False touch prevention high-altitude falling identification method based on deep learning

Abstract

The invention belongs to the technical field of electric power, and relates to a false touch prevention high-altitude falling identification method based on deep learning, which comprises the following steps of: acquiring kinematics data and dynamics data of human body descending actions such as squatting, bending, falling and falling, denoising the data by using a Butterworth filter, and obtaining a false touch prevention high-altitude falling identification result; according to the method, the feature information capable of rapidly identifying the falling action is extracted, the deep learning network model is constructed, and the door mechanism training of the long-short term memory neural network model is guided by using the space-time parameters, so that the identification speed and accuracy are improved, and the intelligent wearable protection system starts to work only when the falling occurs, and false triggering is avoided.

Description

Technical Field

[0001] The present invention belongs to the field of electric power technology and relates to an intelligent detection technology, in particular to an anti-accidental-touch high-altitude fall recognition method based on deep learning. Background Art

[0002] High-altitude power work is a vital part of the power industry, which includes the installation, maintenance and overhaul of power lines. This work is usually carried out at an altitude of tens or even hundreds of meters, and the working environment is complex and changeable, including dangerous factors such as strong winds, extreme weather, and high-voltage electric fields. In addition to facing the risk of falling from heights, operators must also deal with possible electric shock hazards and physical fatigue caused by long-term high-altitude work. Although traditional safety protection measures, such as safety ropes and safety belts, provide a certain degree of protection, they still have limitations in practical applications and cannot completely avoid accidents. For example, safety ropes may fail due to wear or improper operation, and safety belts may not provide adequate protection in time in certain emergency situations. Therefore, the safety issue of high-altitude power work has always been the focus of the industry, and it is necessary to continuously explore and improve more effective safety technologies and equipment.

[0003] With the advancement of technology, intelligent wearable protection systems, especially airbag suits, provide new safety guarantees for workers working at heights. These systems can monitor the workers' motion status and environmental changes in real time by integrating sensors and algorithms. Once the risk of falling is detected, the system can respond quickly and automatically deploy airbags to provide additional cushioning protection for the workers. The introduction of intelligent protection systems not only improves the safety of high-altitude operations, but also reduces the occurrence of accidents through intelligent monitoring and early warning. In addition, the intelligent protection system can also communicate with the ground monitoring center in real time. Once an emergency occurs, the rescue procedure can be started immediately to improve the rescue efficiency. These characteristics of the intelligent protection system make it play an increasingly important role in high-altitude operations.

[0004] Although the intelligent protection system has made some progress in the safety of aerial work, there are still some shortcomings in practical applications. For example, some sensor-based high-altitude fall detection systems may trigger the airbag by mistake due to mis-touch or mis-operation.

[0005] Patent CN117746572A proposes a high-altitude fall accident detection method based on a smart safety helmet. By collecting the acceleration of the smart safety helmet on the X-axis, Y-axis and Z-axis, it is determined whether the construction worker is weightless, and whether a fall has occurred. However, due to the complexity of human body movements, actions such as squatting and falling may produce acceleration changes similar to falling, causing the system to misjudge it as a fall accident, thereby erroneously triggering the airbag. This false triggering will not only reduce the reliability of the protection system, but may also cause unnecessary troubles and risks to the operators. Therefore, how to improve the accuracy of the detection system and reduce false triggering is a direction that needs further research and improvement in the current field of high-altitude work safety technology.

[0006] Through searching the published patent documents, no published patent documents identical to the present patent application were found. Summary of the invention

[0007] The purpose of the present invention is to provide a method for preventing accidental touch and high-altitude fall recognition based on deep learning to solve the problems raised in the background technology.

[0008] The present invention solves the technical problem by the following technical solutions:

[0009] A method for preventing accidental touch and high-altitude fall recognition based on deep learning, comprising the following steps:

[0010] Step 1, collecting kinematic and dynamic data of the human body's descending action through a data acquisition module;

[0011] Step 2: The kinematic and dynamic data of the human body's falling action are preprocessed by the data processing module and then input into the deep learning network model for analysis to ensure that the smart wearable protection system deploys the airbag only when a real fall occurs, avoiding false triggering due to accidental touch.

