Application research on vehicle type identification based on acoustic detection method based on neural network

    Abstract: This paper introduces the learning and training method based on multi-layer feedforward neural network and the identification system that uses noise measurement principles and methods to identify vehicle types, which provides a powerful basis and means for vehicle traffic management and monitoring statistics.

    Keywords: automotive neural network noise recognition

With the rapid development of highway and automobile technology in our country, the speed of automobiles is getting higher and higher, and a more effective traffic management system is urgently needed. Such a system should be able to automatically identify cars and accurately determine the type of car. Since the 1950s, point test equipment such as ring-forming circle detectors has been mainly used for traffic control and traffic data collection at intersections. This detection system usually uses an inductor coil buried under the road surface to identify cars through electromagnetic induction. However, this system has the disadvantages of expensive installation, difficult maintenance, and inability to classify cars. Using acoustic geodesics and training with neural networks, the types of cars passing on the road can be effectively identified.

1 Acoustic detection vehicle type identification system

The working principle block diagram of the acoustic detection vehicle type identification system is shown in Figure 1.

The working principle of the acoustic car type identification system is: when a car passes, the microphone transmits the sound pressure signal of the sound wave it generates through the connector to the classification system, and the sound pressure signal is converted into a series of discrete discrete numbers through the A/D converter. digital signal and perform spectrum conversion (FFT conversion) in a spectrum analyzer.

The neural network provides a classification indicator for each vector received by the attenuator. Each vector displays a predetermined pause, which is a brief time interval (0.1 seconds) allowed for the object producing the sound. This way, the neural network produces a classification indicator for each sound every 0.1 seconds, classifying each vector independently. The time totalizer is a processor that classifies indicator traffic after neural network analysis. That is, the time totalizer is combined with the operation results multiple times to produce the final classification result output by the entire system.

The acoustic detection car type identification system based on neural network uses low-cost sensor materials, which has a good effect on car type identification, and the application coverage is wider and is not affected by weather and light.

2 Neural network learning and training algorithm

After preprocessing the original noise, Yu's recognition work is implemented by the neural network. The network designed in this article is a three-layer feedforward network. Its structure is: the input layer has n nodes, n=Np×N; the hidden layer has P nodes, and the output layer has q nodes. The q output nodes respectively correspond to q types of car models. The input value used in the input layer is the value after Fourier transform, and the activation function of the hidden layer and output layer is the Sigmoid function. After training, when the output value of the i-th node in the output layer is greater than 0.99 and the output values ​​of other nodes are less than 0.11, the model identified this time is considered to be the i-th car. The network structure is shown in Figure 2.

The training algorithm used is BP algorithm, and the error function in this algorithm is:

Among them, dk and rk are the expected output amount and actual output vector of the network respectively, and M is the training sample pair.

A prominent shortcoming of the BP algorithm is the slow learning speed. The reasons are multifaceted, such as related to the structure of the network and the shortcomings of the learning algorithm itself. When the above error function is used to adjust the weight, it is known from derivation that the weight adjustment always includes the following factors:

As can be seen from the above formula, when the actual output ri of the output layer unit i is close to 0 or 1, the factor ri(1-ri) in the error signal makes the error signal become very small. At this time, if the output layer unit i When the actual output ri is very different from the expected output value di, no strong error is generated to correct the weight, thus prolonging the learning process. In addition, since the excitation function f(x)=(1+e -x)-1 is a saturated function, when it approaches the saturated state, the derivative is close to zero, thus slowing down the convergence speed. From this, we can consider removing the factor ri (1-ri) from the result of the partial differential of the error function with respect to the weight, so the error function of the BP algorithm can be improved to:

Because 0

Using the chain differential rule we get:

The first term on the right represents the influence of the i-th component in the output vector on the deviation, and the second term represents the influence of the weight coefficient on the output component. Calculate the partial differential of the first term according to equation (4):

Calculate the second partial differential of equation (5):

Among them, bj>0 is the implicit output above; equation (9) is the negative gradient direction on the sk influence surface in wij space.

In this way, any element wij in the weight matrix [w] has:

This shows the relationship between the correction of weights and the gradient descent method, which shows that the adjustment of weights by this error correction method is convergent. Utilizing the improvement of the error function to equation (3), the BP algorithm determines to eliminate the ri(1-ri) factor from the result of the partial differential of the error function to the weight. From the actual training point of view, when using this algorithm to modify the weights, it can converge quickly.

The BP algorithm and the improved BLP algorithm in the paper were separately trained on the network (4 types of vehicle models were selected, that is, q=4), and tested on actual traffic lines. The samples of each vehicle model are 4 sets of samples in different situations in actual applications (small cars, medium cars, heavy cars, super heavy cars). In this way, 16 sets of noise are used as training samples, and the node of the hidden layer is set to p =16. The flow chart of the online learning process is shown in Figure 3.

When the network structure is exactly the same, using the error function of formula (1) for training requires cross-training 20,000 times before it can be used in practice, and the output accuracy can reach 0.99; while using the improved error function ( 3) It only takes 200 times to train using the formula to achieve the same accuracy, and can be used to identify car types online on traffic lines. The test results are shown in Table 1.

    Judging from the test results, the improved BP algorithm can speed up the learning speed of the network and can be easily applied to actual production lines. Even if a new car model appears on the traffic line, the network training can be completed quickly. The test results show that the recognition error rate is less than 1%, and the recognition effect of cars is better than the calculated method. Using the improved BP algorithm improves learning and provides an effective method for the practical application of neural network recognition, especially on traffic lines with large traffic volumes. From the perspective of practical applications, neural network recognition is more fault-tolerant than statistical recognition using special patches, and is also better than feature statistical methods. In practice, if there are new car models to be added, just slightly improve the network structure.