A fabric garment dynamic performance prediction and evaluation system and method based on multi-dimensional sensing

Abstract

This invention discloses a system and method for predicting and evaluating the dynamic performance of fabrics and garments based on multi-dimensional sensing. The system includes: a control module for generating composite motion commands to drive an actuator to apply a composite dynamic load, including axial tension, radial torsion, and normal impact, to a fabric sample; a multi-modal sensing module for simultaneously acquiring image signals, acoustic signals, and mechanical signals from the fabric sample; a feature extraction module for preprocessing the multi-modal signals and extracting quantized feature vectors; an AI prediction module with a built-in deep learning model for inputting the quantized feature vectors and outputting the predicted dynamic performance of the garment; and an output module for visualizing the prediction results. This invention can simulate the complex mechanical states exerted on fabrics by real human activity, achieve objective quantification of subjective indicators such as feel and style, and accurately predict garment performance through a deep learning model. It can be widely applied in fabric research and development, garment design, and quality control.

Description

Technical Field

[0001] This invention belongs to the field of textile testing technology based on artificial intelligence, and specifically relates to a system and method for predicting and evaluating the dynamic performance of fabrics and garments. Background Technology

[0002] With the rapid development of the textile industry and garment engineering technology, fabric performance testing and garment quality assessment have become crucial links in the garment supply chain. However, there is a significant disconnect between the existing fabric testing system and the actual application needs of garments, mainly in the following aspects: Traditional fabric testing instruments (such as tensile strength testers, abrasion testers, and drape testers) mostly adopt "static" or "unidirectional" testing modes. For example, tensile testing usually only involves uniaxial stretching along the warp or weft direction; abrasion testing usually only involves friction along a single trajectory on a plane; current testing systems mainly rely on mechanical sensors to collect "force-displacement" data, and for key sensory indicators that influence consumer purchasing decisions—such as "hand feel" (softness, smoothness) and "style" (drape, firmness)—they often rely on subjective human touch evaluation, lacking objective quantitative standards. Existing test results are usually a set of isolated physical parameters (such as breaking strength 300N, air permeability 500mm / s), which are mainly used for quality control (QC) rather than design assistance; existing equipment is mostly a simple data acquisition tool, lacking built-in intelligent algorithms to deeply mine massive amounts of test data.

[0003] In conclusion, developing a fabric evaluation system that can simulate complex mechanical states, integrate multimodal sensing technology, and possess AI prediction capabilities for garment performance is an urgent need to address the aforementioned industry pain points and promote the intelligentization of textile testing. Summary of the Invention

[0004] The purpose of this invention is to provide a system and method for predicting and evaluating the dynamic performance of fabrics and garments based on multidimensional sensing, aiming to solve the following problems existing in the prior art: traditional fabric testing modes are singular and cannot simulate complex dynamic mechanical states; key sensory indicators such as hand feel and style lack objective quantitative standards; test data are isolated and lack intelligent analysis and garment performance prediction capabilities.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A system for predicting and evaluating the dynamic performance of fabric garments based on multidimensional sensing, comprising:

[0007] The control module is used to generate composite motion commands according to a preset scene mode. The composite motion commands are used to drive the actuator to apply a composite dynamic load to the fabric sample, which includes at least two of the motions of axial tension, radial torsion and normal impact.

[0008] A multimodal sensing module, comprising at least two of a visual sensor, an acoustic sensor, and a mechanical sensor, is used to simultaneously acquire image signals, acoustic signals, and mechanical signals of a fabric sample under the combined dynamic load.

[0009] The feature extraction module, connected to the multimodal sensing module, is used to preprocess the acquired multimodal signals and extract quantized feature vectors characterizing the physical properties of the fabric. The quantized feature vectors include at least two of the following: visual texture features, acoustic tactile features, and dynamic mechanical features.

[0010] The AI ​​prediction module has a built-in trained deep learning model, which takes the quantized feature vector as input and outputs the prediction result of the dynamic performance of the garment after model inference.

[0011] The output module is used to present the prediction results in a visual form.

