CN106228979A - A kind of abnormal sound in public places feature extraction and recognition methods - Google Patents

Abstract

The invention relates to a method for extracting and identifying abnormal sounds in public places. It improves the pole symmetric mode decomposition (ESMD), referred to as D‑ESMD. The influence of background noise in small public places on the feature extraction of abnormal sounds; Aiming at the problem that the original ESMD has a poor decomposition effect when decomposing abnormal sounds, a symmetrical midpoint interpolation method is proposed to replace the extreme value midpoint parity interpolation method to improve the decomposition efficiency and recognition of abnormal sounds Aiming at the defects of the original ESMD in effective decomposition mode selection, a permutation entropy algorithm is proposed to detect the complexity of the modes obtained by ESMD decomposition, and adaptively obtain the effective mode components of abnormal sounds. The invention can fully describe the characteristics of abnormal sounds, obtain better classification and recognition results, can more accurately extract the characteristics of abnormal sounds, and has better robustness to environmental background noise.

Description

A method for feature extraction and recognition of abnormal sounds in public places

technical field

The invention belongs to the technical field of audio signal feature extraction and pattern recognition, and in particular relates to a method for feature extraction and identification of abnormal sounds in public places.

Background technique

Public places such as squares, bus stations, subways, etc. have the characteristics of large flow of people and wide areas, and the safety precautions of public places have always been widely concerned by governments and people of all countries. At present, the monitoring technology mainly based on video monitoring has played a positive role in the safety of public places. However, video monitoring technology has problems such as monitoring dead spots and blurry monitoring in rainy days. As we all know, when abnormal events occur, they are often accompanied by abnormal sounds such as screams, gunshots, broken glass, explosions, etc. Therefore, the cooperative operation of audio monitoring and video monitoring has become the development direction of the field of public place security monitoring. At present, the existing audio monitoring system only includes simple sound collection and transmission, etc., and lacks effective identification of abnormal sounds. The reason is that the core theory and technology of audio monitoring have not been broken through. Abnormal sound recognition technology in public places is the core technology of the audio monitoring system. Therefore, the research on this technology has important social significance and research value.

At present, there are problems in extracting abnormal sound features in public places by Extreme-point Symmetric Mode Decomposition (ESMD) method: ① Abnormal sounds in public places are composed of two parts: abnormal sound signal and background noise signal. The background noise signal will block the local features of the abnormal sound. Using ESMD to decompose the abnormal sound in public places, the modal components obtained must contain the background noise component, which will lead to deviations in the characteristics of the abnormal sound. ② When ESMD decomposes the signal, it constructs 1, 2, 3 or more interpolation curves according to the midpoint of the extreme value to improve the decomposition effect, that is, ESMD_I, ESMD_II, ESMD_III methods. Since the interpolation method has a great influence on the effect of mode decomposition, by comparing the above three interpolation methods, it is found that with the increase of the interpolation line, the number of modes will decrease, the degree of symmetry will decrease, the amplitude change will increase, and the decomposition efficiency will increase. ③ When ESMD decomposes abnormal sounds in public places with background noise, it uses the remaining number of extreme points as the judgment condition for the termination of the decomposition mode, which will keep the decomposed low-frequency noise and obtain pseudo-feature components that do not belong to abnormal sounds . ④ When the ESMD method decomposes abnormal sound signals in public places, set the number of screening K to change within the interval [K _min , K _max ], then use different screening times to repeatedly decompose abnormal sound signals, and finally use the principle of least squares to calculate the best screening times, it takes a long time for ESMD to decompose abnormal sound signals.

To sum up, ESMD decomposition technology has some room for improvement.

Contents of the invention

Aiming at the deficiencies in the prior art above, the present invention proposes a public place abnormal sound feature extraction and recognition method based on the improved ESMD (D-ESMD) decomposition technology, by adding noise to the input signal of ESMD, the internal interpolation method and decomposition model are improved. The characteristics of abnormal sounds in public places at different scales are obtained by using the judgment conditions of modal termination and the screening times of modal components.

