Unsupervised learning-based industrial edge rolling process unknown fault early warning method and system

Abstract

The invention discloses an industrial edge rolling technology unknown fault early warning method and system based on unsupervised learning. The method comprises the steps that multi-dimensional time sequence track data in the operation process of an edge rolling robot is acquired; dividing the data into independent track segments; extracting a group of statistical features covering a centralized trend, a dispersion degree and a distribution pattern from each track segment to construct a high-dimensional feature vector; a sorting method based on feature importance is adopted, a key feature subset with the most discriminative power is screened out from all the features, and therefore optimal feature representation is constructed; training a deep auto-encoder model only by using sample features in a normal state, so that the deep auto-encoder model learns the accurate reconstruction capability of normal feature vectors; during state judgment, a reconstruction error between a new sample feature vector and a model reconstruction result is calculated to serve as an abnormal score; and a decision threshold is scientifically determined through receiver operation characteristic curve analysis and a Youden index maximization criterion, and the abnormal score is compared with the threshold, so that real-time and automatic judgment of the process state is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent manufacturing and industrial artificial intelligence, and particularly relates to a real-time, automatic and high-precision monitoring and abnormal early warning method for a rolling edge process state without relying on abnormal sample labels, based on unsupervised deep learning and data reconstruction principles, by extracting features from multi-dimensional time series data generated during the operation of a robot and deep learning modeling, which is suitable for quality control and predictive maintenance of industrial robots in the body manufacturing process in the case of difficulty in measuring abnormal data. BACKGROUND

[0002] In modern industrial manufacturing processes, the rolling edge process is mainly used for the edge compression of components such as doors, engine covers and trunk lids, and is a key link that determines the appearance quality. The process is usually performed by an industrial robot, which precisely controls the end roller to compress the edge of a sheet metal part along a predetermined trajectory, and the quality directly affects the appearance consistency, sealing and structural strength of the product.

[0003] Currently, the monitoring of the rolling edge quality mainly relies on post-optical detection or manual sampling inspection. Optical measurement refers to sampling and detecting workpieces by a coordinate measuring machine (CMM) or a laser scanning device after production is completed. Although this method has high precision, it cannot cover all products, and the detection results have obvious hysteresis, which cannot provide real-time feedback on abnormalities in the production process, resulting in a high risk of batch quality problems. Manual visual inspection is a visual inspection or tactile judgment of the rolling edge workpiece by experienced workers. This method is highly subjective, low in efficiency, and prone to missed detection or misjudgment due to fatigue or negligence.

[0004] Another problem that cannot be ignored is that although intelligent monitoring methods based on supervised learning have been proposed, in actual production, normal state samples account for the vast majority, the probability of abnormal state occurrence is extremely low and the types are diverse, making it difficult to collect a sufficient number of labeled samples to train a high-robustness classification model, which seriously restricts the application practice of supervised learning methods. In addition, accurate labeling of fault data requires deep domain knowledge, and manual labeling is costly and inconsistent. Therefore, there is an urgent need for a technology that can achieve efficient anomaly detection without relying on abnormal sample labels, using only a large amount of easily accessible normal process data, to break through the limitations of supervised learning in application scenarios. SUMMARY

[0005] The present application aims to solve the problem of dependence on abnormal state label data in supervised learning and provides an anomaly detection method that only requires normal state data for modeling.

[0006] To achieve the above objectives, this invention provides a method for early warning of unknown faults in industrial hemming processes based on unsupervised learning. The core of this method lies in its independent use of labeled data from abnormal states, instead employing a deep learning model to fully learn the feature distribution patterns composed solely of data from normal states. When new production data deviates from the learned normal pattern, it is identified as an anomaly. Essentially, it compares the differences between the input data and the model-reconstructed data. The core ideas of this method are consistent with those of claim 1, summarized as follows:

[0007] Data acquisition, processing, and feature vector construction steps: Similar to mainstream data analysis methods, first acquire and process time-series data, and then extract structured feature vectors.

[0008] Model training steps: A deep autoencoder model is trained using only feature vectors representing normal process conditions. The goal of the model is to learn how to accurately reconstruct the normal input feature vectors.

