A semiconductor device machine hand life prediction method

Abstract

The application provides a semiconductor equipment manipulator life prediction method, motion parameters of a manipulator joint are dynamically collected within a preset time; a multi-dimensional feature vector is extracted based on the motion parameters, a Mahalanobis distance is obtained according to the multi-dimensional feature vector and a reference state, and the Mahalanobis distance is mapped into a health degree index; based on a state space model including a state equation and an observation equation, a current time degradation rate parameter and a current wear state are obtained according to the health degree index, a number of cycles when a predicted wear amount reaches a predetermined failure threshold is obtained according to the degradation rate parameter and the current wear state, and a physical reference life is obtained according to the number of cycles and a time length corresponding to the cycles; a time sequence of the health degree index, the physical reference life and wet environment temperature and humidity data are input into a long short-term memory network model to obtain a predicted health degree trajectory; and based on a time point when the predicted health degree trajectory reaches a preset safety threshold, a remaining service life of the manipulator joint is determined.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, and more specifically to a method for predicting the lifespan of a semiconductor equipment robotic arm. Background Technology

[0002] In semiconductor manufacturing processes, wet front-end equipment is mainly used for processes such as wafer or mask cleaning, spin coating, development, and resist removal. The material handling system is a core component of this wet equipment, typically employing multi-jointed robotic arms to frequently move substrates between different process chambers and storage devices. These typical wet process environments often have specific temperature and humidity requirements and may contain volatile chemicals. The robotic arms, operating under these complex conditions for extended periods, undergo high-intensity, frequent vertical lifting, short-distance extension and contraction, or rotational movements. Consequently, core moving components such as reducers and servo motors at the joints inevitably experience wear and aging.

[0003] However, if a robotic arm malfunctions unexpectedly during operation, it can not only cause unplanned downtime of the entire wet processing equipment, but may also lead to the breakage, fall, or scrapping of expensive wafers and photomasks that are being handled or processed, resulting in economic losses. Therefore, it is necessary to monitor the health status of the robotic arm and predict its remaining service life.

[0004] Most existing equipment relies on theoretical component lifespan models for simple lifespan statistics to determine when to maintain and replace robotic arms, lacking accurate lifespan prediction capabilities. However, the operating conditions of semiconductor equipment robotic arms vary greatly, their load states change frequently when unloaded and handling different substrates, and their motion trajectories differ due to varying process formulations. Traditional statistical methods based on theoretical component lifespans struggle to accurately predict the actual lifespan of robotic arms, resulting in significant and uncontrollable random prediction errors. This makes it impossible to accurately predict component failure times and fails to meet the semiconductor manufacturing industry's requirements for equipment operating efficiency and reliability. Summary of the Invention

[0005] The problem this invention addresses is that existing statistical methods for predicting the actual lifespan of robotic arms based on theoretical component lifespan are difficult to accurately predict.

[0006] To address the above problems, this invention provides a method for predicting the lifespan of a semiconductor device robotic arm, comprising: Dynamically collect motion parameters of the robotic arm joints within a preset time period; Based on the motion parameters, a multidimensional feature vector is extracted, and the Mahalanobis distance is obtained based on the multidimensional feature vector and the baseline state. The Mahalanobis distance is then mapped to a health index. Based on the constructed state-space model containing state equations and observation equations, the degradation rate parameter and current wear state at the current moment are obtained according to the health index. The number of cycles when the predicted wear amount reaches the predetermined failure threshold is obtained according to the degradation rate parameter and the current wear state. The physical reference lifetime is obtained according to the number of cycles and the time length corresponding to the cycle. The time series of the health indicators, the physical baseline lifespan, and the wet environment temperature and humidity data are input into a long short-term memory network model to obtain the predicted health trajectory. The remaining service life of the robotic arm joints is determined based on the time point at which the predicted health trajectory reaches a preset safety threshold.

[0007] Optionally, the motion parameters include target position, target velocity, target acceleration, actual measured vibration acceleration, and actual servo motor current; the multidimensional feature vector includes gear meshing frequency energy; the extraction of the multidimensional feature vector based on the motion parameters includes: The theoretical gear meshing frequency is calculated based on the target speed and the gear ratio of the manipulator joint reducer. The discrete spectrum is obtained by performing a fast Fourier transform on the time-series signal of the vibration acceleration. The energy of the gear meshing frequency is obtained by integrating the square of the spectral amplitude of the vibration acceleration within a specific frequency band centered on the theoretical gear meshing frequency and its preceding harmonics.

[0008] Optionally, the multidimensional feature vector may also include an additional work integral; The additional work integral is calculated by integrating the product of the absolute value of the difference between the actual current of the servo motor and the reference current that performs the same action under the factory health condition of the device, the torque constant of the servo motor, and the target speed over time within a preset time window.

[0009] Optionally, the extraction of multidimensional feature vectors based on the motion parameters includes: The root mean square and kurtosis are obtained by performing time-domain statistical analysis on the amplitude of the vibration acceleration in the motion parameters. The root mean square, the kurtosis, the gear meshing frequency energy, and the additional work integral are concatenated into the multidimensional feature vector in a fixed order.

