Wafer processing fault early warning method and system based on hybrid sequence decomposition and mamba architecture

Abstract

The application discloses a wafer process fault early warning method and system based on mixed sequence decomposition and Mamba architecture, which processes long-range dependence and non-stationary trend characteristics presented by multi-channel sensor time series data in semiconductor wafer processing. First, the trend and periodicity of each sensor signal are decomposed, and only the residual part is reserved as input. Then, a double-branch autoencoder structure is used for anomaly detection: the first branch is a linear state space model, which efficiently models long-term dynamics; the second branch is an attention mechanism, which captures local and contextual associations between sensors; an adaptive gating mechanism is introduced to fuse the outputs of the two branches, dynamically highlighting abnormal patterns according to actual conditions; finally, the abnormal score is calculated based on the reconstruction error of the autoencoder. The method has higher fault early warning performance and running efficiency, can be used for real-time monitoring and early warning of faults in wafer manufacturing process, improves the safety and yield of the production process, and has important practical value.

Description

Technical Field

[0001] This invention relates to the field of multi-sensor time-series data processing technology in semiconductor manufacturing, and particularly to a wafer process fault early warning method and system based on hybrid sequence decomposition and Mamba architecture. Background Technology

[0002] In high-value industrial processes such as semiconductor manufacturing, hundreds or even thousands of sensor signals (such as pressure, flow, and temperature) may be collected in each wafer fabrication cycle, generating complex multivariate time-series data streams. Existing traditional anomaly detection methods struggle to handle such high-dimensional, nonlinear time-series data with long-range dependencies. For example, algorithms such as OneClassSVM and IsolationForest can provide basic anomaly labeling, but they tend to fail when dealing with highly nonlinear data or data with long-term dependencies; Principal Component Analysis (PCA) focuses on the global covariance structure of the data and is not sensitive enough to abrupt or non-stationary drift anomalies.

[0003] While recent deep learning methods (such as LSTM autoencoders, variational autoencoders, and graph neural networks) have improved detection accuracy to some extent, most can only address a subset of characteristics in long-term dependent or multi-sensor correlation scenarios, and suffer from high computational complexity and difficulty in efficient deployment. Traditional state-space models (SSMs) are linear time-invariant systems, while the Selective State-Space Model (SSM) introduced by Mamba upgrades them to input-dependent linear time-varying systems. This means the model can dynamically decide which information to retain and which to ignore based on the current input, similar to the "focusing" in attention mechanisms, while also reducing computational cost. This innovation enables Mamba to achieve linear time complexity when processing long sequences, avoiding the quadratic computational burden of the Transformer self-attention mechanism, and demonstrating significant advantages in tasks involving long time sequences.

[0004] Furthermore, existing solutions often ignore global trends and periodic changes in sensor signals. Directly training the model on the original sequence can easily lead to overfitting of these common patterns, thus missing subtle anomalous signals. At the same time, single-branch structures (such as simple RNNs or attention networks) are difficult to take into account both global and local information, and it is difficult to ensure high sensitivity to both sudden anomalies and slow drifts.

[0005] Therefore, the industry urgently needs a new architecture that can simultaneously balance long-range dependency modeling, sensor correlation feature extraction, and computational efficiency. Summary of the Invention

[0006] To address the challenges of existing technologies in simultaneously handling long-range dependencies, multi-sensor correlations, and non-stationary trends in multi-sensor timing anomaly detection during semiconductor wafer fabrication, this invention proposes a wafer fabrication process fault early warning method and system based on hybrid sequence decomposition and Mamba architecture, thereby improving fault response capabilities and production capacity efficiency in the manufacturing process.

[0007] This invention adopts the following technical solution: a wafer fabrication process fault early warning method based on hybrid sequence decomposition and Mamba architecture, comprising the following steps:

[0008] S1. Preprocess the raw time-series data of the multi-channel sensors of the semiconductor wafer manufacturing equipment, extract the trend component and periodic component of each sensor signal, and obtain the residual sequence.

[0009] S2. Input the residual sequence of the sensor signals from each channel into the dual-branch autoencoder model (TideMamba dual-branch model) for reconstruction; the dual-branch autoencoder model includes a state space branch and an attention mechanism branch, and is equipped with an adaptive gating module for fusing the outputs of the two branches;

[0010] S3. In the state space branch, the hidden state is updated with a learnable linear state space equation according to the residual sequence, and the hidden state is mapped to an output stream of the same dimension as the input.

