Visible light indoor position sensing system based on biological inspiration optimization width learning

Abstract

The invention discloses a visible light indoor position sensing system based on a biological inspiration optimization width learning network model. The method comprises the specific steps that firstly, a photoelectric detector collects illumination intensity at different positions in a 4m * 4m * 3m three-dimensional space, and a database is established in real time; inputting the training data set into a network model for model training; and finally, inputting the test set into the trained network model to obtain a corresponding positioning result, and sending the positioning result to a receiving end. According to the method, the biological heuristic algorithm and the width learning network are combined, so that high-dimensional and large-scale data challenges in the field of machine learning are overcome. Besides, aiming at the problem that the weight and bias of width learning input are randomly selected and can influence the test performance of the algorithm, the invention adopts a biological inspiration algorithm so as to effectively improve the positioning precision and stability.

Description

Technical Field

[0001] The present invention belongs to the technical field of visible light positioning, and in particular relates to a visible light indoor position perception system based on biologically inspired optimized width learning. Background Art

[0002] In recent years, indoor location sensing technology has been widely used in smart homes, indoor navigation and other fields. Traditional indoor location sensing technology has limited accuracy and high hardware requirements, such as wireless signal strength measurement, ultrasonic positioning, laser ranging, visual recognition, magnetic field positioning, etc. Compared with traditional indoor location sensing technology, visible light indoor location sensing has outstanding advantages such as higher accuracy and anti-electromagnetic interference. In addition, existing deep learning is complex, the network training speed is slow, and the accuracy is low. Summary of the invention

[0003] In order to solve the problem that the above algorithm is not optimized enough, resulting in low positioning accuracy and large amount of sampling data in a large environment, a visible light indoor position perception system based on bio-inspired optimized width learning is proposed.

[0004] The present invention is implemented by the following technical scheme: a visible light indoor position perception system based on biologically inspired optimized width learning, the specific steps of which are:

[0005] Step 1: Build a three-dimensional indoor space of 4m×4m×3m.

[0006] Step 2: The photoelectric detector collects light intensity at heights of 0m, 0.5m, 1m, and 1.5m to establish a training data set.

[0007] Step 3: Train the network model for biologically inspired optimized width learning.

[0008] Step 4: Input the test set into the trained network model to obtain the positioning result.

[0009] Specifically, in step 1, the DC gain of the first-order reflection of the optical channel of the visible light indoor channel propagation model is The expression is:

[0010]

[0011] Where m t 、m r They represent the Lambertian luminescence order of the transmitting end and the receiving end respectively, d1 and d2 are the distances from the LED to the reflection point and from the reflection point to the receiver, and dS f is the area element of the reflector, A is the effective area of ​​the photodetector, ρ is the material coefficient of the reflector, T sis the gain of the optical filter, α and β are the incident and exit angles of the reflected light signal, and θ is the emission angle relative to the vertical axis of the LED. is the incident angle relative to the receiving axis. According to the characteristics of the photodiode, the demodulated electrical signal current is proportional to the intensity of the received optical power. The relationship between the LED's optical radiation power and the electrical power received by the receiving end PD is:

[0012]

[0013] Where R is the responsiveness of the receiving PD, P t is the optical radiation power of a single LED, H(0) is the DC gain of the optical channel, is the total noise variance.

[0014] Specifically, in step 2, the photoelectric detector collects data every 10 cm.

[0015] Among them, the working principle of the photoelectric detector is that the microcontroller reads the light intensity sensor, obtains the actual light intensity after calculation, and transmits the data through the serial port.

[0016] Among them, the TTL to 485 module is a circuit module that can convert TTL signals to RS485 signals, making communication between devices easier. It uses TTL level signals (usually 5V or 3.3V) as input and converts them to RS485 level signals (usually 12V or 24V), making communication between devices easier.

[0017] Specifically, in step 3, in the width learning network model, feature mapping is first performed on the input data X to obtain n groups of mapping features, and the formula is as follows:

[0018]

[0019] in is the randomly generated connection weight coefficient, is a randomly generated bias, and φ is a linear transformation. Then Z a =[Z1,Z2,...,Z a ] is the feature node formed by the combination of the first a groups of mappings. Assuming that each group of features has η1 nodes, there are η1×a feature nodes in group a. The feature nodes are then enhanced to enhanced nodes with randomly generated weights. The formula is as follows:

[0020]

[0021] in and is also a randomly generated connection weight coefficient and bias, and ξ is a nonlinear activation function. Then E b=[E1,E2,...,E b ] is b groups of enhanced nodes, and the feature nodes are converted to generate a total of η2×b enhanced nodes.

