Method for testing ostomy bag

Abstract

The invention relates to the technical field of medical instrument testing, and discloses an ostomy bag testing method which comprises the following steps: step 1, establishing a thermodynamic equilibrium testing environment, and obtaining environment compensation parameters; step 2, executing test piece clamping based on the environment compensation parameters; 3, dynamic pressure excitation is applied according to the clamp pose data; 4, synchronously acquiring spatial and temporal distribution data of the pressure field, the deformation field and the temperature field; 5, performing fluid-structure interaction field conjoint analysis on the spatial and temporal distribution data; step 6, constructing a defect feature tensor based on the stress-strain features; and step 7, identifying the leakage mode type through the defect feature tensor. By constructing a thermodynamic equilibrium environment and a dynamic pressure excitation system, accurate and synchronous loading of a temperature field and a deformation field is realized, complex load conditions in a real physiological environment are effectively simulated, and the fidelity of test working conditions is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical device testing, in particular to a testing method for an ostomy bag. BACKGROUND

[0002] As an important device in the field of medical care, the sealing performance and service life of the ostomy bag directly affect the quality of life of patients. The existing testing method mainly adopts a combination of static pressure loading and visual inspection, which is difficult to truly simulate the complex dynamic load environment generated by human activities. In the long-term use process, the material of the ostomy bag will degrade due to factors such as temperature change and mechanical stress cycle, and the traditional testing method cannot effectively capture the potential failure risk caused by these time-varying characteristics.

[0003] The current mainstream testing system usually controls temperature, pressure, deformation and other parameters independently, and lacks research on the coupling mechanism of multiple physical fields. The parameter drift caused by environmental disturbance in the testing process is not effectively inhibited, resulting in significant deviation between experimental data and actual working conditions. Especially when simulating physiological activities such as human breathing and movement, the existing method is difficult to generate dynamic load waveforms that meet the biomechanical characteristics, affecting the clinical guidance value of the test results.

[0004] In the prior art, the problem of insufficient control accuracy of environmental parameters is particularly prominent. The temperature fluctuation range of the traditional constant temperature box is large, and it cannot compensate for the change of heat conduction caused by the deformation of the test piece in real time. The pressure loading system mostly adopts an open-loop control strategy, which lacks self-adaptive adjustment ability for the stiffness change of the test piece. This extensive control method leads to significant differences between the test environment and the actual human body environment, making it difficult to accurately evaluate the long-term reliability of the ostomy bag in complex use scenarios. SUMMARY

[0005] The purpose of the present application is to provide a testing method for an ostomy bag, which solves the problems of the existing ostomy bag testing method that cannot truly simulate the complex physiological environment of the human body, is difficult to accurately detect early micro-leakage, and lacks time-varying reliability evaluation.

[0006] To achieve the above purpose, the present application is implemented by the following technical scheme: a testing method for an ostomy bag, comprising the following steps: Step one, establish a thermodynamic equilibrium test environment, and obtain environmental compensation parameters; Step two, based on the environmental compensation parameters, perform test piece clamping, and generate clamp pose data; Step three, apply dynamic pressure excitation according to the clamp pose data, and generate pressure field distribution; Step four, synchronously collect the spatio-temporal distribution data of the pressure field, deformation field and temperature field; Step five, perform fluid-solid coupling field joint analysis on the spatio-temporal distribution data, and output stress-strain characteristics; Step six, constructing a defect feature tensor based on the stress-strain characteristics; Step seven, identifying the leakage mode type through the defect feature tensor; Step eight, verifying the system reliability according to the leakage mode and generating a feedback signal.

[0007] Preferably, the step one comprises: When establishing a thermodynamic equilibrium test environment, perform temperature compensation calibration: In the formula: κ0 is the calibrated thermal diffusivity; α T is the temperature drift compensation factor; ΔT = T env -T set is the environmental temperature difference; β is the gradient history compensation coefficient.

[0008] Preferably, the step two comprises: Performing adaptive adjustment of clamp contact force: In the formula: is the reference contact stiffness; u is the specimen displacement vector; ∈ c is the critical strain threshold; γ is the friction change rate gain coefficient.

