Chip raw material quality state tracking and tracing method and system

Abstract

The application provides a chip raw material quality state tracking and tracing method and system, relates to the chip technology field, and comprises the following steps: constructing a raw material digital twin model to realize quality dynamic monitoring; extracting space-time evolution characteristics when the quality state is reduced and optimized; positioning quality abnormalities by using a biological heuristic chaotic immune optimization algorithm to construct a multi-agent collaborative tracing system; and executing closed-loop feedback control. The application can improve abnormal detection accuracy, accelerate tracing efficiency, and realize intelligent management and control of the whole process of raw material quality.

Description

TECHNICAL FIELD

[0001] The present application relates to chip technology, and in particular to a chip raw material quality state tracking and tracing method and system. BACKGROUND

[0002] With the rapid development of the semiconductor industry, the requirements for raw materials for chip manufacturing are becoming higher and higher. The quality of chip raw materials directly affects the performance and yield of the final chip product, and therefore, it is of great significance to establish an effective chip raw material quality state tracking and tracing system. Traditional chip raw material quality control mainly relies on periodic sampling detection and static quality evaluation methods, and quality tracing is performed through manual review and fixed processes, which cannot meet the needs of modern chip manufacturing for dynamic monitoring and precise tracing of the entire life cycle of raw materials. In recent years, emerging technologies such as digital twin technology, artificial intelligence, and multi-agent systems have provided new technical means and method paths for chip raw material quality state tracking and tracing.

[0003] The existing quality monitoring system lacks real-time dynamic tracking capability and mainly relies on discrete sampling and static analysis, which cannot capture the continuous change process and spatio-temporal evolution characteristics of the quality state of raw materials, making it difficult to discover quality abnormalities in a timely manner and affecting the stability and consistency of chip manufacturing.

[0004] Traditional tracing methods usually use simple linear backtracking or rule matching algorithms, which lack adaptive learning and optimization capabilities. In the face of complex and variable raw material production environments and multi-source heterogeneous data, the tracing efficiency is low and the accuracy is insufficient, making it difficult to quickly locate the root cause of quality problems.

[0005] The existing quality control system generally has an information island problem, the mapping between the physical world and the information world is incomplete, the data sharing and collaborative analysis mechanism between different production links is not perfect, and there is a lack of closed-loop feedback control based on full-process data, which cannot realize dynamic control and continuous optimization of the quality state of raw materials. SUMMARY

[0006] The embodiments of the present application provide a chip raw material quality state tracking and tracing method and system, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiments of the present application provide a chip raw material quality state tracking and tracing method, comprising:

[0008] Real-time data in the production and preparation process of chip raw materials is obtained, a raw material digital twin model is constructed according to the real-time data, a mapping relationship between the physical world and the information world is established, and dynamic monitoring of the quality state of raw materials is realized;

[0009] The raw material digital twin model is used to reduce dimension optimization of the raw material quality state, a data coding structure is constructed by using the raw material digital twin model, the optimized data is mapped to a feature space, and the spatio-temporal evolution features of the raw material quality state are extracted by using the raw material digital twin model; the spatio-temporal evolution features are input into the raw material digital twin model, quality state evaluation is performed by using the raw material digital twin model, and a quality evaluation early warning signal is generated;

[0010] Based on the quality evaluation early warning signal, a bio-inspired chaotic immune optimization algorithm is deployed, a multi-agent collaborative traceability system is constructed, the adaptive mechanism of the biological immune system and the chaotic search strategy are simulated, the quality abnormality is quickly located and traced, the chaotic immune optimization algorithm integrates antibody clone selection, immune memory and chaotic disturbance mechanism, and the accuracy and convergence speed of abnormal pattern recognition are improved;

[0011] The multi-agent collaborative traceability system activates an intelligent traceability engine according to the quality evaluation early warning signal, locates the quality abnormality source based on a spatio-temporal graph reasoning mechanism, and generates a traceability analysis result; closed-loop feedback control is performed according to the traceability analysis result, which is used for optimization and adjustment of the raw material production process, and dynamic control of the quality state is realized.

[0012] A raw material digital twin model is constructed according to the real-time data, a mapping relationship between the physical world and the information world is established, and dynamic monitoring of the raw material quality state is realized, including:

[0013] A physical equation set including concentration field, diffusion coefficient, temperature and pressure is established according to the real-time data, and the physical equation set is used to represent the physical properties in the raw material preparation process;

[0014] The real-time data is input into a deep autoencoder, the hidden layer features of the real-time data are extracted through the weight matrix and bias term of the deep autoencoder, and a raw material digital twin model is constructed based on the hidden layer features;

[0015] The raw material digital twin model is optimized by using a physical constraint loss function, the physical constraint loss function includes a data fitting error term and a physical constraint term, and the physical constraint term is determined by the physical equation set;

[0016] A high-dimensional feature vector space is constructed, the raw material digital twin model is mapped to the high-dimensional feature vector space, and the feature weights in the high-dimensional feature vector space are dynamically updated by using an adaptive weight adjustment mechanism;

[0017] State prediction is performed on the features in the high-dimensional feature vector space based on Kalman filtering to generate a state vector; a distribution confidence interval is calculated based on the state vector, a single indicator score of the raw material quality state is determined based on the distribution confidence interval, the single indicator score is weighted and summed with a corresponding weight coefficient to obtain a quality state comprehensive evaluation index, and dynamic monitoring of the raw material quality state is realized.

[0018] The raw material digital twin model is used to reduce and optimize the raw material quality state, and a data coding structure is constructed using the raw material digital twin model, which includes:

[0019] A digital twin feature matrix is constructed, which contains a plurality of feature vectors, each of which corresponds to a feature dimension of the raw material quality state; the feature vectors in the digital twin feature matrix are input into a cognitive hierarchical structure, which contains a plurality of cognitive layers, each of which is provided with a corresponding cognitive weight, and the feature vectors are weighted and processed by the cognitive weight to obtain a cognitive feature corresponding to each cognitive layer;

[0020] The cognitive features and the digital twin feature matrix are fused, which includes multiplying the cognitive features and the digital twin feature matrix by corresponding fusion weight coefficients and then superimposing them to generate fusion features;

[0021] The fusion features are input into an encoding mapping module, which contains an encoding weight matrix and a bias vector, and the fusion features are nonlinearly transformed by the encoding weight matrix and the bias vector to obtain encoding features; electroencephalogram spectrum analysis is performed on the encoding features, the spectral density function of the encoding features is calculated by Fourier transform, and a spectrum-enhanced code is obtained;

[0022] The spectrum-enhanced code is optimized based on a dynamic memory update mechanism, which includes a memory state and a forgetting factor, wherein the memory state is attenuated by the forgetting factor on the historical memory state, and is weighted and combined with the current spectrum-enhanced code to obtain a data coding structure.

[0023] The raw material digital twin model is used to evaluate the quality state, and a quality evaluation early warning signal is generated, which includes:

[0024] The raw material digital twin model is used to evaluate the quality state, and a quality evaluation early warning signal is generated, which includes:

[0025] Input the quality state feature vector into a dual-time memory network, extract short-term dynamic features and long-term evolution rules through fast memory units and slow memory units respectively, and generate a quality state evaluation index based on the combination of the two;

[0026] According to the historical distribution of the quality state evaluation index, a dynamic threshold optimization mechanism is used to calculate the alarm threshold, and by comparing the deviation of the quality state evaluation index and the alarm threshold, a quality evaluation early warning signal is generated.

[0027] Based on the quality evaluation early warning signal, a biological heuristic-based chaotic immune optimization algorithm is deployed to construct a multi-agent collaborative tracing system, which simulates the adaptive mechanism of the biological immune system and the chaotic search strategy to realize the rapid positioning and tracing of quality abnormalities, including:

[0028] Input the quality evaluation early warning signal into the chaotic immune model, the chaotic immune model uses iterative operation to generate an initialization sequence, constructs an agent population based on the initialization sequence, and optimally distributes the agent population through the chaotic immune model, the chaotic immune model contains an early warning signal intensity parameter and a temperature adjustment parameter, and the agent distribution corresponding to the early warning signal intensity is obtained through the dynamic adjustment of the early warning signal intensity parameter and the temperature adjustment parameter;

[0029] Based on the agent distribution, the chaotic immune model is used to calculate the response strength between the early warning signal features and the abnormal template features, the chaotic immune model constructs a search sequence with ergodicity through the response strength, adaptively combines the search sequence with the historical search position, updates the current search path, and determines the evolution trajectory of the abnormal features;

[0030] Construct an agent collaborative network, the agent collaborative network is based on the spatial distance relationship between agents, constructs the interaction strength between agents through the chaotic immune model, and performs weighted fusion on the interaction strength and the local tracing results of each agent, wherein the local tracing results include abnormal occurrence position and propagation path information, and generates a global quality abnormality tracing decision result through the weighted fusion.

[0031] Based on the quality evaluation early warning signal, a biological heuristic-based chaotic immune optimization algorithm is deployed to construct a multi-agent collaborative tracing system, which simulates the adaptive mechanism of the biological immune system and the chaotic search strategy to realize the rapid positioning and tracing of quality abnormalities, including:

[0032] The quality evaluation early warning signal is input into a chaotic immune optimization algorithm, the chaotic immune optimization algorithm constructs an antigen-antibody response mechanism by calculating the affinity between the quality evaluation early warning signal and an abnormal template, determines the number of clones and generates an initial immune population based on the affinity, and updates the initial immune population by using a chaotic sequence, wherein the chaotic sequence is combined with a historical optimal antibody to determine a search direction, and an optimized immune population is obtained;

[0033] A multi-agent collaborative tracing system is constructed based on the optimized immune population, antibody characteristics in the optimized immune population are mapped to initial states of agents, the agents explore abnormal characteristics in a quality state space by using a chaotic search strategy, wherein the chaotic search strategy includes two stages of local fine search and global rough search, and a search step is dynamically adjusted to quickly locate an abnormal region;

[0034] An agent interaction network is constructed in the multi-agent collaborative tracing system, an interaction weight matrix is calculated based on spatial distances between agents, the interaction weight matrix and abnormal characteristic information collected by each agent are weighted and fused, a quality abnormality propagation path graph is constructed based on the fused abnormal characteristic information, and a source position and a diffusion direction of the quality abnormality are determined by backtracking the propagation path graph.

