Deep learning-based unsafe behavior monitoring and early warning method and device

Abstract

The present application relates to the field of monitoring and early warning, in particular to an unsafe behavior monitoring and early warning method and device based on deep learning; containing a wearable monitoring device, a fluid impedance tomography monitoring module and a fluid-solid coupling wave equation inverse solution model; the wearable monitoring device integrates a low-pressure pneumatic flexible network, a magneto-rheological fluid layer and a negative Poisson ratio expansion skeleton, real-time collects pressure pulsation spectrum and electromagnetic parameters, and uses deep learning to invert the normalized stiffness index of the human contact part; the core is to link and control the magnetic field PWM and the air pressure frequency according to the stiffness determination result, use the air pressure pretightening force to induce the phase change of the magneto-rheological fluid and generate high-frequency vibration early warning; the present application uses the geometric antagonistic mechanism of the expansion skeleton to ensure that the device is closely attached and the signal is high-fidelity, and realizes accurate tactile feedback monitoring.

Description

Technical Field

[0001] This invention relates to the field of monitoring and early warning, specifically to a method and apparatus for monitoring and early warning of unsafe behaviors based on deep learning. Background Technology

[0002] In high-risk confined space operations such as inside wind turbine blades and underground mines, safety monitoring is a core element in ensuring the safety of workers. These extreme environments often have severe visual obstruction and high-intensity mechanical noise interference, making conventional monitoring methods ineffective.

[0003] Currently, monitoring technologies for such scenarios mainly rely on video surveillance, voice intercom, and wearable physiological monitoring devices. However, in actual operations, existing monitoring methods have significant limitations. First, the complex structure of confined spaces and dusty environments often obstruct the camera's field of view, while high-decibel noise makes it easy for audible and visual alarm signals to be masked, often preventing workers from receiving real-time danger warnings. Second, existing physiological monitoring devices are mostly based on surface electromyography (EMG) signal acquisition. In high-temperature and high-humidity working environments, skin sweat can severely interfere with the conductivity of the electrodes, leading to distortion of the signal coupling interface and making it difficult to accurately obtain physiological data of workers.

[0004] Furthermore, existing safety early warning systems primarily focus on monitoring environmental or basic physiological parameters, lacking the ability to deeply identify workers' muscle fatigue, hidden violations, and stamina-building states. Because they cannot quantify the microscopic muscle changes caused by complex movements in confined spaces, the systems struggle to provide effective early warnings during the critical millisecond period before an accident. Simultaneously, existing feedback devices often employ rigid structures or simple vibration motors, resulting in poor wearing comfort and severe signal attenuation under heavy industrial protective clothing, leading to low sensitivity and failing to ensure clear and timely capture of warning information by workers under extreme conditions.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for monitoring and warning of unsafe behaviors based on deep learning, so as to solve the problems mentioned in the background art. The technical solution of this invention includes:

[0007] S1. A wearable monitoring device is installed, which includes a low-pressure pneumatic flexible network as a basic skeleton, a magnetorheological fluid layer encapsulated in the cavity of the low-pressure pneumatic flexible network, and a negative Poisson's ratio tensile skeleton embedded in the airbag wall of the low-pressure pneumatic flexible network; the low-pressure pneumatic flexible network is connected to a pneumatic drive unit, and the magnetorheological fluid layer is in a controllable magnetic field environment.

[0008] S2. Activate the wearable monitoring device and use the fluid impedance tomography monitoring module to collect high-frequency pressure pulsation spectrum data inside the low-pressure aerodynamic flexible network in real time, and simultaneously collect the inductance or capacitance change rate data of the magnetorheological fluid layer.

[0009] S3. Input the high-frequency pressure pulsation spectrum data and the inductance or capacitance change rate data into the preset fluid-structure interaction wave equation inverse solution model, and calculate the normalized stiffness index of the human body contact part through deep learning algorithm.

[0010] S4. Based on the comparison between the normalized stiffness index and the preset safety threshold, the wearable monitoring device is subjected to physical field medium regulation; when an unsafe behavior is determined, the PWM duty cycle of the controllable magnetic field and the output frequency of the pneumatic drive unit are regulated in conjunction, and the internal air pressure of the low-pressure pneumatic flexible network is used as a pre-tightening force to cause the magnetorheological fluid layer to undergo a phase change and generate a rigid structure capable of transmitting 200-300Hz vibration signals.

[0011] Preferably, the step S1 includes the following steps before: S1.1, by constructing a physical model, with the optimization objectives of maximizing the contact area and homogenizing the local pressure when the low-pressure pneumatic flexible network is inflated, simulating and calculating the structural reentry angle and density parameters of the negative Poisson's ratio tensile skeleton, so that the negative Poisson's ratio tensile skeleton generates a lateral expansion force when the air pressure increases to counteract the normal expansion of the airbag.