[0009] Preferably, the step three comprises: Generating a dynamic pressure excitation function: And performing closed-loop feedback control: In the formula: A i is the amplitude modulation coefficient; is the physiological activity characteristic phase angle.

[0010] Preferably, the step five comprises establishing an anisotropic viscoelastic constitutive model: In the formula: is the fourth-order orthogonal anisotropic stiffness tensor; η is the strain rate sensitivity coefficient; β T is the thermal expansion coupling coefficient; T0 is the reference temperature.

[0011] Preferably, the step five further comprises solving with adaptive time step: And performing parallel computing acceleration: In the formula: ∈ c is the convergence threshold; τmax is the maximum allowed step size; P is the number of parallel computing nodes; is the p-th sub-domain stiffness matrix.

[0012] Preferably, the step six comprises constructing a multi-scale defect feature tensor: wherein: is a wavelet transform basis function; and m is a multi-resolution weight coefficient; s is a differential order; and M is a wavelet decomposition layer number.

[0013] Preferably, the step seven further comprises constructing a spatio-temporal graph attention network: wherein: is a spatio-temporal neighborhood set of node i; and (l) is a trainable weight matrix of the i-th layer; β is a time gradient coupling coefficient; and σ is a ReLU activation function.

[0014] Preferably, the step eight comprises calculating a time-varying reliability index: and performing failure mode identification: wherein: β is a Weibull shape parameter; and η is a characteristic life parameter; is a cumulative stress gradient; and M k is a pre-trained failure mode template tensor; and w k is a mode weight coefficient.

[0015] Preferably, the step eight further comprises dynamic life prediction: wherein: R th is a failure threshold; γ is a reliability decay compensation factor; and t c is a current detection time point.

[0016] In summary, the present application comprises at least one of the following beneficial technical effects: 1. The present application realizes accurate synchronous loading of temperature field, pressure field and deformation field by constructing a thermodynamic equilibrium environment and a dynamic pressure excitation system. The closed-loop compensation algorithm is used to eliminate environmental disturbance, and the biological characteristic waveform synthesis technology is combined to effectively simulate complex load conditions under real physiological environment, thereby significantly improving the fidelity of test working conditions.

[0017] 2. The leakage mode identification method based on the space-time diagram attention network of the application fuses multi-scale differential features and wavelet energy information, and breaks through the limitations of traditional threshold methods. The local abnormal propagation characteristics are captured through the space-time neighborhood attention mechanism, and the accurate positioning and classification of early micro leakage are realized.

[0018] 3. The application adopts parallel fluid-solid coupling solving algorithm and adaptive grid technology, and the contact stiffness adaptive adjustment function of the intelligent clamping system is realized under the premise of ensuring the calculation accuracy, the number of manual intervention is reduced, and the automation level of the test process is improved.

[0019] 4. The application constructs a multi-source data space-time coding system, converts the complex physical field evolution process into an interpretable engineering parameter through tensor decomposition and feature visualization technology, supports multi-dimensional correlation analysis of test results, and provides clear guidance for material improvement and structure optimization. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The figure is a schematic diagram of the method of the application. DETAILED DESCRIPTION

[0021] The following will be combined with the Figure 1 The application will be further described in detail.

[0022] The application provides a test method for a stoma bag, which realizes accurate synchronous loading of temperature field, pressure field and deformation field by constructing a thermodynamic equilibrium environment and a dynamic pressure excitation system. The closed-loop compensation algorithm is used to eliminate environmental disturbance, and the biological feature waveform synthesis technology is combined to effectively simulate the complex load conditions in the real physiological environment, thereby significantly improving the fidelity of the test working condition.

[0023] As Figure 1 shown, the test method for the stoma bag can include the following steps: Step 1, establish a thermodynamic equilibrium test environment, and obtain environmental compensation parameters; Step 2, perform specimen clamping based on the environmental compensation parameters, and generate clamp pose data; Step 3, apply dynamic pressure excitation according to the clamp pose data to generate pressure field distribution; Step 4, synchronously collect space-time distribution data of the pressure field, deformation field and temperature field; Step 5, perform fluid-solid coupling field joint analysis on the space-time distribution data, and output stress-strain characteristics; Step 6, construct a defect feature tensor based on the stress-strain characteristics; Step 7, identify the leakage mode type through the defect feature tensor; Step 8, verify the system reliability according to the leakage mode and generate a feedback signal.