[0012] Furthermore, the control module includes:

[0013] The trajectory generation unit is used to generate trajectories based on a preset scene pattern using formulas. Calculate the composite displacement vector applied to the fabric. ,in, These represent unit vectors in three dimensions. It is a periodic stretching function used to simulate the reciprocating motion of limbs; The torsion angle function is used to simulate shear deformation; The pulse function generated by the normal tapping; parameters These are the weighting coefficients for different testing scenarios;

[0014] A PID controller is used to measure the difference between the target tension and the measured tension. ,pass Output correction amount to maintain constant pretension during the test, wherein, , , These are the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller, respectively.

[0015] Furthermore, the multimodal sensing module also includes a preprocessing unit, which is used for:

[0016] Adaptive histogram equalization (CLAHE) and edge feature extraction are performed on the image signals acquired by the vision sensor;

[0017] The acoustic signal acquired by the acoustic sensor is subjected to short-time Fourier transform (STFT) and noise reduction is performed using spectral subtraction.

[0018] The mechanical signals acquired by the mechanical sensor are smoothed using a Savitzky-Golay digital filter.

[0019] Furthermore, the visual texture features extracted by the feature extraction module include at least the contrast calculated based on the gray-level co-occurrence matrix (GLCM). Entropy The contrast is used to characterize the surface roughness of the fabric, and the entropy is used to quantify the severity of pilling and fuzzing of the fabric.

[0020] Furthermore, the acoustic tactile features extracted by the feature extraction module are Mel frequency cepstral coefficients (MFCCs), which are filtered by a Mel-scale filter bank. It was obtained by simulating the characteristics of human hearing.

[0021] Furthermore, the dynamic mechanical features extracted by the feature extraction module include at least the energy loss in a single stretch-recovery cycle. ,in, For stress, Strain is used to characterize the elastic recovery ability of a fabric.

[0022] Furthermore, the deep learning model in the AI ​​prediction module is a CNN-LSTM hybrid network architecture, and its input vector... From the perspective of vision ,acoustics and mechanics Features pieced together: ;

[0023] The deep learning model also includes an adaptive modal attention mechanism, which automatically assigns weights to different modal features based on the prediction target. and through Feature fusion is performed, where, The fused feature vector is represented by Dense(·), which indicates processing by a fully connected layer.

[0024] Furthermore, the output module includes:

[0025] The similarity retrieval unit is used to calculate the feature vector of the fabric to be tested in the feature space. Compared with the feature vector of the standard sample in the database cosine similarity It also outputs style similarity conclusions.

[0026] The time-series prediction unit is used to recursively predict the performance evolution curve of the fabric after long-term use based on short-term test data using the LSTM network in the AI ​​prediction module.

[0027] Furthermore, the prediction results generated by the output module are visualized as a generative evaluation report, which includes at least one of the following information: fabric performance score, percentage of similarity to the standard sample, performance degradation curve throughout the entire life cycle, and appearance rating conclusion.

[0028] This invention also provides a method for predicting and evaluating the dynamic performance of fabrics and garments based on multidimensional sensing, applied to the aforementioned system, comprising the following steps:

[0029] Step 1: Generate composite action commands through the control module to drive the actuator to apply composite dynamic loads to the fabric sample;

[0030] Step 2: Simultaneously acquire image signals, acoustic signals, and mechanical signals of the fabric sample under composite dynamic load using a multimodal sensing module, and perform preprocessing.

[0031] Step 3: The preprocessed signal is mathematically transformed by the feature extraction module to extract a quantized feature vector containing visual texture features, acoustic tactile features and dynamic mechanical features.

[0032] Step 4: Input the quantized feature vector into the deep learning model of the AI ​​prediction module, and output the prediction result of the dynamic performance of the garment through model inference.

[0033] Step 5: Present the prediction results in a visual form through the output module.

[0034] Beneficial Effects: This invention's multi-dimensional sensing-based fabric and garment dynamic performance prediction and evaluation system is a standardized digital foundation technology that helps break down data barriers between fabric R&D, garment design, and end-sales. It provides the textile industry with a new intelligent manufacturing tool, driving the industry's transformation from an experience-based traditional model to a data-driven intelligent model. Significantly Reduces R&D Costs and Promotes Sustainable Development in the Textile Industry: In traditional garment R&D processes, verifying fabric performance requires repeated iterations of "sampling-fitting-modification," consuming significant manpower and resources and generating waste. This invention, through precise predictive capabilities, allows companies to anticipate garment effects at the fabric sample stage, drastically reducing the number of physical sample garment productions. This not only shortens the product launch cycle and reduces company costs but also reduces textile material waste, aligning with the social needs of green manufacturing and sustainable development. Enhances Consumer Experience and Supports the Development of E-commerce and Virtual Fitting Technology: This invention can output highly objective "feel" and "style" data (such as the percentage similarity to silk). With the increasing popularity of e-commerce and virtual try-on in the apparel industry, this data can be transformed into digital descriptions that consumers can perceive, helping them to understand product characteristics more accurately without being able to touch the physical item, thereby reducing return rates and improving the online shopping experience. Attached Figure Description