[0010] Optionally, the hidden state in the state-space model is defined as the actual wear depth inside the manipulator joint reducer; obtaining the degradation rate parameter and current wear state at the current moment based on the health index includes: Establish a state equation to characterize the recursive relationship of the actual wear depth between adjacent time steps under the influence of system process noise; An observation equation was established to characterize the relationship between the actual internal wear depth and the corresponding health index under the influence of measurement noise. The degradation rate parameter is obtained by solving the state equation and the observation equation simultaneously.

[0011] Optionally, the degradation rate parameter is an estimated degradation rate parameter. Based on the degradation rate parameter and the current wear state, the number of cycles required to predict the wear amount reaching a predetermined failure threshold is obtained. Then, based on the number of cycles and the corresponding time length of each cycle, the physical reference lifetime is obtained, including: Based on the estimated degradation rate parameter and the current wear state, the predicted wear amount for each future cycle is recursively calculated until the predicted wear amount reaches the predetermined failure threshold, so as to determine the number of cycles required to reach the predetermined failure threshold. The physical reference lifetime is calculated based on the number of cycles and the duration of a single cycle.

[0012] Optionally, obtaining the Mahalanobis distance based on the multidimensional feature vector and the baseline state, and mapping the Mahalanobis distance to a health index, includes: The Mahalanobis distance is calculated based on the multidimensional feature vector extracted from the current time window, the mean and covariance matrix of the feature values ​​when the device is in a factory health state; The Mahalanobis distance is exponentially mapped using preset parameters to obtain the health index with a value range normalized to zero to one.

[0013] Optionally, determining the remaining lifespan of the robotic arm joint based on the time point when the predicted health trajectory falls below a preset safety threshold includes: Based on the predicted health trajectory of the robotic hand joint in each future cycle output by the Long Short-Term Memory Network Model, the number of simulation cycles required for the predicted health trajectory to first fall below the preset safety threshold from the current moment is determined. Multiplying the number of simulation cycles by the duration of a single handling cycle yields the remaining service life of the robotic arm joint at the current moment.

[0014] Optionally, the offline training data of the long short-term memory network model covers the entire lifecycle data of the robotic arm joint. The entire lifecycle data includes data in load scenarios such as no load, wafer handling, and mask handling, data in trajectory scenarios of movement operations at different distances, and data in different wet process environment scenarios.

[0015] Optionally, during the offline training phase, the long short-term memory network model uses a loss function that combines mean squared error and physical monotonicity penalty term for parameter optimization. The mean squared error term is used to measure the deviation between the health status predicted by the model and the actual health status label; the physical monotonicity penalty term is used to force the predicted health status trajectory output by the model to exhibit a monotonically decreasing trend in the time dimension.

[0016] Compared with the prior art, the technical solution of the embodiments of the present invention has the following advantages: The semiconductor device robotic arm life prediction method provided in this application dynamically collects the motion parameters of the robotic arm joints and obtains health indicators based on these motion parameters, which can accurately and dynamically determine the subtle degradation of the semiconductor device. Furthermore, by using a state-space model containing state equations and observation equations, it obtains the degradation rate parameters and current wear state at the current moment, and based on this, obtains the number of cycles and physical baseline life, ensuring the physical rationality of the life prediction, conforming to the joint wear mechanism, and guaranteeing the theoretical rigor of the prediction. It can track the actual degradation state of the device in real time, overcoming the defect of traditional static models where errors increase over time. Moreover, by inputting the time series of health indicators, physical baseline life, and wet environment temperature and humidity data into a long short-term memory network model to obtain a predicted health estimate, it determines the remaining service life of the robotic arm joints. By integrating physical baseline life and time series environmental data, it not only relates to the physical laws of device degradation but also includes dynamic correlations under complex working conditions, effectively reducing the prediction bias of a single model and the influence of environmental interference, significantly improving the reliability and accuracy of the remaining life assessment results. Attached Figure Description

[0017] Figure 1 A schematic flowchart illustrating the semiconductor device robot life prediction method provided in this application embodiment; Figure 2 A schematic diagram of the hardware application environment structure for the robot life prediction method provided in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the data flow of multidimensional feature and model fusion processing provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the Long Short-Term Memory network model provided in the embodiments of this application; Figure 5 This is a schematic diagram of the trajectory prediction of joint health of a robotic arm provided in an embodiment of this application. Detailed Implementation

[0018] To make the above-mentioned objectives, features and structures of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below.

[0019] Please see Figure 1 , Figure 1This is a flowchart illustrating the semiconductor equipment robot lifespan prediction method provided in this application. The semiconductor equipment robot lifespan prediction method of this application is mainly applied to semiconductor manufacturing equipment, especially wet process equipment, such as various cleaning, spin coating, and developing equipment, as well as integrated systems combining the above equipment. Besides equipment used in wafer manufacturing, this method is also applicable to mask equipment. The overall architecture of wet process equipment includes a material handling system, an environmental control system, process chamber-related systems, etc. Furthermore, it is understood that in addition to aging and lifespan depletion caused by the robot's own movement, the environmental conditions within the process chamber containing volatile chemicals and specific temperature and humidity conditions can also accelerate the aging of mechanical components.

