The manufacturing platforms aim to satisfy the needs of the shop floor. This manufacturing platform can reduce production costs, detect anomalies, schedule production plans, or manage production processes in real-time [17]. Platforms generally allow users to customize, extend, and easily access system services. The emerging manufacturing platform is a cloud-based platform classified as a service-oriented manufacturing model whose primary purpose is to cooperate or share resources [18]. Liu et al. [19] utilized the industrial internet of things (IoT) technologies in a cloud manufacturing system to operate the core functions of a manufacturing system. The system is designed to provide high efficiency in the decision-making processes of the manufacturing. However, the SMEs cannot adopt the cloud manufacturing system because of its limitation in launching digital transformation. Digital transformation includes physical changes on the shop floor and virtual elements of the manufacturing systems [20]. To initiate these changes, producers must be able to install sensors or IoT devices on the shop floor to collect or control real-time data of the manufacturing automatically. Regarding the limitations in the innovation of SMEs, Andalib et al. [21] proposed a blockchain-based manufacturing platform for SMEs in Bangladesh to save innovation costs. Han et al. [22] implemented a data science platform to provide collaboration opportunities for SMEs. Data science platforms are expected to achieve data mining, analysis, sharing, and pre-processing. Along with the concept of a service-oriented manufacturing platform and low-priced solutions, we design and implement the novel concept of a manufacturing service platform in this paper.

Numerous kinds of research have been conducted to solve the problem of shop floor scheduling, also known as the job shop scheduling problem (JSSP). The JSSP is proved to be the NP-hard problem [23]. Solving the NP-hard problem is to find a way to allocate n jobs on m machines efficiently. Each job Ji is divided into O_ij operations, and O_ij are assigned to the machine M_ij with processing time Time_ij, as shown in Table 1. The well-planned schedule achieves the minimum makespan compared to other schedule results. The scheduling results are represented in Gantt chart format as shown in Fig. 1. Optimal scheduling between tasks and machines contributes to efficient production by minimizing the waste of resources and production time in factories. The JSSP is based on the static environment. The process of each job has conditions that it can be performed on specific machines, and one job cannot be performed on multiple machines at the same time.

This aspect of the JSSP has some limitations on the existing field systems [24]. The GA can be applied to various tasks or problems to function as an optimizer. Omar et al. [25] utilized stopping criteria in GA to solve the JSSP. The proposed GA includes a critical block neighborhood and distance measuring method to stop the crossover and mutation step iteration. Although these stopping criteria are used to reduce the running time of the GA, the data used in the study are small and, therefore, unsuitable for large-scale problems. The GA first randomly generates a population of chromosomes. The generated chromosomes are evaluated for the effectiveness of the solution. The next chromosome is then generated by selecting and combining the original solution to compare the original solution with the newly generated solution. Generating the new generation requires the operation of three leading operators; reproduction (or selection), crossover, and mutation. The reproduction makes copies of improved strings for generating a new population. To make this platform optimal for textile manufacturing, we applied the GA to solve the JSSP in the proposed system and set the deadline constraints to determine possible production schedules for each producer.

Scheduling textile manufacturing has been studied in numerous studies, and the textile processes used are different. Saydam et al. [26] focused on the dyeing process to efficiently schedule by maximizing the utilization of machines. Guo et al. [27] used genetic optimization and mixed integer programming to solve the scheduling problem for apparel manufacturing. Mourtos et al. [28] proposed the scheduling method for the weaving process in textile manufacturing by splitting the jobs and setup the resource constraints. These resource constraints limit machines from processing all of them simultaneously because of the limited number of workers on the shop floor. Wang et al. [29] utilized GA for generating an automatic scheduling method to minimize downtime and gaiting load to obtain efficiency in weaving enterprises. Perret et al. [30] proposed a two-stage algorithm using tabu search and job-wise shift operator to utilize it in the mass customized production processes and decrease the production cost. These approaches were made to minimize the production makespan andbenefit the manufacturer by lowering the production cost, energy, and workforce. In this sense, scheduling the textile manufacturing process to minimize production costs and the energy consumption is vital for SMEs. This approach will be beneficial not only for textile manufacturing SMEs but it will be significantly impact general manufacturing SMEs. In this paper, the test data focuses on the weaving process of a small textile manufacturing company because of the difficulty of actual data collection.

