Data collection.

1. Power parameter monitoring. The fabric tear testing machine conducts trouser-shaped specimen tear tests using the CRE constant-rate elongation tester. For fabric tear performance testing, real-time monitoring of the power usage of the fabric tear testing machine is conducted. This includes parameters such as current, voltage, and power. Throughout the fabric tear performance testing process, the operational states of the CRE constant-rate elongation tester are categorized as standby, stretching, resetting, and sample changing, as detailed in Table 1. Considering these states, the collection of power parameters needs to span the entire experimental process. Particularly during the stretching operation of the equipment, the current and power consumption in the power parameters are maximal. Furthermore, according to international standards ‘ISO 13937-2-2000 Textile fabrics—Tear properties—Part 2: Determination of tear strength of trouser specimens (single seam)’, for constant-rate elongation instruments, if force and elongation records are obtained through data acquisition chips and software, the data acquisition frequency should be at least 8Hz. Following the Nyquist sampling theorem, to faithfully reconstruct the original signal without distortion, the sampling frequency should be greater than twice the highest frequency of the signal. Therefore, the designed frequency for collecting power parameters should be no less than 16Hz to real-time reflect changes in monitoring equipment power loads.
2. Temperature and humidity monitoring. Environmental temperature and humidity can impact the physical properties of fabrics. In high-temperature environments, the strength and toughness of fabrics decrease, while in humid conditions, the softness and elasticity of fabrics decrease. Fabric tear performance testing requires testing under specific temperature and humidity conditions. The international standard ‘ISO 139-1973 Textiles—Standard atmospheres for conditioning and testing’ stipulates the atmospheric conditions for testing: an atmospheric temperature of 20.0℃ with a tolerance of ±2.0%, and a relative humidity of 65.0% with a tolerance of ±4.0%. To obtain more accurate test results, it is necessary to install temperature and humidity sensors in the testing area of the fabric tear performance testing equipment. This helps prevent the test results from being affected by local temperature and humidity conditions. Therefore, when selecting the model of temperature and humidity sensors, sensors with an accuracy of 0.1 should be chosen.
3. Operator behavior detection. In fabric tearing performance testing experiments, it is necessary to monitor in real-time whether individuals are present during different operational stages. Based on the four working states of individuals during equipment operation, as shown in Table 2, which are: sample preparation stage, sample humidity adjustment stage, sample experimentation stage, and data recording stage, two types of personnel presence need to be monitored. One is the monitoring of the fabric operator’s behavior state, and the other is the monitoring of the equipment operator’s behavior state and data recording. The fabric operator needs to be present throughout the sample preparation and data recording, while the equipment operator needs to be present throughout the sample experimentation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Operational states of the CRE constant-rate elongation tester.

https://doi.org/10.1371/journal.pone.0302037.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Four working states of personnel during equipment operation.

https://doi.org/10.1371/journal.pone.0302037.t002

Data preprocessing.

In response to the multi-source sensor data collected during the data acquisition stage, three major categories are obtained: power parameter monitoring data, environmental temperature and humidity monitoring data, and operator behavior detection data. Due to the different sampling frequencies of various sensors and the prolonged duration of fabric tear performance testing activities, the collected data are all time-series data. Therefore, preprocessing operations such as data cleaning, data normalization, and sliding window segmentation are required for the collected data.

