TWI728632B - Positioning method for specific sound source - Google Patents

Abstract

A positioning method for a specific sound source is provided. Acoustic signals are collected through one sensor at various locations on a preset path. An algorithm is performed on the acoustic signals to obtain multiple signal characteristics. The signal characteristics are used as an input of a deep learning model, and the deep learning model is used for signal identification to obtain multiple specific audio signals at each position. An autocorrelation function operation is performed on these specific audio signals obtained at the same position to obtain multiple autocorrelation coefficients. A representative value is selected from these autocorrelation coefficients as a representative coefficient corresponding to each position. A specific sound source position is found according to the representative coefficient of each position.

Description

Location method of specific sound source

The present disclosure relates to a positioning method, and particularly relates to a positioning method of a specific sound source.

Most of the current methods for detecting the occurrence point of a specific sound source require the inspector to use a handheld sensing device to gradually perform fixed-point sensing, and then use the experience of the inspector to identify whether the sensed signal is a specific audio signal. Once it is confirmed that the specific audio is detected, the inspector will then determine whether the location is the location where the specific audio is emitted based on their experience. However, the identification by the user is easily affected by subjective consciousness, external environment and other factors, which makes the identification accuracy rate poor.

The present disclosure provides a method for locating a specific sound source, which can effectively improve the recognition accuracy.

The method for locating a specific sound source disclosed in the present disclosure includes: collecting a sound wave signal at each of a plurality of positions on a preset path through a sensor; Execute algorithms to obtain multiple signal features; use the signal features as the input of the deep learning model, and use the deep learning model for signal identification to obtain multiple specific audio signals at each location; These specific audio signals are subjected to autocorrelation function operations to obtain multiple autocorrelation coefficients; among these autocorrelation coefficients, representative values are selected as representative coefficients corresponding to each position; and specific audio source positions are found according to the representative coefficients of each position.

In an embodiment of the present disclosure, the step of performing an algorithm on the acoustic signal to obtain signal characteristics includes: performing a Mel cepstrum calculation on the acoustic signal, and converting the obtained Mel cepstrum coefficients and first-order differential Mel inverts. Spectral coefficients and second-order differential Mel cepstral coefficients are used as signal characteristics.

In an embodiment of the present disclosure, the step of finding the position of a specific sound source according to the representative coefficient of each position includes: finding the maximum value among the representative coefficients of each position; and corresponding the representative coefficient as the maximum value. The position is judged to be the position of the specific sound source.

In an embodiment of the present disclosure, the step of finding a specific sound source position based on the representative coefficients of each position includes: calculating the difference between two representative coefficients of two adjacent positions, and when the difference is greater than a threshold, determining two The position corresponding to the larger of the representative coefficients is the specific sound source position.

In an embodiment of the present disclosure, the above-mentioned deep learning model is a convolutional neural network.

In an embodiment of the present disclosure, the sensor obtains a plurality of sampling signals based on the sampling frequency and the sampling period, and uses the sampling signals as the acoustic signal.

In an embodiment of the present disclosure, using the deep learning model to perform signal recognition includes: using the deep learning model to determine whether the sound wave signal belongs to a specific audio signal; and when determining that the sound wave signal belongs to the specific audio signal, using the sampled signal as the specific audio signal.

In an embodiment of the present disclosure, the above-mentioned representative value is the maximum value of the autocorrelation coefficient.

Based on the above, the measured acoustic signal is recognized in real time through the deep learning model, and the signal correlation analysis method is used to locate the location of the signal. Through the measurement of audio, a specific event can be diagnosed in real time and the location of the event can be located to shorten event processing time and improve event processing efficiency.

1~9: location

100: positioning device

100A: host

110: processor

120: storage device

130: Sensor

300A: schematic diagram

300B: curve graph

310: Preset path

D: Moving direction

S205~S230: The steps of the location method of a specific sound source

FIG. 1A is a block diagram of a device for locating a specific sound source according to an embodiment of the disclosure.

FIG. 1B is a block diagram of a device for locating a specific sound source according to another embodiment of the disclosure.