In this paper, the weaving process data of SME-owned manufacturing are used to test the GA model. There are three main operations in textile manufacturing: weaving and knitting, dyeing, and sewing and post-processing. First, the weaving process produces simple fabrics for clothes or shoes. This simple fabric results from various patterns and designs that require a high level of technology. Depending on the type of weaving, the processing time may be longer than other designs for the finest and sheer fabric quality. Second, the dyeing process dyes the fabrics. This process requires diverse temperature settings that require more energy than other processes. Finally, the sewing and post-processing finalize the textile manufacturing. On the other hand, if the product the buyer wants to order is fabric, sewing and post-processing are not always required—only apparel goods are required to proceed with the process. Since the weaving process is longer than other processes, efficient scheduling is required. Generally, the weaving process takes more than three months on one machine. In some cases, it lasts more than six months. These practices of the weaving process can delay the production time of manufacturing. According to a condition of JSSP, the operation should be declared before executing the scheduling. However, as the study focuses on solving real-world problems, textile manufacturing does not always cover all three processes on one production site. Thus, parallel operation of the same job on different machines is allowed in this weaving process scheduling.

The proposed approach is tested by scheduling examples from the SME-owned weaving factory. The raw data were collected for one year, from January 1, 2020, to December 31, 2020. The total data are 62,452 records, and we used 11% of the total data in this study. There are three conditions in data selection. First, the data selection was based on machine numbers from 1 to 20. Second, among the selected machines, the jobs need to have lasted more than three days. Third, the chosen machines have produced two or more products on each machine. Utilizing the inputs for the GA model, the raw data are required to be preprocessed. There are seven columns in raw data: work date Wd, machine name, work groups Wg, product name p, density Den_p, rpm Rpm, and demand. Each product has its Den_p value. In addition, depending on p, Wd, Wg, and machine, Rpm is different from each other by a difference of ±3. Hence, we calculate the mean value of Rpm for input data generation. Preprocessing the raw data for GA model input requires a name/information of each product Prod_i, machine number m, density of Prod_i DenProdi, mean of rpm of Prod_i RpmProdi¯, and processing time PTTij data. The notations of input data for JSSP are defined in Table 2.

There are three steps in raw data preprocessing. First, the number of products is divided based on the input data m set by the producer. Second, the processing time is the primary input data for GA and is generated through Algorithm 1. After dividing the J_n, assigning the processing time of each M_m is essential. The division process is defined and described in Algorithm 1 lines 19. Each divided task T_ij has different processing time values depending on the total demand of the orders. The processing time PTTij of the j^th task for a Prod_i is calculated as Algorithm 20 line 24. Calculating the PTTij requires average rpm. Through the function in Algorithm 1 lines 1 to 7, the value of RpmProdi¯ can be obtained. The data generation function calculates the processing time based on the history data set of each product prod_l, machine list ml, demand of the order, and work start date StProdi. The number of orders can be listed in each order from one to infinite. Therefore, each order has different prod_l, ml, DenProdi, RpmProdi¯, demand, and StProdi. In this sense, it is important to compute the PTTij and divide T_ij based on the m. Moreover, since the weaving process has only one operation, we defined the operation in JSSP based on the real dataset by allocating jobs J_n on M_m, as shown in Table 3.

The final step of the raw data preprocessing is to save a list of input data into the .csv file. After this process, the GA model is ready for the execution process. In the execution process of the GA model, setting up the three values is important. The three inputs are population size, iteration count, and mutation rate. The initiation of the model begins with setting up the population size to generate a set of chromosomes so that the first operation of the algorithm, selection, can be used. Iterations are initiated depending on the population size, and each iteration randomly generates a set of chromosomes within the range multiplied by m and J_n. These randomly generated chromosomes are repeated within the iteration count until an optimal solution is found. However, when a buyer provides the work end date in the platform, the scheduling iteration of GA terminates when the target value of the work end date has been met. Finally, the mutation operator randomly generates new information in the children’s generation of the chromosome; some are inherited from the parent generation. The mutation function first swaps the two elements of the parent generation with the next generation. Then it randomly selects an element and sets it to its children’s generation. This mutation operator is expected to diversify the population and is beneficial for local search. Operating the mutation function requires mutation probability, which defines the size of the parts that can be changed during a mutation function. During the selection, the calculation of makespan computes a fitness value to rank the best solution. The population stacks the next best solution during the three main processes of algorithms by calculating the makespan of each population.

Then the producer can execute the GA model. According to the Table 3, seven products are grouped as Order 1 to generate the job order and processing timetable, which are required to execute the scheduling model of the platform. Depending on each J_n, there are restrictions on machine allocations due to beam and textile types. In this sense, not all J_n are evenly distributed to every M in this study. The job shop scheduling engine execution process requires input m, StProdi, and Prod_i. In addition, DdProdi is required to schedule execution process to determine what job can be done within the buyer’s expected DdProdi. As the scheduling execution returns result of makespan and Gantt chart data, the parameter of the result includes x-axis as processing time, y-axis as machine number, producer P, StProdi, DdProdi, and J_n.

