Powder metallurgy production quality safety monitoring management system

Abstract

The quality safety monitoring management system for powder metallurgy production comprises a production parameter intelligent sensing module, a quality safety intelligent analysis module, a quality traceability management module and a man-machine interaction module; the production parameter intelligent sensing module collects production parameters and completes data preprocessing, and sends the results to the quality safety intelligent analysis module; the quality safety intelligent analysis module receives production parameters, obtains key quality characteristics through semantic ontology driven adaptive sampling, realizes fault diagnosis by applying a mixed enhanced learning method of tensor decomposition, dynamically adjusts quality management strategies through fractional order chaos optimization, and generates quality early warning information, diagnosis decisions and management plans; the quality traceability management module establishes a quality traceability chain covering the whole life cycle of raw materials, processes and products by using blockchain technology, receives quality data and completes safety evidence; the man-machine interaction module realizes visual display of system functions and data, accepts user instructions and coordinates other modules to complete quality monitoring management tasks.

Description

Technical Field

[0001] This invention relates to the field of quality and safety monitoring and management systems, and more specifically, to a quality and safety monitoring and management system for powder metallurgy production. Background Technology

[0002] Powder metallurgy is an advanced material forming and processing technology that uses powdered raw materials such as metals and ceramics, which are mixed, pressed, and sintered, to manufacture high-precision components with complex shapes and excellent performance. It is widely used in the automotive, aerospace, and machinery manufacturing industries. However, powder metallurgy production processes are complex, involving multiple stages such as raw material proportioning, pressing, and sintering. Quality is prone to fluctuations during production, resulting in a high product defect rate. Traditional quality management relies mainly on manual sampling and statistical analysis, which often only provides reactive solutions and is insufficient for precise early warning and proactive control.

[0003] In recent years, with the development of new-generation information technologies such as sensing technology, big data analysis, and artificial intelligence, intelligent manufacturing has become a new way to improve the quality control level of powder metallurgy enterprises. By deploying intelligent sensors on production equipment, real-time online monitoring of the production process can be achieved. Domestic and international scholars have conducted quality-related modeling and analysis research on powder metallurgy production data. Some scholars have applied wavelet packet energy spectrum and BP neural networks to establish a soft measurement model for billet quality, realizing quality prediction during the pressing process. Other scholars have constructed a diagnostic knowledge base for powder metallurgy parts defects based on Bayesian theory and expert experience, enabling root cause analysis and statistical classification of common defects.

[0004] However, existing quality management systems and methods still have shortcomings: First, production parameter collection is incomplete, preprocessing analysis is not intelligent, and the characterization of production status is not precise enough; second, quality early warning models have poor generalization ability, lack prior knowledge guidance, and the interpretability of early warning results is weak; third, defect diagnosis rules are rigid and lack continuous optimization and learning capabilities; fourth, decision optimization methods are not suitable for the dynamic and complex control of production processes. Furthermore, most existing systems adopt a centralized management model, which is insufficient in supporting supply chain traceability and cross-enterprise collaboration. There is an urgent need to explore a completely new quality management system architecture that integrates data collection, modeling and analysis, control and optimization, and traceability. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a quality and safety monitoring and management system for powder metallurgy production. The system utilizes innovative methods such as semantic ontology-driven adaptive sampling, tensor decomposition-based hybrid reinforcement learning, fractional-order chaotic optimization for quality management, and blockchain-driven quality traceability to achieve intelligent control over the entire production quality process.

[0006] This invention provides a quality and safety monitoring and management system for powder metallurgy production, including a production parameter intelligent sensing module, a quality and safety intelligent analysis module, a quality traceability management module, and a human-computer interaction module. The production parameter intelligent sensing module collects production parameters and performs data preprocessing, sending the results to the quality and safety intelligent analysis module. The quality and safety intelligent analysis module receives the production parameters, obtains key quality features based on semantic ontology-driven adaptive sampling, implements fault diagnosis using a hybrid reinforcement learning method with tensor decomposition, and dynamically adjusts quality management strategies through fractional-order chaotic optimization to generate quality early warning information, diagnostic decisions, and management plans. The quality traceability management module utilizes blockchain technology to establish a quality traceability chain covering the entire lifecycle of raw materials, processes, and products, receives quality data, and completes secure data storage. The human-computer interaction module visualizes system functions and data, accepts user commands, and coordinates other modules to complete quality monitoring and management tasks.

[0007] Specifically, the intelligent sensing module for production parameters includes multi-source heterogeneous sensors, an edge gateway, and a time-sensitive network. The multi-source heterogeneous sensors collect raw material properties, process parameters, and equipment status signals, and transmit the data to the edge gateway in real time. The edge gateway performs preprocessing such as cleaning and normalization on the sensor data, and temporarily stores it after classifying it by time and semantics. The time-sensitive network transmits the sensing data from the edge gateway to the quality and safety intelligent analysis module with a delay of no more than 10ms.

[0008] Specifically, the semantic ontology-driven adaptive sampling unit in the quality and safety intelligent analysis module obtains key quality features in the following way:

[0009] A production process semantic ontology O = (C, R, A) is formally defined, consisting of semantic categories and relationships related to raw materials, equipment, processes, and quality. Here, C is the concept set, R is the relation set, and A is the axiom set. O is described using the ontology language OWL.