[0011] S4. In the attention mechanism branch, the residual sequence is encoded using a simplified Mamba self-attention structure to capture the contextual dependencies between each time step and the sensor.

[0012] S5. In the adaptive gating module, a channel-by-channel gating vector is generated based on the splicing result of the current outputs of the two branches, and the outputs of the two branches are weighted and fused.

[0013] S6. The fused output vector is mapped back to the original sensor space through the decoder to obtain the reconstructed signal;

[0014] S7. Calculate the anomaly score based on the error between the reconstructed signal and the original residual sequence. When the anomaly score exceeds the threshold, determine that a fault has occurred in the corresponding wafer process cycle and output a warning message.

[0015] Furthermore, the trend / period decomposition preprocessing uses methods such as moving average or filtering to perform long-period filtering on each sensor data to estimate its slow-changing trend and extract long-term trend components; and identifies periodic components through process cycle analysis or spectrum analysis results to extract the periodic components of the signal.

[0016] The original signal is represented as the sum of trend components, periodic components, and residuals. Only the residual part is retained, the residual sequence is calculated, and then fed into the subsequent model.

[0017] This step removes common slow drift and known periodic effects, allowing the model to focus on genuine anomalous perturbations.

[0018] Furthermore, the dual-branch autoencoder model employs a parallel dual-branch autoencoder structure.

[0019] The first branch is the selective SSM branch, which uses the learnable state transition matrix and the input matrix to perform linear state updates on the temporal residuals. This is equivalent to introducing a hidden state similar to an RNN, which can accumulate long-term information without the gradient decay problem.

[0020] The second branch is the attention mechanism branch, which uses a simplified Mamba self-attention mechanism to encode the input sequence and outputs a feature representation that reflects cross-time and cross-channel correlations.

[0021] The two branches extract information in parallel, each excelling in different scenarios: the selective SSM branch effectively captures stable long-range dynamics, while the attention branch is sensitive to sudden local anomalies and multi-sensor correlations.

[0022] Furthermore, an adaptive gating unit is introduced to perform adaptive gating fusion. A gating vector is automatically calculated based on the current input features, with each component corresponding to a sensor signal, which is used to control the output weights of the SSM branch and the attention branch on that channel.

[0023] Under normal circumstances, the selective SSM branch can reconstruct the signal well, and the gating tends to trust the state-space prediction. However, when some channels experience anomalies, the attention branch can more sensitively capture context-related anomalies. In this case, the gating mechanism dynamically shifts the weights of the corresponding channels to the attention branch, thereby highlighting the anomalous signals. Through this mechanism, the advantages of the two branches are complemented and integrated.

[0024] Furthermore, anomaly reconstruction and scoring are performed, and the fused representation is mapped back to the original signal space through the decoder to calculate the reconstruction error with the actual residual.

[0025] The model is trained on normal data using the mean squared error (MSE) loss function, enabling it to learn the typical patterns of normal data. During the inference phase, an anomaly score is defined based on the magnitude of the reconstruction error at each time step. If the anomaly score exceeds a preset threshold, the corresponding wafer period is considered an anomaly.

[0026] This scheme can effectively utilize the synergistic effect of SSM and attention branch, with low reconstruction error for normal samples and a sharp increase in error for abnormal samples, thus achieving a clear distinction between normal and abnormal samples.

[0027] Furthermore, for unsupervised learning and interpretability, the TideMamba two-branch model does not require anomaly labels during training, using only normal operating data. By analyzing the reconstruction error and the component values ​​of the gating vector, it is possible to intuitively determine which sensor deviated and when, thus achieving interpretable localization of anomalies. For example, when a sensor experiences a surge in error while the corresponding gating value approaches 0 (the gating parameter...),... When the value is approximately 0.9, it indicates that the anomaly is mainly identified by the attention branch, which captures the contextual anomalies associated with other sensors.

[0028] The present invention also provides: a wafer fabrication process fault early warning system based on hybrid sequence decomposition and Mamba architecture, for implementing the aforementioned method, comprising: a data acquisition module, a data preprocessing module, a fault detection module, and an alarm module.