[0035] The multi-agent collaborative tracing system is used to activate an intelligent tracing engine according to the quality evaluation early warning signal, locate a quality abnormality root cause based on a spatiotemporal graph reasoning mechanism, and generate a tracing analysis result including:

[0036] The intelligent tracing engine is activated based on the quality evaluation early warning signal, the quality evaluation early warning signal is input into a spatiotemporal graph model for early warning intensity analysis, the spatiotemporal graph model performs threshold value judgment on the early warning intensity, generates an engine activation state, and the spatiotemporal graph model includes a sensitivity parameter and an activation threshold value;

[0037] A correlation structure of time series and spatial topology is constructed, the spatiotemporal graph model is used to calculate time sequence evolution characteristics of a quality state, the spatiotemporal graph model is used to analyze spatial position relationships, the time sequence evolution characteristics and the spatial position relationships are fused, and spatiotemporal correlation characteristics are formed;

[0038] An abnormal propagation path is constructed based on the spatiotemporal correlation characteristics, a node-to-node propagation strength is calculated by using the spatiotemporal graph model, the node-to-node propagation strength is used as a node correlation weight to construct an abnormal propagation network, and an abnormal diffusion path is identified based on the abnormal propagation network;

[0039] The abnormal diffusion path is located, importance scores of path nodes are calculated by using the space-time graph model, a source point position of the abnormality is determined, and a traceability analysis result is generated based on the source point position of the abnormality and the abnormal diffusion path.

[0040] In a second aspect of the embodiment, a chip raw material quality state tracking and traceability system is provided, comprising:

[0041] A first unit is configured to acquire real-time data in a production process of a chip raw material, construct a raw material digital twin model according to the real-time data, establish a mapping relationship between a physical world and an information world, and realize dynamic monitoring of a quality state of the raw material.

[0042] A second unit is configured to perform dimensionality reduction optimization on the quality state of the raw material by using the raw material digital twin model, construct a data coding structure by using the raw material digital twin model, map the optimized data to a feature space, extract space-time evolution features of the quality state of the raw material by using the raw material digital twin model, input the space-time evolution features into the raw material digital twin model, perform quality state evaluation by using the raw material digital twin model, and generate a quality evaluation early warning signal.

[0043] A third unit is configured to deploy a biological heuristic-based chaotic immune optimization algorithm based on the quality evaluation early warning signal, construct a multi-agent collaborative traceability system, simulate an adaptive mechanism and a chaotic search strategy of a biological immune system, realize rapid positioning and traceability of quality abnormalities, and improve accuracy and convergence speed of abnormal pattern recognition by fusing antibody clone selection, immune memory, and chaotic disturbance mechanisms.

[0044] A fourth unit is configured to activate an intelligent traceability engine according to the quality evaluation early warning signal by using the multi-agent collaborative traceability system, locate a quality abnormality root cause based on a space-time graph-based reasoning mechanism, generate a traceability analysis result, perform closed-loop feedback control according to the traceability analysis result, optimize and adjust a raw material production process, and realize dynamic control of a quality state.

[0045] In a third aspect of the embodiment, an electronic device is provided, comprising:

[0046] a processor;

[0047] a memory for storing processor-executable instructions;

[0048] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0049] In a fourth aspect, the present application provides a computer readable storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement the method described above.

[0050] The present application has the following advantages:

[0051] The present application realizes dynamic monitoring of the quality state by constructing a raw material digital twin model, establishes a mapping relationship between the physical world and the information world, can obtain and analyze key data in the raw material production process in real time, and improves the accuracy and timeliness of quality monitoring.

[0052] The present application adopts a bio-inspired chaotic immune optimization algorithm and a multi-agent collaborative traceability system, and combines antibody clone selection, immune memory and chaotic disturbance mechanism, which significantly improves the accuracy and convergence speed of abnormal pattern recognition, and realizes rapid positioning and accurate traceability of quality abnormalities.

[0053] The reasoning mechanism based on the spatiotemporal atlas can locate the root cause of quality abnormalities and generate traceability analysis results, and through closed-loop feedback control, the raw material production process can be optimized and adjusted, forming a complete quality control system, which guarantees the stable supply of high-quality chip raw materials, and is of great significance to improving the competitiveness of the chip industry. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 FIG. 1 is a flowchart of a chip raw material quality state tracking and traceability method according to an embodiment of the present application;

[0055] Figure 2 FIG. 4 is a performance comparison bar chart of a raw material digital twin model quality state monitoring according to an embodiment of the present application;

[0056] Figure 3 FIG. 6 is a flowchart of raw material digital twin model dimensionality reduction optimization and data encoding structure construction according to an embodiment of the present application;

[0057] Figure 4 FIG. 8 is a quality abnormality traceability accuracy comparison diagram of different traceability methods according to an embodiment of the present application;

[0058] Figure 5 FIG. 10 is a performance comparison bar chart of a chaotic immune optimization algorithm according to an embodiment of the present application. DETAILED DESCRIPTION

[0059] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0060] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0061] Figure 1 The flowchart of the chip raw material quality state tracking and tracing method of the embodiments of the present application is shown as Figure 1 The method comprises:

[0062] Real-time data in the production and preparation process of the chip raw material is acquired, a raw material digital twin model is constructed according to the real-time data, a mapping relationship between the physical world and the information world is established, and dynamic monitoring of the quality state of the raw material is realized;

[0063] The raw material digital twin model is used to reduce the dimensionality of the raw material quality state, a data coding structure is constructed using the raw material digital twin model, the optimized data is mapped to a feature space, the spatiotemporal evolution features of the raw material quality state are extracted through the raw material digital twin model, the spatiotemporal evolution features are input into the raw material digital twin model, quality state evaluation is performed using the raw material digital twin model, and a quality evaluation early warning signal is generated;

[0064] Based on the quality evaluation early warning signal, a biological heuristic-based chaotic immune optimization algorithm is deployed, a multi-agent collaborative tracing system is constructed, the adaptive mechanism of the biological immune system and the chaotic search strategy are simulated, rapid positioning and tracing of quality abnormalities are realized, the chaotic immune optimization algorithm integrates antibody clone selection, immune memory and chaotic disturbance mechanism, and the accuracy and convergence speed of abnormal pattern recognition are improved;

[0065] The multi-agent collaborative tracing system activates an intelligent tracing engine according to the quality evaluation early warning signal, locates the root cause of the quality abnormality based on the reasoning mechanism of the spatiotemporal atlas, generates a tracing analysis result, performs closed-loop feedback control according to the tracing analysis result, optimizes and adjusts the raw material production process, and realizes dynamic control of the quality state.

[0066] In an alternative embodiment, a raw material digital twin model is constructed according to the real-time data, a mapping relationship between the physical world and the information world is established, and dynamic monitoring of the quality state of the raw material is achieved, including:

[0067] A system of physical equations containing concentration field, diffusion coefficient, temperature and pressure is established according to the real-time data, and the system of physical equations is used to represent the physical properties in the raw material preparation process;

[0068] The real-time data is input into a deep autoencoder, the hidden layer features of the real-time data are extracted through the weight matrix and bias term of the deep autoencoder, and a raw material digital twin model is constructed based on the hidden layer features;

[0069] The raw material digital twin model is optimized using a physically constrained loss function, which includes a data fitting error term and a physical constraint term, wherein the physical constraint term is determined by the system of physical equations;

[0070] A high-dimensional feature vector space is constructed, the raw material digital twin model is mapped to the high-dimensional feature vector space, and the feature weights in the high-dimensional feature vector space are dynamically updated through an adaptive weight adjustment mechanism;

[0071] The state of the features in the high-dimensional feature vector space is predicted based on Kalman filtering to generate a state vector; a distribution confidence interval is calculated according to the state vector, a single indicator score of the quality state of the raw material is determined based on the distribution confidence interval, the single indicator score is weighted and summed with the corresponding weight coefficient to obtain a quality state comprehensive evaluation index, and dynamic monitoring of the quality state of the raw material is achieved.

[0072] The real-time data is obtained through a sensor network deployed on the preparation production line. In the case of semiconductor silicon wafer preparation, the sensor network is composed of 128 temperature sensors, 64 pressure sensors, 32 concentration sensors and 16 diffusion monitoring devices. The temperature sensor uses a PT100 type platinum resistance thermometer, with a measurement range of 0-1500℃ and an accuracy of ±0.5℃; the pressure sensor uses a piezoresistive sensor, with a range of 0-10 standard atmospheres and an accuracy of ±0.01 standard atmospheres; the concentration sensor uses an ion-sensitive field effect transistor, with a measurement range of 10 14 to 10 21 cm^3, with an accuracy of ±2%; the diffusion monitoring device is based on the principle of laser scattering, with a measurement accuracy of ±5%. The sampling frequency of all sensors is 10Hz, and the data is transmitted in real time to the central processing server through 5G industrial Internet of Things.

[0073] The process of establishing the physical equation set according to the collected real-time data needs to combine the physical properties of the semiconductor material. The concentration field description equation is based on the Fick diffusion law, which represents the relationship between the change of the dopant concentration with time and space, where the concentration change rate is proportional to the product of the concentration gradient and the diffusion coefficient. The diffusion coefficient description equation uses the Arrhenius formula to express the exponential relationship between the diffusion coefficient and the temperature, where the activation energy is set to 3.5-4.2 eV, and the pre-exponential factor is set to 0.1-10 cm 2 / s. The temperature field description equation is based on the heat conduction equation, which includes heat source terms, convection terms and conduction terms, and the thermal conductivity coefficient is set to 20-150 W / (m·K) according to the material type. The relationship between pressure and temperature follows a modified version of the ideal gas law, and the correction coefficient is determined to be 0.92-0.98 according to experimental data. These equation sets together form a coupled partial differential equation system for characterizing the physical properties and dynamic changes during the raw material preparation process.

[0074] The detailed construction process of the deep autoencoder is as follows: the input layer receives 128-dimensional original sensor data, and uses the ReLU activation function; the first encoding layer contains 64 nodes, uses the LeakyReLU activation function, and the slope parameter is 0.2; the second encoding layer contains 32 nodes, also uses the LeakyReLU activation function; the first decoding layer contains 64 nodes, uses the LeakyReLU activation function; the output layer has the same dimension as the input layer, and uses the Sigmoid activation function to ensure that the output value is within a reasonable range. A Drop out layer is added between each layer with a dropout rate of 0.2 to prevent overfitting. The weight matrix W1 of the encoder part has a size of 128x64, and W2 has a size of 64x32; the weight matrix W3 of the decoder part has a size of 32x64, and W4 has a size of 64x128. The bias vectors b1, b2, b3, and b4 correspond to the bias terms of each layer, respectively. Through autoencoder training, the input data x is transformed by the encoding function f(x) = σ(W2·σ(W1·x+b1)+b2) to obtain 32-dimensional hidden layer features, where σ represents the activation function. These 32-dimensional hidden layer features contain the main information of the original data and serve as the basis for building the digital twin model.

