Semiconductor equipment fault prediction and maintenance method and system based on data fusion

Abstract

The present invention provides a semiconductor equipment fault prediction and maintenance method and system based on data fusion, which relates to the technical field of semiconductor equipment fault prediction and maintenance, including collecting multi-source sensor data, building a process-stress-material multidimensional state model, and calculating the failure probability of key components. When the failure probability exceeds the threshold, the fault diagnosis process is triggered, fault mode identification is performed based on the digital twin model, and the fault development trend is predicted in combination with historical data. Finally, a maintenance objective function is constructed, and the optimal maintenance plan is determined using an adaptive optimization algorithm and a maintenance decision rule base, and maintenance guidance is performed. The present invention realizes continuous optimization of fault prediction and maintenance capabilities by constructing a maintenance knowledge graph, thereby improving the reliability and production efficiency of semiconductor equipment.

Description

Technical Field

[0001] The present invention relates to semiconductor equipment fault prediction and maintenance technology, and in particular to a semiconductor equipment fault prediction and maintenance method and system based on data fusion. Background Art

[0002] Semiconductor manufacturing equipment plays a vital role in modern integrated circuit production. With the increasing complexity of production processes and the continuous expansion of production scale, the impact of equipment failure on production efficiency, product quality and enterprise cost control has become more and more significant. Therefore, how to improve equipment reliability, reduce the occurrence of failures and perform maintenance in a timely manner has become a major challenge facing the semiconductor manufacturing industry.

[0003] At present, traditional equipment fault detection methods mostly rely on regular manual inspections and empirical judgments, which cannot fully monitor the status changes of equipment in real time, and often cannot predict equipment failures in advance. These traditional methods have problems such as slow response time, low prediction accuracy, and poor maintenance plan execution efficiency, which are difficult to meet the needs of efficient, accurate, and intelligent equipment management in modern semiconductor production environments.

[0004] Therefore, there is an urgent need for a new method to achieve accurate prediction and efficient maintenance of semiconductor equipment failures, thereby reducing the frequency of failures and improving equipment reliability and production efficiency. Summary of the invention

[0005] The embodiments of the present invention provide a semiconductor equipment fault prediction and maintenance method and system based on data fusion, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A semiconductor equipment fault prediction and maintenance method based on data fusion is provided, comprising:

[0008] Collect vibration sensor data, temperature sensor data, pressure sensor data and current sensor data of semiconductor equipment, construct a feature space through dimensionality reduction mapping, establish a process-stress-material multidimensional state model in the feature space, calculate process parameter drift values, stress distribution values ​​and material degradation values ​​based on the process-stress-material multidimensional state model, establish a device state mapping relationship according to the process parameter drift values, stress distribution values ​​and material degradation values, calculate the failure probability of key components of the semiconductor equipment, and trigger a fault diagnosis process when the failure probability exceeds a preset threshold;

[0009] After the fault diagnosis process is triggered, the target monitoring part is determined according to the equipment state mapping relationship, the multi-dimensional monitoring data of the target monitoring part is collected, the feature extraction of the multi-dimensional monitoring data is performed to obtain the comprehensive fault feature, the comprehensive fault feature is input into the pre-trained digital twin model to perform fault mode recognition to obtain the fault diagnosis result, a fault prediction model is constructed based on the fault diagnosis result, and the historical fault evolution data is input to train the fault prediction model to obtain the fault development trend;

[0010] A maintenance objective function is constructed according to the fault diagnosis result and the fault development trend, and an adaptive optimization algorithm is used to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set. The optimal solution set is input into a preset maintenance decision rule library for comprehensive evaluation, and the solution with the highest comprehensive evaluation score is selected as the optimal maintenance plan. The optimal maintenance plan is converted into operation instructions for maintenance guidance, and maintenance process data and equipment status data before and after maintenance are collected. A maintenance knowledge graph is constructed according to the maintenance process data and equipment status data, and a fault prediction model and a maintenance decision rule library are updated based on the maintenance knowledge graph to achieve continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0011] In an optional embodiment,

[0012] Calculating process parameter drift values, stress distribution values, and material degradation values ​​based on the process-stress-material multidimensional state model, establishing a device state mapping relationship according to the process parameter drift values, stress distribution values, and material degradation values, calculating the failure probability of key components of the semiconductor device, and triggering a fault diagnosis process when the failure probability exceeds a preset threshold includes:

[0013] Constructing a three-dimensional state tensor including process parameters, stress distribution, and material properties in the feature space, wherein the first dimension of the three-dimensional state tensor corresponds to the process parameters, the second dimension corresponds to the stress distribution, and the third dimension corresponds to the material properties, and uniformly sampling the three-dimensional state tensor according to the time dimension to obtain a state sequence of multiple time nodes;

[0014] Perform multi-layer decomposition on the three-dimensional state tensor of each time node in the state sequence of multiple time nodes, and extract the characteristic data matrix of process parameters, the characteristic data matrix of stress distribution, and the characteristic data matrix of material properties respectively;

[0015] Perform kernel function mapping on the characteristic data matrix of process parameters and the preset reference process state, and calculate the process parameter drift value of the characteristic data matrix of process parameters relative to the reference process state;

[0016] Establishing a probability transmission mapping of the characteristic data matrix of stress distribution between adjacent time nodes, and calculating the stress distribution values ​​of the characteristic data matrix of stress distribution at adjacent time nodes;

[0017] Establishing a state evolution equation for a characteristic data matrix of material properties through a probability density function, and solving the state evolution equation to obtain a material property degradation value of the material property;

[0018] Building a dynamic probability graph model based on the process parameter drift value, stress distribution value and material property degradation value, wherein the dynamic probability graph model includes state transition probabilities between nodes and association probabilities between nodes;

[0019] The state probability distribution of the semiconductor device at a future moment is calculated according to the dynamic probability graph model, and the failure probability is calculated based on the state probability distribution. When the failure probability exceeds a preset threshold, the fault diagnosis process is triggered.

[0020] In an optional embodiment,

[0021] Calculating the state probability distribution of the semiconductor device at a future time according to the dynamic probability graph model, and calculating the failure probability based on the state probability distribution includes:

[0022] Combining process parameter drift values, stress distribution values, and material property degradation values ​​of semiconductor devices into state vectors, arranging the state vectors in chronological order to form a state vector sequence, and establishing a time-series dependency relationship of the state vector sequence in a dynamic probability graph model;

[0023] Constructing a conditional variational posterior distribution network, wherein the conditional variational posterior distribution network comprises an encoder network and a decoder network, wherein the encoder network maps the state vector sequence to a latent variable space to obtain a latent variable representation, and the decoder network reconstructs the latent variable representation into a state vector;

[0024] Constructing a recursive neural network including a long short-term memory unit and an attention mechanism unit, inputting the latent variable representation into the recursive neural network, and outputting a mean vector and a covariance matrix of a conditional variational posterior distribution;

[0025] Calculate the relative entropy between the conditional variational posterior distribution and the true posterior distribution as the first loss function, calculate the mean square error between the reconstructed state vector and the original state vector as the second loss function, construct the weighted sum of the first loss function and the second loss function as the total loss function, and optimize the total loss function using a stochastic gradient descent algorithm with a momentum term;

[0026] The current state vector is input into the optimized recursive neural network, and the state probability distribution representing the process parameters, stress distribution and material properties at the future time is output;

[0027] Constructing a multidimensional mixed Gaussian distribution according to the statistical characteristics of the state probability distribution, extracting state samples from the multidimensional mixed Gaussian distribution, calculating the probability density ratio of each state sample under the state probability distribution and the multidimensional mixed Gaussian distribution, and obtaining the importance weight of each state sample;

[0028] The state samples and the corresponding importance weights are input into a failure probability estimator built based on a deep neural network. The failure probability estimator maps the state samples to a failure probability space through a multi-layer perceptron network to obtain the failure probability of the semiconductor device.

[0029] In an optional embodiment,

[0030] Inputting the comprehensive fault features into the pre-trained digital twin model to perform fault mode recognition to obtain fault diagnosis results including:

[0031] Inputting historical operation data of actual equipment into a physical knowledge guidance network, wherein the physical knowledge guidance network includes a physical constraint layer and a data driving layer, wherein the physical constraint layer uses the equipment dynamics equation to construct a loss function, and the data driving layer uses a time series feature extraction module to extract data association features, and inputting the loss function and the data association features into a digital twin model training unit to obtain a pre-trained digital twin model;

[0032] The comprehensive fault features are input into a pre-trained digital twin model to output prediction features, and at the same time, an improved deep residual network is used to extract features from real-time monitoring data, wherein the deep residual network includes multiple residual learning units, and the residual learning unit uses deep separable convolution to extract local features to obtain real-time fault features;

[0033] Adaptively fusing the prediction features with the real-time fault features, calculating the correlation weights of the two types of features through a cross-attention mechanism, and fusing the features based on the correlation weights to obtain enhanced fault features;

[0034] Constructing a working condition perception module, inputting working condition information into a working condition encoder to obtain a working condition encoding vector, generating dynamic convolution kernel parameters based on the working condition encoding vector, and using the dynamic convolution kernel parameters to perform an adaptive convolution operation on the enhanced fault feature to obtain a working condition adaptation feature;

[0035] Inputting the working condition adaptation feature into a multi-task learning network, wherein the multi-task learning network includes a fault classification branch and a fault degree assessment branch, wherein the fault classification branch outputs a fault type probability distribution, and the fault degree assessment branch outputs a fault degree score, to obtain a preliminary diagnosis result;

[0036] A Bayesian neural network is used to model the uncertainty of the preliminary diagnosis result, and a prediction distribution is obtained through multiple random samplings. A confidence interval is calculated based on the prediction distribution, and a diagnosis result with a confidence interval higher than a preset confidence threshold is taken as the final fault diagnosis result.

