People counting method in underground restricted area based on LoRa adaptive model

Abstract

The invention discloses a personnel counting method in an underground restricted area based on a LoRa adaptive model. A signal transmitter and a signal receiver of a counting system are deployed in an underground restricted area of ​​a coal mine. The signal receiver is used to receive a LoRa signal including static and active signals of an area environment and all personnel emitted by the signal transmitter. The deployment method is simple and the mine workers do not need to wear any equipment. The signal receiver uses two antennas to receive signals and combines signal quotient and multi-frequency hopping to expand the signal dimension and enrich signal data. The received original LoRa signal is downsampled, amplitude data is extracted, and normalization processing is performed after rearrangement. A lightweight neural network model based on domain confrontation is constructed. The model is trained and tested, so that it can be used for accurate counting of the number of personnel, and relatively ideal accuracy and reliability can be achieved for counting personnel in the underground restricted area.

Description

Technical Field

[0001] The invention relates to a method for counting people in a limited-personnel area in an underground coal mine, in particular to a method for counting people in a limited-personnel area in an underground coal mine based on a LoRa adaptive model, and belongs to the technical field of safe production in coal mines. Background Art

[0002] As my country's strategic core energy, coal has long provided a solid backing for the stable and rapid economic growth. However, the coal mining process has a complex underground mining environment and complicated mining process. With the continuous increase in coal mining intensity and the continuous extension of mining depth, accompanied by various uncertainties in the coal mining process, the risk of coal mine accidents has gradually increased, and coal mine safety accidents occur from time to time. In order to prevent special and major accidents in coal mines, coal mines are strictly prohibited from over-staffing production. Real-time monitoring of the number of underground workers in coal mines is one of the effective measures to prevent over-staffing production. Especially in the coal mining face, tunneling face and other personnel-limited areas with potential impact ground pressure hazards, the number of personnel must be strictly controlled. When the number of personnel exceeds the specified number, the ground computer displays an alarm message, so that the administrator can grasp the number of personnel in the tunnel in real time and take relevant measures immediately.

[0003] The traditional method of counting people in the restricted area of ​​underground coal mines is usually to count people by carrying identification cards, but the identification card system cannot recognize the phenomenon of multiple people on one card, and often misses the reading, so accurate counting cannot be achieved. With the popularization and application of wireless network technology, wireless network technologies such as radio frequency identification (RFID) technology, wireless local area network (Wi-Fi) technology, and millimeter wave radar technology have emerged in the prior art to realize the task of wirelessly sensing the number of people in the restricted area of ​​underground coal mines. By setting up card readers in the restricted area of ​​underground coal mines and wirelessly sensing the electronic cards with unique identification codes equipped by underground coal mine workers (generally the electronic cards are stored in the safety helmets of underground coal mine workers), the personal information and entry time of underground coal mine workers are transmitted to the ground control system. However, the communication signals of the above-mentioned wireless perception are usually affected by two aspects: ① The harsh environment underground in coal mines causes serious attenuation of wireless signals and severely limits the perception distance. Taking Wi-Fi technology as an example, although its communication distance can reach tens of meters, its perception distance is usually less than 8 meters because it analyzes weak reflected signals. This makes Wi-Fi technology used for crowd counting tasks almost only applicable to small indoor scenes, and the application scope of crowd counting based on Wi-Fi wireless signals is severely limited; ② The temporal instability of wireless signals leads to poor perception performance. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a personnel counting method in a restricted area underground based on a LoRa adaptive model, which can achieve ideal accuracy and reliability in counting personnel in the restricted area underground.