[0022] Specifically, in step 3, in the biologically inspired optimization algorithm, it is assumed that the target problem corresponds to a d-dimensional solution space, and N individuals P = (X1, X2, ..., X N ) are distributed in this solution space, and the position of the i-th individual in the solution space can be expressed as a d-dimensional vector X i =(x i1 ,x i2 ,...,x id ) T Obviously, the position of each individual corresponds to a potential solution to the target problem. In addition, the bio-inspired optimization algorithm divides the entire population P into sub-populations SP1 (N1 individuals) and sub-populations SP2 (N2 individuals) according to the fitness value, and the migration operator and adjustment operator act on the two sub-populations respectively to select the best individual.

[0023] Specifically, in step 4, sending the positioning results to the transmitter requires an ESP32 development board and a serial port screen. Connect the serial port screen module to ESP32, and ensure that the TX and RX pins are connected, as well as the power supply and ground. When writing the ESP32 firmware, use Arduino to generate the initialization code and configure the serial port communication and GPIO.

[0024] The present invention utilizes a biologically inspired optimization width learning network model to effectively improve the accuracy and efficiency of visible light indoor position perception.

[0025] Preferably, the biologically inspired optimized width learning network is trained by:

[0026] Step 3.1: Set the initial generation t = 0, set the maximum number of iterations MaxGen, the population size N, set the subpopulation 1 size to N1, and the subpopulation 2 size to N2. Set the adjustment rate BAR, the migration period peri, and the migration rate p;

[0027] Step 3.2: Use the average precision of width learning as the fitness function. Calculate the fitness value of each butterfly through the np.apply_along_axis function.

[0028] Step 3.3: According to the update rule of the population optimization algorithm, calculate the updated value of the butterfly population position and find the best and worst in the group by sorting fitness;

[0029] Step 3.4: Determine whether the maximum number of iterations is met. If so, store the optimal bias and weight. If not, repeat 2-3 until the optimal solution is obtained.

[0030] Step 3.5: Provide optimal weights and biases for width learning in the test set;

[0031] Step 3.6: In the width learning model, the input data is estimated by matrix multiplication after being processed by weights and bias terms, and random features are obtained by ReLU activation function.

[0032] Step 3.7: Concatenate the original input data and random feature levels together to form enhanced features;

[0033] Step 3.8: Calculate the output weights by the pseudo-inverse matrix of the enhanced features and the training target;

[0034] Step 3.9: Calculate the error and update the weights. If the conditions are met, the predicted value is obtained. If the conditions are not met, repeat this step until the conditions are met.

[0035] The indoor location perception system based on biologically inspired optimized width learning of the present invention has the following advantages and beneficial effects:

[0036] Common deep learning increases the network's learning ability for complex nonlinear functions by continuously increasing the number of network layers when facing richer and larger-dimensional data sets, but most neural networks are time-consuming due to their cumbersome structures and redundant hyperparameters. The introduction of the width learning system has become a new machine learning concept to replace deep neural networks. The width learning system does not require stacking network layers, and its algorithm complexity is low and the training speed is fast. It solves the problems caused by the excessive amount of data and data dimensions of traditional neural networks, and eliminates the process of repeated training, solving or adjusting parameters. Applying the width learning system to visible light indoor position perception can effectively improve positioning performance and provide a stable and reliable method for indoor position perception. The present invention uses a learning algorithm that combines a bio-inspired algorithm and a width learning algorithm to train the network model. The bio-inspired algorithm optimizes the weights and biases randomly selected by width learning, effectively improving the stability of the system and the accuracy of indoor visible light positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of a visible light indoor position perception system based on biologically inspired optimized width learning of the present invention;

[0038] Figure 2 It is a visible light channel propagation model; the four LEDs emit sinusoidal signals of the same frequency;

[0039] Figure 3 It is a schematic diagram of the width learning network model of the present invention. BLS maps input samples to a suitable feature space, processes high-dimensional and time-varying data, and retains the mechanism of randomly generating hidden layer node weights;

[0040] Figure 4 The test results of the network model after training using the training method of the present invention are shown in the figure. The cumulative probability distribution of the positioning error is given;

[0041] Figure 5 The figure shows the test results of the network model after training using the particle swarm optimization width learning training method. The cumulative probability distribution of the positioning error is given in the figure. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with the embodiments and drawings.

