Artificial intelligence (AI) in textile industry operational modernization

Abstract

Purpose

With the help of basic physics, the application of computer algorithms in the form of recent advances such as machine learning and neural networking in textile Industry has been discussed in this review. Scientists have linked the underlying structural or chemical science of textile materials and discovered several strategies for completing some of the most time-consuming tasks with ease and precision. Since the 1980s, computer algorithms and machine learning have been used to aid the majority of the textile testing process. With the rise in demand for automation, deep learning, and neural networks, these two now handle the majority of testing and quality control operations in the form of image processing.

Design/methodology/approach

The state-of-the-art of artificial intelligence (AI) applications in the textile sector is reviewed in this paper. Based on several research problems and AI-based methods, the current literature is evaluated. The research issues are categorized into three categories based on the operation processes of the textile industry, including yarn manufacturing, fabric manufacture and coloration.

Findings

AI-assisted automation has improved not only machine efficiency but also overall industry operations. AI's fundamental concepts have been examined for real-world challenges. Several scientists conducted the majority of the case studies, and they confirmed that image analysis, backpropagation and neural networking may be specifically used as testing techniques in textile material testing. AI can be used to automate processes in various circumstances.

Originality/value

This research conducts a thorough analysis of artificial neural network applications in the textile sector.

Citation

Sikka, M.P., Sarkar, A. and Garg, S. (2024), "Artificial intelligence (AI) in textile industry operational modernization", Research Journal of Textile and Apparel, Vol. 28 No. 1, pp. 67-83. https://doi.org/10.1108/RJTA-04-2021-0046

1. Introduction

Textile manufacturing industries have long supported major fashion, clothing and cosmetics businesses in Asia, particularly China, India, Bangladesh and Vietnam. As labor prices in Asia climbed in recent decades, many workers moved east to work in textile production. With access to historical and real-time operational data, textile manufacturers will use artificial intelligence (AI) to improve productivity and workforce capacity as industrial automation spreads (Kotsiantis, 2017; Gong and Chen, 1999). It is necessary to define AI to comprehend it and associate it with its applications. AI systems may be classified into many categories as shown in Figure 1. They are all different, yet they often work together to complete a programming job. In the textile sector, expert systems and artificial neural networks (ANN) are being used. AI is defined as “the theory and development of computer systems capable of doing tasks ordinarily requiring human intelligence, such as visual perception, speech recognition, decision-making, and language translation,” according to the Oxford Dictionary. Although there appears to be no general adoption of AI – even in industrialized countries – the introduction of AI technology in the textile sector is still fairly early, with few applications. Instead, we are investigating the use of AI in the textile industry today. AI has been brought into the textile industry throughout the previous three decades based on trials and limited application. From the fiber sorting process to the logistic management of the final garment to the consumers, AI has the potential to be implemented at every stage of the process. The major goal of this review is to shed light on the use of AI in critical procedures such as spinning, weaving and coloration.

2. Organs of artificial intelligence

Machine learning is a branch of AI that allows computers to learn and improve on their own without having to be explicitly programmed. Machine learning is concerned with the creation of computer programs that can access data and learn on their own. The learning process starts with observations or data, such as examples, direct experience or instruction, to find patterns in data and make better decisions in the future based on the examples we provide. The fundamental goal is for computers to learn on their own, without the need for human involvement, and to change their behavior accordingly.

2.1 Some machine learning methods

Supervised machine learning algorithms can use labeled examples to apply what they have learned in the past to fresh data and predict future events (Figure 2). The learning algorithm creates an inferred function to generate predictions about the output values based on the examination of a known training data set. After enough training, the system can provide targets for any new input. The learning algorithm can also compare its output to the correct, intended output and identify faults so that the model can be modified appropriately (Lloyd et al., 2013).

When the data used for training is not categorized or labeled, unsupervised machine learning techniques are used. Unattended learning investigates how structures might derive a function from unlabeled data to describe a hidden structure. The system does not identify the optimal performance, but it studies the data and can extract information from data sets to define unlabeled structures (Lloyd et al., 2013; Hu et al., 2017).

