Production line fault prediction method and system

Abstract

The invention discloses a production line fault prediction method and system, and belongs to the technical field of intelligent industry, and the method comprises the steps: constructing a virtual production line corresponding to an entity production line based on a digital twinning technology, so as to provide an additional training data set; setting an LSTM-FCN model based on the power spectrum density parameter; and processing production line data output by the entity production line based on the LSTM-FCN model so as to output fault prediction data. The method comprises the following steps: constructing a virtual production line matched with physical characteristics of an entity production line based on a digital twinning technology, and generating virtual data with working condition coverage; the virtual data and the historical data of the entity production line are fused to construct a training set, so that the problem of scarcity of industrial field data is effectively solved; setting an LSTM-FCN model based on the power spectrum density parameter; and processing the production line data output by the entity production line based on the LSTM-FCN model so as to output fault prediction data, thereby accurately predicting the occurrence of faults and reducing the loss caused by shutdown.

Description

Technical Field

[0001] The present invention relates to the field of intelligent industrial technology, and in particular to a production line fault prediction method and system. Background Art

[0002] In modern manufacturing, the efficiency and stability of production lines are key factors in improving a company's competitiveness. In high-end manufacturing, unplanned downtime caused by equipment failures can account for approximately 30% of operating time, significantly impacting production efficiency and product quality. With the rapid development of Industry 4.0 and intelligent manufacturing technologies, data-driven fault prediction and diagnosis methods have become an important approach to solving such problems. Predictive maintenance can reduce unplanned downtime by 30% to 50% through real-time monitoring and analysis of equipment sensor data, significantly improving overall operational efficiency. This shows that timely fault identification and effective staffing are crucial to ensuring the continuous operation of production lines and maximizing output. Existing prediction methods have limited available fault samples, suffer from severe dataset skewness, and slow prediction times. Summary of the Invention

[0003] The present invention proposes a production line fault prediction method and system to solve the problems of inconvenient data collection and slow prediction time in the prior art.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A production line fault prediction method includes: constructing a virtual production line corresponding to a physical production line based on digital twin technology to provide an additional training data set; setting an LSTM-FCN model based on power spectral density parameters; and processing production line data output by the physical production line based on the LSTM-FCN model to output fault prediction data.

[0006] Furthermore, the construction of a virtual production line corresponding to the physical production line based on digital twin technology includes: virtual model construction: based on the key processes of the actual production line, the geometric structure and motion trajectory of the corresponding production line equipment are constructed, and the operation path and sensor data of the production line are output; signal point configuration: based on the signal points of the S7-PLCSIM Advanced virtual PLC controller, to record the sensor data of the production line equipment; logic verification: using TIA Portal software to write and debug the virtual PLC program, and download it to the virtual PLC controller for testing; dynamic simulation and optimization: simulating the operation process of each process of the production line through a virtual environment, generating virtual data to supplement the deficiency of on-site data.

[0007] Furthermore, the setting of the LSTM-FCN model includes: collecting fault data; data preprocessing: detecting missing values and processing them through interpolation or mean filling, detecting and deleting outliers based on the box plot method, and defining fault characteristic variables and fault types related to the fault based on the sensor configuration and sensor data in the production line; data enhancement: resampling the fault data using random window offset; and training the model: training the LSTM-FCN model based on the fault data.

[0008] Furthermore, the collection of fault data includes: determining a value interval of the actual fault data based on the time of the actual fault; sampling the value interval based on a time window to obtain a fault data set, the dimension of the fault data set including the number of samples in the data set, the length of the time series and the number of features; and normalizing the fault data set to obtain the fault data.

[0009] Furthermore, the setting of the LSTM-FCN model based on the power spectral density parameter includes: calculating the power spectral density of the fault characteristic variable from the fault data; analyzing the power spectral density to determine the change rules of different fault types and corresponding power spectral densities to obtain a fault training data set; and training the LSTM-FCN model using the fault training data set to obtain an LSTM-FCN fault alarm model.