[0024] In this embodiment, precise control of the testing environment is achieved by constructing a closed-loop thermodynamic equilibrium system. The establishment of the thermodynamic equilibrium testing environment includes three core modules: distributed temperature sensor array arrangement, dynamic compensation algorithm design, and closed-loop control of environmental parameters. The distributed temperature sensor array is arranged in a three-dimensional mesh topology within the testing chamber, with the spacing between adjacent sensors not exceeding 1 / 5 of the specimen's characteristic dimension. Each sensor collects ambient temperature data T in real time. env The temperature field distribution T(x,y,z,t) is reconstructed using a spatial interpolation algorithm, where an improved Kriging algorithm is employed for spatial interpolation. In the formula, the interpolation weight coefficient λ i Satisfying the unbiasedness condition ∑λ i =1 and the semivariogram γ(h) is fitted using an exponential model.

[0025] Temperature compensation calibration is achieved through a multiphysics coupling model, specifically by performing the following calculation process: First, obtain the set temperature T. set With ring ambient temperature T env The real-time temperature difference ΔT = T env -T set Then, the dynamic compensation value of the thermal diffusivity is calculated: The reference thermal diffusivity κ0 was obtained through calibration experiments on standard specimens under steady-state conditions, and the temperature drift compensation factor α... T It exhibits a linear correlation with the material's coefficient of thermal expansion. The integral term in the gradient history compensation term characterizes the cumulative effect of spatial non-uniformity in the temperature field, and the compensation coefficient β is dynamically adjusted based on the test chamber volume and heat flow rate. The environmental control module employs a feedforward-feedback composite control strategy, and its control law can be expressed as: The feedforward term in the formula Obtained through real-time solution of the thermal flow field perturbation model, specifically including: Convectional heat transfer disturbance caused by specimen clamping; The adiabatic compression heat term generated by the pressure excitation system; The Joule thermal term of the data acquisition equipment. Each disturbance term is estimated online using a pre-established transfer function model.

[0026] Preferably, the temperature sensor array contains at least three measurement planes, and the sensors in each plane are distributed in an equilateral triangle. In the temperature field reconstruction process, the mirror virtual node method is used to process the boundary area to eliminate the reconstruction error caused by the edge effect. The dynamic compensation algorithm performs iterative calculation every 10 milliseconds to ensure that the fluctuation range of the environmental parameters is controlled within ±0.3°C.

[0027] In this embodiment, the acquisition of the environmental compensation parameters includes the following key steps: first, the reference parameter calibration is performed in the system preheating stage to obtain the initial value of κ0; then in the dynamic test stage, the time integral term of the temperature gradient is calculated by the sliding window method, and the window length is adaptively adjusted according to the thermal relaxation time; finally, the compensated thermal diffusion coefficient is input into the environmental control module to drive the heating / cooling unit to perform power regulation.

[0028] In this embodiment, the contact state between the test piece and the fixture is optimized through an intelligent clamping system. The test piece clamping process includes three core links of multi-degree-of-freedom pose adjustment, contact stiffness dynamic matching, and clamping force closed-loop control.

[0029] The clamping system drives the fixture actuator by a six-degree-of-freedom parallel mechanism, and its kinematics model can be represented as: where q is the joint space coordinate vector, X d = [x, y, z, α, β, γ] T is the target pose in the Cartesian space. During the pose adjustment process, the laser displacement sensor is used to monitor the surface topography of the test piece in real time to generate a displacement compensation vector Δu.

[0030] The contact force adaptive adjustment is realized based on a nonlinear stiffness model, and its update strategy is: where the reference contact stiffness is calculated according to the elastic modulus of the test piece material, and satisfies where A c is the contact area, and h0 is the initial clamping gap. The hyperbolic tangent function term realizes the smooth transition of the contact stiffness, and the critical strain threshold ∈ c is related to the yield strength of the material, and is determined by ∈ c = 0.8 ∈ yield .