[0029] The data acquisition module is used to collect multi-channel sensor data during semiconductor wafer fabrication.

[0030] The data preprocessing module is used to perform trend / periodic decomposition on the acquired sensor time-series signals to obtain residual sequences;

[0031] The fault detection module includes a dual-branch autoencoder and a gated fusion unit, used to detect anomalies in the input residual sequence and output anomaly scores; the dual-branch autoencoder includes a state space branch and an attention branch, as well as a neural network for synthesizing gated weights;

[0032] The alarm module outputs a fault warning based on the comparison result of the anomaly score and the threshold.

[0033] Furthermore, the data acquisition module includes a low-pass filter for extracting trend components and a periodic analysis unit for extracting periodic components, as well as a differencer for calculating residuals.

[0034] The alarm module uses the anomaly score to make a judgment on each wafer process cycle, and outputs a fault warning notification when an anomaly score higher than the threshold is detected.

[0035] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0036] 1. The method of this invention combines sequence decomposition with a dual-branch fusion autoencoder to perform anomaly monitoring on high-dimensional multi-sensor data from wafer fabrication processes, achieving significantly superior results. Experimental results show that, on actual CVD process datasets, the method described in this invention has higher detection accuracy and efficiency compared to existing advanced methods (such as OmniAnomaly, InterFusion, AnomalyTransformer, and MTAD-GAT graph attention network).

[0037] 2. The gating fusion mechanism of this invention gives the model good interpretability, enabling engineers to specify the sensing channel and time when anomalies occur, which is often difficult to achieve in traditional black-box models; at the same time, it provides an accurate, reliable and easy-to-interpret wafer process fault early warning method, which can significantly improve the fault response capability and capacity efficiency of the production process. Attached Figure Description

[0038] Figure 1 This diagram illustrates the main stages and anomaly types in semiconductor wafer fabrication.

[0039] Figure 2 This diagram illustrates the application background and main application stages of the present invention.

[0040] Figure 3 This is a flowchart illustrating the application of the semiconductor wafer process fault early warning method of the present invention.

[0041] Figure 4 This is an anomaly detection curve of the present invention on a real semiconductor wafer fabrication process dataset.

[0042] Figure 5 This is a schematic diagram illustrating an anomaly detection case of the present invention on a real semiconductor wafer fabrication process dataset. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0044] In existing industrial processes, the main stages and anomaly types of semiconductor manufacturing processes are as follows: Figure 1 As shown. Among them, Figure 1(a) shows the main process stages of wafer manufacturing, which mainly include stage 1 cleaning, stage 2 oxidation, stage 3 deposition, stage 4 photolithography, stage 5 etching and stage 6 inspection. Figure 1 (b) shows typical defects that may occur on the wafer surface or thin film during the manufacturing process, including particle contamination, crack formation, film thickness variation and surface roughness variation; Figure 1 (c) illustrates typical anomaly types at the device structure or cross-sectional level, including thin film inhomogeneities, pattern misalignments, particulate contamination, scratch defects, void defects, and pinhole defects. These anomalies can originate from either sudden local disturbances or slow drifts in process parameters. Therefore, continuous monitoring of multi-sensor time-series data is necessary to enable timely identification and early warning of anomalies at different process stages.

[0045] Based on this, the present invention proposes a wafer fabrication process fault early warning method and system based on hybrid sequence decomposition and Mamba architecture, the application background and main application stages of which are as follows: Figure 2 As shown.

[0046] The application background of the method of this invention includes four aspects: wafer fabrication process, data foundation, fault monitoring model, and subsequent extended applications. First, as... Figure 2 Part a involves continuously collecting process data during wafer fabrication steps such as deposition, photolithography, etching, cleaning, interconnection, testing, and packaging; secondly, such as Figure 2 Part b of the document establishes a database foundation based on a cloud server; then, as... Figure 2 Part c in the document will contain the collected time series samples. to Input the Mamba fault monitoring model and obtain the corresponding output. to And further form feature representation It is used for distinguishing between abnormal and normal states and for fault early warning tasks. Furthermore, such as... Figure 2 In the d part of the model, relevant features can be input into the MLP to establish a film thickness mapping function in model extension applications. Output The results are used to predict wafer film thickness and perform virtual measurement.