In this paper, the data are divided into individual orders according to the average production cycle (three months/six months) of each product and 20 machines from the actual data. We test the feasibility of the GA model in textile manufacturing by comparing manual scheduling from field experts in the weaving process in textile manufacturing owned by SMEs, i.e., manual scheduling. We used Intel Core i-7 with 32 GB memory at a speed of 2.9 GHz and Python 3.9 to solve the JSSP and run the platform. Since the products are divided into multiple J_n, the blank area of each machine can process other products or be empty to prepare for the next J_n. This aspect is also adopted in the GA model in this paper. Table 4 indicates the processing time of each order. Fig. 3 is the scheduling result in Gantt chart format. Fig. 3a is the Order 1 data that utilizes 40 jobs and 20 machines to schedule. The work start date starts on May 10, 2020 and ends on September 6, 2020. Fig. 3b is the Order 2 data that utilizes 38 jobs and 13 machines to schedule. Fig. 3c is the Order 3 data that utilizes 27 jobs and 8 machines to schedule. Orders 1, 2, and 3 overlap each other. The scheduling assign tasks to not to affect the existing Orders by saving the former schedule in the job shop scheduling engine database and setting up the distinct work start dates. Finally, Fig. 3d is the Order 4 data that utilizes 63 jobs and 19 machines to schedule.

As shown in Fig. 4, the comparative results between the GA model and manual scheduling are similar, but the values are decreased. Order 1 has a smaller makespan of 152 than 155 of manual scheduling. However, according to Table 4, GA is not always having a minimum makespan compared to the manual scheduling. The manual scheduling for Order 3 is smaller at a makespan of 135 than for the GA model of 136. This result was obtained because the jobs were evenly assigned to 8 machines for three months. Although there is a difference of 1-day, it is an acceptable period that does not significantly impact the overall end date of orders. Therefore, the use of the GA in this platform is feasible since the job shop scheduling engine aims to minimize the machine usage per order and obtain optimal scheduling. In addition, in this paper, the condition that other products do not preempt the operation sequence of each product was defined in advance, including the machine usage for the textile manufacturing-based JSSP. This feature is closely related to the use of GA. It is crucial to rearrange the population of textile manufacturing JSSP instance in order of each product through GA. Since the GA has a fitness value to rank the best solution between the randomly generated populations, the textile manufacturing JSSP instances will be well suited for obtaining the best solution regarding the precondition of this paper. Moreover, the scheduling system takes 2 s to schedule productions on average. This speed in production scheduling will save workers time and labor costs in scheduling jobs.

The job shop scheduling execution is desired to get the optimized schedule for the production plan. Optimized schedules are based on the shortest makespan. Producers are expected to be able to run as many jobs as possible. When the scheduling engine generates the makespan, producers select the jobs that end within expected due dates. The job shop scheduling execution process requires several input data to be prepared for its algorithm. We used python language and Django Web application framework for scheduling and matching systems in this platform. Fig. 5a is the screen that shows how producers set their input data to the job shop scheduling engine. The input data required by the JSSP algorithms is processing time and machine allocation settings. In addition, the input data includes the start date and end date of the job as the criteria by which the job can be processed. Fig. 5b shows how the scheduling results came out after executing the scheduling function with the input data for the weaving process. If the jobs are deployed before the expected end date set by the producer, it will appear on the right side of the screen, so the producer can send a notification to the user to check what producers are available.

The JSSP algorithm in the platform is based on the GA model. Shao et al. [16] proposed self-supervised long-short term memory (SS-LSTM) to minimize the makespan by predicting the following action with high feature extraction of hidden patterns. The method used GA to generate the optimized solution for SS-LSTM and test datasets. This method outperformed when it was tested with the large-scale JSSP instances. Chang et al. [31] also utilized the GA for generating a hybrid genetic algorithm (HGA) for solving the distributed and flexible job-shop scheduling problem (DFJSP) in manufacturing. This proposed method used benchmark data of various manufacturing and obtained its feasibility by comparing the scheduling results with the previous studies that used the same benchmark manufacturing data. Liu et al. [32] proposed a refined genetic algorithm to solve the problem of the traditional GA algorithm and generate an improved encoding method to reflect the real-world environment fully. This newly developed encoding method is utilized to diversify chromosomes in GA, and the scheduled results were close to real fastener manufacturing systems in Taiwan. Although these GA models have been tested for feasibility in real-world manufacturing, they are limited in practical usage. Unlike the studies mentioned above, the proposed platform in this paper tests its feasibility in the weaving process of SMEs textile manufacturing and is also implemented as a platform that buyers and producers can use. Therefore, since the platform includes real-world data from actual manufacturing and tests the feasibility of the GA model based on it, it has significance in directly applied to real-world manufacturing. Table 5 compares related studies with the scheduling and utilization methods of this paper.