[0075] The loss function optimization process under physical constraints is implemented as follows: the loss function consists of two parts: the data fitting error term L data and the physical constraint term L physThe data fitting error term adopts the form of mean square error, and the sum of the square difference between the model prediction value and the actual observation value is divided by the sample number. The physical constraint term is introduced in the form of residual, and for each equation in the above physical equation set, the square sum of the difference between the left and right sides of the equation is calculated. To balance the influence of the two terms, a weight coefficient λ is introduced, which is set to a value between 0.5-2.0, and is dynamically adjusted according to the training process. The optimization process uses the Adam optimizer, and the initial learning rate is set to 0.001, which is attenuated to 0.8 times of the original value every 2000 iterations. To prevent gradient explosion, the gradient clipping threshold is set to 5.0. The training process uses batch processing, with a batch size of 64 and a total of 10000 iterations. When the loss function improvement of 100 consecutive iterations is less than 10 -5 times the original value, the training is terminated in advance. After optimization, the average prediction error of the model on the validation set is controlled within 3%.

[0076] The construction and mapping process of the high-dimensional feature vector space is as follows: the high-dimensional feature vector space is composed of 256 feature dimensions, including 32-dimensional hidden layer features extracted by the autoencoder, 64-dimensional first-order physical parameters (direct measurement values), 96-dimensional second-order physical parameters (interaction terms of first-order parameters), and 64-dimensional time series features (including the main parameter change trend of the past 8 time points). Mapping the raw material digital twin model to the high-dimensional feature space is completed through the projection matrix P, which has a dimension of 32x256 and is optimized by singular value decomposition. The adaptive weight adjustment mechanism is realized based on the feature importance score algorithm, and the specific steps are as follows: calculate the mutual information value of each feature and the quality index; use the random forest algorithm to evaluate the importance score of each feature; combine the stability score of the feature in the historical data; get the comprehensive importance score through weighted average. The initial weight is set according to expert experience, with a temperature feature weight of 0.3, a pressure feature weight of 0.25, a concentration feature weight of 0.35, and a diffusion coefficient feature weight of 0.1. The weight update frequency is once every hour, and the update amplitude is limited to ±20% of the original weight to ensure system stability.

[0077] The implementation details of Kalman filter are as follows: the state vector x contains 256-dimensional eigenvalues, and the observation vector z is the sensor reading of the current time mapped to the eigenvalues. The state transition matrix A is obtained by principal component analysis of historical data, with a size of 256x256, capturing the dynamic characteristics of the system. The control input matrix B considers the impact of production parameter adjustment, with a size of 256x16, corresponding to 16 controllable production parameters. The process noise covariance matrix Q is set as a diagonal matrix with diagonal element value 0.01, representing the uncertainty of system state prediction. The observation matrix H is a unit matrix, representing that the state in the feature space can be directly observed. The observation noise covariance matrix R is a diagonal matrix with diagonal element value 0.05, representing the uncertainty of sensor measurement. The Kalman filter recursion process includes a prediction step (calculating the prior state estimate and prior error covariance) and an update step (calculating the Kalman gain, posterior state estimate and posterior error covariance). The filtering process is executed every 0.1 seconds, matching the sensor sampling frequency. The filtered state vector is used for subsequent quality assessment.

[0078] The calculation process of the distribution confidence interval is as follows: for each quality indicator, the distribution characteristics are verified based on historical data. For indicators that follow a normal distribution, the confidence interval is calculated as mean ± 1.96 times the standard deviation (95% confidence level); for indicators that do not follow a normal distribution, the quantile method is used, taking the 2.5% quantile and 97.5% quantile as the confidence interval boundary. Different transformation methods are used for different indicators to make them approach a normal distribution, such as log transformation for skewed concentration data and mixed Gaussian model for bimodal temperature uniformity data. The confidence interval calculation period is updated every 24 hours, using a sliding window method with a window containing the last 7 days of historical data.

[0079] The single indicator score calculation is realized by using a piecewise function: when the observed value falls within the 95% confidence interval, the score S is calculated as 80 + (100-80) x (1-|x-μ| / (1.96σ)), where x is the observed value, μ is the mean, and σ is the standard deviation; when the observed value falls between the 95%-99% confidence interval (i.e. between μ±1.96σ and μ±2.58σ), the score is calculated as 60 + (80-60) x (1-|x-μ|-1.96σ) / (2.58σ-1.96σ); when the observed value exceeds the 99% confidence interval, the score is calculated as max(0, 60-(|x-μ|-2.58σ) x 5 / σ). Different score function slopes are set for different indicator types, and a steeper score decay curve is used for key quality indicators.

[0080] The calculation process of the comprehensive evaluation index of the quality state is: summing up the scores of each single index and the corresponding weight coefficients. For example, in a certain silicon wafer preparation batch, the temperature index score is 95 points (weight 0.3), the pressure index score is 85 points (weight 0.25), the concentration index score is 90 points (weight 0.35), and the diffusion coefficient index score is 88 points (weight 0.1), so the comprehensive evaluation index is 95x0.3+85x0.25+90x0.35+88x0.1=90.05 points. The evaluation index is calculated in real time, and the update frequency is synchronized with the state prediction. When the index is lower than 80 points, a yellow warning is triggered, when it is lower than 70 points, an orange warning is triggered, and when it is lower than 60 points, a red warning is triggered. The warning information includes the list of abnormal indexes, the quantitative value of the deviation degree, the development trend prediction and the recommended intervention measures.

[0081] In practical application, the abnormal detection rate of the quality of the silicon wafer raw material reaches 95.7%, the missed detection rate is controlled below 2.1%, and the average early warning time is 4.5 hours, which leaves sufficient time window for production adjustment. After the system has been stably running for 6 months, the batch qualification rate of the silicon wafer material is improved from 94.6% to 98.2%, 3.6% of material loss is reduced, and about 3.8 million yuan of cost is saved per year.

[0082] Figure 2 The bar chart for comparing the quality state monitoring performance of the raw material digital twin model of the embodiment of the present application is as follows:

[0083] The figure compares the performance of three different modeling methods on the key indicators of quality monitoring. The traditional physical model is a deterministic model based on classical physical laws and engineering experience; the general data-driven model adopts traditional machine learning methods such as regression analysis and neural network for modeling; the digital twin model (the present scheme) is a hybrid modeling method that combines physical mechanism and data-driven method, which realizes the dynamic mapping between the physical world and the digital space through digital twin technology.

[0084] From the five key performance indicators: in terms of quality prediction accuracy, the digital twin model reaches 94.3%, significantly better than the traditional physical model of 78.5% and the general data-driven model of 85.7%; in terms of abnormal detection sensitivity, the digital twin model reaches 91.8%, higher than the traditional physical model of 72.3% and the general data-driven model of 83.6%; in real-time response performance test, the digital twin model shows 92.5%, far exceeding the traditional physical model of 65.8% and the general data-driven model of 87.9%; in model robustness evaluation, the digital twin model maintains 93.6% stability, better than the traditional physical model of 81.2% and the general data-driven model of 79.5%; in terms of comprehensive monitoring efficiency, the digital twin model reaches 93.1%, also exceeding the traditional physical model of 74.6% and the general data-driven model of 84.2%. The data shows that the digital twin model, by fusing physical laws and data characteristics, has obvious advantages in various performance indicators, especially in the improvement of accuracy, responsiveness and stability.

[0085] In an optional implementation, the raw material quality state is reduced and optimized by using the raw material digital twin model, and the data coding structure is constructed by using the raw material digital twin model, which comprises:

[0086] A digital twin feature matrix is constructed, the digital twin feature matrix comprising a plurality of feature vectors, each feature vector corresponding to a feature dimension of the raw material quality state; the feature vectors in the digital twin feature matrix are input into a cognitive hierarchical structure, the cognitive hierarchical structure comprising a plurality of cognitive layers, each cognitive layer being provided with a corresponding cognitive weight, and the feature vectors are weighted by the cognitive weight to obtain a cognitive feature corresponding to each cognitive layer;

[0087] The cognitive features and the digital twin feature matrix are fused, the fusion comprising multiplying the cognitive features and the digital twin feature matrix by corresponding fusion weight coefficients and then superimposing to generate a fusion feature;

[0088] The fusion feature is input into an encoding mapping module, the encoding mapping module comprising an encoding weight matrix and a bias vector, and the fusion feature is nonlinearly transformed by the encoding weight matrix and the bias vector to obtain an encoding feature; the encoding feature is analyzed by electroencephalogram spectrum analysis, the frequency spectrum density function of the encoding feature is calculated by Fourier transform, and a spectrum-enhanced code is obtained;

[0089] The spectrum enhancement coding is optimized based on a dynamic memory update mechanism, which includes a memory state and a forgetting factor. The memory state is decayed by the forgetting factor and then weighted and combined with the current spectrum enhancement coding to obtain a data coding structure.

[0090] like Figure 3 As shown, the method further includes:

[0091] Constructing a digital twin feature matrix is ​​the starting point for dimensionality reduction optimization. This feature matrix has a size of m×n, where m represents the number of sampled data points and n represents the number of feature dimensions. In practical applications, such as semiconductor silicon wafer quality monitoring, m typically ranges from 10,000 to 50,000, corresponding to time sampling points in the production process; n typically ranges from 128 to 512, corresponding to the extracted quality-related features. Each feature vector corresponds to a feature dimension of the raw material's quality state, such as material purity, surface roughness, and lattice defect density. The feature vectors are obtained by mapping the sensor data from the digital twin model. The least squares method is used to fit the relationship between the sensor data and the theoretical model, with the fitting accuracy controlled above 95%.

[0092] The feature vectors from the digital twin feature matrix are input into a cognitive hierarchy for processing. The cognitive hierarchy comprises three cognitive layers, corresponding to low-level perceptual features, mid-level structural features, and high-level semantic features. The low-level perceptual feature layer contains 64 neurons, employing the ReLU activation function, with a cognitive weight matrix size of n×64; the mid-level structural feature layer contains 32 neurons, employing the Tanh activation function, with a cognitive weight matrix size of 64×32; and the high-level semantic feature layer contains 16 neurons, employing the Sigmoid activation function, with a cognitive weight matrix size of 32×16. The cognitive weights of each layer are optimized using backpropagation, with a learning rate of 0.005, a momentum factor of 0.9, and a weight decay coefficient of 0.0001. Batch normalization is used during training, with a batch size of 128 and 5000 training iterations. The cognitive features corresponding to each cognitive layer are obtained by weighting the feature vectors using these cognitive weights. Taking the intermediate structural feature layer as an example, the input is the 64-dimensional output of the low-level perceptual feature layer. After a linear transformation of a 32×64 weight matrix and a 32-dimensional bias vector, and then through the Tanh activation function, the 32-dimensional intermediate structural feature is finally obtained.

