CN116321008A - Indoor position fingerprint positioning method based on triplet self-encoder - Google Patents

Abstract

The invention relates to an indoor position fingerprint positioning method based on a triplet self-encoder, and belongs to the technical field of positioning. The method comprises the following steps: acquiring a group of WIFI signal intensity data at each position by using a WIFI receiver to acquire the WIFI signal intensity of each position point; preprocessing the acquired data by using a Gaussian filtering mode; constructing a self-encoder-based triplet neural network and a loss function and a BP neural network for final positioning; training the constructed triplet self-encoder by the pretreatment indoor position fingerprint data set, and reserving the structure and parameters of a network model; extracting the preprocessed data set by using a feature extraction network to obtain a dimensionality-reduced data set, training a BP neural network, and storing the feature extraction network and the trained BP neural network as a Tri-Sae model; the method solves the problems of high feature extraction difficulty, low positioning accuracy, more bad data and the like.

Description

An Indoor Location Fingerprint Location Method Based on Triplet Autoencoder

technical field

The invention belongs to the technical field of positioning, and relates to an indoor position fingerprint positioning method based on a triplet autoencoder.

Background technique

Indoor location fingerprint positioning is based on pattern matching, machine learning, deep learning and other technologies, and with the help of computer processing technology, it performs real-time matching of location fingerprints in the database. At present, indoor location fingerprint positioning technology is mainly realized by using neural network and machine learning, using BP neural network to capture the deep features of location fingerprints, and then using machine learning to compare the Euclidean distance between features , to achieve the final position positioning. However, using BP neural network as the feature extraction layer of the indoor positioning system is difficult to capture the relationship between samples, which makes the positioning accuracy of the final system very unsatisfactory.

Contents of the invention

In view of this, the object of the present invention is to provide a method for indoor position fingerprint positioning based on a triplet autoencoder.

To achieve the above object, the present invention provides the following technical solutions:

A method for indoor position fingerprint positioning based on a triple autoencoder, the method comprising the following steps:

S1: Use the WIFI receiver to obtain a set of WIFI signal strength data at each location, and obtain the WIFI signal strength of each location point;

S2: Use Gaussian filtering to preprocess the collected data;

S3: Construct the triplet neural network and loss function based on the autoencoder and the BP neural network for final positioning;

S4: The preprocessed indoor location fingerprint data set trains the constructed triplet autoencoder, prunes the trained model to obtain the feature extraction network, and retains the structure and parameters of the network model;

S5: Use the feature extraction network to extract the preprocessed data set, obtain the data set after dimension reduction, and train the BP neural network, and save the feature extraction network and the trained BP neural network as a Tri-Sae model.

S6: Use the Tri-Sae model and the distance measure to jointly vote for positioning, and finally evaluate the accuracy of the indoor positioning system.

Optionally, in said S2, Gaussian filtering is performed on the collected data, bad data in the sample is removed, effective values are selected from the high probability area of the RSSI value, and the average value of the output value is calculated to improve the measurement efficiency. From precision, its procedure is:

RSSI值的均值μ的计算公式为/>

σ²的计算公式为/>

如果RSSI值在[μ-σ，μ+σ]的范围内，则为高概率置信区间。通过高斯滤波器计算出的平均RSSI值减少环境干扰的影响。Suppose the total number of n RSSI values obeys Gaussian distribution, the mean value of RSSI value is μ, the variance of RSSI value is σ ² , and the probability density function of RSSI value is f(x), so

The formula for calculating the mean value μ of the RSSI value is />

The calculation formula of σ ² is />

If the RSSI value is in the range [μ-σ, μ+σ], then it is a high-probability confidence interval. The average RSSI value calculated by a Gaussian filter reduces the influence of environmental interference.