Semisupervised master learning algorithms fall midway between controlled and unsupervised learning algorithms in that they use both labeled and unlabeled data for training, with small amounts of labeled data and a substantial amount of unlabeled data. The accuracy of learning will be considerably improved by systems that use this strategy. Semisupervised learning is used when the gathered data necessitates the use of qualified and appropriate resources to train/learn from it. Obtaining unlabeled data, on the other hand, usually does not necessitate additional resources (Zhu et al., 2009).

Reinforcement machine learning algorithms are a type of learning algorithm that interacts with its surroundings by generating actions and detecting failures or rewards. The most important elements of reinforcement learning are trial-and-error search and delayed reward. This technology enables machines and software agents to automatically select the best behavior for a given situation to enhance performance. For the agent to learn which action is better, simple reward feedback is required; this is known as the reinforcement signal (Botvinick et al., 2019).

Deep learning is an AI job that mimics the human brain's data processing and decision-making routines. Deep learning, also known as profound neural education or deep neural networks, is the AI machine learning category that includes networks capable of learning unstructured or unstructured input. Deep learning is based on massive volumes of unstructured data that people must ordinarily comprehend and analyze for decades. Individual items were detected by the early deep models in tightly sliced, extremely small pictures (Samek et al., 2017). The size of images that neural networks can analyze has gradually expanded since then. Networks for identifying modern items analyze large pictures and do not require a snapshot taken close to the object to be identified. Similarly, earlier networks could only distinguish between two types of objects (or, in some cases, the absence or presence of a single type), whereas modern networks can recognize over 1,000 different object categories (https://www.sas.com/en us/insights/articles/big-data/artificial-intelligence-machine-learning-deep-learning-and-beyond.html). When a major rise in profound learning happens, a coevolutionary network solves this difficulty by a huge margin, lowering state-of-the-art error rates from 26.1% to 15.3% (Haragrave, 2020).

Neural networks are a type of software or hardware that performs activities comparable to those performed by human brain neurons. Deep learning and machine learning are examples of technology used in neural networks. ANNs are the most common machine learning technique. These are systems that are created by the brain's inspiration of neurons and are designed to mimic how humans learn. An input and output layer, as well as a hidden layer of input-to-output units that can use the output value, make up neural networks. These are the approaches for finding and training programmers to recognize patterns in the machine that are numerous and complex. A vast number of parallel processors are stacked in layers in an ANN. Raw data, such as the optical nerves of the human eye, is fed into the first layer. Each layer collects raw input data as a result of the previous layer, similar to optic nerve neurons receiving signals from the close layer. The final layer is responsible for production. Below are several layers (Hardesty, 2017). Neural networks can be adjusted to their specific requirements, for example, by adapting to their training and running in parallel to provide greater global knowledge. If the network produces the “desired” output, the qualified input data must not be changed and vice versa. When the network delivers an “unwanted” output resulting in an error, the program altered the qualified input data to improve the outcomes (Hagan, 1995).

3. Artificial intelligence in textiles industries

The ultimate objective of AI is to create systems that think and act like humans. The ANN methodology uses backpropagation with a variable learning rate and several linear regressions. It can help with fiber grading, yarn quality prediction, fabric issue diagnostics and dye recipe prediction. Its applicability in various aspects of the textile industry, from raw materials to finished textiles, has sparked significant controversy. Fabrication, dyeing and yarn production are all reviewed in this research. Adaptive neuro-fuzzy inference system (ANFIS) has recently been applied to assess yarn characteristics. The textile sector also uses ANN to anticipate utility attributes including moisture and heat transfer rate in fabrics, as well as air permeability. The use of neural networks in textile coloring and printing also has reduced the need for human contact by predicting dye formulas, matching colors and identifying dyeing faults.

3.1 Artificial intelligence in yarn production

Fabric and other textile materials have been made from yarn for centuries. Except for nonwovens, most items are made of yarn. Fabric qualities are influenced to a large extent by yarn properties, hence finding faults is crucial to maintaining yarn and thus fabric quality. A human inspector would struggle to discover and manage over 70% of reported defects if they were bigger than 2 m and moving faster than 30 m per minute. Researches claim that thresholding can detect 90% of defects in a plain textile. Until the 1960s, AI was an algorithm-based system (Seckin et al., 2019; Artificial Intelligence in the Textile Industry-Current and Future Application, 2019). Fabrication began using computers and algorithms after the 1980s (at least in R&D sections). Image analysis is being used to assess the yarn quality. This approach examines hair area index, hair longitude index, mean absolute hair-length deviation index, normal hair-length index deviation and hair-length index coefficient. Hair area is defined as the total area of single fibers divided by the total area of the core. The total number of field pixels may be calculated from binary frames by adding the total area of the fibers and the entire area of the yarn core.