[0005] To achieve the above purpose, the personnel counting method in the underground limited area based on the LoRa adaptive model specifically includes the following steps:

[0006] Step 1, the signal transmitter and signal receiver of the counting system are fixedly deployed in the restricted area of ​​the underground tunnel. The signal transmitter is used to transmit linear frequency modulation LoRa signals. The signal transmitter includes a LoRa node and a directional antenna electrically connected to the power supply. The signal receiver is used to receive the LoRa signal transmitted by the signal transmitter, which includes the static and active signals of the regional environment and all personnel. The signal receiver includes a software radio platform electrically connected to the power supply and two receiving antennas, receiving antenna I and receiving antenna II, which both receive LoRa signals. After the signal transmitter and signal receiver are fixedly deployed, start the signal transmitter and signal receiver;

[0007] Step 2, establish the mapping relationship between the number of personnel and the amplitude, control the LoRa node to transmit signals in a multi-frequency hopping cycle within the working frequency band, receive the LoRa signal transmitted by the directional antenna of the signal transmitter through the two receiving antennas of the signal receiver, generate a new LoRa signal using the signal quotient, and collect the received signal data at multiple frequencies. The obtained LoRa receiving signal matrix is ​​expressed as:

[0008]

[0009] Wherein: each row of the matrix represents the signal data of receiving antenna I, receiving antenna II and signal quotient; each column of the matrix represents the signal data at different frequencies; f represents the number of frequencies;

[0010] Step 3, first pre-process the collected LoRa signal, divide the collected LoRa signal by using the downsampling method, extract the amplitude information of the original signal, and rearrange and combine the divided amplitude signals for standardization, then label the pre-processed data, divide it into training set and test set, and perform training;

[0011] Step 4, extract features from the data, build a lightweight coal mine underground limited area personnel counting model based on domain adversarial, and train and test it to obtain the specific number of people in the current area.

[0012] Furthermore, in Step 2, the signal quotient operation formula is as follows:

[0013]

[0014] Where: Q1 is the signal received by receiving antenna I; Q2 is the signal received by receiving antenna II; Q r It is the quotient operation of the signals received by the two receiving antennas.

[0015] Furthermore, the specific steps of Step 3 are as follows:

[0016] Step 3.1, set the sampling interval to S, and after determining the number of sampling times, downsample and divide the collected LoRa data;

[0017] Step 3.2, extract amplitude information from the downsampled original data, and arrange the divided S-segment amplitude data into an S×C matrix;

[0018] Step 3.3, calculate the average value of each row of all S×C matrices to obtain a (3×S)×f Signal matrix Data, calculate the mean and standard deviation of the amplitude measurement values ​​of each row, and perform normalization operations, as shown below:

[0019]

[0020] Where: a is the normalized amplitude data; is the original amplitude measurement value; μ is the mean of the original amplitude measurement value; σ is the standard deviation of the original amplitude measurement value;

[0021] Step 3.4, use the data collected for the first time in the scene as the source domain data set, and the data collected in the next time period as the target domain data set. For the data collected at different times, divide each data segment into a training set and a test set. During the training process, input the labeled source domain data and the unlabeled target domain data for training.

[0022] Furthermore, when the collected LoRa data is downsampled and divided in Step 3.1, the following processing is performed:

[0023]

[0024] Where: C is the number of downsampling operations each time; L is the length of the original LoRa received signal; Represents a round-down operation;

[0025] Starting from the first data of the original signal, sampling is performed every S data. When the sampling is completed, start from the second data of the original signal and repeat the previous operation until the sampling starting from the S-1th data of the original signal is completed.

[0026] Furthermore, in Step 3.4, each segment of data is divided into training set and test set in a ratio of 9:1.

[0027] Furthermore, when extracting features from data in Step 4,

[0028] The feature extraction module for extracting deep features from the original data of the source domain and the target domain includes 1 1D convolution block and 2 1D inverted residual blocks. The input size of the feature extraction module is 3×S×f. The convolution block includes a convolution layer, a BN layer, and a ReLU layer. The kernel size of the convolution block is 2 and the step size is 1. ReLU6 is used as the activation function, and its function definition is as follows:

[0029] ReLU6=min(max(0,x),6)

[0030] Where: x is the output vector of the previous layer or the original input vector;

[0031] The first layer of the inverted residual block is a 1×1 convolution, which maps the previously extracted features to a high-dimensional space. Through the second layer of depth convolution, a single convolution filter is applied to each input channel to perform lightweight filtering, followed by a third layer of 1×1 point convolution, which is responsible for creating new features by calculating the linear combination of the input channels. The mathematical expression is as follows:

[0032] Y=f(f(X*W1)*W2)*W3

[0033] Where: f(.) is the activation function; X is the input feature map; W1 is the weight matrix of 1×1 convolution; W2 is the weight matrix of depth-wise separable convolution; W3 is the weight matrix of 1×1 convolution; Y represents the final output feature map.

