AI Powered Anomaly Detection and IoT Automation for Improving Textile Manufacturing Quality Management and Productivity Levels

Maintaining consistent fabric quality is a major challenge for the textile industry because it relies on labor-intensive traditional processes that are prone to human error. High rejection rates, irregularities in the dyeing process, variations in fabric texture and irregularities in fiber strength are all examples of these inconsistencies. Manual inspection techniques are ineffective and frequently miss small flaws in real time, which increases material waste and lowers production efficiency, which is shown in Figure 2. Without an integrated monitoring framework, quality control is reactive rather than proactive, which exacerbates production delays and costs. This study suggests an IoT-driven real-time process monitoring and quality control system that combines automation, smart sensors and predictive analytics to address these issues. By streamlining the spinning, weaving, dyeing and finishing processes, this system seeks to increase operational efficiency and quality consistency.

Fig. 2.

Quality control process of clothes

The data collected in this study are now explicitly linked to specific textile processes and product types, rather than only to the geographical location. For each unit, measurements were conducted during clearly defined process steps: in Tiruppur, yarn tension was recorded on ring-spun cotton yarn during winding, and dye consistency was measured during knitted fabric dyeing; in Coimbatore, fabric density and humidity were monitored during grey woven fabric formation and finishing; in Surat, machine vibration was measured during shuttleless weaving, and fiber strength was tested prior to weaving; in Ludhiana, the temperature in the knitting hall and defect detection during knitted fabric inspection were collected; and finally, in Bhilwara, moisture content and thread elasticity were assessed during dyeing and finishing processes. This approach ensures that the sensor data collected are process-specific and directly relevant to quality evaluation, addressing the reviewer's concern regarding ambiguity in the previous dataset presentation. An overview of the information gathered is provided in Table 1.

The fabrics quality was evaluated using a number of key performance indicators (KPIs) that were obtained from sensor data. Fabric density was recorded in grams per square meter (GSM) and yarn tension was measured with optical sensors in Newtons per meter (N/m). Through the use of spectrophotometric sensors which measured color variance in terms of Delta E values, dye consistency was examined. In order to ensure ideal processing conditions during the weaving and finishing stages, humidity sensors were used to measure moisture levels in percentage terms. Accelerometers and thermal sensors were used to record machine vibration and temperature, respectively, which helped anticipate possible mechanical failures before they happened. High accuracy was guaranteed and thorough defect analysis made easier with this structured data measurement approach.

The issue is exacerbated by the fact that many sources - different sensors in factories in Tiruppur Coimbatore Surat Ludhiana and Bhilwara - make the databases in data collection areas extremely vulnerable to noise, as well as missing and inconsistent data. In other words, the first question that needs to be answered is how can large amounts of raw data be extracted from sensors in textile factories in Tiruppur, Coimbatore, Surat, Ludhiana, and Bhilwara and accurate and clean information about operation parameters be supplied. Figure 3 lists a number of data preprocessing methods. As seen in Figure 3-a, data cleaning can be used to eliminate noise and fix inconsistencies in the data. As seen in Figure 3-b, data integration combines information from various sources to create a logical data store like a data warehouse. Data reduction as illustrated in Figure 3-c can decrease the size of data by clustering, removing redundant features or aggregating. Variables typically have ranges that differ significantly from one another, however. To standardize the degree to which each variable influences the outcomes, data miners frequently normalize their numerical variables. Data transformation is the term for this procedure, which is depicted in Figure 3.

Fig. 3.

Data preprocessing

The research methodology covers problem identification, multi-stage data collection, sensor calibration, AI algorithm development, and experimental validation across 50 textile units over six months. Each production stage - spinning, weaving, dyeing, and finishing - was monitored using IoT-enabled sensors measuring yarn tension, fabric density, moisture content, machine vibrations, and dye absorption levels. Data preprocessing techniques included noise filtering, normalization, and dimensionality reduction to ensure analytical accuracy. For AI modelling, deep autoencoders and LSTM networks were implemented for anomaly detection and predictive maintenance, while federated learning enabled decentralized model training. Statistical methods such as Pearson correlation, ANOVA, and five-fold cross-validation ensured reliability and robustness. Comparative analysis with traditional methods and benchmarking against global studies provided additional validation of the Hybrid IoT–AI Framework (HIAF) proposed. This systematic approach ensures reproducibility, transparency, and methodological rigour.