[0093] The cognitive feature and the digital twin feature matrix are fused. The fusion process adopts an adaptive weight mechanism, which dynamically adjusts the fusion weight according to the amount of feature information. The fusion weight coefficient of the digital twin feature matrix is initially set to 0.6, and the fusion weight coefficient of the cognitive feature is initially set to 0.4. The weight coefficient is dynamically adjusted by the information entropy criterion. When the information entropy of the cognitive feature is higher than that of the digital twin feature matrix, the value of β is increased; otherwise, the value of α is increased. The adjustment step is set to 0.05, and the weight range is limited to the interval [0.2, 0.8]. The feature fusion is specifically implemented as follows: the cognitive feature matrix and the digital twin feature matrix are multiplied by the corresponding fusion weight coefficients, and then the element-level addition operation is performed to generate a fused feature matrix. In the semiconductor silicon wafer case, the cognitive feature dimension is 16, the digital twin feature dimension is 128, and the fused feature dimension is 128. The 16-dimensional cognitive feature is expanded to 128-dimensional through the dimension alignment technology.

[0094] The fused feature is input into the encoding mapping module for processing. The encoding mapping module adopts a variational autoencoder structure, including an encoder and a decoder. The encoder consists of three fully connected neural networks, with the number of hidden layer neurons being 64 and 32, respectively, and the output layer being a 16-dimensional latent space vector. The encoding weight matrix is 128x64, 64x32, and 32x16, and the corresponding bias vector dimensions are 64, 32, and 16. The activation function uses LeakyReLU, and the negative slope parameter is set to 0.2. The encoder performs nonlinear transformation on the fused feature, compresses and maps the 128-dimensional feature to a 16-dimensional latent space, and obtains an encoded feature. The training of the encoding mapping adopts a combination of reconstruction loss and KL divergence loss, with the reconstruction loss weight coefficient set to 0.8 and the KL divergence loss weight coefficient set to 0.2. The optimizer uses Adam, with an initial learning rate of 0.001 and a training period of 100 rounds, each round containing 200 batches of data.

[0095] The EEG frequency spectrum of the encoded features is analyzed. This step is inspired by the human brain's cognitive mechanism and enhances the feature representation capability through frequency domain analysis. Specifically, for each element in the 16-dimensional encoded feature vector, it is treated as a time series signal, and the Fast Fourier Transform algorithm is used to calculate its frequency spectrum. The transform window length is set to 16, and the Hanning window is used to reduce spectral leakage. The calculated spectral density function contains amplitude and phase information, with the amplitude reflecting the energy distribution of the signal and the phase reflecting the structural characteristics of the signal. The spectrum is divided into five frequency bands: δ (0.5-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (> 30 Hz), each corresponding to different information characteristics. The energy proportion of each frequency band is calculated to construct a frequency band energy feature vector. Finally, the frequency band energy feature vector is concatenated with the original encoded features to form a 24-dimensional spectrum-enhanced encoding. In practical applications, spectral analysis significantly improves the model's ability to recognize periodic quality fluctuations, with detection accuracy increasing from 87.3% to 94.1%.

[0096] The spectrum-enhanced encoding is optimized based on a dynamic memory updating mechanism. The dynamic memory updating mechanism simulates the long-term and short-term memory characteristics of the human brain, and includes two key components: a memory state matrix M and a forgetting factor γ. The memory state matrix M has a dimension of k x 24, where k is the memory capacity, set to 64, representing the storage of 64 historical feature prototypes. The forgetting factor γ has a value range of [0, 1], with an initial value of 0.9, controlling the retention degree of historical memory. The memory updating process is as follows: first, calculate the similarity between the current spectrum-enhanced encoding and each prototype in the memory state matrix, using cosine similarity measure; then, select the most matching memory prototype according to the similarity, and apply the forgetting factor γ to the prototype to achieve exponential decay of historical memory; finally, fuse the decayed memory prototype with the current spectrum-enhanced encoding in proportion to update the memory state. The fusion ratio λ of memory and current encoding is adjusted adaptively according to the similarity, with higher similarity resulting in larger λ value, ranging from 0.3 to 0.7. After dynamic memory updating, the final 24-dimensional data encoding structure is obtained.

[0097] In practical application cases, a certain integrated circuit enterprise uses the above method to monitor the quality of wafer manufacturing process. The original feature dimension is 512, and after dimensionality reduction optimization, a 24-dimensional data encoding structure is obtained, with an information retention rate of 94.2%. The computational efficiency of the reduced features is improved by 21 times, and the model training time is reduced from 47 hours to 2.2 hours. In the quality anomaly detection task, the monitoring model based on the optimized encoding structure has a sensitivity of 95.3%, a specificity of 97.1%, and an F1 score of 0.962, which is better than the model built with the original unoptimized features (F1 score of 0.887). Especially for slight quality fluctuations in low signal-to-noise ratio environments, the detection capability is significantly improved, with a success rate increasing from 71.2% to 89.5%.

[0098] The embodiment realizes efficient dimension reduction optimization of the quality state of the raw material through six key steps of digital twin feature matrix construction, cognitive hierarchy processing, feature fusion, encoding mapping, electroencephalogram spectrum analysis and dynamic memory updating, and constructs a data encoding structure with strong expression ability, thereby providing a solid foundation for subsequent quality state evaluation and early warning.

[0099] In an optional embodiment, the quality state evaluation using the raw material digital twin model includes:

[0100] The layered attention mechanism of the raw material digital twin model is used to extract spatial features and time sequence features from the real-time data of the quality state of the raw material, and a quality state feature vector with spatio-temporal weights is obtained.

[0101] The quality state feature vector is input into a double-time memory network, short-term dynamic features and long-term evolution rules are extracted through fast memory units and slow memory units respectively, and a quality state evaluation index is generated based on the combination relationship between the two.

[0102] According to the historical distribution of the quality state evaluation index, a dynamic threshold optimization mechanism is used to calculate an alarm threshold, and a quality evaluation early warning signal is generated by comparing the deviation of the quality state evaluation index and the alarm threshold.

[0103] The raw material digital twin model is constructed to realize real-time monitoring and evaluation of the quality state of the raw material, and then generate a quality evaluation early warning signal.

[0104] When the raw material digital twin model is used for quality state evaluation, the system uses a layered attention mechanism to extract spatial features and time sequence features from the real-time data of the quality state of the raw material. The layered attention mechanism includes a spatial attention layer and a time sequence attention layer. The spatial attention layer mainly processes the correlation between data collected by sensors arranged at different positions of the raw material, and the time sequence attention layer processes dynamic change features in time sequence. Taking steel raw material as an example, assume that 10 sensor points are arranged on the surface of the steel, and each point collects 5 parameters such as temperature, hardness and stress, forming a spatial feature matrix with a shape of (10, 5). For each feature point, the system calculates the correlation degree with other feature points to obtain a weight coefficient. For example, the temperature parameter of the first sensor point has a high correlation with the stress parameter of the third sensor point, so a high weight 0.85 is given; and the correlation with the hardness parameter of the eighth sensor point is low, so a low weight 0.12 is given. These weights are used to adjust the importance of the original features to form weighted spatial features.

[0105] The temporal attention layer processes the sampled data over 24 hours, and for each time point, the system evaluates the degree of influence on the current state. For example, if the raw material temperature rises rapidly in the past 4 hours, the system assigns a higher weight of 0.78 to the temperature data in this period, and a lower weight of 0.05 to the data 24 hours ago. In this way, the system captures the key change patterns in the time dimension. Finally, the spatial features and temporal features are fused by weighting to form a quality state feature vector with a dimension of 128, and each element in the vector contains comprehensive information in the space-time dimension.

[0106] After inputting the quality state feature vector into the dual-time memory network, the system can capture both short-term dynamic features and long-term evolution rules. The dual-time memory network consists of a fast memory unit and a slow memory unit. The fast memory unit uses a forgetting gate control mechanism to process data in the last 30 minutes, quickly responding to sudden changes in the raw material state. For example, when the steel temperature suddenly rises from the standard temperature of 750°C to 830°C during processing, the fast memory unit can quickly capture this change pattern. The unit contains 64 neurons, extracts short-term features through a three-layer structure, and outputs a 64-dimensional short-term feature vector.

[0107] The slow memory unit focuses on the long-term change trend of the raw material quality state, processes historical data from the past 7 days, and identifies slow-developing quality degradation patterns. For example, the hardness of the steel slowly decreases from the initial hardness value of HRC 58 to HRC 55.5 due to slow oxidation in the environment, and this small but continuous change will be captured by the slow memory unit. The unit uses 32 neurons, retains long-term information through a memory update mechanism, and outputs a 32-dimensional long-term feature vector.

[0108] The system combines the outputs of the fast memory unit and the slow memory unit through the feature fusion layer, and adopts an adaptive weight scheme, giving a higher weight of 0.7 to long-term features and a weight of 0.3 to short-term features in stable states; while detecting abnormal changes, adjusting the weight of short-term features to 0.8 and the weight of long-term features to 0.2. After fusion, a comprehensive quality state evaluation index is generated, with a value range of 0-100, where 80-100 indicates a good quality state, 60-80 indicates a normal quality state but needs attention, 40-60 indicates a slight abnormality, and 0-40 indicates a serious quality problem.

[0109] According to the historical distribution of the quality state evaluation index, the system adopts a dynamic threshold optimization mechanism to calculate the alarm threshold. First, the system statistics the distribution characteristics of the quality state evaluation index in the past 90 days, including the mean, standard deviation, kurtosis and skewness. For stable production of raw materials, the system obtains the distribution characteristics of the mean of 78.6 and the standard deviation of 5.2. Based on these statistics, the system uses an adaptive algorithm to calculate the alarm threshold, which is usually set to the mean minus 2.5 times the standard deviation, which is 65.6 in this case. The system also dynamically adjusts the threshold sensitivity according to factors such as seasonal changes in production and equipment maintenance cycles, for example, during the stable period after equipment maintenance, the threshold coefficient can be adjusted from 2.5 to 2.2, making the threshold increase to 67.2, to reduce false positives.