[0009] State determination steps: For new trajectory data, the model attempts to reconstruct its feature vectors. If the new data is normal, the reconstruction effect will be good, and the reconstruction error will be low; if the new data is abnormal, because its pattern is different from the normal pattern learned by the model, the reconstruction effect will be poor, and the reconstruction error will be high. By comparing this error with a preset threshold, it can be determined whether the current process is in an abnormal state.

[0010] It is important to emphasize that the training process of the core model in this invention is completely unsupervised and does not rely on any anomaly labels. The final decision threshold is determined through supervised calibration using a validation set containing a small number of known samples. This hybrid strategy of "unsupervised training + supervised calibration" solves the problem of insufficient anomaly samples during training while ensuring the objectivity and optimality of the judgment criteria during final deployment, representing the best combination of theory and practice.

[0011] Furthermore, the present invention also provides a system for implementing the above method, and a computer-readable storage medium for storing related computer programs.

[0012] A method for early warning of unknown faults in industrial hemming processes based on unsupervised learning includes:

[0013] S1: Data acquisition, obtaining multi-dimensional time-series trajectory data of the robot end effector for multiple work cycles in the hemming process;

[0014] S2: Data segmentation, which uses a periodic detection algorithm to automatically segment the multidimensional time-series trajectory data into trajectory segments corresponding to a single, independent rolling edge task;

[0015] S3: Feature extraction, for each of the trajectory segments, a set of statistical features is calculated, and the set of statistical features is combined into a feature vector, wherein the statistical features comprehensively cover multiple quantities describing the trend, dispersion degree and distribution form of the data set, so as to capture and establish the complex mode and distribution boundary of the normal state in the richer data features;

[0016] S4: Model training, a training data set containing only feature vectors corresponding to the normal process state is provided, a deep learning model based on reconstruction is trained using the training data set, the deep learning model based on reconstruction is a variational autoencoder, the deep learning model based on reconstruction is trained to minimize the reconstruction error between its input and output, and a trained deep learning model is obtained;

[0017] S5: State judgment:

[0018] (a) The newly acquired trajectory segment is used to obtain a new feature vector by steps S1-S3, and the feature vector is input into the trained deep learning model to obtain a reconstructed feature vector;

[0019] (b) The reconstruction error value between the new feature vector and the reconstructed feature vector is calculated, and the reconstruction error is obtained by calculating the Euclidean distance between the two;

[0020] (c) The reconstruction error value is compared with a preset decision threshold to determine the state of the new trajectory segment, wherein the decision threshold is set based on the reconstruction error distribution on the normal state verification set.

[0021] Further, the multi-dimensional time sequence trajectory data in the S1 step is derived from at least two robots in different physical positions, so as to construct a state discrimination model that can be applied to multiple robots.

[0022] Further, the robot work cycle in the S1 step is defined by a cycle signal in the control program, one cycle corresponds to the completion of a single edge rolling task of a workpiece, and the multi-dimensional time sequence trajectory data obtained in each cycle includes a three-dimensional coordinate (X, Y, Z) sequence of the center of the robot flange, an angle sequence of the six joints, and a torque sequence of the six joints.

[0023] Further, in the S3 step, the statistical features include: mean, maximum, minimum, 25% quantile, 75% quantile, standard deviation, variance, range, interquartile range, coefficient of variation, skewness, and kurtosis.

[0024] Further, in the S3 step, to improve the efficiency and robustness of model training, after the feature vector is constructed, at least one filtering-based feature selection algorithm is used to reduce the dimension of the feature vector, and the filtering-based feature selection algorithm is selected from variance filtering, correlation filtering, t-test or Fisher score.

[0025] Further, in the S4 step, the hyperparameters of the variational autoencoder include the number of network layers of the encoder and the decoder, the dimension of the hidden layer and the dimension of the latent space, and the reconstruction error performance of the model on normal samples is evaluated through five-fold cross-validation.

[0026] Further, when the deep learning model based on reconstruction in the S4 step is a variational autoencoder, the loss function includes a reconstruction loss and a KL divergence loss term for learning the probability distribution of normal state data in the latent space.

[0027] Further, in the (c) step of S5, the decision threshold is determined by ROC curve analysis based on the maximum criterion of Youden index from the validation set containing normal and abnormal samples; when the reconstruction error value in S5(b) exceeds the decision threshold, it is judged as an abnormal state.

