold in step 2 is: first calculate the mean μ and standard deviation σ of the smoothed envelope, then calculate two thresholds, the starting threshold is μ, and the maximum threshold is μ+2σ.

[0031] The calculation method of the average energy feature and the zero-crossing rate feature described in step four is: pass the effective pulse waveform segment signal through a band-pass filter group composed of 10 band-pass filters to obtain waveforms in 10 frequency bands, and then calculate the average energy feature and the zero-crossing rate feature respectively in the 10 frequency bands, and obtain a total of 10-dimensional average energy features and 10-dimensional zero-crossing rate features.

[0032] The passband frequency ranges of the 10 bandpass filters in the bandpass filter bank are 2kHz-3kHz, 3kHz-4kHz, 4kHz-5kHz, 5kHz-6kHz, 7kHz-8kHz, 8kHz-9kHz, 9kHz-10kHz, 10kHz-11kHz, 11kHz-12kHz, respectively. The transition bandwidth is 500Hz, and the stopband suppression is not less than 40dB.

[0033] The linear discriminant analysis method described in step 4 should map the 10-dimensional average energy characteristics and 10-dimensional zero-crossing rate characteristics of the effective pulse waveform segment respectively.

[0034] When the linear discriminant analysis method is used, the mapping of the first two dimensions of the average energy characteristics of the effective pulse waveform segment and the mapping of the first dimension of the zero-crossing rate characteristics are finally retained.

[0035] Compared with the prior art, the present invention has the following beneficial effects: non-destructive detection of wood borers is achieved based on the active sound of wood borers. Different frequency band features are extracted by bandpass filtering, the complexity of the algorithm is reduced by feature dimension reduction, and finally the SVM algorithm based on the kernel function is used to quickly detect whether the wood contains insects, thereby improving the efficiency of wood borer quarantine for imported wood. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flowchart of data preparation for an embodiment of the present invention;

[0037] Figure 2 The flowchart of training and testing of an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is further illustrated below in conjunction with the embodiments and drawings. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0039] like Figure 1 As shown, a method for nondestructive detection and identification of wood borers based on acoustic signals comprises the following steps:

[0040] Step 1: Collect wood sound signals: Use a microphone attached to the wood surface to collect sound signals and obtain a data set.

[0041] Step 2: Extract valid pulse waveform segments from wood sound signals: First, extract the envelope of the wood sound signal, and then smooth and filter the extracted envelope twice. According to the smoothed envelope, calculate the pulse starting threshold and peak threshold. Starting from the first data point of the smoothed envelope, traverse the smoothed envelope and find the point where the envelope changes from less than or equal to the starting threshold to greater than the starting threshold as the pulse starting point. Starting from the pulse starting point, find the point where the envelope changes from greater than the starting threshold to less than or equal to the starting threshold as the pulse ending point. Find the maximum envelope value between the pulse starting point and the pulse ending point. If the value is greater than the peak threshold, this pulse is considered to be a valid pulse waveform segment, otherwise it is discarded. Extract all valid pulse waveform segments that meet the conditions.

[0042] Step 3: Perform data enhancement on the extracted effective pulse waveform segment signal.

[0043] Step 4: Extract the features of the effective pulse waveform segment signal: Calculate the average energy features and zero-crossing rate features of the effective pulse waveform segment. Use the linear discriminant analysis method to reduce the dimensions of the extracted average energy features and zero-crossing rate features of the effective pulse waveform segment, and combine the reduced-dimensional signal features into a feature vector.

[0044] Step 5: SVM model training and detection: Label the extracted valid pulse waveform segments to see if they contain insects, and randomly divide the data set into a training set and a test set. Input the feature vector of the training set data into the SVM model using the Gaussian kernel function to train the model. Input the feature vector of the test set data into the trained model, classify the valid pulse waveforms of the test set, and determine whether they are indeed insect sounds.

[0045] In a specific embodiment, the detailed steps of a method for nondestructive detection and identification of wood borers based on acoustic signals disclosed in the present invention are as follows:

[0046] First, a microphone is attached to the surface of imported wood to collect sound signals. The wood insect sound audio signals involved in the embodiment of the present invention include the audio of three types of borers, namely, powder beetles, wood borers, and fruit flies, and the noise audio involved includes human voices, wind sounds, bird sounds, and noise audio emitted by wireless devices that are easily mixed into the actual sampling signals in actual sampling.