[0012] Preferably, in step S3: the inverse solution model of the fluid-structure interaction wave equation adopts an LSTM-CNN deep neural network architecture, wherein the deep learning algorithm uses the pressure wave reflection coefficient and phase lag as input features and the microscopic isometric contraction rate of human muscles as output features.

[0013] Preferably, step S4 includes: when the normalized stiffness index is in the first range, adjusting the PWM duty cycle of the controllable magnetic field to increase the viscosity of the magnetorheological fluid layer to provide damping; when the normalized stiffness index is in the second range, locking the controllable magnetic field with full power output and adjusting the pneumatic drive unit to generate high-frequency pulsation; the negative Poisson's ratio stretched skeleton forces the low-pressure pneumatic flexible network to unfold tangentially during the pressurization process of the pneumatic drive unit, maintaining a constant contact area with human skin.

[0014] Preferably, in step S1, the magnetorheological fluid layer is composed of soft magnetic particles suspended in a carrier liquid, and the volume fraction of the soft magnetic particles is 20%-40%.

[0015] Preferably, in step S4, the 200-300Hz vibration signal corresponds to the sensitive frequency range of the human Pacinian body.

[0016] Preferably, in step S1, the low-pressure pneumatic flexible network has a honeycomb cavity structure, and the magnetorheological fluid layer fills the gaps in the honeycomb cavity.

[0017] A deep learning-based unsafe behavior monitoring and early warning device includes:

[0018] A low-pressure aerodynamic flexible network serves as the main driving force field framework, used for conformal contact with human skin.

[0019] A pneumatic drive unit, connected to the low-pressure pneumatic flexible network, is used to regulate the internal air pressure and generate high-frequency pressure waves.

[0020] A magnetorheological fluid layer, encapsulated within the cavity structure of the low-pressure aerodynamic flexible network, is used to provide a variable physical intervention field;

[0021] A magnetic field generator is disposed around the magnetorheological fluid layer to generate a high-frequency alternating magnetic field for controlling the phase change of the fluid.

[0022] A negative Poisson's ratio expansion skeleton is embedded in the airbag wall of the low-pressure aerodynamic flexible network to generate tangential expansion force when the air pressure changes.

[0023] The edge computing control terminal is electrically connected to the pneumatic drive unit and the magnetic field generator, respectively, and is used to execute the calculation and control command output of the inverse solution model of the fluid-structure interaction wave equation.

[0024] Preferably, the unsafe behavior monitoring and early warning device further includes a fluid impedance tomography monitoring module, which includes a pressure sensor array arranged inside the low-pressure aerodynamic flexible network and an electromagnetic induction probe for monitoring the state of the magnetorheological fluid layer.

[0025] Compared with the prior art, the present invention has the following improvements and advantages:

[0026] 1. This solution embeds a negative Poisson's ratio tensile skeleton into the airbag wall, utilizing its unique geometric antagonistic mechanism to generate a lateral expansion force to counteract normal expansion when air pressure increases. This ensures that the wearable monitoring device can closely conform to the curvature of the human body under different pressure levels, not only improving wearing comfort but also guaranteeing the constancy and high fidelity of the signal acquisition interface;

[0027] 2. This scheme employs fluid impedance tomography (FIT) technology, treating the air pressure pulsations within the pneumatic network as signal carriers. By introducing a curvature-acoustic dispersion coupling correction algorithm, it eliminates the phase distortion caused by pipe bending during wear. The formula for calculating the corrected wavenumber is as follows: By combining magnetorheological fluid electroacoustic coupling correction, it is possible to accurately extract the microscopic isometric contraction rate that characterizes the muscle's power storage state from complex fluid-structure interaction fluctuation data, effectively avoiding the interference of environmental noise and limb movement.

[0028] 3. This scheme uses a physical information-guided composite loss function to train the deep neural network. By introducing physical residual constraints from the wave equation, the model output is forced to conform to the physical laws of fluid-structure interaction, significantly improving the system's generalization ability under unseen operating conditions and achieving millisecond-level prediction of unsafe behaviors;

[0029] 4. This solution employs tiered intervention based on the normalized stiffness index. Upon identifying a high-risk situation, it utilizes the phase transition characteristics of magnetorheological fluids activated by a magnetic field, and uses air pressure as a pre-tensioning force to instantly form a rigid structure. This structure can transmit high-frequency vibration signals without loss, and this frequency band precisely corresponds to the sensitive range of Pacinian bodies deep within human skin. Even under heavy protective clothing and background noise interference, this warning signal can be clearly perceived by workers, greatly improving the response speed in emergency situations.