In order to improve textile manufacturing processes, this study presents a Hybrid IoT-AI Framework (HIAF) that blends edge computing real-time sensor networks and predictive analytics based on deep learning. Fabric quality yarn tension moisture content and machine vibrations are all captured in real time by the method suggested, which combines a multi-layered sensor array made up of RFID tags, optical defect sensors, piezoelectric vibration sensors and humidity detectors. Equations 1 to 5 are expressed as follows

Anomaly Detection via the Autoencoder Loss Function: (1) LAE=1n∑i=1n(xi−x^ i)2 {L_{AE}} = {1 \over n}\sum\nolimits_{i = 1}^n {{{({x_i} - {{\hat x}_i})}^2}} where x_i is the original sensor input, x̂_i the reconstructed output, and L_AE represents the mean squared reconstruction error used for anomaly identification.

Predictive Maintenance Using LSTM-Based Time Series Forecasting: (2) LAE=1n∑i=1n(xi−x^ i)2 {L_{AE}} = {1 \over n}\sum\nolimits_{i = 1}^n {{{({x_i} - {{\hat x}_i})}^2}} where h_t is the hidden state, x_t the input sensor data at time t, W_h and W_x weight matrices, and b_h is the bias term.

Adaptive Federated Learning Model for Edge Processing: (3) LAE=1n∑i=1n(xi−x^ i)2 {L_{AE}} = {1 \over n}\sum\nolimits_{i = 1}^n {{{({x_i} - {{\hat x}_i})}^2}} where θtk \theta _{\rm{t}}^{\rm{k}} represents model parameters for edge device k at iteration t; η is the learning rate, and ∇Fk(θtk) \nabla {\rm{Fk}}(\theta _{\rm{t}}^{\rm{k}}) denotes the gradient update.

Digital Twin Optimization for Machine Parameters: (4) LAE=1n∑i=1n(xi−x^ i)2 {L_{AE}} = {1 \over n}\sum\nolimits_{i = 1}^n {{{({x_i} - {{\hat x}_i})}^2}} where C_opt is the optimal machine parameter set, Q_i represents observed quality metrics, and Qi∧ {\rm{Q}}_{\rm{i}}^ \wedge is the predicted quality output.

Defect Reduction Metric Calculation: (5) LAE=1n∑i=1n(xi−x^ i)2 {L_{AE}} = {1 \over n}\sum\nolimits_{i = 1}^n {{{({x_i} - {{\hat x}_i})}^2}} where DR is the defect reduction percentage, D_base the defect count in the conventional system, and DHIAF is the defect count using the HIAF approach proposed.

The efficiency of the HIAF suggested in improving defect detection predictive maintenance and process optimization in smart textile manufacturing is validated by this mathematical formulation. As opposed to traditional systems, HIAF processes sensor data at the edge layer using an adaptive federated learning algorithm which lowers latency and improves computational efficiency. The framework uses an anomaly detection model based on deep autoencoders to find flaws and irregularities in the spinning, weaving, dyeing and finishing processes. After processing, the data are sent to an analytics engine in the cloud which makes use of an LSTM network to predict possible flaws and suggest fixes. Furthermore, by simulating production scenarios a digital twin module makes proactive maintenance and process optimization possible. This hybrid framework maximizes real-time defect detection, predictive insights and machine control efficiency while minimizing data transmission overhead.

Despite the demonstrated benefits of IoT-enabled smart manufacturing, several implementation barriers were identified during pilot deployment. Integration issues primarily arose when linking legacy textile machinery with IoT-based sensors and edge-computing modules, as older production lines lacked standardized communication protocols such as MQTT or OPC-UA. To overcome this, middleware gateways were deployed for protocol translation, ensuring real-time interoperability without replacing existing infrastructure. Another key challenge involved staff competencies, since traditional operators lacked experience in handling IoT-enabled dashboards, predictive maintenance alerts, and AI-driven analytics platforms. A structured training program was introduced, covering sensor calibration, data interpretation, and cyber-physical system management, enabling a 30% reduction in operator intervention errors within three months of training completion. From a data security perspective, the large-scale collection of real-time production data raised concerns about potential cyber threats and intellectual property leakage. To address this, all data transmissions between edge devices and cloud servers employed AES-256 encryption and blockchain-based access control for secure storage and tamper-proof traceability. Furthermore, federated learning eliminated the need for centralized raw data storage, enhancing privacy by enabling on-device AI model training while sharing only model parameters across units. These mitigation strategies ensured seamless system integration, improved workforce adaptability, and robust data security, paving the way for large-scale, secure, and efficient adoption of IoT-enabled textile manufacturing.