[0047] Secondly, perform envelope extraction on the audio signal: first find the zero-crossing point of the audio signal and record its position.

[0048] The amplitude of the audio signal at the zero-crossing point is set to 0, and the absolute value of the amplitude of the audio signal between the zero-crossing points is taken.

[0049] Find the maximum value between zero crossing points and record its position and amplitude. Finally, connect all the maximum values ​​between zero crossing points to form the audio envelope.

[0050] Then, the audio envelope is smoothed and filtered: traverse the data points of the input signal, find all the maximum points of the input signal, and record their positions p k and amplitude x k (k=1, 2, 3, ...).

[0051] For each maximum point, calculate the new value where x0 takes the value of the first data point.

[0052] Linear interpolation is performed according to the following formula:

[0053]

[0054] Among them, x′0=x0.

[0055] After two smoothing filters, the smoothed envelope of the audio signal is obtained.

[0056] Then the mean μ and standard deviation σ of the smoothed envelope are calculated.

[0057] The starting threshold is μ and the maximum threshold is μ+2σ.

[0058] Traverse each data point of the smoothed envelope, and find the point where the envelope data changes from less than or equal to the starting threshold to greater than the starting threshold, and record this position as the starting point of the current audio pulse.

[0059] Starting from the starting point of the audio pulse, find the first data point where the envelope data is less than or equal to the starting threshold, and record this position as the end point of the current audio pulse.

[0060] Find the maximum value between the start and end points of the audio pulse. If the maximum value is greater than the peak threshold, the audio pulse is retained, otherwise it is discarded.

[0061] Traverse the entire smoothed envelope and capture all valid audio pulse segments that meet the requirements.

[0062] All extracted valid pulse waveform segments are labeled, with insect sound pulse waveforms marked as 1 and noise pulse waveforms marked as 0.

[0063] By adding Gaussian white noise to each valid pulse waveform and randomly aliasing two similar pulses, the extracted valid pulse waveform samples were data enhanced. Finally, 6000 insect sound pulse segments and 6000 noise pulse segments were obtained to form the data set.

[0064] The effective pulse fragments are passed through a bandpass filter bank. The bandpass filter bank consists of 10 bandpass filters, and the passband frequency ranges are 2kHz-3kHz, 3kHz-4kHz, 4kHz-5kHz, 5kHz-6kHz, 7kHz-8kHz, 8kHz-9kHz, 9kHz-10kHz, 10kHz-11kHz, 11kHz-12kHz. The transition bandwidth is 500Hz, and the stopband suppression is 40dB. The waveforms of the effective pulse samples in these 10 frequency bands are obtained.

[0065] Calculate the zero-crossing rate z on 10 frequency bands respectively f (f=1,2,……,10).

[0066] Calculate the average energy in 10 frequency bands respectively where x fi It represents the value of the i-th sampling point of the waveform in the f-th frequency band, and N represents the length of the effective pulse waveform.

[0067] The 10-dimensional average energy feature is input into the LDA algorithm to generate its feature map we.

[0068] The 10-dimensional zero-crossing rate feature is input into the LDA algorithm to generate its feature map wz.

[0069] Select the first two dimensions of we and the first dimension of wz to form a 3D feature vector.

[0070] The data set is randomly divided into a training set and a test set. Both the training set and the test set contain 3000 insect sound pulses and 3000 noise pulses.

[0071] The 3D feature vectors of the effective pulses in the training set are input into the SVM model, which uses the Gaussian kernel function. The model is trained and the hyperplane is obtained.

[0072] The test set is input into the trained SVM model to obtain the predicted label. The pulse with the predicted label of 1 is the insect sound pulse, otherwise it is a noise pulse.

[0073] In this embodiment, the test set predicted label accuracy reached 92.67%.

[0074] In summary, the advantages of the present invention in the wood insect sound collection data provided in the embodiment are: non-destructive detection of wood borers is performed based on the active sound of wood borers, effective extraction of audio features is ensured through multi-band feature extraction, the LDA dimension reduction method is used to reduce the algorithm complexity, and finally, the SVM classification algorithm based on the kernel function realizes non-destructive and efficient detection of whether the wood contains wood borers. In the embodiment test, the accuracy rate exceeded 92%, which significantly improved the efficiency of non-destructive quarantine of wood borers.