[0020] The core concept of this application is to improve the accuracy of robot life prediction by dynamically determining model parameters through machine learning based on the established robot life prediction model. By integrating the physical state-space model and the machine learning model, the multi-dimensional temporal characteristics of the robot are processed to achieve dynamic prediction of the remaining life of the weakest component. Based on this more accurate remaining life prediction result, targeted maintenance and replacement strategies are set, or subsequent process parameters are optimized, thereby optimizing and improving the actual remaining service life.

[0021] The specific process of the semiconductor equipment robot life prediction method is as follows: S110. Dynamically collect motion parameters of the robotic arm joints within a preset time.

[0022] The motion parameters of each joint of the robotic arm are acquired within a preset time window. In some embodiments, the time window is a sliding time window of fixed length. The motion parameters collected in real time within this time window include target position, target velocity, target acceleration, encoder position deviation, vibration acceleration, and actual servo motor current, etc. At a minimum, vibration acceleration and actual servo motor current should be included. These multi-dimensional motion parameters together constitute a complete data input for assessing the health status of the robotic arm joints.

[0023] The target refers to the key components of specific joints of the robotic arm being studied, such as motors or reducers.

[0024] Vibration acceleration refers to the dynamic physical response generated by key components at specific joints of a robotic arm when performing the aforementioned movements. By collecting various motion data, the true motion state of the robotic arm can be comprehensively reflected. Furthermore, the generation of vibration acceleration is also related to the target position, target velocity, and target acceleration in the motion data.

[0025] Encoder position deviation is used to characterize the motion trajectory following error caused by wear of the transmission mechanism. It refers to the difference between the target position command read from the servo drive or host controller and the actual feedback position of the encoder, i.e., the following error.

[0026] The actual current of the servo motor directly reflects the driving torque required to overcome mechanical friction and load.

[0027] Understandably, research shows that methods that use second or third derivative calculations of the position signals fed back by encoders to obtain acceleration or jerk, and then infer the vibration state of the equipment, cannot replace directly acquiring vibration acceleration or calculating vibration acceleration from the vibration state. This is because the target acceleration issued by the control system or the actual acceleration calculated by the encoder reflects the rigid body kinematic characteristics of the manipulator joints on the macroscopic motion trajectory. Its frequency is usually low and mainly controlled by the closed-loop regulation of the servo system. However, critical internal components of the manipulator, such as reducers or bearings, generate high-frequency impacts and friction at the microscopic level when experiencing early wear, fatigue, spalling, or poor lubrication. These impacts then propagate as elastic waves within the mechanical structure, forming the acquired high-frequency vibration acceleration. This high-frequency vibration acceleration is a direct dynamic physical response of the mechanical structure to internal physical damage, and its frequency range far exceeds the bandwidth of the servo control system, making it impossible to effectively represent in the macroscopic kinematic parameters. Relying solely on calculations to obtain vibration will completely lose these characteristics, reducing the accuracy of the prediction model.

[0028] Therefore, the vibration acceleration in this application is preferably detected by an independent sensor, such as... Figure 2 As shown, Figure 2 This diagram illustrates the hardware application environment of the robot life prediction method provided in this application. Vibration acceleration is measured using physical sensors. Specifically, vibration acceleration refers to the dynamic acceleration response of the robot's joint surfaces in three orthogonal directions during handling tasks. To achieve high-precision measurement, triaxial microelectromechanical system (MEMS) accelerometers can be directly mounted or embedded in key force-bearing and transmission components such as the reducer housings of the robot's joints or the motor flanges of the servo motors. These sensors can directly sense the real physical vibrations generated by gear meshing and bearing operation within the mechanical structure and convert them into electrical signals. This operation not only acquires high-frequency vibration signals containing rich degradation characteristics but also, by combining synchronously acquired kinematic context such as target position and target velocity, accurately distinguishes between normal vibrations and vibrations caused by abnormal wear under specific working conditions, facilitating the identification of key wear characteristics.

[0029] S120. Extract multidimensional feature vectors based on motion parameters, obtain Mahalanobis distance between the multidimensional feature vectors and the baseline state, and map the Mahalanobis distance into a health index.

[0030] Understandably, a multidimensional feature vector is a one-dimensional array used to comprehensively characterize the current health status of a robotic hand's joints. This multidimensional feature vector is defined based on several key parameters extracted from motion data, including at least the root mean square of vibration acceleration, kurtosis, gear meshing frequency energy, and the integral of additional work done. These parameters provide quantitative indicators of mechanical degradation, offering reliable data support for subsequent health mapping. The one-dimensional array can also include original target position, target velocity, and target acceleration data.

[0031] Specifically, time-domain statistical analysis is performed on the amplitude of vibration acceleration in the collected motion parameters to extract the root mean square and kurtosis.

[0032] The root mean square (RMS) is used to characterize the average magnitude of vibrational energy over a given time period. The formula for calculating the RMS is: ; Where N is the total number of sampling points of the vibration acceleration signal within the time window; n is a positive integer satisfying 1≤n≤N; and x(n) is the vibration acceleration amplitude of the nth sampling point.