FIG. 2 is a flowchart of a method for locating a specific sound source according to an embodiment of the disclosure.

FIG. 3 is a schematic diagram of detecting a specific audio source according to an embodiment of the disclosure.

FIG. 1A is a diagram of a device for locating a specific sound source according to an embodiment of the present disclosure Block diagram. FIG. 1B is a block diagram of a device for locating a specific sound source according to another embodiment of the disclosure.

In FIG. 1A, the positioning device 100 includes a processor 110, a storage device 120 and a sensor 130. The processor 110 is coupled to the storage device 120 and the sensor 130. The processor 110 is, for example, a central processing unit (CPU), a physical processing unit (PPU), a programmable microprocessor (Microprocessor), an embedded control chip, and a digital signal processor (Digital Signal Processor). Processor, DSP), Application Specific Integrated Circuits (ASIC) or other similar devices.

The storage device 120 is, for example, any type of fixed or removable random access memory (Random Access Memory, RAM), read-only memory (Read-Only Memory, ROM), flash memory (Flash memory), hard disk Dish or other similar device or combination of these devices. One or more code snippets are stored in the storage device 120. After the code snippets are installed, they will be executed by the processor 110 to implement the specific sound source location method described later.

The sensor 130 is used to collect the acoustic signal at each of a plurality of positions on the predetermined path. In one embodiment, only one sensor 130 needs to be provided, but it is not limited to this. A plurality of positions are preset on the preset path. The positioning device 100 starts to move along the preset path from the initial position. Whenever it moves to the set position, the positioning device 100 collects the acoustic signal therethrough through the sensor 130 Then, the processor 110 can use the sound wave signal to find out the location where the specific audio signal is emitted (hereinafter referred to as the specific audio source location).

In another embodiment, as shown in FIG. 1B, the positioning device 100 includes a host 100A and a sensor 130 provided independently. In this embodiment, the processor 110 and the storage device 120 are arranged in the same host 100A, and the sensor 130 is an independent component. After the sensor 130 collects the sound wave signal, it transmits the sound wave signal to the host 100A through a wired or wireless transmission method. Here, one or more code snippets are stored in the storage device 120. After the code snippets are installed, they will be executed by the processor 110 to implement the specific sound source location method described later.

FIG. 2 is a flowchart of a method for locating a specific sound source according to an embodiment of the disclosure. FIG. 3 is a schematic diagram of detecting a specific audio source according to an embodiment of the disclosure. In FIG. 3, the upper half of the schematic diagram 300A is used to show that the sensor 130 collects acoustic signals at a plurality of positions 1 to 9 on the preset path 310, and the lower half of the graph 300B is based on the corresponding position 1 ~The graph obtained by the representative coefficient at position 9.

2 and 3, in step S205, the acoustic signal is collected at each of a plurality of positions 1 to 9 on the predetermined path 310 through the sensor 130. That is, when the sensor 130 moves to the position 1, it will stop and sample to collect the acoustic signal at the position 1. Then, the sensor 130 moves to the position 2 and stops to sample to collect the acoustic signal at the position 2, and so on to collect the acoustic signal at the position 1 to the position 9.

Next, in step S210, an algorithm is performed on the acoustic signal to obtain multiple signal characteristics. The processor 110 executes an algorithm on the acoustic signal collected at each location. In one embodiment, the processor 110 performs Mel-Frequency Cepstrum (MFC) calculation on the acoustic signal, and calculates the obtained Mel-Frequency Cepstrum (MFC) The coefficients, the first-order differential Mel cepstral coefficients, and the second-order differential Mel cepstral coefficients are used as the signal characteristics of the acoustic signal.

After that, in step S215, the processor 110 uses the signal feature as the input of the deep learning model, and uses the deep learning model to perform signal recognition to obtain a plurality of specific audio signals at each location. The deep learning model is, for example, a Convolutional Neural Network (CNN) model. The signal feature extracted in step S210 is used as the input of the CNN model, and the output result of the CNN model is the identification result.