[0047] In one embodiment of the present invention, the wafer fabrication process dataset provided is chemical vapor deposition (CVD) process data provided by a semiconductor manufacturing plant. This dataset contains hundreds of different types of sensor channels (including actuator voltage, RF power, chamber pressure, gas flow rate, temperature, position signals, etc.), totaling 1596 dimensions of high-dimensional time-series data, with multiple samples acquired per second. Some channels are analog signals (such as temperature, pressure, and voltage signals), while others are discrete I / O status signals (such as PM2.IO.DI, PM2.IO.DO, etc.). All channel sampled data are normalized (zero mean, unit variance) before being used for modeling.

[0048] It is known that various failure modes exist in the process (such as pressure sensor failure, unstable RF power, valve leakage, etc.). These failures manifest in the data as abnormal fluctuations or systematic drift of individual sensor signals.

[0049] The application scenario of this embodiment is illustrated as follows: Figure 3 The diagram illustrates the overall workflow of multi-sensor data acquisition, preprocessing, and fault diagnosis in wafer fabrication. First, the 1596-dimensional input signal at each sampling time point is preprocessed and decomposed. Then, the decomposed residual sequence is input into the TideMamba dual-branch model, where state-space branching and attention branching are encoded in parallel. After fusion via a gating mechanism, the encoded data is decoded and reconstructed. Finally, the reconstruction error is calculated to obtain an anomaly score, and an alarm is triggered.

[0050] Specifically, Figure 3 Part a shows how the wafer fabrication equipment collects multi-source process information such as air pressure, power, volume, temperature, feedback, and airflow, and forms multi-sensor time sequence 1 to sequence N; Figure 3 Part b shows that the input sequence is processed by the Mamba module in sequence through linear flattening, HP trend filter, feature embedding and trend decomposition. In the trend decomposition Mamba module, Mamba modeling, period estimation, moving average and linearization are further performed. The output is obtained after decomposing and reconstructing the trend term, period term and residual information. Figure 3 Part c in the diagram shows the attention reconstruction branch. The input passes through an attention layer containing Q (query), K (key), and V (value), multi-head attention MHA, Add & Norm, attention reconstruction output, a Mamba module in the form of Selective SSM, layer normalization, connection and gating control, and outputs the reconstruction result.

[0051] In this embodiment, the specific operation steps are as follows:

[0052] S1. Data preprocessing and decomposition: For each sensor signal, perform trend and periodic analysis in the time dimension.

[0053] Step S1.1: Obtain the trend component of the signal by using long window moving average or low-pass filtering, and estimate the periodic component by analyzing the known process cycle or frequency.

[0054] For example, a 1000-second sliding window can be applied to smooth the temperature curve sampled per second, and a pause filter can be used to remove the periodic component of about 60 seconds. The original measurement value can be represented in the form of trend component + periodic component + residual component.

[0055] Step S1.2: Calculate the residual sequence :

[0056] ;

[0057] in, For the input signal, , The sequence consists of trend components and periodic components;

[0058] The residuals mainly contain high-frequency and irregular variations, representing the part most likely to contain fault signals. By removing known stationary drifts and periodic oscillations, the residual sequence is approximately stationary, enhancing the model's sensitivity to anomalous disturbances.

[0059] Specifically, in this embodiment, all 1596 channels undergo the above decomposition process independently to generate residual sequence matrices of the same length. , which serves as the input for subsequent processing of the model.

[0060] S2. Dual-branch autoencoder model structure: residual sequence matrix Divided into multiple lengths The ontology model is input via a sliding window and trained in an unsupervised manner.

[0061] In this embodiment, the dual-branch autoencoder model structure includes parallel SSM branches and attention branches, as well as a subsequent fusion decoding module.

[0062] First, the wafer fabrication equipment collects multi-source process variables such as air pressure, power, volume, temperature, feedback, and airflow to form a multi-sensor time series. Then, the input sequence is linearly flattened, followed by an HP trend filter to separate trend information, and then fed into the Mamba module for trend decomposition, where it undergoes feature embedding, Mamba modeling, period estimation, moving average, and linearization to extract trend terms, periodic terms, and residual information.

[0063] On the other hand, the input data also enters the attention reconstruction branch, where attention weights are calculated using Q, K, and V in the attention layer. The attention reconstruction output is obtained through MHA and Add&Norm, and then combined with the Selective SSM-style Mamba module, layer normalization, connection, and gating control structure to finally output the reconstruction result used for anomaly detection and fault warning.