[0033] Kurtosis is a statistical measure reflecting the distribution characteristics of vibration signals. It is highly sensitive to impact signals and can be used to measure wear caused by impact response during use. The formula for calculating kurtosis is: ; in, This represents the average value of the vibration acceleration signal. Kurtosis is used to characterize the presence of sudden impacts such as tooth breakage.

[0034] In some other embodiments, the multidimensional feature vector may also include gear meshing frequency energy. It is understood that early functional degradation can be detected through gear meshing frequency energy. When gears inside the reducer wear early, microscopic metal spalling and increased gear clearance will generate high-frequency vibrations immediately. Therefore, by extracting gear meshing frequency energy, signs of degradation can be detected very early.

[0035] The calculation of gear meshing frequency energy depends on the target velocity and vibration acceleration in the aforementioned collected motion data. First, the theoretical gear meshing frequency is calculated based on the target velocity of the servo motor and the gear ratio of the reducer. Then, a Fast Fourier Transform is performed on the vibration acceleration time-series signal to convert the time-domain signal into a frequency-domain signal, obtaining a discrete spectrum. Finally, the square of the spectral amplitude is integrated within a specific frequency band centered on the theoretical gear meshing frequency and its preceding harmonics. The formula for calculating gear meshing frequency energy is: ; in, Let M be the energy at the gear meshing frequency, M be the harmonic order considered, and j be a positive integer satisfying 1 ≤ j ≤ M. This is the theoretical gear meshing frequency; The set half-width of the frequency band; This represents the frequency domain amplitude of the vibration acceleration signal.

[0036] Simultaneously, the additional work integral is calculated based on the actual current of the servo motor and the target speed. The additional work integral is used to quantify the extra mechanical work output by the motor to overcome the resistance of mechanical aging and wear when the robot completes one handling operation. Because if the robot's reducer wears or jams, the motor must output a larger electromagnetic torque to reach the set target speed, which manifests as an actual current that is significantly higher than the reference current. Therefore, calculating this integral can represent the extra mechanical work done by the motor to overcome degradation resistance within the time window.

[0037] The additional work integral is calculated by integrating the product of the absolute difference between the actual current of the servo motor and the reference current performing the same action under factory health conditions within a preset time window, the torque constant of the servo motor, and the target speed over time. Specifically, the additional work integral is: ; in For extra work done, The time window length, This represents the actual current of the servo motor. This is the reference current when the equipment performs the same operation under factory-safe conditions. Let be the torque constant of the motor. The target speed.

[0038] The calculated parameters, namely the root mean square of vibration acceleration, kurtosis, gear meshing frequency energy, additional work integral, and optionally other motion data, are concatenated and combined in a fixed order to form a complete multidimensional feature vector.

[0039] The process of obtaining Mahalanobis distance based on multidimensional feature vectors and baseline states, and mapping Mahalanobis distance to a health index, includes: calculating Mahalanobis distance based on multidimensional feature vectors extracted from the current time window, the mean and covariance matrix of feature values ​​when the device is in a factory health state; and performing exponential mapping on the Mahalanobis distance using preset parameters to obtain a health index with a value range normalized to zero to one.

[0040] The formula for extracting multidimensional feature vectors and calculating Mahalanobis distance is as follows: ; Among them, MD t The Mahalanobis distance, This is the multidimensional feature vector extracted for the current time window. and These are the mean and covariance matrices of the characteristic values ​​when the equipment is in its factory healthy state. After calculating the Mahalanobis distance, an exponential function is used to map the Mahalanobis distance to a health index. That is: ; in, Let k be the health index at time k. These are preset parameters. Let be the Mahalanobis distance at time k. The range of the health index is normalized to between zero and one using the above formula. The normalized health index solves the problem that the original temperature, humidity, and vibration data cannot directly characterize lifespan, and can be directly used for lifespan prediction.

[0041] S130. Based on the constructed state-space model containing state equations and observation equations, the degradation rate parameter and current wear state at the current moment are obtained according to the health index. The number of cycles when the predicted wear amount reaches the predetermined failure threshold is obtained according to the degradation rate parameter and the current wear state. The physical reference lifetime is obtained according to the number of cycles and the time length corresponding to the cycle.

[0042] It should be noted that constructing a state-space model that includes state equations and observation equations aims to establish the relationship between the physical model and lifetime parameters, providing physical constraints for subsequent deep learning models and avoiding the divergence of purely data-driven models.

[0043] In some embodiments, the hidden state in the state-space model is defined as the actual physical wear depth inside the manipulator joint reducer. Obtaining the degradation rate parameter and current wear state at the current moment based on the health index includes: establishing a state equation to characterize the recursive relationship of the actual wear depth at adjacent moments under the influence of system process noise; establishing an observation equation to characterize the relationship between the actual wear depth and the corresponding moment's health index under the influence of measurement noise; and solving the state equation and observation equation simultaneously to obtain the degradation rate parameter.