A method for feature extraction and identification of abnormal sounds in public places, the specific implementation steps are as follows:

Step 1: Input the abnormal sounds to be identified in public places and perform preprocessing.

Step 2: Use the improved pole-symmetric mode D-ESMD decomposition method to decompose the abnormal sound signal to be identified to obtain the modal components of each order, and each order modal component contains the characteristics of the abnormal sound signal in different frequency bands.

Step 3: Calculate the energy ratio of the modal components of each order obtained in step 2 relative to the original abnormal sound signal, and combine them into a vector form for normalization, as the feature vector of the abnormal sound signal to be identified.

Step 4: Determine whether the feature vector is valid; if not, go to step 3; if valid, go to step 5.

Step 5: The identification process of abnormal sounds to be identified in public places: First, randomly select each type and a certain number of training samples in the abnormal sound library that has been established, and obtain the eigenvectors of the training samples through steps 2 and 3 and Establish the SVM classification model; then, use the established SVM classification model to classify the feature vectors of the abnormal sounds to be recognized, and obtain the classification recognition results.

The D-ESMD decomposition method is based on the pole symmetric mode ESMD decomposition method, adding a random T-distributed noise sequence to the abnormal sounds to be identified in public places, and using the symmetrical midpoint interpolation method to replace the extreme value midpoint parity interpolation of ESMD method, calculate the permutation entropy value for the decomposed modal components, and improve the modal component screening times, and then complete the complexity detection of each mode, and adaptively obtain the effective modal components of abnormal sounds.

The library of abnormal sounds includes explosions, screams, gunshots, and broken glass.

Specifically, the concrete process of described D-ESMD decomposition method is:

Step 2.1 determines the number of times N to add T-distributed random noise;

Step 2.2 Assuming that the abnormal sound signal to be identified is x, add a random T distribution sequence to the sound signal x to be identified, and obtain the abnormal sound signal Xi with noise _added ;

Step 2.3 Find the extreme points of the abnormal sound signal X _i after adding noise, connect the adjacent extreme points, mark the midpoint of the line segment as F _i , supplement the left and right boundary points F ₀ and F _n , and adopt the symmetrical midpoint interpolation method The extreme midpoint odd-even interpolation method that replaces ESMD constructs an interpolation curve L ^* for n+1 extreme midpoints;

Step 2.4 takes X _i -L ^* as input, repeats the above step 2.3 until the number of screening reaches the maximum value, and obtains the first-order modal component M ₁ ⁱ , and calculates the value of the permutation entropy of the modal component; if the permutation entropy of the signal is If the value is greater than the agreed threshold θ, it is considered to be an abnormal sound modal component, otherwise it is considered to be a noise component;

Step 2.5 If the modal component M ₁ ⁱ is an abnormal sound modal component, then take Xi-M _{1 i} as the input signal, and repeat steps ^2.3-2.4 until the decomposed modal component M _n ⁱ is a noise component;

Step 2.6 If i<N, set i=i+1, repeat step 2.2 to step 2.5, the T distribution noise signal added each time is different, until N times of decomposition are performed, and all modal components are obtained Calculate the overall average and use the result as the final modal component M _k of the signal to be decomposed:

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}$

In the above formula, k is the order of the modal component, and N is the number of times of adding noise.

Specifically, the specific steps of the symmetric midpoint interpolation method are:

Step 3.1, assuming that the input signal is y, find all the maximum value points y _max and minimum value points y _min of y;

Step 3.2, connect all adjacent extreme points, and calculate the extreme midpoint y _mean ;

y _mean ＝(y _max +y _min )/2

In step 3.3, the symmetrical midpoint y _m of the adjacent extreme midpoints is obtained, and the cubic spline interpolation method is used to interpolate y _m to obtain the final interpolation curve.

Specifically, the optimal number of screening times in step 2.4 is 12.

Specifically, the specific calculation process of permutation entropy is as follows:

Assuming a time series signal x(i) with length N, i=1, 2, ..., N, and reconstructing it with delay, the following time series is obtained:

In the formula, l is the time delay, m is the reconstruction dimension, and the m elements in X(i) are arranged in ascending order to obtain:

X _i '={x(i+(j ₁ -1)*l)≤x(i+(j ₂ -1)*l)

≤…≤x(i+(j _m -1)*l)}

Therefore, each vector X(i) has a set of permutations:

Sg＝{j ₁ ,j ₂ ,j ₃ ,…j _m }

In the formula, j represents the index of the column where each element in the reconstruction component is located.