[0012] Preferably, the human body descending movements include squatting, bending over, falling down and dropping.

[0013] Preferably, the data acquisition module is integrated into the intelligent wearable protection system, and the data acquisition module includes an inertial measurement unit and a pressure sensor. The inertial measurement unit includes an accelerometer and a gyroscope for real-time acquisition of kinematic data of human body descent action, and the pressure sensor is used to acquire dynamic data.

[0014] Preferably, the kinematic and dynamic data include acceleration, relative positions of joints and pressure values ​​between the person and the airbag suit.

[0015] Preferably, the data processing module is used to pre-process the collected kinematic and dynamic data of the human body's descending action, including filtering, denoising and data standardization, so as to improve the quality and usability of the data.

[0016] Preferably, the deep learning network model, including an improved model of spatiotemporal embedded long short-term memory deep learning network (STLSTM), is used to analyze and identify the preprocessed kinematic and dynamic data of the human body's descent action; the model is trained by a large amount of labeled human body descent action data so that it can accurately distinguish between real falling actions and other descent actions.

[0017] The advantages and beneficial effects of the present invention are:

[0018] This deep learning-based anti-accidental high-altitude fall recognition method uses advanced deep learning technology to accurately identify the actions of high-altitude workers. Compared with traditional recognition methods based on preset thresholds, deep learning models can learn complex patterns and features from a large amount of kinematic and dynamic data, thereby more accurately distinguishing between falls and other descending actions. This high degree of accuracy means that the smart wearable protection system will only activate the airbag when a real fall event is detected, greatly reducing the possibility of false triggering. This advantage not only improves the safety of workers, but also reduces the waste of resources and work interruptions that may be caused by false triggering of airbags.

[0019] Since the present invention can effectively distinguish between falling and non-falling actions, the intelligent wearable protection system will not mistakenly activate the airbag when the operator is performing normal operations (such as squatting or bending over). This provides greater operational freedom for high-altitude operators. In addition, reducing the false triggering of the airbag also means that the operator can enjoy a higher degree of comfort during long-term operations because they do not need to frequently adjust or re-equip the airbag that is deployed due to false triggering. This is crucial to improving work efficiency and the job satisfaction of operators. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of a method for preventing accidental touch and high-altitude fall recognition based on deep learning in an embodiment of the present invention;

[0021] Figure 2 This is a flow chart of PCA feature dimensionality reduction in an embodiment of the present invention;

[0022] Figure 3 This is a structural diagram of the LSTM model in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0024] The present invention proposes a deep learning-based anti-accidental high-altitude fall recognition system, which collects kinematic and dynamic data of the human body in different descending actions (squatting, bending, falling and falling) to train and verify the recognition algorithm. Therefore, it is necessary to carry out descending action experiments.

[0025] Recruit 10 healthy adult volunteers (aged 20-30 years old, without motor impairment or balance problems) and ensure that they fully understand the experimental procedures and potential risks. Obtain written consent from the participants before the experiment and ensure that the experimental process complies with ethical standards.

[0026] The experimental equipment includes a. Inertial measurement unit: including accelerometers and gyroscopes, used to collect kinematic data; b. Pressure sensor: used to collect dynamic data; c. Data acquisition system: including data acquisition card and software, used to record sensor data in real time; d. Airbag simulator: used to simulate the deployment of airbags during falls; e. High-speed camera: used to record the movements of participants for subsequent analysis; f. Experimental clothing: special clothing for installing inertial measurement units and pressure sensors; g. Drop test bench: used to simulate falling movements, equipped with a safety net to ensure the safety of participants.

[0027] Before the experiment, the experimental process and purpose were explained to the participants, and informed consent was obtained. The participants were asked to wear inertial measurement units on their wrists and waists, and pressure sensors were attached between the human body and the airbag suit.

[0028] The experiment described above requires the completion of four descending movements. Squatting movement: participants slowly squat to the lowest point from a standing position, then return to a standing position, and repeat 10 times; bending movement: participants bend forward from a standing position, try to touch the toes, and then return to a standing position, and repeat 10 times; falling movement: on a safety mat, participants simulate falling forward and backward, and repeat 10 times; simulated falling movement: under safety supervision, participants fall freely forward and backward from a low-height (such as 1.5 meters) platform to simulate falling, and repeat 10 times.