[0064] S3, SSM Branch Encoder: Referencing the Mamba framework, a state-space branch is defined... Hidden state vector of dimension The state update equation is in linear state-space form:

[0065] ;

[0066] in, for The state transition matrix, for The input matrix, For the first Moment 3D residual input (in this example) ).

[0067] initial state The hidden state can be set as a zero vector or a learnable parameter. Information accumulates over time, analogous to the hidden state of an RNN, but it avoids the vanishing gradient problem due to the use of linear transitions. In the Mamba framework, [the hidden state is...]. As learnable parameters, they are updated during model training, while certain structural constraints are employed to maintain stability.

[0068] Set the output mapping matrix ,Will Map back to the same dimension as the input:

[0069] ;

[0070] in, It can be viewed as the encoding of the system behavior at the current moment by the state space branch.

[0071] S4, Attention Branch Encoder: The attention branch, which runs parallel to the SSM branch, uses a self-attention mechanism to encode the residual sequence.

[0072] will sequence It is embedded into query, key, and value vectors, and information from different time points is aggregated by calculating attention weights.

[0073] In this embodiment, single-head scaling of dot product attention is used:

[0074] ;

[0075] Among them, query ,key ,value ; For trainable projection matrices, For normalization function, For dimensions.

[0076] The attention branch output It can capture the contextual association of sudden events in long-term dependencies. For example, when there is an anomaly that multiple sensors are occurring simultaneously at a certain moment, the attention mechanism can directly connect that moment with historical similar patterns, thereby highlighting its abnormal characteristics.

[0077] S5, Adaptive Gated Fusion: This converts the output of the SSM branch... and attention branch output The final encoded representation is then fused together.

[0078] Step S5.1: Concatenate the two along the channel dimension as input to the gating network, pass them through a small fully connected layer, and then pass them through an element-wise sigmoid function to generate the gating vector. :

[0079] ;

[0080] in, For trainable parameters, Activated for Sigmoid. The Quantity Indicates the time of the first... The degree to which the SSM branch needs to be trusted in the output signal of the sensor.

[0081] Step S5.2, using Weight the outputs of the two branches:

[0082] ;

[0083] in, This is an element-wise multiplication.

[0084] This gating mechanism allows the model to flexibly switch between two encoding methods: when When it is close to 1, it indicates a prediction of the branch of the trust state space; when When the value is close to 0, it becomes more dependent on the contextual information of the attention branch.

[0085] This embodiment observes that, under normal circumstances It tends to favor the SSM branch (smooth model reconstruction), but when anomalies occur... The corresponding position decreases, triggering the attention branch to respond to the anomaly.

[0086] S6. Decoding and Reconstruction: The Merged Representation The decoder maps the data back to the original residual space. The decoder can be implemented as a mapping matrix to the encoder output. The transpose of the same linear projection, or a simple fully connected mapping:

[0087] ;

[0088] in, Can be adopted Or an independently trainable matrix, such that Compared with the original residual They have the same dimensions.

[0089] Specifically, in this embodiment, the TideMamba dual-branch model is trained by minimizing the reconstruction error, and the loss function is... Using the mean square error of all time points and all channels:

[0090] ;

[0091] in, This represents the total number of moments. This represents the total number of channels. For the original residual, This represents the residual after reconstruction.

[0092] During training, normal operating data is primarily used to enable the model to learn normal behavioral patterns.

[0093] In particular, to avoid oversimplifying the decoder, this embodiment may introduce appropriate regularization (such as for...). (Sparse constraints), but in actual experiments, the basic MSE is sufficient for the model to learn stable reconstruction capabilities.

[0094] S7. Anomaly Detection and Early Warning: After the model training is completed, online detection is performed on the newly collected real-time data stream.

[0095] First, the real-time sensor residual sequence is input into the TideMamba dual-branch model using the same preprocessing and windowing methods to obtain the results for each time step. Reconstruction .

[0096] Then, the error distribution is reconstructed based on the pre-statistically obtained normal patterns to determine the anomaly score threshold. For example, let the anomaly score... The normalized value of the sum of squared reconstruction errors for all channels at the current moment:

[0097]

[0098] like Exceeding the threshold (This can be taken as a high quantile of normal error, such as the 99th percentile) then it is considered that the 1st... Anomalies can occur at any time.