[0010] Furthermore, the calculation of the power spectrum density of the fault characteristic variable includes: using the Welch method to calculate the power spectrum density: segmenting the data, applying discrete Fourier transform to estimate the spectrum of each segment, and then averaging the spectrum of each segment to smooth the result.

[0011] Furthermore, the analyzing the power spectral density includes: analyzing the power spectral density based on a kernel-based change point detection method, wherein the kernel function is a radial basis function.

[0012] Furthermore, the optimizer of the LSTM-FCN fault alarm model selects the Adam algorithm.

[0013] Furthermore, the LSTM-FCN fault alarm model performs network search by determining network weight initialization, number of iterations, number of batch samples, and randomly discarding node units.

[0014] Production line fault prediction system, including:

[0015] The first module is used to build a virtual production line corresponding to the physical production line based on digital twin technology to provide additional training data sets;

[0016] The second module sets the LSTM-FCN model based on the power spectral density parameters;

[0017] The third module processes the production line data output by the physical production line based on the LSTM-FCN model to output fault prediction data.

[0018] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0019] 1. The present invention constructs a virtual production line based on digital twin technology that matches the physical characteristics of the physical production line, generating virtual data with comprehensive operating condition coverage. This virtual data is then integrated with historical data from the physical production line to construct a training set, effectively addressing the scarcity of industrial field data. An LSTM-FCN model is configured based on power spectral density parameters. The LSTM-FCN model processes the production line data output by the physical production line to output fault prediction data. This output of fault prediction data can accurately predict the occurrence of faults and reduce losses caused by downtime. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the production line fault prediction method proposed by the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] like Figure 1 The production line fault prediction method shown includes: S1. Building a virtual production line corresponding to the physical production line based on digital twin technology to provide an additional training data set; S2. Setting an LSTM-FCN model based on power spectral density parameters; S3. Processing the production line data output by the physical production line based on the LSTM-FCN model to output fault prediction data.

[0023] Mainstream production lines have a low degree of digitization and insufficient intelligence, so they are not efficient in collecting on-site data and the data samples generated are insufficient. If the data samples are insufficient, the accuracy of the trained model cannot be guaranteed.

[0024] By building a virtual production line through digital twin technology, the physical production line can be converted into a virtual production line. On the one hand, it can simulate the operation of the on-site production line and supplement data samples. On the other hand, it can comprehensively collect on-site data for subsequent analysis and processing.

[0025] The LSTM-FCN model is well suited for this solution. Specifically, the LSTM-FCN model outputs fault prediction data based on changes in power spectral density parameters, supplemented by other potentially related factors.

[0026] The data collected on-site is output as production line data, and the LSTM-FCN model is used to process the output fault prediction data. This can accurately predict the time and type of fault occurrence. Predicting faults allows measures to be taken in advance to reduce the risk and loss of shutdowns.

[0027] The virtual production line corresponding to the physical production line based on digital twin technology includes: virtual model construction: based on the key processes of the actual production line, the geometric structure and motion trajectory of the corresponding production line equipment are constructed, and the operation path and sensor data of the production line are output; signal point configuration: based on the signal points of the S7-PLCSIM Advanced virtual PLC controller, the sensor data of the production line equipment is recorded; logic verification: with the help of TIA Portal software, the virtual PLC program is written and debugged, and downloaded to the virtual PLC controller for testing; dynamic simulation and optimization: through the virtual environment, the operation process of each process of the production line is simulated to generate virtual data to supplement the deficiency of on-site data.

[0028] The setting of the LSTM-FCN model includes: collecting fault data; data preprocessing: detecting missing values and processing them through interpolation or mean filling, detecting and deleting outliers based on a box plot method, and defining fault characteristic variables and fault types related to the fault based on sensor configuration and sensor data in the production line; data enhancement: resampling the fault data using a random window offset; and model training: training the LSTM-FCN model based on the fault data.

[0029] The collecting of fault data includes: determining a value interval of the actual fault data based on the time of the actual fault; sampling the value interval based on a time window to obtain a fault data set, wherein the dimensions of the fault data set include the number of samples in the data set, the length of the time series, and the number of features; and normalizing the fault data set to obtain the fault data.