[0034] Furthermore, each convolutional layer is followed by a BN layer.

[0035] Furthermore, when constructing a lightweight coal mine underground limited area personnel counting model based on domain adversarial, the domain adversarial method is adopted to solve the cross-time domain problem. Specifically:

[0036] A domain discriminator is added after the feature extraction module and connected to the feature extraction module through a gradient reversal layer. For labeled data from the source domain, the network learns how to accurately predict categories by minimizing the loss of the category predictor. The domain discriminator processes all data from the source domain and the target domain. The domain discriminator first determines whether the incoming information is from the source domain or the target domain and calculates the corresponding loss. Then the gradient reversal layer passes the gradient of the reversed domain discriminator to the feature extraction module, so that the features learned by the feature extractor make it impossible for the domain discriminator to correctly determine the domain source of the data.

[0037] Finally, when the domain discriminator can no longer distinguish between source domain samples and target domain samples, the feature extraction module completes the task; then the domain data features obtained through adversarial training are used to predict the number of people through the category predictor.

[0038] Furthermore, when the domain data features obtained by the adversarial method are used to predict the number of people through the category predictor, the softmax function is used to convert the prediction results into a probability distribution P:

[0039] P = softmax(W4Y+b)

[0040] Where: W4 is the classification weight matrix; b is the bias vector;

[0041] Final output number It is determined by selecting the category with the highest probability. The mathematical expression is as follows:

[0042]

[0043] Where: i represents the number of people in the area.

[0044] Compared with the prior art, the present method for counting people in the restricted area of ​​underground mine based on the LoRa adaptive model only needs to deploy the signal transmitter and signal receiver of the counting system in the restricted area of ​​underground coal mine. The signal receiver is used to receive the LoRa signal emitted by the signal transmitter, which contains the static and active signals of the regional environment and all personnel. This deployment method is simple, and the mine workers do not need to wear any equipment. The signal receiver uses two antennas to receive signals and combines the signal quotient and multi-frequency hopping, which can expand the signal dimension and enrich the signal data, and can solve the problem of low accuracy in detecting the number of people due to the single LoRa receiving signal mode and the inability to provide rich amplitude information of personnel activities. The received original LoRa signal is downsampled, the amplitude data is extracted, and normalized after rearrangement, which can facilitate model learning and increase the number of training data samples. The constructed lightweight neural network model based on domain adversarial can solve the time instability of wireless signals and the real-time detection. The model can be trained and tested to achieve accurate counting of the number of people, and can achieve relatively ideal accuracy and reliability for counting people in the restricted area of ​​underground mine. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the perception scene of the present invention;

[0046] Figure 2 Experimental flow chart of the present invention;

[0047] Figure 3 The architecture diagram of the lightweight coal mine underground limited-personnel area personnel counting model based on domain adversarial technology of the present invention.

[0048] In the figure: 1. Receiver, 2. Transmitter, 3. Underground tunnel, 4. Operator DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings.