[0030] 5. The honeycomb cavity structure adopted in this solution not only enhances the structural stability, but also prevents the sedimentation of soft magnetic particles through the micro-cavity design. The magnetorheological fluid maintains low viscosity under normal conditions to adapt to limb movement, while generating huge magnetostrictive yield stress under intervention conditions. This design provides a great dynamic adjustment range between being imperceptible during daily wear and providing strong intervention at critical moments. Attached Figure Description

[0031] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0032] Figure 1 This is a schematic diagram of the external structure of a wearable monitoring device;

[0033] Figure 2 This is a schematic diagram of the internal honeycomb cavity structure distribution of the low-pressure pneumatic flexible network of the wearable monitoring device;

[0034] Figure 3 This is a schematic diagram of the internal microstructure and structural location of the low-pressure aerodynamic flexible network of the wearable monitoring device.

[0035] Figure 4 This is a schematic diagram of the process flow of the method of the present invention.

[0036] In the figure: 100, wearable monitoring device; 110, low-pressure pneumatic flexible network; 111, honeycomb cavity; 112, airbag wall; 120, magnetorheological fluid layer; 121, soft magnetic particles; 130, negative Poisson's ratio tensile skeleton; 140, highly elastic film; 200, pneumatic drive unit; 300, magnetic field generator; 400, fluid impedance tomography monitoring module; 410, pressure sensor array; 420, electromagnetic induction probe. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0038] Example 1:

[0039] Please see Figures 1-4 This invention provides a deep learning-based method for monitoring and warning of unsafe behaviors, including:

[0040] S1. A wearable monitoring device 100 is installed, which includes a low-pressure pneumatic flexible network 110 as a basic skeleton, a magnetorheological fluid layer 120 encapsulated in the cavity of the low-pressure pneumatic flexible network 110, and a negative Poisson's ratio tensile skeleton 130 embedded in the airbag wall 112 of the low-pressure pneumatic flexible network 110; the low-pressure pneumatic flexible network 110 is connected to a pneumatic drive unit 200, and the magnetorheological fluid layer 120 is in a controllable magnetic field environment.

[0041] S2. Start the wearable monitoring device 100 and use the fluid impedance tomography monitoring module 400 to collect high-frequency pressure pulsation spectrum data inside the low-pressure aerodynamic flexible network 110 in real time, and simultaneously collect the inductance or capacitance change rate data of the magnetorheological fluid layer 120.

[0042] S3. Input the high-frequency pressure pulsation spectrum data and the inductance or capacitance change rate data into the preset fluid-structure interaction wave equation inverse solution model, and calculate the normalized stiffness index of the human body contact part through deep learning algorithm.

[0043] S4. Based on the comparison between the normalized stiffness index and the preset safety threshold, the wearable monitoring device 100 is subjected to physical field medium regulation; when an unsafe behavior is determined, the PWM duty cycle of the controllable magnetic field and the output frequency of the pneumatic drive unit 200 are regulated in conjunction, and the internal air pressure of the low-pressure pneumatic flexible network 110 is used as the pre-tightening force to cause the magnetorheological fluid layer 120 to undergo a phase change and generate a rigid structure capable of transmitting 200-300Hz vibration signals.

[0044] This embodiment addresses the visual obstruction and high-intensity noise interference problems in high-risk confined spaces, such as inside wind turbine blades and in underground mine operations, by proposing a tactile feedback monitoring scheme based on dual-physical-field coupling. In these extreme environments, workers often cannot perceive their own muscle fatigue or impending unsafe postures, rendering traditional audible and visual alarms completely ineffective. This method constructs a rigid-flexible composite medium carrier through step S1, in which the low-pressure aerodynamic flexible network 110 serves as a macroscopic configuration field, responsible for providing clothing-like compliance under normal conditions, ensuring that it does not interfere with the complex joint movements of workers; while the magnetorheological fluid layer 120 serves as a microscopic property field, latent within the aerodynamic framework, ready to undergo sudden changes in physical properties at any time.

[0045] Steps S2 to S3 change the traditional monitoring logic based on electromyography signals. By using fluid impedance tomography, the air pressure pulsation inside the pneumatic network is regarded as a carrier wave. By monitoring the reflection and distortion of this carrier wave at the contact interface, the microscopic isometric contraction rate of human muscles is inverted. This method of extracting latent features avoids the interference of sweat on the electrodes. In the warning triggering stage of S4, the system does not simply superimpose air pressure and magnetic field, but uses the high-pressure gas generated by the air pressure drive unit 200 as a pre-tightening force to apply basic compressive stress to the magnetorheological fluid layer 120.

[0046] To ensure that the air pressure preload can be effectively transmitted to the fluid layer, a highly elastic film 140, such as a 0.1mm thick Ecoflex silicone film, is used as a physical diaphragm between the pneumatic network cavity and the magnetorheological fluid layer 120. This film has extremely low bending stiffness and can transmit the hydrostatic pressure in the air cavity to the magnetorheological fluid without loss, so that it can be activated by the magnetic field under the state of compressed microparticle spacing. This greatly increases the yield stress threshold of the magnetorheological fluid at the moment of magnetic field activation, thereby forming a rigid structure that can transmit 200-300Hz high-frequency vibration signals without loss without increasing the volume, ensuring that the warning signal can directly penetrate the thick protective clothing and be perceived by the workers.