[0029] In the data processing module, there will be a large number of interference signals during the data acquisition process, which will affect the accuracy of intention recognition. Generally, these interference signals are eliminated by filtering. The Butterworth filter belongs to the IIR type digital filter, which has the advantages of fast operation speed and good real-time performance. In addition, the Butterworth filter can effectively suppress the zero drift phenomenon generated during the kinematic and dynamic data acquisition process.

[0030] The frequencies of kinematic and dynamic data are concentrated between 50Hz and 300Hz, so the cutoff frequencies of the designed filters are 20Hz and 400Hz respectively; in addition, there is 50Hz power frequency interference in the noise interference, and the collected kinematic and dynamic data are then filtered by a 50Hz notch filter. The bandpass and notch filters are designed as follows:

[0031] Step 1: Determine the parameters of the digital filter.

[0032] According to the filtering requirements of kinematic and dynamic data, the passband cutoff frequency was set to 20 Hz, the stopband cutoff frequency was set to 500 Hz, and the notch frequency was set to 50 Hz.

[0033] Step 2: Calculate the filter order N.

[0034] According to formula (1), the filter order N=4 is calculated.

[0035]

[0036] In the formula, |H a (jΩ) 2 is the magnitude square function of the Butterworth filter, Ω represents the frequency, and Ω c Indicates the cutoff frequency.

[0037] A digital filter that meets the above requirements is designed using MATLAB software. Signals with frequencies between 20 Hz and 400 Hz are basically retained, and signals at 50 Hz are filtered out.

[0038] Since the present invention needs to quickly identify the descending action, it is necessary to avoid complex calculus calculations or transformations when selecting features, and select feature parameters that are simple to calculate and have high real-time performance. Therefore, the present invention extracts the time domain features of the kinematic and dynamic data of the descending action as the input of the deep learning network model. The selected time domain features include the mean absolute value (MAV), the root mean square value (RMS), the variance value (VAR), the zero crossing value (ZC) and the slope sign change value (SSC), which are calculated as follows:

[0039] (1) MAV stands for mean value, which is defined as:

[0040]

[0041] Among them, x i For kinematic and kinetic data.

[0042] (2)RMS indicates the degree of signal change within a certain period of time, which is related to the intensity of limb activity and is defined as:

[0043]

[0044] (3) VAR indicates the degree of deviation of the signal variable from its mean value, which can reflect the severity of the signal change and thus the degree of limb activity. The larger the variance value, the greater the limb movement and the faster the movement speed. It is defined as:

[0045]

[0046] (4) ZC represents the number of times the sensor signal passes through the zero point in a data sampling window, which can reflect the fluctuation of the signal in the time series and is defined as:

[0047]

[0048] (5) SSC represents the number of changes in the sign of the sensor signal slope, which is defined as:

[0049]

[0050] In order to speed up the recognition of intention, fewer and more critical features need to be screened out. Principal component analysis (PCA) is a multivariate dimensionality reduction method, which is often used to study the correlation between multiple variables. The present invention uses PCA to perform linear analysis on the initial gait parameters and selects feature parameters with lower linear correlation to replace the complex original features. The flowchart of PCA is shown in Figure 2 As shown. First, calculate the mean of each feature, and then subtract the corresponding mean from each data point in the data set to centralize the data. Since PCA is very sensitive to the scale of the data, it is usually necessary to standardize the data so that the mean of each feature is 0 and the standard deviation is 1; then, calculate the covariance matrix, which is a matrix that describes the strength of the linear relationship between the features in the data; next, calculate the eigenvalues ​​and eigenvectors of the covariance matrix, the eigenvalues ​​represent the variance of each feature direction, and the eigenvectors represent these feature directions. The eigenvector determines the direction of the new coordinate system, and the eigenvalue determines the scale of these new coordinate axes; next, sort the eigenvalues ​​from large to small, and the eigenvector corresponding to each eigenvalue represents the maximum variance of the data in this direction; finally, select the principal components, determine the number of principal components m according to the amount of information to be retained or a specific standard (such as the cumulative contribution rate reaches 90%), and select the eigenvectors corresponding to the first m principal components. These eigenvectors constitute a new time domain feature set to reflect action information with less data.