[0010] Based on the Tableau algorithm, reasoning is performed on the semantic ontology O to obtain the semantic association matrix W∈RN×N, where the matrix element wij∈[0,1] represents the semantic association strength between concept i and concept j, and N is the number of semantic concepts in the ontology;

[0011] The semantic importance weights of the sensor-acquired parameters xi are determined based on the correlation matrix W. The calculation formula is Design a round-rotation matrix compression sampling operator Φ∈RM×M, where the matrix elements Φij follow a Bernoulli distribution. sigmoid(x)=(1+e -x ) -1 Here, λ is the Sigmoid function, and λ is the scaling factor.

[0012] Adaptive sampling of the sensor parameter vector x yields the compressed sensing vector y = Φ·x. i y∈RM is a low-dimensional semantic feature vector.

[0013] Specifically, the fault diagnosis unit based on tensor decomposition and hybrid reinforcement learning in the quality and safety intelligent analysis module achieves fault diagnosis through the following steps:

[0014] A three-dimensional semantic cubic tensor T∈R of fault-symptom-component is constructed based on equipment operation logs, alarm information, and quality defect data. I×J×K Where I is the number of failure modes, J is the number of symptom modes, and K is the number of associated components;

[0015] The semantic tensor T is decomposed using a multimodal tensor decomposition algorithm to obtain the fault factor matrix. Symptom factor matrix S∈R J ×R², component factor matrix V∈R K ×R3, and the core tensor The expression T≈G×1U×2S×3V is satisfied, where × n Let R1, R2, and R3 represent the n-mode tensor product, where R1, R2, and R3 are the ranks of the three dimensions, respectively.

[0016] Constructing integrated prior knowledge Θ K and data-driven knowledge Θ D A hybrid fault diagnosis strategy network π(a|s;Θ), where state s∈R J Let J be the observation vector of the symptom of the J-dimensional symptom, and let a ∈ R. I For the fault mode diagnosis probability vector, the network parameter Θ = (Θ K ,Θ D );

[0017] Prior knowledge Θ K A graph convolutional network with tensor kernels is used for modeling, and state feature propagation employs a dynamic bidirectional attention propagation mechanism. Where H (k) ∈R N×F Let F be the node feature matrix of the k-th layer network, F be the node feature dimension, σ(·) be the activation function, and Ai and This is a dynamic attention matrix;

[0018] Data-driven knowledge Θ D Long Short-Term Memory (LSTM) network modeling is used to model the fault symptom observation sequence (s1, s2, ..., s...). T Encode as a hidden state vector h T and memory cell state vector c T ;

[0019] The loss function of the hybrid diagnostic strategy network is:

[0020] Where γ∈(0,1) is the discount factor, r t As a reward for diagnostic accuracy, y t Q represents the real state minus the action value. π (s t ,a t ) represents the estimated state-action value.

[0021] Specifically, the quality management strategy dynamic adjustment unit based on fractional-order chaotic optimization in the quality and safety intelligent analysis module achieves optimal management strategy search through the following method:

[0022] Construct an N-dimensional mass state vector Q = [q1; q2; ...; q N ], where q i ∈R represents the i-th quality feature value; construct an M-dimensional management decision vector D = [d1; d2; ...; d...]. M ], where d j ∈Ω represents the j-th adjustable management parameter, and the parameter space Ω represents the range of parameter values.

[0023] The fractional-order dynamic characteristics of the mass state are described using Caputo fractional differential equations:

[0024]

[0025] in dτ is the Caputo fractional differential operator, and Γ(-) is the Gamma function;

[0026] Discretizing the fractional differential equation yields the quality state prediction model:

[0027] Q(t+1)=Q(t)+hα·∑(k=0:t)ω k ,t+1·f(Q(tk),D(tk));·

[0028] The cumulative quality loss J in future T steps T =∑(t=1:T)L(Q(t),Q * To optimize the quality management decision-making model, a quality management optimization model is constructed, where L(:) is the measure of the deviation of the quality state Q from the target Q. * The loss function;

[0029] A chaotic optimization algorithm based on the Logistic mapping z(k+1)=μ·z(k)·(1-z(k)) is introduced to encode and map the management decision D through the chaotic variable z: d j =d minj +z·(dmaxj -d minj And search for the optimal decision vector D in the parameter space Ω. * ; by the optimal decision vector D * Decode and generate a quality management plan for the next T steps to guide quality improvement in the production process.

[0030] Specifically, the quality traceability management module constructs a three-layer blockchain quality traceability network in the following ways: it builds raw material blockchain, process blockchain, and product blockchain, covering the entire production life cycle; it uses hash pointer technology to link each layer of blockchain to achieve end-to-end quality traceability; it uses elliptic curve cryptography algorithm to perform asymmetric encryption and digital signature on quality data to ensure data security; and it supports dynamic expansion of no less than 100 nodes, with consensus latency controlled within 1 second.

[0031] Specifically, the human-computer interaction module is designed based on a B / S architecture and supports cross-platform Web access; a single service node supports no less than 1,000 concurrent HTTP requests;

[0032] It adopts the WebGL graphics rendering engine to realize the 3D visualization of data such as production parameters, quality characteristics, diagnostic results, and traceability information, with a rendering frame rate of no less than 60FPS; it has a built-in data mining algorithm library to support data analysis functions such as correlation analysis, comparative analysis, and trend prediction to assist in quality management decisions.

[0033] Specifically, the multi-source heterogeneous sensor has the following parameters: the sensor sampling frequency is not less than 1kHz to meet the requirements of online quality monitoring; the sensor range covers the key parameter range of the production process, and the indication error is not greater than 1%; the edge gateway adopts an industrial-grade embedded computing platform, configured with no less than 4-core CPU, 8GB memory and 128GB solid-state drive, and supports multi-protocol access and edge computing.