Online monitoring and process management are currently used by most big yarn production enterprises. According to the narrative, AI can detect defects, scientifically analyze failure rates and adjust control settings to maximize spinning. In most situations, AI has been used to analyze the programmable quality attributes of yarn. The picture of a yarn flaw is captured by a high-resolution camera in the system and analyzed by a preprogrammed computer system (Padmanabhan and Balasubramanian, 1990; Semnani et al., 2006; Marati and Semnani, 2011). This approach is enhanced when photographs are processed using machine learning. Deep learning and neural networks is employed to enhance the system's error analysis.

In neural networks and deep learning, a system learns a huge number of defect combinations to identify any yarn fault (Nateri et al., 2014; Sharma and Sindhe, 2016; Gultekin et al., 2019; Czimmermann et al., 2020). Poorer apparent quality is determined by irregular fiber arrangement on the yarn's surface. ASTM D 2255 splits yarn surface defects into four classes. This standard grades yarn on fuzziness, nepiness, unevenness and apparent foreign materials. In 2006, Semnani et al., 2006) verified the same technique for assessing yarns. This approach works for any kind of fiber since it is based only on the physical look of the yarn surface. They analyze pictures using a neural network. They acquire the best mistakes by experimenting with the starting weights for each category. After generating initial weights, a Perceptron ANN processes them. Each category has its neural network, with its weights and grade indicators. The network has four core nodes and one fuzzy node. It uses 10,000 epochs at a 0.1 training rate to learn the initial weights and indicator values.

Carvalho et al. (2009) developed a technique that combined image processing and AI to autonomously assess yarn quality and fabric appearance. This solution addresses the disadvantages of typical yarn tests by being cheap cost, low volume, low weight, portable, high resolution and technologically stable. Furthermore, the less sophisticated measuring apparatus needed allows for online control throughout manufacture. Fabijanska et al. (2012) used image processing to quantify yarn hairiness (Fabijanska and Jackowska-Strumitto, 2012). They first used a Boykov and Jolly (2001) algorithm to segment the yarn core. On the core axis, Cybulska recommended scanning each picture line perpendicular to it and measuring its edges from the longest foreground pixel intervals. This is done by adjusting the initial bounds using predefined curves whose points create the yarn center edge (Cybulska, 1999). Nateri et al. developed an image processing method to identify yarn problems in 2014. Then, using Otsu's method, pictures are converted to binary integers (Otsu, 1979).

σ(t) = q1(t) σ1(t) + q2(t) σ2(t)

The machine was taught using a simple set of images in this approach. As a result, this technology is incapable of identifying any type of flaw or yarn surface features. To determine yarn metrics including diameter, diameter variance, number of thick and thin squares, hairiness indices and hair zone indexes, Sengupta et al. devised a computerized technique in 2014 (Sengupta et al., 2015). Hairiness is more exact than other tests that measure hair quality or hair count. Because image processing involves a measuring notion, choosing the correct context is critical. Because the observable parameter may be modified pixel by pixel, the system designed has better accuracy. The developed approach works well for colored threads on a white backdrop. Temperature and humidity do not affect the system's functioning.

In 2021, an ANFIS has been developed to forecast yarn tenacity and unevenness using six input cotton fiber parameters: strength, elongation, upper half mean length, uniformity index, fineness and short fiber content. The ANFIS paradigm blends fuzzy control with neural networks. Fuzzy control uses a learning and computational neural network. However, fuzzy control provides high knowledge and fuzzy rules for use in neural networks. Using an experimental data set, the ANFIS model is trained to predict yarn tenacity and unevenness. Its prediction accuracy is evaluated using five statistical metrics: correlation coefficient, mean absolute percentage error, root-mean-square error, coefficient of efficiency and variance performance index. ANFIS models may be used to forecast different yarn quality parameters using fiber qualities as input variables based on permissible values (Das and Chakraborty, 2021).