[0037] In an optional embodiment,

[0038] Constructing a maintenance objective function according to the fault diagnosis result and the fault development trend, performing multi-objective optimization on the maintenance objective function using an adaptive optimization algorithm to obtain an optimal solution set, inputting the optimal solution set into a preset maintenance decision rule library for comprehensive evaluation, and selecting the solution with the highest comprehensive evaluation score as the optimal maintenance plan includes:

[0039] Based on the fault degree score and fault development trend prediction result in the fault diagnosis result, an exponential function is used to calculate the fault rate function value, and the equipment reliability objective function is constructed based on the fault rate function value;

[0040] Obtain the preventive maintenance times and failure maintenance times of semiconductor equipment, calculate the total maintenance cost by combining the preventive maintenance cost coefficient and the failure maintenance cost coefficient, and construct a maintenance cost objective function;

[0041] Obtaining a mean time between failures and a mean time to repair of a semiconductor device, calculating a device availability value according to a ratio of the mean time between failures and the mean time to repair, and constructing a device availability objective function;

[0042] A maintenance objective function is constructed by combining the equipment reliability objective function, the maintenance cost objective function and the equipment availability objective function, and the maintenance objective function is input into an adaptive optimization model. The adaptive optimization model adaptively adjusts a crossover probability parameter and a mutation probability parameter according to the Euclidean distance value from the population individual to the optimal frontier, and obtains a non-dominated solution set through non-dominated sorting iterative calculation;

[0043] Partitioning the solution space of the non-dominated solution set according to the fault type code, calculating the grid boundary threshold based on the density distribution of the solution in the partition, and adaptively meshing each subspace according to the grid boundary threshold to obtain a grid partition solution set;

[0044] Constructing a maintenance decision rule base, calculating the reliability membership value, cost membership value and availability membership value corresponding to each solution in the grid partition solution set, and calculating the comprehensive evaluation score in combination with the corresponding indicator weight coefficient;

[0045] A state deviation value is calculated according to real-time monitoring state data of semiconductor equipment, and the indicator weight coefficient and membership function parameter are adjusted based on the state deviation value to obtain a revised comprehensive evaluation value, and a non-dominated solution with the largest revised comprehensive evaluation value is selected as the optimal maintenance plan.

[0046] In an optional embodiment,

[0047] Convert the optimal maintenance plan into operation instructions for maintenance guidance, collect maintenance process data and equipment status data before and after maintenance, build a maintenance knowledge graph based on the maintenance process data and equipment status data, update the fault prediction model and maintenance decision rule base based on the maintenance knowledge graph, and realize the continuous optimization of semiconductor equipment fault prediction and maintenance capabilities, including:

[0048] The maintenance plan is converted into a maintenance operation instruction sequence, wherein the maintenance operation instruction sequence includes operation part information, operation type information and operation parameter information, and a spatial positioning algorithm is used to track the maintenance scene in real time, and the maintenance operation instruction sequence is displayed on the equipment body through three-dimensional holographic projection;

[0049] Acquire operation video stream information of the maintenance process, extract the operation action sequence from the operation video stream information based on a human posture estimation algorithm, collect maintenance process data including operation tool usage information, operation duration information, operation sequence information, and equipment sensor data and semiconductor equipment component image data before and after maintenance as equipment status data;

[0050] Establish a maintenance knowledge ontology model, the maintenance knowledge ontology model includes equipment entity information, fault type information, maintenance operation information, and status feature information; extract knowledge triples from the maintenance process data and equipment status data according to preset rules and statistical methods; integrate the knowledge triples into a maintenance knowledge graph through entity alignment and relational reasoning;

[0051] Vectorizing the entity information and relationship information in the maintenance knowledge graph to obtain a feature vector; inputting the feature vector into a graph neural network, optimizing the parameters of the fault prediction model, and obtaining an optimized fault prediction model;

[0052] The maintenance decision rule base is updated by using a reinforcement learning method, and the maintenance effect information is used as a reward signal. A strategy search is performed in the maintenance knowledge graph based on a Monte Carlo tree search algorithm to generate an updated maintenance decision rule base.

[0053] Fault prediction of equipment is performed using the optimized fault prediction model to obtain prediction results; a new maintenance plan is determined from the updated maintenance decision rule library based on the prediction results, thereby achieving continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0054] In an optional embodiment,

[0055] The maintenance decision rule base is updated by using a reinforcement learning method, and the maintenance effect information is used as a reward signal. A strategy search is performed in the maintenance knowledge graph based on a Monte Carlo tree search algorithm to generate an updated maintenance decision rule base including:

[0056] Construct a Markov decision process model for maintenance decision-making, construct the equipment state vector and historical maintenance operation sequence as the state space, construct the optional maintenance operations as the action space, and obtain the state transition probability based on relational reasoning in the knowledge graph;

[0057] Obtaining equipment status indicator information and maintenance cost information before and after maintenance, and calculating an immediate reward value based on the improvement degree of the equipment status indicator information and the maintenance cost; obtaining equipment operation time information of the next maintenance cycle, and calculating a delayed reward value based on the equipment operation time information; combining the immediate reward value with the delayed reward value to obtain a reward signal;

[0058] Establish a search tree in the maintenance knowledge graph, use the maintenance status information as the tree node information, and use the maintenance operation information as the connection edge information; select the node to be expanded based on the confidence upper limit criterion, generate child nodes through the relationship information of the maintenance knowledge graph, and predict the maintenance operation effect through knowledge graph reasoning, and update the access count information and reward value information of each node;

[0059] Extracting historical maintenance case information from the knowledge graph and establishing a heuristic function for allocating prior probabilities of maintenance operations; using the heuristic function for node selection in the search tree;

[0060] The optimal maintenance strategy sequence obtained by the search is converted into a rule form containing trigger conditions and maintenance operation sequences, rule conflict detection and conflict resolution are performed, and confidence levels are assigned to the rules based on historical application results.

[0061] Collect maintenance effect data, and update rule confidence information according to the maintenance effect data; delete rules whose confidence is lower than a preset threshold, and summarize the maintenance experience information to obtain new rules; add the new rules to the maintenance decision rule base to generate an updated maintenance decision rule base.

[0062] According to a second aspect of the embodiments of the present invention,

[0063] A semiconductor equipment fault prediction and maintenance system based on data fusion is provided, comprising:

[0064] The first unit is used to collect vibration sensor data, temperature sensor data, pressure sensor data and current sensor data of the semiconductor device, construct a feature space through dimensionality reduction mapping, establish a process-stress-material multidimensional state model in the feature space, calculate the process parameter drift value, stress distribution value and material degradation value based on the process-stress-material multidimensional state model, establish an equipment state mapping relationship according to the process parameter drift value, stress distribution value and material degradation value, calculate the failure probability of the key components of the semiconductor device, and trigger the fault diagnosis process when the failure probability exceeds a preset threshold;

[0065] The second unit is used to determine the target monitoring part according to the equipment state mapping relationship after the fault diagnosis process is triggered, collect multi-dimensional monitoring data of the target monitoring part, extract features of the multi-dimensional monitoring data to obtain comprehensive fault features, input the comprehensive fault features into a pre-trained digital twin model to perform fault mode recognition to obtain a fault diagnosis result, build a fault prediction model based on the fault diagnosis result, input historical fault evolution data to train the fault prediction model to obtain a fault development trend;

[0066] The third unit is used to construct a maintenance objective function according to the fault diagnosis result and the fault development trend, use an adaptive optimization algorithm to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set, input the optimal solution set into a preset maintenance decision rule library for comprehensive evaluation, select the solution with the highest comprehensive evaluation score as the optimal maintenance plan, convert the optimal maintenance plan into operation instructions for maintenance guidance, collect maintenance process data and equipment status data before and after maintenance, construct a maintenance knowledge graph according to the maintenance process data and equipment status data, update the fault prediction model and the maintenance decision rule library based on the maintenance knowledge graph, and realize continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0067] According to a third aspect of the embodiments of the present invention,

[0068] An electronic device is provided, comprising:

[0069] processor;

[0070] a memory for storing processor-executable instructions;

[0071] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0072] A fourth aspect of the embodiments of the present invention is:

[0073] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0074] In this embodiment, by fusing multiple sensor data, building a multidimensional state model, and combining the digital twin model and fault evolution data for fault prediction, the equipment status can be more comprehensively reflected, and the accuracy and timeliness of fault prediction can be improved. Based on the fault diagnosis results and prediction trends, a maintenance objective function is constructed and multi-objective optimization is performed. Combined with the maintenance decision rule base for comprehensive evaluation, a more scientific and optimized maintenance plan can be formulated, maintenance efficiency can be improved, and maintenance costs can be reduced. By collecting maintenance process data and equipment status data, a maintenance knowledge graph is constructed, and it is used to update the fault prediction model and maintenance decision rule base, so as to achieve continuous optimization of fault prediction and maintenance capabilities and continuously improve the intelligence level of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A schematic diagram of a flow chart of a semiconductor equipment fault prediction and maintenance method based on data fusion according to an embodiment of the present invention;

[0076] Figure 2 It is a schematic diagram of the structure of a semiconductor equipment fault prediction and maintenance system based on data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0079] Figure 1 FIG. 1 is a flow chart of a semiconductor device fault prediction and maintenance method based on data fusion according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0080] S101. Collect vibration sensor data, temperature sensor data, pressure sensor data and current sensor data of semiconductor equipment, construct a feature space through dimensionality reduction mapping, establish a process-stress-material multidimensional state model in the feature space, calculate process parameter drift values, stress distribution values ​​and material degradation values ​​based on the process-stress-material multidimensional state model, establish a device state mapping relationship according to the process parameter drift values, stress distribution values ​​and material degradation values, calculate the failure probability of key components of the semiconductor equipment, and trigger a fault diagnosis process when the failure probability exceeds a preset threshold;