[0034] Specifically, the fault diagnosis unit has an active learning capability and achieves online updates of the fault knowledge base in the following ways: extracting new fault modes and symptom information based on expert feedback on the diagnosis results; dynamically expanding the dimension of the semantic tensor T and adding new tensor elements; periodically recalculating tensor decomposition and updating prior diagnostic knowledge; and fine-tuning the network parameters of the hybrid diagnostic strategy to continuously improve the accuracy and generalization of fault diagnosis.

[0035] Specifically, the quality traceability module adopts a consortium blockchain model to organize the blockchain network, supporting collaborative quality management through the following mechanisms: production enterprises, suppliers, distributors, regulatory authorities, etc., jointly act as blockchain nodes for identity authentication and authorization; smart contract technology is used to regulate production and transaction behaviors, enabling automatic on-chaining and sharing of quality data; identity-based privacy protection and fine-grained access control balance data sharing and security; and a permission-controlled quality traceability interface is provided for third-party audits, promoting social collaboration in quality management.

[0036] The present invention has the following beneficial effects:

[0037] (1) A semantic ontology-driven adaptive sampling method for production parameters is proposed. By constructing a semantic ontology of the production process, the inherent semantic relationship between parameters and quality is mined, and the sensor sampling rate is adaptively adjusted to improve the acquisition accuracy of quality-related parameters while reducing data redundancy.

[0038] (2) A hybrid reinforcement learning quality diagnosis method based on tensor decomposition is proposed. The semantic representation of fault symptoms is obtained by tensor decomposition, a hybrid neural network integrating prior knowledge and data-driven approaches is constructed, and reinforcement learning is used to optimize the diagnosis strategy, which significantly improves the accuracy of quality warning and defect diagnosis.

[0039] (3) A fractional-order chaotic optimization method for quality management strategy optimization is proposed. By dynamically modeling the quality state through fractional-order differential equations and introducing a chaotic optimization algorithm to search for the optimal control strategy, adaptive dynamic adjustment of management decisions can be achieved.

[0040] (4) A quality traceability system integrating blockchain technology is proposed. Through the consortium blockchain architecture, business collaboration in production, supply, and sales can be achieved, which can support quality management and data sharing throughout the entire life cycle of powder metallurgy products.

[0041] (5) Improved the targeting and efficiency of production quality data collection, significantly enhanced edge computing capabilities, and more accurate semantic feature extraction; greatly improved the accuracy of quality early warning and defect diagnosis, and made it more adaptable to complex production conditions; enhanced the initiative and foresight of quality management, and made decision optimization more in line with the dynamic characteristics of powder metallurgy processes; and built a trustworthy and traceable quality management alliance chain to provide data support for the rapid location of quality problems and accountability.

[0042] In summary, the quality and safety monitoring and management system of this invention provides an innovative theoretical method and information solution for intelligent quality control in the powder metallurgy industry, which is of great significance for improving the quality level and market competitiveness of powder metallurgy products. Attached Figure Description

[0043] Figure 1This is a schematic diagram of the framework of the quality and safety monitoring and management system for powder metallurgy production according to the present invention. Detailed Implementation

[0044] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the quality and safety monitoring and management system for powder metallurgy production proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The specific solution of the quality and safety monitoring and management system for powder metallurgy production provided by this invention is described below in conjunction with the accompanying drawings.

[0046] Please see Figure 1 This invention provides a quality and safety monitoring and management system for powder metallurgy production. The system consists of a production parameter intelligent sensing module 1, a quality and safety intelligent analysis module 2, a quality traceability management module 3, and a human-computer interaction module 4. The modules are closely coupled and collaboratively optimized through data flow and control flow to jointly achieve closed-loop control of quality and safety in the production process.

[0047] The intelligent sensing module 2 for production parameters is responsible for collecting and preprocessing various parameters from the powder metallurgy production process. This module primarily consists of multi-source heterogeneous sensors, an edge gateway, and a time-sensitive network (TSN). In actual deployment, sensors for pressure, temperature, and atmosphere can be installed on equipment such as raw material silos, mixers, presses, and sintering furnaces, according to production process requirements, to collect production parameters at each workstation in real time. The sensors typically employ industrial-grade intelligent sensors with sampling frequencies exceeding 1kHz, meeting the timeliness requirements of online quality inspection. The sensors connect to the nearest edge gateway via industrial Ethernet or bus. The gateway uses a high-performance industrial computer to preprocess the sensor data, including cleaning, normalization, and compression, and categorizes and stores it according to semantic types, preparing for subsequent semantic extraction. Simultaneously, the gateway uploads the preprocessed sensing data to the quality and safety intelligent analysis module via a time-sensitive network (such as TSN). The time-sensitive network provides deterministic data transmission latency, typically controlled within 10ms.

[0048] The Quality and Safety Intelligent Analysis Module 2 is the core of this system. This module comprehensively utilizes artificial intelligence technologies such as semantic extraction, reinforcement learning, and chaotic optimization to uncover hidden quality defect patterns from production parameters, enabling intelligent early warning, diagnosis, and optimization of the production process. This module first employs an adaptive sampling method based on semantic ontology to extract key quality features. The specific steps are as follows:

[0049] Step 1: Construct a semantic ontology for the powder metallurgy production process. Process experts and knowledge engineers jointly define the main semantic concepts, including raw materials, equipment, processes, and quality. The relationships between these concepts are formally described using an ontology language (such as OWL), e.g., how pressing pressure affects billet density and sintering temperature determines grain size. This forms the knowledge graph of the semantic ontology. The number of concept nodes in the ontology is typically in the range of 100 to 1000.

