Manufacturing system implicit disturbance intelligent sensing method based on reconstruction error

Abstract

The invention discloses a manufacturing system disturbance sensing method fusing structured errors and probability confirmation, and belongs to the technical field of intelligent manufacturing. The method is characterized by comprising the following steps: firstly, constructing a self-attention long-short-term memory network auto-encoder (SA-LSTM-AE) model, training on data only containing normal working conditions, and deeply learning a healthy operation normal form of a manufacturing system; and then, in a disturbance sensing stage, adopting an innovative composite reconstruction error measure-feature importance weighted mahalanobis distance, and finally, adopting a two-stage confirmation mechanism to carry out disturbance judgment. According to the method, hidden disturbance which cannot be found by a traditional method and is hidden in high-dimensional data can be effectively perceived, recoverable production fluctuation and disturbance which cannot be self-adjusted by the system can be intelligently distinguished, and the disturbance perception sensitivity and the reliability of the system in a real industrial environment are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and artificial intelligence technology, specifically relating to a deep learning-based method for monitoring industrial production processes. More specifically, it is a method that uses a self-attention long short-term memory network autoencoder to perceive disturbances in a manufacturing system and combines interpretability analysis to trace the source of the disturbances. Background Technology

[0002] With the deepening development of Industry 4.0 and intelligent manufacturing, modern manufacturing systems, especially those with fixed process routes and continuous production cycles, such as semiconductor manufacturing and precision assembly lines, are evolving towards high automation, integration, and complexity. However, this complexity also brings new technological challenges: how to effectively perceive and quickly trace various disturbances in the manufacturing process. These disturbances include not only explicit disturbances such as equipment failures and material shortages, but also implicit disturbances such as minute drifts in process parameters, early performance degradation of equipment components, and gradual congestion in production cells. These implicit disturbance signals are weak, cumulative, and hidden beneath massive amounts of high-dimensional data background noise. If they are not detected and intervened in a timely manner, they will gradually accumulate, ultimately seriously affecting product quality, production efficiency, and even equipment lifespan, becoming a key technological bottleneck restricting the improvement of efficiency and the stability of product quality in such manufacturing systems.

[0003] Currently, existing technical approaches for anomaly detection and disturbance perception in industrial manufacturing systems mainly include model-based, knowledge-based, and data-driven methods. Model-based methods identify deviations by establishing precise physical or mathematical models of the production process; however, the construction and maintenance of high-fidelity models are extremely difficult when dealing with modern complex nonlinear manufacturing systems. Knowledge-based methods rely on the knowledge and experience of domain experts for reasoning, but their completeness of knowledge acquisition and generalization ability to unknown disturbance patterns are often limited.

[0004] In recent years, with the improvement of data acquisition capabilities, data-driven methods have shown great potential due to their ability to learn complex patterns directly from massive amounts of data. Among them, unsupervised learning methods that mainly rely on normal operation data for modeling have received widespread attention. In unsupervised methods, autoencoders (AEs) are a commonly used neural network structure. By learning the effective compression representation and reconstruction capabilities of normal data, they can effectively identify any data points that deviate from the learned "normal" pattern. Liu Huihai et al. constructed an autoencoder network model based on vibration signal monitoring for wind turbine gearbox problems. (Detailed reference: Liu Huihai, Zhao Xingyu, Zhao Hongshan, et al. Fault detection of wind turbine gearbox based on deep autoencoder network model [J]. Journal of Electrical Engineering, 2017, 32(17): 156-163.) However, conventional autoencoders have limited ability to process time-series data with rich dynamic information.