[0030] The LSTM-FCN model is set based on the power spectrum density parameter, including: calculating the power spectrum density of the fault characteristic variable from the fault data; analyzing the power spectrum density to determine the change rules of different fault types and corresponding power spectrum densities to obtain a fault training data set; and training the LSTM-FCN model using the fault training data set to obtain an LSTM-FCN fault alarm model.

[0031] The power spectrum density of the fault characteristic variable is calculated by using the Welch method to calculate the power spectrum density: segmenting the data, applying discrete Fourier transform to estimate the spectrum of each segment, and then averaging the spectrum of each segment to smooth the result.

[0032] The analyzing the power spectrum density includes: analyzing the power spectrum density based on a kernel change point detection method, wherein the kernel function is a radial basis function.

[0033] The optimizer of the LSTM-FCN fault alarm model selects the Adam algorithm.

[0034] The LSTM-FCN fault alarm model performs network search by determining network weight initialization, number of iterations, number of batch samples, and randomly discarding node units.

[0035] Production line fault prediction system, including:

[0036] The first module is used to build a virtual production line corresponding to the physical production line based on digital twin technology to provide additional training data sets;

[0037] The second module sets the LSTM-FCN model based on the power spectral density parameters;

[0038] The third module processes the production line data output by the physical production line based on the LSTM-FCN model to output fault prediction data.

[0039] To ensure the effective construction of the digital twin system, we first conducted a detailed analysis of the on-site production line architecture. This solution, centered around an intelligent master dispatch system, designed a layered production line control system to achieve efficient data communication and production task scheduling. The intelligent master dispatch system's communication architecture is divided into three layers: the host computer, the communication method, and the slave computer.

[0040] Host Computer: As the decision-making center of the intelligent overall scheduling system, the host computer is responsible for receiving task instructions from the production order system, analyzing and decoding the tasks, and issuing specific work instructions to the slave computers. By dynamically adjusting task priorities and execution order, the host computer optimizes the production process and improves production line efficiency.

[0041] Communication: The system uses TCP / IP to efficiently communicate between the host computer and the MySQL database, transmitting important data such as task status and equipment feedback in real time. The MySQL database is used to store key information during the production process, supporting subsequent scheduling decisions.

[0042] Lower computer: The lower computer system includes the workstation industrial computer system and the AGV control console system. The former is responsible for welding and assembly tasks at each workstation, while the latter coordinates the AGV's material transportation and distribution. Through the linkage between the lower and upper computers, the entire production process is highly automated and digitized.

[0043] Digital twin system modeling is a core component of this solution. By virtualizing the physical production line, it enables intelligent production management that integrates the real and the virtual. Based on virtual simulation technology, it is possible to simulate the production process in a virtual environment and monitor and optimize equipment.

[0044] Virtual Model Construction: Utilizing 3D modeling software, we constructed a virtual model of key processes, including loading, welding, and handling, to visually demonstrate the operational paths and sensor data of production equipment. The model includes the geometry and motion paths of all physical equipment, ensuring high consistency with the actual equipment.

[0045] Signal point configuration: To generate virtual data, signal points for all equipment on the production line are added to the virtual model to record device sensor data. These signal points are generated by the S7-PLCSIM Advanced virtual PLC controller, providing control signals consistent with those on the actual production line.

[0046] Logic Verification: Using TIA Portal software, we wrote and debugged the PLC program and downloaded it to the virtual PLC for testing. This testing verified the correctness of the electronic control logic and ensured the reliability of the virtual model.

[0047] Dynamic Simulation and Optimization: Finally, the system simulates each production line process in a virtual environment, generating high-quality virtual data to complement in-situ data. The system can simulate a variety of scenarios, including normal equipment sensor data and potential failure scenarios, providing rich data support for model training and optimization.