[0035] Figure 1 This is a flowchart of a method for predicting the properties of dynamic fabrics. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] This invention provides a system for predicting and evaluating the dynamic performance of fabrics and garments based on multi-dimensional sensing, comprising: a control module, a multi-modal sensing module, a feature extraction module, an AI prediction module, and an output module.

[0038] The control module generates composite action commands based on preset scene modes, driving the actuator to apply a composite dynamic load, including axial tension, radial torsion, and normal impact, to the fabric sample. The multimodal sensing module, comprising visual, acoustic, and mechanical sensors, simultaneously acquires image, acoustic, and mechanical signals from the fabric sample under the composite dynamic load. A feature extraction module, connected to the multimodal sensing module, preprocesses the acquired multimodal signals and extracts quantized feature vectors characterizing the fabric's physical properties. The AI ​​prediction module incorporates a trained deep learning model, using the quantized feature vectors as input and outputting a prediction of the garment's dynamic performance after model inference. The output module presents the prediction results in a visual format.

[0039] Based on the above system, this invention proposes a dynamic prediction method for fabric and garment performance based on multidimensional physical field coupling and multimodal deep learning. This method aims to solve the problem that existing single physical index tests cannot reflect the complex dynamic performance of garments. The specific implementation process of this method includes the following five closely linked steps:

[0040] Step 1: Construct a nonlinear composite force action model (control layer):

[0041] Real human activities (such as walking, bending over, and turning) often exert multidirectional and time-varying forces on fabrics. To reproduce this complex state in a laboratory environment, this step does not use simple constant-speed stretching, but instead uses an algorithm to generate coordinated control commands for multi-degree-of-freedom motors to drive the test fixture to perform compound movements.

[0042] Parametric generation of complex motion trajectories: The system decomposes the complex movements during garment wearing into three orthogonal motion components: axial stretching (simulating limb extension), radial torsion (simulating joint rotation), and normal impact (simulating slapping / contact). The system control module calculates the parameters at any given moment based on a preset scene mode (e.g., "vigorous motion mode"). Composite displacement vector applied to the fabric Its mathematical expression is as follows:

[0043]

[0044] In this formula, These represent unit vectors in three dimensions. Specifically, It was designed as a periodic stretching function to simulate the reciprocating motion of limbs; The torsion angle function is used to simulate shear deformation; This is the pulse function generated by the normal tapping. Parameters Weighting coefficients were assigned to different testing scenarios, enabling flexible simulation of various wearing conditions.

[0045] Constant Tension Closed-Loop Control Based on PID: During dynamic testing, the fabric is prone to slack (creep), leading to distortion of subsequent test data. To address this, the system introduces a real-time feedback mechanism, utilizing a PID control algorithm to maintain pretension. The constant. The control error is defined as the difference between the target tension and the measured tension. Controller output correction amount for:

[0046]

[0047] in, , , These are the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller, respectively.

[0048] This step ensures that each test is conducted under standardized initial tension, thereby guaranteeing the repeatability and comparative value of the data.

[0049] Step 2: Synchronous acquisition and preprocessing of multimodal heterogeneous signals (sensing layer):

[0050] While performing the aforementioned complex actions, the system needs to perceive changes in the fabric, much like human senses. Since the collected raw data contains environmental noise and has different modalities (images are two-dimensional matrices, and sound is a one-dimensional waveform), the core task of this step is to perform spatiotemporal alignment and noise reduction on these heterogeneous data.

