CN107451651A - A kind of driving fatigue detection method of the H ELM based on particle group optimizing - Google Patents

Abstract

The invention discloses a kind of driving fatigue detection method of the H ELM based on particle group optimizing；Specially EEG signals are driven using the collection of brain wave acquisition equipment；The EEG signals collected are pre-processed；Discrete Fourier transform is carried out to pretreated data adding window；Power spectral density is sought according to the data after discrete Fourier transform, and frequency band division is carried out according to the frequency band of EEG signals, feature is used as using the power of each frequency band；Transfinite learning machine progress classification learning, identification are learnt using multilayer to the feature of extraction；Optimized by classification of the particle cluster algorithm to the learning machine that transfinites, recognition effect.The present invention is detected using the H HELM graders after PSO optimizations to driving fatigue, can effectively improve classification and Detection accuracy rate.

Description

A driving fatigue detection method based on particle swarm optimization H-ELM

technical field

The invention relates to a driving fatigue detection method, in particular to a driving fatigue detection method based on particle swarm optimization multi-layer learning extreme learning machine.

Background technique

Extreme Learning Machine (ELM) is a single hidden layer feedforward neural network (Single-Hidden Layer Feedforward Neural Networks, SLFNs), only one hidden layer. Compared with the traditional BP neural network, which requires multiple iterations to adjust parameters, the training speed is slow, it is easy to fall into local minimum, and it cannot reach the global minimum. The ELM algorithm randomly generates the connection weights of the input layer and the hidden layer. Bias, and there is no need to iteratively adjust the hidden layer parameters during the training process, only need to set the number of hidden layer neurons and the activation function, and then the output weight can be obtained by minimizing the square loss function.

The multi-layer learning extreme learning machine (H-ELM) first converts the input into a random feature space, and then extracts the high-level features of the data through multi-layer unsupervised learning, and then uses the ordinary extreme learning machine to process the features. Learn and classify. The choice of the norm and scale factor of the ELM directly affects the classification effect of the ELM.

Contents of the invention

The purpose of the present invention is on the basis that power spectrum carries out feature extraction to signal, combines particle swarm optimization algorithm (PSO) to carry out iterative optimization to norm and scale factor of multilayer learning extreme learning machine, proposes a kind of based on particle Group-optimized multilayer learning extreme learning machine (PSO-HELM) approach to driver fatigue detection.

According to the technical scheme provided by the present invention, a kind of multi-layer learning extreme learning machine driving fatigue detection method based on particle swarm optimization is proposed, comprising the following steps:

Step 1. Use EEG acquisition equipment to collect 32 channels of driving EEG signals;

Step 2. Preprocessing the collected EEG signals, including frequency reduction and noise reduction;

Step 3, performing discrete Fourier transform on the preprocessed data window;

Step 4, calculate the power spectral density according to the data after the discrete Fourier transform, and divide the frequency bands according to the frequency bands of the EEG signal, and use the power of each frequency band as a feature;

Step 5, using the multi-layer learning extreme learning machine to classify and identify the extracted features;

Step 6: Optimizing the classification and recognition effects of the extreme learning machine through the particle swarm optimization algorithm.

Carry out frequency band division according to the frequency band of EEG signal in the described step 4, specifically: extract five frequency bands in the frequency domain signal of each channel, be respectively δ (0.1-3Hz), θ (4-7Hz), α (8-15Hz), β(16-31Hz), γ(32-50Hz);

In the described step 5, the steps of classification learning and recognition performed by the multi-layer learning extreme learning machine are as follows:

5-1. Transform the input into a random feature space;

5-2. After K layers of hidden layers, each layer of hidden layers carries out unsupervised learning, and the output H _K represents the high-level features of the input data, and then the features are learned and classified by common extreme learning machines;

The output of each hidden layer can be expressed as

H _i =g(H _i-1 ·β),

Wherein, H _i is the output of the i-th hidden layer, H _i-1 is the output of the i-1 hidden layer, g ( ) is the activation function of the hidden layer, and β is the output weight of the hidden layer;

In the process of hidden layer self-learning, use the infinite approximation method to design the hidden layer self-encoding formula

Among them, O _β is the minimum error between the input data X and the output data of the hidden layer, H is the random output mapping of the self-encoder, β is the output weight of the hidden layer, and l1 is the norm optimization parameter.