[0044] The state equation is as follows: ; in Let be the actual internal wear depth at time k, which is the actual internal wear depth at the current time. This represents the actual internal wear depth at time k-1, which is the actual internal wear depth at the previous time step. and For degradation rate parameters, This represents system process noise. The exponential term in this recursive formula incorporates the parameter k to reflect the physical law of accelerated wear in the later stages of equipment aging.

[0045] The observation equation is .

[0046] in Let k be the health index at time k. Let C be the actual internal wear depth at time k, and C be the conversion factor. For measuring noise.

[0047] The aforementioned state equation and observation equation construct a continuous degradation physical model for a single life stage. Instead of using a piecewise function, it naturally simulates the entire process from slow wear in the early stage to accelerated degradation in the later stage through the nonlinear growth of the exponential term, so as to smoothly reflect the real physical degradation trajectory of the manipulator joint in a wet environment. This is mainly because it does not need to consider the numerical mutation and prediction discontinuity problems that may occur at the switching points of the piecewise model.

[0048] In some embodiments, the degradation rate parameter is an estimated value of the degradation rate parameter. Based on the degradation rate parameter and the current wear state, the number of cycles required for the predicted wear amount to reach a predetermined failure threshold is obtained. The physical reference lifetime is obtained based on the number of cycles and the time length corresponding to each cycle. This includes: recursively calculating the predicted wear amount for each future cycle based on the estimated degradation rate parameter and the current wear state until the predicted wear amount reaches the predetermined failure threshold, so as to determine the number of cycles required to reach the predetermined failure threshold; and calculating the physical reference lifetime based on the number of cycles and the time length of a single cycle.

[0049] Specifically, health indicators are used as observed values ​​to correct the degradation rate parameters in the equation of state, and the physical baseline lifespan of each joint is calculated. By combining the equation of state and the observation equation, and using the available observations, namely the health indicators, the degradation rate parameters in the equation of state are estimated.

[0050] In other embodiments, mature particle filtering algorithms or Kalman filtering algorithms are used to solve and estimate the parameters in the above equations. For example, when using a particle filtering algorithm for parameter estimation and state update, the particle swarm is first initialized to generate a set of particles for the degradation rate parameter and initial wear state. Then, state prediction is performed, using the state equation to predict the wear depth of each particle at the current moment. Next, the theoretical health is calculated by multiplying the predicted wear depth by a conversion coefficient using the observation equation to obtain the theoretical health of each particle. At the same time, the actual observed health is obtained, which is the current health index obtained by mapping Mahalanobis distance and exponential function. Then, the residual is calculated by comparing the theoretical health and the actual observed health, and the weight of each particle is updated using the Gaussian likelihood function. After that, the particles are resampled according to the weight, eliminating low-weight particles and replicating high-weight particles. Finally, parameter estimation is performed, calculating the mean of the particle swarm after resampling to obtain the optimal degradation rate parameter estimate and the current wear state at the current moment. Thus, the required degradation rate parameter estimate is obtained.

[0051] After obtaining the optimal parameter estimates, a recursive extrapolation is performed towards future time. The final formula for calculating the predicted wear amount in the m-th future period after elimination is as follows: ; in, and This is the optimal degradation rate parameter estimate for the current moment; Let be the estimated wear state at time k, which is considered as the current wear state; j is a positive integer satisfying 1≤j≤m.

[0052] Calculate the predicted wear amount for the m-th future cycle until the calculated predicted wear amount reaches the set failure threshold. Calculate the physical reference lifetime. .in The duration of a single handling cycle.

[0053] It should be noted that this physical baseline lifetime serves as a benchmark for the physical wear evolution rate under ideal conditions. This closed-loop correction mechanism enables real-time tracking of the actual degradation state of the equipment, overcoming the limitation of traditional static models where errors increase over time. By calculating and outputting this physical baseline lifetime, a coarse lifetime estimate under ideal conditions is essentially provided as input to the neural network. This shifts the neural network's focus from learning fundamental physical laws to learning precise nonlinear corrections to this coarse estimate based on complex environmental data, a fundamental difference from existing lifetime prediction methods.

[0054] Can Figure 3 As shown, Figure 3This is a schematic diagram illustrating the data flow of multidimensional feature and model fusion processing provided in the embodiments of this application.

[0055] S140. Input the time series of health indicators, physical baseline lifespan and wet environment temperature and humidity data into the long short-term memory network model to obtain the predicted health trajectory.

[0056] For example, the time series operating condition data of the health index is concatenated with the physical baseline lifespan to form a joint vector. In the basic embodiment of this application, the operating condition data is only illustrated by including wet environment temperature and humidity data. The wet environment temperature and humidity data and the time series of the health index together constitute the input data matrix.

[0057] It is understandable that the semiconductor wet process chamber is often accompanied by the volatilization of corrosive gases and extreme temperature and humidity changes. These environmental factors will directly accelerate the aging of the joint seals of the robot and the deterioration of the lubricating grease, thereby significantly changing the physical degradation rate of the reducer. Therefore, using the ambient temperature and humidity as the core operating condition data input can greatly correct the prediction deviation of the physical model under ideal conditions.