[0005] To process temporal data, Long Short-Term Memory (LSTM) networks, due to their unique gated recurrent structure, have shown advantages in capturing and learning long-term temporal dependencies in sequential data and have been applied to construct LSTM autoencoders. Mei et al. successfully used LSTM networks for fault localization and life prediction of aircraft turbofan engines under complex operating conditions. (Detailed reference: Yuan, Mei, Yuting Wu, and Li Lin. "Fault diagnosis and remaining useful life estimation of aero engine using LSTM neural network." 2016 IEEE international conference on aircraft utilitysystems (AUS). IEEE, 2016.) Although LSTM performs excellently in handling temporal dependencies, it may still encounter problems such as information bottlenecks or insufficient attention to key information when dealing with extremely long sequences or when more finely capturing the complex relationships between different parts of the sequence. To enhance the model's ability to capture key information within the sequence, attention mechanisms, especially self-attention, have been introduced into the model. This allows the model to dynamically evaluate and focus on the importance of different parts of the sequence when processing it. Zhang et al. proposed a novel deep learning model, DSA-CNN, which utilizes a dual-channel self-attention mechanism and CNN to detect subtle anomalies in time-series sensors of autonomous vehicles. (Detailed reference: Zhang, Ze, et al. "Time Series Anomaly Detection in Vehicle Sensors Using Self-Attention Mechanisms." IEEE Transactions on Intelligent Transportation Systems (2024).)

[0006] However, both LSTM-based and attention-based deep learning models share a core limitation—their "black box" nature. Their complex nonlinear decision-making processes are difficult for humans to intuitively understand, hindering their widespread adoption in industrial decision-making requiring high reliability and credibility. When disturbances occur in a manufacturing system, simply identifying the disturbance is far from sufficient. For production line operators and managers, understanding the causes of the disturbance—that is, the changes in input features the model bases its judgment on—is crucial. This transparent understanding of the model's decision-making basis is essential for enhancing trust in intelligent systems, quickly and accurately locating the root cause of disturbances, and formulating effective intervention decisions. Therefore, introducing explainable artificial intelligence (XAI) technologies, especially advanced methods like SHAP (SHapley Additive ex Planations) that provide feature-level attribution analysis for complex models, is of paramount importance for improving the practicality, credibility, and ultimately industrial adoption of intelligent disturbance perception systems. Qian et al. have combined SHAP with Support Vector Machines (SVM) for root cause analysis of production anomalies in discrete manufacturing workshops. (Detailed literature: Qian, Weiwei. "ARoot Cause Analysis Method for Abnormal Production in DiscreteManufacturing Workshop." 2023International Conference on Advanced Enterprise Information System (AEIS). IEEE, 2023.)

[0007] Any complex manufacturing system is not simply a collection of independent sensors, but an organic network of tightly coupled and dynamically interacting functional subsystems (such as processing units, material handling, and quality control). A disturbance in any subsystem can gradually spread and even affect the normal operation of the entire manufacturing system. Research on self-encoders, LSTMs, and self-attention mechanisms in industrial manufacturing systems has largely focused on equipment-level fault diagnosis, particularly based on the analysis of physical signals such as vibration and acoustics, achieving significant results. However, existing research generally suffers from a limitation: it confines these tools or algorithms to "microscopic" component-level fault diagnosis, neglecting the possibility of applying them to the higher dimension of "macroscopic" disturbance perception within the manufacturing system. The "disturbance" of a manufacturing system is fundamentally different from the "failure" of a single device: the former is a slow imbalance in the collaborative relationships between multiple variables and subsystems; while the latter is often a physical failure of a single component. No existing technologies have been found to address this problem in manufacturing systems. This patent innovatively transfers and reshapes these technologies to solve the problems of collaboration and balance at the manufacturing system level. Summary of the Invention

[0008] The purpose of this invention is to address the problems of insensitive disturbance perception and opaque model decision-making processes in manufacturing systems by proposing an intelligent perception method for implicit disturbances in manufacturing systems based on reconstruction errors. This method organically integrates the self-attention mechanism of LSTM autoencoder (SA-LSTM-AE) with unsupervised disturbance perception and interpretability tracing technology of SHAP analysis.

[0009] The technical solution of this invention is:

[0010] A smart sensing method for latent disturbances in a manufacturing system based on reconstruction error, characterized by comprising the following steps:

[0011] (1) Obtain multi-dimensional time-series sensor data of equipment from the Manufacturing Execution System (MES), clean and standardize the data, and construct sequence samples using the sliding window technique; construct a self-attention long short-term memory network autoencoder (SA-LSTM-AE) model, which contains an encoder and a decoder; train the model using only data samples under normal operating conditions to enable it to learn the distribution pattern and reconstruction capability under normal operating conditions.