[0110] When the real-time calculated quality state evaluation index is lower than the alarm threshold, the system generates a quality evaluation warning signal. The warning signal contains three levels: prompt, warning and emergency alarm, corresponding to different degrees of threshold deviation. For example, when the index value is 64.5, slightly lower than the threshold 65.6, the system generates a "prompt" level warning; when the index value drops to 58.2, the system upgrades to "warning" level; if the index further decreases to 45.0, the "emergency alarm" is triggered. Each warning signal contains information such as abnormal time, abnormal type, impact parameter and suggested treatment scheme, and is directly pushed to the quality management system and the mobile terminal of the relevant operator, realizing early warning and intervention of raw material quality problems.

[0111] In an alternative embodiment, based on the quality evaluation warning signal, a bio-inspired chaotic immune optimization algorithm is deployed to build a multi-agent collaborative traceability system, which simulates the adaptive mechanism of the biological immune system and the chaotic search strategy to realize rapid positioning and traceability of quality abnormalities, including:

[0112] The quality evaluation warning signal is input into the chaotic immune model, which generates an initialization sequence through iterative operation, builds an agent population based on the initialization sequence, and optimally distributes the agent population through the chaotic immune model. The chaotic immune model contains a warning signal intensity parameter and a temperature adjustment parameter, and the agent distribution corresponding to the warning signal intensity is obtained through dynamic adjustment of the warning signal intensity parameter and the temperature adjustment parameter;

[0113] Based on the agent distribution, the chaotic immune model calculates the response strength between the warning signal characteristics and the abnormal template characteristics, and constructs a search sequence with ergodicity through the response strength. The search sequence is adaptively combined with the historical search position, the current search path is updated, and the evolution trajectory of the abnormal characteristics is determined;

[0114] An agent collaborative network is constructed, the agent collaborative network is constructed based on spatial distance relationship between agents, an action strength between the agents is constructed through the chaotic immune model, the action strength is weighted and fused with a local traceability result of each agent, wherein the local traceability result contains abnormal position and propagation path information, and a traceability decision result of a global quality abnormality is generated through the weighted fusion.

[0115] A quality evaluation early warning signal is input into the chaotic immune model for processing. The quality evaluation early warning signal contains information such as abnormal type, abnormal strength, occurrence time and influence range. The early warning signal is represented in the form of a vector, with a dimension of 24, of which the first 8 dimensions represent the abnormal type, the middle 8 dimensions represent the abnormal strength and time information, and the last 8 dimensions represent the influence range. The chaotic immune model first generates an initialization sequence using the Logistic mapping, with the mapping parameter r set to 3.9 to ensure that the system is in a chaotic state. The initial seed value x0 is set to 0.4, and after 200 iterations, the last 100 values are taken as the initialization sequence. Based on the initialization sequence, an agent population is constructed, with a population size of 200, and each agent represents a candidate solution in the solution space.

[0116] The agent coding adopts real number coding, with a coding length of 32, the first 24 dimensions corresponding to the weights of each dimension of the early warning signal, and the last 8 dimensions corresponding to the traceability strategy parameters. The agent population is optimized and distributed through the chaotic immune model, and two key parameters are considered in the distribution process: the early warning signal strength parameter a and the temperature adjustment parameter β. The early warning signal strength parameter a takes a value in the range of [0.1, 0.9], with an initial value of 0.5, which increases with the increase of the early warning signal strength; the temperature adjustment parameter β takes a value in the range of [0.01, 1.0], with an initial value of 0.8, which controls the diversity of agent distribution.

[0117] In the optimization and distribution process, for early warning signals with high strength (such as abnormal strength exceeding 75 points), the value of a increases to 0.7-0.9, and the value of β decreases to 0.3-0.5, so that the agent distribution is more concentrated in the high-risk area; for early warning signals with low strength (such as abnormal strength less than 50 points), the value of a decreases to 0.2-0.4, and the value of β increases to 0.6-0.8, so that the agent distribution is more uniform and the search range is expanded. Through the dynamic adjustment of the early warning signal strength parameter and the temperature adjustment parameter, the agent distribution corresponding to the early warning signal strength is obtained. In actual application, for a semiconductor production line temperature abnormal early warning signal (strength of 82 points), through parameter setting (a = 0.8, β = 0.4), 65% of the agents are distributed to the temperature control system related area, 25% of the agents are distributed to the upstream material supply link, and 10% of the agents are distributed to the environmental impact factor area.

[0118] Based on the agent distribution, the response strength between the early warning signal features and the abnormal template features is calculated using the chaotic immune model. The abnormal template feature library contains the feature vectors of historical abnormal cases, with a total of 128 templates covering 12 common quality abnormalities. The response strength calculation uses cosine similarity measurement, considering both the angle difference and amplitude difference of the feature vectors. The response strength R ranges from 0 to 1, and the larger the R value, the more similar the current early warning signal to the template features.

[0119] The chaotic immune model constructs a search sequence with ergodicity through response strength. The search sequence is generated using Tent mapping, with the control parameter μ set to 1.99 and the initial point taken from the feature vectors corresponding to the top 3 templates with the highest response strength. The search sequence length is set to 300, which is the boundary of the problem solution space. The search sequence is adaptively combined with the historical search positions, with the combination weight λ dynamically adjusted based on the historical search results. λ ranges from 0.2 to 0.8, with an initial value of 0.5. When the historical search path discovers a high response strength region, λ increases, enhancing the use of historical experience. When the historical search falls into a local optimum, λ decreases, enhancing the randomness of chaotic search.

[0120] The current search path is updated through adaptive weight combination, and the response strength of the new path is evaluated after each update. The top 20 positions with the highest response strength are retained. After 300 iterations, the evolution trajectory of the abnormal features is determined. In a certain photoresist quality abnormality case, this method traces the complete evolution trajectory of the abnormal features from the supplier's raw material purity fluctuation, through solvent ratio deviation, to the final photoresist viscosity abnormality, with a response strength peak of 0.87 and a tracing accuracy of 92%.

[0121] A multi-agent collaborative network is constructed to realize multi-agent collaborative tracing. The multi-agent collaborative network is established based on the spatial distance relationship between agents. The spatial distance is calculated using the Euclidean distance, and the distance threshold dth is set to 25% of the diameter of the solution space. When the distance between two agents is less than dth, a connection is established, forming a dynamic network with variable topology. The interaction strength between agents is constructed using the chaotic immune model. The interaction strength S is inversely proportional to the distance between agents and proportional to the agent fitness (response strength). In the specific calculation formula, the distance decay coefficient is set to 2.0, and the fitness gain coefficient is set to 1.5. The interaction strength S ranges from 0 to 1, and the larger the S value, the stronger the collaboration effect between two agents. The interaction strength is weighted and fused with the local tracing results of each agent. The local tracing results include the abnormal occurrence position coordinates (positioning in the production process parameter space) and the propagation path information (represented as a directed graph structure, with nodes as key process parameters and edges as influence relationships).

[0122] The weighted fusion adopts a soft voting mechanism, and the voting weight of each agent is proportional to its response strength and the average action strength with other agents. The global quality anomaly tracing decision result is generated by weighted fusion, including the most abnormal source location (confidence score), abnormal propagation key path and impact range prediction. In practical application, a certain integrated circuit manufacturing enterprise uses this method to trace the wafer oxide layer thickness anomaly. After 500 iterations of 200 agents, the temperature controller parameter offset of the oxidation furnace is successfully identified as the root cause, with a confidence score of 0.89, and the tracing time is shortened from 4 hours of the traditional method to 22 minutes.

[0123] The specific implementation of the chaotic immune optimization algorithm includes the following key technical points: the cloning selection operation simulates the cloning proliferation process after the antibody recognizes the antigen in the biological immune system. The cloning number is proportional to the response strength of the agent, and the upper limit of the cloning number is set to 10. The cloning ratio coefficient is set to 0.8, the agent with the highest response strength is cloned 10 times, and the agent with the lowest response strength is cloned 1 time. The high mutation operation performs large-scale mutation on the low response strength agent, the mutation proportion is set to 0.3, and the mutation amplitude is set to 30% of the diameter of the solution space; the low mutation operation performs small-scale fine mutation on the high response strength agent, the mutation proportion is set to 0.1, and the mutation amplitude is set to 5% of the diameter of the solution space.

[0124] The immune memory mechanism saves the historical optimal solution, and the memory bank capacity is set to 20, which is updated every 10 iterations. The new solution replacement strategy adopts the combination of elite reservation and roulette, and the reservation rate is set to 0.2. The chaotic disturbance mechanism is triggered when the algorithm falls into local optimum, and the trigger condition is that the optimal solution improvement amplitude is less than 0.1% for 30 consecutive iterations. The disturbance adopts Logistic mapping to generate a disturbance sequence, and the disturbance strength decreases with the increase of iteration number. The initial disturbance strength is 20% of the diameter of the solution space, and finally decreases to 2%. The convergence criterion is that the optimal solution improvement amplitude is less than 0.01% for 50 consecutive iterations or the maximum iteration number is 1000.

[0125] The multi-agent collaborative tracing system adopts a hierarchical architecture: the perception layer is composed of 200 sensing nodes, responsible for real-time data acquisition; the analysis layer includes a central processing unit and an edge computing node, responsible for abnormal feature extraction and pattern recognition; the decision layer is composed of chaotic immune optimization algorithm and expert knowledge base, responsible for generating tracing strategy; the execution layer includes feedback control module, responsible for the implementation of quality measures. The system communication adopts industrial Ethernet protocol, the data transmission rate is 100Mbps, and the delay is controlled within 10ms. The system fault tolerance mechanism includes data redundancy storage and backup decision module, which automatically switches to the backup system when the response time of the main system exceeds the set threshold (200ms). The system security is ensured by data encryption (AES-256 algorithm) and access control.

[0126] In an application case at a semiconductor wafer manufacturing company, the system successfully handled 93 quality anomaly events, achieving an average traceability accuracy of 91.7% and an average traceability time of 15.4 minutes, representing an efficiency improvement of approximately 76% compared to traditional methods. Particularly for complex, multi-factor coupled anomalies, the identification rate increased from 65% to 87% using traditional methods. After six months of system operation, the company's first-pass yield increased by 4.2 percentage points, quality costs decreased by approximately 15%, resulting in annualized cost savings of approximately 5.8 million yuan.