3.2 Artificial intelligence in fabric production

Fabric formation is the second stage in the textile production process. Most weaving businesses have been focused on automation. However, other operations, such as fabric inspection, are still regarded to be done by a human. It implies that certain processes necessitate the use of human intelligence. Humans, on the other hand, are prone to making mistakes. As a result, neural networking and machine learning have become increasingly significant. Fault detection is the most common application of AI. In addition, logistic management in loom shade, pattern recognition, physical property prediction and other sectors can benefit greatly from AI. Table 1 shows the AI modeling in the area of fabric manufacturing.

Tsai et al. (1995) developed a neural network-based method for detecting fabric faults. The performance and accuracy of a neural network-based fabric fault detection system were studied in numerous areas. The network detected four categories of common textile defects. With the learning legislation history algorithm, clothing defects may be identified. The cooccurrence-based technique is used to retrieve six function parameters. Among these are three spatial displacements and four fabric defect labeling sites.

Ribolzi et al. (1993) used optoelectronics to discover defects in real-time. In the two-dimensional power spectrum, the local peak values of the textile picture fitted the null order and maximum diffraction. Comparing null and first-order values from both normal and defective materials reveals defects. But Ribolzi's method could only uncover a few frequent problems.

Chen et al. (1998) used a neural network system to build a new model based on the same theory in 1998. Various types of samples with the same defect often yield various power spectrums, resulting in a unique kind of textile sheet defect. As a result, identifying the various sorts of garment faults when studying the power continuum might be difficult. Nonetheless, they thought that the sample variances were due to modest changes in the sample's general structure. Yet a neural network can categorize items into groups with minor variations. They examined 12 problems in their report. Using this logic, they started with ten standard textile spectrums and ten defect-type spectra. The system was built using nondefect samples, and five of the ten tests for each fault sample were used for the device inspection. As a consequence, 65 two-dimensional spectrums were available for training and 65 for testing. They used a MatLab backpropagation neural network to identify and categorize defects in fresh textile samples. Twelve types of faults were detected and assessed, each with five examples. The device database has 9 out of 12 defects that can be discovered and evaluated appropriately. Each sample is recognized and classified in 0.2 s. So, this system should be able to operate in real-time.

Mallik-Goswami and Datta (2000) devised an image-based approach for evaluating fabric defects. The imaging technology set them apart from previous investigations. For example, they used morphological processes like erosion and opening of testing fabrics to find weaknesses. Mathematical morphology is used to extract and characterize regional shapes. The morphological techniques modify a planar geometric structure by a structural element. The structuring element's size and form are important morphological steps. Oxidation and dilatation are key functions that delete or add binary pixels dependent on the next pixel patterns. Textiles with tiny flaws perform better when morphological faults are detected on spatially filtered photographs of fabrics. The test fabric's diffraction pattern is acquired visually using a collimated laser beam. In the Fourier plane, a spatial filter removes the fabric's periodic grating structure. It involves morphological methods with a properly selected structural element after proper preprocessing. To discover fabric faults, Tilocca et al. (2002) constructed a neural network in 2002. The sample's grey levels and three-dimensional range profile data were used to classify faults and the double set of patterns collected provides essential sample information which is clear. Because no further data is needed to categorize, the system's output is fast and suited for high-volume online fabric monitoring.

Park et al. (2000) evaluated the hand value of knitted cloth using two methods. One uses backpropagation neural networks, while the other uses fuzzy prediction and the resultant total hand values. The second strategy uses a fuzzy membership function, a weighted factor vector and a backpropagation algorithm to repeat mistakes. They are linked to essential mechanical qualities such as stretchiness, bulkiness, strength, distortion, weight and surface roughness. For fuzzy and neural networks, subjective evaluation yields better results than the Kawabata evaluation system for fabric (KES-FB) approach. Excessive mechanical characteristics are required to foresee the entire hand’s worth of knitted outerwear. In each example, the projected error is smaller than the scanning electron microscopy (SEM). Initially, the fuzzy-neural system was differentiated from the neural network system. There are three kinds of fuzzy transform membership functions used to turn the process into fuzzy values. After fuzzification, they examined a fabric with seven unknown primary mechanical characteristics, and the rest of the method was similar. The fuzzy neural system rates the knit fabric using human perceptions. Then, using the KES-FB system, they evaluated 47 knits' mechanical characteristics and sent the results into neural and fuzzy-neural network backpropagation. For assessing the overall hand value of outerwear fabrics, fuzzy logic and neural networks performed well. This simulator's second benefit is its adaptability to a wide range of textile markets and fabric varieties.