[0050] like Figure 2 As shown, the personnel counting method in the underground restricted area based on the LoRa adaptive model specifically includes the following steps:

[0051] Step 1, such as Figure 1 As shown, the signal transmitter and signal receiver of the counting system are fixedly deployed in the restricted area of ​​the underground tunnel. The signal transmitter and signal receiver can be deployed at any position that does not affect the tunnel activities, such as hanging and fixing on the top wall of the tunnel. The signal transmitter is used to transmit linear frequency modulation LoRa signals. The signal transmitter includes a LoRa node electrically connected to the power supply and a 6dbi directional antenna. The signal receiver is used to receive the LoRa signal emitted by the signal transmitter, which contains the static and active signals of the regional environment and all personnel. The signal receiver includes a USRP X310 software radio platform electrically connected to the power supply and two receiving antennas, receiving antenna I and receiving antenna II, which both receive LoRa signals. After the signal transmitter and signal receiver are fixedly deployed, start the signal transmitter and signal receiver.

[0052] Step 2, control the LoRa node to transmit signals in a multi-frequency hopping cycle within the working frequency band, receive the LoRa signal transmitted by the directional antenna of the signal transmitter through the two receiving antennas of the signal receiver, use the signal quotient to generate a new LoRa signal, and collect the received signal data at multiple frequencies to expand the characteristic dimension of the number of people, enrich the received signal information, and improve the detection accuracy.

[0053] The main working frequency band of LoRa nodes is 915MHz, and the receiving signal mode is single. It is difficult to accurately express the number of people in the scene using only one-dimensional amplitude. In order to expand the signal dimension and establish a mapping relationship between the number of people and the amplitude, the LoRa node transmits signals in a multi-frequency hopping manner within the working frequency band. The carrier frequencies of the transmitted signals are 902MHz, 910MHz, 915MHz, 920MHz and 928MHz respectively. The LoRa node transmits signals cyclically in this order. The channel bandwidth is 125KHz, the spreading factor is 12, and the sampling rate is 500KHz.

[0054] The signal quotient operation formula is as follows:

[0055]

[0056] Where: Q1 is the signal received by receiving antenna I; Q2 is the signal received by receiving antenna II; Qr It is the quotient operation of the signals received by the two receiving antennas.

[0057] By using different antennas, signal quotients, multi-frequency hopping and other combinations to receive signal data, the characteristic dimension of the number of people is expanded to improve the detection accuracy. The obtained LoRa receiving signal matrix is ​​expressed as:

[0058]

[0059] Among them: each row of the matrix represents the signal data of receiving antenna I, receiving antenna II and signal quotient respectively; each column of the matrix represents the signal data at frequencies of 902MHz, 910MHz, 915MHz, 920MHz and 928MHz respectively; f represents the number of frequencies in the multi-frequency hopping method.

[0060] Step 3, in order to facilitate the subsequent processing of the machine learning model, the collected LoRa signal is first preprocessed, and the collected LoRa signal is divided using the downsampling method to extract the amplitude information of the original signal, and the divided amplitude signal is rearranged and combined for standardization, and then the preprocessed data is labeled and divided into training set and test set, and training is performed. The details are as follows:

[0061] Step 3.1, downsample the collected LoRa data and set the sampling interval to S. To ensure that the length of all sampled data of each data is equal, perform the following processing:

[0062]

[0063] Where: C is the number of downsampling operations each time; L is the length of the original LoRa received signal; Represents a floor operation.

[0064] After determining the number of sampling times, downsampling is performed. Starting from the first data of the original signal, sampling is performed every S data. When sampling is completed, the second data of the original signal is started again, and the previous operation is repeated until the sampling starting from the S-1th data of the original signal is completed. Based on this, the original data can be divided into S segments to increase the number of samples of the training data.

[0065] Step 3.2, extract the amplitude information from the downsampled original data, and arrange the divided S-segment amplitude data into an S×C matrix.

[0066] Step 3.3, calculate the average value of each row of all S×C matrices to obtain a (3×S)×f Signal matrix data.

[0067] Calculate the mean and standard deviation of the amplitude measurements for each row and perform normalization, as shown below:

[0068]

[0069] Where: a is the normalized amplitude data; is the original amplitude measurement value; μ is the mean of the original amplitude measurement value; σ is the standard deviation of the original amplitude measurement value.