[0047] Before step S1, the steps include: S1.1, by constructing a physical model, with the optimization objectives of maximizing the contact area and homogenizing the local pressure when the low-pressure aerodynamic flexible network 110 is inflated, the structural reentry angle and density parameters of the negative Poisson's ratio tensile skeleton 130 are simulated and calculated, so that the negative Poisson's ratio tensile skeleton 130 generates a lateral expansion force when the air pressure increases to counteract the normal expansion of the airbag.

[0048] This embodiment aims to address the inherent physical bottleneck of the spherical expansion effect of pneumatic flexible actuators. Traditional airbags tend to expand and become rounded in the direction normal to and perpendicular to the skin during inflation, resulting in a sharp reduction in the contact area with the skin and stress concentration at the contact point. This not only causes discomfort but, more seriously, leads to distortion of the signal coupling interface in fluid impedance tomography. Step S1.1 introduces a geometric antagonistic mechanism based on the negative Poisson's ratio effect. Specifically, this physical model is a mechanical equilibrium solver built based on the constitutive equations of anisotropic hyperelastic materials; the model defines the internal pressure of the airbag. 1. Skeleton geometry and topology, determined by weaving angles and density Characterization and Skin Strain Tensor The nonlinear mapping relationship between them; its core logic lies in solving the deformation compatibility equations. The boundary conditions, where, The radial strain corresponds to the normal expansion. Let be the equivalent Poisson's ratio function of the skeleton. The tangential strain corresponding to the lateral expansion of the expansion skeleton;

[0049] Regarding the missing function definition issue pointed out in Question 1, this embodiment explicitly discloses... The analytical expression: Based on the micromechanical model of the reentrant cellular structure, the function is defined as:

[0050] ;

[0051] in, The aspect ratio of the skeletal unit cell structure is taken as 2.4 in this embodiment. For fiber braiding angles;

[0052] Substituting this specific constitutive equation into the deformation compatibility equation, and calculating the nodal displacements using the finite element method, the model quantifies the effect of the weaving angle. Within a specific range, such as When the structure exhibits a large negative Poisson's ratio effect, the transverse and tangential expansion force Ftan generated when the skeleton is under tension can precisely offset the normal expansion component of the airbag, forcing the low-pressure pneumatic flexible network to conform to the curvature of the human body even under high pressure. This design ensures that the wearable monitoring device 100 and the human body maintain a constant and maximized signal acquisition interface at any pressure level, providing a physical basis for the high accuracy of subsequent muscle stiffness inversion.

[0053] In step S3: the inverse solution model of the fluid-structure interaction wave equation adopts the LSTM-CNN deep neural network architecture. The deep learning algorithm takes the pressure wave reflection coefficient and phase lag as input features and the microscopic isometric contraction rate of human muscles as output features.

[0054] The core objective of this model is to establish the propagation and dissipation law of pressure waves at the interface between soft tissue and fluid, so as to inversely deduce the physical stiffness of deep muscles by measuring wave distortion without invading the skin.

[0055] This embodiment specifies the algorithm logic for non-contact muscle state perception; addressing the issue of missing feature extraction logic mentioned in the task list, this embodiment clarifies the signal processing flow from raw data acquisition to model input features: the system uses the improved dual-sensor transfer function method to perform wavefield separation based on the data from the pressure sensor array 410 in the fluid impedance tomography monitoring module 400.

[0056] To address the compatibility issue between the algorithm model and the physical scenario pointed out in Problem 1—namely, the inability of the rigid straight tube formula to adapt to flexible curved waveguides—this embodiment introduces a curvature-acoustic dispersion coupling correction algorithm. Although the wearable monitoring device 100 has a reserved straight tube waveguide channel, the channel will inevitably bend with the limbs during wear. The system uses the deformation data of the S1 negative Poisson's ratio skeleton to calculate the geometric curvature of the waveguide channel in real time and dynamically compensates for the acoustic propagation constant: The corrected wavenumber is defined as follows:

[0057] ;

[0058] in, The uncorrected fundamental sound propagation constant. Waveguide geometric dispersion correction factor, in units Its numerical value represents the square of the equivalent characteristic length of the curved path. The real-time geometric curvature of the waveguide channel; substituting the correction parameters into the formula:

[0059] ;

[0060] in, Let be the measured frequency domain transfer function between the two sensors. The equivalent physical distance between the two sensors along the direction of sound wave propagation. This is the theoretical spatial phase delay factor for the positive propagation of the pressure wave between the two sensors;

[0061] This eliminates the phase distortion caused by bending, ensuring the physical fidelity of the reflection coefficient calculation; based on this, the rate of change of inductance or capacitance acquired by S2 is used... For example, the rate of change of capacitance The fundamental acoustic phase is corrected a second time, and the calculation formula is as follows:

[0062] ;

[0063] Here, The fundamental acoustic phase difference caused by purely physical distortion before the addition of electroacoustic coupling parameters. Let be the electroacoustic coupling coefficient of the magnetorheological fluid, taking into account the changes in the physical properties of the magnetorheological fluid under different magnetic field strengths. It is based on the PWM duty cycle of the current controllable magnetic field. The function variables are dynamically calibrated; the system pre-stores fitting relationships determined experimentally.