[0031] Preferably, the gain coefficient γ in the friction force change rate term is determined by pre-calibration method: a step change normal load is applied at the end of the clamping system, the friction force response curve is collected, and the least squares method is used to fit to obtain γ = k1·Ra + k2, where Ra is the surface roughness parameter of the test piece, and k1, k2 are material property related coefficients.

[0032] The displacement vector u is obtained by multi-sensor data fusion: the laser displacement sensor array measures the local deformation, the six-dimensional force sensor collects the contact force components, and the Kalman filter is used for data fusion: where the observation matrix H is established according to the geometry of the fixture, and the Kalman gain K is updated online to optimize the estimation accuracy. The differential term of the displacement vector is calculated by the five-point central difference method: In this embodiment, the clamping force closed-loop control executes the following process: first, generate the initial clamping path according to the specimen CAD model, then dynamically adjust the fixture pose to avoid local over-stress by real-time monitoring of the contact stress distribution. When the contact stress is detected to exceed the threshold σ max = 0.7σ yield , trigger the safety protection mechanism, back off the fixture and re-plan the path.

[0033] Preferably, the calculation of the contact area A c uses the Hertz contact theory correction model, considering the influence of the specimen surface curvature radius and the fixture geometry. The clamping gap h0 is measured by a laser range finder in the initial positioning stage, with a measurement accuracy better than 0.01 mm. The stiffness adjustment algorithm performs iterative calculation every 5 milliseconds to ensure the stability of the contact state.

[0034] In this embodiment, dynamic load application is achieved through multi-frequency complex pressure excitation and adaptive control. The generation of dynamic pressure excitation includes three core links: physiological characteristic waveform synthesis, pressure field closed-loop regulation and anti-interference compensation.

[0035] The dynamic pressure excitation function adopts an exponential decay type multi-frequency modulation model, and its mathematical expression is: where the base pressure P0 is determined according to the rated working pressure of the specimen, satisfying P0 = 1.2P works . The amplitude modulation coefficient A i ∈ [0.1, 0.5] is distributed according to the equal ratio sequence, satisfying A i+1 = 0.8A i . The characteristic frequency f i corresponds to the base frequency of human physiological activity: f1 = 0.3 Hz simulates the respiratory rhythm, f2 = 1.2 Hz represents the heartbeat period, and f3 = 2.5 Hz reflects the intestinal rumbling frequency.

[0036] The physiological activity characteristic phase angle The real-time biological signal acquisition system is obtained, specifically including: the R-R interval is extracted after the ECG signal is filtered by a band-pass filter, the respiratory signal is measured by the impedance method to measure the change of the thoracic volume, and the intestinal sound is collected by a contact accelerometer. The phase angle calculation formula is: Where s i (t) is the pre-processed biological signal time domain waveform, the FFT operation adopts a sliding window strategy, and the window length matches the excitation frequency.

[0037] The closed-loop feedback control system adopts a gain scheduling PID algorithm, and the control law can be decomposed as: Where the reference gain k0 is calibrated by a step response experiment, and the temperature gradient term Real-time reflects the surface heat flow distribution state of the test piece. The integral time constant τ i is related to the thermal inertia of the system, and satisfies τ i = 0.5t thermal , where t thermal is the thermal relaxation time of the test piece.

[0038] Preferably, the attenuation coefficient α is dynamically adjusted according to the pressure maintenance time requirement, and when the test piece deformation rate is detected to exceed the threshold value, the value of λ is automatically increased to accelerate the pressure decay. The five-point central difference method is used for the calculation of the differential term: In this embodiment, the pressure field distribution monitoring is realized by a ring-shaped pressure sensor array, the sensor interval angle is 120 degrees, and the sampling rate is not less than 10 times the excitation frequency. The data acquisition system adopts a synchronous triggering mechanism to ensure the phase consistency of the time domain waveform. When the local pressure deviation is detected to exceed ±5%, the gain rescheduling mechanism is triggered, and K p parameters are updated and the integral term is reset.

[0039] The anti-interference compensation module includes two core measures: A comb filter is designed to suppress characteristic frequency noise for hydraulic system pulsation interference; A pressure-volume compensation model is established for the volume change caused by the deformation of the test piece: Where the fluid bulk modulus β is determined according to the characteristics of the working medium, and the reference volume V0 is calibrated in the system initialization stage.