[0051] After data collection and data processing, the time domain feature set of the kinematic and dynamic data of the descent action is obtained, and the time domain feature set is input into the deep learning network model to train the model. After the deep learning network model is trained, when the time domain feature set of the kinematic and dynamic data of different descent actions (squatting, bending, falling and falling) is input, the output is the correct descent action label, such as squatting (label a), bending (label b), falling (label c) and falling (label d).

[0052] The deep learning network model construction is improved on the basis of the LSTM model.

[0053] The human body will generate time-related sequence data during the falling process, and the different falling actions will form unique data forms. These sequence data are the basis for studying the movement rules of the human body under different falling actions. The LSTM model is a type of recurrent neural network and has significant advantages in processing time series data. The LSTM model structure is as follows: Figure 3 As shown in the figure, a group of recurrently connected blocks are introduced into the LSTM model, called memory cells, each of which consists of a state unit c k And three multiplication units, namely input gate i k , output gate o k and forget gate f k . Both the input gate and the forget gate are used to update the internal state of the state unit. For example, the input gate determines whether to add the input feature to the state unit, thereby preventing irrelevant inputs from perturbing the network; the forget gate is usually composed of a sigmoid activation function, which determines the information discarded from the cell; the output gate will play a final limiting role on the actual output of the cell based on the changes in the input gate and the state unit. The three gates provide a constant data stream, called a constant error, to keep the memory cell stable. After training, the input gate and output gate will learn to open and close access to the error stream in the constant error of each state unit, eliminating the possibility of action errors due to pre-set time windows, and then accurately modeling complex multivariate time series data.

[0054] Information c of the current state unit of the LSTM model k is the state unit information c of the previous moment k-1 , input gate information i k , the information currently input to the status unit and forget gate f k The output gate is used to control the output value h of the memory cell. k From the state unit c kThe amount of information obtained, so the LSTM model can cause changes in the output through the adjustment of long-term memory, and finally obtain the information of the output gate. The mathematical description of each gate is:

[0055] f k =σ(W f [h k-1 , T k ]+b f ) Formula (8)

[0056] i k =σ(W i [h k-1 , T k ]+b i ) Formula (9)

[0057]

[0058] o k =σ(W o [h k-1 , T k ]+b o Formula (11)

[0059] h k =o k ☉φ(c k ) Formula (12)

[0060] Among them, ⊙ is the multiplication operation of the elements at the corresponding position, W i , W f , W o Represents the weight matrices of the input gate, forget gate, and output gate respectively; b i , b f and b o are the biases of the input gate, forget gate and output gate respectively; T k is the input value, h k-1 and h k is the output of the memory cell at time k-1 and k; σ is the sigmoid function; φ() is the hyperbolic tangent function.

[0061] The movement of the human body in different descending actions will accumulate different motion trajectory information, and there will be differences between these motion trajectory information. Each descending action has a separate and regular trajectory expression. The initial motion data of the human body can effectively identify the subsequent actions and descending actions.

[0062] Given the real partial descent action sequence data, the process of generating a new descent action sequence is as follows: first fix the real descent action data in the previous k-1 time period, predict and generate the position of the joint to be reached at the next moment k according to the generation model, and iterate continuously until a position change sequence with a sequence length and shape that conforms to the real trajectory is generated.

[0063] Although the LSTM model has achieved good research results in human motion intention recognition, to realize its complete gate function, a large amount of data is needed to support the training of the model, and this embodiment can only use the data of the descending stage, so it is difficult to meet the training requirements of the LSTM model. The time interval information and spatial position interval information in the descending action trajectory are highly correlated with the descending action. This correlation can be used as potential information to guide the training of the gate function in the LSTM model. Based on this, the present invention proposes a spatiotemporal embedded deep learning network (STLSTM) model.