The neural network may be used to evaluate additional properties such as comfort and handle of the fabric. The feed-forward backpropagation ANN can estimate the thermal insulation of materials using inputs like weave, thread density, yarn count, area density and thickness. The network predicted the thermal insulation of woven textile materials based on these characteristics before they were manufactured.

For woven fabrics, Cay et al. (2007) developed an air permeability and water content prediction model. They sewed 30 samples with varied warp and weft densities. They tested air permeability and water content in woven dry conditions. The structural features of the fabrics and observed values were associated using multiple linear regression and ANN. Generalized regression ANNs, in particular, can estimate fabric air permeability and hence fabric water content after vacuum drying. The performance of the linked models was evaluated by comparing anticipated and experimental values. The nonlinear models projected values that matched the experimental data. The air permeability and water content modeling helped forecast the fabric's physical qualities based on the design specifications. Vacuum drying performance assessment helped optimize industrial drying procedures.

Bu et al. (2008) developed computer vision-based automated fabric failure detection. In Yu et al. (2010) developed an artificially intelligent system that measures cloth hand-value. To train and test the algorithm, 30 people rated 10 fabric samples. They found the approach was 80% accurate. Their study shows that their suggested FNN can accurately predict fabric hand. The reduced fuzzy set may help fashion designers make decisions. To forecast woven fabric heat transmission properties, Kothari and Bhattacharjee (2011) defined ANNs and their construction, development and application to woven fabric thermal characteristics. Based on parameters such as yarn count, weave, thread density thickness and GSM, feed-forward backpropagation may predict certain physical behavior of textile materials. The neural network was also built to forecast the thermal insulation properties of woven textiles. For thermal applications, it is found that ANNs may save time and money provided the material's thermal insulation behavior is understood before manufacture and testing.

In 2017, Sparavigna used image segmentation to study textile texture (Sparvigna, 2017). They noticed that looking at many images of the same cloth and analyzing the corresponding distributions might help quantify the difference. Another research used deep learning to automatically identify defects in woven cloth, with a classification accuracy of 78.1% for holes, 81.6% for stains, 84.7% for selvedge tails and 74.6% for snarls. This research concluded that using automated fault detection and identification systems in garment inspection systems will increase the production and quality of apparel (Das et al., 2021).

ANN can predict various physical attributes from a fabric's structural data, such as air permeability, thermal insulation and bending rigidity. It saves time and money by predicting the behavior of material before it is manufactured.

3.3 Artificial intelligence in dying process

In the textile business, coloration or dyeing has always been a precise process. Before computers and automation, people prepared dye mixtures, applied dyes to fabrics and inspected for dyeing. Initially, soft computing replaced human-driven procedures. AI, the next generation of soft computing, can now achieve anything humans could do. With AI, reliable dye formulations, color matching and fault detection are now possible without human involvement. The Kubelka-Munk (K-M) model was used in early 1940s recipe prediction systems. In the K-M theory, colorants are characterized by their absorption and scattering coefficients, K and S. The K-M theory is a two-flux version of a multiflux technique. In spite of more exact theories, the K-M theory is still extensively used because of its simplicity and ease of determining K and S coefficients. The K-M hypothesis is used to forecast the color of textiles, paints and printing inks. This study's version of the theory is for transparent printing inks, where K and S must be fixed absolutely. Using this model, the K_i and S_i contributions for each colorant i are assumed to be additive for mixture j thus