[0127] Figure 4 This is a schematic diagram comparing the accuracy of quality anomaly tracing using different tracing methods in this invention:

[0128] This figure compares the average tracing time of three different anomaly tracing methods when facing anomalies of varying complexity. Traditional manual tracing represents a method relying entirely on human experience and manual analysis; general intelligent tracing algorithms employ basic machine learning models such as decision trees and random forests for anomaly analysis; and this technical solution is an intelligent tracing system based on chaos-immune optimization. Test results for anomaly scenarios ranging from level 1 to level 6 complexity show that at level 1 complexity, the three methods took 70.8 minutes, 41.5 minutes, and 10.3 minutes, respectively; as the complexity increased to level 3, the times increased to 142.6 minutes, 75.3 minutes, and 21.5 minutes, respectively; and at the highest level 6 complexity, traditional manual tracing required 312.8 minutes, general intelligent tracing methods required 168.7 minutes, while this technical solution only required 42.5 minutes. The data demonstrates that this technical solution maintains a significant time advantage across all complexity levels, and its relative advantage becomes more pronounced as the anomaly complexity increases. Especially in high-complexity scenarios (level 5-6), the tracing time of this technical solution shows a significantly slower growth trend, demonstrating excellent scalability and robustness.

[0129] In one optional implementation, based on the quality assessment early warning signal, a bio-inspired chaotic immune optimization algorithm is deployed to construct a multi-agent collaborative tracing system. By simulating the adaptive mechanism and chaotic search strategy of the biological immune system, rapid location and tracing of quality anomalies are achieved, including:

[0130] The quality assessment warning signal is input into the chaotic immune optimization algorithm. The chaotic immune optimization algorithm constructs an antigen-antibody response mechanism by calculating the affinity between the quality assessment warning signal and the abnormal template. Based on the affinity, the number of clones is determined and an initial immune population is generated. The initial immune population is perturbed and updated using a chaotic sequence. The chaotic sequence is weighted and combined with the historical best antibody to determine the search direction, resulting in an optimized immune population.

[0131] An optimized immune population is constructed based on the optimized immune population, and antibody characteristics in the optimized immune population are mapped to initial states of intelligent agents. The intelligent agents explore abnormal characteristics in a quality state space through a chaotic search strategy, wherein the chaotic search strategy includes two stages of local fine search and global rough search, and a search step is dynamically adjusted to quickly locate an abnormal area.

[0132] An intelligent agent interaction network is constructed in the multi-agent collaborative tracing system. An interaction weight matrix is calculated based on spatial distances between intelligent agents, the interaction weight matrix is weighted and fused with abnormal characteristic information collected by each intelligent agent, a quality abnormality propagation path graph is constructed based on the fused abnormal characteristic information, and a source position and a diffusion direction of the quality abnormality are determined by backtracking the propagation path graph.

[0133] A quality evaluation warning signal is input into a chaotic immune optimization algorithm for processing. The quality evaluation warning signal is represented as a 32-dimensional vector, including abnormal type encoding (8 dimensions), abnormal intensity index (6 dimensions), time characteristics (10 dimensions), and spatial distribution characteristics (8 dimensions). The chaotic immune optimization algorithm constructs an antigen-antibody response mechanism by calculating the affinity between the quality evaluation warning signal and the abnormal template. The affinity calculation uses cosine similarity measurement, with a value range of [0, 1], and a larger value indicates a higher matching degree. The abnormal template library contains 200 predefined templates, covering 15 common quality abnormality patterns. Different weights are assigned to different characteristic dimensions when calculating the affinity. The abnormal type weight is 0.4, the abnormal intensity weight is 0.3, the time characteristic weight is 0.2, and the spatial distribution characteristic weight is 0.1. For a practical case in a semiconductor lithography process, the affinity of a certain warning signal with the "photoresist uniformity abnormality" template is 0.87, the affinity with the "incomplete development" template is 0.34, and the affinity with other templates is less than 0.3.

[0134] The number of clones is determined based on the affinity and an initial immune population is generated. The number of clones is proportional to the affinity, and the calculation formula is that the number of clones is equal to the maximum number of clones (set to 10) multiplied by the affinity, and the result is rounded up. Nine antibodies with an affinity of 0.87 are cloned, four antibodies with an affinity of 0.34 are cloned, and the remaining antibodies are cloned according to their respective affinities. The initial immune population size is set to 100, including antibodies selected from the template library and their clones.

[0135] Each antibody in the population is represented as a 32-dimensional vector, consistent with the dimensionality of the early warning signals. The initial immune population is updated by perturbation using a chaotic sequence. The chaotic sequence is generated using the Logistic map, with a control parameter r set to 3.99 and an initial value x0 set to 0.63. After 200 iterations, the last 100 values are taken as the chaotic sequence. The chaotic sequence has a value range of [0, 1] and is mapped to the actual value range of each parameter through linear transformation. The search direction is determined by the weighted combination of the chaotic sequence and the historical best antibody, with the combination weight α dynamically adjusted with the number of iterations. The initial value is 0.3, which gradually increases to 0.8 as the number of iterations increases, enhancing the use of historical experience. During the perturbation update process, the update amplitude of each antibody is inversely proportional to its affinity. Antibodies with high affinity are searched with small amplitude, while antibodies with low affinity are searched with large amplitude. After 100 iterations, the optimized immune population is obtained, with the average affinity increasing from 0.42 to 0.78.

[0136] Based on the optimized immune population, a multi-agent collaborative traceability system is constructed. The antibody features in the optimized immune population are mapped to the initial state of the agent. During the mapping process, the 32-dimensional feature vector of the antibody is converted to the state vector of the agent, including position coordinates (coordinates in the quality state space), search direction, moving speed, and perception range. The total number of agents is 100, consistent with the size of the immune population. Agents explore abnormal features in the quality state space through chaotic search strategy. The quality state space is an 8-dimensional space representing the state combination of key quality parameters. The chaotic search strategy includes two stages: local fine search and global coarse search.

[0137] In the local fine search stage, agents search in high affinity areas with small step sizes, set as 1% to 5% of the space diameter; in the global coarse search stage, agents explore in low affinity areas with large step sizes, set as 10% to 30% of the space diameter. The search step size is dynamically adjusted during the search process, with the adjustment formula being the current step size equal to the base step size multiplied by the chaotic factor and then multiplied by the inverse function of the affinity. The base step size gradually decreases with the number of iterations, with an initial value of 20% of the space diameter and a final value of 2%. The chaotic factor is generated through Tent mapping, with a control parameter μ set to 1.97. In the photoresist uniformity anomaly case, through dynamic step size adjustment, the agent group successfully locates the abnormal area after 50 iterations, with a positioning accuracy of 97.3%.

[0138] An agent interaction network is constructed in the multi-agent collaborative traceability system. An interaction weight matrix is calculated based on the spatial distance between agents, and the spatial distance is calculated using the Euclidean distance. The interaction weight is inversely proportional to the distance, and the calculation formula is that the interaction weight is equal to the distance threshold divided by the sum of the actual distance and the distance threshold, and the distance threshold is set to 15% of the spatial diameter. When the distance between two agents is less than the distance threshold, the interaction weight is greater than 0.5; when the distance is equal to the distance threshold, the interaction weight is equal to 0.5; and when the distance is greater than the distance threshold, the interaction weight is less than 0.5 but always greater than 0.

[0139] The interaction weight matrix is a 100x100 square matrix, and the diagonal elements are set to 1, indicating the interaction weight of the agent with itself. The interaction weight matrix is weighted and fused with the abnormal feature information collected by each agent, including parameter deviation value, change rate, and fluctuation frequency data. The weighted fusion adopts a weighted average method, and the fused abnormal feature is represented as the weighted sum of all agent features, and the weight is the corresponding interaction weight.

[0140] A propagation path graph of quality abnormality is constructed according to the fused abnormal feature information. The propagation path graph adopts a directed graph structure, the nodes represent quality parameters, the edges represent the influence relationship between parameters, and the weight of the edge represents the influence strength. The graph construction process is based on a causal inference algorithm, which determines the causal direction by analyzing the time sequence correlation of parameter changes. The node importance is evaluated by the centrality index, including degree centrality, closeness centrality and betweenness centrality, with weights of 0.3, 0.3 and 0.4 respectively. The source location and diffusion direction of the quality abnormality are determined by backtracking the propagation path graph. The backtracking algorithm starts from the abnormal performance node and searches in the reverse direction along the incoming edge, calculates the cumulative weight of each path, and selects the path with the maximum cumulative weight as the optimal abnormal propagation path. The source location is the starting node of the path, and the diffusion direction is determined by the path direction.

[0141] In the practical application of a semiconductor manufacturing plant, the method is used to trace the abnormality of photoresist uniformity. By analyzing the traceability results, it is determined that the abnormal source is the parameter drift of the temperature control unit of the photoresist supply system, and the cumulative path weight is 0.86, which is significantly higher than the 0.52 of the suboptimal path. The abnormal diffusion direction is temperature control unit-photoresist viscosity-photoresist uniformity-pattern size deviation. The entire traceability process takes 17 minutes, which is more than 85% more efficient than the traditional manual traceability method (which usually takes 2-4 hours). Six months of operation data show that the traceability accuracy of this method is 93.7%, which is much higher than the 78.2% of the traditional method. Especially for multi-factor coupled abnormalities, the accuracy is significantly improved from 65.4% to 89.1%.

[0142] The practical application effect of the method is remarkable: after a certain integrated circuit manufacturing enterprise applies the method, the product yield is increased by 3.5 percentage points, from 91.2% to 94.7%; the abnormal response time is shortened by an average of 75%, from 3.2 hours to 0.8 hours; the quality cost is reduced by about 12%, and the annual savings reach 4.3 million yuan. The method is especially suitable for the fields of high-precision, high-value semiconductors, precision optics and advanced materials manufacturing, and has important value for improving product quality stability and production efficiency.

[0143] Figure 5 The performance comparison analysis column chart of the chaotic immune optimization algorithm of the embodiment of the application:

[0144] The figure shows the performance comparison data of three different traceability methods. Among them, the traditional method represents a traceability algorithm based on statistical analysis, mainly relying on data statistics and correlation analysis for traceability; the general intelligent algorithm adopts traditional machine learning methods such as artificial neural network and support vector machine; the chaotic immune optimization algorithm is a new type of intelligent optimization algorithm that combines biological immune mechanism and chaotic dynamic characteristics. In terms of specific performance indicators, in terms of traceability accuracy, the chaotic immune optimization algorithm reaches 93.8%, which is significantly higher than the 76.5% of the traditional method and the 85.7% of the general intelligent algorithm; in terms of abnormal detection rate, the chaotic immune optimization algorithm reaches 95.2%, which is also superior to the 82.3% of the traditional method and the 88.6% of the general intelligent algorithm. Especially in the efficiency indicator of average traceability time, the chaotic immune optimization algorithm only needs 23.5 minutes to complete the traceability, which is significantly more time-efficient than the 168.4 minutes of the traditional method and the 68.2 minutes of the general intelligent algorithm. The data shows that the chaotic immune optimization algorithm combines biological heuristic mechanism and chaotic search strategy, greatly improving the traceability efficiency while ensuring high accuracy.