Optionally, in the S3, the triplet network construction is divided into design sub-networks, and the process of adding a loss function loss to construct a complete network:

S31: The sub-network of the triplet network is a stacked autoencoder, and the network structure of [input layer + [256, 128, 64, 128, 256] + output layer] is designed. The autoencoder is divided into upper, middle and lower layers. The upper and lower layers share weights, and the activation function of each layer is set to

S32: Define the loss function loss of the new neural network model. Figure 1 shows the calculation idea of the loss function loss and the triplet neural network architecture, which are used to complete the construction of the neural network model. The loss function loss of the neural network model The process is as follows:

Input the location fingerprint data into the stacked autoencoder to extract the feature vector of each sample data, and the feature dimension of each sample data is the same. Three samples are selected for each training as the training data, which are the representative feature anchor of each category (obtained by simply calculating the average value), the positive sample positive of the anchor (from the same position as the anchor), and the negative sample negative of the anchor (with the anchor from different positions), and input them to the stacked autoencoder (a sub-network of the triplet network) to obtain the corresponding feature vector.

What the present invention wants is to allow the Euclidean distance of the vector obtained by anchor and positive to be as small as possible; to allow the Euclidean distance of the vector obtained by anchor and negative to be as large as possible; the formula can be made

成立，；同时positive和negative的样本来自原始数据集本身，经过了堆叠自编码器后，还原到了与输入数据相同的维度，想让输入结果和输出结果尽量保持一致，就需要满足/>

β为任意大于零的实数，模型就具有了较好的效果，所以loss用公式表达为

Established; at the same time, the positive and negative samples come from the original data set itself. After passing through the stacked autoencoder, they are restored to the same dimension as the input data. If you want to keep the input results and output results as consistent as possible, you need to satisfy />

β is any real number greater than zero, and the model has a better effect, so the loss is expressed as

S33: A BP neural network model is designed for the final positioning function. This neural network is a simple multi-layer structure, and its network structure is [input layer + [128, 128] + output layer], and the output layer uses sigmoid Activation function, the remaining layers use the ReLU activation function, and add a Dropout layer with a discard rate of 0.1 to prevent overfitting;

Optionally, the S4 includes a training process and a pruning process:

和验证数据集/>

其中训练数据集/>

中含有n个张量样本，用式表达为/>

则训练数据集

其中验证数据集/>

中含有T-n个张量样本，用式表达为/>

则训练数据集/>

然后要构建训练数据集，本方案中首先利用每个位置的训练数据集构造该位置的anchor指纹张量，构造anchor指纹张量的过程为：将R_i的训练数据集/>

中的RSSI指纹张量求均值，得到F_i，并作为模型输入之一的anchor，其计算公式为/>

上式中的F_i即为R_i的anchor张量，所有N个位置的anchor即构成了指纹集合F，用式表达为F＝{F₁，F₂，…，F_i，…，F_N}，然后为了保证训练的充分性，选择将训练集中的指纹张量与指纹集合F中的指纹张量两两配对得到训练数据集K₁和验证数据集K₂，K₁表达式为K₁＝{T_1，1，1，T_i，j，k，T_N，n，N}，其中元素T_i，j，k表示由训练集中i个位置Ri的第j个指纹张量/>

与第k个位置R_k的anchor指纹张量F_k构造的样本对，可以用式表达为/>

K₂表达式为K₂＝{V_1，1，1，V_i，j，k，V_N，T-n，N}，其中元素V_i，j，k表示由验证集中Ri的第j个指纹张量/>

与R_k的anchor指纹张量F_k构造的样本对，可以用式表达为/>

进而，使用新数据集训练三元组网络模型；S41, in the training model stage, assume that there are N positions available for positioning, and each position point has T samples. After data processing, a training data set is generated for the position R _i

and the validation dataset />

where the training dataset />

Contains n tensor samples, expressed as />

Then the training data set

where the validation dataset />

Contains Tn tensor samples, expressed as />

Then the training data set />

Then to construct the training data set, in this scheme, first use the training data set of each position to construct the anchor fingerprint tensor of the position, the process of constructing the anchor fingerprint tensor is: the training data set of R _i