[0051] Visual image texture enhancement and ROI extraction: The system utilizes a miniature camera to capture video streams of microscopic deformations on the fabric surface at high frame rates. Due to potential slight fluctuations in ambient lighting during testing, each frame of the image was processed to accurately capture the details of fiber pilling and fuzzing. The adaptive histogram equalization (CLAHE) algorithm is used for processing, and edge features are extracted in combination with the Canny operator:

[0052]

[0053] in, and These represent the mean and standard deviation of the local window, respectively. This step significantly enhances the contrast of the fabric texture, highlighting minor physical defects.

[0054] Acoustic signal spectral subtraction noise reduction: The sound generated by fabric friction is very weak and easily drowned out by the background noise of motor operation. In order to extract pure friction sound (i.e., "fabric language"), the system first acquires the original sound wave. The signal is then converted to the frequency domain using a short-time Fourier transform (STFT). Subsequently, the estimated background noise spectrum is subtracted from the total signal using spectral subtraction. To obtain a pure fabric sound spectrum :

[0055]

[0056] By adjusting the over-reduction factor The system can effectively filter out motor noise while retaining subtle acoustic characteristics that reflect the roughness and stiffness of the fabric.

[0057] Smoothing of mechanical hysteresis loops: for stress-strain data sequences acquired by force sensors A Savitzky-Golay digital filter is used for smoothing. This method removes high-frequency mechanical jitter noise while perfectly preserving the peak and trough characteristics of the mechanical hysteresis loop, preventing signal distortion caused by filtering.

[0058] Step 3: Multiphysics Feature Extraction and Quantization (Feature Layer):

[0059] Even after preprocessing, the amount of data remains enormous and difficult to understand directly. This step aims to transform the massive amount of raw signals into a set of feature vectors that can characterize the physical properties of the fabric through mathematical transformations, providing "fuel" for subsequent AI inference.

[0060] Visual texture feature extraction based on GLCM: In order to quantify the microscopic changes on the fabric surface (such as pilling, wear), the gray-level co-occurrence matrix (GLCM) of the image is calculated. Based on this matrix, two key indicators are extracted:

[0061] Contrast: This indicator reflects the depth of the grooves in the image texture, which directly corresponds to the surface roughness of the fabric.

[0062] Entropy: This metric reflects the degree of messiness in the image and is used to quantify the severity of pilling.

[0063] MFCC-based acoustic tactile feature mapping: To enable machines to "understand" tactile sensation, the system introduces Mel-frequency cepstral coefficients (MFCCs) from the field of speech recognition. A Mel-scale filter bank is used to simulate the nonlinear perceptual characteristics of the human ear regarding sound frequencies.

[0064]

[0065] The extracted first MFCC coefficients This constitutes the fabric's "acoustic fingerprint," which is highly correlated with subjective evaluations of "smoothness" and "gritty feel."

[0066] Dynamic mechanical viscoelastic characteristics: Calculate the energy loss (i.e., hysteresis loop area) in a single stretch-recovery cycle. This index characterizes the fabric's elastic recovery ability:

[0067]

[0068] in, For stress, In response to the situation.

[0069] The larger the hysteresis loop area, the more easily the fabric undergoes plastic deformation, and the more likely the garment will have poor shape retention (such as a bulge at the knee).

[0070] Step 4: Performance prediction model based on hybrid attention mechanism (algorithm layer):

[0071] Simple feature extraction cannot directly lead to the final conclusion. Therefore, this step constructs a deep neural network that mimics the comprehensive judgment logic of human experts, integrates the fragmented features mentioned above, and outputs a prediction of the garment's performance.

[0072] Heterogeneous Neural Network Architecture Design: Considering that the input data includes images (spatial information) and temporal signals (temporal information), a CNN-LSTM hybrid network architecture was constructed. Model input vector. From the perspective of vision ,acoustics and mechanics It is composed of three features pieced together:

[0073]

[0074] In this layer, 1D-CNN is used to extract local features of acoustic and mechanical signals; the LSTM layer utilizes its memory units. Capture the cumulative fatigue effect of fabrics during continuous dynamic testing.

[0075] Adaptive Modal Attention Mechanism: When evaluating different performance metrics, the importance of each modality varies (e.g., evaluating "balling" primarily considers visual aspects, while evaluating "feel" primarily considers acoustics and mechanics). Therefore, an attention mechanism is introduced to automatically assign weights. :

[0076]

[0077] in, The fused feature vector is represented by Dense(·), which indicates processing by a fully connected layer.

[0078] This mechanism enables the model to act like a human expert, automatically focusing on the most valuable data channels based on the prediction objective.