The sensor 110 obtains a plurality of sampling signals based on a sampling frequency and a sampling period, and uses the sampling signals as the acoustic signal. For example, the sampling frequency of the sensor 130 is 8KHz, the sampling period is 1 second, and the number of samples is 185 pen. It shows that the sound wave signal includes 185 sampled signals. After performing steps S210 and S215, it is determined that the 185 sampled signals belong to the specific audio signal, and these sampled signals are regarded as the specific audio signal, and then the autocorrelation function calculation is performed (to be included later). Details).

The type of the specific audio signal is, for example, water leakage sound, traffic sound, electric appliance sound, air leakage sound, or mechanical operation sound, but it is not limited to this. In one embodiment, in the training process, the CNN model is trained by using signal characteristics calculated by Mel cepstrum, such as water leakage sound, traffic sound, electrical sound, air leakage sound, mechanical running sound, and so on. According to this, after inputting the signal characteristics to the CNN model, the CNN model can be used to identify the signal characteristics as water leakage sound, traffic sound, electrical sound, air leakage sound or mechanical operating sound.

For example, input the signal characteristics of the sonic signals of electrical appliance sound, traffic sound and water leakage sound respectively, and identify them through the CNN model, and obtain the identification results shown in Table 1.

In Table 1, there are 53 acoustic signals with the actual category of traffic sounds, and 53 of the predicted categories obtained through the CNN model are traffic sounds. There are 66 acoustic signals with the actual category of electrical sounds, and 66 of the predicted categories obtained through the CNN model are electrical sounds. There are 66 acoustic signals with the actual category of leaking sound, 64 of the predicted categories obtained through the CNN model are leaking sounds, and 2 of them are misjudged as electrical sounds. It can be seen that, in this embodiment, the accuracy of identifying the signal characteristics of the acoustic signal using the CNN model can reach 98.9%.

The deep learning model is used to determine whether the sound wave signal belongs to a specific audio signal. When it is determined that the sound wave signal belongs to a specific audio signal, the sampled signal is used as the specific audio signal. In one embodiment, the main purpose is to detect the location of the water leakage. In this case, the sound wave signal identified as the water leakage sound is selected as the specific audio signal. After that, in step S220, the processor 110 performs an autocorrelation function operation on a plurality of specific audio signals obtained in one position to obtain a plurality of autocorrelation coefficients. After that, in step S225, the processor 110 selects a representative value among these autocorrelation coefficients as the representative coefficient corresponding to each position. The processor 110 uses the autocorrelation coefficients of multiple specific audio signals sampled at the same position to quantize the periodic signal. The strength of the number. Specifically, in one embodiment, the representative value is a maximum value among the autocorrelation coefficients, that is, the maximum value is taken out of the multiple autocorrelation coefficients of multiple specific audio signals sampled at the same position. As the representative coefficient of the position, but not limited to this.

In the example of FIG. 3, it is described that the sound emitted from the position 1 to the position 9 arranged along the movement direction D of the preset path 300 belongs to the specific audio. First, when the sensor 130 moves to position 1 along the moving direction D, at position 1, the processor 110 can determine that the sound wave signal collected by the sensor 130 belongs to a specific audio signal through the CNN model, and then it is directed to the position 1 The sampled multiple specific audio signals are subjected to an autocorrelation function operation, thereby obtaining multiple autocorrelation coefficients corresponding to position 1. After that, the maximum value among the multiple autocorrelation coefficients corresponding to position 1 is taken as the representative coefficient corresponding to position 1. Then, the sensor 130 moves along the moving direction D to position 2, and the processor 110 can determine through the CNN model that the sound wave signal collected by the sensor 130 belongs to a specific audio signal, and then a plurality of specific audio signals sampled at position 2 Perform autocorrelation function calculations to obtain multiple autocorrelation coefficients corresponding to position 2. After that, the maximum value among the multiple autocorrelation coefficients corresponding to position 2 is taken as the representative coefficient corresponding to position 2. By analogy, when moving to position 3~position 9 step by step, the representative coefficient corresponding to position 3~position 9 can be calculated.

