CN114740091A - Watermelon maturity detection method and system based on acoustic analysis and machine learning - Google Patents

Abstract

The invention discloses a watermelon maturity detection method and system based on acoustic analysis and machine learning. The method comprises: acquiring a data set composed of a tap sound signal and weight of a watermelon sample, classifying the watermelon samples according to maturity, and classifying the data set Divide into training set and test set; build a maturity detection model to calculate the maturity of the target watermelon sample; train the maturity detection model with the training set, then test the maturity detection model with the test set, and finally calculate the accuracy and record it, and continue to execute The steps of training the maturity detection model with the training set until the accuracy is the highest, and the trained maturity detection model is obtained; the tap sound signal and weight of the watermelon test sample are obtained, and the trained maturity detection model is input to obtain the watermelon test The maturity of the sample. The invention fulfills the requirement of watermelon maturity detection, improves the detection accuracy, can complete the training and testing of the model in the small sample data set, and improves the applicability.

Description

Watermelon maturity detection method and system based on acoustic analysis and machine learning

technical field

The invention relates to the field of audio analysis, in particular to a watermelon maturity detection method and system based on acoustic analysis and machine learning.

Background technique

my country's watermelon production is large, but there are problems such as insufficient competitiveness, small export volume and low export unit price in the world market. The reason is that my country's watermelon maturity detection system is backward, resulting in mixed watermelons of various maturity levels, reducing the overall quality.

At present, the detection method based on acoustic characteristics is mainly used for the detection of watermelon maturity, because the sound signal generated by tapping the watermelon is easier to obtain, which can reflect the watermelon pulp hardness, absorption and reflection characteristics, size, water content and other information, so that it can be judged Ripeness of watermelon.

At present, most of the detection methods based on acoustic characteristics use the acoustic analysis method of a single frequency point. First, the Fourier transform or fast Fourier transform is used to analyze the sound signal in the frequency domain to extract a single frequency point, and then the single frequency point is compared with the single frequency point. Mathematical operations are performed on the quality characteristics to obtain evaluation indicators. By comparing the evaluation indicators and the size of the set range, the maturity level is determined. A large number of data samples are needed to determine the evaluation indicator range. Because it is difficult to express the acoustic feature distribution of the sample with a single frequency point, it is only a maximum frequency component in the frequency domain, ignoring other frequency domain features, and completely missing the variation of the sound signal in the time domain. At the same time, the size of the evaluation index is easily affected by the error of the calculation result of a single frequency point, because it is only obtained through a simple mathematical operation relationship. Therefore, the current detection methods based on acoustic characteristics have poor applicability, are easily affected by the error of a single frequency point of the sample, and have low accuracy.

However, other watermelon ripening detection methods, such as those based on infrared spectrum detection and machine vision detection, have shortcomings such as complex detection equipment, high cost, and difficulty in obtaining evaluation indicators.

Machine learning is one of the most intelligent features of artificial intelligence and one of the most cutting-edge research areas, which can solve problems that are too complicated or no known algorithms exist using traditional methods. Its application fields are wide and have good application prospects. How to apply machine learning methods to watermelon maturity detection is also a problem worth studying.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is: in view of the technical problems existing in the prior art, the present invention provides a watermelon maturity detection method and system based on acoustic analysis and machine learning, which realizes the demand for watermelon maturity detection and improves the Detection accuracy, the training and testing of the model can be completed in a small sample data set, which improves the applicability.

In order to solve the above-mentioned technical problems, the technical scheme proposed by the present invention is:

A watermelon maturity detection method based on acoustic analysis and machine learning, comprising the following steps:

Obtain the tap sound signal and weight of the watermelon samples, and form a data set, classify the watermelon samples in the data set according to maturity, and divide the data set into a training set and a test set;

Build a maturity detection model, the maturity detection model is used to extract the time-frequency features of the watermelon sample percussion sound signal, and then calculate the distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency characteristics and the corresponding weight, Finally, among the categories corresponding to the K watermelon samples that are closest to the target watermelon sample, the category with the highest frequency is used as the maturity of the target watermelon sample;

Use the training set to train the maturity detection model, then use the test set to test the maturity detection model, finally calculate the accuracy and record, repeat this step until the accuracy is the highest, and obtain the trained maturity detection model;

Obtain the tap sound signal and weight of the watermelon test sample, and input the trained maturity detection model to obtain the maturity of the watermelon test sample.

Further, the specific steps of extracting the time-frequency characteristics of the watermelon sample percussion sound signal include:

Calculate the FFT spectrum and extract the spectral features according to the tapping sound signals of all the watermelon samples;

； According to the extracted spectral features, the value range of the spectral resolution parameter f of the STFT window function is determined, and the STFT window function is constructed according to the values of the time resolution parameter τ and the spectral resolution parameter f

;

和每个西瓜样本的敲击声音信号，计 算每个西瓜样本的时频矩阵。 According to the STFT window function

and the percussion sound signal of each watermelon sample, calculate the time-frequency matrix of each watermelon sample.