[0070] Step 3.4, label the data and divide it into training set and test set. The data collected for the first time in the scene is used as the source domain data set, and the data collected in the next time period is used as the target domain data set. For the data collected at different times, each data segment is divided into training set and test set in a ratio of 9:1. During the training process, input the labeled source domain data and unlabeled target domain data for training.

[0071] Step 4, extract features from the data, build a lightweight coal mine underground limited area personnel counting model based on domain adversarial, train and test it, and obtain the specific number of personnel in the current area. The main process is as follows:

[0072] Step 4.1, extract features from the data.

[0073] In the feature extraction stage, Figure 3 As shown in Figure 1, the feature extraction module for extracting deep features from the raw data of the source and target domains includes 1 1D convolution block and 2 1D inverted residual blocks. The input size of the feature extraction module is 3×S×f, and the convolution block includes a convolution layer, a BN layer, and a ReLU layer. The kernel size of the convolution block is 2 and the stride is 1 to expand the number of feature channels.

[0074] ReLU6 is used as the activation function to improve the numerical resolution, and its function is defined as follows:

[0075] ReLU6=min(max(0,x),6)

[0076] Where: x is the output vector of the previous layer or the original input vector.

[0077] The first layer of the inverted residual block is a 1×1 convolution, which maps the previously extracted features to a high-dimensional space. Through the second layer of depth convolution, a single convolution filter is applied to each input channel to perform lightweight filtering. This is followed by a third layer of 1×1 point convolution, which is responsible for creating new features by calculating the linear combination of the input channels and expanding the dimension of the model. The mathematical expression is as follows:

[0078] Y=f(f(X*W1)*W2)*W3

[0079] Where: f(.) is the activation function; X is the input feature map; W1 is the weight matrix of 1×1 convolution; W2 is the weight matrix of depth-wise separable convolution; W3 is the weight matrix of 1×1 convolution; Y represents the final output feature map.

[0080] The three-layer operation of the inverted residual block is used to replace the full convolution operator. This structure maps the features to a high-dimensional space for fine-grained feature extraction, which can significantly reduce the number of parameters of the convolution kernel and better distinguish noise signals to reduce noise interference, thereby improving detection accuracy.

[0081] Through the above structural design, the feature extraction module can effectively extract the deep features of the source domain and the target domain, and provide rich and stable feature input for the subsequent domain adversarial model training. It not only improves the effect of feature extraction, but also achieves a good balance between computational efficiency and model complexity.

[0082] To reduce computational overhead, each convolutional layer is followed by a batch normalization (BN) layer to keep the input of each layer of the neural network with the same distribution.

[0083] Step 4.2: After the model training is completed, use the test set to test and check the detection accuracy

[0084] In the classification stage, because there are certain differences between the source domain and the target domain data, the effect of directly predicting the target domain through the source domain data is not ideal. In order to obtain accurate and robust detection performance, the domain confrontation method is adopted to solve the cross-time domain problem. A domain discriminator is added after the feature extraction module and connected to the feature extraction module through a gradient reversal layer. During the training process, the entire model contains two main tasks: category prediction and domain identification.

[0085] For labeled data from the source domain, the network learns how to accurately predict the class by minimizing the loss of the class predictor. This step ensures that the model performs well on the task on the source domain.

[0086] The domain discriminator processes all data from the source domain and the target domain. Its task is to determine whether the incoming information is from the source domain or the target domain and calculate the corresponding loss. The training goal of the domain discriminator is to correctly classify the input information into the source domain or target domain category as much as possible.

[0087] However, due to the existence of the gradient reversal layer, the goal of the feature extraction module is exactly the opposite of the domain discriminator. The gradient reversal layer passes the gradient of the reversed domain discriminator to the feature extraction module, forcing the features learned by the feature extraction module to prevent the domain discriminator from correctly determining the domain origin of the data. This mechanism forces the features extracted by the feature extraction module to remain consistent between the source domain and the target domain.