[0028] An industrial rolling process unknown fault early warning system based on unsupervised learning, comprising a processor and a memory, the memory storing computer executable instructions, when the processor executes the instructions, the system is configured to:

[0029] A data acquisition module for collecting multi-dimensional time series trajectory data of the robot end effector in the rolling process;

[0030] A data processing module for dividing the multi-dimensional time series trajectory data into trajectory segments corresponding to single and independent rolling tasks;

[0031] A feature vector construction module for calculating a set of statistical features for each trajectory segment and combining the set of statistical features into a feature vector, wherein the statistical features include at least quantities describing the trend, dispersion and distribution of the data set, and further configured to use a filtering-based feature selection algorithm to reduce the dimension of the initial feature vector;

[0032] A model training module for training a deep learning model based on a training data set containing only normal process state feature vectors, and the training module of the deep learning model is further configured to:

[0033] 1) Use Bayesian optimization algorithm to automatically optimize the hyperparameters of the deep learning model, including the number of network layers of the encoder and the decoder, the dimension of the hidden layer and the dimension of the latent space;

[0034] 2) Training is performed using a loss function that includes reconstruction loss and KL divergence loss to minimize reconstruction error;

[0035] The state discrimination module is used to load the trained deep learning model, and obtain a new feature vector from the newly acquired trajectory segment through steps S1-S3. The feature vector is then input into the trained deep learning model to obtain a reconstructed feature vector. The Euclidean distance between the new feature vector and the reconstructed feature vector is calculated as the reconstruction error value and compared with the decision threshold to output the judgment result of the trajectory state. The decision threshold is determined from the validation set through ROC curve analysis and based on the Yoden exponent maximization criterion.

[0036] The alarm and linkage control module is used to perform at least one of the following operations when it receives the classification results of abnormal status: triggering an alarm on the human-machine interface, automatically marking the current workpiece as to be re-inspected, or sending control commands to the production line central control system (PLC).

[0037] Compared with the prior art, the present invention has the following significant advantages:

[0038] 1. Real-time online monitoring: By analyzing process data in real time, quality judgment can be shifted from post-inspection to in-process monitoring, enabling the immediate detection of process drift or sudden anomalies, effectively reducing the generation of defective products and achieving predictive maintenance.

[0039] 2. Reduced inspection costs and complexity: This invention does not require additional expensive optical inspection hardware, but makes full use of the data generated by the robot controller itself, which is a low-cost, non-invasive, and highly efficient quality control solution.

[0040] 3. Reduced dependence on fault samples: This invention uses unsupervised learning, which only requires normal production data for modeling. It solves the common problem of difficulty in collecting and labeling fault samples in industrial applications and lowers the application threshold.

[0041] 4. Improved accuracy and robustness: By transforming dynamic time-series data into stable statistical feature vectors and combining them with the powerful non-linear feature learning capabilities of deep autoencoders, it can accurately capture the boundaries of normal states and also has good detection capabilities for previously unseen anomaly types.

[0042] 5. Provide data-driven decision support: The model can not only determine the state, but also quantify the severity of the anomaly by the magnitude of its reconstruction error, providing data support for process optimization and fault diagnosis, and promoting the intelligentization of the production process.

[0043] 6. Capable of detecting unknown faults: Unlike supervised models that can only identify fault types already present in the training set, this invention discovers anomalies by learning the boundaries of normal patterns. This means that any trajectory deviating from the normal state, regardless of whether its fault mode has ever occurred, can theoretically be detected. This ability to detect "zero-day faults" has immeasurable value for ensuring production line safety and rapidly responding to new problems. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart of an unknown fault early warning method for industrial hemming process based on unsupervised learning according to Embodiment 1 of the present invention;

[0046] Figure 2 Box plots and scatter plots of the reconstruction error of the model dataset provided in Embodiment 2 of the present invention;

[0047] Figure 3 This is an ROC curve analysis diagram provided according to Embodiment 2 of the present invention;

[0048] Figure 4 This is a block diagram of a rolling edge trajectory state judgment system based on unsupervised learning.