The sensor data gathered from multiple textile production facilities offered insightful information about the characteristics of the fabric and the operational circumstances in various places, which is shown in Table 2. Key differences in the textile production process were identified through the examination of the following factors: fabric type, yarn tension, moisture content, dye absorption, fabric density, machine vibration defect rate, temperature humidity and machine speed. Cotton fabric from Tiruppur had the following characteristics: a machine vibration of 55 Hz, a defect rate of 3%, a yarn tension of 11%, a moisture content of 6%, a dye absorption of 85%, and a fabric density of 180 gsm. Despite having a lower yarn tension of 9.8 N/m and moisture content of 4.7 percent, the polyester fabric from Coimbatore showed a higher dye absorption of 92.1 percent and fabric density of 220 gsm.

Additionally, the fabric also maintained lower vibration levels and defect rates. In addition to having the highest yarn tension (13.2 N/m) and dye absorption (87.8 percent), Surat, which is well-known for producing silk, also had the highest defect rate (4.1 percent). With a yarn tension of 12.0 N/m and defect rate of 3.7 percent, Bhilwaras rayon fabric showed balanced properties, whereas Ludhiana, mostly used in wool textiles, had a moderate yarn tension of 10.9 N/m and dye absorption of 89.6 percent. According to these variances, various fabric types needed ideal processing conditions in order to preserve effectiveness and quality.

Table 3 examines the correlation between machine vibration and the defect rate. The highest correlation coefficient (0.87) was found in Tiruppurs spinning machines, which also had the highest vibration frequency (55.3 Hz) and defect rate (3.2 percent). Coimbatore weaving machines had a slightly lower correlation of 0.81, a vibration rate of 48.7 Hz and a defect rate of 2.9 percent. With a correlation coefficient of 0.92, weaving machines for Surat showed the highest vibration of 59.8 Hz and the highest defect rate of 4.1 percent, underscoring the crucial role that machine stability plays in preventing defects.

Table 3.

Machine Type	Machine Details	Vibration (Hz)	Defect Rate (%)	Correlation Coefficient (R)	Thread Tension (N/m)	Power Consumption (kW)	Efficiency (%)
Spinning		55.3	3.2	0.87	11.2	7.5	92.4
Weaving		48.7	2.9	0.81	9.9	6.8	91.2
Dyeing		59.8	4.1	0.92	13.5	8.2	89.7
Finishing		52.1	3.5	0.85	10.8	7.1	90.5
Mixed Process		57.4	3.7	0.88	12.1	7.8	91.0

Strong correlations between vibration and the defect rate were also observed in Ludhiana finishing machines and Bhilwara mixed-process machines with respective values of 0.85 and 0.88. The significance of vibration monitoring in lowering defect rates and enhancing machine efficiency was emphasized by this study.

The reduction in machine failure rates across spinning, weaving, dyeing, and finishing units was directly attributed to the continuous real-time monitoring of critical fabric and machine parameters, including yarn tension stability (N/m), moisture content (%), dye absorption consistency (%), machine vibration frequency (Hz), and thermal operating conditions (°C). For instance, IoT-enabled yarn tension sensors maintained operational limits within ±1.5 N/m compared to ±3.0 N/m under conventional monitoring, thereby preventing thread breakages that previously led to loom stoppages. Similarly, moisture content fluctuations were reduced by 12.1% after IoT deployment (Table 9), ensuring optimal fabric elasticity and reducing roller slippage incidents. Vibration monitoring detected early mechanical imbalance, correlating strongly (R = 0.87–0.92) with defect rates across weaving and dyeing machines (Table 3), while thermal sensors prevented overheating-driven shutdowns by maintaining operational temperatures within 26–29°C. Collectively, these parameters enabled predictive maintenance algorithms to anticipate faults, schedule timely interventions, and achieve a 75% failure rate reduction in finishing units and an overall 63.6% reduction in connected machine processes (Table 4).

AI-based defect detection accuracy varied across fabric types, with the highest detection accuracy recorded for cotton at 98.3%, followed closely by rayon at 97.8%, which is shown in Table 5. Polyester and wool fabrics achieved accuracies of 97.6% and 97.2%, respectively, while silk exhibited the lowest detection accuracy of 96.9%. The processing time for AI-based detection ranged from 115 ms for polyester to 130 ms for silk.

Table 5.