In the described step 6, the classification and identification effect of the extreme learning machine are optimized by the particle swarm optimization algorithm, and the specific steps are:

6-1. In the D-dimensional space, initialize the initial positions and velocities of M particles, including setting the initial parameters c ₁ and c ₂ of the particle swarm, and determining the position range of each particle and the speed range of each particle;

6-2. Define the optimal position of each particle and the global optimal position of the entire particle swarm:

pbest _i ＝(p _i1 ,p _i2 ,...,p _iD )

gbest _i ＝(g _i1 ,g _i2 ,...,g _iD )

Where pbest is the optimal position of the i-th particle, gbest is the global optimal position of the population, i=1,2,...M;

6-3. Use the initialized parameters to establish a multi-layer learning extreme learning machine, train the model according to the training samples, and calculate the fitness function value;

6-4. Obtain the initial individual and global optimal positions according to the initial fitness value of the particles;

6-5. According to the iterative update formula of the particle's position and velocity, update the state of the particle:

x _ij (t+1)＝x _ij (t)+v _ij (t+1)

v _ij (t+1)＝ωv _ij (t)+c ₁ r ₁ (pbest _ij (t)-x _ij (t))+c ₂ r ₂ (gbest _j (t)-x _ij (t))

Among them, x _ij (t) is the position of particle i in generation j, v _ij (t) is the velocity of particle i in generation j, r ₁ and r ₂ are random numbers in [0,1];

6-6. Calculate the fitness value of the particle, and update the individual and global optimal positions;

6-7. Keep updating iteratively until the maximum number of iterations is reached or the required error condition is met. At this time, the global optimal position is the optimal solution of the parameters;

6-8. Construct a multi-layer learning extreme learning machine with the parameters at this time to detect driving fatigue.

The beneficial effects of the present invention are as follows:

After using the power spectral density for feature extraction, the classification recognition results based on PSO-HELM were compared with those of single HELM classification recognition, traditional SVM classification recognition results, and traditional kNN classification recognition results. The results show that using PSO-optimized H- The HELM classifier has a higher correct rate of detecting driving fatigue, which effectively improves the recognition rate of classification detection.

Description of drawings

Fig. 1 is a schematic diagram of a multi-layer learning extreme learning machine;

Fig. 2 is the flow chart of PSO-HELM iterative optimization.

detailed description:

The present invention will be further described below in conjunction with specific examples. The following description is only for demonstration and explanation, and does not limit the present invention in any form.

The steps that the present invention realizes as shown in Figure 1 and Figure 2 are as follows:

Step 1. Use EEG acquisition equipment to collect driving EEG signals;

Step 2. Preprocessing the collected EEG signals, including frequency reduction and noise reduction;

Step 3, performing discrete Fourier transform on the preprocessed data window;

Step 4, calculate the power spectral density according to the data after the discrete Fourier transform, and divide the frequency bands according to the frequency bands of the EEG signal, and use the power of each frequency band as a feature;

Step 5, using the multi-layer learning extreme learning machine to classify and identify the extracted features;

Step 6: Optimizing the classification and recognition effects of the extreme learning machine through the particle swarm optimization algorithm.

In the described step 4, in the described step 4, the frequency band is divided according to the frequency band of the EEG signal, specifically: five frequency bands are extracted from the frequency domain signal of each channel, which are respectively δ (0.1-3Hz), θ (4-7Hz), α(8-15Hz), β(16-31Hz), γ(32-50Hz) five frequency bands, at this time, the signal feature dimension of each channel has been reduced to five dimensions.