1. Data cleaning. Data cleaning involves noise reduction and data alignment.
   Noise Reduction:Since multi-source sensor data are collected from different sensors, the data collected from laboratory temperature and humidity sensors and operator detection devices may be subject to interference from the environment or their own performance, leading to data disturbances. For instance, long working hours of temperature and humidity sensors can cause data drift. To address this issue, median filtering is employed to eliminate pulse signal anomalies in the collected data. Median filtering, a non-linear signal processing method, utilizes a moving window with an odd number of points. The central value in the window is replaced with the median of the values within the window. By determining whether the signal value is a local maximum or minimum within the filtering window, median filtering effectively removes pulse noise and abrupt changes, as shown in Eq (1). (1) Here, represents the denoised data, N is the window size, and i is the starting position of the window. Operator detection devices are typically used to determine the presence of individuals in a specific area. Their output is usually a Boolean value (0 or 1), indicating the absence or presence of a fanyperson. However, these devices may generate frequent false alarms or sustained erroneous states. Although the output of operator detection devices is in Boolean form, continuous detection results can be smoothed using logical AND operations in signal smoothing techniques. This helps reduce false positives and false negatives. If the detection result is consistently ‘1’ for k consecutive times, it is confirmed as a presence of an individual. This can be represented by Eq (2). (2) Where x_t is the original detection result (Boolean) for theinstance, and y_t is the result after smoothing. Data Alignment:Data alignment involves interpolating the data collected from different sensors based on the start and end times of the time window, ensuring that the data from different sensors is not affected by the timing of the collection. Additionally, due to prolonged sensor operation leading to slower response times, missing values may occur during data collection. Linear interpolation is employed to fill in these missing values, as shown in Eq (3). (3) Here, represents missing points, and the window size is N = [a, b]. Once the data from all sensors has been interpolated to the same time interval, they can be considered synchronized.
2. Data normalization. The laboratory environmental temperature and humidity monitoring data are based on the threshold set for experimental temperature and humidity under the standard atmosphere mentioned in international standards to distinguish between normal and abnormal conditions during the experiment. The data from the operator behavior detection is boolean data, which is also judged based on the set threshold. Since different electrical parameters have different dimensions (voltage is measured in volts, current in amperes, and power in watts), their diverse dimensions can impact data analysis and processing. Normalization is employed to transform them into a uniform dimensionless indicator range, facilitating data analysis and comparison. This paper uses the most commonly used method in data processing, namely Min-Max normalization. The original data is linearly transformed using Eq (4), mapping the results to the range [0, 1]. (4)
3. Sliding window processing. In the overall activity design for fabric tearing performance testing, the sliding window technique is used to process and analyze the real-time collected time-series data. Let the minimum sampling frequency be f_min Hz, which corresponds to a sampling period T_sample = f_min second. Accordingly, the size of the sliding window N can be set to one or more complete sampling periods, which is expressed by Eq (5). (5) where k denotes the number of complete sampling cycles selected, determined according to the actual demand.
   To ensure continuity and capture subtle changes in the data, we set the step size S of the sliding window to half the size of the window, which is expressed by Eq (6). (6)
   When applying this sliding window algorithm, for any time point t, the current sliding window w_t as well as the position of the next window can be defined on the data series. The range of the i th window is [t_i − S, t_i], and the starting position of the next window w_i+1 is updated to t_i + S.
   The advantage of the sliding window approach is that, for a large number of continuous real-time data sets, it is able to perform segmentation while maintaining data continuity, which facilitates the extraction of potential pattern and trend information. This means that the system can update the data sub-sequence within the sliding window in real-time, and monitor and analyze it, so as to effectively identify and track the changes in the working status of the fabric tearing performance testing activities, and enhance the accuracy and traceability of the testing process.

Situation understanding.

Based on the problem of classifying equipment operation states in fabric tear performance testing addressed in this paper, and given that the input involves multi-source data, the classification structure is not a simple binary classification problem. Therefore, a DTSVM multi-class method has been proposed, representing the optimal solution for classifying equipment operation states within the situational awareness system. It is a parallel decision-level fusion method where multiple sensor data undergo parallel decision-level fusion. The method employs multiple threshold detectors and support vector machine models, executing sequential decisions. Each decision stage relies on the output of the preceding stage and operates in the form of a decision tree. Each model receives distinct input data, generates corresponding classification results, and the final decision results are fused through the outputs of multiple models.

As illustrated in Fig 5 of the DTSVM algorithm structure, the process of the DTSVM algorithm model structure is as follows: (1)Input the collected 7-dimensional vector dataset into the first threshold discriminator to classify non-1A test environment temperature and humidity vector data and 1B test environment temperature and humidity vector data.(2)Input the 1A test environment temperature and humidity vector data into the second threshold discriminator to classify 2A non-test operating state vector data and 2B test operating state vector data.(3)Input the 2B test operating state vector data into the first SVM to classify 3A standby state vector data and non-standby state vector data.(4)Input non-standby state vector data into the second SVM to classify 3B unmanned stretching experiment state and 3C manned stretching experiment state.(5)Input the standby state vector data into the third decision tree to classify 4B manned fabric preparation state and 4A unmanned fabric preparation state.(6)Construct the DTSVM algorithm model based on the above content. Through analysis, the working states can be categorized into 6 classes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. DTSVM algorithm model structure.