[0049] Figure 5 This is a schematic diagram of a system for determining the state of rolling edge data trajectory based on a supervised learning model according to Embodiment 3 of the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that the following detailed description is illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0051] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0052] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0053] Example 1: A method for determining the state of automobile rolling edge data trajectory based on an unsupervised learning model

[0054] like Figure 1 As shown, this embodiment provides a method for determining the trajectory status of vehicle rolling edge data based on an unsupervised learning model, including the following steps:

[0055] Step S101: Data acquisition. Obtain multi-dimensional time-series trajectory data of one or more work cycles of the robot end effector in the hemming process. The data can be acquired in real time at a preset frequency or read in batches from a historical database.

[0056] The multidimensional time-series trajectory data includes at least:

[0057] The sequence of three-dimensional coordinates (X, Y, Z) of the robot end flange center in the world coordinate system or tool coordinate system.

[0058] Angle sequence of the robot's six joints (J1-J6).

[0059] Torque sequence of the robot's six joints (mtq1-mtq6).

[0060] To improve the model's generalization ability and robustness, data needs to be collected from robots at at least two different physical locations on the production line, and this data should be used together for subsequent model training. The model trained in this way will not overfit to minor characteristics of a specific hardware component and can be more widely applied to multiple workstations.

[0061] Step S102: Data segmentation. The original acquired data is a continuous time-series stream. By using an automated periodic detection unit to identify specific signals in the robot control program, the original data can be segmented into trajectory segments corresponding to a single, independent rolling task. For example, the process from the start of rolling to the end of rolling the edge of the door cover constitutes a trajectory segment.

[0062] Step S103: Feature extraction. Directly using raw time-series trajectory segments of varying lengths as model input faces the problems of dimensionality curse and high computational complexity. Therefore, this step calculates a set of statistical features that comprehensively describe the intrinsic attributes of each segmented trajectory segment, and combines these features into a high-dimensional feature vector with fixed dimensions. The statistical features may include, but are not limited to:

[0063] Measures of central tendency: mean, maximum, minimum, 25th percentile (Q1), 75th percentile (Q3).

[0064] Measures of dispersion: Standard Deviation, Variance, Range, Interquartile Range, and Coefficient of Variation.

[0065] Distribution morphology measures: Skewness and Kurtosis. Skewness can be used to detect the asymmetry of the trajectory, while kurtosis can be used to detect the steepness or smoothness of the trajectory. These are closely related to the robot's operating state.

[0066] Because the extracted features have a high dimensionality, this invention further introduces a feature selection step to remove redundant and irrelevant features, improving the efficiency and accuracy of the model in learning normal pattern distributions in the latent space. This step aims to select the most effective subset of features for state judgment from the high-dimensional features, thereby reducing model complexity and preventing overfitting. A filtering-based feature selection algorithm can be used, for example:

[0067] Variance filtering: Remove features with small variance across all samples.

[0068] Relevance filtering: Remove features that are low in relevance to the target label.

[0069] t-test or Fisher score: A statistical test method used to assess the discriminative power of a single feature.

[0070] Step S104: Model Training. The core of this step is to train a deep learning model, allowing it to learn the inherent patterns and distribution of the feature vectors of a normal rolling edge trajectory. The training process is completely different from traditional supervised classification:

[0071] 1. Constructing the training set: Only collect and use trajectory data that has been validated as "normal" in production. Process these trajectory segments into feature vectors according to S103, and further divide them into training and validation sets.

[0072] 2. Model Selection and Optimization: A deep learning model was chosen. This is a neural network consisting of an encoder and a decoder. The training objective is to make the reconstructed vector as consistent as possible with the original vector, i.e., to minimize the reconstruction error. To obtain the best model performance, a Bayesian optimization algorithm was used to automatically optimize the network structure parameters and training process parameters. A Dropout term was added to ensure model robustness and prevent overfitting. Cross-validation was used to evaluate the model's reconstruction error on the normal validation set, and the model with the smallest error was selected.

[0073] 3. Final Training: Using the optimized combination of hyperparameters, train the deep learning model on a training dataset with all normal states. Use the Adam optimizer to update the network weights based on the gradient of the loss function. After each or every few training epochs, evaluate the model's loss on the validation set. Finally, select the model that achieves the lowest loss value on the validation set as the fixed state discrimination model that can accurately reconstruct normal data.