[0064] ;

[0065] This eliminates the nonlinear interference of fluid density changes on wave velocity and phase; among which, the fitting constant... Depending on the specific formulation of the magnetorheological fluid, in this embodiment, when a suspension with a soft magnetic particle volume fraction of 30% is used, the range of values ​​for the dimensionless constant is:

[0066] , , ;

[0067] After the above corrections, accurate reflection coefficients and phase lags are obtained. In response to the problem of loss of key frequency domain information caused by feature engineering as pointed out in Problem 2, this embodiment reconstructs the input tensor construction method of the deep learning model: abandoning the original wideband averaging strategy and adopting a full-spectrum feature preservation scheme.

[0068] Given that changes in muscle state are often reflected in alterations in the shape of the impedance spectrum, such as resonance peak shifts, the system discretizes the 200-300Hz frequency band into... Subband, in this example Preserve the amplitude at each frequency point and phase At this point, the size of the constructed input tensor is... ,in It includes 32 amplitude channels, 32 phase channels, and 1 capacitance rate channel; regarding the missing tensor space mapping logic issue pointed out in Question 1, it is hereby clarified that: The dimension does not directly correspond to the physical sensor coordinates, but is a virtual feature plane generated through spatial reconstruction;

[0069] Because the physical pressure sensor is limited by the distribution of 110 nodes in the low-pressure aerodynamic flexible network (e.g., 64 discrete points), the system employs a radial basis function interpolation algorithm to map the discrete spectral features to a standard 16×16 Cartesian grid. This constructs a continuous spatial feature map for the CNN to extract texture features; the CNN layer is responsible for... Spatial dimensions and Extracting stiffness distribution features in the frequency domain, the LSTM layer is responsible for... It captures evolutionary trends in dimensions; this high-dimensional input fully preserves the spectral fingerprint, enabling the model to accurately identify subtle frequency response changes that characterize microscopic isometric shrinkage rates;

[0070] Regarding the construction of the model training dataset, especially the method for obtaining the microscopic isometric contraction rate as the output label, this embodiment uses an ultrasound-pressure synchronous calibration experiment for data acquisition. Specifically, a high frame rate plane wave ultrasound imaging system, such as the Verasonics Vantage, is vertically aligned with the target muscle group, and data is acquired synchronously with the wearable monitoring device 100. The actual microscopic displacement of the muscle bundle is calculated using an ultrasound image speckle tracking algorithm, and the dimensionless microscopic strain, i.e., the aforementioned microscopic isometric contraction rate, is obtained from the differential of the displacement field. This is used as the truth label, forming a supervised learning sample pair with the input pressure wave features. Regarding the issue of the undisclosed loss function mentioned in Question 3, this embodiment provides the following explanation: To overcome the multiple solutions to the inversion problem, the model training employs a physical information-guided composite loss function, defined as:

[0071] ;

[0072] in, The mean square error between the predicted value and the true ultrasound value. This is a balance coefficient, such as 0.1. Based on the standard acoustic wave equation The physical residual constraint term in the formula This refers to the real-time dynamic sound pressure inside the cavity of the low-pressure aerodynamic flexible network 110. The equivalent propagation velocity of the pressure wave in the current coupling medium is given by the loss function. This loss function forces the model output to satisfy the physical conservation law of fluid-structure interaction, thereby ensuring the model's generalization ability under unseen conditions. This deep learning model can extract the interference caused by environmental noise and body movement from complex fluid-structure interaction wave data, accurately invert the microscopic isometric contraction rate representing the human body's energy storage state or the precursor to spasm, and achieve millisecond-level prediction of unsafe behavior.

[0073] Step S4 includes: when the normalized stiffness index is in the first range, adjusting the PWM duty cycle of the controllable magnetic field to increase the viscosity of the magnetorheological fluid layer 120 to provide damping; when the normalized stiffness index is in the second range, locking the controllable magnetic field with full power output, that is, the control unit outputs a DC signal with a 100% duty cycle, causing the magnetic field generator 300 to generate a constant saturated magnetic field, forcing the magnetic particles in the magnetorheological fluid to instantly form a penetrating particle chain structure, and adjusting the pneumatic drive unit 200 to generate high-frequency pulsation; during the pressurization process of the pneumatic drive unit 200, the negative Poisson's ratio stretched skeleton 130 forces the low-pressure pneumatic flexible network 110 to expand tangentially, maintaining a constant contact area with human skin; in step S4, the 200-300Hz vibration signal corresponds to the sensitive frequency range of the human Pacinian body;

