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

Abstract

The application belongs to the technical field of electric power and relates to a false touch prevention high-altitude falling identification method based on deep learning. The method collects kinematics data and dynamics data of human body lowering actions such as squatting, bending, falling down and falling, removes noise from the data by using a Butterworth filter, extracts feature information capable of realizing rapid identification of falling actions, constructs a deep learning network model, trains a gate mechanism of a long short-term memory neural network model by using space-time parameters, improves the identification speed and accuracy, and enables an intelligent wearable protection system to start working only when falling occurs, thereby avoiding false triggering.

Description

Technical Field

[0001] This invention belongs to the field of power technology and relates to an intelligent detection technology, particularly a deep learning-based method for preventing accidental touches and high-altitude fall detection. Background Technology

[0002] High-altitude electrical work is a crucial part of the power industry, encompassing the installation, maintenance, and repair of power lines. This work typically takes place at heights of tens or even hundreds of meters, in complex and variable environments, including strong winds, extreme weather, and high-voltage electric fields. Workers not only face the risk of falls from heights but also the dangers of electric shock and the physical fatigue resulting from prolonged work at heights. Traditional safety measures, such as safety ropes and harnesses, while providing a degree of protection, have limitations in practical application and cannot completely prevent accidents. For example, safety ropes may fail due to wear or improper handling, and harnesses may not provide sufficient protection in certain emergency situations. Therefore, the safety of high-altitude electrical work remains a focus of industry attention, requiring continuous exploration and improvement of more effective safety technologies and equipment.

[0003] With advancements in technology, intelligent wearable protective systems, especially airbag suits, have provided new safety guarantees for workers at heights. These systems, through the integration of sensors and algorithms, can monitor the worker's movement and environmental changes in real time. Once a fall risk is detected, the system can respond quickly and automatically deploy airbags to provide additional cushioning protection. The introduction of intelligent protective systems not only improves the safety of working at heights but also reduces accidents through intelligent monitoring and early warning. Furthermore, intelligent protective systems can communicate in real time with ground monitoring centers, enabling immediate initiation of rescue procedures in emergencies and improving rescue efficiency. These features make intelligent protective systems play an increasingly important role in working at heights.

[0004] While intelligent safety systems have made some progress in ensuring safety during high-altitude operations, some shortcomings still exist in practical applications. For example, some sensor-based fall detection systems may erroneously trigger airbags due to accidental touches or misoperation.

[0005] Patent CN117746572A proposes a method for detecting high-altitude fall accidents based on a smart safety helmet. This method determines whether a worker experiences weightlessness by collecting the helmet's acceleration along the X, Y, and Z axes, thus indicating whether a fall has occurred. However, due to the complexity of human movements, actions such as squatting or falling can produce acceleration changes similar to a fall, leading the system to misjudge the incident as a fall and incorrectly trigger the airbag. Such false triggering not only reduces the reliability of the safety system but also poses unnecessary distress and risk to workers. Therefore, improving the accuracy of the detection system and reducing false triggering are directions that require further research and improvement in the field of high-altitude work safety technology.

[0006] A search of publicly available patent documents revealed no documents identical to this patent application. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based method for preventing accidental touches and falling from heights, in order to solve the problems mentioned in the background art.

[0008] The technical problem solved by this invention is achieved through the following technical solution:

[0009] A deep learning-based method for preventing accidental touches and fall detection from heights includes the following steps:

[0010] Step 1: Collect kinematic and dynamic data of the human body's descent motion using the data acquisition module;

[0011] Step 2: The data processing module preprocesses the kinematic and dynamic data of the human body's descent, and then inputs it into the deep learning network model for analysis to ensure that the smart wearable protective system deploys airbags only when a real fall occurs, avoiding accidental triggering due to unintended contact.

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

[0013] Preferably, the data acquisition module is integrated into the intelligent wearable protective system. The data acquisition module includes an inertial measurement unit and a pressure sensor. The inertial measurement unit includes an accelerometer and a gyroscope, which are used to collect kinematic data of the human body's descent in real time. The pressure sensor is used to collect dynamic data.

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

[0015] Prior to this, the data processing module is used to preprocess the collected kinematic and dynamic data of the human descent motion, including filtering, denoising, and data standardization, to improve the quality and usability of the data.