[0058] In a preferred embodiment of this application, the operating condition data may further include the path trajectory and load data at the current moment. Since the dynamic physical response of the robot arm has already been sufficiently characterized by vibration acceleration in the aforementioned multidimensional feature vector, there is no need to repeatedly introduce any vibration or chattering features into the operating condition data here, avoiding feature redundancy and model overfitting. Adding path trajectory and load data to the operating condition data can help the model learn the nonlinear effects of different handling distances and different wafer or mask weights on the robot arm's lifespan, thereby achieving higher accuracy in lifespan prediction in the optimal embodiment.

[0059] Please continue reading. Figure 4 , Figure 4 This is a schematic diagram of the Long Short-Term Memory (LSTM) network model provided in an embodiment of this application. The LTM network model is preferably deployed on an edge computing industrial control computer. Specifically, the LTM network model includes an input layer, three hidden layers, and one fully connected output layer. The input layer receives the concatenated joint vector, the dimension of which depends on the time window length and the number of features. All three hidden layers use LTM units; the first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons. A random deactivation layer with a dropout rate of 20% is set after each hidden layer to prevent overfitting when dealing with complex working conditions. The fully connected output layer contains one neuron to output the predicted health value for each future time step.

[0060] The offline training phase of the Long Short-Term Memory (LSTM) network model used training data covering the entire lifecycle of multiple robotic arms' joints. This lifecycle data included diverse load scenarios such as idle wafer handling and mask handling, diverse trajectory scenarios involving frequent up-and-down movements over long and short distances, and diverse wet process environment scenarios ranging from ambient temperature and humidity to high temperature and humidity. During training, all collected lifecycle data were first normalized to their minimum and maximum values, mapping them to the zero-to-one range to eliminate the impact of different units on gradient descent. The dataset was then divided into training and validation sets at a ratio of 80% to 20%. The model training employed an adaptive moment estimation optimizer with an initial learning rate of 0.001, using cosine annealing for learning rate decay. The batch size was set to 64.

[0061] In some embodiments, during the offline training phase, the Long Short-Term Memory (LSTM) network model employs a loss function combining mean squared error (MSE) and a physical monotonicity penalty term for parameter optimization. The MSE term measures the deviation between the model's predicted health status and the actual health status label; the physical monotonicity penalty term forces the predicted health status trajectory output by the model to exhibit a monotonically decreasing trend over time. The formula for calculating the loss function is as follows: ; Where N is the total number of time steps for prediction, and i is a positive integer not less than 1. For accurate health labels, For the health level predicted by the model, The penalty coefficient is used. The penalty term in this loss function forces the predicted health trajectory output by the model to exhibit a monotonically decreasing trend over time. This aligns with the physical law of irreversible wear and tear on mechanical parts, thus effectively avoiding the illogical fluctuations in predictions that might occur with purely data-driven models. An early stopping mechanism is implemented during training: when the loss function on the validation set no longer decreases for 20 consecutive rounds, training stops and the current weights are saved as the optimal model.

[0062] The machine learning process described above essentially involves inputting a joint vector into a Long Short-Term Memory (LSTM) network model, and then performing nonlinear fine-tuning of the physical baseline lifetime by combining historical degradation trends with current complex operating conditions. It's important to clarify that the output of the LTM network model is not the degradation parameter, nor is it the physical baseline lifetime itself, but rather a predicted health indicator for future time steps.

[0063] The physical baseline lifetime calculated from real data is essentially the input, not the output, for model training and application. The physical baseline lifetime is fed as input to the neural network because purely data-driven models converge slowly when faced with complex mechanical wear, and their prediction accuracy cannot be guaranteed. The physical baseline lifetime actually provides the neural network with a theoretical reference line for pure mechanical wear based on a physical model. The training and application process of the neural network in this application can be viewed as a nonlinear correction of this theoretical reference line based on complex environmental data. Especially after inputting the semiconductor equipment's process environment, the model actually learns and outputs the trend of lifetime changes under varying process environments and other inputs, thus obtaining the subsequent health status and health status trajectory. Particularly under harsh operating conditions of high temperature, high humidity, and high concentration of chemical vapors, the neural network may significantly correct the physical baseline lifetime based on the patterns in the training set.

[0064] During the offline training phase, this application employs a sliding window to construct training data inputs that include historical health status sequences, current operating condition data, and physical baseline lifespan. The output labels are simply the actual health status indicators for the next cycle or a fixed short window in the future. Using these actual health status labels to guide the network's loss function calculation ensures the model's absolute convergence. Then, to obtain the predicted health status trajectory throughout the entire process, the model utilizes an autoregressive iterative method for prediction, progressively advancing through iterative iterations to obtain the trajectory of the health status indicators.

[0065] Specifically, the Long Short-Term Memory (LSTM) network model first outputs the predicted health index for the first future period based on the joint vector of the current time window. Then, this predicted health index value is used as part of known historical data and re-inputted into the model to predict the health index for the second future period, until the predicted health index reaches or falls below a preset safety threshold. Piecing together these iteratively predicted discrete points forms a complete predicted health index trajectory.

[0066] S150. Based on the time point when the predicted health trajectory reaches the preset safety threshold, determine the remaining service life of the robotic arm joint.