[0074] Step S105: Status Determination

[0075] After the trained deep learning model is deployed to the production site, the system can perform real-time online status assessments. The core of this step is to transform the continuous reconstruction error values ​​output by the model into discrete "normal" or "abnormal" states with clear business implications, using a scientifically determined decision threshold. The detailed process is as follows:

[0076] 1. Data Reconstruction: For a newly acquired hemming trajectory from the production line, the system first executes steps S101-S103 in real time to generate its feature vector. Next, this new feature vector must be standardized using the mean and standard deviation saved during model training in step S104. This is a crucial step to ensure that the data used for online judgment follows the same distribution as the data used during model training. Subsequently, the standardized feature vector is input into the trained deep learning model to obtain a reconstructed feature vector.

[0077] 2. Calculate the reconstruction error: The system calculates the numerical difference between the original, standardized new feature vector and the reconstructed feature vector output by the model, using this as a quantitative indicator of the anomaly score. In this embodiment, the reconstruction error value is specifically obtained by calculating the Euclidean distance between the two. A high reconstruction error value indicates that the features of the current trajectory significantly deviate from the normal pattern learned by the model.

[0078] 3. Threshold Determination: To ensure the accuracy and robustness of the determination, the preset decision threshold is not subjectively set based on experience, but is determined offline before model deployment through a data-driven optimization process. This process utilizes a validation or test dataset containing known labels, and the specific steps are as follows:

[0079] (a) Error and label registration: Prepare a dataset containing normal and abnormal samples, and assign a true label (0 represents normal and 1 represents abnormal) to the reconstruction error of each sample.

[0080] (b) ROC Curve Generation: Iterate through all possible thresholds, calculating the True Positive Rate (TPR) and False Positive Rate (FPR) at each threshold to classify samples as anomalous. Plot the ROC curve with FPR on the x-axis and TPR on the y-axis. The area under the ROC curve can be used to evaluate the overall discriminative ability of the model.

[0081] (c) Optimal Threshold Selection: To maximize anomaly detection while minimizing false positives for normal samples, this embodiment uses the Youden index as the selection criterion for the optimal point. The Youden index is defined as J = TPR - FPR. The threshold corresponding to the Youden index at each point on the ROC curve is calculated, and the threshold corresponding to the point with the largest value is selected as the optimal decision threshold. In addition to the Youden index, alternative optimization criteria include finding the threshold closest to the upper left corner (0,1) of the ROC curve, or the threshold that maximizes the balance between accuracy and error.

[0082] The optimal decision threshold determined by the ROC optimization method described above is fixed in the system. For the reconstruction error value of the new trajectory calculated in step 2, the system performs a final comparison and judgment: if the reconstruction error value is greater than the optimal decision threshold, the trajectory is judged as "abnormal state". If the reconstruction error value is less than or equal to the optimal decision threshold, it is judged as "normal state".

[0083] Through this series of steps, the present invention not only realizes the judgment from data to state, but more importantly, it ensures that the decision threshold is scientifically set based on the statistical characteristics of data and the principle of optimal classification performance, thereby guaranteeing the high accuracy and reliability of the entire method in practical applications.

[0084] Example 2: A method for determining the status of automotive helix data trajectory using a variational autoencoder.

[0085] This embodiment provides a method for determining the trajectory status of vehicle rolling edge data using a variational autoencoder (VAE). Building upon Embodiment 1, this method further introduces a variational inference mechanism, enhancing the model's ability to characterize normal data distributions and its robustness in anomaly detection through latent variable modeling.

[0086] 1. Data preparation and feature selection

[0087] Multidimensional time-series trajectory data was collected from the controllers of the hemming robot at two different physical locations. This included the three-dimensional coordinates (X, Y, Z) sequence of the robot's end flange center, and the angle and torque sequences of the six joints. The continuous time-series data was automatically segmented into independent trajectory segments according to the work cycle, with each segment corresponding to one complete hemming task. Fifteen predefined statistical features were extracted for each variable, forming a 15*12=180-dimensional feature vector. A multi-stage feature selection strategy (variance filtering, correlation filtering, t-test, and Fisher score) was employed to select the 40-dimensional feature subset with the highest discriminative power from the high-dimensional features, thereby reducing model complexity and avoiding overfitting.

[0088] 2. VAE Model Construction

[0089] Constructing a variational autoencoder with an encoder-decoder structure:

[0090] Encoder: Receives a 40-dimensional feature vector as input, extracts high-level features through two fully connected layers (dimensions 512 and 256 respectively), and finally outputs the mean vector μ and the log-variance vector log(σ²) of the latent space, both with a dimension of 32.