[0050] Step 2: Use the Tableau algorithm to infer the semantic association strength. The Tableau algorithm is a knowledge reasoning method based on deductive logic. Through a resolution refutation mechanism, it can deduce implicit conceptual relationships from an ontology. Assuming the ontology contains N concept nodes, Tableau reasoning will generate an N×N semantic association matrix W, where the matrix elements w... ij ∈[0,1] represents the quantified value of the semantic association strength between concept and j. ij The larger the value of w, the closer the semantic relationship between i and j. Generally, w... ij A value greater than 0.6 indicates that the relationship between the two concepts is significant.

[0051] Step 3: Calculate the semantic importance weights of the production parameters. Let the production parameters collected by the sensors be x1, x2, ..., x... M Then parameter x i semantic importance weight Through formula The calculation yields w. i j is x i The strength of the association between the semantic concept and the quality concept j. The parameter x was quantized. i Significance in terms of quality semantics. The range of its value is [0,1], and The larger x is i The higher the quality relevance, the better. In practice, Parameters greater than 0.5 can be considered key quality influencing factors.

[0052] Step 4: Construct a round-robin sampling operator based on parametric semantic weights. To efficiently extract key parameters related to quality semantics during online data acquisition, the system designs a round-robin matrix compression sampling operator Φ. Let Φ be an M×M matrix (M is the total number of parameters), and matrix elements Φ... ij Follows Bernoulli distribution The scaling coefficient λ controls the semantic weights. The steepness of the mapping to the sampling probability can be simply taken as 5 to 10. The Sigmoid function will... Mapping from the (0,1) interval to the (0.5,1) interval facilitates probabilistic mapping. Step 5: Map the production parameter vector x = [x1; x2; ...]. M Adaptive semantic sampling is performed. y = Φ·x, where y is the compressed low-dimensional semantic feature vector. Because Φ ij by Using a probability distribution, the sampling process retains key parameters with high semantic weights with a higher probability, while sampling relatively less important parameters with a lower probability, thus achieving semantic adaptation in parameter acquisition. Simulation tests show that when the total number of parameters M = 10000, the sampling compression ratio can reach 10-100 times, meaning that the final semantic feature dimension is only 1 / 10 to 1 / 100 of the original parameters, greatly reducing data transmission and storage overhead.

[0053] Step 5: Generate the production parameter vector x = [x1; x2; ...; x...]. M Adaptive semantic sampling is performed. y = Φ·x, where y is the compressed low-dimensional semantic feature vector. Because Φ ij by Using a probability distribution, the sampling process retains key parameters with high semantic weights with a higher probability, while sampling relatively less important parameters with a lower probability, thus achieving semantic adaptation in parameter acquisition. Simulation tests show that when the total number of parameters M = 1000, the sampling compression ratio can reach 10 to 100 times, meaning that the final semantic feature dimension is only 1 / 10 to 1 / 100 of the original parameters, greatly reducing data transmission and storage overhead.

[0054] The tensor decomposition-based hybrid reinforcement learning fault diagnosis unit in the Quality and Safety Intelligent Analysis Module 2 employs an innovative method to achieve high-precision fault diagnosis. The specific steps are as follows:

[0055] First, various production logs and quality reports are collected from information sources such as equipment management systems and MES. Then, structured information such as fault patterns, fault symptoms, and affected components are extracted using text mining methods. A three-dimensional semantic tensor T∈R of fault-symptom-component is constructed. I×J×K Where I, J, and K represent the number of fault modes, the number of symptom modes, and the number of associated components, respectively.

[0056] Next, the system employs an innovative Multimodal Tensor Decomposition (MTD) algorithm to decompose the semantic tensor T. Unlike traditional Tucker decomposition, the MTD algorithm can simultaneously consider multiple modal characteristics of tensor data, better capturing the complex correlations between faults, symptoms, and components. The goal of the MTD algorithm is to find the fault factor matrix. Symptom Factor Matrix Component factor matrix and core tensor This makes T≈G×1U×2S×3V. Here, × n Let R1, R2, and R3 represent the ranks of the three dimensions, which are typically much smaller than the original dimensions I, J, and K, thus achieving dimensionality reduction and feature extraction. The core idea of ​​the MTD algorithm is to introduce a multimodal regularization term based on the traditional Tucker decomposition to fully utilize the complementary information between different modes. The optimization objective function of the algorithm can be expressed as:

[0057]

[0058] Among them ||·|| F Let Ω1(U,S,V) represent the Frobenius norm, Ω2(G) be the multimodal regularization term of the factor matrix, Ω2(G) be the sparse regularization term of the core tensor, and λ1 and λ2 be trade-off parameters. Solving this optimization problem using the Alternating Least Squares (ALS) method yields the required factor matrix and core tensor. The U, S, and V matrices obtained by the MTD algorithm characterize the latent semantic themes of fault modes, symptoms, and components, respectively, while the core tensor G reflects the high-order interactions between these themes. These decomposition results contain important prior diagnostic knowledge that will guide the subsequent training of the neural network.