(1) Kj= ΣicijKi

Sj= ΣicijSi

According to a report by Westland (2002), using a neural network to formulate a recipe was simply a trial until 1994, and it was not even an established method. He also highlighted the creation of colorimetric online ANN equipment that is affordable. By computing arbitrary mappings between the readings of different instruments in spectrophotometer and colorimeter measurements, ANNs may also find a way to improve the interinstrument dependability. Westland (2002) showed in 2002 that the K-M theory can map a colored vector c and a reflection vector r to represent the color prediction issue. Multilayer perceptrons (MLPs) may properly replicate any continuous feature to any degree. MLP stands for the multilayer processing unit. The first or input layer units take input from a real-world vector, while the final or output layer units use the network's output. Between the input and output layers, there may be hidden tiers of units. The findings show that although ANNs can learn to map colorant concentrations and spectral reflectance in theory, they struggle to beat the K-M model in reality. They showed the value of using separate training and test sets to assess overall network performance.

Almodarresi et al. (2019) developed a neural network-based scanner for color matching reactive colored cotton. Colorimetric and spectrophotometric matching are K-M color matching methods. The former is based on Allen's approach, which concerns tristimulus values equalization under specified observation circumstances. More accurate color formulation prediction of reactive dyed cotton samples using neural network and scanner than spectrophotometer. The scanner scanned fabric samples at 150, 300 and 600 dpi. The photographs' histograms fed a neural network. The input layer trained neural networks with 15, 30 and 60 input vectors, three output neurons, and a hidden layer with varying numbers of neurons. The least mean square error for a neural network using 60 input vectors from photographs at 300 dpi and 24 hidden neurons was 3.319710–5. The best neural network also has the low ternary relative error. Almost 80% of the testing data for the top neural network had a color difference of less than 1.5.

Yang et al. (2018) developed an ANN method to assess a mélange yarn's color mix. The most important and difficult step in creating combination films is blending precolored fibres. The recruitment estimates a list of precolored fiber percentages that are used to forecast. The ANN is regarded to be a more effective recipe prediction method than the traditional one. The modular artificial neural network (MANN) is considered as a novel approach for resetting top-dyed mixed yarn from reflecting spectrums. According to the study's findings, the prediction formula was generated by combining various subnetwork estimates and processing them in tandem. The MANN model predicted a better success rate than the ANN, with fewer error factors and a faster instruction time. The color difference between CMC (2:1) and the anticipated spectrum derived by the MANN model was 1.26, which is bigger (0.60) than reality.

In Haji and Vadood (2021), the polyester cloth was dyed with madder, an eco-friendly natural dye. Using a Box-Behnken experimental design, 46 samples were colored with varying amounts of five parameters: dye concentration, dyebath pH, temperature, duration and liquor ratio. Each parameter's effect on color strength was examined using multiple analyses of variance. The observed K/S values were predicted using ANN and fuzzy logic models. To increase model accuracy, the genetic algorithm was used in both models. The top ANN and fuzzy models could predict K/S values with mean absolute percentage errors of 2.52 and 3.01. Using partial derivatives, the contribution of each input parameter on ANN was found, with dye concentration and liquor ratio having the highest and smallest effects, respectively. Furthermore, Gladys and Olalekan (2021) created a model for textile color demand forecasting. The suggested approach uses a convolutional neural network (CNN) to extract hidden information from a photo collection. The K-means approach was used to extract colors from the recovered features. The research indicated that both techniques worked well and that correct color forecasting may considerably increase textile industry productivity and sales. Table 2 lists some dyeing studies.

In printing, Golob et al. (2008) created an ANN-based system to select pigment combinations for printing. They demonstrated the ability to use counterpropagation neural networks to identify color or pigment combinations in textile printing. To train the neural network, they used 1,430 cloth samples printed with ten colors. Once the neural network has been trained, determining new unknown data is simple and quick. As a result, a huge number of pigment and dye combinations can be determined.

Apart from color recipe prediction various researchers have focused on dyeing defects and dyeing difficulties. Huang and Yu (2001) classified seven categories of dyeing errors using image processing and fuzzy neural networks. Inconsistent coloring on the selvage and filling band in the shade are among the defects. Fuzzy neural classification system is built by an expert system using a neural network as the fuzzy inference engine. The inference engine is trained on sample data. This fuzzy neural network system uses fuzzy logic and neural networks to intelligently tackle pattern recognition and categorization problems. Color differences between samples are vital in defect segmentation and feature extraction. The standard color is used to differentiate fault areas from the dyeing environment. It may also be used to weaken a weak zone. It must thus be achieved under the same illumination circumstances. Fuzzification expands the feature space by converting input values into fuzzy sets. The right fuzzy sets may help differentiate classes in the function space.