[0099] Specifically, to accommodate the practical need for defining anomalies on a batch (cycle) basis in wafer fabrication processes, anomalies are typically defined for all time points within a single operating cycle. Find the maximum or average value to represent the overall degree of abnormality in that period.

[0100] The dual-branch design of this invention reduces errors within the normal period. All are relatively small, and at certain moments during the abnormal cycle It will surge significantly, thus forming a clear peak on the timeline, making it easy to distinguish.

[0101] Example of implementation effect: In this embodiment, the above method was used to conduct experiments on the CVD dataset.

[0102] Specifically, a multivariate time-series dataset based on industrial wafer CVD processes was collected during the semiconductor manufacturing deposition stage to evaluate TideMamba. The raw data contained 1596 recorded variables, which were reduced to approximately 100 independent sensor features after preprocessing. These features cover key process measurements such as temperature, chamber pressure, gas flow rate, RF power / voltage, and various status signals. The dataset contains 2,245 timestamps, sampled at a frequency of 10 Hz, spanning multiple production runs, and was concatenated into a continuous sequence after removing gaps.

[0103] A temporal allocation was used: the first 80% (~1,570 points) was used for training / validation, and the remaining 20% ​​(~675 points) for testing, to maintain temporal dependence and prevent data leakage. Notably, no major anomalies appeared during training, while all 32 anomalies appeared during testing, reflecting the realistic environment of the model being trained on mostly normal historical data and evaluating subsequent anomalies. This setup allows the model to learn normal behavior, while the validation set captures normal reconstruction errors at a set anomaly threshold. Each anomaly in the dataset is labeled with its type (e.g., "pressure surge," "RF power failure," or "gas flow error"). These anomalies are grouped into anomaly categories for detection, but we use the labels to validate whether the model detects all categories.

[0104] To conduct a rigorous and comprehensive evaluation of TideMamba, it was further compared with ten representative anomaly detection baselines: InterFusion, IsolationForest, LOF, LSTM_AE, MTAD-GAT, OmniAnomaly, OneClassSVM, PCA_Recon, USAD, and AnomalyTransformer, covering both classic machine learning methods and state-of-the-art deep neural approaches. All baseline models were implemented in Python or copied from publicly available code, and their hyperparameters were carefully tuned on the validation set. Evaluation was performed at both the point and event levels in a standard binary anomaly detection setting, with each time step classified as normal or anomalous, and all points within a consecutive anomaly segment considered anomalous. Detection quality was quantified using precision, recall, F1 score, and F2 score, with F2 emphasizing recall in scenarios where missing anomalies are particularly costly.

[0105] This embodiment also reports ROC-AUC and PR-AUC as threshold-independent metrics across operation points, with PR-AUC being particularly useful under severe class imbalance, while G-mean reflects the sensitivity and specificity of balance while preventing excessive false positives. Finally, execution times were recorded under the same hardware settings, including average inference time per sequence or per time point, as well as significant differences in training costs and computational requirements, to evaluate the practical applicability of each method in real-time monitoring applications.

[0106] In terms of ranking quality, TideMamba achieved the highest AUC-ROC of 0.989 and the highest PR-AUC of 0.791, surpassing the strongest competing baselines in both metrics. The closest AUC-ROC was achieved by OmniAnomaly (0.986), and the closest PR-AUC by LSTM-AE (0.776). This result is particularly important for imbalanced industrial anomaly detection, as the quality of anomaly ranking is often more informative than the single thresholder operating point. TideMamba also achieved the highest accuracy (0.840), tied with MTAD-GAT (0.840), indicating that its alarms are relatively reliable and less prone to false triggers. Meanwhile, TideMamba maintains competitive point detection performance. Its F1 score reaches 0.737, the highest among all neural baselines. The highest F1 score in the table is 0.746, achieved by Isolation Forest, LOF, and Single-Class SVM, respectively. However, these classic methods significantly underperformed TideMamba in both AUC-ROC and PR-AUC, indicating their competitiveness in the final binary decision-making process during this split, while the overall distribution of outlier scores was less robust. A similar pattern was observed in F2: methods with higher recall rates, such as OmniAnomaly (0.719) or the classic baseline (0.688 recalls), had slightly higher recall-oriented F2 values, while TideMamba favored more precise operating intervals, with an F2 of 0.686.