[0012] (2) Input the test sample into the trained SA-LSTM-AE model to obtain the reconstructed sample; based on the feature covariance matrix calculated on normal working condition data and the preset feature importance weight, calculate the weighted Mahalanobis distance between the test sample and the reconstructed sample, and define the distance as its reconstruction error.

[0013] (3) Two-stage disturbance determination mechanism: First, the Gaussian mixture model (GMM) is used to perform probability density analysis on the reconstruction error of normal samples to adaptively set a statistical warning line; when the reconstruction error of the sample to be tested exceeds the warning line for the first time, a disturbance confirmation module is started; the disturbance confirmation module analyzes the time series pattern of the recent reconstruction error sequence and finally confirms whether the system has experienced a real disturbance by probabilistic inference.

[0014] (4) For samples that are determined to have real perturbations, the Gradient Explainer in the SHAP method is used to analyze the determination results of the SA-LSTM-AE model; the contribution of each input feature in the sample to the high reconstruction error is quantitatively calculated, and the key source features that cause the perturbation are determined according to the magnitude of the contribution.

[0015] in:

[0016] (1) Data Acquisition and Preprocessing. This refers to acquiring raw, multi-dimensional time-series data from multi-channel sensors on the production line from sources such as Manufacturing Execution Systems (MES). The dataset is constructed primarily from three dimensions: first, equipment-level data reflecting the physical state of each processing unit, such as real-time power, temperature, and vibration; second, logistics-level data, such as material flow rate and processing waiting time; and third, buffer-level data reflecting the balance between processes, such as capacity utilization and work-in-process quantity. The data undergoes necessary cleaning and standardization. Then, using the sliding window technique, continuous time-series data is constructed into multi-dimensional sequence samples with fixed time steps, and divided into training and testing sets.

[0017] (2) Model construction and training. The model follows an overall encoder-decoder paradigm, but its input layer, internal structure and output layer are designed to achieve deep representation and accurate reconstruction of high-dimensional time series data.

[0018] The model's input layer receives a three-dimensional tensor with dimensions of (batch size, time step, number of features). Here, "time step" corresponds to the width of the sliding window in the above steps, and "number of features" corresponds to the number of multi-dimensional state and performance parameters collected from multiple levels such as equipment, logistics, and buffers.

[0019] The encoder employs a cascaded structure of "temporal key point enhancement" to achieve deep understanding and information compression of the input sequence. First, the input sequence passes through an LSTM layer, which is responsible for capturing the temporal dependencies and dynamic evolution patterns between data points. Subsequently, the output sequence of the LSTM layer is fed into a multi-head self-attention module, which dynamically assigns higher attention weights to time steps containing more critical information by calculating the correlation between time points within the sequence, thereby focusing on key information. Finally, a residual connection fuses the focused information with the original temporal context, and a bottleneck LSTM layer compresses the entire enhanced sequence information into a fixed-length low-dimensional latent vector.

[0020] The decoder is responsible for reconstructing the low-dimensional latent vector generated by the encoder back to the original data space. First, the vector is copied in the time dimension through a repeating vector layer to match the length of the output sequence. Then, the sequence is decoded through one or more LSTM layers to gradually recover its temporal structure. Finally, a fully connected layer is used as the output layer to output a vector with the exact same feature dimensions as the input sequence at a certain time point, completing the reconstruction task.

[0021] Considering the large number of model parameters and the complexity of industrial data, during the training phase, only training set samples under normal operating conditions are used. The loss function is minimized by minimizing the mean square error between the reconstructed value and the original true value of the model output. The optimizer selected is Adam.

[0022] (3) Perturbation Perception. New data samples are input into the trained SA-LSTM-AE model, and its reconstruction error is calculated. However, traditional reconstruction error functions (such as mean squared error, MSE) have a fundamental flaw: they assume that all feature dimensions are independent of each other, and therefore cannot effectively capture the implicit perturbations caused by the disruption of correlations between multiple variables. To address this, this invention designs a novel composite error metric—feature importance-weighted Mahalanobis distance. Instead of calculating simple numerical differences, Mahalanobis distance is used to quantify the statistical distance between the reconstructed value and the true value. The covariance matrix of all features is pre-calculated based on a normal training dataset. This matrix represents the legal cooperative relationship between all process variables in a healthy system. By introducing the inverse of this covariance matrix, Mahalanobis distance can extremely sensitively identify any "structural anomalies" that violate this normal cooperative pattern, even implicit perturbations where the deviation of a single variable is not significant.