[0016] Prior to this, the deep learning network model includes an improved model of a spatiotemporally embedded long short-term memory deep learning network (STLSTM) for analyzing and identifying the kinematic and dynamic data of preprocessed human descent movements; the model is trained with a large amount of labeled human descent movement data so that it can accurately distinguish between real fall movements and other descent movements.

[0017] The advantages and beneficial effects of this invention are as follows:

[0018] This deep learning-based method for preventing accidental triggering of fall detection at heights employs advanced deep learning technology to accurately identify the movements of workers at heights. Compared to traditional methods based on preset thresholds, deep learning models can learn complex patterns and features from vast amounts of kinematic and dynamic data, thus more accurately distinguishing falls from other descent movements. This high level of accuracy means that the intelligent wearable protective system will only activate the airbag when a genuine fall event is detected, significantly reducing the possibility of false triggering. This advantage not only enhances worker safety but also reduces potential resource waste and work interruptions caused by accidental airbag activation.

[0019] Because this invention can effectively distinguish between fall and non-fall actions, the intelligent wearable protective system will not mistakenly deploy airbags when workers are performing normal operations (such as squatting or bending over). This provides greater operational freedom for workers at heights. Furthermore, reduced false airbag deployment also means greater comfort for workers during extended periods of operation, as they do not need to frequently adjust or re-equip themselves with airbags that have deployed due to false triggering. This is crucial for improving work efficiency and worker job satisfaction. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a deep learning-based method for preventing accidental touches and falling from heights in an embodiment of the present invention.

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

[0022] Figure 3 This is a diagram of the LSTM model structure in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0024] This invention proposes a deep learning-based system for preventing accidental falls from heights. It collects kinematic and dynamic data of the human body during different descent movements (squatting, bending, falling, and crashing) to train and validate the recognition algorithm. Therefore, experiments on descent movements are necessary.

[0025] Recruit 10 healthy adult volunteers (aged 20-30, without motor impairments or balance problems) and ensure they fully understand the experimental procedures and potential risks. Written consent from participants must be obtained before the experiment, and the experimental process must comply with ethical standards.

[0026] The experimental equipment includes: a. Inertial Measurement Unit (IMU): comprising accelerometers and gyroscopes for acquiring kinematic data; b. Pressure Sensors: for acquiring dynamic data; c. Data Acquisition System: comprising data acquisition cards and software for real-time recording of sensor data; d. Airbag Simulator: for simulating airbag deployment during a fall; e. High-Speed ​​Camera: for recording participants' movements for subsequent analysis; f. Experimental Suit: a specially designed suit for mounting the IMU and pressure sensors; g. Fall Test Platform: for simulating fall actions, equipped with a safety net to ensure participant safety.

[0027] Before conducting the experiment, the participants were informed of the procedure and purpose, and their informed consent was obtained. Inertial measurement units were fitted to the participants' wrists and waists, and pressure sensors were attached between the body and the airbag suit.

[0028] The experiment requires participants to complete four descent maneuvers: Squatting: Participants slowly squat down from a standing position to their lowest point, then return to a standing position, repeating 10 times; Bending: Participants bend forward from a standing position, attempting to touch their toes, then return to a standing position, repeating 10 times; Falling: On a safety mat, participants simulate falling forward and backward, repeating 10 times; Simulated fall: Under safety supervision, participants freely fall forward and backward from a low-height (e.g., 1.5 meters) platform, simulating a fall, repeating 10 times.

[0029] The data processing module encounters numerous interference signals during data acquisition, affecting the accuracy of intent recognition. These interference signals are typically eliminated through filtering. The Butterworth filter, an IIR-type digital filter, offers advantages such as high processing speed and good real-time performance. Furthermore, the Butterworth filter effectively suppresses zero-point drift phenomena generated during kinematic and dynamic data acquisition.

[0030] The kinematic and dynamic data frequencies are concentrated between 50Hz and 300Hz. Therefore, the cutoff frequencies of the filters are designed to be 20Hz and 400Hz, respectively. Additionally, there is 50Hz power frequency interference in the noise, so the acquired kinematic and dynamic data are further filtered using a 50Hz notch filter. The bandpass and notch filter designs are as follows:

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

[0032] Based on the filtering requirements of kinematic and dynamic data, the bandpass cutoff frequency is set to 20Hz, the stopband cutoff frequency to 500Hz, and the notch filter frequency to 50Hz.