Among them, there must be m for m different symbols! Different permutations; calculate the probability p ₁ , p ₂ , ... p ₃ of each permutation in X(i), then the normalized permutation entropy is:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; lgp</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; lg</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; !</span><span\; class="notranslate">\; )</span>$

where N is the time series length, m is the reconstruction dimension and l is the time delay.

The effective gain effect is:

When the present invention decomposes abnormal sounds in public places based on D-ESMD, random T-distributed noise sequences are added to the abnormal sound signals in public places to be decomposed, and the decomposition deviation caused by background noise is reduced from the source, so that abnormal sounds in public places Voice recognition ability has been greatly improved. Moreover, the present invention combines the characteristics of abnormal sounds and background noise in public places, and proposes the D-ESMD method for feature extraction and identification of abnormal sounds in public places, and decomposes abnormal sounds in public places into a series of modal components with relatively single frequency components. Theoretically, the internal interpolation method of ESMD, the judgment condition of decomposition modal termination and the screening times of modal components are improved to ensure that the modal components obtained by decomposition can reflect the characteristics of abnormal sounds in public places at different scales.

Description of drawings

Fig. 1: A flow chart of a method for feature extraction and identification of abnormal sounds in public places proposed by the present invention;

Figure 2: The ESMD interpolation method decomposes the analog signal diagram;

Fig. 3: the improved interpolation method proposed by the present invention decomposes the analog signal diagram;

Fig. 4: Comparison of Receiver Operating Characteristic (ROC) curves between the present invention and several other abnormal sound feature extraction methods.

detailed description

The present invention is further described in detail below in conjunction with the accompanying drawings.

The core technology of the present invention is the D-ESMD decomposition method. The D-ESMD decomposition method is an improvement based on the ESMD decomposition method, and its improvement points are:

1. The ESMD decomposition method based on T distribution is used to weaken the background noise component in the modal component, so as to better extract the characteristics of abnormal sounds. Specific steps are as follows:

Add random T distribution sequence to the sound signal to be recognized, weaken the background noise component in the modal component, reduce the decomposition deviation caused by the background noise from the source, and improve the feature extraction ability of abnormal sound. The specific processing process is:

Assuming that the abnormal sound signal in public places is X(t), it generally consists of the real abnormal sound signal x(t) and the background noise signal N(t), namely:

X(t)=x(t)+N(t)

When X(t) is decomposed by ESMD, the obtained mode M(t) also contains abnormal sound signal component m(t) and background noise signal component c(t), namely:

$\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>$

In the formula, n is the number of modal components, and r(t) is the decomposition remainder.

After adding k times of different T noise sequences n _i (t) to the signal X(t), the series of formulas can be expressed as:

X(t)+n ₁ (t)=m ₁₁ (t)+m ₁₂ (t)+...+m _1n (t)+c ₁₁ (t)+c ₁₂ (t)+...+c _1n (t) +r ₁ (t)

X(t)+n ₂ (t)＝m ₂₁ (t)+m ₂₂ (t)+...+m _2n (t)+c ₂₁ (t)+c ₂₂ (t)+...+c _2n (t) +r ₂ (t)

………

X(t)+n _i (t)＝m _i1 (t)+m _i2 (t)+…+m _in (t)+c _i1 (t)+c _i2 (t)+…+c _in (t) +r _i (t)

………

X(t)+n _k (t)＝m _k1 (t)+m _k2 (t)+…+m _kn (t)+c _k1 (t)+c _k2 (t)+…+c _kn (t) +r _k (t)

Adding up the above N formulas, we can get:

$\begin{array}{c}\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; ...</span>}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; ...</span>}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>\end{array}$

It can be seen from the above formula that when k～∞, k*N(t)+n ₁ (t)+n ₂ (t)+…n _k (t) and c _ij (t) both tend to zero, then The above formula is transformed into the following:

$\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>$

It can be seen from the above formula that by adding k times of random T-distributed noise sequences to abnormal sounds in public places, and using ESMD to take the mean value of each order mode obtained by its decomposition, the background noise component c(t) has been eliminated, thereby reducing The effect of background noise in public places on the decomposition of abnormal sounds.

