Impact of green technology innovation based on IoT and industrial supply chain on the promotion of enterprise digital economy

Corporate digital innovation research

Most scholars do not have a unified understanding of digital innovation in manufacturing. Most give different explanations based on different perspectives, mainly based on the three levels of the innovation process, the outcome and the integrated study of both. At the level of digital innovation results, Sahut, Dana & Laroche (2020) consider digital innovation as the recombination of digital and physical components to produce new products, which implies value creation and performance improvement through digital technology and business model innovation. At the level of digital innovation process, Lyytinen, Yoo & Boland (2016) on the other hand, consider digital technologies and tools as operational resources and embed them in the innovation process through digital connectivity and digital convergence. Nambisan et al. (2017) argue that digital technologies and digital tools give rise to digital innovation, that the creation of new products and innovative business processes through digital technology is the essence of digital innovation, and that the results of digital innovation can be achieved without digitization, as long as the innovation process directly or indirectly utilises digital technology. Leydesdorff & Ivanova (2016) argue that the evolution of the open innovation model in the digital era is no longer exactly like traditional innovation in general, and Oborn et al. (2019) argue that innovation in the digital era is more likely to occur through the spatial and temporal distribution of emerging heterogeneous networks, forming a dynamic process of feedback loops. In emerging heterogeneous networks, developing a platform innovation model and innovation will become more uncertain and complex due to the heterogeneity and uncertainty of innovation agents in the network platform (Oborn et al., 2019). Digital platforms created by digital technologies blur the boundaries of innovation processes and outcomes and encourage the growth of distributed and combinatorial innovation, according to Yoo, Henfridsson & Lyytinen (2010), who argue that digital technologies will change the nature of products and services and that digital innovation is essentially a reorganization of digital technology at the integrated consideration of digital innovation outcomes and processes.

The current digital economy is mainly a forward-looking study of technology’s development and application scope based on the above innovation process and results. With the continuous application of IoT, the above ideas are realised through artificial intelligence and other methods. Applying cutting-edge technology to business development can encourage and support the digital economy.

Enterprise supply chain optimization research

Inventory optimization is using inventory control to reduce a company’s inventory costs, which is crucial to reducing the capital tied up in the company by lowering inventory costs. Inventory control involves the issues of when to restock raw materials and goods, determining the restocking amount, and safety stock. Gunasekaran et al. used three different MCDM methods for ABC analysis to determine the appropriate product for each inventory category. They applied it to inventory management at a large automotive company (Kartal et al., 2016). Sitompul et al. (2008) studied the relationship between demand variability, capacity constraints, and safety stock and proposed an alternative model for setting up safety stock in a supply chain with a sufficient capacity. Galli et al. (2021) proposed an inventory optimization model for the medical inventory domain to minimise inventory levels and the need for emergency replenishment. Duarte, Gonçalves & Santos (2021) developed a mixed integer linear programming model and solved it with a solver based on a branch-and-cut algorithm. Finding the optimal production and inventory strategy in a multi-product baking unit was solved. System dynamics models have also been used to solve inventory problems due to the complex dynamics of supply chain multilevel inventories, such as nonlinearity, dynamics, and delays. Poles (2013) used a system dynamics simulation modelling approach to model the production and inventory system for remanufacturing. They explored the dynamic remanufacturing process to give inventory and production improvement strategies for the simulated system. Making decisions for uncertainty is an essential issue in supply chains, where the processes in the flow of supply chain goods and services contain many complex decision processes and information barriers. To improve supply chain visibility, many leading organizations share information from both ends of their supply chains with their suppliers, and supply chain management is becoming increasingly data intensive (Ali et al., 2017). Realising the growing importance of data in supply chains, supply chain practitioners have tried every possible way to improve them. Machine learning is considered one of the methods. Machine learning algorithms have become a new hot spot for supply chain optimization solutions research due to their nonlinear solid data processing capabilities and predictive capabilities in unsupervised domains. Cavalcante et al. (2019) studied different machine learning algorithms in supplier selection, combining machine learning and simulation techniques to improve supplier delivery reliability. Priore et al. (2019) used a random forest algorithm to determine the optimal replenishment strategy for different commodities in a three-level supply chain to reduce the supply chain bullwhip effect. Wang et al. considered the effects of power optimization and cost factors. They designed and developed a statistical cost-based approach based on multiple costing frameworks using machine learning models (SCM-MLM) (Wang & Zhang, 2020). Martínez et al. (2020) developed a machine learning framework for predicting whether a company’s customers will buy again in the future. Huber et al. (2019) proposed a data-driven machine learning and quantile regression-based prediction method and applied it to point-of-sale data from a large German bakery chain.

It is clear from the study of supply chain optimization and the growth of the corporate digital economy that various industries are investigating using artificial intelligence (AI) technology throughout the whole cycle of their enterprise operation to increase efficiency. The application of a large amount of data also brings the data economy come into being. Scheduling for the corporate supply chain provides inventory optimization, identifies commodity products while optimizing inventory, and allows the supply chain to play a part in the enterprise’s digital economy. Only accurate scheduling can achieve green development.