Among them, in step 5, the steps of classifying learning and identifying by the multi-layer learning extreme learning machine are as follows:

5-1. Transform the input into a random feature space;

5-2. Through K layers of hidden layers, each hidden layer carries out unsupervised learning, and the output H _K represents the high-level features of the input data, and then the features are learned and classified by common extreme learning machines;

The output of each hidden layer can be expressed as

H _i =g(H _i-1 ·β),

Wherein, H _i is the output of the i-th hidden layer, H _i-1 is the output of the i-1 hidden layer, g ( ) is the activation function of the hidden layer, and β is the output weight of the hidden layer;

In the process of hidden layer self-learning, use the infinite approximation method to design the hidden layer self-encoding formula

Among them, O _β is the minimum error between the input data X and the output data of the hidden layer, H is the random output mapping of the self-encoder, β is the output weight of the hidden layer, and l1 is the norm optimization parameter.

In the described step 6, the classification and recognition effect of the extreme learning machine are optimized by the particle swarm optimization algorithm, and the specific steps are: 6-1. In the D-dimensional space, initialize the initial positions and speeds of the M particles, including setting Determine the initial parameters c ₁ and c ₂ of the particle swarm, determine the position range of each particle and the speed range of each particle;

6-2. Define the optimal position of each particle and the global optimal position of the entire particle swarm:

pbest _i ＝(p _i1 ,p _i2 ,…,p _iD )

gbest _i ＝(g _i1 ,g _i2 ,…,g _iD )

Where pbest is the optimal position of the i-th particle, gbest is the global optimal position of the population, i=1,2,...M;

6-3. Use the initialized parameters to establish a multi-layer learning extreme learning machine, train the model according to the training samples, and calculate the fitness function value;

6-4. Define the initial individual and global optimal positions according to the initial fitness value of the particles;

6-5. According to the iterative update formula of the particle's position and velocity, update the state of the particle:

x _ij (t+1)＝x _ij (t)+v _ij (t+1)

v _ij (t+1)＝ωv _ij (t)+c ₁ r ₁ (pbest _ij (t)-x _ij (t))+c ₂ r ₂ (gbest _j (t)-x _ij (t))

Among them, x _ij (t) is the position of particle i in generation j, v _ij (t) is the velocity of particle i in generation j, r ₁ and r ₂ are random numbers in [0,1];

6-6. Calculate the fitness value of the particle, and update the values of pbest and gbest;

6-7. Keep updating iteratively until the maximum number of iterations is reached or the required error condition is met. At this time, the global optimal position is the optimal solution of the parameters;

6-8. Construct a multi-layer learning extreme learning machine with the parameters at this time to detect driving fatigue.

Taking 480 driving EEG samples as training data and 960 EEG samples as testing data, the kNN, SVM, H-ELM and PSO-HELM algorithms are used for classification respectively, and the classification results are shown in Table 1 below.

Table 1 Comparison of classification accuracy of four classification algorithms

By comparing the classification recognition rates of the four algorithms, it can be clearly seen that the H-ELM classification algorithm has a better classification effect than the traditional kNN algorithm and SVM algorithm, and PSO-HELM further optimizes the parameters on the basis of the H-ELM algorithm , which increases the classification accuracy by about 2.5%, indicating that PSO-HELM obtains the optimal parameters and effectively improves the performance of the multi-layer learning extreme learning machine.

Claims (4)

1. A kind of 1. driving fatigue detection method of the H-ELM based on particle group optimizing, it is characterised in that this method specifically include as Lower step：

   Step 1, the driving EEG signals using brain wave acquisition equipment 32 channels of collection；

   Step 2, the EEG signals collected are pre-processed, including frequency reducing, noise reduction；

   Step 3, discrete Fourier transform is carried out to pretreated data adding window；

   Step 4, power spectral density is sought according to the data after discrete Fourier transform, and frequency band is carried out according to the frequency band of EEG signals Division, feature is used as using the power of each frequency band；

   Step 5, the feature to extraction learn transfinite learning machine progress classification learning, identification using multilayer；