Calculate the mean value of the RSSI fingerprint tensor in F _i , and use it as one of the model input anchors, the calculation formula is />

F _i in the above formula is the anchor tensor of R _i , and all the anchors of N positions constitute the fingerprint set F, expressed as F={F ₁ , F ₂ ,..., F _i ,...,F _N }, and then in order to ensure the adequacy of training, choose to pair the fingerprint tensor in the training set with the fingerprint tensor in the fingerprint set F to obtain the training data set K ₁ and the verification data set K ₂ , the expression of K ₁ is K ₁ = {T _{1, 1, 1} , T _{i, j, k} , T _{N, n, N} }, where the element T _{i, j, k} represents the jth fingerprint tensor from the i position Ri in the training set />

The sample pair constructed with the anchor fingerprint tensor F _k of the k-th position R _k can be expressed as />

The expression of K ₂ is K ₂ ={V _{1, 1, 1} , V _{i, j, k} , V _{N, Tn, N} }, where the element V _{i, j, k} represents the jth fingerprint sheet of Ri in the verification set Quantity />

The sample pair constructed with the anchor fingerprint tensor F _k of R _k can be expressed as />

Furthermore, the triplet network model is trained using the new dataset;

S42: After the triplet network is trained, cut off the main network of the triplet, and only keep the first half of the stacked autoencoder, which is saved as the feature extraction network of the indoor positioning system.

Optionally, in S5, constructing a complete neural network model is to connect the saved feature extraction network to the trained BP neural network, and finally save the model and its parameters.

Optionally, in S6, in the testing phase, after obtaining y ₁ sets of fingerprint tensors at the location L _i to be tested, use the same Gaussian filtering method in S2 to preprocess new data to obtain y ₂ samples. First, input the obtained data into the Tri-Sae model in turn, and finally put y ₂ positioning results; then combine the calculated anchor fingerprint set F={F ₁ , F ₂ ,..., F _i ,..., F _N } with y ₂ samples constitute the test sample pair set D={D ₁ ,...,D _j ,...,D _y2 }, where D _j ={k ₁ ,k ₂ ,..., _ki ,...,k _N }, where k _i = (t, F _i ), t is a piece of data in _{the 2} samples of y, and then input this test pair into two well-trained stacked codes, compare the Euclidean distance of the two test pairs, the smallest That is, the positioning result of this test pair, and then vote to get the final result of sample t, iterate y ₂ times, and y ₂ structures will be obtained; based on the above, each position L _i to be tested will get 2×y ₂ results, by voting, to get the final predicted label of the position L _i to be tested.

The beneficial effects of the present invention are:

1. An autoencoder training method based on triplet network is proposed, which provides a new technology for extracting indoor location fingerprint feature vectors. It no longer extracts sample features through simple training of autoencoders, and solves the problem of training set Few, difficult feature extraction, low positioning accuracy and other issues.

2. The location fingerprint database no longer uses the original data of the location fingerprint to store the location information, but converts it into the feature vector corresponding to the location for storage, which not only protects the privacy of the indoor space, reduces the storage space, but also in the positioning process.

3. With the continuous update of subsequent data, the feature extraction model can be continuously trained and improved. It only needs to pass the training data through the feature extraction network, and then it can be saved in the database used to train the positioning model.

4. In the offline stage of positioning, voting is used to improve the overall accuracy of the indoor positioning system and avoid the impact of individual bad data that has not been removed on the final positioning result.

Other advantages, objects and features of the present invention will be set forth in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the investigation and research below, or can be obtained from It is taught in the practice of the present invention. The objects and other advantages of the invention may be realized and attained by the following specification.

Description of drawings

In order to make the purpose of the present invention, technical solutions and advantages clearer, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

Fig. 1 is a schematic diagram of the evaluation process and a schematic diagram of the network architecture of the loss function loss of the present invention;

Fig. 2 is a schematic diagram of the positioning process of the present invention.