[0081] S102. After the fault diagnosis process is triggered, the target monitoring part is determined according to the equipment state mapping relationship, the multi-dimensional monitoring data of the target monitoring part is collected, the feature extraction of the multi-dimensional monitoring data is performed to obtain a comprehensive fault feature, the comprehensive fault feature is input into the pre-trained digital twin model to perform fault pattern recognition to obtain a fault diagnosis result, a fault prediction model is constructed based on the fault diagnosis result, and the historical fault evolution data is input to train the fault prediction model to obtain a fault development trend;

[0082] S103. Construct a maintenance objective function according to the fault diagnosis result and the fault development trend, use an adaptive optimization algorithm to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set, input the optimal solution set into a preset maintenance decision rule library for comprehensive evaluation, select the solution with the highest comprehensive evaluation score as the optimal maintenance plan, convert the optimal maintenance plan into operation instructions for maintenance guidance, collect maintenance process data and equipment status data before and after maintenance, construct a maintenance knowledge graph according to the maintenance process data and equipment status data, update the fault prediction model and the maintenance decision rule library based on the maintenance knowledge graph, and realize continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0083] Among them, vibration sensor data is the vibration information collected by the vibration sensor during the operation of the equipment, reflecting the operating status of the mechanical parts inside the equipment, possible wear and tear, and signs of failure. In semiconductor equipment, abnormal vibration may indicate mechanical failure or process instability. Temperature sensor data is the temperature information of each key component or specific area of ​​the equipment measured by the temperature sensor, which can help monitor whether the equipment has overheating, poor cooling and other phenomena during operation. These temperature anomalies are usually early signs of failure. Pressure sensor data reflects the pressure inside the semiconductor equipment or process area, and is mainly used to monitor key parameters such as gas flow and vacuum environment inside the equipment. Current sensor data is the equipment current information measured by the current sensor, reflecting the power operation of the equipment.

[0084] The process parameter drift value reflects the parameter changes during the equipment process; the stress distribution value characterizes the stress distribution of different components during the operation of the equipment; the material degradation value describes the degree of material aging of key components of the equipment over time. These data help establish a mapping relationship between the equipment status, which is used to calculate the failure probability of key components of the equipment and to judge possible faults in advance. When the failure probability exceeds the set threshold, the fault diagnosis process is automatically triggered. Among them, key components usually include transmission system components: such as motors, gears, bearings and belts, heating / cooling system components such as heaters, coolant pumps and temperature control systems, air pressure system components: such as air pumps, pressure valves, etc., power system components: such as power modules, current control devices, etc., responsible for power supply. Vacuum system components: such as vacuum pumps and related valves, as well as components that directly affect equipment stability, production efficiency and product quality. The failure of these components may cause equipment downtime, production interruption or quality problems.

[0085] The target monitoring location refers to the specific monitoring area or location selected according to the equipment status mapping relationship during the fault diagnosis process. These locations are usually high-risk areas for faults to occur, and collecting their data helps to accurately locate the fault.

[0086] The fault prediction model is a model trained on the basis of diagnosis by combining historical fault evolution data. It is used to predict the evolution trend of equipment faults and future failure time.

[0087] In an optional implementation, based on the process-stress-material multidimensional state model, a process parameter drift value, a stress distribution value, and a material degradation value are calculated, a device state mapping relationship is established according to the process parameter drift value, the stress distribution value, and the material degradation value, and a failure probability of a key component of the semiconductor device is calculated. When the failure probability exceeds a preset threshold, a fault diagnosis process is triggered, including:

[0088] Constructing a three-dimensional state tensor including process parameters, stress distribution, and material properties in the feature space, wherein the first dimension of the three-dimensional state tensor corresponds to the process parameters, the second dimension corresponds to the stress distribution, and the third dimension corresponds to the material properties, and uniformly sampling the three-dimensional state tensor according to the time dimension to obtain a state sequence of multiple time nodes;

[0089] Perform multi-layer decomposition on the three-dimensional state tensor of each time node in the state sequence of multiple time nodes, and extract the characteristic data matrix of process parameters, the characteristic data matrix of stress distribution, and the characteristic data matrix of material properties respectively;

[0090] Perform kernel function mapping on the characteristic data matrix of process parameters and the preset reference process state, and calculate the process parameter drift value of the characteristic data matrix of process parameters relative to the reference process state;

[0091] Establishing a probability transmission mapping of the characteristic data matrix of stress distribution between adjacent time nodes, and calculating the stress distribution values ​​of the characteristic data matrix of stress distribution at adjacent time nodes;

[0092] Establishing a state evolution equation for a characteristic data matrix of material properties through a probability density function, and solving the state evolution equation to obtain a material property degradation value of the material property;

[0093] Building a dynamic probability graph model based on the process parameter drift value, stress distribution value and material property degradation value, wherein the dynamic probability graph model includes state transition probabilities between nodes and association probabilities between nodes;

[0094] The state probability distribution of the semiconductor device at a future moment is calculated according to the dynamic probability graph model, and the failure probability is calculated based on the state probability distribution. When the failure probability exceeds a preset threshold, the fault diagnosis process is triggered.

[0095] Exemplarily, first, a three-dimensional state tensor is constructed to describe the state of the semiconductor device. The three dimensions of this tensor represent process parameters, stress distribution, and material properties. For example, process parameters may include temperature, pressure, deposition rate, etc.; stress distribution may include mechanical stress, thermal stress, electrical stress, etc.; material properties may include resistivity, dielectric constant, Young's modulus, etc. Assume that we monitor four parameters: temperature, pressure, mechanical stress, and resistivity.

[0096] Next, this three-dimensional state tensor is uniformly sampled according to the time dimension to obtain a state sequence of multiple time nodes. For example, if data is collected once every hour for 24 hours, a state sequence of 24 time nodes is obtained. Each time node corresponds to a three-dimensional state tensor. Assume that at a certain point in time, the temperature is 25 degrees Celsius, the pressure is 1 standard atmosphere, the mechanical stress is 10MPa, and the resistivity is 10 ohm cm.

[0097] Then, the three-dimensional state tensor of each time node is multi-layered decomposed to extract the characteristic data matrix of process parameters, the characteristic data matrix of stress distribution, and the characteristic data matrix of material properties. Multi-layer decomposition can be understood as decomposing a complex data structure into multiple simple and easy-to-analyze data structures. For example, the temperature and pressure data can be extracted to form the characteristic data matrix of process parameters, the mechanical stress and electrical stress data can be extracted to form the characteristic data matrix of stress distribution, and the resistivity data can be extracted to form the characteristic data matrix of material properties.

[0098] Next, the characteristic data matrix of the process parameters is compared with the preset reference process state to calculate the drift value of the process parameters. The reference process state can be understood as the ideal or standard process parameter value. For example, assume that the reference temperature is 20 degrees Celsius and the reference pressure is 0.9 standard atmospheres. Then, the current temperature of 25 degrees Celsius drifts 5 degrees Celsius relative to the reference temperature of 20 degrees Celsius, and the current pressure of 1 standard atmosphere drifts 0.1 standard atmosphere relative to the reference pressure of 0.9 standard atmospheres.

[0099] Then, the characteristic data matrix of stress distribution is compared between adjacent nodes in time to calculate the change value of stress distribution. For example, assuming that the mechanical stress at the previous time node is 9MPa and the mechanical stress at the current time node is 10MPa, the mechanical stress has increased by 1MPa.

[0100] Next, the characteristic data matrix of the material properties is analyzed to calculate the degradation value of the material properties by analyzing its change pattern over time. For example, if the resistivity at the previous time node is 9.5 ohm·cm and the resistivity at the current time node is 10 ohm·cm, then the resistivity has increased by 0.5 ohm·cm.

[0101] Based on the calculated process parameter drift values, stress distribution change values, and material property degradation values, a dynamic probability graphical model is constructed. This model describes how different parameters affect each other and how these parameters change over time. For example, an increase in temperature may lead to an increase in mechanical stress, which in turn leads to degradation of material properties.

[0102] According to the dynamic probability graph model, the state probability distribution of the semiconductor device at the future moment can be calculated, and the failure probability can be calculated based on this. Finally, if the calculated failure probability exceeds the preset threshold, such as 10%, the fault diagnosis process is triggered. The fault diagnosis process can include a series of inspections and tests to determine the specific cause of the equipment failure and take appropriate measures to repair it.

[0103] In this embodiment, the influence of three factors, namely process parameters, stress distribution and material properties, on the equipment status is comprehensively considered, and the failure probability of the equipment can be predicted more accurately. By real-time monitoring and prediction of the equipment status, an early warning can be issued before the equipment fails, thereby avoiding major losses. By timely fault diagnosis and repair, the reliability and service life of the equipment can be improved.

[0104] In an optional implementation, calculating the state probability distribution of the semiconductor device at a future time according to the dynamic probability graph model, and calculating the failure probability based on the state probability distribution includes:

[0105] Combining process parameter drift values, stress distribution values, and material property degradation values ​​of semiconductor devices into state vectors, arranging the state vectors in chronological order to form a state vector sequence, and establishing a time-series dependency relationship of the state vector sequence in a dynamic probability graph model;

[0106] Constructing a conditional variational posterior distribution network, wherein the conditional variational posterior distribution network comprises an encoder network and a decoder network, wherein the encoder network maps the state vector sequence to a latent variable space to obtain a latent variable representation, and the decoder network reconstructs the latent variable representation into a state vector;

[0107] Constructing a recursive neural network including a long short-term memory unit and an attention mechanism unit, inputting the latent variable representation into the recursive neural network, and outputting a mean vector and a covariance matrix of a conditional variational posterior distribution;

[0108] Calculate the relative entropy between the conditional variational posterior distribution and the true posterior distribution as the first loss function, calculate the mean square error between the reconstructed state vector and the original state vector as the second loss function, construct the weighted sum of the first loss function and the second loss function as the total loss function, and optimize the total loss function using a stochastic gradient descent algorithm with a momentum term;

[0109] The current state vector is input into the optimized recursive neural network, and the state probability distribution representing the process parameters, stress distribution and material properties at the future time is output;

[0110] Constructing a multidimensional mixed Gaussian distribution according to the statistical characteristics of the state probability distribution, extracting state samples from the multidimensional mixed Gaussian distribution, calculating the probability density ratio of each state sample under the state probability distribution and the multidimensional mixed Gaussian distribution, and obtaining the importance weight of each state sample;

[0111] The state samples and the corresponding importance weights are input into a failure probability estimator built based on a deep neural network. The failure probability estimator maps the state samples to a failure probability space through a multi-layer perceptron network to obtain the failure probability of the semiconductor device.