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

[0034] The filter order N = 4 is calculated according to equation (1).

[0035]

[0036] In the formula, |H a (jΩ) 2 Let Ω be the squared amplitude function of the Butterworth filter, where Ω represents the frequency. c Indicates the cutoff frequency.

[0037] Using MATLAB software, a digital filter that meets the above requirements can be designed. Signals with frequencies between 20Hz and 400Hz will be basically preserved, while signals with a frequency of 50Hz will be filtered out.

[0038] Since this invention requires rapid identification of descent movements, complex calculus calculations or transformations must be avoided when selecting features. Instead, simple and real-time feature parameters are chosen. Therefore, this invention extracts the temporal features of the kinematic and dynamic data of the descent movement as input to the deep learning network model. The selected temporal features include Mean Absolute Value (MAV), Root Mean Square Value (RMS), Variance Value (VAR), Zero Crossing Value (ZC), and Slope Sign Change Value (SSC), which are calculated as follows:

[0039] (1) MAV represents the average value, which is defined as:

[0040]

[0041] Where, x i For kinematic and dynamic data.

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

[0043]

[0044] (3) VAR represents the degree of deviation of a signal variable from its mean, reflecting the intensity of signal changes and thus the degree of limb activity. A larger variance indicates greater limb movement and faster speed. It is defined as:

[0045]

[0046] (4) ZC represents the number of times the sensor signal crosses zero within a data sampling window, reflecting the fluctuation of the signal in the time series. It is defined as:

[0047]

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

[0049]

[0050] To accelerate intent recognition, fewer, more critical features need to be filtered out. Principal Component Analysis (PCA) is a multivariate dimensionality reduction method commonly used to study the correlation between multiple variables. This invention utilizes PCA to perform linear analysis on the initial gait parameters, selecting feature parameters with lower linear correlation to replace complex original features. The flowchart of PCA is as follows... Figure 2 As shown. First, the mean of each feature is calculated, and then the mean is subtracted from each data point in the dataset to centralize the data. Since PCA is highly sensitive to data scale, data standardization is usually required, ensuring that the mean of each feature is 0 and the standard deviation is 1. Next, the covariance matrix is ​​calculated, which describes the strength of the linear relationship between features in the data. Then, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. Eigenvalues ​​represent the magnitude of the variance in each feature direction, while eigenvectors represent these feature directions. Eigenvectors determine the orientation of the new coordinate system, while eigenvalues ​​determine the scale of these new coordinate axes. Next, the eigenvalues ​​are sorted from largest to smallest, and the eigenvector corresponding to each eigenvalue represents the maximum variance of the data in that direction. Finally, principal components are selected. The number of principal components, *m*, is determined based on the required amount of information to retain or a specific criterion (e.g., a cumulative contribution rate of 90%). The eigenvectors corresponding to the first *m* principal components are selected; these eigenvectors constitute a new time-domain feature set, reflecting action information with less data.

[0051] After data acquisition and processing, a set of temporal features of the kinematics and dynamics of the descent motion is obtained. This set of temporal features is then input into a deep learning network model for training. Once the deep learning network model is trained, when the set of temporal features of the kinematics and dynamics of different descent motions (squatting, bending over, falling, and dropping) is input, the output is the correct label for the descent motion, such as squatting (label a), bending over (label b), falling (label c), and dropping (label d).

[0052] The deep learning network model described is an improvement on the LSTM model.

[0053] The human body generates time-related sequential data during descent, forming unique data formats depending on the different descent movements. This sequential data is fundamental for studying the motion patterns of the human body under different descent movements. The LSTM model, a type of recurrent neural network, has significant advantages in processing time-series data. The structure of an LSTM model is as follows: Figure 3 As shown, a set of circularly connected blocks, called memory cells, are introduced into the LSTM model. Each memory cell consists of a state unit c. k It consists of three multiplication units, namely the input gate i k Output gate o k and the Gate of Oblivion 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 input features to the state unit, thus preventing irrelevant inputs from perturbing the network; the forget gate, typically composed of a sigmoid activation function, determines the information discarded from the cell; and the output gate ultimately limits the cell's actual output based on changes in the input gate and the state unit. These three gates provide a constant data stream, called constant error, which keeps the memory cell stable. Through training, the input and output gates learn to open and close access to the error stream within the constant error of each state unit, eliminating the possibility of errors caused by pre-set time windows, thereby enabling accurate modeling of complex multivariate time series data.