[0074] This embodiment illustrates a graded physical intervention strategy and its physiological basis. Addressing the logical connection between the S3 output characteristics and the S4 input criteria, this embodiment establishes a physical mapping relationship from microscopic isometric contraction rate to stiffness modulus K. Based on the Hill muscle mechanics model, muscle stiffness K and microscopic isometric contraction rate... This is defined as a nonlinear mapping existing between dimensionless strains:

[0075] ;

[0076] in, For the relaxed state foundation stiffness, This represents the maximum rigidity during contraction. It is a nonlinear correction factor, a dimensionless positive number, characterizing the exponential hardening trend of muscle stiffness with respect to the rate of contraction; because It has been clearly defined as a dimensionless physical quantity, ensuring the exponent term. In this embodiment, the typical range of values ​​for the physical meaning and dimensional validity of the operation is as follows: The system uses this mapping to convert the shrinkage rate output by the deep learning model into a physical stiffness value, and then quantifies and defines the normalized stiffness exponent: using the normalized stiffness exponent as the physical unit, the calculation formula is as follows:

[0077] ;

[0078] in, This refers to the current stiffness modulus calculated using the above mapping, in kPa; based on this, the system sets a clear trigger boundary: the first interval is defined as... Corresponding to the statistical normal distribution to The deviation range represents abnormal muscle tension accumulation; the second interval is defined as... This corresponds to the physiological threshold of tetanic contraction. In the first interval, for minor risks such as non-standard posture, the system only adjusts the PWM duty cycle to thicken the magnetorheological fluid, providing a viscous damping sensation for joint movement, serving as a non-mandatory tactile cue. In the second interval, for high risks such as impending fall or misoperation, the system employs a time-division multiplexing asynchronous coupling strategy to specifically implement the aforementioned linkage control logic. The linkage control is characterized by microsecond-level asynchronous execution in terms of timing, prioritizing the structural deployment driven by air pressure and executing the rigid locking excited by the magnetic field.

[0079] Here, the aforementioned linkage control refers to the edge computing control terminal issuing commands simultaneously at the decision level using air pressure frequency and magnetic field PWM as linkage variables. At the physical execution level, in order to resolve the physical conflict between rigid locking and tangential expansion, the command is decomposed into microsecond-level timing: the edge computing control terminal executes the expansion phase, which lasts for about 20-50ms. During this time, the magnetic field is kept closed, and the pressure surge of the air pressure drive unit 200 drives the negative Poisson's ratio expansion skeleton 130 to fully expand tangentially in order to maximize the contact area.

[0080] Upon entering the locked phase, the full-power activated magnetic field instantly solidifies the fluid, locking the unfolded geometry into a rigid shell. High-frequency pulsations are superimposed inside the rigid shell. At this point, the negative Poisson's ratio tensile skeleton 130 plays a crucial role, forcing the wearable monitoring device 100 to unfold tangentially when the air pressure surges, uniformly coupling the generated 200-300Hz high-frequency vibrations to a large area of ​​skin. The 200-300Hz frequency band was chosen because it precisely corresponds to the sensitive peak of the Pacinian bodies deep in human skin. Even under the interference of heavy industrial protective clothing and external machine noise, the vibrations in this frequency band can be clearly identified by the nervous system as an emergency alarm, maximizing the signal-to-noise ratio at the biological perception level.

[0081] In step S1, the magnetorheological fluid layer 120 is composed of soft magnetic particles 121 suspended in a carrier liquid, and the volume fraction of the soft magnetic particles 121 is 20%-40%; in step S1, the low-pressure pneumatic flexible network 110 has a honeycomb cavity structure 111, and the magnetorheological fluid layer 120 fills the gaps in the honeycomb cavity 111.

[0082] This embodiment optimizes and defines the microstructure of the composite material. The low-pressure pneumatic flexible network 110 adopts a honeycomb cavity structure 111. This biomimetic configuration not only provides extremely high structural stability and prevents airway blockage during joint bending, but more importantly, it provides segmented micro-cavities for the magnetorheological fluid, effectively preventing the long-term sedimentation of the soft magnetic particles 121 under gravity and ensuring the uniformity of material properties. Regarding the volume fraction of the soft magnetic particles 121, the selection of 20%-40% is based on the balance between zero-field viscosity and magnetostrictive yield stress: if it is below 20%, the magnetostrictive effect is insufficient to generate effective mechanical braking; if it is above 40%, the basic viscosity of the fluid in the unactivated state is too high, which will significantly increase the physical exertion of the worker. This ratio ensures a large dynamic adjustment range between the imperceptible wearing state in the flexible obstacle avoidance state and the strong intervention in the rigid locking state.

[0083] Example 2:

[0084] Please see Figures 1-3 A deep learning-based unsafe behavior monitoring and early warning device includes:

[0085] The low-pressure aerodynamic flexible network 110 serves as the main driving force field framework and is used to make conformal contact with human skin.

