Human body posture detection method and system based on 1DCNN-GRU model
By adopting the human posture detection method based on the 1DCNN-GRU model and the UWB tag card embedded system in the coal mine, the problem of insufficient real-time and accuracy of attitude detection in the coal mine is solved, and efficient and low-cost attitude monitoring is achieved, improving the Coal mine safety and production efficiency.
Patent Information
- Application Number
- CN202411477127.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-22
- Publication Date
- 2024-12-31
AI Technical Summary
In the coal mining industry, it is difficult for the existing technology to achieve real-time and accurate detection of miners' human postures, and the equipment is complex or expensive, making it difficult to achieve lightweight deployment and portable applications, resulting in challenges in security monitoring.
Using the human posture detection method based on the 1DCNN-GRU model, combined with the embedded system in the UWB tag card, the attitude detection is realized through six-axis sensor data acquisition, preprocessing and fusion model prediction, and the abnormal state is sent to the base station through the UWB protocol to reduce the Power consumption and cost.
It improves the accuracy and real-time performance of posture detection, reduces the detection cost, improves the safety and efficiency of coal mine production, is suitable for complex environments, and provides safety guarantees for high-risk professionals.
Smart Images
Figure CN119226931A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of coal mine safety technology, and in particular to a human posture detection method and system based on a 1DCNN-GRU model. Background Art
[0002] In the coal mining industry, ensuring the safety of miners has always been a top priority. However, due to the particularity of the underground working environment in coal mines, such as insufficient light, dust, small space and various potential dangerous factors, the safety monitoring of miners has become an extremely challenging task. At present, although some safety monitoring measures have been applied to coal mine production, there are still many deficiencies in the real-time and accurate detection of miners' body postures, and there are problems such as complex equipment or high costs, making it difficult to achieve lightweight deployment and portable applications. Summary of the invention
[0003] Based on the technical problems existing in the background technology, the present invention proposes a human posture detection method and system based on the 1DCNN-GRU model, which significantly improves the accuracy and real-time performance of miners' human posture detection, reduces the detection cost, and improves the safety and efficiency of coal mine production.
[0004] The human body posture detection method based on the 1DCNN-GRU model proposed in the present invention detects the human body posture by the embedded system in the UWB tag card. The detection method specifically includes the following steps:
[0005] Step 1: The person wears a UWB tag card, and the embedded system obtains and caches the six-axis sensor data of human activities;
[0006] Step 2: regularly obtain the ADC data of the original six-axis sensor from the cache, convert the binary ADC data into a floating-point two-dimensional array, and perform data preprocessing on the floating-point two-dimensional array;
[0007] Step 3: Input the preprocessed floating-point two-dimensional array into the fusion model to output the action type of the human body posture. When the action type is determined to be abnormal, the UWB tag card sends the abnormal posture status to the base station through the UWB protocol, and the base station transmits it to the well monitoring platform.
[0008] Furthermore, in step 2, the binary ADC data is converted into a floating point two-dimensional array, specifically:
[0009] The binary ADC data is converted into a floating-point two-dimensional array of 500 rows and 6 columns. The 500 rows represent the time series data of 500 acquisition points, and the 6 columns represent the x, y, and z three-axis acceleration data and x, y, and z three-axis angular velocity data of the six-axis sensor.
[0010] Furthermore, in step 2, data preprocessing is performed on the floating-point two-dimensional array, specifically: Butterworth filtering and normalization are performed on the floating-point two-dimensional array in sequence to obtain a preprocessed floating-point two-dimensional array.
[0011] Furthermore, in step three, the fusion model is a one-dimensional convolutional neural network and a gated recurrent unit series fusion model.
[0012] Furthermore, the fusion model runs on an embedded microcontroller, where the fusion model is first quantized and pruned by the Tensflow-Lite method, and then gated by the IMU’s mode switching scheme;
[0013] The mode switching scheme of the IMU is specifically as follows:
[0014] There are two IMU operation modes: low power mode and normal attitude detection mode. Threshold A and threshold B are set. Threshold A is smaller than threshold B.
[0015] In low power consumption mode, it is determined whether the total acceleration of the acquired three-axis acceleration data is greater than the threshold value thresholdA. If it is greater than the threshold value thresholdA, the IMU operation mode is switched to the operation mode during normal posture detection. If it is less than or equal to the threshold value thresholdA, the IMU operation mode remains in the low power consumption mode.
[0016] Determine whether the combined acceleration of the three-axis acceleration data is greater than the threshold thresholdB. If it is greater than the threshold thresholdB, the IMU operation mode remains in the normal attitude detection operation mode. If it is less than or equal to the threshold thresholdB, the IMU operation mode is switched to the low power consumption mode.
[0017] The preprocessed floating-point two-dimensional array corresponding to the value exceeding the threshold thresholdB is passed through the fusion model to output the action type of the human body posture.