Further, the functional expression of the time-frequency matrix is:

为所构建的STFT 窗函数，其中τ为时间分辨参数，f为频谱分辨率参数，[f _i , f _i+1]之间的频谱间隔代表频谱 分辨率的带宽。 In the above formula, x ( t ) is the percussion sound signal of the watermelon sample,

is the constructed STFT window function, where τ is the time resolution parameter, f is the spectral resolution parameter, and the spectral interval between [ f _i , f _{i +1} ] represents the bandwidth of the spectral resolution.

Further, the specific steps for calculating the distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency feature and weight include:

Calculate the time-frequency matrix distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency matrix of the target watermelon sample and other watermelon samples in the training set, and calculate the target watermelon sample according to the weight of the target watermelon sample and other watermelon samples in the training set. The weight feature distance from other watermelon samples in the training set;

Normalize the time-frequency matrix distance and weight feature distance between the target watermelon sample and other watermelon samples in the training set;

According to the information fusion weight parameter λ , as well as the normalized time-frequency matrix distance and weight feature distance, the distance between the target watermelon sample and other watermelon samples in the training set is calculated.

Further, other watermelon samples in the training set include the kth nearest watermelon sample of the target watermelon sample, and the function expression of the distance between the target watermelon sample and the kth nearest watermelon sample is:

In the above formula, λ is the information fusion weight parameter, m _k is the time-frequency matrix distance between the target watermelon sample and the kth nearest watermelon sample, n _k is the distance between the target watermelon sample and the kth nearest watermelon sample. Weight feature distance.

Further, before using the training set to train the maturity detection model, it also includes the steps of optimizing the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ , specifically including:

If the values of the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ do not reach the preset maximum value, increase the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter with a preset step size The value of λ , executes the step of training maturity detection model with described training set;

If the value of the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ reaches the preset maximum value, find the highest accuracy from all the recorded accuracies, and use the time resolution parameter τ , the spectral resolution parameter The values of f and the information fusion weight parameter λ are adjusted to the value corresponding to the highest accuracy, and the step of training the maturity detection model with the training set is performed.

Further, after acquiring the tapping sound signal of the watermelon sample, it also includes the step of extracting an effective signal, which specifically includes:

Calculate the energy of each frame of the percussion sound signal, retain the frames whose energy is greater than the preset threshold, and obtain a filtered percussion sound signal;

The percussion sound signal filtered once is filtered to obtain the percussion sound signal filtered for the second time, which is used as an effective signal of the percussion sound signal.

The present invention also proposes a watermelon maturity detection system based on acoustic analysis and machine learning, including:

a data acquisition unit for acquiring the tap sound signal and weight of the watermelon samples, and forming a data set, classifying the watermelon samples in the data set according to maturity, and dividing the data set into a training set and a test set; and It is used to obtain the tap sound signal and weight of the watermelon test sample, and input the trained maturity detection model to obtain the maturity of the watermelon test sample;

The model building unit is used to build a maturity detection model, and the maturity detection model is used to extract the time-frequency characteristics of the tap sound signal of the watermelon sample, and then calculate the target watermelon sample and other watermelons in the training set according to the time-frequency characteristics and the corresponding weight. The distance between samples, and finally, in the categories corresponding to the K watermelon samples that are closest to the target watermelon sample, the category with the highest frequency is used as the maturity of the target watermelon sample;

The model training and testing unit is used to train the maturity detection model with the training set, then use the test set to test the maturity detection model, finally calculate the accuracy and record, and continue to use the training set to train the maturity detection model, Until the accuracy is the highest, the trained maturity detection model is obtained.

Further, when the maturity detection model extracts the time-frequency features of the watermelon sample tapping sound signal, the model building unit is configured to perform the following steps:

Calculate the FFT spectrum and extract the spectral features according to the tapping sound signals of all the watermelon samples;

According to the extracted spectral characteristics, determine the value range of the spectral resolution parameter f of the STFT window function, and construct the STFT window function according to the values of the time resolution parameter τ and the spectral resolution parameter f ;

According to the STFT window function and the percussion sound signal of each watermelon sample, the time-frequency matrix of each watermelon sample is calculated.

Further, when the maturity detection model calculates the distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency feature and weight, the model building unit is configured to perform the following steps:

Calculate the time-frequency matrix distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency matrix of the target watermelon sample and other watermelon samples in the training set, and calculate the target watermelon sample according to the weight of the target watermelon sample and other watermelon samples in the training set. The weight feature distance from other watermelon samples in the training set;

Normalize the time-frequency matrix distance and weight feature distance between the target watermelon sample and other watermelon samples in the training set;

According to the information fusion weight parameter λ , as well as the normalized time-frequency matrix distance and weight feature distance, the distance between the target watermelon sample and other watermelon samples in the training set is calculated.

Compared with the prior art, the advantages of the present invention are:

The invention realizes the requirement of watermelon maturity detection based on acoustic analysis, extracts the time-frequency characteristic matrix including time information and frequency information as the acoustic characteristic of the sample, and solves the problem that a single frequency point is easily interfered as an acoustic characteristic. Moreover, the calculation formula of the time-frequency matrix is optimized by improving the STFT window function, which reduces the dimension of the time-frequency feature matrix, reduces the subsequent calculation amount, and provides features that are more representative of the sample for the subsequent classification algorithm.