2. Use symmetrical midpoint interpolation instead of extreme midpoint parity interpolation to improve ESMD decomposition efficiency and decomposition accuracy from the source of the signal.

The symmetric midpoint interpolation method is:

Step 3.1 Obtain all maximum value points y _max and minimum value points y _min of the original signal;

Step 3.2 connect all adjacent extreme points, and find the extreme midpoint y _mean ;

y _mean ＝(y _max +y _min )/2

Step 3.3 Calculate the symmetrical midpoint y _m of the adjacent extreme midpoints, and at the same time use the cubic spline interpolation method to interpolate y _m to obtain the final interpolation curve.

Do not use symmetrical midpoint interpolation and extreme point parity interpolation to decompose the analog signal z. Assume that the analog signal z consists of three sinusoidal signals with different frequencies and different amplitudes, as follows:

z=sin(20*p*t)+1.5cos(40*π*t)+2.5cos(80*π*t)

It can be seen from Figure 2 that when the ESMD interpolation method is used to decompose the analog signal, the resulting mode appears distorted, and the amplitude deviation between the mode and the original signal is large. Fig. 3 The improved interpolation method proposed by the present invention decomposes the analog signal diagram, which effectively alleviates the distortion problem caused by the ambiguity of the end points of ESMD interpolation.

3. Based on the permutation entropy algorithm, the complexity detection of the modal components obtained by ESMD decomposition is carried out, and this is used as the judgment criterion for distinguishing abnormal sounds from background noise, and the effective abnormal sound components are adaptively obtained.

The specific calculation process of permutation entropy is as follows:

Assuming a time series signal x(i) with length N, i=1, 2, ..., N, and reconstructing it with delay, the following time series is obtained:

In the formula, l is the time delay, m is the reconstruction dimension, and the m elements in X(i) are arranged in ascending order, and we can get:

X _i '={x(i+(j ₁ -1)*l)≤x(i+(j ₂ -1)*l)

≤…≤x(i+(j _m -1)*l)}

Therefore, each vector X(i) has a set of permutations:

Sg＝{j ₁ ,j ₂ ,j ₃ ,…j _m }

In the formula, j represents the index of the column where each element in the reconstruction component is located.

Among them, there must be m for m different symbols! different permutations. Calculate the probability p ₁ , p ₂ , ... p ₃ of each arrangement in X(i), then the normalized arrangement entropy is:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; lgp</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; lg</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; !</span><span\; class="notranslate">\; )</span>$

where N is the time series length, m is the reconstruction dimension and l is the time delay. According to the experimental results, the reconstruction dimension m is generally selected from 3 to 7. The time delay has little effect on the permutation entropy and can generally be selected as 1.

In the present invention, the choice of the mode is determined by determining whether the permutation entropy value H of the modal components of different frequency scales obtained by decomposing the abnormal sound signal in a public place with a random T distribution sequence is greater than the threshold θ. After experiments, it is found that when the value of θ is in the range of 0.25-0.35, the effect of extracting abnormal sound features is better.

Fourth, the number of modal component screening

The number of modal screenings is determined through a large number of experiments to determine the optimal number of screenings, and the preferred value is 12.

The present invention utilizes the above improvements to realize feature extraction and recognition of abnormal sounds in public places. As shown in Figure 1, the method mainly includes three parts: including: decomposition, feature extraction and recognition of abnormal sounds to be recognized in public places.

The specific steps of the method are as follows:

Step 1: Input the abnormal sound signal to be identified in public places and perform preprocessing.

Step 2: Decompose the abnormal sound signal to be identified into a series of modal components by using the improved pole-symmetric mode D-ESMD decomposition method, and each modal component contains the characteristics of the abnormal sound signal in different frequency bands.