[0023] To address the heterogeneity of importance among different parameters, a data-driven feature importance weighting mechanism is introduced. By analyzing the model's performance on a normal validation set, an importance weight is calculated for each feature, which is inversely proportional to the feature's average reconstruction error. This aims to amplify the deviation signals of critical process parameters while suppressing the "noise" generated by normal fluctuations in non-critical parameters.

[0024] Combining the two factors mentioned above, the final reconstruction error is a composite indicator that integrates "structural misalignment" and "critical risk." This indicator not only quantifies the magnitude of the deviation but also profoundly reflects its true impact on the stability of the entire system.

[0025] A two-stage confirmation mechanism is employed for perturbation determination. To avoid the limitations of using fixed or empirical thresholds, this invention introduces a Gaussian mixture model (GMM) to fit the probability density of the reconstruction error distribution of normal training samples, thereby adaptively determining a dynamic, probability-based perturbation warning line.

[0026] Then, when the reconstruction error of the sample under test exceeds the warning line for the first time, the system does not immediately determine it as a disturbance state, but marks the system as a suspicious state and starts a disturbance confirmation module.

[0027] The training process of the disturbance confirmation module involves discretizing the reconstruction error sequence generated by the manufacturing system operating under normal conditions and mapping it to a finite number of observation symbols. This invention employs a discretization method based on K-Means clustering. Specifically, the number of clusters is set to 3. The algorithm automatically divides all error values ​​into three data clusters. Based on the magnitude of the centroids of the three data clusters, a discrete observation symbol V = {v1, v2, v3} is assigned to each cluster, representing the three levels of "steady-state operation," "normal fluctuation," and "disturbance state," respectively. The reconstruction error sequence is then converted point by point into a set of discrete observation symbol sequences. Before training the Hidden Markov Model (HMM), the number of hidden states is defined as 3. These are fed into the discrete observation symbol sequence, and the standard Baum-Welch algorithm is used for training. Through an iterative process of expectation maximization, the state transition probability matrix (A), emission probability matrix (B), and initial state probability distribution (π) are continuously optimized and adjusted. The optimal parameter matrix λ = (A, B, π) contains all the knowledge of the normal fluctuation patterns of the manufacturing system, based on which the probability of production fluctuations returning to normal can be distinguished.

[0028] The perturbation confirmation module employs a pre-trained Hidden Markov Model (HMM) to calculate the posterior probability that the current system is in a "true perturbation state" by analyzing the dynamic evolution patterns of recent reconstruction error sequences. Only when this posterior probability exceeds a preset confidence threshold is the system finally confirmed to have experienced a true perturbation requiring intervention.

[0029] (4) Disturbance source tracing.

[0030] After perturbation confirmation, the system first loads the pre-trained SA-LSTM-AE model and selects a representative set of sequence samples under normal operating conditions as the background dataset. Then, a SHAP gradient interpreter suitable for deep learning models is initialized, using the aforementioned model and background data as input parameters.

[0031] The complete perturbation sample that triggered this alarm is taken as the target instance to be interpreted. The target instance is input into the initialized gradient interpreter, and the contribution of each input feature point to the final output of the model is quantitatively attributed by calculating the gradient of the reconstructed value of the model output with respect to each input feature at each time step, and a SHAP value three-dimensional tensor is output.

[0032] To obtain the overall contribution of each input feature throughout the entire time window, the obtained SHAP value tensors are aggregated. First, their absolute values ​​are taken, and then summed along the time step dimension.

[0033] After the above processing, the system obtains a one-dimensional feature importance vector. Each value in this vector quantitatively represents the overall contribution of the corresponding input feature to the reconstruction anomaly of the SA-LSTM-AE model throughout the entire perturbation period.

