Papermaking industry intelligent supply chain management model

Abstract

The invention belongs to the technical field of enterprise intelligent management, and relates to a papermaking industry intelligent supply chain management model, and a supplier model selection model is used for intelligently recommending high-quality suppliers; the bid confusion detection model identifies potential bid confusion risks through enterprise association network analysis and bidding behavior pattern mining based on a graph neural network and an anomaly detection technology; the intelligent bid evaluation model adopts a multi-objective optimization and ensemble learning technology, and semantic understanding and evaluation of bid files are realized through a deep neural network; the supplier evaluation model adopts an LSTM and an attention mechanism, a supplier development trend prediction model is constructed, and historical evaluation and future prediction are realized. Based on a traditional model algorithm, close combination with enterprise management is realized through adjustment and deep learning, panoramic detection and intelligent analysis of supply chain upstream suppliers are realized in combination with knowledge graph and natural language processing, and supply and demand resource cooperation efficiency and resource utilization rate are improved.

Description

Technical Field

[0001] This invention belongs to the field of enterprise intelligent management technology, specifically relating to an intelligent supply chain management model for the paper industry. Background Technology

[0002] In the process of digital transformation of traditional paper manufacturing enterprises, especially newsprint manufacturers, using digital means to achieve supply and demand resource matching and information sharing between upstream enterprises and main enterprises is a technical problem that the paper industry urgently needs to solve.

[0003] Currently, many paper manufacturing companies have added ERP (Enterprise Resource Planning) or SRM (Supplier Relationship Management) systems. Although this has enabled the procurement business to shift from offline to online, there are still technical defects such as scattered data sources, inconsistent data quality, and failure to deeply mine the value of data through models / algorithms. The collaborative capabilities of upstream and downstream enterprises in the paper industry supply chain are still insufficient. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent supply chain management model for paper manufacturing enterprises based on artificial intelligence. The technical solution adopted by this invention is as follows:

[0005] A smart supply chain management model for the paper industry includes a supplier selection model, a bid-rigging detection model, a smart bidding evaluation model, and a supplier evaluation model. The supplier selection model automatically recommends and ranks suppliers that meet the cooperation criteria. The bid-rigging detection model is built based on graph neural networks and anomaly detection technology. Through enterprise relationship network analysis and bidding behavior pattern mining, it identifies potential bid-rigging behaviors. The smart bidding evaluation model adopts multi-objective optimization and ensemble learning techniques. Through deep neural networks, it achieves semantic understanding and evaluation of bid documents. The supplier evaluation model uses LSTM and attention mechanisms to build a future performance prediction LSTM model and a supplier development trend prediction model to achieve historical evaluation and future prediction. The evaluation mechanism design includes a dynamic time window evaluation mechanism to balance historical and recent performance. It also establishes a scenario-differentiated evaluation system for different procurement categories and formulates corresponding evaluation standards.

[0006] Preferably, the supplier selection model includes a data acquisition layer, a feature engineering layer, and a ranking model layer. The data acquisition layer adopts a two-layer selection mechanism. The first layer extracts partner data from the enterprise knowledge base using Kettle, and the second layer uses DeepSeek to access publicly released website data and connects to Tianyancha via API to achieve risk detection for selecting backup suppliers. The feature engineering layer uses the logical sequence of data preprocessing → feature construction → feature dimensionality reduction. Each step matches different technical means according to different data types (numerical, textual, and time-based) to adapt to and meet the required business scenarios. Data preprocessing refers to data cleaning, data transformation, and data anonymization of the accessed data source. Feature weights are obtained through training on historical transaction data, and hyperparameter tuning is performed using Bayesian optimization methods to achieve feature construction. The ranking model layer adopts a fusion of multi-objective ranking models and deep learning models to handle high-dimensional sparse features and capture complex user preferences.

[0007] Preferably, the implementation method of the supplier selection model is to use a multi-dimensional evaluation and ranking algorithm, combined with the BERT pre-trained model and industry knowledge graph, to build a supplier evaluation engine suitable for the characteristics of the paper industry.