Hussain et al. (2005) designed and created a technique to troubleshoot the dying difficulties. It is a “knowledge-based expert system” in which collecting dyeing data was the first stage. The expert system's primary architecture is knowledge. An expert system must accommodate domain knowledge. Soliciting, evaluating and interpreting information to help a human expert solve a problem. They are acquired in five phases. Next, the expert system must be represented. The expert system may be taught knowledge in several ways. They introduced knowledge in the form of rules in the aforementioned article. A rule has an antecedent and a result. The antecedent is the “if” component of a rule. The rule's right-hand side is referred to as the “then”. Hussain et al. (2005) evaluated their system. So although the expert system is not flawless, it is better than individual experts. The key reason for the expert system's superior performance over human experts is because the expert system evaluates all probable causes, whether less prevalent or more common. As a consequence, human experts are more prone to ignore probable mistake reasons than expert systems. The expert system beat human experts in identifying challenging dyeing faults.

SenthilKumar and Selvakumar (2006) suggested using a neural network to create desired shadow depth. The expected shade depth is an important factor to achieve in colored objects. It must be returned for rework or rejected if the depth produced is not anticipated. The neural network, which was trained using input and output parameters related to reactive HE dyes on cotton fabric, has a 1% error in dying time. The trained network offers the same error percent when tested with a different color and cloth. This was verified even when the input and output parameters selected were beyond the range used to train the network. As a consequence, the neural network can predict the main exhaustion and fixing times for high fatigue reactive dyes on cotton cloth. SenthilKumar (2007) proposed a feed-forward neural network model of CIELAB values. The following findings come from research on using neural networks to model CIELAB values. The L*a*b* values of the neural network built using input and output parameters are 2.0% off for vinyl sulfone dyes. The trained network has the same error percent when tested with dyes that were not used for training. The trained network provides the same error % even when the input and output parameters are beyond the range used for training. A value of 0–1.5 for DE*, the difference between calculated and predicted values, is allowed. Most of the samples above have a lightness difference between 0.7 and 0.4, which is acceptable. In spite of the higher lightness difference, all samples are acceptable owing to the lower DE. The built-in neural network model may be used to optimize dyeing parameters for any vinyl sulfone dye.

Khataee et al. (2009) used a neural network to model the removal of a textile dye from contaminated water using microalga Chlamydomonas species biosorption. The potential of Chlamydomonas sp. to biosorb and decolorize C.I. Basic Green 4 was discovered in this work. Decolorization was shown to be influenced by the initial dye concentration, algal concentration, pH, reaction duration and temperature. The decolorization efficiency increased with increasing temperature in the 5–45°C range. Alkaline pH is required for decolorization. Using a range of experimental conditions, the ANN model could explain the complex interaction process. A simulation using an ANN model can predict the process's behavior in different conditions. The neural network connection weights were also employed to generate a measure of input variable saliency, enabling the relative importance of input parameters on color removal efficiency.

Also, note that most dyeing professionals create an expert diagnosis of a dyeing problem based on real observation of the sample, which usually incorporates the “human touch.” The expert system appears to be favorable in that it does not necessitate this type of data. When compared to human experts, the knowledge-based expert system built during this study performed well, and the evaluation findings revealed the system's potential value and application in the cotton dyeing industry. However, like any other piece of software or expert system, the first version is far from ideal, and there is plenty of space for quantitative and qualitative development. This study will serve as a benchmark for establishing new computer-based diagnostic expert systems in textiles and further examining the system's potential for improvement.

4. Conclusion

AI has massive potential advantages. There will be less biased situations if machines are managed by artificial human intelligence-designed systems, which may lead to greater operational precision and product quality. AI in textile operations could help to solve the problem of human precision and quality variation. A human vision-based technique is used for the majority of industrial inspections. It is feasible to catch a variety of errors with 100% accuracy using image processing and neural networks. Because neural networking can handle different images of samples with a single computer coding or programming, there will be no need for various codes for different faults. A computer system can learn billions of combinations of errors that occur on a textile product using machine learning and big data programming. So, whatever the defect is, a preprogrammed system can detect it.