[0112] Exemplarily, first, various data of the semiconductor device during operation are collected, including process parameter drift values, stress distribution values, and material property degradation values. For example, the process parameter drift value may be the gate oxide thickness deviation, the stress distribution value may be the chip operating temperature distribution, and the material property degradation value may be the transistor threshold voltage drift. These data are arranged in chronological order to form a state vector sequence. For example, data may be collected every hour to form a time series. Then, a time-series dependency relationship of the state vector sequence in the dynamic probability graph model is established. For example, it may be assumed that the state vector at the current moment is affected by the state vector at the previous moment.

[0113] Next, a conditional variational posterior distribution network is constructed. The network consists of an encoder network and a decoder network. The encoder network maps the state vector sequence to a low-dimensional latent variable space to obtain a latent variable representation. For example, a multi-layer fully connected neural network can be used as an encoder. The decoder network reconstructs the latent variable representation into a state vector. For example, a multi-layer fully connected neural network can also be used as a decoder.

[0114] Then, a recurrent neural network is constructed, which contains long short-term memory units and attention mechanism units. The latent variable representation obtained by the encoder is input into the recurrent neural network, and the mean vector and covariance matrix of the conditional variational posterior distribution are output. For example, the mean vector can be a vector of length 10, and the covariance matrix can be a 10x10 matrix.

[0115] In order to train the network, two loss functions are defined. The first loss function is the relative entropy between the conditional variational posterior distribution and the true posterior distribution, which is used to measure the accuracy of variational inference. The second loss function is the mean square error between the reconstructed state vector and the original state vector, which is used to measure the accuracy of reconstruction. The weighted sum of these two loss functions is used as the total loss function, and the total loss function is optimized using the stochastic gradient descent algorithm with momentum. For example, the weights of the two loss functions can be set to 0.5 and 0.5 respectively, and the Adam optimizer with a learning rate of 0.001 is used for training.

[0116] After training is completed, the current state vector is input into the optimized recursive neural network, and the state probability distribution representing the process parameters, stress distribution and material properties at the future time is output. For example, the state probability distribution of the next 24 hours can be predicted.

[0117] In order to estimate the failure probability, a multidimensional mixed Gaussian distribution is first constructed based on the statistical characteristics of the state probability distribution. For example, a Gaussian mixture model can be used to fit the state probability distribution. Then, state samples are extracted from the multidimensional mixed Gaussian distribution. For example, 1,000 samples can be extracted. The probability density ratio of each state sample under the state probability distribution and the multidimensional mixed Gaussian distribution is calculated to obtain the importance weight of each state sample.

[0118] Finally, the state samples and the corresponding importance weights are input into the failure probability estimator built based on the deep neural network. The estimator maps the state samples to the failure probability space through the multi-layer perceptron network to obtain the failure probability of the semiconductor device. For example, the failure probability can be defined as the probability that the device will fail in a certain period of time in the future.

[0119] In this embodiment, by combining the dynamic probability graph model and deep learning, the evolution law of the semiconductor device state can be captured more accurately, thereby improving the accuracy of failure probability prediction. Compared with the traditional Monte Carlo simulation method, this method can effectively reduce the number of samples required through the importance sampling technology, thereby reducing the computational cost. By learning the representation of the state vector in the latent variable space, the influence of different factors on device failure can be revealed, thereby enhancing the interpretability of the model.

[0120] In an optional implementation, inputting the comprehensive fault feature into a pre-trained digital twin model to perform fault mode recognition to obtain a fault diagnosis result includes:

[0121] Inputting historical operation data of actual equipment into a physical knowledge guidance network, wherein the physical knowledge guidance network includes a physical constraint layer and a data driving layer, wherein the physical constraint layer uses the equipment dynamics equation to construct a loss function, and the data driving layer uses a time series feature extraction module to extract data association features, and inputting the loss function and the data association features into a digital twin model training unit to obtain a pre-trained digital twin model;

[0122] The comprehensive fault features are input into a pre-trained digital twin model to output prediction features, and at the same time, an improved deep residual network is used to extract features from real-time monitoring data, wherein the deep residual network includes multiple residual learning units, and the residual learning unit uses deep separable convolution to extract local features to obtain real-time fault features;

[0123] Adaptively fusing the prediction features with the real-time fault features, calculating the correlation weights of the two types of features through a cross-attention mechanism, and fusing the features based on the correlation weights to obtain enhanced fault features;

[0124] Constructing a working condition perception module, inputting working condition information into a working condition encoder to obtain a working condition encoding vector, generating dynamic convolution kernel parameters based on the working condition encoding vector, and using the dynamic convolution kernel parameters to perform an adaptive convolution operation on the enhanced fault feature to obtain a working condition adaptation feature;

[0125] Inputting the working condition adaptation feature into a multi-task learning network, wherein the multi-task learning network includes a fault classification branch and a fault degree assessment branch, wherein the fault classification branch outputs a fault type probability distribution, and the fault degree assessment branch outputs a fault degree score, to obtain a preliminary diagnosis result;

[0126] A Bayesian neural network is used to model the uncertainty of the preliminary diagnosis result, and a prediction distribution is obtained through multiple random samplings. A confidence interval is calculated based on the prediction distribution, and a diagnosis result with a confidence interval higher than a preset confidence threshold is taken as the final fault diagnosis result.

[0127] Exemplarily, first, in the stage of pre-training the digital twin model, the historical operation data of the device is input into the physical knowledge guidance network. The physical knowledge guidance network consists of a physical constraint layer and a data drive layer, which is used to pre-train the model based on the dynamic principle and timing correlation characteristics of the device. The physical constraint layer constructs a loss function through the device dynamics equation, and introduces the physical laws of the actual operation of the device into the model training. In this embodiment, taking the mechanical vibration data of semiconductor equipment as an example, based on the vibration transmission path and mechanical dynamics equation of the equipment, the loss function constructed in the physical constraint layer can constrain the vibration data output by the model to ensure that the prediction results of the model meet the physical characteristics of the equipment. Then, the data drive layer mines the timing features in the historical data through the timing feature extraction module to extract the correlation features during the operation of the equipment. For example, key timing features are extracted from the correlation changes between vibration, temperature and current signals. The physical constraint layer and the data drive layer work together to integrate the physical properties and data correlation information of the device into the digital twin model, thereby obtaining a pre-trained model guided by physical knowledge.

[0128] After obtaining the pre-trained digital twin model, the current comprehensive fault characteristics of the device are input into it to output prediction characteristics to help identify possible fault modes of the device. In this embodiment, the comprehensive fault characteristics can be composed of vibration, temperature, pressure and current data monitored in real time. After input, the pre-trained model will analyze these characteristics based on the knowledge of historical operating data and output prediction characteristics to characterize the current possible operating status and fault mode of the device.

[0129] At the same time, in order to more accurately reflect the real-time status of the equipment, the system uses an improved deep residual network to process the real-time monitoring data. The deep residual network consists of multiple residual learning units, each of which extracts local features of real-time data based on deep separable convolution. In this embodiment, the improved deep residual network can be used to perform convolution operations on real-time vibration signals to extract local abnormal features in the vibration signals and generate real-time fault features. Deep separable convolution operations can effectively reduce computational complexity and make real-time data processing more efficient.

[0130] Afterwards, the prediction features and real-time fault features are fused. In this process, the system uses a cross-attention mechanism to calculate the correlation weights of the prediction features and the real-time fault features. The cross-attention mechanism assigns weights by analyzing the correlation between the two types of features to ensure that the features most relevant to the fault receive higher weights. After combining the weights, the system adaptively fuses the two types of features to obtain enhanced fault features. In this embodiment, if the correlation weights of the vibration signal and the temperature signal in the prediction features are high, and the vibration signal in the real-time fault features fluctuates significantly, the cross-attention mechanism will increase the weight of the vibration signal, so that it occupies a larger proportion in the enhanced fault features after fusion, and enhance the fault indication capability of the features.

[0131] Furthermore, in order to adapt to the characteristic differences of the equipment under different working conditions, the system introduces a working condition perception module to adaptively adjust the enhanced fault characteristics. First, according to the current working condition information of the equipment, it is input into the working condition encoder to generate a working condition coding vector. In this embodiment, the working condition information may include working condition parameters such as the current operating temperature and pressure of the equipment. The working condition encoder encodes these parameters into a working condition coding vector, which represents the operating conditions of the current equipment. Then, the system generates dynamic convolution kernel parameters based on the working condition coding vector, so that the convolution kernel parameters can adjust the convolution characteristics according to the current working conditions. The enhanced fault characteristics are convolved using the dynamic convolution kernel parameters to obtain working condition adaptation characteristics. In this embodiment, if the equipment is operating under high temperature conditions, the dynamic convolution kernel parameters will pay more attention to the impact of high temperature on the fault characteristics, thereby enhancing the adaptability of the fault characteristics to high temperature conditions.

