CN113358600A - Gas detection chamber, laser spectrum gas detection system based on artificial neural network and laser spectrum gas detection method based on artificial neural network - Google Patents

Abstract

The invention belongs to the technical field of gas detection, and relates to a gas detection gas chamber, a system and a method. A gas detection gas chamber is characterized in that: two ends of the gas chamber are respectively installed with mirrors with opposite mirror surfaces, and the mirrors are spherical mirrors with a curvature radius of 500 mm. The spherical mirror with a curvature radius of 500mm is installed at both ends of the gas chamber of the present invention, and the spherical mirror is composed of many small spherical mirrors, so that the incident beam can be reflected multiple times in the gas chamber, and the effective optical path is increased. Absorbance of incident light by a gas. In the method of the present invention, after determining the trace gas composition using the neural network model, the concentration of the specific gas is calculated, and finally the trace gas composition and concentration are obtained. It has the advantages of strong nonlinear mapping ability and fast training speed.

Description

Gas detection gas chamber, laser spectrum gas detection system and method based on artificial neural network

technical field

The invention belongs to the technical field of gas detection, and relates to a gas detection gas chamber, a system and a method.

Background technique

The rapid, high-precision identification and quantitative detection of trace constituents in gas mixtures is in great demand in many fields. Exhaust gas discharged from industrial production will pollute the air environment, cause a greenhouse effect and endanger the human respiratory system; toxic or flammable and explosive gases generated in the production process, if randomly discharged or leaked, will cause great social security risks. In addition, in the medical diagnosis of chronic malignant diseases, the conventional diagnostic methods have a long cycle and cause great harm, but by detecting the patient's respiratory gas markers, the purpose of diagnosis can be achieved quickly, effectively and non-invasively. Therefore, the detection of trace gas composition and concentration is of great significance to the fields of ensuring air quality, production safety and medical diagnosis.

In the gas detection technology, although the traditional gas-sensing detection method, meteorological detection method and chemiluminescence method have their own advantages, the measurement form is single-point, limited by the level of optoelectronic technology, and the detection sensitivity and accuracy of these methods are not high. The traditional optical gas detection uses laser spectroscopy technology for gas measurement, which has high accuracy and sensitivity, and strong real-time performance. It has a good identification and quantitative analysis effect for single-component gas or multi-component gas with no obvious spectral aliasing characteristics . However, this technique is still difficult to implement and has low precision for multi-component gases whose characteristic spectral lines are aliased due to the low concentration of trace harmful gases, low absorption and similar chemical bonds of gas components. Therefore, although traditional optical gas detection is considered as a new gas detection technology, its limitations are still very obvious.

Optical frequency comb (Optical Frequency Comb) technology came into being under the above background. This technology has become a high-precision and high-resolution gas detection technology by virtue of its advantages of wide spectrum, narrow pulse width, high frequency stability, and traceability of time-domain and frequency-domain characteristics to the microwave frequency reference. At home and abroad, optical frequency combs have been realized by various methods, such as optical frequency comb cavity ring-down spectroscopy, optical frequency comb cavity enhancement spectroscopy and dual optical frequency comb multi-heterodyne spectroscopy. The foreign research group applied the cavity ring-down optical frequency comb to the gas absorption measurement, which increased the sensitivity of a single spectral line to 7×10 ^-13 cm ^-1 , and realized the absorption spectrum detection of industrial waste gas and atmospheric greenhouse gases. Purification grades provide the solution. Utilizing dual-optical frequency comb multi-heterodyne spectroscopy technology improves measurement accuracy by a factor of 10 and measures trace components in patient exhaled breath. At present, developed countries in Europe and the United States are at the international leading level in the research of optical frequency comb technology. The overall domestic research level is relatively backward, focusing on tracking and expanding foreign related research, especially in the application of optical frequency comb gas detection.