It is a parallel decision-level fusion method where multiple sensor data undergo parallel decision-level fusion. The method employs multiple threshold detectors and support vector machine models, executing sequential decisions. Each decision stage relies on the output of the preceding stage and operates in the form of a decision tree. Each model receives distinct input data, generates corresponding classification results, and the final decision results are fused through the outputs of multiple model.

https://doi.org/10.1371/journal.pone.0302037.g005

Decision tree support vector machine.

There is a difference in average similarity between samples of different classes, and for a classification problem with m classes, the DTSVM is a binary tree with m-1 internal nodes, each of which is an m leaf node and a binary classifier. If the classification performance of the upper nodes of the decision tree is not good, the overall classification performance becomes worse. Therefore, to ensure high generalization ability, the most separable classes should be separated preferentially at the upper nodes of the decision tree. In DTSVM, if each internal node considers using both binary SVM classifiers and thresholding judgments, the process of constructing the model can be regarded as a hybrid strategy, and the selection of each node may be based on the optimal classification hyperplane of SVM or on the threshold division of a certain feature.

For each internal node of the decision tree, suppose that a division based on the threshold t of feature A is considered, when the optimal feature and threshold can be selected based on some measure (e.g., information gain, Gini index, etc.), which can be used to reduce the uncertainty of the node. If the division is based on the threshold value using Eq (7): (7) where Gain(A, t) denotes the gain from dividing the dataset by feature A at threshold t.

Suppose we consider training a binary SVM classifier to classify the two classes of samples, denoted by Eq (8): (8) where w is the normal vector of the hyperplane, b is the displacement term, x_i is the feature vector, y_i ∈ {−1, 1} is the category label, and s is the set of training samples contained in the current internal node.

The specific construction process of the DTSVM model takes place iteratively, and each node is divided based on a specific optimization criterion to decide whether to use the SVM or the threshold judge.

The model was chosen for multi-source data and multiple problems because of its ability to integrate the intuitiveness and interpretability of decision trees and the generalization ability and nonlinear classification advantages of SVM. In complex systems, their state classification is often not a simple binary distinction, but involves multiple categories, and there may be complex boundaries and intrinsic connections between the categories. Individual nodes can either be filtered for classification using a threshold judge, or SVM can be applied for classification. Multiple models are connected in series through the form of decision tree, and each model receives part of the input data and outputs the classification results, and finally the outputs of all models are integrated to realize the effect of integrated learning, which improves the overall classification performance and robustness.

Decision tree method.

The Decision Tree method is a commonly used machine learning algorithm employed for classification and regression tasks. This model systematically partitions data features, ultimately generating a tree-shaped structure that can be utilized for decision-making. The fundamental idea behind a decision tree is to start from the root node and, through a series of branching nodes, classify or predict data. Each branching node corresponds to a feature, and the different paths of the branches represent different values of that feature. Meanwhile, each leaf node represents a classification label or regression value.

The structure of the DTSVM algorithm model is similar to a decision tree, where nodes represent decision or split points, and leaf nodes indicate the final classification results. Due to the specifications outlined in international standard ‘ISO 139-1973 Textiles—Standard atmospheres for conditioning and testing’ regarding the environmental temperature and humidity range in experiments, threshold detectors are employed at these nodes for classification. Simultaneously, the behavior state of the operating personnel is a simple boolean value, hence threshold detectors are also utilized for classification. As for the electrical parameters, support vector machines (SVM) are employed at nodes or leaf nodes to conduct more refined data classification, enhancing the model’s expressive capability, particularly when dealing with complex decision boundaries.

In the process of constructing a decision tree, the algorithm selects the optimal features and split points based on relevant criteria, aiming to make the data as pure as possible in each branch (i.e., samples of the same category are clustered together as much as possible). Information Gain (ID3) is an algorithm used to measure the contribution of features to the decision tree. Given a set D, the proportion of samples for the i-th class in the set is p_i(i = 1, 2, ……, n), and the entropy E(D) is represented as shown in Eq (9): (9)

Assuming m is the outcome of discrete attribute b, after splitting the set based on attribute b, m nodes are formed, and the m-th node is labeled as D^m. The information gain G = (D, b) from partitioning set D based on attribute m is as shown in Eq (10): (10)

Whereas, the larger is G = (D, b), the better the effectiveness of the attribute partitioning.