[0040] In this embodiment, the time and space synchronous acquisition of the pressure field, deformation field and temperature field is realized through a multi-modal sensing network. The data acquisition system includes three core modules of distributed sensing array arrangement, high-precision synchronous triggering mechanism and multi-source data fusion processing.

[0041] The pressure field measurement adopts a MEMS piezoresistive sensor array arranged in an equidistant grid topology on the contact surface of the test piece. The sensor node output signal is conditioned by an adaptive gain amplifier circuit, and the pressure field reconstruction is realized by the following formula: where the spatial weight coefficient w i is dynamically adjusted according to the distance between adjacent nodes, and satisfies ∑w i = 1. The Gaussian kernel function parameter σ p is set to 1 / 2 of the node spacing, ensuring the spatial continuity of the pressure field. The deformation field measurement is realized based on a digital image correlation (DIC) system, and its three-dimensional displacement calculation model is: where Ω is the local calculation window, and I1 and I2 are the speckle images captured by the left and right cameras. Preferably, the speckle pattern is designed with a non-periodic random distribution, with a feature point density of not less than 20 points / mm 2 , and the matching algorithm uses a reverse combination Gaussian-Newton optimization. The temperature field monitoring is realized by fusing an infrared thermal imager and an embedded thermocouple array, and the temperature field reconstruction formula is: where the fusion coefficient α is dynamically adjusted according to the spatial position, and α = 0.3 is taken in the vicinity of the thermocouple, and α = 0.8 is taken in the far field. The kernel function K(d) adopts a cubic spline interpolation form, satisfying K(0) = 1 and K(D) = 0 (D is the influence radius).

[0042] The synchronous triggering mechanism adopts a hierarchical clock distribution architecture: the main controller generates a 5V TTL synchronous pulse signal, which is distributed to each acquisition subsystem through a star topology network. Timestamp alignment is realized by the following strategies: Hardware-level synchronization: each subsystem is equipped with a constant temperature crystal oscillator clock source, with a frequency stability better than ±1ppm Protocol-level synchronization: sub-microsecond network time synchronization is performed using the IEEE 1588 Precision Time Protocol (PTP) Data-level synchronization: the acquisition data header information includes a GPS time code and a sampling period counter Preferably, the data preprocessing module performs the following key operations: Temperature drift compensation: T corr = T raw -k T ·(t-t0), where the drift coefficient k T is calibrated through preheating experiments; Strain filtering: a zero-phase Butterworth filter is used, with a transfer function of: Pressure signal de-noising: based on wavelet threshold de-noising algorithm, the threshold selection rule meets: Where σ is the noise variance estimate, and N is the signal length.

[0043] In this embodiment, the space-time data encoding adopts a four-dimensional tensor storage format, and the dimensions are composed of: X-axis spatial coordinate x Y-axis spatial coordinate x time axis x physical quantity channel. The data compression algorithm adopts multilinear principal component analysis based on Tucker decomposition: Where the core tensor g retains more than 95% of the energy information, and the factor matrix U (i) corresponds to the characteristic basis vector of each dimension. Preferably, the physical quantity channel is arranged in the order of pressure-deformation-temperature, and each channel data is normalized before storage: The scheme realizes accurate synchronous acquisition and efficient storage of multi-physical field data through the above technical means, and provides a complete input data basis for subsequent fluid-solid coupling analysis.

[0044] In this embodiment, the anisotropic viscoelastic constitutive model is established to realize the joint numerical analysis of the fluid-solid coupling field. The fluid-solid coupling field analysis includes material constitutive modeling, nonlinear equation system solving, and parallel computing acceleration.