The invention integrates the time-frequency feature matrix and the quality feature for machine learning classification, establishes the relationship between various information features and watermelon maturity, and solves the problems that the dimensions of the various information features are not uniform and the order of magnitude difference is large. Machine learning classification that combines time-frequency feature matrix and quality feature no longer requires a large number of samples to set the evaluation index range, and can be trained and tested on a small sample data set, which can achieve higher accuracy and improve applicability .

Description of drawings

FIG. 1 is a technical architecture diagram of Embodiment 1 of the present invention.

FIG. 2 is a flowchart of Embodiment 1 of the present invention.

FIG. 3 is a working flowchart of a maturity detection model according to Embodiment 1 of the present invention.

FIG. 4 is a comparison diagram of the time-frequency feature extracted in the first embodiment of the present invention and the time-frequency feature extracted by the conventional method.

FIG. 5 is a comparison diagram of the original signal and the effective signal of the percussion sound signal in the first embodiment of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

Example 1

There are many time-frequency analysis methods based on acoustic analysis, such as Short-Time Fourier Transform (STFT), wavelet transform, S-transform, etc. Among them, STFT is widely used because of its fast calculation speed and good effect of extracting sound features. application. STFT can cut the sound signal into frame signals of a specified time length through a window function, and perform Fast Fourier Transform (FFT) on each frame signal to obtain frequency domain features, and convert the frequency domain features of these frames. Combined, a two-dimensional matrix representing time-frequency features is obtained. However, since the window function of STFT is fixed, the time resolution is unchanged, and the spectral resolution of FFT is also determined, if the signal features are mainly concentrated in the low-frequency region, it will cause data redundancy in the high-frequency region, making the subsequent calculation speed. slow down. At the same time, the time resolution of the window function is also an important factor affecting the quality of the evaluation index, and its determination needs to be adaptively changed according to the sample.

There are many machine learning classification algorithms, such as K-Nearest Neighbor (KNN), support vector machine, logistic regression, random forest, etc. Among them, KNN has the advantages of simple principle and easy calculation, which is suitable for watermelon maturity detection. KNN makes judgments by calculating the distance between the test sample features and the training sample features. In the process of distance calculation, the same weight is assigned to each feature of the sample, which will cause some important feature points to be ignored, thereby reducing the classification accuracy. In the watermelon maturity detection, the two-dimensional matrix of time-frequency features and the weight feature obtained by acoustic analysis are too different in data dimension and order of magnitude, but they have an important impact on the final classification result. It is necessary to balance the two-dimensional matrix and the weight feature. Influence weight, which cannot be achieved by traditional KNN classification algorithm.

In order to solve the above problems, as shown in Figure 1, this embodiment proposes a watermelon maturity detection method based on acoustic analysis and machine learning, establishes an optimized STFT time-frequency feature extraction algorithm and an information fusion KNN classification algorithm, and combines them to achieve To meet the needs of watermelon maturity detection, improve the detection accuracy, and complete the training and testing of the system in a small sample data set to improve the applicability.

As shown in Figure 2, the method for detecting the maturity of watermelon based on acoustic analysis and machine learning in this embodiment includes the following steps:

S1) Obtain the percussion sound signal and weight of the watermelon samples, and form a data set. The watermelon samples in the data set are classified according to their maturity, which are generally divided into [Immature, Ripe] or [Immature, Suitable Ripe, Overripe], etc. category;

S2) Divide the dataset into training set and test set;

S3) Build a maturity detection model, as shown in Figure 3, the maturity detection model uses the optimized STFT time-frequency feature extraction algorithm to extract the time-frequency characteristics of the watermelon sample tapping sound signal, and also uses the information fusion KNN classification algorithm to classify the target. The watermelon sample is used as the test sample, and the distance between the test sample and other watermelon samples in the training set is calculated from the time-frequency features and the corresponding weight. the maturity of the category as a test sample;

S4) Use the training set to train the maturity detection model, then use the test set to test the maturity detection model, finally calculate the accuracy and record, repeat this step until the accuracy is the highest, and obtain the trained maturity detection model;

S5) Obtain the tap sound signal and weight of the watermelon test sample, and input the maturity detection model trained by the above steps to obtain the maturity of the watermelon test sample.

As shown in Figure 1, optimizing the STFT time-frequency feature extraction algorithm specifically includes the following steps:

Step 1: Calculate the FFT spectrum of the data set. The data set contains the tap sound signals of all watermelon samples. The spectrum of the signal will directly reflect the frequency components of the signal. The frequency characteristic distribution of the sound signal can be observed through the spectrum. The expression of the FFT spectrum The formula is as follows:

（1）

(1)

In the above formula, x ( t ) is the percussion sound signal of the watermelon sample, m is the serial number of the watermelon sample, and f is the frequency;

Step 2: Calculate the variance of all watermelon samples at each frequency point f , indicating the difference between the samples at the frequency point. The frequency resolution is divided according to the variance value because the variance can show the watermelon tapping sound signal at each frequency point. The difference of the frequency points, the greater the variance of the frequency points, the greater the difference, and the greater the degree of discrimination of watermelon maturity. The expression of the variance is as follows:

（2）

(2)

In the above formula, M is the total number of samples, m is the serial number of the watermelon sample, X _m ( f ) is the actual value of the frequency point, and X' _m ( f ) is the average value of the frequency point;

，STFT窗函数中，时间分辨参数τ决 定了窗口长度；频谱分辨率参数f取值范围为f=[f ₁ ,f ₂ …f _N ]，确定取值范围f=[f ₁ ,f ₂ …f _N ]表 达式如下 Step 3: Build the STFT window function

, in the STFT window function, the time resolution parameter τ determines the window length; the spectral resolution parameter f has a value range of f = [ f ₁ , f ₂ ... f _N ], and determines the value range f = [ f ₁ , f ₂ ... f _N ] expression is as follows

（3）

(3)

的频谱分辨率参数f是自适应数据集频谱特征分布情况改变的；In the above formula, DX ( f ) is the variance of all watermelon samples at each frequency f , N is the number of STFT spectral bandwidths, and the spectral interval between [ f _i , f _{i +1} ] represents the bandwidth of the spectral resolution , can be obtained from steps 1 and 2, the window function

The spectral resolution parameter f is changed by the spectral feature distribution of the adaptive dataset;

，优化STFT变换的推导公式如下 Step 4: According to the STFT window function

, the derivation formula for optimizing the STFT transformation is as follows

（4）

(4)

为上一步骤中构建的窗函数，x(t)为西瓜样本的敲击声音信 号。 In the above formula, S ( f _i ,τ ) is the time-frequency two-dimensional matrix after the time-frequency feature extraction of the tap sound signal of the watermelon sample,

is the window function constructed in the previous step, x ( t ) is the tap sound signal of the watermelon sample.

The optimized STFT time-frequency feature extraction algorithm solves the problem that the traditional STFT can only change the time resolution but cannot change the spectral resolution, and the spectral resolution is adaptive to the sample spectral feature distribution. Therefore, in step S2) of this embodiment, according to the optimized STFT time-frequency feature extraction algorithm, the specific steps of extracting the time-frequency feature of the watermelon sample tapping sound signal include:

According to the tapping sound signals of all watermelon samples, the FFT spectrum is calculated by formula (1) and formula (2) and the spectrum features are extracted;

； According to the extracted spectral features, the value range of the spectral resolution parameter f of the STFT window function is determined by formula (3), and the STFT window function is constructed according to the values of the time resolution parameter τ and the spectral resolution parameter f

;

和每个西瓜样本的敲击声音信号，由 式（4）计算每个西瓜样本的时频矩阵。 According to the STFT window function

and the percussion sound signal of each watermelon sample, the time-frequency matrix of each watermelon sample is calculated by formula (4).

As shown in Figure 4, Figure 4(a) is the FFT spectrum of all watermelon samples calculated by formula (1), the horizontal axis is the frequency, and the vertical axis is the watermelon sample serial number. Figure 4(b) shows the variance value calculated by formula (2). Combining with Figure 4(a), it can be seen that the parts with large characteristic differences in the spectrum are mainly concentrated in low frequencies, and the differences between each frequency point are not the same. Therefore, the optimized STFT time-frequency feature extraction algorithm of this embodiment divides the frequency resolution according to the variance value. Fig. 4(c) is the time-frequency matrix of the tap sound signal of the watermelon sample transformed according to the conventional STFT, and Fig. 4(d) is the tap sound signal of the watermelon sample based on the optimized STFT time-frequency feature extraction algorithm of the present embodiment The time-frequency matrix after STFT transformation, the horizontal axis is time, the vertical axis is frequency, it can be seen that the time-frequency matrix feature distribution extracted by the optimized STFT time-frequency feature extraction algorithm is more obvious, and the data dimension is reduced from [10,256] to [10, 32], where 10 is determined by the time parameter, which is the total signal time length/time resolution τ , and 32 is the number N of STFT spectral bandwidths in formula (3).

When KNN calculates the distance between the time-frequency matrix and the weight of the watermelon sample, the distance between a watermelon sample w _d ( S _d , y _d ) and its kth nearest neighbor sample w _k ( S _k , y _k ) The Euclidean distance is calculated using the following formula:

（5）

(5)

In the above formula, S _d ( f _i , τ ) and S _k ( f _i , τ ) represent the watermelon samples w _d ( S _d , y _d ) and w _k ( S _k , y _k ), respectively, using the optimized STFT time-frequency characteristics After extracting the time-frequency matrix obtained by the algorithm, y _d and y _k are the weights of the watermelon samples w _d ( S _d , y _d ) and w _k ( S _k , y _k ), respectively.