[0043] The embodiment of the present invention provides a visible light indoor position perception system based on biologically inspired optimized width learning. Figure 1 , including the following steps: First, the photodetector samples at different positions, records the light intensity and the real position at the same time, and builds a training data set in real time; then the training set is input into the biologically inspired optimized width learning network model to train the network. Finally, the test data set is input into the trained network model to obtain the predicted coordinates, and finally the position information is transmitted to the serial port screen through ESP32 to realize the visible light indoor position perception system.

[0044] In order to verify the method of the present invention, this embodiment built an indoor visible light positioning platform. Figure 3 As shown, in this embodiment, the platform is a 4m×4m×3m three-dimensional space. Sampling is performed at planes at 0m, 0.5m, 1m, 1.5m, and 2m in height, and the light intensity and true position are recorded. There are four light sources on the plane at 3m in height.

[0045] In this experiment, the light intensity is used as the input of the network training, and the predicted coordinates are used as the output target. In the test phase, the light intensity at different sampling positions is used as input, and the output of the network is compared with the actual position to analyze the positioning accuracy.

[0046] Combine the following Figure 2 and Figure 3 , the specific network construction and training process of the embodiment of the present invention is described.

[0047] Implementation steps:

[0048] S1: Parameter initialization, set the initial generation t = 0, set the maximum number of iterations MaxGen, the population size N, set the subpopulation 1 size N1, and the subpopulation 2 size N2. Set the adjustment rate BAR, the migration period per i, and the migration rate p;

[0049] S2: Find the fitness value of the individual solution of the training sample, using the BLS average accuracy as the fitness function;

[0050] S3: Find the best and worst in the group by ranking fitness;

[0051] S4: Determine whether the maximum number of iterations is met. If so, store the optimal bias and weight. If not, repeat 2-3 until the optimal solution is obtained.

[0052] S5: Provide optimal weights and biases for width learning in the test set;

[0053] S6: Calculate the error and update the weights. If the conditions are met, the predicted value is obtained. If the conditions are not met, repeat this step until the conditions are met. At this point, the network training is completed.

[0054] Input the random group data into the trained network model, and the cumulative probability distribution of its positioning error is as follows: Figure 4 As shown. Randomly input 50 sets of data into the width learning network, and the cumulative probability distribution of its error is obtained as follows Figure 5 shown.

[0055] Depend on Figure 4 and Figure 5 As we know, the confidence probability of the prediction point error of the biologically inspired optimized width learning network algorithm is less than 0.1m is 75%, and the confidence probability of less than 0.3m is 95%. Compared with the results of the BLS positioning algorithm, the confidence probability of the positioning error less than 0.1m is increased by 8%, and the confidence probability less than 0.3m is increased by 7%. This shows that the prediction error of the width learning network model after biologically inspired optimization is much smaller than the error predicted by the width learning network model, and it has higher positioning accuracy.

[0056] The above is a detailed description of the present invention "Visible light internal position perception system based on bio-inspired optimized width learning", but the implementation methods of the present invention are not limited to the above embodiments, and any other changes that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A visible light indoor position perception system based on biologically inspired optimized width learning, characterized in that: The following steps are involved: Step 1: Build a three-dimensional indoor space of 4m×4m×3m; Step 2: The photoelectric detector collects light intensity at heights of 0m, 0.5m, 1m, and 1.5m to establish a training data set; Step 3: training a network model for biologically inspired optimization width learning; Step 4: Input the test set into the trained network model to obtain the positioning result.

2. According to claim 1, a visible light indoor position perception system based on biologically inspired optimized width learning is characterized in that: In step 1, the DC gain of the first-order reflection of the optical channel of the visible light indoor channel propagation model is The expression is: Where m t 、m r They represent the Lambertian luminescence order of the transmitting end and the receiving end respectively, d1 and d2 are the distances from the LED to the reflection point and from the reflection point to the receiver, and dS f is the area element of the reflector, A is the effective area of ​​the photodetector, ρ is the material coefficient of the reflector, T s is the gain of the optical filter, α and β are the incident and exit angles of the reflected light signal, and θ is the emission angle relative to the vertical axis of the LED. is the incident angle relative to the receiving axis; according to the characteristics of the photodiode, the demodulated electrical signal current is proportional to the intensity of the received optical power. The relationship between the LED light radiation power and the electrical power received by the receiving end PD is: Where R is the responsiveness of the receiving PD, P t is the optical radiation power of a single LED, H(0) is the DC gain of the optical channel, is the total noise variance.