[0018] Furthermore, the low power consumption mode is specifically as follows: the accelerometer in the six-axis sensor has a sampling frequency of 50 Hz and the angular velocity sensor is in standby mode;
[0019] The specific operating mode during normal posture detection is: the acceleration sensor in the six-axis sensor has a sampling frequency of 200HZ, and the angular velocity sensor has a sampling frequency of 200HZ.
[0020] A human posture detection method based on a 1DCNN-GRU model, wherein an embedded system in a UWB tag detects human posture, and the embedded system includes a data acquisition module, a data processing module, a model prediction module, and a data transmission module;
[0021] The data acquisition module is used to obtain and cache the six-axis sensor data of human activities;
[0022] The data processing module is used to obtain the original six-axis sensor data from the cache at a regular interval, parse the binary ADC data, convert it into a floating-point two-dimensional array, and perform data preprocessing on the floating-point two-dimensional array;
[0023] The model prediction module is used to input the preprocessed floating-point two-dimensional array into the fusion model to output the action type of the human body posture;
[0024] The data transmission module is used to send the abnormal posture status to the base station through the UWB protocol when the action type is determined to be abnormal posture, and the base station transmits it to the well monitoring platform.
[0025] The advantages of the human posture detection method and system based on the 1DCNN-GRU model provided by the present invention are: the integration of posture detection technology and UWB identification card is realized, the portability and practicality of the system are improved, and the accuracy and robustness of posture detection are improved by integrating the 1DCNN and GRU fusion model, thereby ensuring the reliability of the detection results; in addition, an ultra-low power consumption design is adopted to reduce the energy consumption of the system, extend the use time, and reduce the use cost; it is particularly suitable for complex environments such as coal mines, and provides a strong guarantee for the safety of high-risk occupational personnel such as miners. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is a schematic diagram of the process of the present invention;
[0027] Figure 2 A block diagram of data transmission for the embedded system in the UWB tag;
[0028] Figure 3 This is an example of the IMU data format after preprocessing the floating-point two-dimensional array;
[0029] Figure 4 This is the flow chart for switching the IMU working mode. DETAILED DESCRIPTION
[0030] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.
[0031] like Figures 1 to 4As shown, the human body posture detection method based on the 1DCNN-GRU model proposed in the present invention, the embedded system in the UWB tag card detects the human body posture, and the detection method specifically includes the following steps:
[0032] Step 1: The person wears a UWB tag card, and the data acquisition module embedded in the system obtains the six-axis sensor data of human activities and caches it;
[0033] When people perform daily activities and simulate abnormal actions such as falling, the data acquisition module will collect the six-axis posture data of the people. The sampling frequency of the IMU acceleration sensor and angular velocity sensor are both set to 200HZ, that is, 200 samples per second, and the collected data is stored in the cache.
[0034] Step 2: regularly obtain the ADC data of the original six-axis sensor from the cache, convert the binary ADC data into a floating-point two-dimensional array, and perform data preprocessing on the floating-point two-dimensional array;
[0035] The data processing module obtains the original six-axis sensor ADC data from the cache in real time every 50ms. After reading, the cache is cleared. 50ms is about 10 acquisition points of data. After reading the data, pre-processing operations such as data analysis, data conversion, data filtering and data normalization are performed.
[0036] Data parsing: after reading the data, the program parses the binary data and converts it into integer. The converted data is ADC data.
[0037] Data conversion: Binary ADC data is converted into a floating-point two-dimensional array of 500 rows and 6 columns. The 500 rows represent the time series data of 500 acquisition points, which is about 2.5 seconds of acquisition data. The 6 columns represent the x, y, and z three-axis acceleration data and the x, y, and z three-axis angular velocity data respectively.
[0038] Data filtering, Butterworth filtering of floating-point two-dimensional arrays to solve the sensor data noise problem. Acceleration and angular velocity sensors are easily interfered by various noises during the measurement process, such as electromagnetic noise, mechanical vibration noise, etc. The stopband attenuation characteristics of the Butterworth filter are very steep, which can effectively filter out these noises above the required frequency range and improve the signal-to-noise ratio of the data;
[0039] Data normalization is to normalize the filtered floating-point two-dimensional array to a range of 1 to -1. Normalization is to increase the generalization ability of the model and improve the convergence speed during model training, eliminate the influence of extreme values or singular samples on model training, and make the fusion model more stable. Figure 3 The data structure shown.
[0040] Step 3: Input the preprocessed floating-point two-dimensional array into the fusion model to output the action type of the human body posture. When the action type is determined to be abnormal, the UWB tag card sends the abnormal posture status to the base station through the UWB protocol, and the base station transmits it to the well monitoring platform.