The traditional KNN Euclidean distance calculation process does not consider the gap between feature dimensions, but the time-frequency feature matrix S ( f _i ,τ ) is a two-dimensional matrix, while the weight data is a data point, and the order of magnitude is not uniform. In order to solve these problems, this embodiment proposes a KNN classification algorithm based on information fusion, as shown in Figure 1, including the following steps:

Step 1: Calculate the time-frequency matrix distance and weight feature distance between the watermelon sample w _d ( S _d , y _d ) and other watermelon samples in the training set, take the watermelon sample w _d ( S _d , y _d ) and its kth most Taking the adjacent watermelon sample w _k ( S _k , y _k ) as an example, calculate the time-frequency matrix distance m _k and the weight characteristic distance n _k of the watermelon sample w _d ( S _d , y _d ) and w _k ( S _k , y _k ) The expression is as follows:

（6）

(6)

（7）

(7)

In formula (6) and formula (7), S _d ( f _i , τ ) and S _k ( f _i , τ ) represent the watermelon samples w _d ( S _d , y _d ) and w _k ( S _k , y _k ), respectively ) using the time-frequency matrix obtained after optimizing the STFT time-frequency feature extraction algorithm, y _d and y _k are the weights of the watermelon samples w _d ( S _d , y _d ) and w _k ( S _k , y _k ) respectively;

Step 2: Normalize the watermelon sample w _d ( S _d , y _d ) with the time-frequency matrix distance and weight feature distance of other watermelon samples in the training set to solve the problem of non-uniform feature magnitudes. The watermelon sample w _d ( S _d , y _d ) and its k -th nearest watermelon sample w _k ( S _k , y _k ) as an example, the time-frequency of watermelon samples w _d ( S _d , y _d ) and w _k ( S _k , y _k ) The normalized expressions of matrix distance m _k and weight feature distance n _k are as follows:

（8）

(8)

（9）

(9)

In formula (8) and formula (9), m _max , m _min respectively represent the maximum and minimum values of the time-frequency matrix distance between the watermelon sample w _d ( S _d , y _d ) and other watermelon samples in the training set, n _max , n _min respectively represent the maximum and minimum distances of the weight feature distance between the watermelon sample w _d ( S _d , y _d ) and other watermelon samples in the training set;

Step 3: Introduce the information fusion weight parameter λ , calculate the Euclidean distance between the watermelon sample w _d ( S _d , y _d ) and other watermelon samples in the training set according to the normalized time-frequency matrix distance and weight feature distance, and use the watermelon sample w _d ( S _d , y _d ) and its k -th nearest watermelon sample w _k ( S _k , y _k ) as an example, the watermelon sample w _d ( S _d , y _d ) and w _k ( S _k , y _k ) ), the Euclidean distance expression is as follows:

（10）

(10)

In the above formula, λ is the information fusion weight parameter, m _k is the time-frequency matrix distance between the watermelon samples w _d ( S _d , y _d ) and w _k ( S _k , y _k ), and n _k is the watermelon sample w _d The weight characteristic distance between ( S _d , y _d ) and w _k ( S _k , y _k ).

After calculating the Euclidean distance between the watermelon sample w _d ( S _d , y _d ) and other watermelon samples in the training set through step 3, sort them in ascending order to find the watermelon sample w _d ( S _d , y _d ) The nearest K watermelon samples are counted, and the category with the highest frequency among the categories of these K watermelon samples is counted, which is the category of the watermelon sample w _d ( S _d , y _d ), that is, the watermelon sample w _d ( S _d , y d ) is obtained . y _d ) maturity.

The KNN classification algorithm based on information fusion further improves the accuracy of the detection system by fusing audio features and quality features, and solves the problem of non-uniform dimensions and large order of magnitude differences between various information features. Therefore, in step S2) of this embodiment, according to the information fusion KNN, the specific steps of calculating the distance between the target watermelon sample and other watermelon samples in the training set from the time-frequency feature and the corresponding weight include:

According to the time-frequency matrix of the target watermelon sample and other watermelon samples in the training set, the time-frequency matrix distance between the target watermelon sample and other watermelon samples in the training set is calculated by formula (6). According to the weight of the target watermelon sample and other watermelon samples in the training set , the weight feature distance between the target watermelon sample and other watermelon samples in the training set is calculated by formula (7);

The time-frequency matrix distance and weight feature distance between the target watermelon sample and other watermelon samples in the training set are normalized by equations (8) and (9);

According to the information fusion weight parameter λ , as well as the normalized time-frequency matrix distance and weight feature distance, the distance between the target watermelon sample and other watermelon samples in the training set is calculated by formula (10).

Step S4) of this embodiment is to test the accuracy and applicability of the model after the model is established to optimize the performance of the model. The maturity detection model after the test set is tested and trained, that is, according to the saved model-related parameters, the maturity detection model performs the feature extraction and classification described in step S2) on the watermelon samples in the test set to obtain its maturity. , the specific process is shown in the dashed box in Figure 3, according to the optimized STFT time-frequency feature extraction algorithm above to calculate the time-frequency feature of each watermelon sample and record the weight feature; Integrate the KNN classification algorithm to calculate the Euclidean distance between the test sample and other watermelon samples in the training set; select the top K watermelon samples with the closest distance in ascending order; finally count the category with the highest frequency among the K watermelon samples, which is the test sample maturity.

Since the watermelon samples in the test set are also classified in step S1), for the watermelon samples in the test set, the number of watermelon samples whose maturity detected by the statistical maturity detection model is consistent with the actual maturity is divided by the watermelon samples in the test set. The total number of samples, the accuracy of the maturity detection model can be obtained.