[0064] Given the input T at time k k , where T k To characterize the feature set of the descending action, the mathematical description of each gate of the STLSTM model is:

[0065] i k =σ(W Ti T k +W hi T k-1 +W si s k +W qi q k +b i ) Formula (13)

[0066] f k =σ(W Tf T k +W hf T k-1 +W sf s k +W qf q k +b f ) Formula (14)

[0067] o k =σ(W To T k +W ho T k-1 +W so s k +W qo q k +b o ) Formula (15)

[0068] Among them, ik is the input gate, f k For the forget gate, k is the output gate, σ is the sigmoid function, T k is the input value, s k is the vector expression of the spatial position interval between time point k-1 and time point k, s 1 =0,q k is the vector expression of the time interval between time point k-1 and time point k, q 1 =0; W Ti Represents the weight matrix of the input gate at time k, W hi Represents the weight matrix of the input gate at time k-1, W si Represents the weight matrix of the spatial variables at the input gate k, W Tf Represents the weight matrix of the time variable of the forget gate at time k; W hf Represents the weight matrix of the forget gate at time k-1, W sf Represents the weight matrix of the spatial variables at the forget gate k, W qf Represents the weight matrix of the time variable of the forget gate at time k; W To Represents the weight matrix of the output gate at time k; W ho Represents the weight matrix of the output gate at time k-1, W so Represents the weight matrix of the spatial variables of the output gate at time k, W qo Represents the weight matrix of the time variable of the output gate at time k, d is the size of the spatiotemporal factor embedding vector, |c| is the dimension of the state unit; b i , b f and b o They are the biases of the input gate, forget gate and output gate respectively.

[0069] If all continuous motion states are expressed mathematically, a large number of parameters need to be learned in the process, which easily leads to data sparsity problems. In order to solve this problem, the present invention divides the time and space positions into equally spaced sub-blocks. For example, 0.01s is the time sub-block interval length, and a descending action time of about 1s can be divided into 100 equally spaced time sub-blocks. Given the time interval v and the space position interval u, the time interval vector q is interpolated by linear interpolation. k and the spatial position interval vector s k Give a description.

[0070]

[0071] Among them, u(v q ) and l(v q ) are the time interval lengths v qThe maximum and minimum values ​​of the block within the range, u(v s ) and l(v s ) are the space interval lengths v s The maximum and minimum values ​​of the block within the range, and are the time factor matrix and the space factor matrix respectively, N q and N s are the number of time sub-blocks and spatial position sub-blocks respectively.

[0072] For the position sequence {T 1 ,T 2 ,···,T k ,T k+1 ,···,T N}, combined with equations (8) to (9), we get the output h at time k k for:

[0073] h k =STLSTM(h k-1 ,T k ,s k ,q k ), k=1,2,···,N Formula (18)

[0074] The position at time k+1 is generated by predicting the position sequence from the initial time to time k.

[0075] P(T k+1 |T 1 ,T 2 ,···,T k )=Softmax(W T h k +b T ) Formula (19)

[0076] Among them, P(T k+1 |T 1 ,T 2 ,···,T k ) is the probability of the position distribution at time k+1, W T is the transformation matrix, b T is the corresponding offset, and the Softmax() function is a normalized exponential function.

[0077] This paper selects recognition accuracy (Precision, Pre) and comprehensive evaluation index (F-Measure, Fm) as performance indicators, compares the recognition accuracy of different intent recognition models, and verifies the superiority of the proposed STLSTM model in intent recognition accuracy. The expressions of Pre and Fm are:

[0078] Pre=TP / (TP+FP) Formula (20)

[0079] Recall = TP / (TP+FN) Formula (21)

[0080] Fm=(2×Recall×Pre) / (Recall+Pre) Formula (22)

[0081] Among them, TP is the number of correctly classified positive samples, TN is the number of correctly classified negative samples, FN is the number of positive samples that are mistakenly classified as negative samples, and FP is the number of negative samples that are mistakenly classified as positive samples.

[0082] In the field of intent recognition, cross-entropy (E tro ) is generally used to evaluate the degree of difference between the actual value and the predicted value. tro The smaller it is, the better the performance of the intent recognition model is. tro The mathematical expression of is shown in formula (23).