[0041] The fusion model is a one-dimensional convolutional neural network 1DCNN and a gated recurrent unit GRU series fusion model, which is obtained by training on the host computer. 1DCNN can effectively capture local relationships in the input sequence, has certain robustness to small changes in the input data, and has good generalization ability; GRU is a variant of LSTM (long short-term memory network), but is simpler in structure, more efficient in calculation, and requires fewer resources for calculation. GRU performs well in processing long-term dependencies through its gating mechanism and memory ability. The fusion model of this embodiment selects 1DCNN and GRU to be fused in series, and uses the characteristics of 1DCNN that can effectively extract complex spatial features to first learn the spatial features in the human posture data, and then uses GRU to effectively learn the characteristics of long-term dependencies in time series data to learn the temporal features in the human posture data. Human posture action data is a time series data. This embodiment fuses the two algorithm models, and through the joint learning of time-space features, the human posture prediction results of a single model of 1DCNN or GRU are more accurate.
[0042] The fusion model is a one-dimensional convolutional neural network and a gated recurrent unit series fusion model. The fusion model is designed as follows:
[0043] 1) Input data:
[0044] The dimension of the input data is [500,6], 500 time steps, and each time step contains 6 features (3-axis acceleration + 3-axis angular velocity);
[0045] 2) One-dimensional convolutional neural network 1DCNN module (spatial feature extraction):
[0046] Convolution layer: 9 convolution kernels; kernel size 3 (processing 3 time steps each time to capture small-scale temporal local features); step size 1;
[0047] Output dimension: Assume that after one layer of convolution, the time dimension remains unchanged, that is, the output dimension is [500,9];
[0048] Pooling layer: Use maximum pooling to reduce the time dimension by half, and the output dimension is [250,9].
[0049] Activation function: Use ReLU activation function;
[0050] 3) Gated recurrent unit GRU module (time-dependent modeling):
[0051] Number of GRU layers: 1 GRU layer;
[0052] Number of hidden units: 64;
[0053] The input dimension is [250,9] and the output dimension is [250,64];
[0054] GRU learns time-dependent features in data through its gating mechanism;
[0055] 4) Fully connected layer and output layer:
[0056] Fully connected layer: Flatten the GRU output into a one-dimensional vector and further process it through a fully connected layer. The size of the fully connected layer is 64;
[0057] Activation function: ReLU activation;
[0058] Output layer: Use softmax activation function for multi-classification. Assuming there are C categories of postures, the output dimension is [C];
[0059] 5) Model training process: Use cross entropy loss function and Adam optimizer for training;
[0060] 6) Loss function: standard cross entropy loss function.
[0061] The fusion model runs on an embedded microcontroller. This embodiment uses the Tensflow-Lite method to quantize and trim the fusion model to reduce the time and power consumption required for model prediction. And the power consumption required for posture detection is further reduced through the IMU mode switching scheme. The IMU mode switching scheme determines whether the combined acceleration of the three-axis acceleration sensor exceeds the threshold to switch the working mode. There are two IMU operating modes, low power consumption mode (acceleration sensor 50HZ sampling frequency, angular velocity sensor standby) and normal posture detection operation mode (acceleration sensor 200HZ sampling frequency, angular velocity sensor 200HZ sampling frequency). The flowchart of IMU mode switching is shown below. Figure 4As shown. Set two thresholds thresholdA=1.5 and thresholdB=4.5, and thresholdA is less than thresholdB. In low power mode, it is determined whether the combined acceleration of the three-axis acceleration sensor is greater than thresholdA. If it is greater, it switches to the normal posture detection operation mode, and if it is less than or equal to, it maintains low power mode operation; the normal posture detection operation mode determines whether the combined acceleration of the three-axis acceleration sensor is greater than thresholdB. If it is greater than or equal to, the operation mode remains unchanged, and if the threshold is less than thresholdB for more than 10 consecutive times, it switches to low power mode operation. Only the posture data exceeding thresholdB will be predicted by the model. According to the power consumption meter, the model prediction module uses the IMU mode switching scheme to reduce power consumption by 50% compared with not using this scheme.
[0062] After obtaining the action type of the human body posture through the fusion model, it detects whether the action type is a fall. If the action type is a fall, it is judged as an abnormal posture. The UWB identification card corresponding to the abnormal posture emits an audible and visual alarm, and the detection result is sent to the base station. The abnormal posture is transmitted to the well monitoring platform through the base station, and the well staff performs the next step of processing.