While training and testing the maturity detection model, it is also necessary to consider the problem of parameter optimization. The parameters that need to be optimized include the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ , so when using the training set to train mature Before the degree detection model, it also includes the steps of optimizing the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ . The grid search algorithm, particle swarm search algorithm, bird flock search algorithm, etc. can be used. For example, the steps of optimizing the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ include:

If the values of the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ do not reach the preset maximum value, increase the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter with a preset step size The value of λ , executes the step of training maturity detection model with described training set;

If the value of the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ reaches the preset maximum value, find the highest accuracy from all the recorded accuracies, and use the time resolution parameter τ , the spectral resolution parameter The values of f and the information fusion weight parameter λ are adjusted to the value corresponding to the highest accuracy, and the step of training the maturity detection model with the training set is performed.

In this embodiment, the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ are all preset with corresponding intervals and step sizes. In step S2), the time resolution parameter τ , the spectral resolution parameter f and The initial value of the information fusion weight parameter λ is the minimum value corresponding to the preset interval. In step S4), the values of the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ are increased by the corresponding step size each time, until The maximum value of the corresponding preset interval is reached. At this time, the value of the time resolution parameter τ , the spectral resolution parameter f and the information fusion weight parameter λ corresponding to the highest accuracy is the optimal value, so the time resolution parameter τ is the optimal value. The maturity detection model trained under the optimal values of , spectral resolution parameter f and information fusion weight parameter λ has the best performance.

Finally, the method for detecting the maturity of watermelon based on acoustic analysis and machine learning in this embodiment considers the influence of the filtering algorithm and the preprocessing steps on noise resistance, and after acquiring the tap sound signal of the watermelon sample, it also includes the step of extracting effective signals. include:

Calculate the energy of each frame of the percussion sound signal, retain the frames whose energy is greater than the preset threshold, and obtain a filtered percussion sound signal;

The percussion sound signal filtered once is filtered to obtain the percussion sound signal filtered for the second time, which is used as an effective signal of the percussion sound signal.

The reason for obtaining a filtered percussion sound signal is that the recording time of the watermelon percussion sound signal is generally relatively long, and the endpoint detection method needs to be used to extract the valid signal from the entire sound signal. Methods such as short-term energy, short-term zero-crossing rate, and frequency domain characteristics can be used. In this embodiment, the method for detecting the endpoint of a sound signal based on short-term energy is adopted. It can calculate the energy of each frame of the sound signal to determine the effective The start and end time points of the signal. The short-term energy calculation formula of the nth frame of the sound signal is as follows:

（11）

(11)

In the above formula, L represents the length of the frame, and x ( m ) represents the tap sound signal of the mth frame of the watermelon sample.

Figure 5(a) is the tap sound signal of a certain watermelon sample. When En _is greater than the preset threshold, the tap sound signal of the corresponding frame is retained. When En _is less than the preset threshold, the tap of the corresponding frame is discarded. The tap sound signal is cut to cut off the sound signal. The final filtered tap sound signal is shown in Figure 5(b), which is beneficial to improve the accuracy of the time-frequency features extracted in the subsequent steps.

For obtaining the tap sound signal after secondary filtering, it is because the tap sound signal of the watermelon sample may be interfered with by environmental noise, which reduces the detection accuracy. Therefore, a filtering algorithm is required to filter the tap sound signal of the watermelon sample. A low-pass filter, a Butterworth filter, an adaptive filter, etc. can be used, and the least mean square adaptive filter is used in this embodiment, which can automatically correct the weight coefficients by means of an effective adaptive algorithm, so as to adapt to the external Changes in the environment keep the filter in the best state, with good adaptability and filtering performance, which can effectively filter out environmental noise and improve the signal-to-noise ratio of effective signals.

Embodiment 2

This embodiment proposes a watermelon maturity detection system based on acoustic analysis and machine learning according to the first embodiment, including:

a data acquisition unit for acquiring the tap sound signal and weight of the watermelon samples, and forming a data set, classifying the watermelon samples in the data set according to maturity, and dividing the data set into a training set and a test set; and It is used to obtain the tap sound signal and weight of the watermelon test sample, and input the trained maturity detection model to obtain the maturity of the watermelon test sample;

The model building unit is used to build a maturity detection model. The maturity detection model uses the optimized STFT time-frequency feature extraction algorithm to extract the time-frequency characteristics of the tap sound signal of the watermelon sample, and also uses the information fusion KNN classification algorithm to extract the target watermelon sample. As a test sample, the distance between the test sample and other watermelon samples in the training set is calculated from the time-frequency features and the corresponding weight. Finally, among the categories corresponding to the K watermelon samples that are closest to the test sample, the category with the highest frequency is used as the maturity of the test sample;

The model training and testing unit is used to train the maturity detection model with the training set, then use the test set to test the maturity detection model, finally calculate the accuracy and record, and continue to use the training set to train the maturity detection model, Until the accuracy is the highest, the trained maturity detection model is obtained.