[0091] Reparameterization: Generating latent variables from the latent distribution through sampling.

[0092]

[0093] Where ε is standard Gaussian noise, ensuring that the model is differentiable.

[0094] Decoder: Maps the latent variable z back to a 40-dimensional reconstructed feature vector. The network structure consists of two fully connected layers (dimensions 256 and 512), and the final output layer uses a linear activation function.

[0095] 3. Model Training

[0096] Using only feature vectors from the normal state as the training set, the dataset is divided into 80% training samples and 20% validation samples. The Adam optimizer (learning rate 0.0005) is used to minimize the following loss function:

[0097] L = Reconstruction loss + β.KL divergence loss

[0098] The reconstruction loss is the mean squared error (MSE), the KL divergence loss is used to constrain the latent distribution to approximate a standard Gaussian distribution, and the hyperparameter β is used to balance reconstruction accuracy and the regularity of the latent space distribution. In this embodiment, it is set to 1.0 through cross-validation to ensure that the model learns a continuous and complete normal state latent space while effectively compressing information, thereby enhancing its sensitivity to anomalies. Early stopping (patience = 20 rounds) and Dropout (0.1~0.2) are used during training to prevent overfitting. Finally, the model with the lowest validation set loss is selected as the state discrimination model, and the discrimination results are as follows: Figure 2 As shown.

[0099] 4. State determination and threshold optimization

[0100] After extracting identical 40-dimensional features from newly acquired trajectory segments and standardizing them, the features are input into a trained VAE model to calculate the reconstruction error. The optimal decision threshold is determined on the validation set using the Youden exponent maximization criterion through ROC curve analysis. This threshold is used to transform continuous reconstruction errors into a binary classification result of "normal / abnormal". The system outputs the status judgment in real time and triggers alarms or linkage control signals when an anomaly is detected, such as... Figure 3 As shown.

[0101] Finally, it should be noted that the method steps of the present invention can also be solidified into a computer program and stored on a computer-readable storage medium, such as a hard disk, solid-state drive, USB flash drive, or cloud server storage space.

[0102] 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.

[0103] Example 3: A Rolling Edge Process Status Monitoring System Based on Unsupervised Learning and Its Implementation Process

[0104] This embodiment provides a specific system architecture and workflow for implementing the unsupervised monitoring method of the present invention in a real production environment, such as... Figure 5As shown. At the physical hardware level, this system is fully compatible with mainstream supervised monitoring systems, sharing the same edge computing gateway, industrial server, and human-machine interface (HMI). However, its internal software logic, data processing flow, and model operation and maintenance mode fully embody the unsupervised core idea of ​​this invention: "learning normal operation and detecting anomalies."

[0105] 1. System Hardware Architecture and Deployment

[0106] The physical deployment of the system is consistent with the aforementioned supervised learning scheme implementation, and also includes three parts: a data acquisition terminal, a model service and processing center, and an application alarm terminal.

[0107] Data acquisition terminal: An industrial edge computing gateway is deployed next to the control cabinet of the rolling robot to collect multi-dimensional time-series data such as the three-dimensional coordinates, joint angles, and joint torques of the robot end effector in real time and at high frequency through the industrial Ethernet protocol.

[0108] Model Service and Processing Center: High-performance industrial servers are deployed in the workshop as the core for data storage, model training, and real-time computing.

[0109] Application alarm terminals: Deploy visual front-ends on the production line HMI or the large screen in the central control room.

[0110] 2. System Workflow

[0111] The workflow of this unsupervised system is clearly divided into two independent but closely related stages: the offline model training and threshold calibration stage and the online real-time monitoring stage.

[0112] (a) Offline model training and threshold calibration phase

[0113] The core objective at this stage is not to distinguish between different categories of samples as in supervised learning, but to construct a mathematical model that represents only the "absolutely normal" state and to scientifically define the boundary between "normal" and "abnormal".

[0114] Step 1: Constructing the Normal Dataset. Unlike supervised learning methods, which require significant effort to find and label various fault samples, this approach involves engineers collaborating with process experts to meticulously select data from historical production data that represents only normal, stable, and high-quality production conditions, constructing a clean, normal dataset. This process strictly excludes data generated during equipment startup, shutdown, maintenance, batch replacement, and any periods known to have experienced process fluctuations.