Step 3: Calculate the energy ratio of the modal components of each order obtained in step 2 relative to the original abnormal sound signal, and combine them into a vector form for normalization, as the feature vector of the abnormal sound signal to be identified.

Step 4: Determine whether the feature vector is valid; if invalid, go to step 3; if valid, go to step 5;

Step 5: The identification process of abnormal sounds to be identified in public places: First, randomly select each type and a certain number of training samples in the abnormal sound library that has been established, and obtain the eigenvectors of the training samples through steps 2 and 3 and Establishing an SVM classification model; then, using the established SVM classification model to classify the feature vectors of the abnormal sounds to be identified, to obtain classification recognition results;

Wherein, the specific steps of the D-ESMD used in feature extraction of abnormal sounds to be identified in public places are as follows:

Step 2.1 determines the number of times N to add T-distributed random noise;

Step 2.2 Assuming that the abnormal sound signal to be identified is x, add a random T distribution sequence to the sound signal x to be identified, and obtain the abnormal sound signal Xi with noise _added ;

Step 2.3 Find the extreme points of the abnormal sound signal X _i after adding noise, connect the adjacent extreme points, mark the midpoint of the line segment as F _i , and supplement the left and right boundary points F ₀ and F _n . A symmetrical midpoint interpolation method is used to replace the ESMD's extreme midpoint even-even interpolation method to construct an interpolation curve L ^* for n+1 extreme midpoints.

Step 2.4 takes X _i -L ^* as input, repeats the above step 2.3 until the number of screening times reaches the maximum value of 12 to obtain the first-order modal component M ₁ ⁱ , and calculates the value of the permutation entropy of the modal component; if the permutation entropy of the signal is If the value is greater than the agreed threshold θ, it is considered to be an abnormal sound modal component, otherwise it is considered to be a noise component;

Step 2.5 If the modal component M ₁ ⁱ is an abnormal sound modal component, then take Xi-M _{1 i} as the input signal, and repeat steps ^2.3-2.4 until the decomposed modal component M _n ⁱ is a noise component;

Step 2.6 If i<N, set i=i+1, repeat step 2.2 to step 2.5, the T distribution noise signal added each time is different, until N times of decomposition are performed, and all modal components are obtained Calculate the overall average and use the result as the final modal component M _k of the signal to be decomposed:

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}$

In the above formula, k is the order of the modal component, and N is the number of times of adding noise.

Fig. 4 is a comparison chart of ROC curves between the present invention and several other abnormal sound feature extraction methods. Among them, ESMD is the pole symmetric mode decomposition method, EEMD is the overall empirical mode decomposition method, SaSEMD is the overall empirical mode decomposition method based on α distribution, and ELMD is the overall local mean decomposition method. D-ESMD is an improved ESMD decomposition method proposed by the present invention.

Claims (5)

1. an abnormal sound in public places feature extraction and recognition methods, it is characterised in that including: public place is to be identified different Chang Shengyin decomposition, feature extraction and identification；Implement step as follows:

Step 1: input public place abnormal sound to be identified also carries out pretreatment；

Step 2: use the pole symmetric mode decomposition D-ESMD method improved to be decomposed by abnormal sound signal to be identified, To each rank modal components, every rank modal components comprises the abnormal sound signal feature in different frequency section respectively；

Step 3: each rank modal components obtained in calculation procedure 2 is relative to the energy ratio of original anomaly acoustical signal, and combines Vector form is become to be normalized, as the characteristic vector of abnormal sound signal to be identified；

Step 4: judging characteristic vector is the most effective；If invalid, jump to step 3；If effectively, performing step 5；

Step 5: the identification process of public place abnormal sound to be identified: first, random in the abnormal sound storehouse having built up Choose each class and a number of training sample, ask for the characteristic vector of its training sample by step 2 and step 3 and build Vertical svm classifier model；Then, utilize the svm classifier model set up that the characteristic vector of abnormal sound to be identified is classified, To Classification and Identification result；

Described D-ESMD decomposition method is on the basis of pole symmetric mode ESMD decomposition method, adds random T partition noise Sequence is in the middle of public place abnormal sound to be identified, and the extreme value midpoint odd even using symmetrical Point Interpolation method to substitute ESMD is inserted Value method, to the modal components calculated permutations entropy decomposed, and improves modal components screening number of times, and then completes each mode Complexity detects, and self adaptation obtains the effective modal components of abnormal sound；

Described abnormal sound storehouse includes explosive sound, shriek, shot, glass breaking sound.