[0059] To fully utilize the prior knowledge obtained from tensor decomposition, an innovative Tensor Factor Graph Convolutional Network (TF-GCN) is designed. Unlike traditional GCNs, TF-GCN introduces a Dynamic Bidirectional Attention Propagation (DBAP) mechanism, which can more effectively capture the dynamic correlations between different features. The state feature propagation rule of the DBAP mechanism is as follows:

[0060]

[0061] Where H( k )∈R N × F Let F be the node feature matrix of the k-th layer network, F be the node feature dimension, and σ(·) be the non-linear activation function (such as ReLU). i and These are dynamically generated attention matrices that control the flow of information along the row and column directions of the features, respectively. Specifically:

[0062] A i =softmax(W1·[U;V])

[0063]

[0064] Here, W1 and W2 are learnable parameter matrices, [·; ·] denote matrix concatenation operations, and G (i) This is the i-th slice of the core tensor G. The softmax function is used to normalize the attention score. In this way, the network can dynamically adjust the importance of feature propagation based on the current state and prior knowledge, improving the model's expressiveness and flexibility.

[0065] The output of the TF-GCN network is concatenated with the hidden state vector of the LSTM network and then input into the fully connected layer to obtain the predicted probability vector of the fault mode. The entire hybrid neural network is trained end-to-end, and the parameters are optimized using the cross-entropy loss function and the Adam optimizer. To further improve diagnostic performance, the system uses reinforcement learning to optimize the network's online diagnostic strategy. Specifically, the fault diagnosis task is modeled as a Markov decision process (MDP), where the state space S is the historical fault symptom observation sequence, the action space A is the set of candidate fault causes, and the reward function R is the accuracy of predicting the fault cause. Under this MDP framework, the network's inference process can be formalized as finding an optimal diagnostic strategy π: S→A, such that taking diagnostic action a in the current fault symptom state s maximizes the expected future cumulative reward.

[0066] The system employs a deep Q-learning (DQN) approach to optimize the diagnostic strategy. Hybrid neural networks serve as an approximation of the Q-value function, i.e.

[0067] Q(sa;θ)=H T (a)

[0068] Where H T (a) represents the predicted probability of fault cause a by the TF-GCN network in state s. During online diagnosis, the system integrates the prediction results of both the TF-GCN and LSTM modules, selecting the fault cause with the largest Q value as the diagnosis result. Based on the feedback (accuracy or error) of the diagnosis result, the network parameters θ are updated, and the policy iteration formula is:

[0069]

[0070] Where y = r + γ·max a′ Q(s′,a′;θ) represents the TD objective, r represents the immediate reward, γ represents the discount factor, and α represents the learning rate. Through multiple training epochs of iteration, the network will gradually strengthen its learning of successful diagnostic strategies.

[0071] Experimental results demonstrate that the hybrid reinforcement learning diagnostic algorithm based on multimodal tensor decomposition and dynamic bidirectional attention propagation achieves an average diagnostic accuracy of 97% when handling typical faults in powder metallurgy production, a 2 percentage point improvement over traditional methods. Particularly when dealing with complex, multi-source faults, the new algorithm exhibits stronger robustness and generalization ability. This is primarily due to the more comprehensive extraction of multidimensional data features by multimodal tensor decomposition, while the dynamic bidirectional attention propagation mechanism enhances the model's ability to capture dynamic correlations between different features. The system incorporates a diagnostic knowledge base covering common fault modes in powder metallurgy production, such as high porosity, dimensional deviations, uneven hardness, crack initiation, and surface roughness. In practical applications, the system can flexibly expand fault modes according to production needs and continuously optimize the diagnostic model through online learning, ensuring that diagnostic performance remains up-to-date.

[0072] Another key innovation of the Quality and Safety Intelligent Analysis Module 2 is its adaptive quality management decision optimization algorithm. Traditional quality management largely relies on qualitative, experience-based decision-making, which struggles to cope with complex and ever-changing production situations. To address this pain point, this system proposes using fractional-order chaos theory and intelligent optimization algorithms to dynamically optimize quality management decisions in the production process in real time.

[0073] First, the system extracts a series of state parameters reflecting product quality attributes (such as appearance, dimensions, hardness, strength, etc.) from the production execution MES system, forming an N-dimensional quality state vector Q = [q1; q2; ...; q N Simultaneously, an adjustable M-dimensional process parameter vector D = [d1; d2; ...; d...] is formed. M [This is used as a management decision variable.] Then, a fractional differential equation is used to describe the evolution of the quality state:

[0074]

[0075] Where α is the fractional order. The fractional differential operator is of the Caputo type, reflecting the nonlocal memory effect and fractional dynamic characteristics of the quality state. f(·) is the nonlinear coupling function between the quality state vector Q and the decision vector D. Using the generalized derivative defined by the fractional differential, this model can characterize the long-range correlation and historical dependence of the quality state, better reflecting the dynamic characteristics of the production process. Considering the difficulty in obtaining analytical solutions to fractional differential equations, the system employs a numerical approximation method for discretization, yielding a time-domain prediction model of the quality state.

[0076]

[0077] Where h is the time step, ω k,t+1The weighting coefficients are defined by Grunwald-Letnikov (GL) for fractional derivatives. Based on this, the quality management optimization objective function is constructed:

[0078]

[0079] Where T represents the time domain for optimizing management decisions, and L(·) represents the deviation of the quality state Q from the target Q. * The loss function can be flexibly defined using forms such as the sum of squared errors (SSE). Meanwhile, considering the physical constraints of process parameters, boundary conditions are added to the decision variables, transforming the quality management decision into a constrained nonlinear programming problem.