[0008] Preferably, the collusion detection model uses big data and artificial intelligence algorithms to automatically alert users to the risk of collusion. The artificial intelligence algorithms include graph convolutional networks and variational autoencoders, which analyze supplier equity relationships and bidding behavior time patterns based on neural networks to identify abnormal bidding alliances.

[0009] Preferably, in the intelligent bidding evaluation model, the comprehensive evaluation score is calculated using a weighted aggregation method, and the technical responsiveness score is scored using a deep semantic matching model.

[0010] Preferably, in the supplier evaluation model, the comprehensive score of each period is calculated according to the weight of quality, delivery, cost and other dimensions. Then, the comprehensive scores of each period are summed according to the time weight to finally obtain the supplier's historical performance total score. Participation in business trade and completion of settlement constitute one period.

[0011] The beneficial effects of this invention are:

[0012] This invention provides a smart supply chain management model for the paper industry. Through deep learning, knowledge graphs, and natural language processing, it achieves 360-degree panoramic monitoring of upstream suppliers, improving the efficiency of supply and demand resource matching. The system architecture adopts a microservice design, where each sub-model can operate independently or collaborate to form a complete supply chain ecosystem for paper enterprises, significantly improving procurement efficiency, accuracy, and compliance. This smart supply chain management model integrates the entire bidding process and seamlessly connects with ERP supply chain modules such as purchase requisitions, purchase orders, contracts, and inventory, realizing the integrated application of artificial intelligence technology across the entire supply chain of paper enterprises. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0014] Figure 1 This is an overall architecture diagram of the intelligent supply chain management model according to an embodiment of the present invention;

[0015] Figure 2 This is a schematic diagram of the evaluation indicators for the supplier evaluation project in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] like Figure 1 As shown, a smart supply chain management model for the paper industry includes four sub-models: supplier selection model, collusion detection model, smart bidding evaluation model, and supplier evaluation model.

[0018] (1) Supplier selection model: Based on bidding and procurement big data, AI intelligent analysis and precise matching of suppliers.

[0019] The supplier selection model adopts a brand-new multi-dimensional evaluation and ranking algorithm, which consists of a three-layer architecture: data acquisition layer, feature engineering layer, and ranking model layer.

[0020] Data Acquisition Layer: A two-tier selection mechanism is employed. The first tier uses KETTLE (also known as Pentaho DataIntegration, an open-source data integration tool for data extraction, transformation, and loading) to extract partner data from the enterprise knowledge base. This primarily includes historical partners, cross-departmental recommendations, and related companies of existing suppliers. The front-end defines the extraction depth as needed. For example, when purchasing a motor of brand A, model B, and specification C, the top three or top ten recommended manufacturers can be extracted. The second tier utilizes DeepSeek to access publicly released website data such as the China Government Procurement Network and the China Bidding and Tendering Public Service Platform. It connects to Tianyancha via API (Application Programming Interface) to perform risk detection for selecting backup suppliers. If the number of backup suppliers selected in the first tier is insufficient, the second tier supplements the number of backup suppliers based on priority indicators such as bidding success rate and reputation within the paper industry.

[0021] Feature Engineering Layer: Utilizing the logical sequence of "data collection preprocessing (data cleaning, data transformation, and data anonymization of the access data source) → feature construction → feature dimensionality reduction", each step matches different technical means according to different data types such as numerical, textual, and temporal types to adapt to and meet the required business scenarios.

[0022] Feature weights are derived through training on historical transaction data, and hyperparameters are tuned using Bayesian optimization methods to construct the features. Model features include: historical performance indicators, customer quality scores, cooperation history scores, target product supply capacity scores, and business geographic distribution scores.

[0023] Principal Component Analysis (PCA) and time-series feature extraction mechanisms are introduced to capture the performance fluctuation trends of backup suppliers. A supplier evaluation standard system specific to the paper industry is designed, and a multi-objective optimization algorithm is adopted to balance multiple objectives such as quality, price, and supply stability.