Detailed ways

Embodiments of the present invention are described below through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation modes, and various modifications or changes can be made to the details in this specification based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments are only schematically illustrating the basic concept of the present invention, and the following embodiments and the features in the embodiments can be combined with each other in the case of no conflict.

Wherein, the accompanying drawings are for illustrative purposes only, and represent only schematic diagrams, rather than physical drawings, and should not be construed as limiting the present invention; in order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, Enlargement or reduction does not represent the size of the actual product; for those skilled in the art, it is understandable that certain known structures and their descriptions in the drawings may be omitted.

In the drawings of the embodiments of the present invention, the same or similar symbols correspond to the same or similar components; , "front", "rear" and other indicated orientations or positional relationships are based on the orientations or positional relationships shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred devices or elements must It has a specific orientation, is constructed and operated in a specific orientation, so the terms describing the positional relationship in the drawings are for illustrative purposes only, and should not be construed as limiting the present invention. For those of ordinary skill in the art, the understanding of the specific meaning of the above terms.

What the present invention discloses is an indoor position fingerprint positioning technology based on a triplet autoencoder, as shown in Figure 2 in detail, and the detailed steps are as follows:

S1: Install wireless signal transmitters evenly in the indoor environment, use the signal reading device that comes with the mobile phone to obtain the RSSI fingerprint tensor of each location, and set the RSSI value of the nodes with extremely weak or unreceived signals to -110, Finally, construct the original fingerprint database;

S2: The labels of all the samples in the original database correspond to different locations. There are N locations. The expression of the location set R is R={R ₁ , R ₂ ,...,R _i ,...,R _N }, each There are n samples at position R _i , assuming that n samples obey the Gaussian distribution, then calculate the mean value μ and variance σ ² , and select the array RSSI fingerprint tensor in the interval [μ-σ, μ+σ] as the training of R _i The sample is iterated N times in turn until all positions are processed by Gaussian filtering, and the filtered RSSI fingerprint tensor forms a new training data set.

此编码器分为上中下层结构，其中上下两层结构共享权重，节省了参数量，提升了性能；然后就是设计使用三元组网络，此网络结构有三个子网络，且都是上述的堆叠式自编码器，三个子网络间共享权值，网络的输入是特别构造过的样本对，样本对会被拆分成三份，分别喂入到三个子网络里面，最后计算loss函数，使用Adam算法，持续更新参数；最后设计BP神经网络，构造[输入层+[128，128]+输出尾]的网络结构，最后一层使用sigmoid激活函数，其余层都使用ReLU激活函数。S3: First construct a stacked autoencoder as a sub-network of the triplet network, design a network structure using [input layer + [256, 128, 64, 128, 256] + output layer], and set the activation function of each layer to

This encoder is divided into an upper, middle and lower layer structure, in which the upper and lower layers share weights, which saves the amount of parameters and improves performance; then it is designed to use a triplet network. This network structure has three sub-networks, and they are all stacked above. Autoencoder, the weights are shared between the three sub-networks. The input of the network is a specially constructed sample pair. The sample pair will be split into three parts and fed into the three sub-networks respectively. Finally, the loss function is calculated and the Adam algorithm is used. , continuously update the parameters; finally design the BP neural network, construct the network structure of [input layer + [128, 128] + output tail], the last layer uses the sigmoid activation function, and the rest of the layers use the ReLU activation function.

计算每个位置R_i的anchor指纹张量，获得指纹集F＝{F₁，F₂，…，F_i，…，F_N}，再使用逐一匹配的方法构造训练数据集K＝{T_1，1，1，T_i，j，k，T_N，n，N}，将构造好的数据集K喂到构造好的三元组网络中，逐轮训练，使用梯度下降算法Adam持续更新模型的参数，模型训练完成后，只保留三元组模型的子网络的前半部分作为特征提取网络。S4: Use the average formula

Calculate the anchor fingerprint tensor of each position R _i , obtain the fingerprint set F={F ₁ , F ₂ ,..., F _i ,...,F _N }, and then use the one-by-one matching method to construct the training data set K={T _{1 , 1} , 1 , T _{i, j, k} , T _{N, n, N} }, feed the constructed data set K to the constructed triplet network, train round by round, and use the gradient descent algorithm Adam to continuously update the model After the model training is completed, only the first half of the sub-network of the triplet model is reserved as the feature extraction network.