[0132] Next, the operating condition adaptation feature is input into the multi-task learning network for further analysis. The multi-task learning network includes a fault classification branch and a fault severity assessment branch. The fault classification branch is used to output the probability distribution of the fault type and determine the possible fault type of the current equipment; the fault severity assessment branch is used to score the severity of the fault. In this embodiment, the fault classification branch identifies the possible fault types of the equipment according to the operating condition adaptation feature, such as material aging failure caused by high temperature, and outputs the probability distribution of the fault type. The fault severity assessment branch evaluates the severity of the fault based on the operating condition adaptation feature, such as determining the degree of aging failure and outputting the corresponding score. The results of these two branches together constitute the preliminary diagnosis results.

[0133] In order to improve the reliability of the diagnosis results, the system models the uncertainty of the preliminary diagnosis results through a Bayesian neural network. The Bayesian neural network can generate a predicted distribution through multiple random sampling, thereby calculating the confidence interval of the diagnosis result. In this embodiment, the system can perform multiple sampling of the aging failure probability of the equipment through a Bayesian neural network based on the preliminary diagnosis results to obtain the distribution of the probability, and calculate the confidence interval based on the distribution. If the confidence interval is higher than the set confidence threshold, the result is output as the final fault diagnosis result. The final fault diagnosis result provides a reliable basis for the maintenance and management of semiconductor equipment.

[0134] In this embodiment, the physical knowledge is used to guide the network and integrate the physical constraints of the equipment into the model training, so that the diagnostic model is more consistent with the physical characteristics of the actual equipment and the accuracy of fault identification is improved. The real-time fault features are extracted using a deep residual network and combined with the predictive features of the pre-trained model. Adaptive fusion is performed through a cross-attention mechanism, so that the system can dynamically respond to changes in fault features and improve the real-time and sensitivity of diagnosis. The operating condition perception module is introduced to adaptively adjust the enhanced fault features to ensure that the system can maintain high-precision fault detection under different operating conditions, thereby improving the system's operating condition adaptability. The uncertainty of the diagnostic results is modeled through a Bayesian neural network, and the confidence interval is calculated to ensure the reliability of the diagnostic results and provide a highly credible basis for maintenance decisions.

[0135] In an optional implementation, a maintenance objective function is constructed according to the fault diagnosis result and the fault development trend, an adaptive optimization algorithm is used to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set, the optimal solution set is input into a preset maintenance decision rule library for comprehensive evaluation, and the solution with the highest comprehensive evaluation score is selected as the optimal maintenance plan, which includes:

[0136] Based on the fault degree score and fault development trend prediction result in the fault diagnosis result, an exponential function is used to calculate the fault rate function value, and the equipment reliability objective function is constructed based on the fault rate function value;

[0137] Obtain the preventive maintenance times and failure maintenance times of semiconductor equipment, calculate the total maintenance cost by combining the preventive maintenance cost coefficient and the failure maintenance cost coefficient, and construct a maintenance cost objective function;

[0138] Obtaining a mean time between failures and a mean time to repair of a semiconductor device, calculating a device availability value according to a ratio of the mean time between failures and the mean time to repair, and constructing a device availability objective function;

[0139] A maintenance objective function is constructed by combining the equipment reliability objective function, the maintenance cost objective function and the equipment availability objective function, and the maintenance objective function is input into an adaptive optimization model. The adaptive optimization model adaptively adjusts a crossover probability parameter and a mutation probability parameter according to the Euclidean distance value from the population individual to the optimal frontier, and obtains a non-dominated solution set through non-dominated sorting iterative calculation;

[0140] Partitioning the solution space of the non-dominated solution set according to the fault type code, calculating the grid boundary threshold based on the density distribution of the solution in the partition, and adaptively meshing each subspace according to the grid boundary threshold to obtain a grid partition solution set;

[0141] Constructing a maintenance decision rule base, calculating the reliability membership value, cost membership value and availability membership value corresponding to each solution in the grid partition solution set, and calculating the comprehensive evaluation score in combination with the corresponding indicator weight coefficient;

[0142] A state deviation value is calculated according to real-time monitoring state data of semiconductor equipment, and the indicator weight coefficient and membership function parameter are adjusted based on the state deviation value to obtain a revised comprehensive evaluation value, and a non-dominated solution with the largest revised comprehensive evaluation value is selected as the optimal maintenance plan.

[0143] For example, first, based on the fault diagnosis results of the equipment, the severity of the fault is evaluated, and combined with the predicted fault development trend, the probability of the fault occurring is calculated. The higher the probability of the fault occurring, the larger the corresponding function value. For example, based on expert experience, the fault severity score is 8, and the predicted probability of the fault occurring within the next week is 0.1, then the fault rate function value can be set to a higher value, such as 0.8.

[0144] Secondly, in order to evaluate the maintenance cost, it is necessary to obtain the number of preventive maintenance and breakdown maintenance of the equipment. For example, in the past year, a certain equipment has undergone 4 preventive maintenance and 2 breakdown maintenance. Multiply these times by the corresponding cost coefficients, for example, the preventive maintenance cost coefficient is 1,000 yuan / time, and the breakdown maintenance cost coefficient is 5,000 yuan / time, and the total maintenance cost can be calculated: (4*1000)+(2*5000)=14,000 yuan.

[0145] Then, in order to evaluate the availability of the equipment, it is necessary to obtain the mean time between failures and the mean time to repair of the equipment. For example, the mean time between failures of a certain equipment is 1000 hours and the mean time to repair is 10 hours. Divide the mean time between failures by the sum of the mean time between failures and the mean time to repair to calculate the equipment availability value: 1000 / (1000+10)≈0.99.

[0146] Next, the previously calculated failure rate function values, total maintenance costs, and equipment availability values ​​are integrated together to construct a comprehensive maintenance objective function. The goal of this function is to minimize the failure rate and maintenance costs while maximizing equipment availability.

[0147] The maintenance objective function is then input into an adaptive optimization model. The model automatically adjusts parameters based on the distance between the current solution and the optimal solution set to improve search efficiency. For example, if the current solution is far from the optimal solution set, the model will increase the amplitude of parameter adjustment to explore the solution space faster; otherwise, it will reduce the amplitude of adjustment to search for the optimal solution more finely. Through multiple iterative calculations, the model will eventually obtain a set of non-dominated solutions that achieve the best balance between failure rate, maintenance cost, and equipment availability.

[0148] In order to further screen the optimal solution, the non-dominated solution set needs to be divided. The solution set is divided into multiple subspaces according to the fault type. For example, the solution set can be divided into three subspaces according to whether the fault is a mechanical fault, an electrical fault, or a software fault. Then, the grid is further divided according to the distribution density of the solutions in each subspace. For example, if the density of solutions in a subspace is high, the subspace can be divided into a finer grid.

[0149] Build a maintenance decision rule base containing multiple rules. For example, "If the failure rate is lower than threshold A and the availability is higher than threshold B, then choose preventive maintenance", "If the failure rate is higher than threshold C, then choose failure maintenance", etc. Then, based on these rules, conduct a comprehensive evaluation of each solution, calculate the reliability membership value, cost membership value, and availability membership value of each solution, and calculate the comprehensive evaluation score in combination with the pre-set weight coefficient.

[0150] Finally, according to the real-time monitoring data of the equipment, the indicator weight coefficient and the membership function parameters are adjusted. For example, if the real-time state deviation of the equipment is large, the weight coefficient of the reliability indicator can be increased and the membership function parameters can be adjusted to pay more attention to the reliability of the equipment. Finally, the solution with the highest comprehensive evaluation score after correction is selected as the optimal maintenance plan.

[0151] In this embodiment, a comprehensive maintenance objective function is constructed based on the fault diagnosis results, fault development trends and real-time equipment status data, and an adaptive optimization algorithm is used to solve it, avoiding the subjectivity and one-sidedness of human experience and improving the scientificity and accuracy of maintenance decisions. By optimizing the maintenance strategy, the maintenance cost is minimized while ensuring the reliability and availability of the equipment. By balancing the number of preventive maintenance and fault maintenance, over-maintenance or under-maintenance can be effectively avoided, thereby reducing the total maintenance cost. By comprehensively considering the failure rate, maintenance cost and equipment availability, and adjusting according to the real-time status of the equipment, the equipment failure rate can be effectively reduced, the average time between failures of the equipment can be increased, and the average repair time can be shortened, thereby improving the reliability and availability of the equipment.

[0152] In an optional implementation, the optimal maintenance plan is converted into an operation instruction for maintenance guidance, maintenance process data and equipment status data before and after maintenance are collected, a maintenance knowledge graph is constructed according to the maintenance process data and equipment status data, and a fault prediction model and a maintenance decision rule base are updated based on the maintenance knowledge graph to achieve continuous optimization of semiconductor equipment fault prediction and maintenance capabilities, including:

[0153] The maintenance plan is converted into a maintenance operation instruction sequence, wherein the maintenance operation instruction sequence includes operation part information, operation type information and operation parameter information, and a spatial positioning algorithm is used to track the maintenance scene in real time, and the maintenance operation instruction sequence is displayed on the equipment body through three-dimensional holographic projection;

[0154] Acquire operation video stream information of the maintenance process, extract the operation action sequence from the operation video stream information based on a human posture estimation algorithm, collect maintenance process data including operation tool usage information, operation duration information, operation sequence information, and equipment sensor data and semiconductor equipment component image data before and after maintenance as equipment status data;

[0155] Establish a maintenance knowledge ontology model, the maintenance knowledge ontology model includes equipment entity information, fault type information, maintenance operation information, and status feature information; extract knowledge triples from the maintenance process data and equipment status data according to preset rules and statistical methods; integrate the knowledge triples into a maintenance knowledge graph through entity alignment and relational reasoning;

[0156] Vectorizing the entity information and relationship information in the maintenance knowledge graph to obtain a feature vector; inputting the feature vector into a graph neural network, optimizing the parameters of the fault prediction model, and obtaining an optimized fault prediction model;

[0157] The maintenance decision rule base is updated by using a reinforcement learning method, and the maintenance effect information is used as a reward signal. A strategy search is performed in the maintenance knowledge graph based on a Monte Carlo tree search algorithm to generate an updated maintenance decision rule base.