[0024] The feature engineering layer enables anomaly pattern recognition based on an attention mechanism, allowing for the early detection of potential risks in selected backup suppliers.

[0025] Ranking Model Layer: It adopts a combination of multi-objective ranking model and deep learning model, which can handle high-dimensional sparse features, capture complex user preferences, and optimize multiple business objectives at the same time, avoiding the decline of other indicators caused by single-objective optimization.

[0026] Implementation Process: This supplier selection model changes the traditional internet search + manual sourcing model by employing a multi-dimensional evaluation and ranking algorithm. It combines a BERT (a bidirectional pre-trained language model based on the Transformer architecture) pre-trained model and an industry knowledge graph to construct a supplier evaluation engine suitable for the characteristics of the paper industry. The supplier selection model uses an improved Learning to Rank (LTR) algorithm, combined with Random Forest and Gradient Boosting Decision Tree (GBDT) to achieve multi-dimensional scoring.

[0027]

[0028] The formula means that the supplier's total score is equal to the weighted sum of the evaluation indicators from the 1st to the nth.

[0029] W i This represents the weight of the i-th evaluation indicator, reflecting the importance of that indicator in the overall score;

[0030] F i (x i ) represents the original data x for the i-th indicator. i The raw data is processed using a function (such as standardization or normalization, transforming it into a form usable for scoring). The processed value of each indicator is multiplied by its corresponding weight, and then all results are summed to obtain the supplier's overall score. This weighted summation method integrates multiple factors, providing a more comprehensive and reasonable evaluation of the supplier.

[0031] The supplier training data indicator library has been updated with new information including historical performance indicators, customer quality scores, cooperation history scores, target product supply capacity scores, and business geographic distribution scores. This model also integrates with ERP master data, obtaining supplier product quality and non-conformity rates, and importing basic enterprise information fields from the Tianyancha database. Furthermore, future performance is crucial for accurate enterprise decision-making; therefore, an LSTM model is used for prediction.

[0032]

[0033] h t : Represents the hidden state of the LSTM network at time t, containing information about the network up to the current time;

[0034] σ: is the Sigmoid activation function, which maps the input value to the (0,1) interval and controls the "pass-through ratio" of information. Here, it determines how much of the current input information can be updated in the cell state.

[0035] W xh : This is inputting xt to enter the hidden state ht The weight matrix is ​​used to weight the input x at the current time. t Adjust the weights;

[0036] x t : is the input data of the LSTM at time t;

[0037] W hh : is in hidden state h t-1 to hidden state h t The weight matrix is ​​used to weight the hidden state h from the previous time step. t-1 Adjust the weights;

[0038] h t-1 : This is the hidden state of the LSTM network at time t-1, carrying information from previous times;

[0039] b h : This is a bias term used to adjust the output of the activation function and increase the expressive power of the model;

[0040] W hy It is the weight matrix from the hidden layer to the output layer, used to weight the hidden state h. t Perform a linear transformation;

[0041] b y It is the bias term of the output layer, used to adjust the baseline value of the output.

[0042] The entire equation, through such a linear combination, combined with possible activation functions (although the activation function is not explicitly stated in this equation, possibly in subsequent steps or because the equation itself is a linear output part), yields the final output y. t It is commonly used in sequence models (such as RNN, LSTM, etc.) to generate predictive outputs of sequences.

[0043] When a company publishes its procurement needs before bidding, the system automatically matches the supplier selection model and triggers the corresponding rules, automatically helping managers recommend suppliers that meet the cooperation conditions and achieve automatic ranking.

[0044] This study uses the 2024 procurement network platform system bidding of a paper industry group as a case study. A horizontal comparison was conducted to manually determine and analyze the time spent on various key indicators before supplier selection, including basic information of each supplier, match with current needs, creditworthiness, and implementation team capabilities. The time consumed in the review process was then summarized and analyzed. The manual time spent on each item is shown in Table 1 below.