S5: Connect the constructed BP neural network to the back of the feature extraction network, use the preprocessed data set to train the network, and only use the gradient descent algorithm Adam to update the parameters of the BP neural network. After the iteration is completed, the model is finally saved as Tri- Sae network model;

取dist最小的为最终结果，经过多次迭代，最终也将获得n个实验结果，然后通过投票的方式，裁定最终定位结构，根据待测位置标签来评估室内定位系统的精度。S6: First feed the preprocessed pieces of real-time RSSI data into the Tri-Sae model, and get n experimental results; then construct the preprocessed pieces of real-time RSSI data and the anchor fingerprint set into a test sample pair k _i =(t, F _i ) and input to the feature extraction network, compare the Euclidean distance between the two samples

The smallest dist is taken as the final result. After multiple iterations, n experimental results will be finally obtained, and then the final positioning structure will be determined by voting, and the accuracy of the indoor positioning system will be evaluated according to the location label to be tested.

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should be included in the scope of the claims of the present invention.

Claims (6)

1. A method for indoor position fingerprint positioning based on triplet autoencoder, characterized in that: the method may further comprise the steps: S1: Use the WIFI receiver to obtain a set of WIFI signal strength data at each location, and obtain the WIFI signal strength of each location point; S2: Use Gaussian filtering to preprocess the collected data; S3: Construct the triplet neural network and loss function based on the autoencoder and the BP neural network for final positioning; S4: The preprocessed indoor location fingerprint data set trains the constructed triplet autoencoder, prunes the trained model to obtain the feature extraction network, and retains the structure and parameters of the network model; S5: Use the feature extraction network to extract the preprocessed data set, obtain the data set after dimension reduction, and train the BP neural network, save the feature extraction network and the trained BP neural network as a Tri-Sae model; S6: Use the Tri-Sae model and the distance measure to jointly vote for positioning, and finally evaluate the accuracy of the indoor positioning system. 2. A kind of indoor position fingerprint location method based on triplet autoencoder according to claim 1, it is characterized in that: in described S2, the processing of Gaussian filter is taken to the data collected, removes the bad part in the sample. Data, select effective values from the high probability area of RSSI value, and calculate the average value of the output value to improve the ranging accuracy, its process is: Suppose the total number of n RSSI values obeys Gaussian distribution, the mean value of RSSI value is μ, the variance of RSSI value is σ ² , and the probability density function of RSSI value is f(x), so

The formula for calculating the mean μ of the RSSI value is

The calculation formula of σ ² is />

If the RSSI value is in the range of [μ-σ, μ+σ], it is a high-probability confidence interval; the average RSSI value calculated by the Gaussian filter reduces the influence of environmental interference. 3. a kind of indoor position fingerprint location method based on triplet autoencoder according to claim 1, it is characterized in that: in described S3, triplet network structure is divided into design sub-network, adds loss function loss structure The process of the complete network: S31: The sub-network of the triplet network is a stacked autoencoder, and the network structure of [input layer + [256, 128, 64, 128, 256] + output layer] is designed. The autoencoder is divided into upper, middle and lower layers. The upper and lower layers share weights, and the activation function of each layer is set to

S32: Define the loss function loss of the new neural network model, the process of the loss function loss of the neural network model is as follows: Input the location fingerprint data into the stacked autoencoder to extract the feature vector of each sample data, and the feature dimension of each sample data is the same; each training selects three samples as training data, which are representatives of each category The feature anchor, the positive sample positive of the anchor, and the negative sample negative of the anchor, input them to the stacked autoencoder to obtain the corresponding feature vector; makes the formula