Claims

1. A method for nondestructive detection and identification of wood borers based on acoustic signals, characterized in that: The method comprises the following steps: Step 1: Collect wood sound signals: Place the sound amplification device on the surface of the wood sample to collect wood sound signals; Step 2: Detect and extract the effective pulse waveform in the wood sound signal: Extract the envelope of wood sound signal, and then perform two smoothing filters on the extracted envelope; According to the smoothed envelope, set the starting threshold and peak threshold of the effective pulse; Starting from the first data point of the smoothed envelope, traverse the entire envelope and find the point where the envelope line changes from less than or equal to the starting threshold to greater than the starting threshold, which is taken as the pulse starting point; Starting from the pulse starting point, find the point where the envelope changes from greater than the starting threshold to less than or equal to the starting threshold, which is taken as the end point of the pulse; Find the maximum peak value of the envelope between the pulse start point and the pulse end point. If the value is greater than the set peak value threshold of the valid pulse, then this pulse is considered to be a valid pulse waveform, otherwise it is discarded. Extract all valid pulse waveform segments that meet the conditions; Step 3: Perform data enhancement on the extracted effective pulse waveform segment; Step 4: Extract the features of the effective pulse waveform segment: For the detected effective pulse waveform segment, extract the average energy features and zero-crossing rate features in 10 frequency bands respectively, use the linear discriminant analysis method to reduce the dimension of the extracted average energy features and zero-crossing rate features of the effective pulse waveform segment respectively, and combine the signal features after dimensionality reduction into a feature vector; Step 5: Use the SVM model for training and testing: The extracted effective pulse waveform segments are labeled into two categories: "with insects" and "without insects"; Randomly divide the dataset with labeled information into training set and test set; Input the feature vector of the training set data into the SVM model based on the Gaussian kernel function and perform model training; The feature vector of the test set data is input into the trained model, and the test set is classified and tested to obtain the identification result of whether there are insects.

2. The method for nondestructive detection and identification of wood borers based on acoustic signals according to claim 1 is characterized in that: The smoothing filtering method described in step 2 comprises the following steps: Step 21: Find the maximum value: traverse the data points of the input signal, find all the maximum points of the input signal, and record their positions p k and amplitude x k (k=1, 2, 3, ...); Step 22: Average adjacent maxima: For each maximum point, calculate a new value; the new value at the kth maximum Where x0 takes the value of the first data point of the input signal; Step 2 and 3: Linear interpolation: The value y of the pth data point of the filtered signal satisfies the formula: Among them, x′0=x0.

3. The method for nondestructive detection and identification of wood borers based on acoustic signals according to claim 1, characterized in that: The calculation method of the starting threshold and the peak threshold in step 2 is specifically: First, the mean μ and standard deviation σ of the smoothed envelope are calculated, and then two thresholds are calculated. The starting threshold is taken as μ and the peak threshold is taken as μ+2σ.

4. The method for nondestructive detection and identification of wood borers based on acoustic signals according to claim 1 is characterized in that: The calculation method of the average energy feature and the zero-crossing rate feature described in step 4 is specifically: The effective pulse signal is passed through a band-pass filter group consisting of 10 band-pass filters to obtain waveforms in 10 frequency bands, and then the average energy feature and the zero-crossing rate feature are calculated respectively in the 10 frequency bands to obtain a total of 10-dimensional average energy feature and 10-dimensional zero-crossing rate feature; The 10 bandpass filters of the bandpass filter group have passband frequency ranges of 2kHz-3kHz, 3kHz-4kHz, 4kHz-5kHz, 5kHz-6kHz, 7kHz-8kHz, 8kHz-9kHz, 9kHz-10kHz, 10kHz-11kHz, and 11kHz-12kHz respectively; the transition bandwidth is 500Hz, and the stopband suppression is not less than 40dB.

5. The method for nondestructive detection and identification of wood borers based on acoustic signals according to claim 1, characterized in that: The linear discriminant analysis method described in step 4 is specifically: Feature mapping is performed on the 10-dimensional average energy features and 10-dimensional zero-crossing rate features of the effective pulse waveform segment respectively; The linear discriminant analysis method is used to retain the mapping features of the first two dimensions of the average energy characteristics of the effective pulse waveform segment and the mapping features of the first dimension of the zero-crossing rate characteristics.

Images