So far, although optical frequency comb technology can effectively improve the accuracy of trace gas detection, it cannot effectively decouple spectral line aliasing by using this technology alone, resulting in failure to achieve high-precision identification of various components in multi-component gases. , which has become an urgent scientific problem to be solved in optical frequency comb technology. In this regard, the advantages of Artificial Neural Network technology are expected to effectively solve the above problems. The artificial neural network can provide a powerful solution algorithm, which can automatically generate a black box function through repeated training and learning of the mapping relationship between a large number of input and output, so as to realize the gas identification without establishing the equation expression of the gas response. There have been some reports on the application of artificial neural network to optical gas detection in foreign countries: in power production, the technology is used to detect the content and composition of gas in oil in transformer oil to eliminate major safety hazards that may be caused by gas leakage. ; The United States and Japan have successfully realized the identification and content detection of up to eight gases (H2, CO, CH4, C2H4, C2H2, C2H6, CO2, O2). Using intelligent olfactory technology based on artificial neural network and support vector machine to solve the aliasing of absorption spectrum of multi-component gas molecules. In biomedical related research, the application of artificial neural network optical gas detection system has been able to successfully distinguish the exhaled gas markers of cancer patients and healthy people, providing a new research direction for cancer diagnosis research. From the international frontier analysis, the deep integration of artificial neural network and optical frequency comb technology should be an important development direction in the future.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a system and method integrating optical frequency comb absorption spectrum gas identification and concentration measurement based on the artificial neural network based on the deficiencies of the above-mentioned prior art. The invention has the advantages of strong non-linear mapping capability, fast training speed, effective identification of trace gas components, and high-precision detection of gas concentration.

In order to solve the technical problem of the present invention, the present invention first provides a gas detection gas chamber, two ends of the gas chamber are respectively installed with mirrors facing opposite mirrors, and the mirrors are spherical mirrors with a curvature radius of 500 mm.

Further, the spherical mirror is a spherical mirror array composed of several small spherical mirrors.

Further, the small spherical mirror is fixed on the angle adjustment bracket.

Further, the angle adjustment bracket includes a bottom plate and a mirror holder distributed in an array on the bottom plate, the mirror holder is hinged with the bottom plate, and the small spherical mirror is inlaid on the mirror holder.

Further, an incident/exit hole with a diameter of 5 mm is provided at the input end of the air chamber, and the incident/exit hole is located in the center of the reflector.

The present invention solves its technical problem and also provides a laser spectrum gas detection system based on artificial neural network, including the gas chamber, laser light source, data acquisition card, photodetector and gas detection unit; the gas chamber and the optical frequency comb connection; the photodetector is connected with the gas chamber, and the outgoing light of the gas chamber is converted into electrical signals and transmitted to the data acquisition card; the data acquisition card is connected with the gas detection unit, and the collected data is transmitted to the gas detection unit ; The gas detection unit preprocesses the input data to complete gas type identification and concentration calculation.

As a preferred mode of the present invention, the gas detection unit includes a gas identification module and a concentration calculation module, wherein the gas identification module is an artificial neural network model.

In order to solve the technical problem of the present invention, the present invention also provides a laser spectrum gas detection method based on an artificial neural network, including: (1), constructing an artificial neural network model for identifying various types of gases;

(2) Introduce the gas to be trained into the gas chamber in turn, the data acquisition card will transmit the collected data of the gas to be trained to the gas detection unit, and process it into an input feature vector, input the artificial neural network model, and complete Identification of the type of gas to be measured;

(3), according to the identified composition of the gas to be measured, calculate the concentration x of the gas to be measured:

in,

为吸收线型，采用voigt线型；S(T)为气体分子吸收线强，

为1f归一化的2f信号强度，P为气室的压强，L为气体在气室内的光程长度。

is the absorption line type, using the voigt line type; S(T) is the gas molecular absorption line intensity,

is the 2f signal intensity normalized to 1f, P is the pressure of the gas chamber, and L is the optical path length of the gas in the gas chamber.

Further, in the construction of the artificial neural network, the acquisition method of the training sample is:

Introduce m kinds of trace gases into the gas chamber in turn to collect data; set a category label Y _j for each gas, j=1, 2, ..., m, representing the jth gas among the m kinds of gases;

i∈(0,n)，j∈(1,m)。For each gas, the data acquisition card transmits the n sets of data of the gas to the gas detection unit, and the gas detection unit normalizes each set of data: frequency X ₁ , second harmonic X ₂ , and adds the category label Y _j ; then get the feature vector of the labeled training sample:

i∈(0,n), j∈(1,m).