[0045] Table 1: Statistical Time Consumption for Each Item

[0046]

[0047] The sales representatives used traditional supplier sourcing methods to conduct a horizontal comparison of the basic information of the above six companies and their suitability for the content of this tender. The total time spent was as follows:

[0048] T(total)=sum(T1)+sum(T2)+sum(T3)……+sum(Tn)

[0049] T(total)=(0.5+0.4+0.5+0.4+0.5+0.5)+0.5*6+0.2*6+1.5*6+1.2*6=23.2(h)

[0050] Note: T (total) is the total time spent (h represents hours), T1...Tn is the total time spent on each item.

[0051] By using the supplier selection model to conduct a selection survey of the above six companies, the time required for just two indicators—company qualification review and supply-demand matching—was only 10 minutes, and the accuracy of the information was greatly improved.

[0052] This supplier selection model integrates various indicators such as supplier comprehensive strength, supply status, team capabilities, supply-demand matching degree, and cooperation evaluation into the data warehouse, further improving the supplier credit scoring model and covering more than 20 dimensions including legal proceedings, operational anomalies, and performance capabilities. It integrates with ERP systems to achieve data indicator consolidation for product supply, quality inspection, and delivery information, creating a 360-degree panoramic profile of suppliers. Taking a company's procurement activities in 2024 as an example, the supplier selection model saved 2,500 effective working days, saved 1.25 million yuan in wages per year, achieved a supplier risk assessment accuracy rate of 92%, shortened risk warning response time to real-time, and intelligent sourcing recommendations covered 80% of frequently procured items, improving supplier matching efficiency by 60%.

[0053] (2) Collusion detection model: intelligently identifies the relationship between bidding suppliers to prevent the risk of collusion.

[0054] In the bidding process, collusion and bid-rigging are frequent occurrences, causing significant problems for procuring entities and seriously affecting market order. This paper addresses this issue by employing big data and artificial intelligence algorithms to automatically alert procuring and bid-rigging risks, thereby preventing unfair bidding practices. The collusion and bid-rigging detection model is built upon Graph Neural Networks (GNNs) and anomaly detection technology. It identifies potential collusion and bid-rigging behaviors through enterprise relationship network analysis and bidding behavior pattern mining. Core technologies include Graph Convolutional Networks (GCNs) and Variational Autoencoders (VAEs). Enterprise relationship calculation utilizes an improved PageRank algorithm, while collusion and bid-rigging risk scoring is based on anomaly detection algorithms using Isolation Forest technology. By developing a multi-level enterprise relationship network construction algorithm that integrates multi-dimensional relationships such as business, bidding, and cooperation, a spatiotemporal sequence anomaly pattern extraction technique is designed to capture anomaly patterns in historical bidding behavior, realizing a meta-path-based implicit relationship discovery mechanism to reveal indirect relationships.

[0055] By analyzing historical bidding behavior of bidding companies using big data, companies exhibiting abnormal behavior can be identified. For example, by analyzing dimensions such as the total number of bids, the probability of winning bids, the scope of bidding, and competitors, a statistical model of bidding behavior can be constructed to score the bidding behavior of companies and classify them into different levels, thereby quickly and accurately identifying professional bid-rigging companies.

[0056] The enterprise correlation calculation uses an improved PageRank algorithm:

[0057]

[0058] The above formula is the formula for calculating the PageRank value of webpage u in the PageRank algorithm. The meanings of each symbol are as follows: R(U) is the PageRank value of webpage u, d is the damping coefficient, B(u) is the set of webpages that link to webpage u, R(v) is the PageRank value of webpage v, W(v,u) is the link weight from webpage v to webpage u, and L(v) is the number of outgoing links from webpage v.

[0059] The risk scoring for bid-rigging is based on anomaly detection algorithms and uses a multi-level optimization technique called Isolation Forest.