Through this review, it is found that yarn defects may be easily recognized using image processing and a preprogrammed system. After feeding the system with standard pictures, the machine learning system can detect yarn flaws such as slub, twist variation, snarl, foreign contamination and so on. Nonlinear regression, in combination with ANN, is useful in fabric testing and production control picture processing.

Further fabric flaws can be discovered using image processing, and nonlinear regression and ANN can predict various physical attributes from the fabric's structural data, such as air permeability and bending rigidity. For thermal applications, it is found that ANNs may save time and money by predicting the fabrics thermal insulation behavior before manufacture and testing. AI in fault detection and identification systems for garment inspection increases the production and quality of apparel.

The expert system outperforms human experts because it considers all possible causes, minor or major. As a result, human experts make more mistakes than expert algorithms. In difficult dyeing flaw detection, the expert system gives more reliable feedback. Image processing can individualize distinct shades and discover a problem with a small size in the dying process. ANN and machine learning can be used to prepare dyeing recipes, shade matching, dyed water treatment and shade depth determination. For resetting top-dyed mixed yarn from reflecting spectrums, MANN is considered a unique solution with less error factors and shorter instruction time which projects higher success percentage.

AI can also help to boost productivity, assist with fiber identification, create safer working conditions and estimate demand. Future applications of AI in the textile industry include virtual modeling of yarn from fiber properties (Cornell), prediction of yarn tensile properties and yarn unevenness (Fraunhofer Institute). In a nutshell, once the industry accepts and implements AI, it will be able to replace all human efforts with better accuracy. But one of the major limitations of AI is that the machines can perform only those tasks which they are designed or programmed to do, anything out of that they tend to crash or give irrelevant outputs.

Figures

Figure 1.

Classification of artificial intelligence

Figure 2.

Detailed classification of machine learning methods

Table 1.

Artificial intelligence modelling for fabric manufacturing

Process	Artificial intelligence method	Model output	Reference
Sizing	• Fuzzy • Neuro-fuzzy • ANN	Exit moisture, size add-on, number of end breaks, warp breakage rate in the weaving process	Dorrity et al. (1994), Kim and Vachtsevanos (2000), Zhang et al. (2015), Yao et al. (2005)
Weaving	• Fuzzy • ANFIS • ANN	Weft yarn insertion velocity, compressed air consumption, the strength transfer efficiency of warps and wefts, air permeability	Dayik and Colak (2004), Hussain et al. (2014), Malik and Malik (2010), Malik et al. (2017)
Knitting	• Fuzzy and ANN • ANN • Fuzzy • ANFIS	Residual bagging bend height, spirality, bursting strength	Jaouachi et al. (2010), Murrells et al. (2009), Shahid and Hossain (2015), Jamshaid et al. (2013), Hossain et al. (2016a, 2016b)

Table 2.

Artificial intelligence for overcoming dyeing problems

Type of problem	Artificial intelligence application	Reference (Year)
Environmentally benign dyeing of polyester fabric with madder: modelling by artificial neural network and fuzzy logic optimized by genetic algorithm	• ANN and fuzzy models	Haji and vadood (2021)
Color matching prediction for textiles dyeing and printing	• Automatic color matching prediction model, CMR-color, by incorporating three neural network models including typical CNN, MLP and ResNet to improve the capability of extracting high-dimensional features from spectral data	Chen et al. (2021)
Dyeing recipe prediction model for cotton fabric dyeing	• Based on the hyperspectral color measurement and a deep learning algorithm	Zhang et al. (2021)
To find the appropriate dyes to mix and their exact concentrations	• A new approach based on linear programming optimization is proposed to solve the textile color formulation problem	Moussa (2021)
Prediction of color difference for dyed fabrics (woven plain polyester fabrics)	• Evaluation model based on a support vector machine (SVM) Accuracy predicted if 98.2%	Zhang and Yang (2014)
Classify seven kinds of dyeing defects including filling band in the shade, dye and carrier spots, mist, oil stain, tailing, listing and uneven dyeing on the selvage	• Image processing and fuzzy neural network	Huang and Yu (2001)