[0086] The pneumatic drive unit 200 is connected to the low-pressure pneumatic flexible network 110 and is used to regulate the internal air pressure and generate high-frequency pressure waves.

[0087] A magnetorheological fluid layer 120, encapsulated in a cavity structure of a low-pressure aerodynamic flexible network 110, is used to provide a variable physical intervention field.

[0088] A magnetic field generator 300 is disposed around the magnetorheological fluid layer 120 to generate a high-frequency alternating magnetic field for controlling the phase change of the fluid.

[0089] A negative Poisson's ratio expansion skeleton 130 is embedded in the airbag wall 112 of the low-pressure aerodynamic flexible network 110 to generate tangential expansion force when the air pressure changes.

[0090] The edge computing control terminal is electrically connected to the pneumatic drive unit 200 and the magnetic field generator 300, respectively, and is used to perform calculations and output control commands for the inverse solution model of the fluid-structure interaction wave equation.

[0091] This edge computing control terminal integrates a high-performance processor specifically designed for running real-time convolutional long short-term memory networks;

[0092] The unsafe behavior monitoring and early warning device also includes a fluid impedance tomography monitoring module 400, which includes a pressure sensor array 410 arranged inside a low-pressure aerodynamic flexible network 110 and an electromagnetic induction probe 420 for monitoring the state of the magnetorheological fluid layer 120.

[0093] This embodiment provides a hardware device architecture for implementing the above method. The device uses a low-pressure pneumatic flexible network 110 as the main driving force field skeleton, which solves the problem of mismatch between traditional rigid exoskeletons and human soft tissue, and achieves conformal contact. The pneumatic drive unit 200 is not only the source of power, but also the signal generator. The high-frequency pressure wave generated by it, together with the pressure sensor array 410 arranged inside the network, belongs to the fluid impedance tomography monitoring module 400, and forms a complete sensing loop. Regarding the hardware details of the high-frequency pneumatic signal generator pointed out in question 2, this embodiment clarifies that the pneumatic drive unit 200 adopts a voice coil motor direct-drive piston structure, rather than a traditional solenoid valve or diaphragm pump. Since conventional air pumps cannot generate high-frequency signals of 200-300Hz, this embodiment utilizes the high-frequency dynamic response characteristics of VCM, with a response frequency >500Hz, to directly drive a lightweight piston to perform micro-amplitude high-frequency reciprocating motion, thereby realizing high-frequency vibration early warning.

[0094] The piston stroke within the pneumatic drive unit 200 is This generates a stable 200-300Hz pressure pulsation in the pneumatic circuit. Simultaneously, to avoid signal attenuation caused by the viscoelastic damping of the flexible pipe, the VCM drive unit is directly coupled to the inlet node of the low-pressure pneumatic flexible network via a short-stroke rigid conduit, with a length controlled within 50mm, ensuring lossless injection of high-frequency energy. The magnetic field generator 300 is typically composed of a flexible coil array, tightly fitted around the magnetorheological fluid layer 120. Addressing the insufficient disclosure of structural features mentioned in Problem 2, this embodiment clarifies the specific construction of the magnetic circuit system: the flexible coil array adopts a planar Archimedean spiral winding topology.

[0095] The topology is designed such that the direction of the generated magnetic field vector is perpendicular to the flow / shear plane of the magnetorheological fluid layer 120, i.e., it penetrates along the thickness direction. This perpendicular magnetic field distribution is crucial because, in the shear mode, the magnetic particles can form the most shear-resistant interlayer particle chain only when the magnetic field lines are perpendicular to the shear direction, thereby generating a sufficiently large shear yield stress to achieve the rigid locking of S4.