[0067] Specifically, this includes: predicting the health trajectory of the robotic joint for each future cycle based on the output of the long short-term memory network model; determining the number of simulation cycles required for the predicted health trajectory to first fall below the preset safety threshold from the current moment; and multiplying the number of simulation cycles by the length of a single handling cycle to obtain the remaining service life of the robotic joint at the current moment.

[0068] The remaining lifespan is determined by identifying the point in time when the predicted health trajectory first falls below a preset safety threshold. The formula is: ; in Let be the remaining lifespan of the i-th robotic joint at the current time t. It represents the number of cycles projected from the current moment into the future. The i-th robotic joint output by the Long Short-Term Memory network model in the future Predictive health indicators at any given time The safety threshold is typically set between 0.1 and 0.2 according to the equipment's factory specifications. The duration of a single handling cycle.

[0069] Understandably, this application iterates along the monotonically decreasing health trajectory predicted by the model cycle by cycle. Once it finds that the predicted health index of a certain future cycle is less than or equal to the safety threshold, it multiplies the difference between the current cycle and the single cycle time to obtain the remaining service life accurate to the time unit. This addresses the short-board effect of the robotic arm and avoids unplanned downtime caused by the sudden failure of local parts.

[0070] For example, please continue reading Figure 5 , Figure 5 This diagram illustrates the prediction of the health index (HI) trajectory of a robotic arm joint provided in this application embodiment. The diagram shows the degradation trajectory of the HI of three key components in the robotic arm of a semiconductor wet processing equipment over a future prediction period. Component A is the base rotary reducer, component B is the Z-axis lifting lead screw bearing, and component C is the arm extension servo motor. The vertical axis represents the normalized health index (0 to 1). It is important to note that in the actual operating environment of a semiconductor wafer fab, due to different maintenance cycles and replacement strategies (asynchronous maintenance), the initial health of each component is not the same at the current assessment time. In the diagram, component C may have just been replaced, with its HI close to 1.0, while components A and B have experienced different degrees of wear, with HIs of 0.85 and 0.95, respectively. The predicted trajectory shown in the diagram is a broken line with real-world operating conditions, output by a Long Short-Term Memory (LSTM) network model.

[0071] It should be noted that because the model integrates dynamic operating condition data such as temperature and humidity in wet environments and diverse loads, and introduces a physical monotonicity penalty term in the loss function, the predicted trajectory exhibits a localized step-like decline and an overall strictly monotonically decreasing trend. This not only realistically reflects the equipment's performance under high-intensity handling tasks or extreme temperature and humidity fluctuations, but also reflects the accelerated step-like decline in health. A horizontal dashed line with a height of 0.15 is set as a safety threshold in the figure. According to the minimum value logic of this application, the system tracks the predicted trajectories of all components in parallel. As can be seen from the figure, since component A has the fastest degradation rate under future operating conditions, its predicted trajectory is the first to fall below the 0.15 safety threshold. The horizontal axis time corresponding to this intersection point is determined as the failure time point of the entire robotic arm system, which means the remaining service life can be derived.

[0072] It should be noted that the predicted trajectories and initial health levels of components A, B, and C in the embodiments of this application are merely illustrative examples and do not constitute a limitation on the scope of protection of this application. In practical applications, the remaining service life of a component is not directly determined by the level of its initial health index. Regardless of the initial health index level, it is necessary to make a judgment based on its predicted health trajectory and a preset safety threshold.

[0073] After determining the specific remaining service life of each joint, the method of this application further includes a structured maintenance instruction generation step based on a large language model and a local knowledge base. While achieving proactive control and life extension, the edge computing industrial control computer converts the current multi-dimensional feature vector of the predicted weakest component and the health degradation curve output by the long short-term memory network model into prompt words, which are then input into a locally deployed lightweight large language model. This large language model is externally coupled with a vector database built based on historical maintenance records of semiconductor wet process equipment and original manufacturer technical manuals, thus forming a retrieval-enhanced generative architecture large language model, which performs technical attribution on the aforementioned time-series features and predicted lifespan. The analysis output includes a structured technical work order containing specific failure mechanisms, recommended spare parts list numbers, and suggested downtime maintenance time windows. This work order is then pushed to the wafer fab's manufacturing execution system. The specific failure mechanism could be, for example, fatigue wear of the flexspline in a harmonic reducer. By using objective sensor data and physical lifespan predicted by a long short-term memory network as input, the large language model outputs technical instructions to guide the maintenance of physical entities. This effectively solves the technical problem that the attribution of complex semiconductor equipment failures is highly dependent on human experience. It achieves closed-loop management from lifespan prediction to intelligent maintenance, significantly improving equipment maintenance efficiency and reducing reliance on senior maintenance personnel.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of a semiconductor equipment robotic arm, characterized in that, include: Dynamically collect motion parameters of the robotic arm joints within a preset time period; Based on the motion parameters, a multidimensional feature vector is extracted, and the Mahalanobis distance is obtained based on the multidimensional feature vector and the baseline state. The Mahalanobis distance is then mapped to a health index. Based on the constructed state-space model containing state equations and observation equations, the degradation rate parameter and current wear state at the current moment are obtained according to the health index. The number of cycles when the predicted wear amount reaches the predetermined failure threshold is obtained according to the degradation rate parameter and the current wear state. The physical reference lifetime is obtained according to the number of cycles and the time length corresponding to the cycle. The time series of the health indicators, the physical baseline lifespan, and the wet environment temperature and humidity data are input into a long short-term memory network model to obtain the predicted health trajectory. Based on the time point when the predicted health trajectory reaches the preset safety threshold, the remaining service life of the robotic hand joint is determined. The hidden state in the state-space model is defined as the actual physical wear depth inside the manipulator joint reducer. Obtaining the degradation rate parameter and current wear state based on the health index includes: Establish a state equation to characterize the recursive relationship of the actual wear depth between adjacent time steps under the influence of system process noise; An observation equation was established to characterize the relationship between the actual internal wear depth and the corresponding health index under the influence of measurement noise. The degradation rate parameter is obtained by solving the state equation and the observation equation simultaneously.