[0158] Fault prediction of equipment is performed using the optimized fault prediction model to obtain prediction results; a new maintenance plan is determined from the updated maintenance decision rule library based on the prediction results, thereby achieving continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0159] Exemplarily, first, the optimal maintenance plan is converted by the system into a sequence of operating instructions, which include information such as the operating part, operation type, and operating parameters. The operating part information indicates the equipment parts that need to be maintained, the operation type information includes operations such as cleaning, repair, and replacement, and the operation parameter information specifically describes the specific parameters required for each operation. For example, a semiconductor device may need to clean the cooling system, and the operating instructions will specify the specific location and cleaning method of the cooler in detail. In order to ensure the accuracy and efficiency of the operation, the system also uses a spatial positioning algorithm to track the maintenance scene in real time, and displays the operating instructions on the equipment body through three-dimensional holographic projection to guide operators to accurately perform maintenance tasks.

[0160] While the maintenance operation is being performed, the system will obtain the operation video stream information during the maintenance process. Through the human posture estimation algorithm, the operation action sequence is extracted from the video stream and the operator's action trajectory is recorded. This information includes the use of operating tools, operation duration and operation sequence, etc., which is a key part of the maintenance process data. At the same time, the sensor data of the equipment and the image data of the equipment components before and after maintenance are also collected and recorded. These equipment status data provide an important basis for subsequent maintenance effect evaluation and knowledge update.

[0161] In order to make full use of these data, the system established a maintenance knowledge ontology model. The model contains equipment entity information, fault type information, maintenance operation information and status feature information. Equipment entity information includes the various components of the equipment and their attributes, fault type information records the failure mode of the equipment in history, maintenance operation information includes all maintenance measures performed and their effects, and status feature information describes the characteristics of the equipment in different states. These data are extracted from maintenance process data and equipment status data through preset rules and statistical methods to obtain knowledge triples (i.e., entities, attributes, and relationships). These triples are integrated into a complete maintenance knowledge graph through entity alignment and relationship reasoning.

[0162] After the maintenance knowledge graph is constructed, the system vectorizes the entity information and relationship information therein to obtain feature vectors. These feature vectors represent the deep associations between equipment status, failure mode, and maintenance process. On this basis, the system inputs the feature vectors into the graph neural network and obtains the optimized fault prediction model by optimizing the parameters of the fault prediction model. Through this optimized model, the system can more accurately predict possible equipment failures and give corresponding maintenance suggestions in a timely manner.

[0163] In addition, the system also uses reinforcement learning methods to update the maintenance decision rule base. In this process, maintenance effect information is used as a reward signal, and the reinforcement learning model continuously adjusts the strategy based on the feedback during the maintenance process to improve the effectiveness of maintenance decisions. Specifically, maintenance effect information includes the equipment status after maintenance, operator feedback, and further development of the fault. Based on the Monte Carlo tree search algorithm, the system performs strategy search in the maintenance knowledge graph to generate an updated maintenance decision rule base.

[0164] Finally, the system uses the optimized fault prediction model to predict equipment failures and generates corresponding prediction results. Based on these prediction results, the system selects new maintenance plans from the updated maintenance decision rule base. In this way, by continuously optimizing the fault prediction model and maintenance decision rule base, the fault prediction and maintenance capabilities of semiconductor equipment can be continuously improved, realizing intelligent and dynamic maintenance within the life cycle of the equipment.

[0165] Taking a semiconductor production line as an example, when a certain type of vibration abnormality occurs in the equipment, the system identifies the possible fault type of the equipment by learning from historical fault data, and performs real-time monitoring based on the optimized fault prediction model. During the real-time monitoring process, the system guides maintenance personnel to operate the equipment through operating instructions, and records all data of the maintenance process, including operation time, use of operating tools, etc. These data will be further supplemented to the maintenance knowledge graph, and the fault prediction model and maintenance decision rule base will be optimized through the graph neural network, ultimately achieving intelligent prediction and precise maintenance, significantly improving production efficiency and equipment reliability.

[0166] In this embodiment, by directly displaying the operating instructions on the equipment in the form of three-dimensional holographic projection, the time for technicians to find information and understand instructions is reduced, maintenance efficiency is improved, human errors can be reduced, and the standardization and accuracy of maintenance operations can be ensured, thereby improving maintenance quality. Through the knowledge graph-driven fault prediction model and maintenance decision rule library, equipment failures can be predicted in advance, and the best maintenance plan can be formulated to avoid unexpected equipment downtime and reduce maintenance costs.

[0167] In an optional implementation, a maintenance decision rule base is updated by using a reinforcement learning method, maintenance effect information is used as a reward signal, and a strategy search is performed in the maintenance knowledge graph based on a Monte Carlo tree search algorithm. The updated maintenance decision rule base is generated, including:

[0168] Construct a Markov decision process model for maintenance decision-making, construct the equipment state vector and historical maintenance operation sequence as the state space, construct the optional maintenance operations as the action space, and obtain the state transition probability based on relational reasoning in the knowledge graph;

[0169] Obtaining equipment status indicator information and maintenance cost information before and after maintenance, and calculating an immediate reward value based on the improvement degree of the equipment status indicator information and the maintenance cost; obtaining equipment operation time information of the next maintenance cycle, and calculating a delayed reward value based on the equipment operation time information; combining the immediate reward value with the delayed reward value to obtain a reward signal;

[0170] Establish a search tree in the maintenance knowledge graph, use the maintenance status information as the tree node information, and use the maintenance operation information as the connection edge information; select the node to be expanded based on the confidence upper limit criterion, generate child nodes through the relationship information of the maintenance knowledge graph, and predict the maintenance operation effect through knowledge graph reasoning, and update the access count information and reward value information of each node;

[0171] Extracting historical maintenance case information from the knowledge graph and establishing a heuristic function for allocating prior probabilities of maintenance operations; using the heuristic function for node selection in the search tree;

[0172] The optimal maintenance strategy sequence obtained by the search is converted into a rule form containing trigger conditions and maintenance operation sequences, rule conflict detection and conflict resolution are performed, and confidence levels are assigned to the rules based on historical application results.

[0173] Collect maintenance effect data, and update rule confidence information according to the maintenance effect data; delete rules whose confidence is lower than a preset threshold, and summarize the maintenance experience information to obtain new rules; add the new rules to the maintenance decision rule base to generate an updated maintenance decision rule base.

[0174] Exemplarily, first, the system combines the equipment's status information with historical maintenance operation data to construct a Markov decision process model. The equipment status consists of multiple dimensions, such as temperature, pressure, vibration, etc. This information is collected in real time by sensors and forms a state vector of the equipment. For example, the current state of the equipment may be "temperature: 75°C, vibration: 0.5mm / s, operating time: 500 hours", and this information will be used as the state space in the model. At the same time, the system will define the action space of maintenance operations, such as "replace the cooling fan" or "clean the filter" based on the current state and historical records of the equipment. In this way, the system can select the most appropriate maintenance operation based on its state when an abnormality occurs in the equipment.

[0175] Next, the system will obtain the changes in the equipment status indicators before and after maintenance, as well as the cost information required for the maintenance operation. Specifically, the equipment status indicators include temperature, vibration, working hours, etc., which can reflect the effectiveness of the maintenance operation. For example, assuming that the temperature of the equipment drops from 75°C to 65°C and the vibration drops from 0.5mm / s to 0.2mm / s after the "clean cooling system" operation is performed, the system can calculate the immediate reward value of the operation. The reward value is calculated as the ratio of the degree of improvement in the equipment's status before and after the maintenance operation (such as temperature change, vibration reduction, etc.) to the operating cost. Assuming the cost is 500 yuan, the temperature improves by 10°C, and the vibration is reduced by 0.3mm / s, the immediate reward value can be evaluated by the ratio of the improvement degree to the cost, reflecting the benefit of the operation.

[0176] After the system obtains the immediate reward value, it also needs to calculate the delayed reward value. The delayed reward value is calculated based on the operating time information of the equipment's next maintenance cycle. For example, the system knows that after the equipment continues to run for 500 hours, it may experience problems such as increased temperature and increased vibration, causing damage to the equipment. The calculation of the delayed reward value takes into account the impact of the equipment's operating time on the risk of future failures. Therefore, by combining the immediate reward value with the delayed reward value, the system is able to form a comprehensive reward signal to ensure that maintenance decisions not only consider the current status, but also predict future risks.

[0177] The system then builds a search tree in the maintenance knowledge graph, where each node represents a device state and each edge represents a possible state transition after a maintenance operation is performed. Using the confidence upper bound criterion, the system selects nodes to be expanded in the maintenance knowledge graph. During the expansion process, the system uses the device's historical data and the reasoning function of the knowledge graph to predict the effects of different maintenance operations. For example, the system can predict that the "replace cooling fan" operation will reduce the device temperature by 10°C, while "cleaning the cooling system" may only reduce the temperature by 5°C. Through reasoning, the system will give priority to more effective operations.

[0178] During the reasoning process, the system also combines historical maintenance data and cases to assign a priority to each node of the search tree through a heuristic function. The heuristic function guides the system's decision-making process based on historical maintenance results. For example, if historical data shows that "Replace the cooling fan" can significantly reduce the temperature, while the "Clean the cooling system" operation has a poor effect, the heuristic function will assign a higher priority to the "Replace the cooling fan" operation.