[0079] Loss function optimization: The model is trained through supervised learning, and the goal is to minimize the predicted score. Compared with real expert ratings The mean squared error (MSE) between the two is calculated, and an L2 regularization term is introduced to prevent overfitting and enhance the model's generalization ability.

[0080] Step 5: Dynamic Reasoning and Generative Evaluation Report (Application Layer)

[0081] The system transforms the algorithm's reasoning results into charts and conclusions that users can intuitively understand, completing a closed loop from "testing" to "decision assistance".

[0082] Style retrieval based on cosine similarity: The system not only outputs absolute values ​​but also provides "relative positioning." In the high-dimensional feature space, the feature vector of the fabric to be tested is calculated. Feature vectors of standard samples (such as top-grade silk) in the database Cosine similarity:

[0083]

[0084] Based on this, the system can output intuitive conclusions such as "the feel of this fabric is 92% similar to that of 100% mulberry silk," to assist designers in material selection.

[0085] Time-series performance degradation prediction: By leveraging the time-series extrapolation capabilities of the LSTM network, the system can recursively predict the state of the fabric after long-term use based on short-term test data (e.g., inputting data from the first 1000 washes to predict the wrinkle level after the 5000th wash), thereby generating a performance evolution curve for the entire life cycle of the garment.

[0086] To more intuitively illustrate the overall process of the above five steps and the data and control flow between each module, the following explanation is provided in conjunction with the accompanying diagrams. Figure 1 As shown, the technical route of the method described in this invention unfolds from left to right, summarizing the aforementioned five implementation steps into four closely connected core step modules. Each module is indicated by a thick arrow to indicate the data and control flow. The specific correspondence is as follows:

[0087] Step 1: Construct the nonlinear motion model (control) module. The core of this module is the "motion control unit," which integrates a complex mathematical model, including the "trajectory generation" formula for defining composite motion patterns. The system includes a PID control formula to maintain test stability. The motion control unit issues precise commands based on these models, driving the rotary motor, linear motor, and impact actuator respectively. These three actuators work together to apply a composite dynamic load to the fabric sample under test, shown in the dashed box in the middle of the figure.

[0088] Step 2: Multimodal Signal Acquisition and Preprocessing (Sensing) Module. When the "fabric sample" undergoes a dynamic response under mechanical action, the "sensor array" synchronously acquires its multiphysics field signals. The array employs a three-way parallel architecture: the first channel uses a "visual sensor" to acquire images and performs "image preprocessing" (including CLAHE enhancement and Canny edge detection); the second channel uses an "acoustic sensor" to acquire sound and performs "audio preprocessing" (including STFT transformation and spectral subtraction noise reduction); the third channel uses a "force sensor" to acquire stress data and performs "signal smoothing" (using a Savitzky-Golay filter). The preprocessed heterogeneous data from these three channels are then spatiotemporally aligned and converged within a "synchronous data buffer."

[0089] Step 3: Multiphysics Feature Extraction (Feature) Module. The "Feature Extraction Unit" reads data from the synchronization data buffer and converts the raw signal into quantized features using a specific algorithm. Specifically, this includes: calculating GLCM-based "visual texture" features (such as contrast and entropy); extracting MFCC-based "acoustic tactile" features (Mel frequency cepstral coefficients); and calculating hysteresis energy loss characterizing "mechanical viscoelasticity" (…). These three different types of physical features are ultimately combined into a high-dimensional comprehensive "feature vector". ".

[0090] Step 4: AI Prediction Model Building (Algorithm) Module. The feature vectors generated in the previous step are input into the core "deep learning model" for inference. The internal structure of this model first processes the input visual ( ),acoustics( ) and machinery ( The features are initially mapped using a CNN encoder (2D), a CNN encoder (1D), and a dense encoder, respectively; then they are fed into a Bi-LSTM layer to capture dynamic dependencies in the time series; finally, an attention mechanism is introduced to adaptively calculate and assign weights to each modality feature. The feature vectors, after being weighted and fused, enter the fully connected layer for nonlinear transformation; finally, the output layer outputs the performance prediction of the target garment. ".

[0091] To more clearly illustrate the technical effects of the present invention, the application of the present invention will be described in detail below through two specific embodiments.