[0080] To efficiently solve the aforementioned quality management optimization model and overcome the tendency of traditional numerical optimization methods to get trapped in local optima, this system innovatively introduces a chaotic optimization algorithm based on Logistic mapping. Utilizing the pseudo-randomness and ergodicity of chaotic motion, management decision variables are encoded and mapped through chaotic variables:

[0081] d i =a i +z i ·(b i -a i ), i = 1, 2, ..., M

[0082] Among them, a i and b i Let z be the lower and upper bounds of the i-th decision variable, respectively. i For chaotic variables, satisfying the Logistic mapping:

[0083]

[0084] μ is a chaos parameter, with a value between 3.56 and 4. During the iterative optimization process, the chaotic variable z dynamically traverses the search space in a pseudo-random manner and adaptively adjusts the search step size, so that the particles continuously jump out of local optima in the solution space and eventually converge to the global optimum with a relatively high probability.

[0085] The algorithm flow is as follows:

[0086] Step 1: Randomly initialize the chaotic variable z(0)∈(0,1), set the maximum number of iterations K, and the current number of iterations k=0;

[0087] Step 2: Generate an initial management decision variable D(k) based on the chaotic variable z(k);

[0088] Step 3: Substitute D(k) into the quality state prediction model to calculate the cumulative quality loss J(D(k)) in the next T steps;

[0089] Step 4: Update the current optimal solution D * =argmin{J(D(k)),J(D * )};

[0090] Step 5: Determine if the convergence condition is met. If it is, output D. * If the decision is optimal, terminate the iteration; otherwise, let k = k + 1 and go to Step 6.

[0091] Step 6: Generate new chaotic variables based on the Logistic mapping z(k+1)=4·z(k)·(1-z(k)), and go to Step 2.

[0092] Simulation tests show that when the maximum number of iterations K = 100, the chaotic optimization algorithm can quickly converge to the optimal management decision, and the average prediction error of the optimized quality state can be reduced by more than 50%. The optimal decision variable D... * By decoding and combining the current production status, a quality control plan for a future period (such as the next day) can be generated, including control values ​​for key process parameters, personnel and equipment scheduling suggestions, etc., forming an operable and closed-loop quality management decision optimization scheme.

[0093] The Quality Traceability Management Module 3 constructs a quality traceability chain covering the entire lifecycle of powder metallurgy products. The system adopts a consortium blockchain architecture, with manufacturers, suppliers, distributors, and regulatory agencies acting as blockchain nodes to record and store quality data at each stage of production, distribution, and sales. The system features a three-layer blockchain structure: raw materials, process, and product. The raw materials blockchain records procurement and inspection information for raw materials and auxiliary materials; the process blockchain records production and processing data; and the product blockchain records finished product testing and sales information. These three sub-chains are linked in an ordered manner via hash pointers, allowing for rapid identification of problematic batches and their distribution routes in the event of a quality issue. The blockchain network uses the PBFT consensus algorithm, supporting large-scale dynamic node expansion with consensus latency controllable to the second level. Simultaneously, the system utilizes the Elliptic Curve Cryptography (ECC) algorithm to encrypt and sign block data, eliminating the risk of tampering. This consortium blockchain model balances commercial privacy protection with the orderly sharing of quality information, providing a trusted collaborative platform for all parties involved in production management.

[0094] This invention further optimizes the system's performance in terms of human-computer interaction and intelligent sensing. The system adopts a B / S architecture for its human-computer interaction module 4, supporting cross-platform and mobile access. The introduction of WebGL and data mining algorithm libraries enables the system to have smooth 3D visualization interaction and one-click statistical analysis capabilities. At the intelligent sensing level, sensor selection fully considers indicators such as accuracy, sampling rate, and stability. Meanwhile, the system's collaborative optimization framework is open and scalable, allowing for flexible addition or removal of functional modules according to application scenarios.

[0095] The following section provides a specific application example to further illustrate the system's workflow. In a powder metallurgy gear production workshop, the system deploys 100 intelligent sensors on key equipment in processes such as pressing, sintering, post-processing, and finished product inspection, with a data acquisition frequency of 1kHz. These sensors are connected to five multi-protocol edge gateways via the Industrial Ethernet protocol. The gateways perform semantic classification of the data and upload it to the MES server via the TSN private network. The quality and safety intelligent analysis module is deployed on the MES server.

[0096] (1) Semantic ontology-driven sampling: The system constructs a semantic ontology of the production process, which includes 500 concepts such as equipment, process, material, quality, and defects. Through ontology reasoning, the top 10 production parameters with the highest semantic importance weight are extracted, mainly including pressing pressure, sintering temperature, holding time, quenching temperature, etc., accounting for 10% of the total parameters. The sampling frequency of each parameter is adaptively adjusted using weights, and the average compression efficiency reaches 10:1.

[0097] (2) Hybrid Reinforcement Learning Diagnosis: The system collected 1,000 fault records from the enterprise over two years, finding that the faults mainly focused on three modes: dimensional deviation, low hardness, and cracking. Diagnostic knowledge tensors covering these three fault modes were obtained through tensor decomposition, and the TF-GCN fault diagnosis model was trained based on this. In the online application phase, when a batch of gears showed a low hardness defect, the hybrid diagnosis model issued a timely warning, tracing back to find that a sintering temperature 20℃ lower than normal was the root cause, achieving a diagnostic accuracy of 95%.