[0048] Fault prediction model design, including:

[0049] 1. Data processing:

[0050] This solution first performs comprehensive data cleaning and feature analysis on the raw data (collected on-site and from a virtual field) to provide a high-quality data foundation for the LSTM-FCN fault alarm model. This includes missing value and outlier detection, and correlation analysis provides a foundation for model design. Missing data is processed using methods such as interpolation or mean-filling to ensure data continuity and consistency. For outliers, this solution uses boxplots to detect and eliminate outliers, thereby preventing the impact of noise on model training. Based on the sensor configuration and sensor data on the production line (determined based on the actual equipment specifications, most simply, whether normal operation or abnormality occurs), feature variables related to fault alarms are defined. These feature variables encompass multiple aspects, from sensor signals to sensor data, including virtual sensor data generated by the digital twin system. These feature variables reflect the sensor data of key equipment in the production process and are an important component of the model input, as shown in Table 1.

[0051] Table 1 Characteristic variables

[0052]

[0053]

[0054] Furthermore, this solution combines the actual operation of the production line with the simulation capabilities of the digital twin system to provide a detailed classification of possible fault types, providing clear output targets for the model. These fault types are derived from both actual fault conditions of physical equipment on site and fault scenarios simulated in the digital twin simulation environment, as shown in Table 2.

[0055] Table 2 Fault types

[0056]

[0057]

[0058] When training a model, directly inputting all data is inefficient. Since production line process data is time-series data, time window sampling is used for data collection. This captures the temporal dynamics of the data, helping the model better understand the time periods and patterns of fault occurrence. In this solution, time window sampling is used to provide a consistent data structure for the fault alarm model and prepare for model training and testing.

[0059] This solution mainly samples a period of time before a fault occurs. The definition of the time window sampling is as follows:

[0060] (1) Assume that the fault data is a time series D = {d1, d2, ..., dn}, where d i Represents the time point t i Define the window size as w, and through time window sampling, the data can be divided into multiple windows W1, W2, ..., W k ,in,

[0061] (2) For overlapping time window sampling, the sampling step is defined as s, and the starting point of the j-th window is t 1+(j-1)×s Starting from the starting point, extract data segments of length w to form a series of overlapping windows: W j ={d i+(j-1)×s |i=1,2,…,w}

[0062] (3) For non-overlapping time window sampling, the starting point is the end of the previous window plus 1. Starting from the starting point, data segments of length w are extracted backward to form a series of time windows: W j ={d i+(j-1)×w |i=1,2,…,w}

[0063] This solution uses time window sampling to generate a fault dataset with dimensions of (b, m, n), where b is the number of samples in the dataset, m is the length of the time series (i.e., the window size), and n is the number of features. Because the values of feature variables vary significantly when data is collected at different time periods, this solution uses MinMax normalization to scale the data values to the (0, 1) range.

[0064] For any feature data x i,j,k (where i represents the sample index, j represents the sequence index, and k represents the feature index), and its value after MinMax normalization is:

[0065]

[0066] Among them, x i,j,k is the original feature value in the data set, min(x ：,：,k ) is the minimum value of the feature in the entire data set, max(x :,:,k ) is the maximum value of the feature in the entire dataset. This ensures consistent scaling across features, eliminates magnitude differences between feature values, avoids model instability caused by a wide range of feature values, and reduces bias during model training and prediction.

[0067] 2. Data Enhancement

[0068] Because the virtual data generated by the digital twin system may differ from the distribution of real-world data in terms of fault points, this difference is primarily manifested in the fixed pattern of fault time points in the virtual data, while in real production lines, the distribution of fault points is random and volatile. This discrepancy can lead to excellent model training performance but insufficient generalization capabilities when faced with real data. To address this issue, this solution proposes strengthening the virtual data by resampling the virtual fault data using a random window offset (RWO) method to enhance the diversity and randomness of the virtual data, thereby making it more consistent with the distribution characteristics of real data. The specific process is as follows:

[0069] Define the range of random offset δ as [-r, r], assuming the start time of the event is t0 and the window radius is w r , then the start and end times after random offset can be calculated as: t start =t0-w r +δ,t end =t0+w r +δ;

[0070] The new data sample obtained is: S={d|t start ≤d.time≤t end}; where d represents a data point in the dataset.