[0145] In an optional implementation, the multi-agent collaborative traceability system is used to activate an intelligent traceability engine according to the quality evaluation warning signal, locate a quality abnormal root cause based on a space-time graph reasoning mechanism, and generate a traceability analysis result, including:

[0146] The intelligent traceability engine is activated based on the quality evaluation warning signal, the quality evaluation warning signal is input into a space-time graph model for warning intensity analysis, the space-time graph model is used for threshold value judgment on the warning intensity, an engine activation state is generated, and the space-time graph model includes a sensitivity parameter and an activation threshold value;

[0147] The correlation structure of time series and spatial topology is constructed, the space-time graph model is used to calculate the time sequence evolution characteristics of the quality state, the space-time graph model is used to analyze the spatial position relationship, the time sequence evolution characteristics and the spatial position relationship are fused, and a space-time correlation feature is formed;

[0148] Based on the spatio-temporal correlation feature, an abnormal propagation path is constructed, the inter-node propagation strength is calculated through the spatio-temporal graph model, the inter-node propagation strength is taken as a node correlation weight to construct an abnormal propagation network, and an abnormal diffusion path is identified based on the abnormal propagation network;

[0149] The abnormal diffusion path is located, the importance score of the path node is calculated by using the spatio-temporal graph model, the abnormal source point position is determined, and a traceability analysis result is generated based on the abnormal source point position and the abnormal diffusion path.

[0150] The process of activating the intelligent traceability engine based on the quality evaluation early warning signal first converts the quality evaluation early warning signal into a standardized format. The early warning signal is represented as a 32-dimensional vector, containing abnormal type, abnormal degree, abnormal location and timestamp information. The standardized early warning signal is input into the spatio-temporal graph model for early warning intensity analysis. The early warning intensity calculation considers the weighted combination of three factors: abnormal degree, influence range and duration. The abnormal degree weight is set to 0.5, the influence range weight is set to 0.3, and the duration weight is set to 0.2.

[0151] A semiconductor manufacturing line detected an abnormality in the thickness of the silicon wafer oxide layer, the abnormal degree of the early warning signal was 78 points (full score 100 points), the influence range involved 3 production batches and a total of 450 silicon wafers, and the duration was 2.5 hours. The calculated early warning intensity is 78x0.5+85x0.3+65x0.2=76.5 points. The spatio-temporal graph model is used to judge the threshold value of the early warning intensity. The spatio-temporal graph model contains two key parameters: sensitivity parameter a and activation threshold b. The sensitivity parameter a controls the response sensitivity of the model to the early warning signal, with a value range of [0.8, 1.2] and a default value of 1.0. It can be adjusted according to the production environment, with a high-precision production line set to 1.1-1.2 and a general production line set to 0.8-0.9. The activation threshold b is the critical value of the early warning intensity that triggers the traceability engine, with a value range of [60, 80] and a default value of 70. When the value of the early warning intensity multiplied by the sensitivity parameter is greater than the activation threshold, the engine activation state is "start"; otherwise, it is "standby". In the above example, the early warning intensity 76.5 multiplied by the sensitivity parameter 1.1 equals 84.15, which is greater than the activation threshold 70, so the intelligent traceability engine is successfully activated.

[0152] The process of constructing the time series and spatial topology correlation structure involves two parallel analysis paths. The time series analysis uses a sliding window technique with a window size of 24 hours and a sliding step of 1 hour to calculate the time series evolution characteristics of the quality state using the time-space atlas model. The time series feature extraction algorithm is applied to the quality data within the window to extract trend features, periodic features, and mutation features. The trend features are described by the linear regression slope, the periodic features are described by the main frequency component of the Fourier transform, and the mutation features are identified by the change point detection algorithm.

[0153] In the case of the silicon wafer oxide layer, the time series analysis found that the oxide layer thickness showed a weak downward trend (slope -0.5 nm / hour) 12 hours before the anomaly, periodic fluctuations (period about 45 minutes) 4 hours before the anomaly, and a mutation point (thickness dropped 5 nm) 1 hour before the anomaly. At the same time, the time-space atlas model was used to analyze the spatial position relationship. The spatial analysis is based on the construction of a spatial topology network based on the production process flowchart. The nodes represent process stations or equipment, and the edges represent material flow or process dependency relationships. For each node, the process distance (number of process steps passed) and physical distance (actual spatial distance) from the anomaly detection point are calculated.

[0154] The spatial analysis determines the key nodes related to the oxide layer thickness anomaly, including the oxidation furnace (process distance 0, physical distance 0 meters), gas supply system (process distance 1, physical distance 15 meters), temperature control unit (process distance 1, physical distance 5 meters), and cleaning station (process distance 2, physical distance 30 meters). The time series evolution characteristics and spatial position relationship are fused to form the time-space correlation characteristics. The fusion uses a weighted combination method, with the time series feature weight being 0.6 and the spatial feature weight being 0.4. The fused time-space correlation characteristics are represented as a 64-dimensional vector, with the first 32 dimensions encoding the time series information and the last 32 dimensions encoding the spatial information.

[0155] The process of constructing the anomaly propagation path based on the time-space correlation characteristics first calculates the propagation strength between nodes using the time-space atlas model. The propagation strength calculation considers four factors: time series precedence (the order of parameter changes), correlation (statistical correlation of parameter changes), physical dependency (causal relationship based on domain knowledge), and distance attenuation (influence attenuation based on spatial distance). The weights of each factor are 0.3, 0.3, 0.3, and 0.1, respectively.

[0156] For example, the propagation strength calculation between the temperature control unit and the oxide layer thickness: the timing lead score is 0.85 (the temperature fluctuation leads the thickness change by about 40 minutes), the correlation score is 0.78 (the Pearson correlation coefficient), the physical dependence score is 0.9 (according to the expert knowledge of semiconductor manufacturing), the distance attenuation score is 0.8 (based on a physical distance of 5 meters), and the comprehensive propagation strength is 0.85x0.3+0.78x0.3+0.9x0.3+0.8x0.1=0.839.

[0157] The inter-node propagation strength is used as the node association weight to construct an abnormal propagation network, which adopts a directed graph structure, with nodes representing process parameters or device states and edges representing the influence relationship between parameters, and the edge weight being the corresponding propagation strength. The propagation strength threshold is set to 0.5, and only edges with a strength greater than the threshold are retained to ensure that the network structure clearly reflects the main propagation path. Based on the abnormal propagation network, the abnormal diffusion path is identified, and the diffusion path identification uses the maximum flow algorithm, with the abnormal detection point being the sink point and the source point and propagation path being searched. The algorithm considers the minimum propagation strength and path length on the path, and preferentially selects the diffusion path with high minimum propagation strength and short path length. In the case, the main diffusion path identified is: gas supply system-temperature control unit-oxidation furnace-oxide layer thickness, with a minimum propagation strength of 0.72 and a path length of 3.

[0158] The process of root cause positioning of the abnormal diffusion path uses the time-space graph model to calculate the importance score of the path nodes. The importance score considers the in-degree centrality, out-degree centrality, betweenness centrality, and feature vector centrality of the nodes, with weights of 0.2, 0.3, 0.2, and 0.3, respectively. The in-degree centrality represents the degree to which a node is affected by other nodes, the out-degree centrality represents the degree to which a node affects other nodes, the betweenness centrality represents the bridging role of a node in the network, and the feature vector centrality represents the degree to which a node is connected to other important nodes.

[0159] Each centrality index is normalized to have a value range of [0, 1]. In the case, the importance score of the gas supply system is 0.2x0.4+0.3x0.9+0.2x0.6+0.3x0.85=0.725, the importance score of the temperature control unit is 0.2x0.6+0.3x0.85+0.2x0.7+0.3x0.8=0.755, and the importance score of the oxidation furnace is 0.2x0.85+0.3x0.6+0.2x0.5+0.3x0.7=0.66. According to the importance score, the location of the abnormal source point is determined, and the node with the highest score is selected as the abnormal source point. In this case, the temperature control unit has the highest importance score (0.755) and is determined as the abnormal source point.

[0160] The abnormal source point position and abnormal diffusion path are used to generate a traceability analysis result. The traceability analysis result includes four parts: abnormal source point information (including source point name, position, parameter state and abnormal degree), abnormal propagation path (including path node sequence and propagation intensity of each section), impact assessment (including impact range, impact degree and duration) and suggestion (including short-term control measures and long-term scheme). In the case, the traceability analysis result shows that the abnormal source point is the temperature control unit, which is located at the oxidation area A3 station, the temperature control parameter deviates by +3.2℃, and the abnormal degree is serious (85 / 100); the abnormal propagation path is the temperature control unit (propagation intensity 0.839) - oxidation furnace (propagation intensity 0.792) - oxidation layer thickness, which causes the thickness to be thin by 7.5nm; the impact assessment shows that 3 production batches of 450 silicon wafers are affected, the yield rate decreases by 12.5%, and the economic loss is about 375,000 yuan; the suggestions include immediately calibrating the temperature control unit parameters, temporarily adjusting the oxidation time to compensate for the thickness deviation, and long-term increasing the temperature control redundant monitoring points to prevent similar problems.

[0161] The traceability method has achieved remarkable results in the actual application of semiconductor manufacturing enterprises. After applying this method, the accuracy of quality abnormality traceability of a certain 12-inch wafer factory has been improved from 76.3% to 93.8%, and the average traceability time has been shortened from 2.8 hours to 23 minutes, with an efficiency improvement of 86%. For complex multi-factor coupled abnormalities, the traceability accuracy has been improved from 62.5% to 89.7%. The method is particularly suitable for high-precision, high-value semiconductor, precision optical and advanced material manufacturing fields, and provides strong support for improving product quality and production efficiency.