[0045] The anisotropic viscoelastic constitutive model decomposes the stress tensor into elastic, viscous, and thermal coupling terms: Where the fourth-order orthogonal anisotropic stiffness tensor satisfies the material principal axis symmetry condition C ijkl =C jikl =C ijlk , and its independent components are obtained through ultrasonic anisotropy detection experiment calibration. The strain rate sensitivity coefficient η is described by a power law model: Where the reference strain rate is 1x10 -3 s -1 , and the exponential factor n is determined by dynamic mechanical analysis (DMA) frequency scanning experiment. The adaptive time step control algorithm is realized based on strain energy error estimation, and its update strategy is: Where the convergence threshold ∈ c is dynamically adjusted according to the material creep characteristics, and satisfies ∈ c =0.1%∈yield The strain increment norm ||Δ∈|| is calculated by comparing the residuals of the current step with the historical steps: Preferably, the parallel computation acceleration adopts a non-overlapping domain decomposition strategy, dividing the calculation domain into P sub-regions, and each sub-region stiffness matrix and load vector satisfy: Wherein the interface of the sub-region adopts the Lagrange multiplier method to process the compatibility condition, and the interface constraint equation can be expressed as: Γ u Γ =0; In the formula, the constraint matrix B Γ is constructed according to the mapping relationship of adjacent sub-region nodes, and u Γ is the interface displacement vector. In this embodiment, the Newton-Raphson iteration method is adopted for solving the nonlinear equation set, and the Jacobian matrix updating strategy is: When the residual reduction rate γ=||R (k) || / ||R (k-1) ||>0.7, the matrix reconstruction is triggered; When the quasi-Newton method is adopted, the maximum approximate iteration number is limited to 3 times; The line search parameter setting satisfies the Wolfe condition to ensure global convergence.

[0046] The coefficient β T in the thermal expansion coupling term is determined by the thermal mechanical analysis (TMA) experiment, and the temperature dependence is described by a piecewise linear interpolation model: Wherein H(·) is the Heaviside step function, and the turning temperature T m is determined according to the phase transition point of the material. Preferably, the reference temperature T0 takes the measured value of the ambient temperature at the clamping completion. The scheme realizes efficient solution of the fluid-structure coupling problem under complex load through the above technical means, and provides accurate stress-strain field data basis for defect feature extraction.

[0047] In this embodiment, the quantitative characterization of the defect feature is realized through the multi-scale space-time tensor fusion technology. The construction of the defect feature tensor includes three core links of differential field calculation, wavelet multi-resolution analysis and tensor compression.

[0048] The mathematical expression of the multi-scale defect feature tensor is: Wherein the selection of the differential order s is based on the damage evolution characteristics of the specimen, when s=1, the linear change trend is represented, s=2 captures the acceleration characteristics, and s=3 reflects the mutation characteristics. The tensor product operation realizes the pressure field The space-time coupling of displacement field u. The time derivative calculation adopts adaptive difference format, using central difference for stationary signal segment: where difference coefficient Determined by Taylor expansion matching, automatically switch to one-sided difference format when signal mutation is detected to avoid Gibbs phenomenon. Spatial gradient operator Adopt least square weighted reconstruction method to calculate: Weight coefficient w i Exponential decay with distance, Decay coefficient σ g Take local grid characteristic length. Wavelet transform base function Adopt Daubechies wavelet family to realize multi-resolution analysis, whose scale function and mother wavelet satisfy: φ j,k (x)=2 j / 2 φ(2 j x-k); ψ j,k (x)=2 j / 2 ψ(2 j x-k); The number of decomposition layers M is determined according to Nyquist sampling theorem, satisfying M≤log2(N / 2), where N is the signal length. Weight coefficient α m Adopt energy entropy criterion to allocate: where E m is the energy of the mth layer wavelet coefficient, The term is used to suppress the influence of high-frequency noise. Preferably, tensor compression adopts the strategy of combining Tucker decomposition and sparse coding. First, the original tensor is decomposed into rank (R1, R2, R3, R4): Then, the core tensor is executed K-SVD dictionary learning to obtain sparse representation: Where dictionary atom d p is obtained by alternating optimization algorithm training, and sparse coefficient λ p satisfies ||λ||0≤S (non-zero element number limit).

[0049] In this embodiment, the construction process of the feature tensor includes: firstly, normalizing and preprocessing the original pressure and displacement field data to eliminate the dimensional difference; then calculating the differential items and wavelet coefficients of each order in parallel; and finally generating a comprehensive feature representation by tensor splicing and weighted fusion. Preferably, the upper limit of the differential order is adaptively determined according to the Shannon entropy criterion, and the calculation is terminated when the entropy change rate caused by the newly added order is less than 5%.