Established; at the same time, positive and negative samples come from the original data set itself. After stacking the autoencoder, it is restored to the same dimension as the input data. I want to keep the input and output results as consistent as possible, satisfying />

β is any real number greater than zero, and the loss is expressed as:

S33: Design a BP neural network model for the final positioning function. This neural network is a simple multi-layer structure, and its network structure is [input layer + [128, 128] + output layer], and the output layer uses a sigmoid activation function. The remaining layers use the ReLU activation function, and add a Dropout layer with a dropout rate of 0.1 to prevent overfitting. 4. a kind of indoor position fingerprint positioning method based on triplet autoencoder according to claim 1, is characterized in that: described S4 comprises training process and pruning process: S41, in the training model stage, assume that there are N positions available for positioning, and each position point has T samples. After data processing, a training data set is generated for the position R _i

and the validation dataset />

where the training dataset />

Contains n tensor samples, expressed as />

Then the training data set

where the validation dataset />

Contains Tn tensor samples, expressed as />

Then the training data set />

Then to construct the training data set, first use the training data set of each position to construct the anchor fingerprint tensor of the position, the process of constructing the anchor fingerprint tensor is: the training data set of R _i />

Calculate the mean value of the RSSI fingerprint tensor in F _i , and use it as one of the model input anchors, the calculation formula is />

F _i in the above formula is the anchor tensor of R _i , and all the anchors of N positions constitute the fingerprint set F, expressed as F={F ₁ , F ₂ ,..., F _i ,...,F _N }, and then in order to ensure the adequacy of training, choose to pair the fingerprint tensor in the training set with the fingerprint tensor in the fingerprint set F to obtain the training data set K ₁ and the verification data set K ₂ , the expression of K ₁ is K ₁ = {T _{1, 1, 1} , T _{i, j, k} , T _{N, n, N} }, where the element T _{i, j, k} represents the jth fingerprint tensor from the i position Ri in the training set />

The sample pair constructed with the anchor fingerprint tensor F _k of the k-th position R _k is expressed as />

The expression of K ₂ is K2={V _{1, 1, 1} , V _{i, j, k} , V _{N, Tn, N} }, where the element V _{i, j, k} represents the jth fingerprint tensor of Ri in the validation set />

The sample pair constructed with the anchor fingerprint tensor F _k of R _k is expressed as

Train a triplet network model using the new dataset; S42: After the triplet network is trained, cut off the main network of the triplet, and only keep the first half of the stacked autoencoder, which is saved as the feature extraction network of the indoor positioning system. 5. a kind of indoor position fingerprint location method based on triplet autoencoder according to claim 1, it is characterized in that: in described S5, constructing complete neural network model is exactly that the feature extraction network that preserves is connected to The trained BP neural network, and finally save the model and its parameters. 6. a kind of indoor position fingerprint positioning method based on triplet autoencoder according to claim 1, is characterized in that: in described S6, test stage, obtains y ₁ group fingerprint tensors at position L _to be tested Finally, use the same Gaussian filtering method in S2 to preprocess the new data to obtain y ₂ samples; firstly, input the obtained data into the Tri-Sae model in turn, and finally y ₂ positioning results; then the calculated anchor fingerprint The set F={F ₁ , F ₂ ,...,F _i ,...,F _N } and y ₂ samples constitute the test sample pair set D={D ₁ ,...,D _j ,...,D _y2 }, where D _j ={k ₁ , k ₂ ,..., _ki ,...,k _N }, where _ki =(t, F _i ), t is a piece of data in y ₂ samples, and then this test pair is input into two In a stacked code that has been trained, compare the Euclidean distances of two test pairs, the smallest one is the positioning result of this test pair, and then vote to get the final result of sample t, and iterate y ₂ times, you will get y ₂ structures; each location L _i to be tested will get 2×y ₂ results, and the final predicted label of the location L _i to be tested will be obtained by voting.

Images