In this embodiment, when the maturity detection model extracts the time-frequency features of the watermelon sample tapping sound signal, the model building unit is configured to perform the following steps:

According to the tapping sound signals of all watermelon samples, the FFT spectrum is calculated by formula (1) and formula (2) and the spectrum features are extracted;

根据 所述STFT窗函数

和每个西瓜样本的敲击声音信号，由式（4）计算每个 西瓜样本的时频矩阵。 According to the extracted spectral features, the value range of the spectral resolution parameter f of the STFT window function is determined by formula (3), and the STFT window function is constructed according to the values of the time resolution parameter τ and the spectral resolution parameter f

According to the STFT window function

and the percussion sound signal of each watermelon sample, the time-frequency matrix of each watermelon sample is calculated by formula (4).

In this embodiment, when the maturity detection model calculates the distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency feature and weight, the model building unit is configured to perform the following steps:

According to the time-frequency matrix of the target watermelon sample and other watermelon samples in the training set, the time-frequency matrix distance between the target watermelon sample and other watermelon samples in the training set is calculated by formula (6). According to the weight of the target watermelon sample and other watermelon samples in the training set , the weight feature distance between the target watermelon sample and other watermelon samples in the training set is calculated by formula (7);

The time-frequency matrix distance and weight feature distance between the target watermelon sample and other watermelon samples in the training set are normalized by equations (8) and (9);

According to the information fusion weight parameter λ , as well as the normalized time-frequency matrix distance and weight feature distance, the distance between the target watermelon sample and other watermelon samples in the training set is calculated by formula (10).

The above are only preferred embodiments of the present invention, and do not limit the present invention in any form. The core idea of the present invention is to use the combined model of the feature extraction algorithm of time-frequency analysis + machine learning classification algorithm to solve the problem of watermelon maturity detection, using the sound signal generated by tapping the watermelon and the quality of the watermelon as the basis. Other schemes based on machine vision or infrared spectrum detection are quite different from the present invention in terms of information sources and technical schemes, but there are alternative technical schemes for the watermelon maturity detection scheme based on acoustic analysis.

Specifically, taking the time-frequency analysis method of a sound signal as an example, the present invention adopts an improved method based on STFT, and there are other basic time-frequency analysis methods and improved methods such as wavelet transform, S transform, and Mel cepstral analysis. Substitute the method of the present invention, but the essence of these methods is still a time-frequency analysis method, that is, a method that can extract the characteristics of the sound signal in the time domain and the frequency domain at the same time; The improved information fusion algorithm based on KNN can replace KNN with support vector machine, logistic regression, random forest, decision tree, neural network and other classification algorithms, but the essence is to combine the above time-frequency characteristics and quality characteristics, through the training of the model And the test completes the watermelon ripeness classification. The feature extraction algorithm of time-frequency analysis + machine learning classification algorithm can arbitrarily replace and combine the above methods. The combination scheme adopted in the present invention is a combination scheme with better performance after model training and testing on the actual data set, but other The combination scheme can also achieve the purpose of watermelon maturity detection.

Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solutions of the present invention should fall within the protection scope of the technical solutions of the present invention.

Claims (10)

1. A watermelon maturity detection method based on acoustic analysis and machine learning is characterized by comprising the following steps:

acquiring a knocking sound signal and weight of a watermelon sample, forming a data set, classifying the watermelon sample in the data set according to maturity, and dividing the data set into a training set and a testing set;

constructing a maturity detection model, wherein the maturity detection model is used for extracting time-frequency characteristics of a knocking sound signal of a watermelon sample, calculating the distance between the target watermelon sample and other watermelon samples in a training set according to the time-frequency characteristics and corresponding weight, and finally taking the class with the highest occurrence frequency as the maturity of the target watermelon sample in the classes corresponding to K watermelon samples closest to the target watermelon sample;

training a maturity detection model by using the training set, testing the maturity detection model by using the testing set, finally calculating and recording the accuracy, and repeating the steps until the accuracy is highest to obtain the trained maturity detection model;

and acquiring the knocking sound signal and the weight of the watermelon test sample, and inputting the knocking sound signal and the weight into the trained maturity detection model to obtain the maturity of the watermelon test sample.

2. The method for detecting the maturity of the watermelon according to claim 1, wherein the step of extracting the time-frequency characteristics of the knocking sound signal of the watermelon sample comprises the following steps:

calculating an FFT frequency spectrum and extracting frequency spectrum characteristics according to the knocking sound signals of all watermelon samples;

determining a spectral resolution parameter of the STFT window function from the extracted spectral featuresfAccording to the time resolution parameterτAnd spectral resolution parameterfValue of (a) construct an STFT window function

；

According to the STFT window function

And the knocking sound signal of each watermelon sample, and calculating the time-frequency matrix of each watermelon sample.

3. The method for detecting the maturity of watermelon according to claim 2, wherein the function expression of the time-frequency matrix is as follows:

in the above-mentioned formula, the compound has the following structure,x(t) Is the knock sound signal of the watermelon sample,

is a constructed STFT window function, whereinτIs time of dayThe resolution parameters are used to determine the resolution parameters,fis a spectral resolution parameter [ alpha ], [ alpha ]f _i , f _i+1]The spectral interval in between represents the bandwidth of the spectral resolution.