[0034] Finally, the system sorts the features in descending order based on their importance vector values ​​and outputs a SHAP summary chart, providing operators with an intuitive and reliable basis for tracing and locating disturbances.

[0035] The beneficial effects of this invention are:

[0036] This invention can improve the detection sensitivity of latent disturbances in manufacturing systems: by using "weighted Mahalanobis distance" as the reconstruction error, this invention can capture early and weak disturbances caused by the disruption of multivariate cooperative relationships that cannot be detected by traditional MSE, and realize the leap from "measuring the numerical value of the deviation" to "understanding the cooperative relationship".

[0037] This invention can significantly reduce the false alarm rate of the system: through the innovative "two-stage disturbance confirmation mechanism", this invention can distinguish between disturbances that can be self-recovered and adjusted in the manufacturing system and disturbances that require disturbance decision-making, effectively reducing the occurrence of false positive alarms.

[0038] This invention realizes intelligent transformation and upgrading from "through algorithms" to "personalized customization": by automatically learning the covariance structure and feature weights of the production data of the manufacturing system through a data-driven approach, it can customize a set of disturbance evaluation standards most suitable for the production characteristics of a specific manufacturing system. Attached Figure Description

[0039] Figure 1 It is an overall flowchart of the disturbance perception and traceability process in the assembly line production system.

[0040] Figure 2 This is a schematic diagram of the network structure of the SA-LSTM-AE perturbation sensing model.

[0041] Figure 3 This is a schematic diagram of the structure of a multi-head self-attention mechanism.

[0042] Figure 4 This is a schematic diagram of the internal structure of an LSTM cell.

[0043] Figure 5 It is the hidden perturbation sample reconstruction error in the buffer.

[0044] Figure 6 This is a summary diagram of the SHAP analysis of the latent perturbation in the buffer zone.

[0045] Figure 7 It is the original characteristic value of the latent perturbation in the buffer. Detailed Implementation

[0046] The invention will be further explained below with reference to the accompanying drawings and a silicon carbide (SiC) wafer production line as an application example. The core characteristic of complex manufacturing systems like silicon carbide wafer processing production lines lies in the high coupling between multiple functional subsystems (equipment, logistics, buffer zones), and the resulting massive high-dimensional data flow. Therefore, issuing an alarm at the time of a detected latent disturbance is only the external manifestation of disturbance detection. Accurately locating the root cause from hundreds of interrelated characteristic variables of the production line is a prerequisite for effective subsequent intervention and decision-making. An efficient and accurate disturbance detection and tracing mechanism has crucial application value for maintaining production stability.

[0047] like Figure 1-7 As shown in Tables 1-3. Table 1 is the relevant parameter table for the SiC production line buffer; Table 2 is the dataset sample setting table; and Table 3 is the model evaluation index comparison table.

[0048] This embodiment studies a disturbance sensing method for a SiC wafer production line.

[0049] The implementation process is as follows Figure 1 It includes the following specific steps:

[0050] Step 1: Data Acquisition and Preprocessing.

[0051] From the MES system of the silicon carbide wafer production line, acquire multi-dimensional time-series data {S} from multi-channel sensors in each buffer zone of the production line (Table 1). i |i=0,1,...,c}, where c represents the number of sensors deployed on the production line. This dataset is constructed primarily from three dimensions: equipment layer data, reflecting real-time data on the physical state of each processing unit, including real-time power, spindle load, and component temperature; logistics layer data, indicators of the efficiency of work-in-process (WIP) flow within the production line, including the rate of wafers entering and leaving each process, AGV utilization and status, and average wafer waiting time; and buffer layer data, representing key indicators of the balance between processes, including the real-time capacity utilization and absolute quantity of WIP in each buffer. A sliding window method is used to convert continuous multi-dimensional time series data into a series of fixed-length input samples. A window size w is set; for each time point t, data from the previous w-1 time points t and the data from the current time point are taken to form an input sequence X. t =[x t-w+1 ,...,x t ], where each x i It is a vector containing c features. The input sequence X t The corresponding target output y t That is, the original feature vector x at the current time point t. t To ensure that every original data point at the beginning of the dataset can be used as a target for reconstruction, the mean of the first w-1 rows of data is used to fill the beginning of the data sequence, so that a complete input window can be formed starting from the first valid point of the original data. The data required for this model training is shown in Table 2.