Support vector machine method.

Support Vector Machine (SVM) is a supervised binary classification model that excels in datasets with complex decision boundaries. The core idea of SVM is to find an optimal hyperplane that effectively separates data points of different classes while maximizing the margin, which is the minimum distance between the two categories. SVM performs exceptionally well in datasets with complex decision boundaries.

z_i ∈ {−1, 1} represents the classification label, and the discriminant function is defined as shown in Eq (11): (11)

The optimization problem is solved using the Lagrange multiplier method, and the optimal decision function is as shown in Eq (12): (12)

Whereas, N is the number of support vectors, and a_i is the Lagrange multiplier.

If the slack variable ξ_i ≥ 0 is added to the previous Eq, it can be reformulated as shown in Eq (13): (13)

Minimizing the objective function using quadratic programming, i.e. as shown in Eq (14), (14)

The parameter C in the previous Eq represents the penalty factor. The minimized objective function is transformed into a linear problem through nonlinear knowledge, resulting in the optimal classification plane and the expression of the optimal classification function as follows Eq (15): (15)

In the previous Eq, K(x_i, x) is the kernel function, and the kernel function must satisfy the requirements of Mercer’s theorem. The choice of kernel function in support vector machines will affect the classification results. Electrical parameters include multiple dimensions: current, voltage, power. SVM is effective in classifying high-dimensional spaces, maintaining good performance even in cases of high-dimensional feature spaces. Polynomial kernel functions, radial basis kernel functions, and linear kernel functions are commonly used as kernel functions in support vector machines. The radial basis kernel function is defined as shown in Eq (16): (16)

The kernel equation η and the penalty factor C are two key parameters that can impact the performance of support vector machine classification. Therefore, it is necessary to find a balance between achieving the goals of ‘correct classification of training samples’ and ‘maximizing the margin of the decision function’ by adjusting the penalty factor C and parameters of the kernel function. This ensures optimal performance of the model.

Ten-fold cross-validation can be employed to assess the classification performance of SVM by randomly partitioning the training samples of multi-source data into 10 mutually exclusive subsets{S₁, S₂, …, S₁₀}, representing ten folds. Each fold is of approximately equal size, and the process is iterated 10 times. In the j-th iteration, the S_jsubset is chosen as the test set, while the sum of the remaining subsets forms the training set. The decision function is trained on the training set and then tested on the corresponding test set. The number of misclassified samples in each iteration is denoted as I_j. After 10 iterations, the ten-fold cross-validation error {I₁, I₂, …, I₁₀} is calculated as the ratio of the sum of misclassified samples over the total number of samples. The formula is as shown in Eq (17): (17)

In the equations: e represents the estimated error rate; k is the iteration count, where in the context of ten-fold cross-validation k equals 10; n is the total number of samples. Based on the cross-validation error, optimal SVM hyperparameters can be selected, not only to improve the classification performance of SVM but also to prevent overfitting. The introduction of ten-fold cross-validation for training SVM involves randomly dividing the training set into 10 subsets. In each iteration, 9 subsets are used for training, and the remaining subset is used for validation. The average of the results from 10 cross-validation iterations is then used to evaluate the algorithm’s accuracy. Its characteristic lies in the ability to repeatedly use randomly generated sub-samples for training and validation. Each training result is validated once, contributing to improving model accuracy and generalization ability.

Data feature analysis.

Sample Preparation and Humidity Adjustment: On the preceding day, samples are prepared with a preparation time of approximately 15 minutes per set, following the personal operating habits of the operator. These samples are then placed in a standard atmospheric environment for humidity adjustment over a 24-hour period. The experiment is conducted the next day, and no monitoring is performed during this stage of the experiment. Experiment Execution: The experiment begins with the startup of the equipment and preparation for testing. After a preloading period of 30 minutes, the operations commence.