[0088] Finally, when the domain discriminator can no longer distinguish between source domain samples and target domain samples, the feature extraction module completes the task. At this point, the data from the source domain and the target domain have been mixed together in the feature space and cannot be distinguished. This indicates that the feature extraction module has learned a common feature representation that eliminates the distribution difference between the source domain and the target domain, which is beneficial to the performance of the model on the target domain. In this way, through adversarial training between the feature extraction module and the domain discriminator, the model improves its generalization ability on the target domain while maintaining the performance of the source domain task.

[0089] Then, the domain data features obtained by the adversarial process are used to predict the number of people through the category predictor to maximize the detection accuracy. In this process, the softmax function is used to convert the prediction results into a probability distribution P:

[0090] P = softmax(W4Y+b)

[0091] Where: W4 is the classification weight matrix; b is the bias vector.

[0092] The softmax function converts the category into a probability distribution and finally outputs the number of people It is determined by selecting the category with the highest probability. The mathematical expression is as follows:

[0093]

[0094] Where: i represents the number of people in the area.

[0095] This underground restricted area personnel counting method based on the LoRa adaptive model only needs to deploy the signal transmitter and signal receiver of the counting system in the restricted area of ​​the coal mine. The deployment method is simple and the mine workers do not need to wear any equipment. The two antennas of the signal receiver receive signals and combine the signal quotient and multi-frequency hopping to solve the problem of low accuracy in personnel number detection due to the single LoRa receiving signal mode and the inability to provide rich amplitude information of personnel activities. The constructed lightweight neural network model based on domain adversarial can solve the time instability of wireless signals and the real-time detection, and can achieve ideal accuracy and reliability in counting personnel in the restricted area of ​​underground personnel.

Claims

1. A method for counting people in a restricted area underground based on a LoRa adaptive model, characterized in that: The specific steps include: Step 1, the signal transmitter and signal receiver of the counting system are fixedly deployed in the restricted area of ​​the underground tunnel. The signal transmitter is used to transmit linear frequency modulation LoRa signals. The signal transmitter includes a LoRa node and a directional antenna electrically connected to the power supply. The signal receiver is used to receive the LoRa signal transmitted by the signal transmitter, which includes the static and active signals of the regional environment and all personnel. The signal receiver includes a software radio platform electrically connected to the power supply and two receiving antennas, receiving antenna I and receiving antenna II, which both receive LoRa signals. After the signal transmitter and signal receiver are fixedly deployed, start the signal transmitter and signal receiver; Step 2, establish the mapping relationship between the number of personnel and the amplitude, control the LoRa node to transmit signals in a multi-frequency hopping cycle within the working frequency band, receive the LoRa signal transmitted by the directional antenna of the signal transmitter through the two receiving antennas of the signal receiver, generate a new LoRa signal using the signal quotient, and collect the received signal data at multiple frequencies. The obtained LoRa receiving signal matrix is ​​expressed as: Wherein: each row of the matrix represents the signal data of receiving antenna I, receiving antenna II and signal quotient; each column of the matrix represents the signal data at different frequencies; f represents the number of frequencies; Step 3, first pre-process the collected LoRa signal, divide the collected LoRa signal by using the downsampling method, extract the amplitude information of the original signal, and rearrange and combine the divided amplitude signals for standardization, then label the pre-processed data, divide it into training set and test set, and perform training; Step 4: Extract features from the data and build a lightweight coal mine underground limited area personnel counting model based on domain adversarial, train and test it to obtain the specific number of personnel in the current area; When extracting features from data, the feature extraction module that extracts deep features from the original data of the source domain and the target domain includes 1 1D convolution block and 2 1D inverted residual blocks. The input size of the feature extraction module is 3×S×f. The convolution block includes a convolution layer, a BN layer, and a ReLU layer. The kernel size of the convolution block is 2 and the step size is 1. ReLU6 is used as the activation function, and its function definition is as follows: ReLU6=min(max(0,x),6) Where: x is the output vector of the previous layer or the original input vector; The first layer of the inverted residual block is a 1×1 convolution, which maps the previously extracted features to a high-dimensional space. Through the second layer of depth convolution, a single convolution filter is applied to each input channel to perform lightweight filtering, followed by a third layer of 1×1 point convolution, which is responsible for creating new features by calculating the linear combination of the input channels. The mathematical expression is as follows: Y=f(f(X*W1)*W2)*W3 Where: f(.) is the activation function; X is the input feature map; W1 is the weight matrix of 1×1 convolution; W2 is the weight matrix of depth-separable convolution; W3 is the weight matrix of 1×1 convolution; Y represents the final output feature map; Each convolutional layer is followed by a BN layer; When constructing a lightweight coal mine underground limited area personnel counting model based on domain adversarial, the domain adversarial method is adopted to solve the cross-time domain problem. Specifically: A domain discriminator is added after the feature extraction module and connected to the feature extraction module through a gradient reversal layer. For labeled data from the source domain, the network learns how to accurately predict categories by minimizing the loss of the category predictor. The domain discriminator processes all data from the source domain and the target domain. The domain discriminator first determines whether the incoming information is from the source domain or the target domain and calculates the corresponding loss. Then the gradient reversal layer passes the gradient of the reversed domain discriminator to the feature extraction module, so that the features learned by the feature extractor make it impossible for the domain discriminator to correctly determine the domain source of the data. Finally, when the domain discriminator can no longer distinguish between source domain samples and target domain samples, the feature extraction module completes the task; then the domain data features obtained through adversarial training are used to predict the number of people through the category predictor.