[0060]

[0061] This describes a decaying or probabilistic functional relationship that varies with variables x and n, where E(h(x)) represents the expected value or other statistical measure of h(x), c(n) is the coefficient related to n, and the overall value of s(x,n) is defined by exponential operations. Through granular model algorithms, it automatically identifies and analyzes the relationships between bidding companies, such as equity and management structures, providing visualized relationship diagrams and risk warnings. This helps tenderers and supervisors identify potential conflicts of interest and compliance risks, ensuring the fairness and transparency of bidding activities.

[0062] Development steps: Complete the construction of a database of bid-rigging behavior patterns, and improve detection accuracy through continuous learning. By connecting the bid-rigging detection model to an external database and comparing and analyzing the historical bidding records of bidding units, it can effectively provide a basis for monitoring bid-rigging risks. Based on the analysis of supplier equity relationships and bidding behavior time sequence patterns, it can identify abnormal bidding alliances.

[0063] (3) Intelligent bidding model: Multi-dimensional analysis and evaluation of bidding proposals, and scientific and objective recommendation of the best winning bidder.

[0064] Traditional bidding methods in many paper manufacturing companies largely rely on the lowest price as the winning bidder. However, this single-dimensional approach is somewhat one-sided, failing to provide a comprehensive 360-degree assessment of participating companies' actual situations and potentially leading to the misconception of prioritizing price over quality. This paper addresses this issue by employing a weighted aggregation method for calculating the overall bid score and a deep semantic matching model for technical responsiveness scoring. A technology review intelligent auxiliary system based on industry knowledge is developed to improve review accuracy. Automatic calibration of multi-dimensional scoring standards is implemented to reduce the influence of subjective human factors. A differentiated scoring mechanism is designed to accurately identify subtle differences in strengths among suppliers. A decision interpretation system is constructed to ensure the interpretability and traceability of the bid evaluation results. AI is leveraged to embed and train multi-faceted rules, effectively addressing the unfairness caused by subjective human scoring. Through a series of algorithms integrating supplier evaluation, bid comparison, and supplier risk control models, a direct and intuitive horizontal comparison of bidding companies is conducted during the bid opening period.

[0065] The intelligent bid evaluation model employs multi-objective optimization and ensemble learning techniques, using deep neural networks to achieve semantic understanding and evaluation of bid documents. Core technologies include natural language processing, multi-criteria decision analysis (MCDA), and Bayesian optimization. The overall bid evaluation score is calculated using a weighted aggregation method.

[0066]

[0067] Formula explanation:

[0068] Score total : Indicates the final total score;

[0069] Σ m i＝1 : Summation symbol, sums the terms from i=1 to i=m (m represents the total number of terms in the calculation);

[0070] α i : Represents the weight coefficient of the i-th item, used to reflect the importance of different items in the total score;

[0071] Score i : represents the score of the i-th item.

[0072] The technical responsiveness score uses a deep semantic matching model:

[0073]

[0074] This formula quantifies the similarity between two documents at the semantic level by converting them into vector representations using the Transformer model and then calculating their similarity using the ratio of the vector inner product to the norm.

[0075] Sim(Dr,Ds): Represents the similarity between Dr and Ds;

[0076] Transformer (Dr): Apply the Transformer model (a deep learning model widely used in fields such as natural language processing, which is good at processing sequential data and capturing long-distance dependencies) to the document Dr to obtain the vector representation of Dr after Transformer encoding;

[0077] Transformer(Ds): Similarly, it is the vector representation obtained by applying the Transformer model to the document Ds.

[0078] The product values ​​mentioned above represent the inner product operation of vectors. The inner product result can reflect the similarity between two vectors. The larger the inner product value, the more similar the vectors are.

[0079] ||Vector Value||: Represents the norm of a vector (usually the Euclidean norm, i.e. the magnitude of the vector). Dividing the inner product of two vectors by the product of their norms is to normalize the similarity so that the result falls between [-1,1] (if the vectors are unit vectors, the result falls between [0,1]), making it easier to measure similarity.