Framework establishment of Yolov4-based warehouse goods identification analysis

Convolutional neural network (CNN) is the primary tool for processing image information. The training process of CNN models is learning the relationships between image pixels, which are reflected in the model parameters through convolutional operations. The convolution process of 3D convolutional neural networks is shown in Eq. (1).

(1) $$f_{out}(v_{ti})=\sum _{v_{ti}\in B(v_{ti})}{\displaystyle \frac{1}{Z}}fe{a}_{in}(v_{ti})\cdot \omega (l_{ti}(v_{ti}))$$

The 3D convolutional neural network’s unique kernel structure can learn temporal and spatial dimensions data. 3D convolutional neural networks can be trained and used immediately using raw video data as the network’s input, removing the requirement to first remove optical flow and video frame data. Inventory management, sensitive to time information, is one application where 3D neural networks are highly successful at detecting content.

The YOLO target detection algorithm is a family of efficient single-step target detection algorithms proposed by Redmon & Farhadi (2018). The core idea is to perform the bounding box regression and identify the category to which the input image belongs directly in the output layer. With the continuous iteration of the algorithm, the current most popular one is the YOLOv4 method. YOLOv4 has various applications in various industries, such as retail, transportation, and security. YOLOv4 can be used in retail for product recognition, stock management, and customer analysis. For example, it can be used to track the movement of products in a store, analyse customer behaviour, and detect product placement errors. YOLOv4 can be used in transportation for autonomous driving, traffic management, and vehicle tracking. It can detect objects such as pedestrians, vehicles, and traffic signs and provide real-time information to autonomous vehicles. It can also be used to monitor traffic flow and notice accidents. In security, YOLOv4 can be used for surveillance, crowd monitoring, and facial recognition. It can detect suspicious activity, monitor crowds, and identify individuals. It can also be used to track the movement of people and vehicles in restricted areas. Overall, YOLOv4 is a powerful and versatile object detection algorithm that can be used in various applications. It’s high accuracy and speed make it an ideal choice for real-time applications that require fast and reliable object detection.

This version firstly incorporates the cross stage partial network (CSPNet) based on the YOLOv3 backbone network Darknet, which reduces memory consumption. Secondly, YOLOv4 uses the Mish activation function to reduce the computational cost. The expression of the Mishap function is shown in Eq. (2).

(2) $$\mathrm{M}\mathrm{i}\mathrm{s}\mathrm{h}=x\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(\ln \left(1+e^x\right)\right)$$

It can be seen that the Mish activation function does not take an utterly truncated value at negative values, it allows the inflow of negative values with relatively small gradients, and the Mish function is unbounded by this feature, avoiding the problem of gradient saturation (at the limit position, the gradient tends to 1). At the same time, the curve of this function is smooth and continuous compared with the traditional Relu function, and the derivative is better, which helps to improve the localization effect on the target. Besides, adding the Squeeze-and-Excitation Net (SENet) (Hu, Shen & Sun, 2018) to the deep network adds an attention module to the channel feature map to enhance the target’s weight. Finally, the CSPDarknet53 is used as the backbone network. YOLOv4 neck uses the Path Aggregation Network (PANet) with the Spatial Pyramid Pooling (SPP) layer instead of the original Feature Pyramid Network (FPN). This module alleviates the problem of losing the location and contour information of the target to a certain extent. Based on the above advantages, this method improves the target detection accuracy based on YOLOv3.

PSO-SVM-based enterprise demand forecasting

For enterprise demand, its inventory changes according to specific patterns, and these patterns are related to various factors such as season, economic level, etc. The optimal way to solve such problems is to complete demand forecasting based on historical data to explore the data patterns (Huang et al., 2021). This article uses the more classical SVM method in machine learning to complete enterprise demand forecasting. Support vector machines (SVM) is a supervised learning method, and this model can show strong processing ability and high performance in dealing with some data with high dimensionality. For the SVM model, the core of its algorithm is to find the optimal hyperplane, which is to find the parameter that satisfies the distance from the interval sample points to the hyperplane and the maximum $\omega$ and $b$. The classifier corresponding to the decision boundary that maximizes the interval is a hard interval support vector machine. Maximising the interval becomes a problem of minimizing $\omega$ The minimization problem, according to its principle, i.e., to select the optimal $\omega$ and, we can obtain the loss function as shown in Eqs. (3) and (4).