[0050] The extraction of the spatio-temporal correlation feature provides a key input for subsequent leakage mode recognition, and effectively enhances the robustness and interpretability of defect representation by fusing the differential features of multiple physical fields and wavelet detail information. In this embodiment, the intelligent recognition of the leakage mode is realized through a spatio-temporal graph attention network. The leakage mode recognition process includes three key technical links: graph structure modeling, attention mechanism design, and time gradient fusion.

[0051] The node update formula of the spatio-temporal graph attention network is: wherein represents the spatio-temporal neighborhood of node i, defined as all nodes within a spatial radius r s and a time window [t-Δt, t+Δt]. The attention coefficient α ij is calculated through the following mechanism: αij=softmax(LeakyReLU(a T [Wh i ||Wh j ])); wherein the trainable parameter vector implements similarity measurement in the feature space, and || represents the vector splicing operation. Preferably, the spatial neighborhood radius r s is adaptively adjusted according to the size of the test piece, satisfying r s =0.1L max , wherein L max is the maximum feature length of the test piece. The time gradient term is calculated using a high-order difference format: The gradient coupling coefficient β is dynamically adjusted through a learnable parameter matrix, satisfying to realize adaptive fusion of features and gradients.

[0052] The network training adopts a multi-task learning framework to jointly optimize the mode classification and positioning tasks: The classification branch output layer uses a softmax activation, and the loss function is cross-entropy: The branch output layer uses sigmoid activation, and the loss function is Dice loss: The total loss function is the weighted sum: where the weight coefficient λ linearly decays with training rounds, and the initial value is set to 0.7.

[0053] Preferably, the node feature initialization includes the following physical field information: Mean and variance of pressure field; Frobenius norm of displacement gradient tensor; Time derivative of temperature field; Wavelet energy entropy feature.

[0054] The edge weight of the graph structure defines the fusion of spatial proximity and physical field similarity: where σ s controls the decay rate of the space, and the cosine similarity term cos(·) measures the correlation between the node feature vectors f i ,f j .

[0055] In this embodiment, the network architecture includes 4 layers of spatio-temporal attention modules, and the output dimensions of each layer are 64, 128, 256, and 64 in turn. The skip connection mechanism splices the shallow features with the deep features to preserve the multi-scale information. Preferably, the Dropout layer is randomly inactivated with a probability of 0.3, and the batch normalization layer accelerates the training convergence.

[0056] The introduction of the time gradient term effectively enhances the network's ability to capture transient leakage features, and through the fusion of dynamic physical field information, it improves the recognition sensitivity to early micro-leakage.

[0057] In this embodiment, system verification and feedback are achieved through a time-varying reliability model and failure mode identification. The reliability verification process includes three core modules: Weibull life modeling, pattern template matching, and dynamic life prediction.

[0058] The calculation formula of the time-varying reliability index is: where the shape parameter β and the characteristic life η are identified online by the maximum likelihood estimation method, satisfying the likelihood function: The cumulative stress gradient term is calculated by numerical integration, and the adaptive Simpson rule is used to control the integral error to be less than 1%. The failure mode identification function uses a logistic regression model: Among them, the pre-trained template tensor M k Obtained through K-SVD dictionary learning, satisfying: In the formula The set of feature tensors representing historical failure cases is used, with the number of non-zero elements in the sparse coefficient matrix X limited to S = 5. The trace operation Tr(·) measures the similarity between the feature tensor and the template tensor. The dynamic lifetime prediction model considers the impact of reliability degradation rate. Where the reliability threshold R th Based on the safety level setting, the attenuation compensation factor γ is estimated using the sliding window method: Preferably, the feedback signal generation module executes the following logic: When R(t) > 0.9R th When the signal is green, a normal signal is output. When 0.7R th <R(t)≤0.9R th At that time, a yellow alert was triggered; When R(t)≤0.7R th At that time, a red alarm is generated and an emergency protocol is activated.