2. The method for counting people in a restricted area of ​​a well based on a LoRa adaptive model according to claim 1, characterized in that: In Step 2, the signal quotient calculation formula is as follows: Where: Q1 is the signal received by receiving antenna I; Q2 is the signal received by receiving antenna II; Q r It is the quotient operation of the signals received by the two receiving antennas.

3. The method for counting people in a restricted area of ​​a well based on a LoRa adaptive model according to claim 1, characterized in that: Step 3 The specific steps are as follows: Step 3.1, set the sampling interval to S, and after determining the number of sampling times, downsample and divide the collected LoRa data; Step 3.2, extract amplitude information from the downsampled original data, and arrange the divided S-segment amplitude data into an S×C matrix; Step 3.3, calculate the average value of each row of all S×C matrices to obtain a (3×S)×f Signal matrix Data, calculate the mean and standard deviation of the amplitude measurement values ​​of each row, and perform normalization operations, as shown below: Where: a is the normalized amplitude data; is the original amplitude measurement value; μ is the mean of the original amplitude measurement value; σ is the standard deviation of the original amplitude measurement value; Step 3.4, use the data collected for the first time in the scene as the source domain data set, and the data collected in the next time period as the target domain data set. For the data collected at different times, divide each data segment into a training set and a test set. During the training process, input the labeled source domain data and the unlabeled target domain data for training.

4. The method for counting people in a restricted area of ​​a well based on a LoRa adaptive model according to claim 3, characterized in that: When downsampling the collected LoRa data in Step 3.1, the following processing is performed: Where: C is the number of downsampling operations each time; L is the length of the original LoRa received signal; Represents a round-down operation; Starting from the first data of the original signal, sampling is performed every S data. When the sampling is completed, start from the second data of the original signal and repeat the previous operation until the sampling starting from the S-1th data of the original signal is completed.

5. The method for counting people in a restricted area of ​​a well based on a LoRa adaptive model according to claim 3, characterized in that: In Step 3.4, each segment of data is divided into training set and test set in a ratio of 9:

1.

6. The method for counting people in a restricted area of ​​a well based on a LoRa adaptive model according to claim 1, characterized in that: When the domain data features obtained by the adversarial method are used to predict the number of people through the category predictor, the softmax function is used to convert the prediction results into a probability distribution P: P = softmax(W4Y+b) Where: W4 is the classification weight matrix; b is the bias vector; Final output number It is determined by selecting the category with the highest probability. The mathematical expression is as follows: Where: i represents the number of people in the area.

Images