[0071] 3. Fault segment location:

[0072] In complex production line environments, directly analyzing faults based on raw data presents significant challenges, such as data noise and unclear feature changes. To improve fault location accuracy, this solution combines power spectral density analysis with change point detection to gradually process the raw data, ultimately filtering out fault-related time periods for subsequent model training and prediction.

[0073] Power spectral density (PSD) represents the distribution of a signal in the frequency domain and can effectively reveal the periodicity and frequency characteristics of time series data. In production line data, significant changes in PSD are often associated with faults or abnormal equipment behavior. This solution uses the Welch method to calculate PSD: the data is segmented, and then a discrete Fourier transform (DFT) is applied to estimate the spectrum of each segment. The spectrum of each segment is then averaged to smooth the result.

[0074] The power spectral density of each segment is defined as:

[0075] Among them, L is the segment length, U is the normalization factor, Xi (k) is the spectrum of different frequencies, i is the segment index, and k is the frequency index.

[0076] Based on the calculation results of power spectrum density, this solution uses the kernel change-point detection method (KCPD) and uses radial basis function (RBF) as the kernel function to capture the nonlinear change characteristics of the data. Its mathematical definition is: Among them, γ is the parameter that controls the kernel width, ||xy|| 2 Represents the square of the Euclidean distance between two points.

[0077] Analysis of the power spectral density of characteristic variable X9 revealed a significant downward trend before the fault occurred. This phenomenon suggests that the frequency energy density of the characteristic variable had begun to decrease, and the dynamic characteristics of the signal were gradually weakening. Specifically, a decrease in power spectral density may indicate that sensor data has become unstable or that the production line has been interrupted due to certain operations, causing the value of characteristic variable X9 to stagnate. Such stagnation may be an early sign of connector failure. Using a change point detection algorithm can effectively locate the fault segment and reduce the time required for manual fault location.

[0078] 4. LSTM-FCN model establishment:

[0079] This solution uses data from several robot workstations as the training and test sets, with a ratio of 7:3. However, since fault data accounts for a relatively small proportion of actual production data, this leads to data imbalance. This data skew may prevent the model from fully learning fault characteristics during training, thus affecting the model's prediction performance.

[0080] To solve this problem, this solution uses a change point detection algorithm based on the radial basis function (RBF) kernel to process the power spectrum density of production data to determine whether the data segment contains faults and improve the model's ability to learn fault data. The mathematical definition of change point detection is: Among them, γ is the parameter that controls the kernel width, ||xy|| 2Represents the square of the Euclidean distance between two points. In the kernel-based change point detection method (KCPD), the RBF kernel is used to measure the similarity between different points in the time series and evaluate the quality of candidate segmentation points. By using the RBF kernel, KCPD can identify potential change points in time series data to help detect sudden changes in data features. This change point detection method, combined with the model selection strategy, can determine the optimal number of change points, thereby effectively identifying potential fault segments in production data. After completing the change point detection, the screened fault segment data is used as LSTM-FCN for fault classification. The model training process is as follows: data preprocessing, RWO data enhancement, splitting the training set and prediction set, LSTM-FCN model parameter setting, training the model, outputting the parameter file when the epoch is reached, and retraining if it is not reached.

[0081] This solution samples the data set generated by the digital twin system in a time window, filters the fault segment data, and uses the RWO method to strengthen the data set. The strengthened training set is then given to the model for training, and finally the parameter file with better results is saved.

[0082] 5. Results Analysis

[0083] This solution uses several robot workstation fault datasets in the domain verification. The verification dataset has similar scenarios and sensor data as the training dataset. The purpose is to measure the fault detection ability of the model in a known environment.

[0084] During the parameter adjustment process, this scheme performs grid search on the network by determining the network weight initialization, number of iterations, number of batch samples, and randomly discarding node units. The Adam algorithm is selected as the optimizer of the model. The results of the K-fold cross-test of the model are shown in Table 3.

[0085] Table 3 LSTM-FCN model parameter adjustment

[0086]

[0087]

[0088] During the model iteration process, when the number of iterations exceeds 20, the model loss function drops to a lower value and changes slowly. Although the more iterations, the better the prediction effect on the test set, in order to prevent overfitting, this solution decides to use 20 epochs as the number of iterations of the model, and the probability of randomly retaining a node is 0.8.