[0092] Example 1: Prediction of dynamic fatigue and pilling performance of knitted sports fabrics at the knee

[0093] This embodiment aims to use the system of the present invention to simulate the repeated flexion and extension movements of the human knee joint and predict the shape stability and appearance retention (bubbling and pilling) of a new type of knitted sports fabric after long-term wear.

[0094] 1. Test subjects:

[0095] Sample to be tested: A high-elastic cotton / spandex blended knitted fabric (sample number: K-Run-01).

[0096] Prediction target: To simulate the degree of knee "bulge" and surface pilling level after wearing the garment for 3000 knee flexion movements.

[0097] 2. Experimental Procedures and Control Parameter Settings (Step 1):

[0098] The circular fabric sample is fixed onto the fixture of the multi-axis linkage actuator. The "knee fatigue test mode" is set in the main control unit, and the system generates the following composite motion instruction script:

[0099] Movement frequency: 1.5 Hz (simulating brisk walking / jogging frequency).

[0100] Composite trajectory:

[0101] Axial tension The maximum strain is 25%, exhibiting a sinusoidal wave pattern.

[0102] Radial torsion Apply simultaneously when stretched to the maximum point. The reciprocating torsion simulates the shear force during joint rotation.

[0103] Normal contact At the end of the recovery phase, a slight normal frictional contact is applied to simulate the contact between the trouser leg and the skin.

[0104] Tension control: The PID controller maintains the initial pretension at... .

[0105] Duration: 3000 consecutive cycles (approximately 33 minutes).

[0106] 3. Multimodal data acquisition and feature evolution (simulation data) (Steps 2 and 3):

[0107] During the test, the sensor array synchronously acquired data, and the system extracted feature values ​​every 500 cycles. Table 1 shows the simulation feature data records at key moments:

[0108] Table 1

[0109] Test Cycles Mechanical characteristics: Hysteresis loop energy loss Eloss (mJ) Visual feature: GLCM texture entropy value Ent (dimensionless) Data Interpretation 0 (Initial state) 15.2 (Good elasticity, fast response) 0.45 (smooth and flat surface) Baseline state 1000 18.5 (Damage increased by 21.7%, slight fatigue appeared) 0.92 (Increased texture disorder, micro-hair becomes more visible) Early fatigue stage 2000 24.1 (Significantly increased losses, accumulation of plastic deformation) 1.85 (Entropy value increases significantly, small hairballs form) Mid-term fatigue stage 3000 (Final State) 29.8 (Losses nearly doubled, poor recovery ability) 2.65 (High entropy value, severe surface pilling) Severe fatigue stage

[0110] Note: An increase in size indicates that the fabric cannot fully recover after absorbing energy, which suggests that an irreversible "bulge" will appear at the knee of the garment. An increase indicates that the image texture has become chaotic, which directly corresponds to an increase in surface fuzzing and pilling.

[0111] 4. AI Model Inference and Prediction Results (Steps Four and Five):

[0112] Feature fusion and attention allocation: The aforementioned temporal features are input into the trained CNN-LSTM model. Since the prediction targets are "balling" and "deformation," the model's adaptive attention mechanism automatically allocates higher weights to the visual and mechanical modalities (simulation weights: ).

[0113] The system outputs a prediction report:

[0114] Conclusion 1 (morphological prediction): "It is expected that after 3,000 wear cycles, obvious permanent bulge deformation will appear in the knee area, and the recovery rate will drop to below 85%."

[0115] Conclusion 2 (Appearance Rating): "The surface pilling rating is expected to drop to 2.5 (according to GB / T 4802.1 standard, 5 is the best and 1 is the worst)."

[0116] 5. Verification and Effect Comparison:

[0117] To verify the accuracy of the prediction, the same batch of fabric was sent to a standard laboratory for Martindale pilling test (2000 cycles of friction) and constant load elongation recovery test.

[0118] Laboratory test results: The pilling level was rated as between 2 and 3; the recovery rate under constant load was 83%.

[0119] Comparative Conclusion: The AI ​​predictions of this system are highly consistent with those of traditional destructive physics experiments. However, this system can produce a comprehensive conclusion in only about 30 minutes, while traditional full-scale testing takes several hours or even days, significantly improving R&D efficiency.