Further, in the construction of the described artificial neural network, the training of the network model includes the following steps:

(1) Input the feature vector of the training sample to the input layer of the neural network; initialize the network weight matrix θ;

(2) implement forward propagation, and calculate the output activation value a ^l ;

(3) Calculate the cost function

则代价函数对于每层权重的偏导数为

(4) Backpropagation is performed to calculate the output layer error δ ^l =(a ^l -Y).*a ^l′ , the hidden layer error

Then the partial derivative of the cost function with respect to the weights of each layer is

(5) Carry out gradient detection, check whether the gradient of numerical calculation is consistent with the partial derivative calculated by backpropagation in step (4), if the agreement indicates that the backpropagation is running normally, proceed to the next step;

(6) Perform stochastic gradient descent method to optimize the weight matrix, repeat steps (2), (3), (4) until the cost function corresponding to the global optimal weight satisfies the accuracy error, and the neural network model training is completed.

Further, in the training of the network model, the activation function adopts the sigmoid function, namely:

Compared with the prior art, the present invention has the beneficial effects of providing a gas detection gas chamber and a laser spectrum gas detection system based on an artificial neural network, and spherical mirrors with a curvature radius of 500 mm are installed at both ends of the gas chamber, and the spherical mirror is composed of many The small spherical mirrors form an array, so that the incident beam can be reflected multiple times in the gas chamber, increasing the effective optical path. Due to the low concentration of trace gas, the absorption is weak, and the absorption of the incident light by the gas is effectively increased. In the method of the present invention, after determining the trace gas composition by using the neural network model, the concentration of the specific gas is calculated, and finally the trace gas composition and concentration are obtained. It has the advantages of strong nonlinear mapping ability and fast training speed.

Description of drawings

Fig. 1 is the structural connection diagram of the laser spectrum gas detection system based on artificial neural network provided by the present invention;

Figure 2 is a side view of a gas chamber;

Fig. 3 is the schematic diagram of the reflector in Fig. 2;

Figure 4 is a side view of the angle adjustment bracket;

Fig. 5 is the flow chart of the laser spectrum gas detection method based on artificial neural network provided by the present invention;

FIG. 6 is a schematic diagram of the structure of an artificial neural network.

Detailed ways

In order to facilitate understanding of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the invention are shown in the accompanying drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described in this specification. Rather, these embodiments are provided so that a thorough and complete understanding of the present disclosure is provided.

In order to solve the problems existing in gas detection in the prior art, this embodiment provides a laser spectrum gas detection system based on artificial neural network. As shown in FIG. 1 , the system mainly includes: a laser controller 1, an optical frequency comb 2, Gas chamber 3, photodetector 4, modem 5, data acquisition card 6, gas detection unit 7 embedded with artificial neural network algorithm and gas-related parameters for concentration calculation. The laser controller 1 is connected to the optical frequency comb 2, generates a scanning sawtooth wave signal, and controls the laser pulse and temperature of the optical frequency comb through current.

The absorption characteristic lines of various gases are different. Since the optical frequency comb has the advantages of wide spectrum and high precision, in this embodiment, the optical frequency comb of an erbium-doped fiber system is selected as the laser light source.

As shown in FIGS. 2 and 3 , two opposite ends of the air chamber 3 are respectively provided with facing mirrors 31 . The mirrors 31 at each end consist of an array of spherical mirrors measuring 50 x 50 mm ² . The spherical mirror array consists of 64 square spherical mirrors 34 with a size of 6.25×6.25 mm ² . The square spherical mirror 34 is embedded on the mirror holder 33. The mirror holder 33 is connected to the base plate 36 through the universal hinge 35. The mirror holder 33, the universal hinge 35 and the base plate 36 constitute an angle adjustment bracket, as shown in FIG. 4 .

64 square spherical mirrors 34 are fixed on the angle adjustment bracket, and the reflection angle of each square spherical mirror can be changed by manual adjustment. Each square spherical mirror can be independently adjusted horizontally and vertically for beam alignment.

At one end of the air chamber 3, there is an injection/ejection port 32 with a diameter of 5 mm. The injection/ejection port 32 is located in the middle of the mirror 31 at one end. This port is the input port and the output port of the laser.

The laser generated by the optical frequency comb 2 enters the gas chamber 3 through the injection/ejection port 32 through the optical fiber. By adjusting the angles of the square spherical mirrors at both ends of the gas chamber, the number of laser reflections in the gas chamber can be changed. After multiple reflections , and then output through the injection/injection port 32. In general, due to the low concentration of trace gas, the absorption in the common gas chamber is weak, which is not conducive to subsequent concentration detection. With the gas chamber provided in this embodiment, the laser undergoes multiple reflections in the gas chamber, which can effectively increase the absorption of the gas to the laser, which is more conducive to the subsequent detection and calculation of the concentration. The mirror coating has high reflection (>99.98%) in the near infrared range [1550nm, 1825nm].