[0162] In a second aspect of the embodiments of the present application, a chip raw material quality state tracking and traceability system is provided, comprising:

[0163] A first unit is configured to acquire real-time data in the production and preparation process of the chip raw material, construct a raw material digital twin model according to the real-time data, establish a mapping relationship between the physical world and the information world, and realize dynamic monitoring of the raw material quality state;

[0164] A second unit is configured to use the raw material digital twin model to optimize the raw material quality state, use the raw material digital twin model to construct a data coding structure, map the optimized data to a feature space, extract the spatio-temporal evolution features of the raw material quality state through the raw material digital twin model, input the spatio-temporal evolution features into the raw material digital twin model, use the raw material digital twin model to evaluate the quality state, and generate a quality evaluation warning signal;

[0165] The third unit is configured to deploy a biological heuristic-based chaotic immune optimization algorithm based on the quality evaluation early warning signal, construct a multi-agent collaborative traceability system, realize rapid positioning and traceability of quality abnormalities by simulating the adaptive mechanism and chaotic search strategy of the biological immune system, and improve the accuracy and convergence speed of abnormal pattern recognition by fusing antibody clone selection, immune memory and chaotic disturbance mechanism in the chaotic immune optimization algorithm.

[0166] The fourth unit is configured to activate an intelligent traceability engine according to the quality evaluation early warning signal by using the multi-agent collaborative traceability system, locate the root cause of quality abnormalities based on a space-time graph reasoning mechanism, and generate traceability analysis results; and perform closed-loop feedback control according to the traceability analysis results for optimization and adjustment of raw material production processes, and realize dynamic management and control of quality states.

[0167] The third aspect of the embodiment of the present application provides an electronic device, comprising:

[0168] a processor;

[0169] a memory for storing processor-executable instructions;

[0170] The processor is configured to call the instructions stored in the memory to execute the method described above.

[0171] The fourth aspect of the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method described above.

[0172] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for executing various aspects of the present application.

[0173] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A chip raw material quality state tracking and tracing method, characterized in that, The method comprises the following steps: acquiring real-time data in the production process of chip raw materials, constructing a raw material digital twin model according to the real-time data, establishing a mapping relationship between the physical world and the information world, and realizing dynamic monitoring of the quality state of the raw materials; dimensionality reduction optimization of the quality state of the raw materials is performed by using the raw material digital twin model, a data coding structure is constructed by using the raw material digital twin model, the optimized data is mapped to a feature space, and the spatiotemporal evolution characteristics of the quality state of the raw materials are extracted by using the raw material digital twin model; the spatiotemporal evolution characteristics are input into the raw material digital twin model, quality state evaluation is performed by using the raw material digital twin model, and a quality evaluation early warning signal is generated; based on the quality evaluation early warning signal, a multi-agent collaborative traceability system is constructed by deploying a biological heuristic-based chaotic immune optimization algorithm, the adaptive mechanism of the biological immune system and the chaotic search strategy are simulated, and rapid positioning and traceability of quality abnormalities are realized, including: the quality evaluation early warning signal is input into the chaotic immune optimization algorithm, the chaotic immune optimization algorithm constructs an antigen-antibody response mechanism by calculating the affinity between the quality evaluation early warning signal and an abnormal template, determines the number of clones and generates an initial immune population based on the affinity, and updates the initial immune population by using a chaotic sequence, wherein the chaotic sequence and the historical optimal antibody are combined to determine the search direction, and an optimized immune population is obtained; a multi-agent collaborative traceability system is constructed based on the optimized immune population, antibody characteristics in the optimized immune population are mapped to the initial state of an agent, and the agent explores abnormal characteristics in the quality state space by using a chaotic search strategy, wherein the chaotic search strategy includes two stages of local fine search and global rough search, and the search step is dynamically adjusted to realize rapid positioning of abnormal regions; an agent interaction network is constructed in the multi-agent collaborative traceability system, an interaction weight matrix is calculated based on the spatial distance between agents, the interaction weight matrix and abnormal characteristic information collected by each agent are weighted and fused, a quality abnormality propagation path graph is constructed according to the fused abnormal characteristic information, and the source position and diffusion direction of the quality abnormality are determined by backtracking the propagation path graph; the multi-agent collaborative traceability system is used to activate an intelligent traceability engine according to the quality evaluation early warning signal, locate the root cause of the quality abnormality based on the inference mechanism of the spatiotemporal graph, and generate a traceability analysis result; closed-loop feedback control is performed according to the traceability analysis result, which is used for optimization and adjustment of the raw material production process, and realizes dynamic control of the quality state.

2. The method of claim 1, wherein, The raw material digital twin model is constructed according to the real-time data, a mapping relationship between the physical world and the information world is established, and dynamic monitoring of the quality state of the raw materials is realized, including: a physical equation set containing concentration field, diffusion coefficient, temperature and pressure is established according to the real-time data, and the physical equation set is used to represent the physical characteristics in the production process of the raw materials; inputting the real-time data into the deep autoencoder, extracting hidden layer features of the real-time data through a weight matrix and a bias term of the deep autoencoder, and constructing a raw material digital twin model based on the hidden layer features; adopting a physically constrained loss function to optimize the raw material digital twin model, wherein the physically constrained loss function comprises a data fitting error term and a physical constraint term, and the physical constraint term is determined by the system of physical equations; constructing a high-dimensional feature vector space, mapping the raw material digital twin model to the high-dimensional feature vector space, and dynamically updating feature weights in the high-dimensional feature vector space through an adaptive weight adjustment mechanism; based on Kalman filtering, predicting the state of the features in the high-dimensional feature vector space to generate a state vector, calculating a distribution confidence interval according to the state vector, determining a single indicator score of the raw material quality state based on the distribution confidence interval, weighting and summing the single indicator score and the corresponding weight coefficient to obtain a quality state comprehensive evaluation index, and realizing dynamic monitoring of the raw material quality state.

3. The method of claim 1, wherein, adopting the raw material digital twin model to optimize the raw material quality state in a dimension-reduced manner, and constructing a data coding structure using the raw material digital twin model, comprising: constructing a digital twin feature matrix, wherein the digital twin feature matrix comprises a plurality of feature vectors, and each feature vector corresponds to a feature dimension of the raw material quality state; inputting the feature vectors in the digital twin feature matrix into a cognitive hierarchical structure, wherein the cognitive hierarchical structure comprises a plurality of cognitive layers, each cognitive layer is provided with a corresponding cognitive weight, and the feature vectors are weighted and processed through the cognitive weight to obtain a cognitive feature corresponding to each cognitive layer; performing feature fusion on the cognitive feature and the digital twin feature matrix, wherein the feature fusion comprises multiplying the cognitive feature and the digital twin feature matrix by corresponding fusion weight coefficients and then superimposing them to generate a fusion feature; inputting the fusion feature into an encoding mapping module, wherein the encoding mapping module comprises an encoding weight matrix and a bias vector, and the fusion feature is nonlinearly transformed through the encoding weight matrix and the bias vector to obtain an encoding feature; performing electroencephalogram spectrum analysis on the encoding feature, calculating the spectral density function of the encoding feature through Fourier transform, and obtaining a spectrum-enhanced code; based on a dynamic memory updating mechanism, optimizing the spectrum-enhanced code, wherein the dynamic memory updating mechanism comprises a memory state and a forgetting factor, wherein the memory state is attenuated by the forgetting factor and combined with the current spectrum-enhanced code to obtain a data coding structure.

4. The method of claim 1, wherein, using the raw material digital twin model to evaluate the quality state and generate a quality evaluation warning signal, comprising: using a hierarchical attention mechanism of the raw material digital twin model to extract spatial features and time sequence features of raw material quality state real-time data to obtain a quality state feature vector with spatio-temporal weights; The quality state feature vector is input into a dual-time memory network, short-term dynamic features and long-term evolution rules are extracted through a fast memory unit and a slow memory unit respectively, and a quality state evaluation index is generated based on the combination of the two; According to the historical distribution of the quality state evaluation index, a dynamic threshold optimization mechanism is used to calculate the alarm threshold, and by comparing the deviation of the quality state evaluation index and the alarm threshold, a quality evaluation early warning signal is generated.

5. The method of claim 1, wherein, The multi-agent collaborative traceability system is used to activate the intelligent traceability engine based on the quality evaluation early warning signal, and the quality abnormal root is located based on the reasoning mechanism of the space-time graph, and the traceability analysis result is generated, including: Based on the quality evaluation early warning signal, the intelligent traceability engine is activated, the quality evaluation early warning signal is input into the space-time graph model for early warning intensity analysis, the space-time graph model is used to judge the threshold value of the early warning intensity, and the engine activation state is generated, the space-time graph model includes sensitivity parameters and activation threshold; The correlation structure of time series and spatial topology is constructed, the time sequence evolution characteristics of the quality state are calculated by using the space-time graph model, and the spatial position relationship is analyzed by using the space-time graph model, the time sequence evolution characteristics and the spatial position relationship are fused to form the space-time correlation characteristics; Based on the space-time correlation characteristics, an abnormal propagation path is constructed, the inter-node propagation intensity is calculated by using the space-time graph model, the inter-node propagation intensity is used as the node correlation weight to construct an abnormal propagation network, and the abnormal diffusion path is identified based on the abnormal propagation network; The abnormal diffusion path is located, the importance score of the path node is calculated by using the space-time graph model, the abnormal source point position is determined, and the traceability analysis result is generated based on the abnormal source point position and the abnormal diffusion path.

6. A chip raw material quality status tracking and tracing system for implementing the method according to any one of the preceding claims 1-5, characterized in that, It includes: The first unit is used for acquiring real-time data in the production and preparation process of chip raw materials, constructing a raw material digital twin model based on the real-time data, establishing a mapping relationship between the physical world and the information world, and realizing dynamic monitoring of the quality state of raw materials; The second unit is used for dimension reduction optimization of the raw material quality state by using the raw material digital twin model, constructing a data coding structure by using the raw material digital twin model, mapping the optimized data to a feature space, extracting the space-time evolution characteristics of the raw material quality state by using the raw material digital twin model, inputting the space-time evolution characteristics into the raw material digital twin model, and generating a quality evaluation early warning signal by using the raw material digital twin model for quality state evaluation; The third unit is used for deploying a bio-inspired chaotic immune optimization algorithm based on the quality evaluation early warning signal, constructing a multi-agent collaborative traceability system, simulating the adaptive mechanism and chaotic search strategy of the biological immune system, realizing rapid positioning and tracing of quality abnormalities, and improving the accuracy and convergence speed of abnormal pattern recognition by fusing antibody clone selection, immune memory and chaotic disturbance mechanism. A fourth unit is configured to activate an intelligent traceability engine according to the quality evaluation early warning signal, locate a quality abnormal root cause based on a space-time graph reasoning mechanism, and generate a traceability analysis result; and perform closed-loop feedback control according to the traceability analysis result, so as to optimize and adjust a raw material production process and realize dynamic management and control of a quality state.

7. An electronic device, comprising: Comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 5.

Images