[0098] (3) Fractional-order chaotic optimization: The system constructs a fractional-order prediction model for three key quality characteristics of gears: strength, hardness, and size. The order α is set to 0.6, and the average prediction error is less than 5%. When the product sampling inspection finds that the hardness is generally low, the system quickly starts the optimization of quality management parameters. Through 100 iterations, the optimal quenching temperature curve and time are locked, and it is predicted that the hardness value will be improved by more than 10% after the process is implemented.

[0099] (4) Blockchain quality traceability: The system deploys blockchain nodes in the stages of raw material procurement, production and processing, warehousing and logistics, and sales. When a customer complains that individual gears are cracked, the production batch can be quickly located through the product QR code. Then, the blockchain is used to trace back to find that the cause is that the carbon content of a certain batch of steel powder is not up to standard, which provides data support for the company's claims and quality improvement.

[0100] In summary, the quality and safety monitoring and management system for powder metallurgy production proposed in this invention integrates cutting-edge artificial intelligence technologies such as semantic sampling, reinforcement learning, chaotic optimization, and blockchain. It is specifically designed and engineered based on the production characteristics of the powder metallurgy industry, forming a complete intelligent sensing, analysis, traceability, and interaction technology solution. Practical application of the system shows that adaptive sampling improves the efficiency of production parameter collection by 50%; hybrid diagnostics achieves a 95% accuracy rate for quality early warning; and fractional-order optimization increases the first-pass yield by 5 percentage points. The successful development of this system provides a new approach to solving the quality control challenges in the powder metallurgy industry and is of great significance for realizing intelligent and lean powder metallurgy production.

[0101] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A quality and safety monitoring and management system for powder metallurgy production, characterized in that, It includes a production parameter intelligent sensing module, a quality and safety intelligent analysis module, a quality traceability management module, and a human-computer interaction module; the production parameter intelligent sensing module collects production parameters and completes data preprocessing, and sends the results to the quality and safety intelligent analysis module; The intelligent quality and safety analysis module receives production parameters, obtains key quality features based on semantic ontology-driven adaptive sampling, applies a hybrid reinforcement learning method using tensor decomposition to achieve fault diagnosis, and dynamically adjusts quality management strategies through fractional-order chaotic optimization to generate quality early warning information, diagnostic decisions, and management plans. The quality traceability management module uses blockchain technology to establish a quality traceability chain covering the entire lifecycle of raw materials, processes, and products, receives quality data, and completes secure data storage. The human-computer interaction module realizes the visualization of system functions and data, accepts user instructions, and coordinates other modules to complete quality monitoring and management tasks. The semantic ontology-driven adaptive sampling unit in the quality and safety intelligent analysis module obtains key quality features in the following ways: Constructing a semantic ontology for the powder metallurgy production process ,in For a set of concepts, For a set of relations, It is an axiom set; it defines the semantic concepts of raw materials, equipment, processes, and quality, and uses the ontology language OWL to formally describe the relationships between concepts, thus forming a knowledge graph of the semantic ontology; The semantic ontology O is inferred based on the Tableau algorithm to obtain the semantic association matrix. , where matrix elements Representing concepts With concept The strength of semantic association between them The number of semantic concepts in the ontology; Let the production parameters collected by the sensor be... According to the association matrix Determine the parameters collected by each sensor semantic importance weight The calculation formula is: Design a round-rotation matrix compression sampling operator Φ∈ , where matrix elements Follows Bernoulli distribution , For the Sigmoid function, Scale factor; For sensor parameter vectors Adaptive sampling is performed to obtain compressed sensing vectors. It is a low-dimensional semantic feature vector; because by As a probability distribution, the sampling process will retain key parameters with high semantic weight with a higher probability, while sampling relatively minor parameters with a lower probability. The fault diagnosis unit based on tensor decomposition and hybrid reinforcement learning in the intelligent analysis module for quality and safety achieves fault diagnosis through the following steps: Based on equipment operation logs, alarm information, and quality defect data, fault modes, fault symptoms, and affected components are extracted, and a three-dimensional semantic cubic tensor of fault-symptom-component is constructed. ,in Number of failure modes For symptom pattern number, This represents the number of associated components; Using a multimodal tensor decomposition algorithm to analyze semantic tensors Decomposition yields the fault factor matrix. Symptom Factor Matrix Component factor matrix and core tensor ,satisfy ,in express - Modal tensor product, , , The ranks are respectively the three dimensions; Constructing integrated prior knowledge and data-driven knowledge Hybrid Fault Diagnosis Strategy Network , where the state Let J be the J-dimensional fault symptom observation vector, and the action... For fault mode diagnosis probability vectors, network parameters ; In order to fully utilize the prior knowledge obtained from tensor decomposition, prior knowledge The Tensor Factor Graph Convolutional Network (TF-GCN) is used for modeling. Unlike traditional GCNs, TF-GCN employs a dynamic bidirectional attention propagation mechanism for state feature propagation. ,in For the first The node feature matrix of the layer network, For node feature dimensions, For activation function, and This is a dynamic attention matrix; Data-driven knowledge Using long short-term memory network modeling, the fault symptom observation sequence is... Encoded as a hidden state vector and memory cell state vector ; Tensor factor graph convolutional networks employ a dynamic bidirectional attention propagation mechanism. The state feature propagation rule of the dynamic bidirectional attention propagation mechanism is as follows: ， in For the first The node feature matrix of the layer network, For node feature dimensions, ReLU is a non-linear activation function. and These are dynamically generated attention matrices that control the flow of information along the row and column directions of the features, specifically: ， ， here and It is a learnable parameter matrix. This represents a matrix concatenation operation. It is the core tensor The Each slice is used, and the softmax function is used to normalize the attention scores. The output of the TF-GCN network is concatenated with the hidden state vector of the LSTM network and then input into the fully connected layer to obtain the predicted probability vector of the fault mode. The system uses reinforcement learning to optimize the network's online diagnosis strategy, modeling the fault diagnosis task as a Markov decision process (MDP). The state space S is the historical fault symptom observation sequence, the action space A is the set of candidate fault causes, and the reward function R is the accuracy of predicting the fault cause. The network's inference process can be formalized as finding an optimal diagnosis strategy π: S→A, such that taking the diagnosis action a under the current fault symptom state s maximizes the expected future cumulative reward. The system uses a deep Q-learning (DQN) approach to optimize the diagnostic strategy, with a hybrid neural network serving as an approximation of the Q-value function. The loss function of the hybrid diagnostic strategy network is: ,in As a discount factor, Rewards for accurate diagnosis. The actual state minus the action value. For estimated state-action values; The dynamic adjustment unit for quality management strategies based on fractional-order chaotic optimization in the intelligent quality and safety analysis module achieves optimal management strategy search through the following method: Construct an N-dimensional mass state vector ,in This is the i-th quality characteristic value; Constructing an M-dimensional management decision vector ,in For the first One adjustable management parameter, parameter space The range of parameter values; The fractional-order dynamic characteristics of the mass state are described using Caputo fractional differential equations: ， in, For fractional order, It is a Caputo-type fractional differential operator, reflecting the nonlocal memory effect of the mass state and the fractional dynamic characteristics. Mass state vector and decision vector The nonlinear coupling function; Discretizing the fractional differential equation yields the quality state prediction model: , in, For time step, The weighting coefficients defined by Grunwald-Letnikov (GL) for fractional differentials; Construct the quality management optimization objective function: ， in, For optimizing the time domain of management decisions, To measure quality status Deviation from target The loss function; A chaotic optimization algorithm based on Logistic mapping is introduced, which utilizes the pseudo-randomness and ergodicity of chaotic motion to encode and map management decision variables through chaotic variables: , in, and The first The lower and upper bounds of the values ​​of each decision variable. For chaotic variables, satisfying the Logistic mapping: ， For chaotic parameters, in parameter space Search for the optimal decision vector From the optimal decision vector Decode and generate a quality management plan for the next T steps to guide quality improvement in the production process.