[0120] Example 2: Evaluation of the "silk rustle" and hand feel style of imitation silk synthetic fiber fabric

[0121] This embodiment aims to use the system of the present invention, focusing on acoustic and microscopic visual technologies, to objectively evaluate whether a new type of modified polyester filament fabric possesses the unique "scroop" (the crisp and pleasant sound produced when rubbed) and smooth feel of high-end silk.

[0122] 1. Test subjects:

[0123] Sample to be tested: Novel imitation silk polyester fabric (sample number: PolySilk-X).

[0124] Comparison standard: 100% mulberry silk double crepe fabric (sample number: RealSilk-S, its characteristic data has been stored in the database).

[0125] 2. Test Procedure and Control Parameter Settings (Step 1): Set the "Silk Sound / Hand Feel Evaluation Mode" on the main control unit. In this mode, the movements are gentle, focusing on the mutual friction and draping of the fabric itself.

[0126] Compound motion: Control the two clamps to drive the fabric to rub against each other at a low speed of 10cm / s, while simultaneously oscillating freely with a small amplitude, simulating the movement of the fabric when touched and worn by a human hand.

[0127] 3. Key Multimodal Feature Extraction (Simulation Data) (Steps 2 and 3): The system focuses on analyzing the acoustic signals generated by friction and the image features during dynamic suspension.

[0128] Motion-facial characteristics (MFCC): The "silk hum" of silk typically exhibits a specific energy distribution in the 2kHz-5kHz frequency band. The system extracts MFCC characteristics and performs spectral analysis.

[0129] Standard sample (silk) data: A significant energy peak appears around 3.5kHz, and the high-frequency harmonics decay smoothly, resulting in a crisp sound.

[0130] Simulation data of the test sample (PolySilk-X): A similar energy peak was detected near 3.2kHz, but the energy in the ultra-high frequency range >8kHz was slightly higher than that of silk, indicating that although its friction sound has the main tone of "silk hum", it is slightly sharp and not as mellow as silk.

[0131] Visual morphological features (optical flow and GLCM):

[0132] Dynamic drape: The pixel displacement field during fabric movement is tracked using optical flow. Simulation data shows that the dynamic recovery rate vector field of PolySilk-X has an 88% similarity to the silk standard sample, exhibiting excellent instantaneous drape.

[0133] Surface smoothness: GLCM contrast characteristics show that PolySilk-X has extremely low surface contrast (value close to that of silk), indicating that its surface is very smooth and has the basis of a "smooth" feel.

[0134] 4. AI Model Inference and Similarity Matching (Steps Four and Five):

[0135] Attention mechanism: When evaluating “style / feel”, the model automatically increases the weights of acoustic and visual modalities (simulation weights:).

[0136] Similarity retrieval and report generation: The system calculates the cosine similarity between the fused feature vector of PolySilk-X and the feature vector of RealSilk-S in the database.

[0137] System output evaluation report:

[0138] Key findings: "The overall style of this fabric is 89.5% similar to that of 100% mulberry silk crepe de chine."

[0139] Sub-item assessment: "Audio-tactile similarity (silk-like sound) 92% (Note: High frequencies are slightly sharp); Visual drape similarity 88%; Tactile delicacy similarity 95%."

[0140] 5. Verification and Effect Comparison: Five experienced fabric buyers were invited to conduct a double-blind subjective hand feel evaluation.

[0141] Human evaluation results: The expert panel gave an average subjective similarity score of 90 points (out of 100), and noted that "the sound is very similar, the feel is very smooth, but it lacks a bit of the firmness of real silk."

[0142] Comparative Conclusion: The quantitative evaluation results of this system (89.5%) are highly consistent with the subjective perceptions of human experts (90 points) and their qualitative descriptions, proving the system's ability to objectively digitize subjective feelings and effectively replace preliminary manual screening.

[0143] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A system for predicting and evaluating the dynamic performance of fabrics and garments based on multidimensional sensing, characterized in that: include: The control module is used to generate composite motion commands according to a preset scene mode. The composite motion commands are used to drive the actuator to apply a composite dynamic load to the fabric sample, which includes at least two of the motions of axial tension, radial torsion and normal impact. A multimodal sensing module, comprising at least two of a visual sensor, an acoustic sensor, and a mechanical sensor, is used to simultaneously acquire image signals, acoustic signals, and mechanical signals of a fabric sample under the combined dynamic load. The feature extraction module, connected to the multimodal sensing module, is used to preprocess the acquired multimodal signals and extract quantized feature vectors characterizing the physical properties of the fabric. The quantized feature vectors include at least two of the following: visual texture features, acoustic tactile features, and dynamic mechanical features. The AI ​​prediction module has a built-in trained deep learning model, which takes the quantized feature vector as input and outputs the prediction result of the dynamic performance of the garment after model inference. The output module is used to present the prediction results in a visual form.