[0052] Table 1. Relevant parameters for the SiC production line buffer zone.

[0053] Process Name capacity Tumbling × Wire cutting 5 coarse grinding 120 fine grinding 120 chamfer 120 annealing 60 test 60 Wax 60 Mechanical polishing 120 Chemical mechanical polishing 120 Cleaning 60 Detection 60 Package 60

[0054] Table 2 Dataset Sample Settings

[0055]

[0056] Step 2: Model building and training.

[0057] The SA-LSTM-AE perturbation sensing model is constructed, and its specific structure is as follows: Figure 2 As shown.

[0058] The input layer receives a three-dimensional tensor with dimensions (time step, number of features). The encoder is responsible for extracting key features from the input sequence and compressing them. The input sequence first passes through an LSTM layer (...). Figure 4The sequence is then initially encoded using its unit structure. Subsequently, the encoded sequence is fed into a multi-head self-attention module (such as...). Figure 3 As shown, this module dynamically identifies and weights the importance of information at each time point within the input window by calculating the relationships between queries, keys, and values, thereby capturing longer-distance and more complex dependencies. Finally, the sequence features enhanced by the self-attention mechanism are further processed by a bottleneck LSTM layer to compress the sequence information into a fixed-length latent vector z.

[0059] The decoder is responsible for recovering the original input sequence from the latent vector z. Vector z is first expanded temporally using a RepeatVector layer to match the sequence length requirements of subsequent LSTM layers. The expanded sequence is then decoded through one or more LSTM layers to progressively reconstruct the sequence information. Finally, the reconstructed sequence X' is output through a fully connected (Dense) layer.

[0060] The training strategy employs a comprehensive optimization strategy for model training. During the training phase, only training set samples under normal operating conditions are used, and the mean squared error (MSE) is used as the loss function for optimization. For a batch containing N samples, the loss function is MSE, calculated as follows:

[0061]

[0062] In the formula, x i x is the true value vector of the i-th sample. i ′ represents the corresponding reconstructed value vector. The Adam optimizer is used for parameter updates.

[0063] Step 3: Disturbance perception.

[0064] (1) Calculate the innovation reconstruction error

[0065] The perturbation sensing employs a complex intelligent decision-making process to evaluate new data samples. The sample to be tested is input into a trained model to obtain a reconstructed sample. Then, the final reconstruction error is calculated using Mahalanobis distance weighted by feature importance.

[0066] First, after the model training is completed, based on the normal training dataset X... train Calculate the n×n dimensional eigencovariance matrix ∑ and find its inverse matrix ∑ -1 .

[0067]

[0068] Then, the trained model is used to make predictions on the validation set, and the mean reconstruction error (MSE) for each feature i is calculated.val,i And calculate the feature importance weight ω i This forms the weight vector W = [ω1, ω2, ..., ω n ].

[0069]

[0070] For the true value vectors x and x′ of the test sample, the final reconstruction error (squared weighted Mahalanobis distance) is... Calculated using the following formula:

[0071]

[0072] Where d = (xx′) is the difference vector, and ⊙ represents the Hadamard product of element-wise multiplication.

[0073] (2) A two-stage confirmation mechanism is used to determine disturbances.

[0074] In the first stage, a Gaussian mixture model (GMM) is used to perform probability density analysis on the reconstruction error distribution of all normal samples in order to adaptively set a statistical warning line, such as... Figure 5 .

[0075]

[0076] In the formula, K is the number of Gaussian components in the mixture model, and ω k It is the mixing weight (Σω) of the k-th Gaussian component. k =1,ω k ≥0), The mean is μ k The covariance matrix is ​​∑ k The probability density function of the Gaussian distribution.

[0077] In the second phase, the system activates the Hidden Markov Model (HMM) module. This HMM model is trained on the reconstructed error sequence under normal operating conditions and can characterize the error fluctuation patterns of a healthy system. This module is a probabilistic model that models the dynamic behavior of the reconstructed error sequence.