2. The system as described in claim 1, characterized in that, The intelligent sensing module for production parameters includes multi-source heterogeneous sensors, an edge gateway, and a time-sensitive network. The multi-source heterogeneous sensors collect raw material properties, process parameters, and equipment status signals, and transmit the data to the edge gateway in real time. The edge gateway cleans and normalizes the sensor data, classifies it by time and semantics, and then temporarily stores it. The time-sensitive network transmits the sensing data from the edge gateway to the intelligent analysis module for quality and safety with a delay of no more than 10ms.

3. The system as described in claim 1, characterized in that, The quality traceability management module constructs a three-layer blockchain quality traceability network in the following ways: it builds raw material blockchain, process blockchain, and product blockchain, covering the entire production life cycle; it uses hash pointer technology to link each layer of blockchain to achieve end-to-end quality traceability; it uses elliptic curve cryptography algorithm to perform asymmetric encryption and digital signature on quality data to ensure data security; and it supports dynamic expansion of no less than 100 nodes with consensus latency controlled within 1 second.

4. The system as described in claim 1, characterized in that, The human-computer interaction module is designed based on a B / S architecture and supports cross-platform web access; a single service node supports no less than 1,000 concurrent HTTP requests; It adopts the WebGL graphics rendering engine to realize the 3D visualization of production parameters, quality characteristics, diagnostic results, and traceability information data, with a rendering frame rate of no less than 60FPS; it has a built-in data mining algorithm library to support correlation analysis, comparative analysis, and trend prediction data analysis functions to assist in quality management decision-making.

5. The system as described in claim 2, characterized in that, The multi-source heterogeneous sensor has the following parameters: the sensor sampling frequency is not less than 1kHz to meet the requirements of online quality monitoring; the sensor range covers the key parameter range of the production process, and the indication error is not greater than 1%; The edge gateway adopts an industrial-grade embedded computing platform, configured with no less than 4-core CPU, 8GB memory and 128GB solid-state drive, and supports multi-protocol access and edge computing.

6. The system as described in claim 1, characterized in that, The fault diagnosis unit has active learning capabilities and achieves online updates to the fault knowledge base in the following ways: extracting new fault modes and symptom information based on expert feedback on the diagnosis results; dynamically expanding the dimensions of the semantic tensor T and adding new tensor elements; periodically recalculating tensor decomposition and updating prior diagnostic knowledge; and fine-tuning the network parameters of the hybrid diagnostic strategy to continuously improve the accuracy and generalization of fault diagnosis.

7. The system as described in claim 3, characterized in that, The quality traceability module adopts a consortium blockchain model to organize the blockchain network, supporting collaborative quality management through the following mechanisms: production enterprises, suppliers, distributors, and regulatory authorities jointly act as blockchain nodes for identity authentication and authorization; smart contract technology is used to regulate production and transaction behaviors, enabling automatic on-chaining and sharing of quality data; identity-based privacy protection and fine-grained access control balance data sharing and security; and a permission-controlled quality traceability interface is provided for third-party audits, promoting social collaboration in quality management.

Images