2. The system according to claim 1, characterized in that: The control module includes: The trajectory generation unit is used to generate trajectories based on a preset scene pattern using formulas. Calculate the composite displacement vector applied to the fabric. ,in, These represent unit vectors in three dimensions. It is a periodic stretching function used to simulate the reciprocating motion of limbs; The torsion angle function is used to simulate shear deformation; The pulse function generated by the normal tapping; parameters These are the weighting coefficients for different testing scenarios; A PID controller is used to measure the difference between the target tension and the measured tension. ,pass Output correction amount to maintain constant pretension during the test, wherein, , , These are the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller, respectively.

3. The system according to claim 1, characterized in that: The multimodal sensing module further includes a preprocessing unit, which is used for: Adaptive histogram equalization (CLAHE) and edge feature extraction are performed on the image signals acquired by the vision sensor; The acoustic signal acquired by the acoustic sensor is subjected to short-time Fourier transform (STFT) and noise reduction is performed using spectral subtraction. The mechanical signals acquired by the mechanical sensor are smoothed using a Savitzky-Golay digital filter.

4. The system according to claim 1, characterized in that: The visual texture features extracted by the feature extraction module include at least the contrast ratio calculated based on the gray-level co-occurrence matrix (GLCM). Entropy The contrast is used to characterize the surface roughness of the fabric, and the entropy is used to quantify the severity of pilling and fuzzing of the fabric.

5. The system according to claim 1, characterized in that: The acoustic tactile features extracted by the feature extraction module are Mel frequency cepstral coefficients (MFCCs), which are filtered by a Mel scale filter bank. It was obtained by simulating the characteristics of human hearing.

6. The system according to claim 1, characterized in that: The dynamic mechanical features extracted by the feature extraction module include at least the energy loss in a single stretch-recovery cycle. ,in, For stress, Strain is used to characterize the elastic recovery ability of a fabric.

7. The system according to claim 1, characterized in that: The deep learning model in the AI ​​prediction module is a CNN-LSTM hybrid network architecture, and its input vector From the perspective ,acoustics and mechanics Features pieced together: ; The deep learning model also includes an adaptive modal attention mechanism, which automatically assigns weights to different modal features based on the prediction target. and through Feature fusion is performed, where, The fused feature vector is represented by Dense(·), which indicates processing by a fully connected layer.

8. The system according to claim 1, characterized in that: The output module includes: The similarity retrieval unit is used to calculate the feature vector of the fabric to be tested in the feature space. Compared with the feature vector of the standard sample in the database cosine similarity It also outputs style similarity conclusions. The time-series prediction unit is used to recursively predict the performance evolution curve of the fabric after long-term use based on short-term test data using the LSTM network in the AI ​​prediction module.

9. The system according to claim 1, characterized in that: The prediction results generated by the output module are visualized as a generative evaluation report, which includes at least one of the following information: fabric performance score, percentage of similarity to standard sample, performance degradation curve throughout the entire life cycle, and appearance rating conclusion.

10. A method for predicting and evaluating the dynamic performance of fabric garments based on multidimensional sensing, applied to the system described in any one of claims 1 to 9, characterized in that: Includes the following steps: Step 1: Generate composite action commands through the control module to drive the actuator to apply composite dynamic loads to the fabric sample; Step 2: Simultaneously acquire image signals, acoustic signals, and mechanical signals of the fabric sample under composite dynamic load using a multimodal sensing module, and perform preprocessing. Step 3: The preprocessed signal is mathematically transformed by the feature extraction module to extract a quantized feature vector containing visual texture features, acoustic tactile features and dynamic mechanical features. Step 4: Input the quantized feature vector into the deep learning model of the AI ​​prediction module, and output the prediction result of the dynamic performance of the garment through model inference. Step 5: Present the prediction results in a visual form through the output module.

Images