[0115] Step Two: Feature Engineering and Deep Learning Model Training. The system extracts comprehensive high-dimensional statistical features from each independent trajectory segment in the normal dataset, covering multiple dimensions such as central tendency, dispersion, and distribution pattern. Subsequently, algorithms such as variance filtering and correlation analysis may be used to remove redundant features. Finally, the model training module uses only these clean normal feature vectors to train a deep learning model. The goal of the training is to enable the deep learning model to reconstruct the input normal features as accurately as possible, i.e., to minimize the reconstruction error on the normal validation set.

[0116] Step 3: Scientific Calibration of the Optimal Decision Threshold. After the deep learning model is trained, the system cannot be used immediately because it only knows how to reconstruct data, but doesn't know what "reconstruction error" constitutes an "anomaly." Therefore, a crucial threshold calibration step is needed. Engineers will prepare a small-scale validation dataset containing known normal and anomaly labels. The system will have the trained deep learning model reconstruct all samples in this validation set and calculate the reconstruction error for each sample. Then, this is executed via an automated script:

[0117] 1. Using each possible reconstruction error value as a threshold, calculate the corresponding true positive rate (TPR) and false positive rate (FPR).

[0118] 2. Plot the ROC curve and automatically find the point on the ROC curve that maximizes the Youden exponent.

[0119] 3. The reconstruction error value corresponding to this point is calibrated by the system as the "optimal decision threshold" and fixed in the online monitoring module.

[0120] (b) Online real-time monitoring phase

[0121] This stage is responsible for performing real-time, automated status checks.

[0122] Step 1: Data Acquisition and Feature Extraction. The edge gateway captures the complete trajectory data of a new edge-rolling task in real time and sends it to the server. The server performs the exact same feature extraction process as in the offline phase, generating a high-dimensional feature vector.

[0123] Step Two: Data Standardization and Reconstruction. The system uses the mean and standard deviation of the normal dataset saved during the offline training phase to standardize the newly generated feature vectors. Then, the standardized vectors are input into the loaded deep learning model to obtain a reconstructed vector.

[0124] Step 3: Anomaly Score Calculation. The state judgment module calculates the Euclidean distance between the input standardized vector and the model output reconstructed vector, and uses this distance value as the anomaly score to quantify the degree of anomaly in the current trajectory.

[0125] Step 4: Status Judgment and Linkage. The system compares the calculated anomaly score with the optimal decision threshold calibrated in the offline phase: if the score is lower than or equal to the threshold, the system determines the current workpiece status as "normal" and only logs it; if the score is higher than the threshold, the system immediately determines it as "abnormal" and triggers the alarm and linkage control module, popping up an alarm window on the HMI, automatically marking the current workpiece as pending re-inspection, or sending a pause command to the production line PLC.

[0126] Through the above process, this embodiment fully demonstrates how unsupervised methods can achieve efficient and automated monitoring of the rolling edge process status on the same hardware platform as supervised methods, through completely different data processing philosophies and workflows, especially highlighting its advanced nature of not relying on anomaly labels and scientifically set judgment criteria.

[0127] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An unsupervised learning-based method for early warning of unknown faults in an industrial serging process, characterized in that, The method comprises: S1: data collection, obtaining multi-dimensional time sequence trajectory data of a plurality of working cycles of a robot end effector in a rolling process; S2: data segmentation, automatically segmenting the multi-dimensional time sequence trajectory data into trajectory segments corresponding to single, independent rolling tasks through a periodic detection algorithm; S3: feature extraction, for each trajectory segment, calculating a set of statistical features, and combining the set of statistical features into a feature vector, wherein the statistical features comprehensively cover a plurality of quantities describing the trend, dispersion degree and distribution form of the data set, so as to capture and establish the complex mode and distribution boundary of the normal state in a more rich data feature; S4: model training, providing a training data set containing only feature vectors corresponding to the normal process state, training a reconstruction-based deep learning model using the training data set, the reconstruction-based deep learning model being a variational autoencoder, training the reconstruction-based deep learning model to minimize the reconstruction error between its input and output, and obtaining a trained deep learning model; S5: state judgment: (a) obtaining a new feature vector using steps S1-S3 for a newly obtained trajectory segment, and inputting the feature vector into the trained deep learning model to obtain a reconstructed feature vector; (b) calculating the reconstruction error value between the new feature vector and the reconstructed feature vector, the reconstruction error being obtained by calculating the Euclidean distance between the two; (c) comparing the reconstruction error value with a preset decision threshold to determine the state of the new trajectory segment, wherein the decision threshold is set based on the reconstruction error distribution on the normal state validation set.