After that, in step S230, the processor 110 finds the specific sound source position according to the representative coefficient of each position. Specifically, the maximum value is found among the representative coefficients of the position, and the position corresponding to the representative coefficient that is the maximum value is determined as the specific sound source position. In Fig. 3, position 5 is determined as a specific sound source position.

In another embodiment, the processor 110 may also calculate the difference between two representative coefficients at two adjacent positions, and when the difference is greater than a threshold, determine that the position corresponding to the larger of the two representative coefficients is a specific The location of the sound source. That is, the sound wave signals measured in a certain range (such as position 1 to position 9 shown in Figure 3) are judged to be specific audio signals, and the representative coefficients of two adjacent positions are compared with each other until the difference between the two representative coefficients If the value is greater than the threshold, the position with the largest representative coefficient is the specific sound source position.

The specific sound source location method can be applied to the location of water leakage sound, traffic sound, electrical appliance sound, air leakage sound or mechanical operating sound, and it is not limited here. For example, in the location of water leakage sound, audio characteristics are used for identification. When a specific event (a leak occurs in an underground pipeline), a specific audio is generated due to the pressure change of the substance (liquid or gas) in the pipe.

To sum up, the present disclosure only needs to use a single sensor to measure the acoustic signal at multiple locations, and use the deep learning model to recognize the measured acoustic signal in real time, and use the periodicity of the signal from the source of the specific audio signal. It has strong characteristics to calculate the autocorrelation coefficients of adjacent positions to find out the specific sound source position. According to this, only a single sensor is needed to find out the location where the specific audio is emitted (the location of the specific audio source). In addition, by measuring the sound wave signal, it is possible to diagnose whether a specific event has occurred in real time, and to locate the location of the specific event, so as to shorten the event processing time and improve the efficiency of event processing.

S205~S230: The steps of the location method of a specific sound source

Claims (8)

A method for locating a specific sound source includes: collecting a sound wave signal through a sensor at each of a plurality of positions on a predetermined path; executing an algorithm on the sound wave signal to obtain a plurality of signal characteristics; The signal features are used as the input of a deep learning model, and the deep learning model is used to perform signal recognition to obtain a plurality of specific audio signals at each of the positions; to perform a check on the specific audio signals obtained at the same position Operate the autocorrelation function to obtain a plurality of autocorrelation coefficients; select a representative value among the autocorrelation coefficients as a representative coefficient corresponding to each of the positions; and find out according to the representative coefficient of each of the positions A specific sound source location. For example, in the method for locating a specific sound source described in item 1 of the scope of patent application, the step of executing the algorithm on the sound wave signal to obtain the signal characteristics includes: performing a Mel cepstrum calculation on the sound wave signal to obtain The Mel cepstral coefficients, the first-order differential Mel cepstral coefficients, and the second-order differential Mel cepstral coefficients are used as the signal characteristics. According to the method for locating a specific sound source as described in item 1 of the scope of patent application, the step of finding the position of the specific sound source according to the representative coefficient of each of the positions includes: among the plurality of representative coefficients of the positions Find the maximum value; and determine the position corresponding to the representative coefficient as the maximum value as the specific sound source position Set. For the method for locating a specific sound source as described in item 1 of the scope of patent application, the step of finding the position of the specific sound source according to the representative coefficient of each of the positions includes: calculating two adjacent positions of two said positions. When the difference is greater than a threshold, the position corresponding to the larger of the two representative coefficients is determined to be the specific sound source position. In the method for locating a specific sound source as described in item 1 of the scope of patent application, the deep learning model is a convolutional neural network model. In the method for locating a specific sound source as described in the first item of the patent application, the sensor obtains a plurality of sampling signals based on a sampling frequency and a sampling period, and the sampling signals are used as the sound wave signal. For example, in the method for locating a specific sound source as described in item 6 of the scope of patent application, the step of using the deep learning model for signal identification includes: using the deep learning model to determine whether the sound wave signal belongs to a specific sound signal; and determining the sound wave signal When it belongs to the specific audio signal, the sampled signals are used as the specific audio signals. In the method for locating a specific sound source as described in item 1 of the scope of the patent application, the representative value is a maximum value among the autocorrelation coefficients.

Images