The incident/ejection port 32 is connected to the photodetector 4 through an optical fiber. The photodetector 4 converts the reflected light output from the air chamber 3 into an electrical signal, and transmits it to the modem 5 after demodulation and low-pass filtering, and then transmits it to the data acquisition card 6, and the data acquisition card transmits the collected signal to the Gas detection unit 7. The modem 5 is also connected with the optical frequency comb 2 to generate a modulation signal to modulate the high frequency band of the laser output signal and input it into the air chamber 3 .

The gas detection unit 7 includes a gas identification module and a concentration calculation module. Among them, the gas identification module is an artificial neural network model. The gas detection unit 7 preprocesses the input data, generates a feature vector, uses an artificial neural network to perform multiple iterative training on the input feature vector, and learns a gas recognition model through training. Concentration calculation, and finally obtain the trace gas composition and concentration.

The present invention also provides another embodiment. This embodiment is a laser spectrum gas detection method based _on artificial neural network. The method flow is shown in 5. ), methane (CH ₄ ), carbon dioxide (CO ₂ ) four trace gases detection as an example, the specific steps are:

1. Set up the artificial neural network structure

According to the 4 gases to be trained: nitrous oxide (N ₂ O), nitric oxide (NO), methane (CH ₄ ), carbon dioxide (CO ₂ ), set the number of neurons in the output layer to 4, due to the number of input features is 2, so set the number of neurons in the input layer to 2, and increase the bias unit of the input layer, set it to 1. Set the number of hidden layers to 1, set the number of hidden layer neurons to 4, and increase the hidden layer bias unit, set to 1. Complete the settings of the neural network structure, as shown in Figure 6. The activation function of the neural network adopts a sigmoid function, and the temperature of the light source is controlled by the laser controller, the temperature of the gas chamber is controlled, and the gas flow rate is controlled to ensure the pressure of the gas chamber.

2. Turn on the laser controller and the optical frequency comb. The current generated by the laser controller is used to generate the scanning sawtooth wave and control the temperature of the optical frequency comb. The modem generates a modulation signal and adds it to the laser output signal, and modulates the output signal in the high frequency band. The optical fiber is introduced into the air chamber.

For the 4 gases to be trained: nitrous oxide (N ₂ O), nitric oxide (NO), methane (CH ₄ ), carbon dioxide (CO ₂ ), each gas to be trained is introduced into the air chamber in turn, and mixed Nitrogen, control the temperature of the gas chamber, and control the pressure of the gas chamber by controlling the gas flow rate. The laser is reflected in the gas chamber for many times and then emitted from the same port. It is transmitted to the photoelectric detector through the optical fiber, and the signal is photoelectrically converted and then transmitted to the modem. After the modem low-pass filters the electrical signal, the data is transmitted to the data acquisition card.

其中，x、

x_max分别为特征的输入值、平均值和最大值。3. For each gas to be trained, the data acquisition card transmits the n groups of data of the gas to the gas detection unit, and the gas detection unit transmits the data: frequency X ₁ , second harmonic intensity X ₂ , using the normalization formula, to carry out Feature scaling:

Among them, x,

x _max is the input value, mean and maximum value of the feature, respectively.

i∈(0,n)，j∈(1,4)。得到训练样本输入数据如下：A class label Y _j , j ∈ 1,4 is added to each set of data for each gas. Then the eigenvectors of the normalized labeled training samples are obtained:

i∈(0,n), j∈(1,4). The input data of the training sample is obtained as follows:

Among them: the correspondence between the category label and the gas type and the representation form are:

N ₂ O: Y ₁ =(1, 0, 0, 0); NO: Y ₂ =(0, 1, 0, 0);

CH ₄ : Y ₃ =(0, 0, 1, 0); CO ₂ : Y ₄ =(0, 0, 0, 1).

(n为每种气体的样本数量)进行缩放，这保证了网络中所有的神经元最初的输出分布大致相同，并在经验上提高了收敛速度。将偏置全部初始化为0。至此得到每层的初始化权重矩阵θ，θ＝(θ¹，θ²)。4. Set the initialized network weights, which can normalize the variance of each neuron's output to 1 by scaling its weight vector by the square root of its inputs (ie, the number of inputs). Let the initial weight be a random floating point number between 0 and 1, and then pass

(n is the number of samples for each gas), which ensures that all neurons in the network initially have roughly the same output distribution and empirically improves the speed of convergence. Initialize all offsets to 0. So far, the initialization weight matrix θ of each layer is obtained, θ=(θ ¹ , θ ² ).