[0096] The generator is used to generate a high-frequency alternating magnetic field to control the phase change of the fluid; the edge computing control terminal, as the core of the system, processes data from the pressure sensor array 410 and the electromagnetic induction probe 420 in real time and performs complex inverse calculations; in particular, the embedded design of the negative Poisson's ratio expansion skeleton 130 enables the device to automatically balance the normal pressure through tangential expansion force when performing high-intensity physical interventions, such as high-pressure pulsation alarms, avoiding user pain or skin damage caused by excessive local pressure, and realizing the integrated integration of safety monitoring and physical intervention.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based method for monitoring and early warning of unsafe behaviors, characterized in that, include: S1. A wearable monitoring device (100) is installed, which includes a low-pressure pneumatic flexible network (110) as a basic skeleton, a magnetorheological fluid layer (120) encapsulated in the cavity of the low-pressure pneumatic flexible network (110), and a negative Poisson's ratio expansion skeleton (130) embedded in the airbag wall (112) of the low-pressure pneumatic flexible network (110); the low-pressure pneumatic flexible network (110) is connected to a pneumatic drive unit (200), and the magnetorheological fluid layer (120) is in a controllable magnetic field environment; S2. Start the wearable monitoring device (100) and use the fluid impedance tomography monitoring module (400) to collect high-frequency pressure pulsation spectrum data inside the low-pressure aerodynamic flexible network (110) in real time, and simultaneously collect the inductance or capacitance change rate data of the magnetorheological fluid layer (120). S3. Input the high-frequency pressure pulsation spectrum data and the inductance or capacitance change rate data into the preset fluid-structure interaction wave equation inverse solution model, and calculate the normalized stiffness index of the human body contact part through deep learning algorithm. S4. Based on the comparison between the normalized stiffness index and the preset safety threshold, the wearable monitoring device (100) is subjected to physical field medium regulation; when it is determined to be an unsafe behavior, the PWM duty cycle of the controllable magnetic field and the output frequency of the pneumatic drive unit (200) are regulated in conjunction, and the internal air pressure of the low-pressure pneumatic flexible network (110) is used as a pre-tightening force to cause the magnetorheological fluid layer (120) to undergo a phase change and generate a rigid structure that can transmit 200-300Hz vibration signals; In step S3: the inverse solution model of the fluid-structure interaction wave equation adopts the LSTM-CNN deep neural network architecture, wherein the deep learning algorithm takes the pressure wave reflection coefficient and phase lag as input features and the microscopic isometric contraction rate of human muscles as output features.

2. The method for monitoring and early warning of unsafe behaviors based on deep learning according to claim 1, characterized in that, Before step S1, the steps include: S1.1, by constructing a physical model, with the optimization objectives of maximizing the contact area and homogenizing the local pressure when the low-pressure aerodynamic flexible network (110) is inflated, the structural reentry angle and density parameters of the negative Poisson's ratio tensile skeleton (130) are simulated and calculated, so that the negative Poisson's ratio tensile skeleton (130) generates a lateral expansion force when the air pressure increases to counteract the normal expansion of the airbag.

3. The method for monitoring and early warning of unsafe behavior based on deep learning as described in claim 1, characterized in that, The steps of S4 include: when the normalized stiffness index is in the first range, adjusting the PWM duty cycle of the controllable magnetic field to increase the viscosity of the magnetorheological fluid layer (120) to provide damping; when the normalized stiffness index is in the second range, locking the controllable magnetic field with full power output and adjusting the pneumatic drive unit (200) to generate high-frequency pulsation; during the pressurization process of the pneumatic drive unit (200), the negative Poisson's ratio stretched skeleton (130) forces the low-pressure pneumatic flexible network (110) to expand tangentially to maintain a constant contact area with human skin.

4. The method for monitoring and early warning of unsafe behavior based on deep learning as described in claim 1, characterized in that, In step S1, the magnetorheological fluid layer (120) is composed of soft magnetic particles (121) suspended in a carrier liquid, and the volume fraction of the soft magnetic particles (121) is 20%-40%.

5. The deep learning-based method for monitoring and early warning of unsafe behaviors according to claim 1, characterized in that, In step S4, the 200-300Hz vibration signal corresponds to the sensitive frequency range of the human Pacinian body.

6. The method for monitoring and early warning of unsafe behavior based on deep learning as described in claim 1, characterized in that, In step S1, the low-pressure pneumatic flexible network (110) has a honeycomb cavity (111) structure, and the magnetorheological fluid layer (120) fills the gaps in the honeycomb cavity (111).

7. A deep learning-based unsafe behavior monitoring and early warning device, applied to the deep learning-based unsafe behavior monitoring and early warning method according to any one of claims 1 to 6, characterized in that, include: A low-pressure aerodynamic flexible network (110) serves as the main driving force field framework for conformal contact with human skin. A pneumatic drive unit (200) is connected to the low-pressure pneumatic flexible network (110) and is used to regulate the internal air pressure and generate high-frequency pressure waves. A magnetorheological fluid layer (120), encapsulated in the cavity structure of the low-pressure aerodynamic flexible network (110), is used to provide a variable physical intervention field; A magnetic field generator (300) is disposed around the magnetorheological fluid layer (120) to generate a high-frequency alternating magnetic field for controlling the phase change of the fluid. A negative Poisson's ratio expansion skeleton (130) is embedded in the airbag wall (112) of the low-pressure aerodynamic flexible network (110) to generate tangential expansion force when the air pressure changes; The edge computing control terminal is electrically connected to the pneumatic drive unit (200) and the magnetic field generator (300) respectively, and is used to execute the calculation of the inverse solution model of the fluid-structure interaction wave equation and output control commands.

8. The deep learning-based unsafe behavior monitoring and early warning device according to claim 7, characterized in that, The unsafe behavior monitoring and early warning device also includes a fluid impedance tomography monitoring module (400), which includes a pressure sensor array (410) arranged inside the low-pressure aerodynamic flexible network (110) and an electromagnetic induction probe (420) for monitoring the state of the magnetorheological fluid layer (120).

Images