[0179] Once the system obtains the optimal sequence of maintenance operations through reinforcement learning and policy search, it converts these operations into maintenance rules. Each rule contains the triggering conditions and execution steps of the operation. For example, a rule may include "When the device temperature exceeds 75°C, replace the cooling fan." The system then performs rule conflict detection to ensure that conflicting operations are not selected at the same time. Assuming that under the same device state, the two operations of "cleaning the cooling system" and "replacing the cooling fan" will conflict with each other, and the system will select the more appropriate operation based on the principle of optimality.

[0180] After executing the rule, the system will collect maintenance effect data and analyze the impact of the maintenance operation on the device status. Through the feedback mechanism, the system will adjust the confidence of the rule according to the maintenance effect. If an operation does not work as expected in actual application, the system will reduce the confidence of its rule. Conversely, if the operation works well, the confidence will increase. For example, if the "Replace cooling fan" operation successfully reduces the device temperature to within the normal range, the confidence of the related rule will increase.

[0181] The system will also summarize new rules to form an updated maintenance decision library. These new rules are based on actual maintenance experience. For example, if multiple maintenance cases show that "replacing the cooling fan" is an effective way to deal with high temperature problems, the system will generate similar rules and push them to other devices. While the system is constantly updating the maintenance rule library, inefficient rules in the rule library will be deleted to ensure that the rules in the library are always optimized.

[0182] Through the above implementation steps, the system can continuously optimize the maintenance decisions of the equipment and continuously improve the equipment failure prediction and maintenance capabilities. For example, when the equipment has abnormal vibration, the system will automatically select the most appropriate maintenance plan based on the current equipment status and historical data, such as "replace the vibration sensor" or "adjust the vibration source". These optimized maintenance decisions will be continuously adjusted and updated through the reinforcement learning model, thereby reducing the frequency of failures and extending the service life of the equipment.

[0183] In this embodiment, by optimizing the maintenance decision rule base through reinforcement learning, the system can continuously adjust the maintenance strategy according to the real-time status and historical data of the equipment, and improve the operating efficiency and reliability of the equipment. Through dynamically updated maintenance decision rules, the system can accurately predict equipment failures and reduce sudden failures and downtime. By collecting the status data of the equipment in real time and combining it with the maintenance history of the equipment, the system can warn before the failure occurs and take effective maintenance measures in time, thereby extending the service life of the equipment. In addition, through real-time tracking and optimization of the maintenance operation process, the system can continuously accumulate maintenance experience, gradually form an efficient maintenance strategy, improve fault diagnosis accuracy and maintenance efficiency, and achieve continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0184] Figure 2 FIG. 1 is a schematic diagram of the structure of a semiconductor device fault prediction and maintenance system based on data fusion according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0185] The first unit is used to collect vibration sensor data, temperature sensor data, pressure sensor data and current sensor data of the semiconductor device, construct a feature space through dimensionality reduction mapping, establish a process-stress-material multidimensional state model in the feature space, calculate the process parameter drift value, stress distribution value and material degradation value based on the process-stress-material multidimensional state model, establish an equipment state mapping relationship according to the process parameter drift value, stress distribution value and material degradation value, calculate the failure probability of the key components of the semiconductor device, and trigger the fault diagnosis process when the failure probability exceeds a preset threshold;

[0186] The second unit is used to determine the target monitoring part according to the equipment state mapping relationship after the fault diagnosis process is triggered, collect multi-dimensional monitoring data of the target monitoring part, extract features of the multi-dimensional monitoring data to obtain comprehensive fault features, input the comprehensive fault features into a pre-trained digital twin model to perform fault mode recognition to obtain a fault diagnosis result, build a fault prediction model based on the fault diagnosis result, input historical fault evolution data to train the fault prediction model to obtain a fault development trend;

[0187] The third unit is used to construct a maintenance objective function according to the fault diagnosis result and the fault development trend, use an adaptive optimization algorithm to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set, input the optimal solution set into a preset maintenance decision rule library for comprehensive evaluation, select the solution with the highest comprehensive evaluation score as the optimal maintenance plan, convert the optimal maintenance plan into operation instructions for maintenance guidance, collect maintenance process data and equipment status data before and after maintenance, construct a maintenance knowledge graph according to the maintenance process data and equipment status data, update the fault prediction model and the maintenance decision rule library based on the maintenance knowledge graph, and realize continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

[0188] According to a third aspect of the embodiments of the present invention,

[0189] An electronic device is provided, comprising:

[0190] processor;

[0191] a memory for storing processor-executable instructions;

[0192] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0193] According to a fourth aspect of the embodiments of the present invention,

[0194] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0195] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semiconductor equipment fault prediction and maintenance method based on data fusion, characterized in that: include: Collect vibration sensor data, temperature sensor data, pressure sensor data and current sensor data of semiconductor equipment, construct a feature space through dimensionality reduction mapping, establish a process-stress-material multidimensional state model in the feature space, calculate process parameter drift values, stress distribution values ​​and material degradation values ​​based on the process-stress-material multidimensional state model, establish a device state mapping relationship according to the process parameter drift values, stress distribution values ​​and material degradation values, calculate the failure probability of key components of the semiconductor equipment, and trigger a fault diagnosis process when the failure probability exceeds a preset threshold; After the fault diagnosis process is triggered, the target monitoring part is determined according to the equipment state mapping relationship, the multi-dimensional monitoring data of the target monitoring part is collected, the feature extraction of the multi-dimensional monitoring data is performed to obtain the comprehensive fault feature, the comprehensive fault feature is input into the pre-trained digital twin model to perform fault mode recognition to obtain the fault diagnosis result, a fault prediction model is constructed based on the fault diagnosis result, and the historical fault evolution data is input to train the fault prediction model to obtain the fault development trend; A maintenance objective function is constructed according to the fault diagnosis result and the fault development trend, and an adaptive optimization algorithm is used to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set, and the optimal solution set is input into a preset maintenance decision rule library for comprehensive evaluation, and the solution with the highest comprehensive evaluation score is selected as the optimal maintenance plan, and the optimal maintenance plan is converted into an operation instruction for maintenance guidance, and maintenance process data and equipment status data before and after maintenance are collected, and a maintenance knowledge graph is constructed according to the maintenance process data and equipment status data, and a fault prediction model and a maintenance decision rule library are updated based on the maintenance knowledge graph to achieve continuous optimization of semiconductor equipment fault prediction and maintenance capabilities; Inputting the comprehensive fault features into the pre-trained digital twin model to perform fault mode recognition to obtain fault diagnosis results including: Inputting historical operation data of actual equipment into a physical knowledge guidance network, wherein the physical knowledge guidance network includes a physical constraint layer and a data driving layer, wherein the physical constraint layer uses the equipment dynamics equation to construct a loss function, and the data driving layer uses a time series feature extraction module to extract data association features, and inputting the loss function and the data association features into a digital twin model training unit to obtain a pre-trained digital twin model; The comprehensive fault features are input into a pre-trained digital twin model to output prediction features, and at the same time, an improved deep residual network is used to extract features from real-time monitoring data, wherein the deep residual network includes multiple residual learning units, and the residual learning unit uses deep separable convolution to extract local features to obtain real-time fault features; Adaptively fusing the prediction features with the real-time fault features, calculating the correlation weights of the two types of features through a cross-attention mechanism, and fusing the features based on the correlation weights to obtain enhanced fault features; Constructing a working condition perception module, inputting working condition information into a working condition encoder to obtain a working condition encoding vector, generating dynamic convolution kernel parameters based on the working condition encoding vector, and using the dynamic convolution kernel parameters to perform an adaptive convolution operation on the enhanced fault feature to obtain a working condition adaptation feature; Inputting the working condition adaptation feature into a multi-task learning network, wherein the multi-task learning network includes a fault classification branch and a fault degree assessment branch, wherein the fault classification branch outputs a fault type probability distribution, and the fault degree assessment branch outputs a fault degree score, to obtain a preliminary diagnosis result; A Bayesian neural network is used to model the uncertainty of the preliminary diagnosis result, and a prediction distribution is obtained through multiple random samplings. A confidence interval is calculated based on the prediction distribution, and a diagnosis result with a confidence interval higher than a preset confidence threshold is taken as the final fault diagnosis result.

2. The method according to claim 1, characterized in that Calculating process parameter drift values, stress distribution values, and material degradation values ​​based on the process-stress-material multidimensional state model, establishing a device state mapping relationship according to the process parameter drift values, stress distribution values, and material degradation values, calculating the failure probability of key components of the semiconductor device, and triggering a fault diagnosis process when the failure probability exceeds a preset threshold includes: Constructing a three-dimensional state tensor including process parameters, stress distribution, and material properties in the feature space, wherein the first dimension of the three-dimensional state tensor corresponds to the process parameters, the second dimension corresponds to the stress distribution, and the third dimension corresponds to the material properties, and uniformly sampling the three-dimensional state tensor according to the time dimension to obtain a state sequence of multiple time nodes; Perform multi-layer decomposition on the three-dimensional state tensor of each time node in the state sequence of multiple time nodes, and extract the characteristic data matrix of process parameters, the characteristic data matrix of stress distribution, and the characteristic data matrix of material properties respectively; Perform kernel function mapping on the characteristic data matrix of process parameters and the preset reference process state, and calculate the process parameter drift value of the characteristic data matrix of process parameters relative to the reference process state; Establishing a probability transmission mapping of the characteristic data matrix of stress distribution between adjacent time nodes, and calculating the stress distribution values ​​of the characteristic data matrix of stress distribution at adjacent time nodes; Establishing a state evolution equation for a characteristic data matrix of material properties through a probability density function, and solving the state evolution equation to obtain a material property degradation value of the material property; Building a dynamic probability graph model based on the process parameter drift value, stress distribution value and material property degradation value, wherein the dynamic probability graph model includes state transition probabilities between nodes and association probabilities between nodes; The state probability distribution of the semiconductor device at a future moment is calculated according to the dynamic probability graph model, and the failure probability is calculated based on the state probability distribution. When the failure probability exceeds a preset threshold, the fault diagnosis process is triggered.