[0059] In this embodiment, the mode weight coefficient w k The update mechanism integrates online learning with expert knowledge: The online learning portion employs stochastic gradient descent with a learning rate η = 0.01. Expert knowledge is injected through a fuzzy rule base, with the rule format as follows: Weight normalization processing: The confidence level of the lifetime prediction results was assessed using Bootstrap resampling, and the standard deviation σ of 100 sampled predictions was calculated. t When σ t >0.1t fail A manual review request is triggered at that time.

[0060] The feedback signal encoding follows the ISO13374 standard, and the data packet structure includes: 4-byte timestamp (UNIX format); 2-byte reliability metrics (fixed-point count, accuracy 0.01%); 1-byte failure mode encoding; 4-byte predicted remaining lifetime (in seconds); 1-byte alarm level.

[0061] The scheme realizes multi-dimensional evaluation of the reliability of the ostomy bag by fusing the physical failure model and the data-driven method, and provides quantitative basis for product improvement and maintenance decision.

[0062] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method of testing an ostomy bag, characterized by, The method comprises the following steps: Step one, establishing a thermodynamic equilibrium test environment, obtaining environmental compensation parameters; Step two, based on the environmental compensation parameters, performing test piece clamping, generating clamp pose data; Step three, according to the clamp pose data, applying dynamic pressure excitation, generating pressure field distribution; Step four, synchronously collecting the space-time distribution data of pressure field, deformation field and temperature field; Step five, performing fluid-solid coupling field joint analysis on the space-time distribution data, and outputting stress-strain characteristics; Step six, constructing defect feature tensor based on the stress-strain characteristics; Step seven, identifying leakage mode type through the defect feature tensor; Step eight, verifying system reliability according to the leakage mode and generating feedback signal.

2. A method of testing an ostomy bag according to claim 1, characterised in that, The step one comprises: When establishing a thermodynamic equilibrium test environment, performing temperature compensation calibration: where: k0 is the calibrated thermal diffusivity; a T is the temperature drift compensation factor; ΔT = T env -T set is the ambient temperature difference; β is the gradient history compensation factor.

3. A method of testing an ostomy bag according to claim 1, characterised in that, The step two comprises: Performing clamp contact force self-adaptive adjustment: In the formula: K is the reference contact stiffness; u is the specimen displacement vector; ∈ c is the critical strain threshold; γ is the friction change rate gain coefficient.

4. A method of testing an ostomy bag according to claim 1, characterised in that, The step three comprises: Generating dynamic pressure excitation function: And performing closed-loop feedback control: In the formula: A i is the amplitude modulation coefficient; is the phase angle of the physiological activity characteristic.

5. A method of testing an ostomy bag according to claim 1, characterised in that, The step five comprises establishing an anisotropic viscoelastic constitutive model: where: is the fourth order orthotropic stiffness tensor; η is the strain rate sensitivity coefficient; β T is the thermal expansion coupling coefficient; To is the reference temperature.

6. A method of testing an ostomy bag according to claim 1, characterised in that, The step five further comprises solving by using adaptive time step: And performing parallel computing acceleration: where: ∈ c is a convergence threshold; τ max is a maximum allowed step size; P is the number of parallel computing nodes; is the stiffness matrix of the p-th subdomain.

7. A method of testing an ostomy bag according to claim 1, characterised in that, The step six comprises constructing multi-scale defect feature tensor: In the formula: is a wavelet transform base function; α m is a multi-resolution weight coefficient; s is a differential order; and M is a wavelet decomposition layer number.

8. A method of testing an ostomy bag according to claim 1, characterised in that, The step seven further comprises constructing space-time graph attention network: In the formula: is the spatiotemporal neighborhood set of node i; W (l) is the Ith layer trainable weight matrix; β is the time gradient coupling coefficient; and σ is the ReLU activation function.

9. A method of testing an ostomy bag according to claim 1, characterised in that, The step eight comprises: Calculating time-varying reliability index: where β is a Weibull shape parameter; η is a characteristic lifetime parameter; is the cumulative stress gradient; M k is a pre-trained failure mode template tensor; w k is a mode weight coefficient.

10. A method of testing an ostomy bag according to claim 1, characterised in that, And performing failure mode identification: The step eight further comprises dynamic life prediction: wherein: R th is a failure threshold; γ is a reliability decay compensation factor; t c is the current detection time point.

Images