2. The method of claim 1, wherein, The multi-dimensional time sequence trajectory data in the S1 step is derived from at least two robots in different physical positions to construct a state discrimination model that can be applied to multiple robots.

3. The method of claim 1, wherein, The robot working cycle in the S1 step is defined by a periodic signal in the control program, and one cycle corresponds to the completion of a single rolling task for a workpiece. The multi-dimensional time sequence trajectory data obtained in each cycle includes a sequence of three-dimensional coordinates (X, Y, Z) of the center of the robot end flange, a sequence of angles of the six joints, and a sequence of torques of the six joints.

4. The method of claim 1, wherein, In the S3 step, the statistical features include: calculating twelve statistical features for each original variable, including mean, maximum, minimum, 25% quantile, 75% quantile, standard deviation, variance, range, interquartile range, coefficient of variation, skewness, and kurtosis, to form a feature vector.

5. The method of claim 1, wherein, In the S3 step, to improve the training efficiency and robustness of the model, at least one filtering-based feature selection algorithm is used to reduce the dimensionality of the feature vector after the feature vector is constructed, and the filtering-based feature selection algorithm is selected from variance filtering, correlation filtering, t-test or Fisher score.

6. The method of claim 1, wherein, In the S4 step, the hyperparameters of the variational autoencoder include the number of network layers of the encoder and the decoder, the dimension of the hidden layer, or the dimension of the latent space, and the reconstruction error performance of the model on normal samples is evaluated through five-fold cross-validation.

7. The method of claim 1, wherein, When the reconstruction-based deep learning model in the S4 step is a variational autoencoder, the loss function comprises a reconstruction loss and a KL divergence loss term for learning the probability distribution of normal state data in the latent space.

8. The method of claim 1, wherein, The decision threshold in the S5 (c) step is determined from a validation set comprising normal and abnormal samples through ROC curve analysis based on a Youden index maximization criterion; when the reconstruction error value in S5 (b) exceeds the decision threshold, it is determined to be an abnormal state.

9. An unsupervised learning based industrial serging process unknown fault early warning system, characterized in that, The system comprises a processor and a memory, wherein the memory stores computer executable instructions, and the processor executes the instructions to configure the system to: a data acquisition module for collecting multi-dimensional time series trajectory data of a robot end effector in a rolling process; a data processing module for segmenting the multi-dimensional time series trajectory data into trajectory segments corresponding to single, independent rolling tasks; a feature vector construction module for calculating a set of statistical features for each trajectory segment and combining the set of statistical features into a feature vector, wherein the statistical features at least include quantities describing the trend, dispersion, and distribution pattern of the data set, and are further configured to use a filter-based feature selection algorithm to reduce the dimensionality of the initial feature vector; a model training module for training a deep learning model based on a training data set comprising only feature vectors of normal process states, wherein the model training module is further configured to: 1) use a Bayesian optimization algorithm to automatically optimize the hyperparameters of the deep learning model, including the number of network layers of the encoder and decoder, the dimension of the hidden layer, and the dimension of the latent space; 2) use a loss function comprising a reconstruction loss and a KL divergence loss to train the deep learning model to minimize the reconstruction error; a state discrimination module for loading the trained deep learning model and obtaining a new feature vector for a newly acquired trajectory segment using steps S1-S3, inputting the feature vector into the trained deep learning model to obtain a reconstructed feature vector, and outputting a judgment result of the trajectory state by calculating the Euclidean distance between the new feature vector and the reconstructed feature vector as a reconstruction error value and comparing it with a decision threshold, wherein the decision threshold is determined from a validation set through ROC curve analysis based on a Youden index maximization criterion; an alarm and linkage control module for performing at least one of the following operations when receiving the classification result of the abnormal state: triggering an alarm on a human-machine interaction interface, automatically marking the current workpiece as pending review, or sending a control instruction to a production line central control system (PLC).

Images