[0054] Information c of the current state unit of the LSTM model k It is derived from the state unit information c of the previous moment. k-1 Input gate information i k Information currently input to the status unit and the Gate of Oblivion k It is jointly determined. The function of the output gate is to control the output value h of the memory cell. k From state unit c kThe amount of information acquired allows the LSTM model to adjust its long-term memory, thereby changing the output and ultimately obtaining the information from the output gates. The mathematical description of each gate is as follows:

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

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

[0057]

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

[0059] h k =o k ☉φ(c k Equation (12)

[0060] Where ⊙ represents element-wise multiplication, W i W f W o b represents the weight matrices for the input gate, forget gate, and output gate, respectively; i b f and b o These are the biases for the input gate, forget gate, and output gate, respectively; T k h is the input value. k-1 and h k σ represents the output of the memory cell at times k-1 and k; σ is the sigmoid function; φ() is the hyperbolic tangent function.

[0061] The human body accumulates different motion trajectory information during different descent movements, and these trajectories vary. Each descent movement has a unique and predictable trajectory, and subsequent movements and descent movements can be effectively identified by using the initial motion data of the human body.

[0062] Given a real partial descent sequence data, the process of generating a new descent sequence is as follows: First, fix the real descent data within the first k-1 time intervals, predict and generate the position of the joint at the next time k based on 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] While LSTM models have achieved promising results in human motion intention recognition, implementing their full gate function requires extensive data support for model training. However, this embodiment only utilizes data from the descent phase, thus failing to meet the training requirements of LSTM models. Furthermore, the time interval and spatial location interval information in the descent trajectory are highly correlated with the descent action. This correlation can serve as latent information to guide the training of the gate function in the LSTM model. Based on this, this invention proposes a Spatiotemporal Embedded Deep Learning Network (STLSTM) model.

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

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

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

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

[0068] Among them, ik For the input gate, f k For the Gate of Oblivion, o k For the output gate, σ is the sigmoid function, and T is the output gate. k For the input value, s k It is a vector representation of the spatial interval between time point k-1 and time point k, s1=0, q k It is a vector representation of the time interval between time point k-1 and time point k, where q1 = 0; W Ti W represents the weight matrix of the input gate at time k. hi W represents the weight matrix of the input gate at time k-1. si W represents the weight matrix of the spatial variables at time k of the input gate. Tf W represents the weight matrix of the time variables at time k in the forgetting gate. hf W represents the weight matrix of the forget gate at time k-1. sf W represents the weight matrix of the spatial variables at time k of the forgetting gate. qf W represents the weight matrix of the time variables at time k in the forgetting gate. To W represents the weight matrix of the output gate at time k; ho W represents the weight matrix of the output gate at time k-1. so W represents the weight matrix of the spatial variables at time k of the output gate. qo The weight matrix represents the time variable of the output gate at time k. d represents the size of the spatiotemporal factor embedding vector, and |c| represents the dimension of the state unit; b i b f and b o These are the biases for the input gate, forget gate, and output gate, respectively.

[0069] If all continuous motion states are mathematically represented, a large number of parameters need to be learned, which can easily lead to data sparsity. To address this issue, this invention divides time and spatial position into equally spaced sub-blocks. For example, a time sub-block interval of 0.01s means that a descent motion lasting approximately 1s can be divided into 100 equal-length time sub-blocks. Given the time interval v and the spatial position interval u, a linear interpolation method is used to interpolate the time interval vector q. k and spatial position interval vector s k Describe it.

[0070]

[0071] Where u(v) q ) and l(v q ) represent the time interval length v q The maximum and minimum values ​​of the block within the range, u(v)s ) and l(v s ) represent the spatial interval length v s The maximum and minimum values ​​of the block within the range. and These are the time factor matrix and the spatial factor matrix, respectively, N q and N s These represent the number of time sub-blocks and the number of spatial location sub-blocks, respectively.