[0078] The model's input is a recent reconstruction error sequence, such as L past time points O = (o1, o2, ..., o...). t The core task of this module is to calculate the total probability, denoted as P(O|λ), of observing the current input sequence O under a pre-trained Hidden Markov Model (HMM) λ that represents "normal behavior." This probability value represents the likelihood that the current error fluctuation pattern is a normal fluctuation. This probability is efficiently calculated using a standard forward pass algorithm. The calculation process is as follows:

[0079] S = {s1, s2, ..., s}N Let} represent the N hidden states within the model, and α1(i) represent the partial sequence (o1, o2, ..., o3) observed at time point t. t And at this time the hidden state is s i The probability of.

[0080] α1(i)=π i b i (o1), i = 1, 2, 3, ..., N

[0081] Where, π i The initial state is s i The probability, b i (o1) is in state s i The emission probability of the first observed symbol o1. For each subsequent time point, the forward variable is recursively calculated:

[0082]

[0083] Among them, a ij From state s i Transition to state s j The transition probability, b j (o t+1 ) is in state s j The symbol o was observed below. t+1 The probability of emission.

[0084] Ultimately, the total probability P(O|λ) of observing the entire sequence O is equal to the sum of the forward variables of all possible hidden states at the last time point L:

[0085]

[0086] The model outputs P(O|λ), which is compared with a preset confidence threshold of 0.95. The threshold can also be between 0.85 and 0.98. Only when P(O|λ) exceeds this threshold is it considered that the production fluctuation will not be absorbed by the buffer, and only then is a real disturbance confirmed and an alarm triggered.

[0087] Step 4: Perturbation Source Identification. When a sample is identified as a perturbation, the SHAP analysis module is activated. The specific solution process of the algorithm is as follows:

[0088] (1) Select a suitable SHAP interpreter. In this invention, GradientExplainer is used for deep learning models.

[0089] (2) Using the SA-LSTM-AE model and some normal data as input, calculate the SHAP value of each input feature in the perturbation sample that causes the current high reconstruction error.

[0090]

[0091] In the formula:

[0092] f: The model to be explained

[0093] x: The input instance to be explained

[0094] F: The set of all features

[0095] S: A subset of features in F that does not contain feature i.

[0096] |S|: The number of features in subset S

[0097] |F|: Total number of features

[0098] f S∪i (x S∪i ): The model's predicted output when feature i is added to feature subset S.

[0099] f S (x S ): The model's prediction output when the feature subset S does not contain feature i.

[0100] For a single prediction, the sum of the SHAP values ​​of all features equals the actual model output for that prediction instance.

[0101]

[0102] Where M is the number of features. This is the baseline prediction value.

[0103] (3) The SHAP value represents the contribution of the feature to the model output (i.e., reconstruction error). A positive SHAP value indicates that the presence of the feature increases the reconstruction error, while a negative value indicates that it decreases the reconstruction error.

[0104] (4) Use visualization tools (such as SHAP summary chart, see...) Figure 6 The graph displays the SHAP values ​​of all features. It clearly shows which features contribute most to disturbance detection. For example, the SHAP value of the capacity utilization sensor for a certain buffer is much higher than other features, directly indicating that the root cause of the disturbance may be related to that buffer. Finally, the buffer data represented by this feature column is visualized to observe changes in relevant parameters. Figure 7 This indicates that a slow, gradual discharge blockage disturbance occurred in the buffer zone, preventing normal discharge and causing the buffer load to gradually increase.

[0105] (5) The performance evaluation index of the method used in this invention is compared with that of AE and LSTM-AE. Overall, the method of this invention performs better, as shown in Table 3.

[0106] Table 3 Comparison of Model Evaluation Indicators

[0107] Model accuracy Recall rate accuracy F1 value AE 0.8429 0.7468 0.9941 0.7919 LSTM-AE 1.000 0.7215 0.9958 0.8382 This invention 0.9851 0.8354 0.9973 0.9041 .