Initially, a tensile test is conducted on polyester fiber, followed by a blended fiber (nylon, spandex, lyocell) tensile test. Each fabric type comprises 10 specimens, with 5 samples taken in the warp direction and 5 in the weft direction.

For the collected two sets of multi-source data, data preprocessing was performed, and the preprocessed data (excluding data normalization) was visualized.

The multi-dimensional sensor recorded the data obtained by SASFTPD-MSD throughout the entire experiment, documenting the working states of two sets of experiments involving polyester fiber and blended fabrics (nylon, spandex, Lyocell), as shown in Fig 7, The SASFTPD-MSD collects multi-source data after the device is powered on, During this experiment, temperature, humidity, and voltage smoothly fluctuated within reasonable ranges. The variation in the operator’s state, obtained from the detection radar, reflected the working state of the experiment when one operator was performing tasks. The changes in current and power reflected the equipment’s states, including standby, preloading, tearing and stretching, and stop-reset. The standby state involves the initiation of the equipment and the preloading process. During this phase, the equipment power gradually rises from 0 to the standby power. The equipment remains in standby mode until the tearing and stretching experiment is initiated. The first set of experiments involves polyester fiber fabric, with 10 sample groups. From the power data, it is evident that the power during tearing and stretching fluctuates around 100W, while the power required for resetting after each tearing is approximately 300W, lasting for about 10 seconds. Throughout this process, the operator remains near the equipment. After completing the experiment, the equipment returns to standby mode, and the operator performs data operations on the computer at the operating console. The experimental process for the second set of blended fabric (nylon, spandex, Lyocell) samples is similar to the first set of polyester fiber fabric, but it is noticeable that the stretching time and reset power are significantly higher than those in the first set, as shown in Fig 8. The 10 sets of data in the experimental process consist of 5 sets of tearing experiments in the radial direction and 5 sets in the weft direction for each type of fabric. We compared the mechanical time series charts with the system-perceived electrical power time series data, as shown in Figs 9a–9d and 10a–10d. Through the comparison of the experiments, it can be observed that due to the differential mechanical characteristics of tearing in the radial and weft directions of the fabric, fabrics with different elasticity also exhibit differential mechanical characteristics during the tearing experiment. By comparing electrical power with mechanical characteristics, the consistent fluctuation trends of the two indicate that we can use fewer power perception data to reflect the mechanical changes. This demonstrates that the SASFTPD-MSD system can effectively restore the working process.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Time-series state diagram of multi-dimensional perception data.

The SASFTPD-MSD collects multiple sources of data after the device is powered on, including temperature, humidity, voltage, operator status changes, and current and power changes.

https://doi.org/10.1371/journal.pone.0302037.g007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8.

Single tearing experiment power timing chart: (a) Polyester Fiber Fabricl; (b) Blended (Nylon, Spandex, Lyocell) Fabric. This is single tearing experiment power timing chart. (a) is Polyester Fiber Fabricl tearing experiment power timing chart. (b) is Blended (Nylon, Spandex, Lyocell) Fabric tearing experiment power timing chart. They reflect the three working states of the equipment operation, which are: stretch state, reset state and change state.

https://doi.org/10.1371/journal.pone.0302037.g008

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9.

Comparison Chart of Polyester Fiber Fabricss: (a) Radial Mechanical Time Series Chart; (b) Radial Electrical Power Time Series; (c)Weft-wise Mechanical Time Series Chart; (d)Weft-wise Electrical Power Time Series Chart. By comparing the warp and weft mechanical time series diagram of polyester fiber fabric with the warp and weft system sensing electrical power time series data diagram, the fluctuation trend of the two is consistent.

https://doi.org/10.1371/journal.pone.0302037.g009

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10.

Comparison of Blended (Nylon, Spandex, Tencel) Fabrics: (a) Radial Mechanical Time Series Chart; (b)Radial Electrical Power Time Series Chart; (c)Weft Mechanical Time Series Chart; (d)Weft Electrical Power Time Series Chart. By comparing the warp and weft mechanical time series diagram of blended (Nylon, Spandex, Tencel) fabrics with the warp and weft system sensing electrical power time series data diagram, the fluctuation trend of the two is consistent.

https://doi.org/10.1371/journal.pone.0302037.g010