传输过程为：传递过程为x＝a⁽¹⁾→z⁽²⁾→a⁽²⁾→z⁽³⁾→a⁽³⁾，具体为：5. Input the feature vector of the training sample obtained in step 3 into the set neural network input layer, carry out forward propagation, calculate the state value and activation value of the input layer, hidden layer, and output layer, and the final result is the output layer. activation value. make

The transmission process is: the transmission process is x=a ⁽¹⁾ →z ⁽²⁾ →a ⁽²⁾ →z ⁽³⁾ →a ⁽³⁾ , specifically:

a ⁽¹⁾ = x ⁽¹ )

z ⁽²⁾ = θ ² a ⁽¹⁾ + b ⁽²⁾

a ⁽²⁾ = g(z ⁽²⁾ )

z ⁽³⁾ = θ ³ a ⁽²⁾ + b ⁽³⁾

h _θ (x) = a ⁽³⁾ = g(z ⁽³⁾ )

where a ^(l) , l∈1,2,3 represent the activation value of the lth layer, z ^(l) , l∈1,2,3 represent the state value of the lth layer, b ^(l) is the lth layer The bias term, g(z ^(l) ), l∈1,2,3 is the activation function, and the activation function adopts the sigmoid function, namely:

The function h _θ (x) is the output predicted value.

为下面梯度检测做准备。6. Calculate the cost function

Prepare for the gradient detection below.

7. Implement back propagation and calculate the propagation error of each layer. Starting from the output layer, the propagation error for the output layer is:

δ ⁽³⁾ = (a ⁽³⁾ -Y).*g′(z ⁽³⁾ )

where Y represents the class label of the gas.

The error of the hidden layer is:

Wherein: g'(z ^(l) )=g(z ^(l) )*(1-g(z ^(l) ));

.* means dot multiplication, i.e. matrix element multiplication.

Then the partial derivative of the cost function with respect to the weight matrix of layer l:

ε＝1×10^-4,若D＝D^(l)则说明反向传播正常。继续下一步。where bD ^(l) is the partial derivative of the bias term. In order to verify whether the partial derivative calculated by backpropagation is correct, gradient detection is required. The partial derivative of the cost function with respect to the weight is the slope of the cost function when θ takes the value, then

ε=1×10 ^-4 , if D=D ^(l) , it means that the backpropagation is normal. Proceed to the next step.

8. Use the gradient descent method to optimize the weight parameters to minimize the cost function, even if the difference between the output prediction and the actual gas category is minimal. The optimization formula for gradient descent is:

θ ^(l) = θ ^(l) - αD ^(l)

b ^(l) = b ^(l) - αbD ^(l)

α is a training step parameter, which needs to be adjusted manually. Usually, the default initial value is 0.01. If the training is too slow, increase α. If it still fails to converge, decrease α. Each time this weight and bias matrix is updated, another forward and reverse neural network propagation is performed. Determine whether the error meets the accuracy requirements. If it is satisfied, go to the next step and continue to optimize.

9. When the error meets the accuracy requirements, that is, the neural network has been trained, then the neural network model can correctly classify the gas entering the gas chamber.

10. Pass the gas to be tested (one of the training gases) into the gas chamber. For the gas that is passed into the gas chamber, the gas detection unit preprocesses the received data of the gas into a feature vector, and inputs it into the artificial neural network. Identify and obtain the type result of the gas to be tested.

11. For the type of the gas to be measured, call the corresponding parameters, and calculate the concentration x according to the following formula:

in:

为吸收线型，采用voigt线型，S(T)为气体分子吸收线强，

为1f归一化的2f信号强度，P为气室压强，L为气体在气室内的光程长度。

可通过气体检测单元对信号预处理得到；L可通过气室的长度和激光在气室内的反射次数计算得到；P通过向气室中通入待测气体的流速计算得到。

is the absorption line type, using the voigt line type, S(T) is the gas molecular absorption line intensity,

is the 2f signal intensity normalized to 1f, P is the gas chamber pressure, and L is the optical path length of the gas in the gas chamber.

It can be obtained by preprocessing the signal by the gas detection unit; L can be calculated by the length of the gas chamber and the number of reflections of the laser in the gas chamber; P is calculated by the flow rate of the gas to be tested into the gas chamber.