[0107] Baseline comparisons further revealed different error characteristics. InterFusion was very conservative, with precision = 0.833 but recall = 0.469, meaning that despite fewer false positives, it missed a significant number of anomalies. OmniAnomaly showed the opposite trend: it achieved the highest recall (0.719) but significantly lower precision (0.676), indicating a higher false positive rate. LSTM-AE provided relatively balanced characteristics (precision = 0.759, recall = 0.688, F1 = 0.721), making it the strongest deep baseline overall, but still inferior to TideMamba in the most discriminative ranking metrics. Notably, the addition of the newer AnomalyTransformer baseline further strengthened the comparison; although it achieved AUC-ROC = 0.975 and PR-AUC = 0.746, its threshold performance (precision = 0.636, F1 = 0.646) was still lower than TideMamba. From a deployment efficiency perspective, TideMamba is also very advantageous. Its average runtime is 3.830 seconds per sequence, which is the lowest among the ten baseline methods.

[0108] This result demonstrates that TideMamba's performance improvement does not come at the cost of higher computational cost; on the contrary, the model offers a practical balance between ranking accuracy, alarm reliability, and inference speed. This also fully reflects Mamba's advantage in interacting with hardware-aware parallel algorithms (hardware-aware parallelism, highly efficient parallelization), its relatively high memory efficiency, linear memory consumption, and its ability to effectively handle millions of sequence data points.

[0109] Furthermore, through typical case analysis, the anomaly detection results of this invention on actual wafer fabrication process datasets can be explained from four aspects: overall error, label distribution, anomaly score, and typical variable cases.

[0110] Specifically, the performance on a particular test wafer, such as Figure 4 As shown. Among them, Figure 4 Part a of the diagram shows the contribution of multiple sensor variables to the reconstruction error during the monitoring period, and reveals a significant error concentration phenomenon during abnormal periods. Figure 4 Part b in the figure shows the correspondence between the model's predicted labels and the true labels on the time axis, indicating that the abnormal intervals have good matching. Figure 4 Part c in the figure shows the abnormal score change curve and the threshold position. When an abnormality occurs, the abnormal score rises significantly and exceeds the threshold.

[0111] Three specific typical cases are as follows Figure 5 As shown, the abnormal changes in flow rate, pressure, and temperature variables are respectively illustrated, demonstrating that the present invention can not only identify sudden drops and rapid rises, but also identify abnormal patterns such as continuous drift and pulse fluctuations.

[0112] In a typical anomaly case, the pressure sensor signal briefly drops abnormally, accompanied by fluctuations in gas flow. In the TideMamba two-branch model, the attention branch interprets this anomaly as a "context-related anomaly," thus generating a high anomaly score for that moment.

[0113] In another case, the temperature sensor deviates from the normal range at the end, at which point the prediction error of the state space branch gradually increases, providing an early warning of possible equipment malfunction.

[0114] These results demonstrate that the method of this invention maintains high sensitivity to both sudden anomalies (such as spikes and drops) and slow drift. Furthermore, by analyzing the reconstruction error distribution and gating weights of each sensor, faulty sensors can be intuitively located. For example, visualizations such as time-sensor error heatmaps can assist engineers in diagnosing the root causes of faults.

[0115] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A wafer fabrication process fault early warning method based on hybrid sequence decomposition and Mamba architecture, characterized in that, Includes the following steps: S1. Preprocess the raw time-series data of the multi-channel sensors of the semiconductor wafer manufacturing equipment, extract the trend component and periodic component of each sensor signal, and obtain the residual sequence. S2. Input the residual sequence of the sensor signals from each channel into the dual-branch autoencoder model for reconstruction; the dual-branch autoencoder model includes a state space branch and an attention mechanism branch, and is equipped with an adaptive gating module for fusing the outputs of the two branches; S3. In the state space branch, the hidden state is updated with a learnable linear state space equation according to the residual sequence, and the hidden state is mapped to an output stream of the same dimension as the input. S4. In the attention mechanism branch, the residual sequence is encoded using a simplified Mamba self-attention structure to capture the contextual dependencies between each time step and the sensor. S5. In the adaptive gating module, a channel-by-channel gating vector is generated based on the splicing result of the current outputs of the two branches, and the outputs of the two branches are weighted and fused. S6. The fused output vector is mapped back to the original sensor space through the decoder to obtain the reconstructed signal; S7. Calculate the anomaly score based on the error between the reconstructed signal and the original residual sequence. When the anomaly score exceeds the threshold, determine that a fault has occurred in the corresponding wafer process cycle and output a warning message.