(3) $$\underset{\omega ,b}{min}\frac{1}{2}{\omega }^2$$

(4) $$\begin{array}{c}\\ \mathrm{s}.\mathrm{t}.1-y_i\left({\omega }^T{x}_i+\mathrm{b}\right)\le 0,i=1,2,\dots ,n\end{array}$$

The optimization problem represented by the above two functions is a convex quadratic programming problem. By adding Lagrangian multipliers to each constraint $\alpha ={\left({\alpha }_1,{\alpha }_2,\dots ,{\alpha }_m\right)}^T,{\alpha }_i\ge 0$, the objective function is transformed to solve $L\left(\omega ,b,\alpha \right)$ the extreme value problem, concerning $\omega ,\mathrm{b}$ solving for the minimum values and the $\alpha$ solving for extreme values. In practical applications, not all issues are linearly differentiable, so a soft interval is needed for error correction, i.e., the introduction of ${\xi }_i$ Relaxation variables enhances the model capability, and the original problem after the introduction of relaxation variables can be expressed by Eqs. (5) and (6).

(5) $$\underset{\omega ,b,i}{min}\frac{1}{2}{\omega }^2+c{\sum }_{i=1}^n{\xi }_i$$

(6) $$\begin{array}{cc}& \\ & \mathrm{s}.\mathrm{t}.\{\begin{array}{c}{y}_i\left({\omega }^T{x}_i+b\right)\ge 1-{\xi }_i,i=1,2,\dots ,n\\ {\xi }_i\ge 0,i=1,2,\dots ,n\end{array}\end{array}$$

When using the SVM method to complete enterprise demand forecasting based on historical data, we discovered that the SVM method is easily prone to fall into the local optimum, making it challenging to achieve model optimization. For this reason, this article chose the particle swarm method to optimize the model. PSO is a search algorithm with few parameters and is easy to implement, which searches for the optimal solution by iteration. The main steps of its algorithm are as follows (Zhao & Zhao, 2021):

To represent the process of PSO optimization SVM more visually, the algorithm process update is shown in Fig. 3.

By optimizing the parameters of SVM, we achieve efficient fitting based on historical inventory data from warehouses in the supply chain to complete inventory forecasting.

The recognition of the goods in the warehouse

The current application of IoT in warehouse management of the supply chain is mainly video monitoring technology, so this article selects the highest-selling product in this warehouse in the last month for identification. Because of the enormous sales volume of this product, its mobility is fast, which increases the difficulty of model training. In the model construction, the period of the item appearing under the fixed camera during the working period of a week is selected for data annotation to complete the model training. Meanwhile, to accurately illustrate the model’s performance, three indicators, Precision, Recall and F1-score are chosen to evaluate the model.

This study uses 30% of the data as the test, and the standard data division approach is chosen for model validation. The experimental results obtained are shown in Fig. 4.

YOLO, the most widely used method in target detection research, has a significantly higher recognition rate than its predecessor methods. As can be seen in Fig. 4, the precision is as high as 98.3% when YOLOv4 performs target recognition. It is significantly higher than conventional machine learning methods and CNN methods.

The prediction for the demand the goods

To complete the analysis of the forecasted commodity demand for this year, this article combines historical data from the company’s warehouse management process over the previous 5 years. To achieve intelligent warehouse management, the outcomes of commodity demand forecasting under various methodologies are shown in Fig. 5.

According to the results shown in Fig. 5, it can be found that due to the sufficient amount of data and the considerable random seasonal variation of commodities, all types of methods can forecast the commodity demand trend with a slight difference in the main trends. However, a more detailed comparison shows that the PSO-SVM method has better forecasting accuracy. The absolute forecast errors of different techniques in different months are shown in Fig. 6, while the sum of forecast errors for each month is shown in Fig. 7.

According to Figs. 6 and 7, we can see that the PSO-SVM method used in this article performs well in the prediction results. The SVM method in this article will have certain singularities in the prediction process, such as the April and June data in Fig. 6, both of which have large deviations, and after optimization by the PSO method, the performance of the model is greatly improved, and better results are achieved.

The framework application in practice

To test effectiveness of the method proposed in the practical application, practical test experiments were conducted in this article. The test framework is shown in Fig. 8.

In the actual testing process, we tested the model recognition rate of the used YOLOv4. We selected a warehouse working peak day within the video surveillance data for testing, and the comparison process is shown on the left side of Fig. 8. The statistical data of the duty officer was compared with the model recognition data based on the actual records of the warehouse personnel to form the true value. The results of the test are shown in Table 1.

It can be seen that the proposed method recognises higher data than manual counting in this camera view, which is because mental inattention often occurs in the manual counting process, leading to errors in data statistics. Therefore, the YOLO method used in this article is an excellent way to avoid such problems and improve the accuracy of goods recognition.

At the same time, this article forecasted the demand for goods in the latest month and compared it with the value set by the company’s strategy department. In the latest month, the actual need for goods was 47 pieces, while the company’s strategy department was forecast. The forecast of the proposed method was 41.2, rounding the data to 41 pieces, and the difference between the two performances was insignificant. However, during the forecasting process, the strategy department researches various information, such as relevant policies and consumption habits, in detail. This process requires a large amount of human and material resources. The proposed PSO-SVM method completes the demand forecast based on historical data, an essential reference for future policy formulation and strategy adjustment in the strategy department.