[0080] Step 1: Design a dynamic time window evaluation mechanism to reasonably balance historical performance and recent performance;

[0081] Step 2: Develop a bidder profiling system to characterize bidders from multiple dimensions, with 32 pre-defined indicators in the model;

[0082] Step 3: Implement a scenario-based differentiated evaluation system, and adopt different evaluation standards for different procurement categories;

[0083] Step 4: Build a supplier development trend prediction model and integrate it into the bid opening list template page.

[0084] (4) Supplier evaluation model: comprehensively evaluate the supplier's performance capabilities and build a high-quality supplier resource pool.

[0085] Intelligent evaluation of suppliers, from registration to termination, helps improve supply quality and service levels. By leveraging data algorithms to monitor suppliers throughout their entire lifecycle and create a comprehensive profile, enterprise managers can identify potential risks in real time and curb non-compliant supplier behavior. The specific implementation steps are as follows:

[0086] Step 1: Technical Model Construction: LSTM and attention mechanisms are used to optimize the calculation logic of supplier historical performance scores, build an LSTM model for future performance prediction, and simultaneously build a supplier development trend prediction model to achieve historical assessment and future prediction.

[0087] Step 2: Evaluation Mechanism Design: Design a dynamic time window evaluation mechanism to balance historical and recent performance; establish scenario-differentiated evaluation systems for different procurement categories, formulate corresponding evaluation standards, and refer to evaluation indicators. Figure 2 The label.

[0088] Step 3: Algorithm Optimization Design

[0089] Supplier historical performance score calculation steps:

[0090]

[0091] Implementation process:

[0092] The "dimensional comprehensive score" for this period is calculated based on the dimensional weights w of "quality, delivery, cost, and others"; then, the comprehensive score for each period (considering participation in business and trade and completion of settlement as one period) is weighted by time α. t The scores are summed to obtain the supplier's total historical performance score. This method takes into account both "multi-dimensional performance" and "differences in importance across different periods," providing a more comprehensive assessment of the supplier's historical performance. Specifically:

[0093] Summing is performed over a time period t from period 1 to period T (integrating performance across multiple periods, such as historical data from multiple months or quarters);

[0094] α t The time weight of period t (reflecting logic such as "recent performance is more important", e.g., αt is larger in recent periods);

[0095] Qt Quality performance in period t (such as product qualification rate, quality complaint rate, etc.);

[0096] T t Timeliness performance in period t (e.g., on-time delivery rate, order response speed, etc.);

[0097] C t Cost performance in period t (e.g., price competitiveness, cost control level);

[0098] E t Performance in other dimensions (such as environment / service, Environment / Else) in period t (the specific meaning can be defined according to the business scenario);

[0099] w1, w2, w3, w4: Weights for quality, delivery, cost, and other dimensions (reflecting the importance of each dimension in the scoring; the weights are usually 1 or set as needed).

[0100] Step 4: System Development and Implementation: Develop a supplier profiling system to characterize supplier features from multiple dimensions, integrate the above models and mechanisms to form a system with an overall architecture, and identify potential problems and opportunities in advance.

[0101] The intelligent supply chain model is not a single generative model, but rather consists of multiple sub-models covering upstream aspects of the industry chain, including supplier selection, supplier evaluation, bid evaluation, verification of technical and commercial bids, and end-to-end risk control. Its data warehouse extends across the entire industry chain and integrates with DeepSeek. Besides allowing users to actively import data for verification, it achieves close integration and implementation triggering in specific business scenarios within the paper industry, particularly in newsprint. For abnormal states, in addition to providing early warnings, it can also intervene by linking corrective action suggestions, enabling real-time algorithmic feedback.