3. The method according to claim 2, characterized in that Calculating the state probability distribution of the semiconductor device at a future time according to the dynamic probability graph model, and calculating the failure probability based on the state probability distribution includes: Combining process parameter drift values, stress distribution values, and material property degradation values ​​of semiconductor devices into state vectors, arranging the state vectors in chronological order to form a state vector sequence, and establishing a time-series dependency relationship of the state vector sequence in a dynamic probability graph model; Constructing a conditional variational posterior distribution network, wherein the conditional variational posterior distribution network comprises an encoder network and a decoder network, wherein the encoder network maps the state vector sequence to a latent variable space to obtain a latent variable representation, and the decoder network reconstructs the latent variable representation into a state vector; Constructing a recursive neural network including a long short-term memory unit and an attention mechanism unit, inputting the latent variable representation into the recursive neural network, and outputting a mean vector and a covariance matrix of a conditional variational posterior distribution; Calculate the relative entropy between the conditional variational posterior distribution and the true posterior distribution as the first loss function, calculate the mean square error between the reconstructed state vector and the original state vector as the second loss function, construct the weighted sum of the first loss function and the second loss function as the total loss function, and optimize the total loss function using a stochastic gradient descent algorithm with a momentum term; The current state vector is input into the optimized recursive neural network, and the state probability distribution representing the process parameters, stress distribution and material properties at the future time is output; Constructing a multidimensional mixed Gaussian distribution according to the statistical characteristics of the state probability distribution, extracting state samples from the multidimensional mixed Gaussian distribution, calculating the probability density ratio of each state sample under the state probability distribution and the multidimensional mixed Gaussian distribution, and obtaining the importance weight of each state sample; The state samples and the corresponding importance weights are input into a failure probability estimator built based on a deep neural network. The failure probability estimator maps the state samples to a failure probability space through a multi-layer perceptron network to obtain the failure probability of the semiconductor device.

4. The method according to claim 1, characterized in that: Constructing a maintenance objective function according to the fault diagnosis result and the fault development trend, performing multi-objective optimization on the maintenance objective function using an adaptive optimization algorithm to obtain an optimal solution set, inputting the optimal solution set into a preset maintenance decision rule library for comprehensive evaluation, and selecting the solution with the highest comprehensive evaluation score as the optimal maintenance plan includes: Based on the fault degree score and fault development trend prediction result in the fault diagnosis result, an exponential function is used to calculate the fault rate function value, and the equipment reliability objective function is constructed based on the fault rate function value; Obtain the preventive maintenance times and failure maintenance times of semiconductor equipment, calculate the total maintenance cost by combining the preventive maintenance cost coefficient and the failure maintenance cost coefficient, and construct a maintenance cost objective function; Obtaining a mean time between failures and a mean time to repair of a semiconductor device, calculating a device availability value according to a ratio of the mean time between failures and the mean time to repair, and constructing a device availability objective function; A maintenance objective function is constructed by combining the equipment reliability objective function, the maintenance cost objective function and the equipment availability objective function, and the maintenance objective function is input into an adaptive optimization model. The adaptive optimization model adaptively adjusts a crossover probability parameter and a mutation probability parameter according to the Euclidean distance value from the population individual to the optimal frontier, and obtains a non-dominated solution set through non-dominated sorting iterative calculation; Partitioning the solution space of the non-dominated solution set according to the fault type code, calculating the grid boundary threshold based on the density distribution of the solution in the partition, and adaptively meshing each subspace according to the grid boundary threshold to obtain a grid partition solution set; Constructing a maintenance decision rule base, calculating the reliability membership value, cost membership value and availability membership value corresponding to each solution in the grid partition solution set, and calculating the comprehensive evaluation score in combination with the corresponding indicator weight coefficient; A state deviation value is calculated according to real-time monitoring state data of semiconductor equipment, and the indicator weight coefficient and membership function parameter are adjusted based on the state deviation value to obtain a revised comprehensive evaluation value, and a non-dominated solution with the largest revised comprehensive evaluation value is selected as the optimal maintenance plan.

5. The method according to claim 1, characterized in that Convert the optimal maintenance plan into operation instructions for maintenance guidance, collect maintenance process data and equipment status data before and after maintenance, build a maintenance knowledge graph based on the maintenance process data and equipment status data, update the fault prediction model and maintenance decision rule base based on the maintenance knowledge graph, and realize the continuous optimization of semiconductor equipment fault prediction and maintenance capabilities, including: The maintenance plan is converted into a maintenance operation instruction sequence, wherein the maintenance operation instruction sequence includes operation part information, operation type information and operation parameter information, and a spatial positioning algorithm is used to track the maintenance scene in real time, and the maintenance operation instruction sequence is displayed on the equipment body through three-dimensional holographic projection; Acquire operation video stream information of the maintenance process, extract the operation action sequence from the operation video stream information based on a human posture estimation algorithm, collect maintenance process data including operation tool usage information, operation duration information, operation sequence information, and equipment sensor data and semiconductor equipment component image data before and after maintenance as equipment status data; Establish a maintenance knowledge ontology model, the maintenance knowledge ontology model includes equipment entity information, fault type information, maintenance operation information, and status feature information; extract knowledge triples from the maintenance process data and equipment status data according to preset rules and statistical methods; integrate the knowledge triples into a maintenance knowledge graph through entity alignment and relational reasoning; Vectorizing the entity information and relationship information in the maintenance knowledge graph to obtain a feature vector; inputting the feature vector into a graph neural network, optimizing the parameters of the fault prediction model, and obtaining an optimized fault prediction model; The maintenance decision rule base is updated by using a reinforcement learning method, and the maintenance effect information is used as a reward signal. A strategy search is performed in the maintenance knowledge graph based on a Monte Carlo tree search algorithm to generate an updated maintenance decision rule base. Fault prediction of equipment is performed using the optimized fault prediction model to obtain prediction results; a new maintenance plan is determined from the updated maintenance decision rule library based on the prediction results, thereby achieving continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

6. The method according to claim 5, characterized in that The maintenance decision rule base is updated by using a reinforcement learning method, and the maintenance effect information is used as a reward signal. A strategy search is performed in the maintenance knowledge graph based on a Monte Carlo tree search algorithm to generate an updated maintenance decision rule base including: Construct a Markov decision process model for maintenance decision-making, construct the equipment state vector and historical maintenance operation sequence as the state space, construct the optional maintenance operations as the action space, and obtain the state transition probability based on relational reasoning in the knowledge graph; Obtaining equipment status indicator information and maintenance cost information before and after maintenance, and calculating an immediate reward value based on the improvement degree of the equipment status indicator information and the maintenance cost; obtaining equipment operation time information of the next maintenance cycle, and calculating a delayed reward value based on the equipment operation time information; combining the immediate reward value with the delayed reward value to obtain a reward signal; Establish a search tree in the maintenance knowledge graph, use the maintenance status information as the tree node information, and use the maintenance operation information as the connection edge information; select the node to be expanded based on the confidence upper limit criterion, generate child nodes through the relationship information of the maintenance knowledge graph, and predict the maintenance operation effect through knowledge graph reasoning, and update the access count information and reward value information of each node; Extracting historical maintenance case information from the knowledge graph and establishing a heuristic function for allocating prior probabilities of maintenance operations; using the heuristic function for node selection in the search tree; The optimal maintenance strategy sequence obtained by the search is converted into a rule form containing trigger conditions and maintenance operation sequences, rule conflict detection and conflict resolution are performed, and confidence levels are assigned to the rules based on historical application results. Collect maintenance effect data, and update rule confidence information according to the maintenance effect data; delete rules whose confidence is lower than a preset threshold, and summarize the maintenance experience information to obtain new rules; add the new rules to the maintenance decision rule base to generate an updated maintenance decision rule base.

7. A semiconductor equipment fault prediction and maintenance system based on data fusion, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect vibration sensor data, temperature sensor data, pressure sensor data and current sensor data of the semiconductor device, construct a feature space through dimensionality reduction mapping, establish a process-stress-material multidimensional state model in the feature space, calculate the process parameter drift value, stress distribution value and material degradation value based on the process-stress-material multidimensional state model, establish an equipment state mapping relationship according to the process parameter drift value, stress distribution value and material degradation value, calculate the failure probability of the key components of the semiconductor device, and trigger the fault diagnosis process when the failure probability exceeds a preset threshold; The second unit is used to determine the target monitoring part according to the equipment state mapping relationship after the fault diagnosis process is triggered, collect multi-dimensional monitoring data of the target monitoring part, extract features of the multi-dimensional monitoring data to obtain comprehensive fault features, input the comprehensive fault features into a pre-trained digital twin model to perform fault mode recognition to obtain a fault diagnosis result, build a fault prediction model based on the fault diagnosis result, input historical fault evolution data to train the fault prediction model to obtain a fault development trend; The third unit is used to construct a maintenance objective function according to the fault diagnosis result and the fault development trend, use an adaptive optimization algorithm to perform multi-objective optimization on the maintenance objective function to obtain an optimal solution set, input the optimal solution set into a preset maintenance decision rule library for comprehensive evaluation, select the solution with the highest comprehensive evaluation score as the optimal maintenance plan, convert the optimal maintenance plan into operation instructions for maintenance guidance, collect maintenance process data and equipment status data before and after maintenance, construct a maintenance knowledge graph according to the maintenance process data and equipment status data, update the fault prediction model and the maintenance decision rule library based on the maintenance knowledge graph, and realize continuous optimization of semiconductor equipment fault prediction and maintenance capabilities.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images