2. The wafer process fault early warning method according to claim 1, characterized in that, The preprocessing described in step S1 involves obtaining the long-term trend component of each sensor signal using a moving average or low-pass filter, and extracting the periodic component of the signal through process cycle analysis or spectrum analysis. The original time series data is represented as a combination of trend component, periodic component, and residual component, and the residual sequence is calculated.

3. The wafer fabrication process fault early warning method according to claim 1, characterized in that, The state space branch described in step S3 includes a state transition matrix and an input matrix, and the hidden states are updated in a linear recursive manner. State updates use linear state-space equations, expressed as: ； in, , For the first Time, Number The dimension of time is The hidden state vector; for The state transition matrix, for The input matrix, This represents the total number of channels. For the first Moment 3D residual input; By outputting the mapping matrix Will Map back to the same dimension as the input: ； in, It is the output of the state space branch.

4. The wafer process fault early warning method according to claim 3, characterized in that, The attention mechanism branch described in step S4 adopts a scaled dot product self-attention mechanism to encode the input residual sequence, calculate attention weights through the learned query, key, and value matrix, aggregate information from different time points, and capture cross-time series and cross-sensor correlations.

5. The wafer process fault early warning method according to claim 3, characterized in that, The adaptive gating module described in step S5 uses a fully connected layer and sigmoid activation to generate gating vectors. : ； in, For trainable parameters, Activated for Sigmoid. The Quantity Indicates the first Time of the first The degree to which the SSM branch needs to be trusted in the sensor's output signal. It is the output of the attention branch; The gate vector Each component is used to control the weighting ratio of the corresponding channel to the state space branch output and the attention branch output during fusion: ； in, For element-wise multiplication, This is the fused vector representation.

6. The wafer process fault early warning method according to claim 5, characterized in that, The decoder described in step S6 represents the fused vector. The residual signals of each sensor are reconstructed by linear transformation. : ； in, Using the output mapping matrix transpose matrix Or an independently trainable matrix, such that Compared with the original residual input Same dimensions; loss function Using the mean square error of all time points and all channels: ； in, This represents the total number of moments. For the original residual, This represents the residual after reconstruction.

7. The wafer process fault early warning method according to claim 5, characterized in that, In step S7, the reconstruction error at each time point is calculated as an anomaly score, and the wafer corresponding to the time point that exceeds the set threshold is marked as an abnormal process; wherein, the threshold is determined based on the high quantile value of the error distribution of normal samples in the training set.

8. The wafer process fault early warning method according to claim 1, characterized in that, The dual-branch autoencoder model uses only normal data to build the model during the training phase, minimizes the reconstruction error of normal samples through the mean squared error loss function, and does not depend on any abnormal labels.

9. A wafer fabrication process fault early warning system based on hybrid sequence decomposition and Mamba architecture, used to implement the method according to any one of claims 1 to 8, characterized in that, include: Data acquisition module: used to collect multi-channel sensor data during semiconductor wafer fabrication; Data preprocessing module: used to perform trend / periodic decomposition on the acquired sensor time-series signals to obtain residual sequences; Fault detection module: includes a dual-branch self-encoder and a gated fusion unit, used to detect anomalies in the input residual sequence and output anomaly scores; The dual-branch autoencoder includes a state space branch and an attention branch, as well as a neural network for synthesizing gating weights; Alarm module: Outputs fault warnings based on the comparison results of anomaly scores and thresholds.

10. The wafer fabrication process fault early warning system according to claim 9, characterized in that, The data acquisition module includes a low-pass filter for extracting trend components, a periodic analysis unit for extracting periodic components, and a differencer for calculating residuals. The alarm module uses the anomaly score to make a judgment on each wafer process cycle, and outputs a fault warning notification when an anomaly score higher than the threshold is detected.

Images