Fabric Type	Fabric Images	Yarn Breakage (%)	Dye Mismatch (%)	Fabric Density Variance (%)	Detection Accuracy (%)	Processing Time (ms)	False Positives (%)	Model Efficiency (%)
Cotton		3.2	2.5	4.1	98.3	120	1.2	95.5
Polyester		2.9	2.1	3.7	97.6	115	1.5	94.8
Silk		4.1	3.8	5.2	96.9	130	2.1	93.9
Wool		3.5	2.9	4.5	97.2	125	1.8	94.2
Rayon		3.7	3.2	4.8	97.8	118	1.6	94.6

At 1.2% cotton had the lowest false-positive rates, while silk had the highest at 2.1%. Silk registered the lowest efficiency at 93.9 percent, while cotton achieved the highest efficiency at 95.5 percent following a similar pattern. These results demonstrated how reliable AI-driven defect detection systems are at guaranteeing the production of high-quality fabrics.

Table 6 and Figure 7 illustrate how the optimization of digital twins on machine parameters led to notable enhancements in important operational metrics. Following optimization, yarn tension rose by 14%, which resulted in a yield improvement of 39% and a decrease in quality defects from 4% to 2%.

Fig. 6.

Experimental analysis

Fig. 7.

Optimization impact

Defects decreased from 5.1 percent to 3.2 percent, yield increased by 37.3 percent, and moisture content optimization improved by 12.1 percent. A 4.3 % increase in dye absorption led to a 33.9% percent decrease in defects and a 33.9% percent improvement in yield. A decrease in defects from 5.7 percent to 3.5% and a 38.6 percent increase in yield resulted from an 8.6% improvement in fabric density. These outcomes showed how well digital twin technology works in increasing the productivity of textile manufacturing.

The implementation of IoT-based monitoring systems significantly reduced defect rates across various fabric types, which is shown in Figure 8. Cotton initially had a defect rate of 8.5%, which decreased to 5.2%, achieving a 38.8% reduction. Polyester's defect rate dropped from 7.2% to 4.5%, marking a 37.5% improvement. Wool, which initially exhibited a 9.3% defect rate, saw a decline to 6.1%, reflecting a 34.4% reduction.

Fig. 8.

Defect reduction

Similarly, silk's defect rate was reduced from 10.5% to 7.0%, and rayon experienced an identical reduction percentage of 33.3%, with defects lowering from 8.1% to 5.4%. Denim, which started with a 6.7% defect rate, demonstrated a substantial reduction to 4.1%, achieving the highest improvement rate of 38.8%. These results highlighted the effectiveness of IoT in minimizing defects and enhancing fabric quality.

The adoption of IoT-based solutions led to a significant improvement in first-pass yield (FPY) across various production stages, which is shown in Table 7 and Figure 9. In the spinning stage, FPY increased from 85.2% to 92.1%, reflecting an 8.1% improvement. The weaving stage experienced a rise from 82.7% to 90.4%, achieving a 9.3% enhancement. Dyeing demonstrated the most substantial improvement, with FPY increasing from 78.5% to 88.0%, marking a 12.1% gain.

Similarly, the finishing stage saw an increase from 80.1% to 89.2%, resulting in an 11.4% improvement. These enhancements underscored the effectiveness of IoT in optimizing production processes and reducing rework, ultimately improving manufacturing efficiency.

Fig. 9.

FPY before and after IoT

Table 8 highlights the impact of IoT-based monitoring on yarn tension stability across different fabric types, showing a significant reduction in standard deviation, which enhances fabric quality and production efficiency. Before IoT integration, yarn tension deviation ranged from 2.4 N/m (denim) to 3.6 N/m (silk), with intermediate values for cotton, polyester, wool, and rayon.

Traditional tension control methods led to inconsistencies, affecting fabric uniformity and increasing waste. IoT-driven systems improved stability by 41.7%–46.4%, with the highest reduction observed in cotton (46.4%) and denim (45.8%), attributed to real-time data acquisition, automated adjustments, and predictive maintenance. Polyester, wool, and rayon also showed notable improvements of 44.0%, 43.8%, and 43.3%, respectively.

Table 9 examines the moisture content optimization. Wool exhibited the highest moisture optimization of 14.6%, followed by silk at 12.8% and rayon at 11.5%. Cotton and polyester saw improvements of 11.1% and 10.8%, respectively, while denim recorded the lowest optimization at 9.8%. These improvements ensured consistent fabric properties and minimized defects related to moisture fluctuations.