[0083]

[0084] Among them, P(x i ) represents the random variable x i The probability distribution of n represents all n possibilities.

[0085] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for preventing accidental touch and high-altitude fall recognition based on deep learning, characterized in that: The steps include: Step 1, collecting kinematic and dynamic data of the human body's descending action through a data acquisition module; Step 2: The kinematic and dynamic data of the human body's falling action are preprocessed by the data processing module and then input into the deep learning network model for analysis to ensure that the smart wearable protection system deploys the airbag only when a real fall occurs, avoiding false triggering due to accidental touch.

2. According to the deep learning-based anti-accidental-touch high-altitude fall recognition method of claim 1, it is characterized by: The human body descending motion includes squatting, bending, falling and dropping.

3. According to the deep learning-based anti-accidental-touch high-altitude fall recognition method of claim 1, it is characterized by: The data acquisition module is integrated into the intelligent wearable protection system. The data acquisition module includes an inertial measurement unit and a pressure sensor. The inertial measurement unit includes an accelerometer and a gyroscope for real-time acquisition of kinematic data of human body descent action, and the pressure sensor is used to acquire dynamic data.

4. The method for preventing accidental touch and high-altitude fall recognition based on deep learning according to claim 1 is characterized in that: The kinematic and dynamic data of the human body's descending action include acceleration, relative positions of joints and pressure values ​​between the human body and the airbag suit.

5. The method for preventing accidental touch and high-altitude fall recognition based on deep learning according to claim 1 is characterized in that: The data processing module is used to pre-process the collected kinematic and dynamic data of the human body's descending action, including filtering, denoising and data standardization processing to improve the quality and availability of the data.

6. The method for preventing accidental touch and high-altitude fall recognition based on deep learning according to claim 1 is characterized in that: The deep learning network model includes an improved long-short-term memory deep learning network model embedded in time and space, which is used to analyze and identify the preprocessed kinematic and dynamic data of the human body's falling action; the model is trained by a large amount of labeled human body falling action data, so that it can accurately distinguish between real falling actions and other falling actions.

7. The method for preventing accidental touch and high-altitude fall recognition based on deep learning according to claim 1 is characterized in that: The spatiotemporal embedded deep learning network model uses the time interval information and spatial position interval information in the descending action trajectory to have a high correlation with the descending action. This correlation is used as potential information to guide the training of the gate function in the long-short term memory deep learning network. Its input is the characteristic value of the kinematic and dynamic data of the human body's descending action, and its output is the label of the human body's descending action, specifically: Given the input T at time k k , where T k To characterize the feature set of the descending action, the mathematical description of each gate of the spatiotemporal embedding deep learning network model is: i k =σ(W Ti T k +W hi T k-1 +W si s k +W qi q k +b i ) f k =σ(W Tf T k +W hf T k-1 +W sf s k +W qf q k +b f ) o k =σ(W To T k +W ho T k-1 +W so s k +W qo q k +b o ) Among them, i k is the input gate, f k For the forget gate, k is the output gate, σ is the sigmoid function, T k is the input value, s k is the vector expression of the spatial position interval between time point k-1 and time point k, s1 = 0, q k is the vector expression of the time interval between time point k-1 and time point k, q1=0; W Ti Represents the weight matrix of the input gate at time k, W hi Represents the weight matrix of the input gate at time k-1, W si Represents the weight matrix of the spatial variables at the input gate k, W Tf Represents the weight matrix of the time variable of the forget gate at time k; W hf Represents the weight matrix of the forget gate at time k-1, W sf Represents the weight matrix of the spatial variables at the forget gate k, W qf Represents the weight matrix of the time variable of the forget gate at time k; W To Represents the weight matrix of the output gate at time k; W ho Represents the weight matrix of the output gate at time k-1, W so Represents the weight matrix of the spatial variables of the output gate at time k, W qo Represents the weight matrix of the time variable of the output gate at time k, d is the size of the spatiotemporal factor embedding vector, |c| is the dimension of the state unit; b i , b f and b o They are the biases of the input gate, forget gate and output gate respectively.

Images