[0108] All parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. An intelligent sensing method for latent disturbances in a manufacturing system based on reconstruction error, characterized by: It includes the following steps: (1) Obtain multi-dimensional time-series sensor data of equipment from the Manufacturing Execution System (MES), clean and standardize the data, and construct sequence samples using the sliding window technique; construct a self-attention long short-term memory network autoencoder (SA-LSTM-AE) model, which contains an encoder and a decoder; train the model using only data samples under normal operating conditions to enable it to learn the distribution pattern and reconstruction capability under normal operating conditions. (2) Input the test sample into the trained SA-LSTM-AE model to obtain the reconstructed sample; based on the feature covariance matrix calculated on normal working condition data and the preset feature importance weight, calculate the weighted Mahalanobis distance between the test sample and the reconstructed sample, and define the distance as its reconstruction error. (3) Two-stage disturbance determination mechanism: First, the Gaussian mixture model (GMM) is used to perform probability density analysis on the reconstruction error of normal samples to adaptively set a statistical warning line; when the reconstruction error of the sample to be tested exceeds the warning line for the first time, a disturbance confirmation module is started; the disturbance confirmation module analyzes the time series pattern of the recent reconstruction error sequence and finally confirms whether the system has experienced a real disturbance by probabilistic inference. (4) For samples that are determined to have real perturbations, the Gradient Explainer in the SHAP method is used to analyze the determination results of the SA-LSTM-AE model; the contribution of each input feature in the sample to the high reconstruction error is quantitatively calculated, and the key source features that cause the perturbation are determined according to the magnitude of the contribution.

2. The method according to claim 1, characterized in that: The specific construction steps of the SA-LSTM-AE model used in step (1) are as follows: (1) The encoder construction steps include: first, feeding the production data sequence into the first long short-term memory network (LSTM) layer; then, feeding the output sequence of the LSTM layer into a multi-head self-attention module, the multi-head self-attention module containing at least 3 attention heads; then, adding the output of the multi-head self-attention module to the original temporal context element by element to perform residual connection fusion, and compressing the fused information into a latent vector of fixed length 32 through a bottleneck LSTM layer; (2) The steps of constructing the decoder include: firstly, the latent space vector is copied in the time dimension through a repeating vector layer to generate a sequence that matches the preset time step; then, the sequence is decoded through at least one LSTM layer; finally, the reconstructed production data sequence is output through a fully connected layer.

3. The method according to claim 1, characterized in that: The step of calculating the composite reconstruction error in step (2) follows the following process: (1) Establish the characteristic covariance matrix of the health status of the manufacturing system: Based on the training set sample containing only the normal operating state, perform statistical analysis to calculate and establish a characteristic covariance matrix that can characterize the inherent synergistic relationship of each process variable under steady-state production. (2) Determine the importance weights of data-driven process variables: Use the trained model to make predictions on an independent validation set, and determine the importance weight of each process variable based on the reciprocal of the average reconstruction error; among them, the variable with the smaller the average reconstruction error is assigned a higher importance weight. (3) Calculate the final weighted Mahalanobis distance: Combine the feature covariance matrix and importance weights obtained in the first two steps to calculate the weighted Mahalanobis distance between the true value vector of the sample to be tested and the vector reconstructed by the model, and use this distance as the final reconstruction error.

4. The method according to claim 1, characterized in that: The disturbance confirmation module in step (3) uses a hidden Markov model (HMM) pre-trained on the reconstruction error sequence under normal operating conditions. The HMM calculates the posterior probability that the current system is in a "disturbance state" by analyzing the reconstruction error sequence, and only confirms that a real disturbance has occurred when the posterior probability exceeds a preset confidence threshold.

5. The method according to claim 4, characterized in that: The preset confidence threshold ranges from 0.85 to 0.

98.

6. The method according to claim 5, characterized in that: The preset confidence threshold is 0.

95.

7. The method according to claim 1, characterized in that: The multi-dimensional time-series sensor data of the equipment includes equipment status, material data, and buffer occupancy rate; the equipment status reflects the physical status of each processing unit: real-time power, temperature, and vibration; the material flow layer data reflects the material flow rate and processing waiting time; the buffer layer data reflects the balance between processes: capacity utilization and work-in-process quantity.

Images