[0072] For the position sequence {T1,T2,...,T...} k ,T k+1 ,···,T N Combining equations (8) to (9), we obtain 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 Equation (18)

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

[0075] P(T k+1 |T1,T2,···,T k =Softmax(W) T h k +b T Equation (19)

[0076] Among them, P(T) k+1 |T1,T2,···,T k W is the probability of the position distribution at time k+1. T It is a transformation matrix, b T This is the corresponding offset, and the Softmax() function is a normalized exponential function.

[0077] This invention selects recognition accuracy (Pre) and comprehensive evaluation index (F-Measure, Fm) as performance indicators to compare the recognition accuracy of different intent recognition models, verifying the superiority of the proposed STLSTM model in terms of intent recognition accuracy. The expressions for Pre and Fm are as follows:

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

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

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

[0081] Where 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 were misclassified as negative samples, and FP is the number of negative samples that were misclassified as positive samples.

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

[0083]

[0084] Wherein, P(x i ) represents the random variable x i The probability distribution is given by n, where n represents all n possibilities.

[0085] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations 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 deep learning-based method for preventing accidental touches and falling objects from heights, characterized in that: Includes the following steps: Step 1: Collect kinematic and dynamic data of the human body's descent motion using the data acquisition module; Step 2: The data processing module preprocesses the kinematic and dynamic data of the human body's descent, and then inputs it into the deep learning network model for analysis to ensure that the smart wearable protective system deploys airbags only when a real fall occurs, avoiding accidental triggering due to accidental contact. The deep learning network model mentioned above includes an improved model of a long short-term memory deep learning network with spatiotemporal embedding, which is used to analyze and identify the kinematic and dynamic data of preprocessed human descent movements; the model is trained with a large amount of labeled human descent movement data so that it can accurately distinguish between real fall movements and other descent movements. The spatiotemporal embedding deep learning network model utilizes the high correlation between the time interval and spatial location interval information in the descent trajectory and the descent action. This correlation serves as latent information to guide the training of the gating function in the long short-term memory deep learning network. Its input consists of feature values ​​from the kinematic and dynamic data of the human descent action, and its output is the label of the human descent action, specifically: Given time k Input T k ,in T k To characterize the feature set of the descent action, the mathematical description of each gate in the spatiotemporal embedding deep learning network model is as follows: in, i k For input gate, f k For the Gate of Oblivion o k For output gate, σ For the sigmoid function, T k For input values, s k It is a point in time. k -1 and time point k Vector representation of the spatial interval between them. s 1 =0, q k It is a point in time. k -1 and time point k Vector representation of the time interval between them. q 1 = 0; W Ti The input gate represents k The weight matrix at time step, W hi The input gate represents k The weight matrix at time -1 W si Representative input gate k Weight matrix of time-space variables, W Tf Represents the Gate of Oblivion k The weight matrix of time variables at each time step; W hf Representing the Gate of Oblivion k The weight matrix at time -1 W sf Represents the Gate of Oblivion k Weight matrix of time-space variables, W qf Represents the Gate of Oblivion k The weight matrix of time variables at each time step; W To Represents the output gate in k The weight matrix at each time step; W ho Represents the output gate in k The weight matrix at time -1 W so Representative output gate k The weight matrix of the time-space variables. W qo Representative output gate k The weight matrix of the time variable at time step. , d The size of the spatiotemporal factor embedding vector. The dimension of the state unit; b i , b f and b o These are the biases for the input gate, forget gate, and output gate, respectively.

2. The method for preventing accidental touches and falling from heights based on deep learning according to claim 1, characterized in that: The aforementioned human descent movements include squatting, bending over, falling, and dropping.

3. The method for preventing accidental touches and falling from heights based on deep learning according to claim 1, characterized in that: The data acquisition module is integrated into the intelligent wearable protective system. The data acquisition module includes an inertial measurement unit and a pressure sensor. The inertial measurement unit includes an accelerometer and a gyroscope, which are used to collect kinematic data of the human body's descent in real time. The pressure sensor is used to collect dynamic data.

4. The method for preventing accidental touches and falling objects from heights based on deep learning according to claim 1, characterized in that: The kinematic and dynamic data of the human descent include acceleration, relative joint positions, and pressure values ​​between the person and the airbag suit.

5. The method for preventing accidental touches and falling objects from heights based on deep learning according to claim 1, characterized in that: The data processing module is used to preprocess the collected kinematic and dynamic data of human descent, including filtering, noise reduction, and data standardization, to improve the quality and usability of the data.

Images