   Step 6, optimized by classification of the particle cluster algorithm to the learning machine that transfinites, recognition effect.
2. 2. a kind of H-ELM based on particle group optimizing according to claim 1 driving fatigue detection method, its feature exist In：Frequency band division is carried out according to the frequency band of EEG signals in described step 4, is specially：In the frequency-region signal of each channel Extract five frequency bands, respectively δ (0.1-3Hz), θ (4-7Hz), α (8-15Hz), β (16-31Hz), γ (32-50Hz).
3. 3. a kind of H-ELM based on particle group optimizing according to claim 1 driving fatigue detection method, its feature exist In：In described step 5, multilayer study transfinite learning machine carry out classification learning, identification the step of be specially：

   5-1. will be inputted and is converted into a random feature space；

   5-2. passes through K layer hidden layers, and each layer of hidden layer carries out unsupervised learning, the Η of output_KRepresent the high level of input data Feature, now feature is learnt and classified by the common learning machine that transfinites again；

   The output of each of which layer hidden layer can be expressed as

   H_i=g (H_i-1β),

   Wherein, Η_iIt is the output of i-th of hidden layer, Η_i-1It is the output of the i-th -1 hidden layer, g () is the activation of hidden layer Function, β are the output weights of hidden layer；

   During hidden layer self study, hidden layer own coding formula is designed using unlimited approach method

   <mrow> <msub> <mi>O</mi> <mi>&amp;beta;</mi> </msub> <mo>=</mo> <munder> <mi>argmin</mi> <mi>&amp;beta;</mi> </munder> <mo>{</mo> <mo>|</mo> <mo>|</mo> <mi>H</mi> <mi>&amp;beta;</mi> <mo>-</mo> <mi>X</mi> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> <mo>+</mo> <mo>|</mo> <mo>|</mo> <mi>&amp;beta;</mi> <mo>|</mo> <msub> <mo>|</mo> <mrow> <mi>l</mi> <mn>1</mn> </mrow> </msub> <mo>}</mo> </mrow>

   Wherein, O_βFor hidden layer input data X and the minimal error of output data, Η is that the random output of own coding maps, and β is The output weight of hidden layer, l1 are norm optimization parameter.
4. 4. a kind of H-ELM based on particle group optimizing according to claim 1 driving fatigue detection method, its feature exist In：In described step 6, optimized by classification of the particle cluster algorithm to the learning machine that transfinites, recognition effect, specific steps For：

   6-1. initializes the initial position and speed of M particle, including setting population initial parameter c in D dimension spaces₁With c₂, it is determined that each position range of particle and the velocity interval of each particle；

   6-2. defines the optimal location of each particle and the global optimum position of whole population：

   pbest_i=(p_i1,p_i2,...,p_iD)

   gbest_i=(g_i1,g_i2,...,g_iD)

   Wherein pbest is the optimal location of i-th particle, and gbest is the global optimum position of population, i=1,2 ... M；

   6-3. establishes multilayer using the parameter of initialization and learns the learning machine that transfinites, and the model is trained according to training sample, And calculate fitness function value；

   6-4. obtains initial individual and global optimum position according to the initial fitness value of particle；

   Positions and speed iteration more new formula of the 6-5. according to particle, are updated to particle state：

   x_ij(t+1)=x_ij(t)+v_ij(t+1)

   v_ij(t+1)=ω v_ij(t)+c₁r₁(pbest_ij(t)-x_ij(t))+c₂r₂(gbest_j(t)-x_ij(t))

   Wherein, x_ij(t) for particle i in the position in jth generation, v_ij(t) for particle i in the speed in jth generation, r₁And r₂It is [0,1] Random number；

   6-6. calculates the fitness value of particle, and more new individual and global optimum position；

   6-7. keeps iteration renewal, until reaching the iterations of maximum or meeting desired error condition；Now, global optimum Position is the optimal solution of parameter；

   6-8. learns the learning machine that transfinites with parameter structure multilayer now, and driving fatigue is detected.