[0063] Through steps one to three, this embodiment uses the IMU sensor to acquire data, the one-dimensional convolutional neural network 1DCNN and the GRU (gated recurrent unit) fusion model for prediction and ultra-low power consumption quantitative clipping fusion model to adapt to the resource-constrained operating environment, and integrates this technology into the UWB personnel identification card embedded device to achieve real-time and accurate detection of the miner's body posture. The detection equipment of this embodiment has the characteristics of good portability and strong durability, can adapt to the complex environment of coal mines, and provide strong protection for the safe production of miners. Through the implementation of this embodiment, the fusion model can be run on the miner's UWB identification card with extremely low power consumption, which significantly improves the accuracy and real-time performance of miner body posture detection, reduces detection costs, and improves the safety and efficiency of coal mine production.
[0064] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A human posture detection method based on 1DCNN-GRU model, characterized in that: The embedded system in the UWB tag detects the human body posture. The detection method specifically includes the following steps: Step 1: The person wears a UWB tag card, and the embedded system obtains and caches the six-axis sensor data of human activities; Step 2: regularly obtain the ADC data of the original six-axis sensor from the cache, convert the binary ADC data into a floating-point two-dimensional array, and perform data preprocessing on the floating-point two-dimensional array; Step 3: Input the preprocessed floating-point two-dimensional array into the fusion model to output the action type of the human body posture. When the action type is determined to be abnormal, the UWB tag card sends the abnormal posture status to the base station through the UWB protocol, and the base station transmits it to the well monitoring platform.
2. The human body posture detection method based on the 1DCNN-GRU model according to claim 1, characterized in that: In step 2, the binary ADC data is converted into a floating-point two-dimensional array, specifically: The binary ADC data is converted into a floating-point two-dimensional array of 500 rows and 6 columns. The 500 rows represent the time series data of 500 acquisition points, and the 6 columns represent the x, y, and z three-axis acceleration data and x, y, and z three-axis angular velocity data of the six-axis sensor.
3. The human body posture detection method based on the 1DCNN-GRU model according to claim 1, characterized in that: In step 2, data preprocessing is performed on the floating-point two-dimensional array, specifically: Butterworth filtering and normalization are performed on the floating-point two-dimensional array in sequence to obtain a preprocessed floating-point two-dimensional array.
4. The human body posture detection method based on the 1DCNN-GRU model according to claim 1, characterized in that: In step three, the fusion model is a one-dimensional convolutional neural network and a gated recurrent unit series fusion model, and is trained using the cross entropy loss function and the Adam optimizer.
5. The human body posture detection method based on the 1DCNN-GRU model according to claim 2 is characterized in that: The fusion model runs on an embedded microcontroller. The fusion model is first quantized and pruned using the Tensflow-Lite method, and then gated using the IMU mode switching scheme. The mode switching scheme of the IMU is specifically as follows: There are two IMU operation modes: low power mode and normal attitude detection mode. Threshold A and threshold B are set. Threshold A is smaller than threshold B. Determine whether the total acceleration of the acquired three-axis acceleration data is greater than a threshold value thresholdA. If it is greater than the threshold value thresholdA, determine whether the current mode is a low power consumption mode. If it is less than or equal to the threshold value thresholdA, return to step 2. If it is low power mode, the IMU operation mode is switched to the normal attitude detection operation mode. If it is not low power mode, it enters the process of judging whether the combined acceleration of the three-axis acceleration data is less than the threshold value thresholdB. If it is less than the threshold value thresholdB, the IMU operation mode temporarily maintains the operation mode during normal posture detection, and if it is less than the threshold value thresholdB for more than 10 times in a row, the IMU operation mode switches to low power consumption mode; If it is greater than or equal to the threshold thresholdB, the IMU operation mode remains in the normal attitude detection operation mode; The preprocessed floating-point two-dimensional array corresponding to the value exceeding the threshold thresholdB is passed through the fusion model to output the action type of the human body posture.
6. The human posture detection system based on the 1DCNN-GRU model according to claim 5, characterized in that: The low power consumption mode is as follows: the accelerometer in the six-axis sensor has a sampling frequency of 50 Hz, and the angular velocity sensor is in standby mode; The specific operating mode during normal posture detection is: the acceleration sensor in the six-axis sensor has a sampling frequency of 200HZ, and the angular velocity sensor has a sampling frequency of 200HZ.
7. A human posture detection method based on 1DCNN-GRU model, characterized in that: The embedded system in the UWB tag detects human posture, and the embedded system includes a data acquisition module, a data processing module, a model prediction module and a data transmission module; The data acquisition module is used to obtain and cache the six-axis sensor data of human activities; The data processing module is used to obtain the original six-axis sensor data from the cache at a regular interval, parse the binary ADC data, convert it into a floating-point two-dimensional array, and perform data preprocessing on the floating-point two-dimensional array; The model prediction module is used to input the preprocessed floating-point two-dimensional array into the fusion model to output the action type of the human body posture; The data transmission module is used to send the abnormal posture status to the base station through the UWB protocol when the action type is determined to be abnormal posture, and the base station transmits it to the well monitoring platform.