4. The method for detecting watermelon ripeness based on acoustic analysis and machine learning of claim 2, wherein the specific step of calculating the distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency characteristics and the weight comprises:

calculating the time-frequency matrix distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency matrix of the target watermelon sample and other watermelon samples in the training set, and calculating the weight characteristic distance between the target watermelon sample and other watermelon samples in the training set according to the weight of the target watermelon sample and other watermelon samples in the training set;

normalizing the time-frequency matrix distance and the weight characteristic distance between the target watermelon sample and other watermelon samples in the training set;

fusing weight parameters according to informationλAnd calculating the distance between the target watermelon sample and other watermelon samples in the training set according to the normalized time-frequency matrix distance and the normalized weight characteristic distance.

5. The method of claim 4, wherein the other watermelon samples in the training set comprise the first watermelon sample of the target watermelonkThe nearest watermelon sample, the target watermelon sample andkthe functional expression for the distance between the nearest watermelon samples is:

in the above formula, the first and second carbon atoms are,λin order to obtain the information fusion weight parameter,m _k is the target watermelon sample andkthe time-frequency matrix distance between the nearest watermelon samples,n _k is the target watermelon sample andkat the mostThe weight characteristic distance between adjacent watermelon samples.

6. The method of claim 4, wherein training the maturity detection model with a training set further comprises optimizing a temporal resolution parameterτSpectral resolution parameterfAnd information fusion weight parameterλThe method specifically comprises the following steps:

if the time resolution parameterτSpectral resolution parameterfAnd information fusion weight parameterλThe value of (a) does not reach the preset maximum value, and the time resolution parameter is increased by the preset step lengthτSpectral resolution parameterfAnd information fusion weight parameterλPerforming a step of training a maturity detection model with the training set;

if the time resolution parameterτSpectral resolution parameterfAnd information fusion weight parameterλThe value of (d) reaches a preset maximum value, the highest accuracy is found from all recorded accuracies, and the time resolution parameter is usedτSpectral resolution parameterfAnd information fusion weight parameterλIs adjusted to the value corresponding to the highest accuracy, the step of training the maturity detection model with the training set is performed.

7. The method for detecting the maturity of watermelon based on acoustic analysis and machine learning of claim 1, wherein after acquiring the tapping sound signal of the watermelon sample, the method further comprises a step of extracting effective signals, and the method specifically comprises the following steps:

calculating the energy of each frame of the knocking sound signal, and reserving the frames with the energy larger than a preset threshold value to obtain the knocking sound signal after primary filtering;

and filtering the primarily filtered tapping sound signal to obtain a secondarily filtered tapping sound signal which is used as an effective signal of the tapping sound signal.

8. A watermelon maturity detection system based on acoustic analysis and machine learning, comprising:

the system comprises a data acquisition unit, a training unit and a testing unit, wherein the data acquisition unit is used for acquiring knocking sound signals and weight of watermelon samples, forming a data set, classifying the watermelon samples in the data set according to maturity and dividing the data set into a training set and a testing set; the device is also used for acquiring a knocking sound signal and weight of the watermelon test sample, and inputting the knocking sound signal and weight into the trained maturity detection model to obtain the maturity of the watermelon test sample;

the model construction unit is used for constructing a maturity detection model, the maturity detection model is used for extracting time-frequency characteristics of a watermelon sample knocking sound signal, then calculating the distance between a target watermelon sample and other watermelon samples in a training set according to the time-frequency characteristics and corresponding weight, and finally taking the class with the highest occurrence frequency as the maturity of the target watermelon sample in the classes corresponding to K watermelon samples closest to the target watermelon sample;

and the model training and testing unit is used for training the maturity detection model by using the training set, testing the maturity detection model by using the testing set, calculating and recording the accuracy, and continuing training the maturity detection model by using the training set until the accuracy is highest to obtain the trained maturity detection model.

9. The system of claim 8, wherein when the maturity detection model extracts time-frequency features of a watermelon sample tapping sound signal, the model construction unit is configured to perform the following steps:

calculating an FFT frequency spectrum and extracting frequency spectrum characteristics according to the knocking sound signals of all watermelon samples;

determining a spectral resolution parameter of the STFT window function from the extracted spectral featuresfAccording to the time resolution parameterτAnd spectral resolution parameterfConstructing an STFT window function;

and calculating a time-frequency matrix of each watermelon sample according to the STFT window function and the knocking sound signal of each watermelon sample.

10. The acoustic analysis and machine learning based watermelon maturity detection system of claim 9 wherein said maturity detection model, when calculating the distance between the target watermelon sample and other watermelon samples in the training set from time-frequency features and weight, said model building unit is configured to perform the steps of:

calculating the time-frequency matrix distance between the target watermelon sample and other watermelon samples in the training set according to the time-frequency matrix of the target watermelon sample and other watermelon samples in the training set, and calculating the weight characteristic distance between the target watermelon sample and other watermelon samples in the training set according to the weight of the target watermelon sample and other watermelon samples in the training set;

normalizing the time-frequency matrix distance and the weight characteristic distance between the target watermelon sample and other watermelon samples in the training set;

fusing weight parameters according to informationλAnd calculating the distance between the target watermelon sample and other watermelon samples in the training set according to the normalized time-frequency matrix distance and the normalized weight characteristic distance.

Images