[0089] The above results were applied to in-domain validation, and precision, recall, and F1 scores were calculated. The results are shown in Table 4. As can be seen from the table, the model performed very well on in-domain data, with both precision and recall remaining at high levels. This indicates that the model is able to effectively detect faults occurring on the production line, and under similar operating conditions, the probability of both false positives and false negatives is low. High precision indicates that the model is able to effectively reduce false positives, while high recall indicates that the model is highly sensitive to fault detection.

[0090] Table 4 Summary of detection accuracy

[0091]

[0092] The above description is a detailed description of the preferred embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A production line fault prediction method, characterized in that: include: Build a virtual production line corresponding to the physical production line based on digital twin technology to provide additional training data sets; Set up the LSTM-FCN model based on the power spectral density parameters; Based on the LSTM-FCN model, the production line data output by the physical production line is processed to output fault prediction data.

2. The production line fault prediction method according to claim 1, characterized in that: The virtual production line corresponding to the physical production line is constructed based on digital twin technology, including: Virtual model construction: Based on the key processes of the actual production line, the geometric structure and motion trajectory of the corresponding production line equipment are constructed, and the operation path and sensor data of the production line are output; Signal point configuration: Based on the signal points of the S7-PLCSIM Advanced virtual PLC controller, to record the sensor data of the production line equipment; Logic verification: Use TIA Portal software to write and debug the virtual PLC program, and download it to the virtual PLC controller for testing; Dynamic simulation and optimization: Simulate the operation process of each process of the production line in a virtual environment and generate virtual data to supplement the deficiency of on-site data.

3. The production line fault prediction method according to claim 2, characterized in that: The LSTM-FCN model is set up, including: Collect fault data; Data preprocessing: Detect missing values and process them through interpolation or mean filling. Detect and delete outliers based on the boxplot method. Define fault characteristic variables and fault types related to faults based on the sensor configuration and sensor data in the production line. Data enhancement: resampling the fault data using random window offsets; Training model: Based on the fault data, train the LSTM-FCN model.

4. The production line fault prediction method according to claim 3, characterized in that: The collecting of fault data includes: Based on the actual fault time, determine the value range of the actual fault data; Sampling the value interval based on the time window to obtain a fault data set, wherein the dimensions of the fault data set include the number of samples in the data set, the length of the time series, and the number of features; Normalization is performed on the fault data set to obtain the fault data.

5. The production line fault prediction method according to claim 4, characterized in that: The LSTM-FCN model is set based on the power spectrum density parameter, including: Calculating the power spectrum density of the fault characteristic variable from the fault data; Analyzing the power spectrum density to determine the change rules of different fault types and corresponding power spectrum densities to obtain a fault training data set; The LSTM-FCN model is trained using the fault training data set to obtain an LSTM-FCN fault alarm model.

6. The production line fault prediction method according to claim 5, characterized in that: The calculating the power spectrum density of the fault characteristic variable includes: The power spectral density is calculated using the Welch method: After the data is segmented, discrete Fourier transform is applied to estimate the spectrum of each segment, and then the spectrum of each segment is averaged to smooth the result.

7. The production line fault prediction method according to claim 6, characterized in that: The analyzing the power spectrum density includes: The power spectral density is analyzed based on a kernel-based change point detection method, wherein the kernel function is a radial basis function.

8. The production line fault prediction method according to claim 7, characterized in that: The optimizer of the LSTM-FCN fault alarm model selects the Adam algorithm.

9. The production line fault prediction method according to claim 8, characterized in that: The LSTM-FCN fault alarm model performs network search by determining network weight initialization, number of iterations, number of batch samples, and randomly discarding node units.

10. A production line fault prediction system, characterized in that: include: The first module is used to build a virtual production line corresponding to the physical production line based on digital twin technology to provide additional training data sets; The second module sets the LSTM-FCN model based on the power spectral density parameters; The third module processes the production line data output by the physical production line based on the LSTM-FCN model to output fault prediction data.

Images