A kind of abnormal sound in public places feature extraction the most according to claim 1 and recognition methods, it is characterised in that institute The detailed process of the D-ESMD decomposition method stated is:

Step 2.1 determines interpolation T distributed random noise times N；

Step 2.2 assumes that abnormal sound signal to be identified is x, adds random T distribution series in acoustical signal x to be identified, Obtain adding the abnormal sound signal X made an uproar_i；

Step 2.3 asks for adding abnormal sound signal X after making an uproar_iExtreme point, connect adjacent extreme point, and by line segment midpoint labelling For F_i, supplement left and right boundary point F₀With F_n, use symmetrical Point Interpolation method to substitute the extreme value midpoint odd even interpolation method pair of ESMD N+1 extreme value midpoint structure interpolation curve L^*；

Step 2.4 is by X_i-L^*As input, repeat the above steps 2.3, until screening number of times reaches maximum, obtains the first order mode State component M₁ ⁱ, calculate the value of the arrangement entropy of modal components；If the value of the arrangement entropy of this signal is more than agreement threshold θ, then it is assumed that It is abnormal sound modal components, otherwise it is assumed that be noise component(s)；

If step 2.5 modal components M₁ ⁱFor abnormal sound modal components, then by X_i-M₁ ⁱAs input signal, repeat step 2.3- Step 2.4, until decomposing modal components M obtained_n ⁱTill noise component(s)；

If step 2.6 i < N, then making i=i+1, repetition step 2.2 to step 2.5, the T partition noise signal every time added is different, Till carrying out n times decomposition, to obtaining all of modal components M_k ⁱSeek population mean, and using result as signal to be decomposed Final modal components M_k:

$M_k=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{M}_k^i$

In above formula, k is modal components exponent number, and N is for adding number of times of making an uproar.

A kind of abnormal sound in public places feature extraction the most according to claim 1 and 2 and recognition methods, its feature exists In, described symmetrical Point Interpolation method concretely comprises the following steps:

Step 3.1, assume that input signal is y, ask for all maximum point y of y_maxWith minimum point y_min；

Step 3.2, connect all adjacent extreme points, and ask for extreme value midpoint y_mean；

y_mean=(y_max+y_min)/2

Step 3.3, ask for the symmetrical midpoint y at adjacent extreme value midpoint_m, use cubic spline interpolation method to y simultaneously_mCarry out interpolation, Obtain final interpolation curve.

A kind of abnormal sound in public places feature extraction the most according to claim 2 and recognition methods, it is characterised in that step The maximum preferential 12 of number of times is screened in rapid 2.4.

A kind of abnormal sound in public places feature extraction the most according to claim 2 and recognition methods, it is characterised in that row The concrete calculating process of row entropy is as follows:

Assume time series signal x (i) of an a length of N, i=1,2 ..., N, it is carried out delay reconstruction, when obtaining following Between sequence:

In formula, l is time delay, and m is reconstruct dimension, m element in X (i) is carried out ascending order arrangement, obtains:

X′_i={ x (i+ (j₁-1)*l)≤x(i+(j₂-1)*l)

≤…≤x(i+(j_m-1)*l)}

Therefore, each vector X (i) is owned by one group of collating sequence:

Sg={j₁,j₂,j₃,…j_m}

In formula, j represents the index of each element column in reconstruct component；

Wherein, m is necessarily had for m different symbol！Plant different arrangements；Calculate every kind of arrangement mode to occur in X (i) Probability p₁、p₂、…p₃, then the arrangement entropy after normalization is:

$H=(-\underset{i=1}{\overset{k}{\&Sigma;}}{p}_i*{\mathrm{lgp}}_i)\lg (m!)$

Wherein, N is length of time series, and m is time delay for reconstruct dimension and l.