3. According to claim 1, a visible light indoor position perception system based on biologically inspired optimized width learning is characterized in that: In step 2, the photoelectric detector collects data every 10 cm. The working principle of the photoelectric detector is that the single chip microcomputer reads the light intensity sensor, obtains the actual light intensity through calculation, and transmits the data through the serial port.

4. According to claim 1, a visible light indoor position perception system based on biologically inspired optimized width learning is characterized in that: In step 3, in the width learning network model, feature mapping is first performed on the input data X to obtain n groups of mapping features, and the formula is as follows: in is the randomly generated connection weight coefficient, is a randomly generated bias, φ is a linear transformation, then Z a =[Z1,Z2,...,Z a ] is the feature node formed by the combination of the first a groups of mappings. Assuming that each group of features has η1 nodes, there are η nodes in the a group. 1 ×a feature nodes; then enhance the feature nodes into enhanced nodes with randomly generated weights, the formula is as follows: in and is also a randomly generated connection weight coefficient and bias, ξ is a nonlinear activation function; then E b =[E1,E2,...,E b ] is b groups of enhanced nodes, and the feature nodes are converted to generate a total of η2×b enhanced nodes.

5. The visible light indoor position perception system based on biologically inspired optimized width learning according to claim 1, characterized in that: In step 3, in the biologically inspired optimization algorithm, it is assumed that the target problem corresponds to a d-dimensional solution space, and N individuals P = (X1, X2, ..., X N ) are distributed in this solution space, and the position of the i-th individual in the solution space can be expressed as a d-dimensional vector X i =(x i1 ,x i2 ,...,x id ) T ,Obviously, the position of each individual corresponds to a potential solution to the target problem.

6. The visible light indoor position perception system based on biologically inspired optimized width learning according to claim 1, characterized in that: In step 3, the bio-inspired optimization algorithm divides the entire population P into sub-population SP1 (N1 individuals) and sub-population SP2 (N2 individuals) according to the fitness value, and the migration operator and the adjustment operator act on the two sub-populations respectively to select the optimal individual.

7. The visible light indoor position perception system based on biologically inspired optimized width learning according to claim 1, characterized in that: In step 4, sending the positioning results to the transmitter requires an ESP32 development board and a serial port screen. Connect the serial port screen module to ESP32, ensure that the TX and RX pins are connected, as well as power supply and grounding. When writing the ESP32 firmware, use Arduino to generate initialization code and configure serial port communication and GPIO.

8. The visible light indoor position perception system based on biologically inspired optimized width learning according to claim 1, characterized in that: In step 3, the biologically inspired optimization width learning network is trained by: Step 3.1: Set the initial generation t=0, set the maximum number of iterations MaxGen, the population size N, set the size of subpopulation 1 to N1, the size of subpopulation 2 to N2, set the adjustment rate BAR, the migration period peri, and the migration rate p; Step 3.2: Take the average precision of width learning as the fitness function and calculate the fitness value of each butterfly through the np.apply_along_axis function; Step 3.3: According to the update rule of the population optimization algorithm, calculate the updated value of the butterfly population position and find the best and worst in the group by sorting fitness; Step 3.4: Determine whether the maximum number of iterations is met. If so, store the optimal bias and weight. If not, repeat 2-3 until the optimal solution is obtained. Step 3.5: Provide optimal weights and biases for width learning in the test set; Step 3.6: In the width learning model, the input data is estimated by matrix multiplication after being processed by weights and bias terms, and random features are obtained by ReLU activation function. Step 3.7: Concatenate the original input data and random feature levels together to form enhanced features; Step 3.8: Calculate the output weights by the pseudo-inverse matrix of the enhanced features and the training target; Step 3.9: Calculate the error and update the weights. If the conditions are met, the predicted value is obtained. If the conditions are not met, repeat this step until the conditions are met.

Images