For the above four gases, parameters such as S(T) are stored in the concentration calculation module of the gas detection unit in advance. After the neural network recognizes the category of the gas to be tested, it will automatically call and calculate the gas concentration.

12. At this point, the type identification and concentration measurement of the trace gas to be measured are completed.

Claims (10)

1. A gas detection gas chamber, characterized in that: the two ends of the gas chamber are respectively equipped with mirrors opposite to mirror surfaces, and the mirrors are spherical mirrors with a radius of curvature of 500 mm. 2 . The gas detection gas chamber according to claim 1 , wherein the spherical mirror is a spherical mirror array composed of several small spherical mirrors. 3 . 3 . The gas detection gas chamber according to claim 2 , wherein the small spherical mirror is fixed on the angle adjustment bracket. 4 . 4 . The gas detection gas chamber according to claim 3 , wherein the angle adjustment bracket comprises a bottom plate and a mirror holder distributed in an array on the bottom plate, the mirror holder and the bottom plate are hinged; the small spherical mirror is embedded in the bottom plate. 5 . on the mirror holder. 5 . The gas detection gas chamber according to claim 1 , wherein the input end of the gas chamber is provided with an incident/exit hole with a diameter of 5 mm, and the incident/exit hole is located in the reflector. 6 . center of. 6. A laser spectrum gas detection system based on artificial neural network, characterized in that, comprising: a gas chamber as claimed in any one of claims 1-5, a laser light source, a modem, a data acquisition card, a photodetector and a gas detection unit; the photodetector is connected to the sampling air chamber, and the outgoing light of the air chamber is converted into an electrical signal; the modem is connected to the photodetector, receives and demodulates the electrical signal of the photodetector, and transmits it to the data acquisition card the data acquisition card is connected with the gas detection unit, and transmits the collected data to the gas detection unit; the gas detection unit preprocesses the input data to complete the gas type identification and concentration calculation. 7 . The laser spectrum gas detection system based on artificial neural network according to claim 6 , wherein the gas detection unit comprises a gas identification module and a concentration calculation module, wherein the gas identification module is an artificial neural network. 8 . network model. 8. A laser spectral gas detection method based on artificial neural network, comprising: (1) Construct an artificial neural network model to identify various gas species; (2) The gas to be tested is introduced into the gas chamber, and the data acquisition card transmits the collected data of the gas to be trained to the gas detection unit, and processes it into an input feature vector, input the artificial neural network model, and completes the waiting Identification of the type of gas to be measured; (3), according to the identified composition of the gas to be measured, calculate the concentration x of the gas to be measured:

in,

is the absorption line type, using the voigt line type; S(T) is the gas molecular absorption line intensity, s _2f /1f is the 2f signal intensity normalized by 1f, P is the pressure of the gas chamber, and L is the optical path length in the gas chamber . 9. The optical frequency comb laser spectrum gas detection method based on artificial neural network according to claim 8, is characterized in that: in the construction of artificial neural network, the acquisition method of training sample is: Introduce m kinds of trace gases into the gas chamber in turn to collect data; set a category label Y _j for each gas, j = 1, 2, ..., m; For each gas, the data acquisition card transmits the n sets of data of the gas to the gas detection unit, and the gas detection unit normalizes each set of data: frequency X ₁ , second harmonic X ₂ , and adds the category label Y _j ; then get the feature vector of the labeled training sample:

10. The laser spectrum gas detection method based on artificial neural network according to claim 8, is characterized in that: in the construction of artificial neural network, the training of network model comprises the following steps: (1) Input the feature vector of the training sample to the input layer of the neural network; initialize the network weight matrix θ; (2) implement forward propagation, and calculate the output activation value a ^l ; (3) Calculate the cost function

(4) Backpropagation is performed to calculate the output layer error δ ^l =(a ^l -Y).*a ^l′ , the hidden layer error

Then the partial derivative of the cost function with respect to the weights of each layer is

(5) Carry out gradient detection, check whether the gradient of numerical calculation is consistent with the partial derivative calculated by backpropagation in step (4), if the agreement indicates that the backpropagation is running normally, proceed to the next step; (6) Perform stochastic gradient descent method to optimize the weight matrix, repeat steps (2), (3), (4) until the cost function corresponding to the global optimal weight satisfies the accuracy error, and the neural network model training is completed.

Images