2. The semiconductor equipment robot life prediction method as described in claim 1, characterized in that: The motion parameters include target position, target velocity, target acceleration, actual measured vibration acceleration, and actual servo motor current; The multidimensional feature vector includes gear meshing frequency energy; the extraction of the multidimensional feature vector based on the motion parameters includes: The theoretical gear meshing frequency is calculated based on the target speed and the gear ratio of the manipulator joint reducer. The discrete spectrum is obtained by performing a fast Fourier transform on the time-series signal of the vibration acceleration. The energy of the gear meshing frequency is obtained by integrating the square of the spectral amplitude of the vibration acceleration within a specific frequency band centered on the theoretical gear meshing frequency and its preceding harmonics.

3. The semiconductor equipment robot life prediction method as described in claim 2, characterized in that: The multidimensional feature vector also includes an additional work integral; The additional work integral is calculated by integrating the product of the absolute value of the difference between the actual current of the servo motor and the reference current that performs the same action under the factory health condition of the device, the torque constant of the servo motor, and the target speed over time within a preset time window.

4. The semiconductor equipment robot life prediction method as described in claim 3, characterized in that: The extraction of multidimensional feature vectors based on the motion parameters includes: The root mean square and kurtosis are obtained by performing time-domain statistical analysis on the amplitude of the vibration acceleration in the motion parameters. The root mean square, the kurtosis, the gear meshing frequency energy, and the additional work integral are concatenated into the multidimensional feature vector in a fixed order.

5. The method for predicting the lifespan of a semiconductor equipment robot as described in claim 1, characterized in that: The degradation rate parameter is an estimated value. Based on the degradation rate parameter and the current wear state, the number of cycles required for the predicted wear amount to reach a predetermined failure threshold is obtained. The physical reference lifetime is then obtained based on the number of cycles and the corresponding time length of each cycle, including: Based on the estimated degradation rate parameter and the current wear state, the predicted wear amount for each future cycle is recursively calculated until the predicted wear amount reaches the predetermined failure threshold, so as to determine the number of cycles required to reach the predetermined failure threshold. The physical reference lifetime is calculated based on the number of cycles and the duration of a single cycle.

6. The method for predicting the lifespan of a semiconductor equipment robot as described in any one of claims 1 to 5, characterized in that: The step of obtaining the Mahalanobis distance based on the multidimensional feature vector and the baseline state, and mapping the Mahalanobis distance to a health index, includes: The Mahalanobis distance is calculated based on the multidimensional feature vector extracted from the current time window, the mean and covariance matrix of the feature values ​​when the device is in a factory health state; The Mahalanobis distance is exponentially mapped using preset parameters to obtain the health index with a value range normalized to zero to one.

7. The method for predicting the lifespan of a semiconductor equipment robot as described in any one of claims 1 to 5, characterized in that: The determination of the remaining service life of the robotic arm joint based on the time point when the predicted health trajectory falls below the preset safety threshold includes: Based on the predicted health trajectory of the robotic hand joint in each future cycle output by the Long Short-Term Memory Network Model, the number of simulation cycles required for the predicted health trajectory to first fall below the preset safety threshold from the current moment is determined. Multiplying the number of simulation cycles by the duration of a single handling cycle yields the remaining service life of the robotic arm joint at the current moment.

8. The method for predicting the lifespan of a semiconductor equipment robot as described in any one of claims 1 to 5, characterized in that: The offline training data of the Long Short-Term Memory Network model covers the entire lifecycle data of the robotic arm joints. The entire lifecycle data includes data in load scenarios such as no load, wafer handling, and mask handling, data in trajectory scenarios of movement operations at different distances, and data in different wet process environment scenarios.

9. The method for predicting the lifespan of a semiconductor equipment robot as described in any one of claims 1 to 5, characterized in that: During the offline training phase, the long short-term memory network model uses a loss function that combines mean squared error and physical monotonicity penalty term for parameter optimization. The mean squared error term is used to measure the deviation between the health status predicted by the model and the actual health status label; the physical monotonicity penalty term is used to force the predicted health status trajectory output by the model to exhibit a monotonically decreasing trend in the time dimension.

Images