[0102] This management model uses over 17,000 suppliers of a certain enterprise as training data, integrates resources from ERP, BPM, SRM, and online bidding platforms, and externally accesses related data from DeepSeek, Tianyancha, and China Bidding and Procurement Network. It aggregates supply and demand information onto a single platform, leveraging the model to achieve AI-assisted management and control of multiple key business nodes, including AI-driven procurement sourcing, intelligent supplier selection, risk control throughout the bidding process, and intelligent supplier evaluation. This demonstrates the optimization and deep application of the intelligent supply chain model algorithm, achieving the goal of AI-assisted management decision-making. The intelligent supply chain model indicator decomposition and evaluation table is shown in Table 2.

[0103] Table 2: Evaluation Table of Smart Supply Chain Model Indicators

[0104]

[0105] In the embodiments of the present invention, all technical features not described in detail are existing technologies or conventional technical means, and will not be repeated here.

[0106] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A smart supply chain management model for the paper industry, characterized in that, It includes a supplier selection model, a bid-rigging detection model, a smart bidding model, and a supplier evaluation model. The supplier selection model automatically recommends suppliers that meet the cooperation conditions and ranks them. The collusion detection model is built on graph neural networks and anomaly detection technology. It identifies potential collusion behaviors through enterprise relationship network analysis and bidding behavior pattern mining. The intelligent bidding evaluation model employs multi-objective optimization and ensemble learning techniques, using deep neural networks to achieve semantic understanding and evaluation of bid documents. The supplier evaluation model uses LSTM and attention mechanisms to build a future performance prediction LSTM model, while simultaneously constructing a supplier development trend prediction model to achieve historical evaluation and future prediction. The evaluation mechanism design includes designing a dynamic time window evaluation mechanism to balance historical and recent performance, establishing scenario-differentiated evaluation systems for different procurement categories, and formulating corresponding evaluation standards.

2. The intelligent supply chain management model for the paper industry according to claim 1, characterized in that, The supplier selection model comprises a data acquisition layer, a feature engineering layer, and a ranking model layer. The data acquisition layer employs a two-layer selection mechanism: the first layer extracts partner data from the enterprise knowledge base using Kettle, and the second layer utilizes DeepSeek to access publicly released website data and connects to Tianyancha via API to perform risk detection for selecting backup suppliers. The feature engineering layer follows a logical sequence of data preprocessing → feature construction → feature dimensionality reduction. Each step matches different technical methods according to different data types (numerical, textual, and temporal) to adapt to and meet the required business scenarios. Data preprocessing refers to data cleaning, data transformation, and data anonymization of the accessed data sources. Feature weights are derived through training on historical transaction data, and hyperparameter tuning is performed using Bayesian optimization methods to achieve feature construction. The ranking model layer adopts a fusion of multi-objective ranking models and deep learning models to handle high-dimensional sparse features and capture complex user preferences.

3. The intelligent supply chain management model for the paper industry according to claim 2, characterized in that, The supplier selection model is implemented by using a multi-dimensional evaluation and ranking algorithm, combined with a BERT pre-trained model and an industry knowledge graph, to build a supplier evaluation engine suitable for the characteristics of the paper industry.

4. The intelligent supply chain management model for the paper industry according to claim 1, characterized in that, The collusion detection model uses big data and artificial intelligence algorithms to automatically alert users to the risk of collusion. The artificial intelligence algorithms include graph convolutional networks and variational autoencoders. Based on neural network analysis, it identifies abnormal bidding alliances by analyzing the equity relationships of suppliers and the temporal patterns of bidding behavior.

5. The intelligent supply chain management model for the paper industry according to claim 1, characterized in that, In the intelligent bidding evaluation model, the comprehensive evaluation score is calculated using a weighted aggregation method, while the technical responsiveness score is scored using a deep semantic matching model.

6. The intelligent supply chain management model for the paper industry according to claim 1, characterized in that, In the supplier evaluation model, the comprehensive score of each period is calculated according to the weight of quality, delivery, cost and other dimensions. Then, the comprehensive scores of each period are summed according to the time weight to obtain the supplier's total historical performance score. Participation in business trade and completion of settlement constitute one period.

Images