Among various manufacturing processes, packaging demonstrates the highest efficiency, with the lowest defect rate (3.5%), highest resource utilization (91.3%), and minimal energy consumption (180 kWh), which are shown in table 11. Finishing also excels, maintaining a low defect rate (4.1%) and highest resource utilization (90.2%). Knitting and Sewing show moderate performance, with relatively balanced defect rates and energy consumption, which is shown in Table 10. Dyeing consumes the most energy (280 kWh) and has the highest defect rate (7.8%), highlighting inefficiencies. Packaging emerges as the most efficient process, with an optimal production time, low defect rates, and minimal energy consumption.

Table 10.

Process	Process Image	Avg. Production Time (hrs)	Defect Rate (%)	Resource Utilization (%)	Energy Consumption (kWh)
Knitting		11.5	5.2%	87.3%	230
Dyeing		14.2	7.8%	82.5%	280
Printing		12.7	6.3%	85.2%	250
Cutting		9.8	4.9%	89.7%	210
Sewing		13.4	5.6%	86.1%	240
Finishing		10.9	4.1%	90.2%	200
Packaging		8.6	3.5%	91.3%	180

The implementation of IoT-driven solutions has significantly enhanced overall quality metrics in manufacturing processes, as illustrated in Table 11. The integration of IoT technologies has led to a substantial reduction in the overall defect rate from 8.3% to 5.6%, marking a 32.5% improvement. Additionally, the first-pass yield, which indicates the percentage of products meeting quality standards without rework, increased from 81.6% to 90.1%, reflecting a 10.4% improvement. Furthermore, production efficiency saw a notable enhancement, rising from 74.8% to 85.2%, corresponding to a 14.0% improvement.

To evaluate sustainability and resource efficiency, environmental indicators such as energy consumption, water usage, CO₂ emissions, and fabric waste generation were measured before and after IoT deployment across all 50 production units. Results indicated a 24.8% reduction in energy consumption, with average electricity use decreasing from 1250 kWh to 940 kWh per production batch (Table 12). Water consumption during dyeing and finishing processes decreased by 27.1%, falling from 850 L to 620L per 100 m² of fabric processed, primarily due to optimized dye bath cycles and moisture regulation enabled by real-time monitoring sensors. CO₂ emissions per production unit decreased by 25.6%, attributed to lower energy demand and reduced machine downtime. Additionally, defective fabric waste dropped from 8.1% to 5.2%, achieving a 35.8% reduction in textile scrap, thereby lowering landfill burden and improving overall resource utilization. These findings demonstrate that the IoT–AI framework proposed not only enhances production efficiency but also contributes significantly toward sustainable textile manufacturing by minimizing environmental impact.

An economic feasibility analysis was conducted to evaluate the costs and benefits of implementing the IoT–AI system across small, medium, and large-scale textile production units. The cost structure included sensor networks (RFID tags, optical sensors, humidity sensors, vibration sensors), IT infrastructure (edge computing nodes, cloud storage, analytics servers), and annual maintenance. On average, the initial installation cost per production unit was USD 45,000, comprising USD 18,000 for sensors, USD 20,000 for IT infrastructure, and USD 7,000 for software licensing and maintenance. However, operational savings from reduced defect rates (up to 32%), lower downtime (63.6%), and optimized energy/water consumption (24–27%) resulted in an average annual saving of USD 24,800 per unit, leading to a payback period of 18–22 months depending on the production scale.

For large-scale units processing over 10,000 m² of fabric per month, ROI was achieved in less than 18 months due to higher production efficiency gains, while medium and small-scale units required 20–22 months. Table 13 summarizes the cost-benefit analysis across different production scales, confirming the financial viability of IoT-driven textile automation.

In this study, the performance of several AI algorithms was evaluated and compared, as shown in Table 14: Comparative AI Performance. The results revealed that the Hybrid IoT–AI (HIAF) model outperformed all other algorithms across key metrics. It achieved the highest accuracy (98.3%), precision (97.9%), recall (98.1%), and F1-score (98.0%). Furthermore, the HIAF model demonstrated the lowest processing time (120 ms) and the lowest rate of false positives (1.2%), indicating its superior efficiency and reliability. The LSTM Forecasting algorithm followed closely, securing the second-highest scores in most categories, with an accuracy of 97.1%. In contrast, Random Forest and Support Vector Machine (SVM) showed lower performance metrics, with Random Forest achieving an accuracy of 92.8% and SVM 91.2%, while also exhibiting slower processing times and higher false positive rates. Overall, the Deep Autoencoder and LSTM forecasting models performed significantly better than traditional machine learning algorithms like SVM and Random Forest.