Abnormality detection method for large die forging press

Abstract

The present application relates to the field of die forging press anomaly detection, and specifically to a large die forging press anomaly detection method, the steps of which include: collecting multi-dimensional time series training data of the die forging press, aligning timestamps and processing missing values; dividing the training data into multiple segments through a sliding window, and dividing the die forging press operation stages into five categories; labeling the data segments, calculating the Pearson correlation coefficient matrix, and generating the corresponding correlation threshold matrix; training a neural network classification model based on the training data; collecting the detection data of the die forging press in real time to determine the current operation stage; selecting, calculating and comparing the correlation threshold matrix and the real-time correlation coefficient matrix according to the current operation stage; if the real-time correlation coefficient is lower than the correlation threshold matrix coefficient, triggering an anomaly warning. The present application performs anomaly detection on the sensor correlations at different stages, thereby improving the accuracy of the correlation algorithm anomaly detection and reducing the false alarm and missed alarm rates.

Description

Technical Field

[0001] The present application relates to the field of abnormality detection of die forging presses, and in particular to a method for detecting abnormality of large die forging presses. Background Art

[0002] Die forging presses are essential, critical equipment in the metal forging process, applying immense pressure to shape metal billets into the desired form. Malfunctions in these presses not only disrupt the production process and reduce efficiency, but can also lead to product quality issues such as dimensional inaccuracies and surface defects. Furthermore, equipment failures increase the risk of damage, accelerate wear, and potentially cause significant damage to the machine. More importantly, these situations can trigger safety incidents and endanger the operator's personal safety. Frequent technical failures also increase repair time and costs, impacting the company's profitability.

[0003] Although there are some existing abnormality warning technologies, they still have certain limitations. Many early warning methods rely heavily on the operator's experience, which makes them susceptible to individual differences and lacks objectivity and consistency. Traditional monitoring technologies can often only detect problems after the abnormality occurs and cannot provide foresight. Some monitoring systems cannot provide timely warnings because they cannot collect and process data in real time. Rule-based systems may generate false alarms or miss detections due to improper rule settings. For complex die forging press systems, traditional methods have difficulty identifying atypical abnormal patterns. In addition, current systems lack sufficient predictive maintenance capabilities and fail to make full use of historical data to predict potential future failures.

[0004] Existing anomaly detection methods for die forging presses focus on detecting anomalies in multidimensional time series. However, with the development of the Industrial Internet, sensor devices have accumulated a large amount of industrial time series data. This data features diverse patterns and changing operating conditions, placing higher demands on the efficiency, effectiveness, and reliability of anomaly detection methods. Furthermore, in real-world equipment operation, the correlations between different stages of operation vary significantly. Therefore, existing anomaly detection methods for die forging presses fail to classify the equipment according to its different operating stages, resulting in inaccurate correlation analysis. Summary of the Invention

[0005] In order to solve the problem that the existing die forging press abnormality detection method fails to classify according to the different operation stages of the equipment, resulting in inaccurate correlation analysis,

[0006] The present application provides a method for detecting abnormalities in a large die forging press, comprising the following steps:

[0007] Collect sensor data from the die forging press as multidimensional time series training data, align timestamps, and handle missing values;

[0008] Segmenting the training data into a plurality of segments through a sliding window, and classifying the operating phases of the die forging press into five categories;

[0009] Labeling the data segments of each operation stage, calculating the Pearson correlation coefficient matrix between the sensor variables of each operation stage, and generating a correlation threshold matrix corresponding to each operation stage;

[0010] training a neural network classification model based on the training data;

[0011] collecting detection data of the die forging press in real time and determining the current operation stage through the neural network classification model;

[0012] Selecting a corresponding correlation threshold matrix according to the current operation stage, and calculating a real-time correlation coefficient matrix of the detection data;

[0013] Comparing the real-time correlation coefficient matrix with the correlation threshold matrix, if the absolute value of the element in the real-time correlation coefficient matrix is lower than the absolute value of the element at the corresponding position in the correlation threshold matrix, and the abnormal warning triggering condition is met, the abnormal warning is triggered.

[0014] In a feasible implementation, the five operating stages are: shutdown stage, standby stage, no-load operation stage, pressure application stage, and pressure maintaining stage;

[0015] During the shutdown phase, the sensor data is in a static state;

[0016] During the standby phase, the die forging press is powered on but does not perform any processing;

[0017] During the no-load operation phase, the die forging press performs an ascending action, or the die forging press performs a descending action;

[0018] During the pressure application phase, the die forging press is in a pressurized state, or the die forging press is in a pressure relief state;

[0019] During the pressure holding stage, the die forging press maintains a constant pressure.

[0020] In a feasible implementation, the length of the sliding window is dynamically adjusted according to the working cycle of the device, and the window length is k time points, and k≥5.

[0021] In a feasible implementation, the step of generating a correlation threshold matrix corresponding to each of the operating stages includes:

[0022] Variable pairs with absolute values of correlation coefficients greater than or equal to 0.9 were retained, and the rest were set to zero;

[0023] The retained correlation coefficient was multiplied by 0.85 as the abnormality judgment threshold.

[0024] In a feasible implementation, the neural network classification model is a multi-layer perceptron, and the neural network classification model includes: an input layer and an output layer;

[0025] The dimension of the input layer of the neural network classification model is the same as the number of sensors of the die forging press;

[0026] The output layer dimension of the neural network classification model is 5, corresponding to the five types of operating stages.

[0027] In a feasible implementation, the missing value processing includes: filling the missing sensor data with linear interpolation, or directly setting the missing sensor data to zero.

[0028] In a feasible implementation, the abnormal warning triggering condition is:

[0029] If the absolute value of the element in the real-time correlation coefficient matrix is lower than the absolute value of the element at the corresponding position in the correlation threshold matrix for three consecutive time points;

[0030] Or the magnitude of the absolute value of the element at the corresponding position in the correlation threshold matrix at a single time point is greater than 30%.

[0031] In a feasible implementation, the sensor data includes signal data collected by a pressure sensor, a displacement sensor, a temperature sensor, and a vibration sensor.

[0032] In a feasible implementation, the calculation formula of the Pearson correlation coefficient matrix is:

[0033] ;

[0034] in, are the data values of the two sensor variables at data point i, is the mean of the two variable data; represents the Pearson correlation coefficient between variables x and y; i represents the number of data points.

[0035] In a feasible implementation, the method further includes the steps of:

[0036] The newly collected detection data is regularly added to the training data to retrain the neural network classification model and the correlation threshold matrix.

[0037] The present application provides a method for detecting anomalies in large die forging presses. By performing anomaly detection on the correlation of sensors at different stages, the accuracy of anomaly detection of the correlation algorithm is improved, and the false alarm and missed alarm rates are reduced. Through real-time monitoring and early warning mechanisms, potential equipment failures can be discovered in a timely manner, providing operators with sufficient time to diagnose and troubleshoot the faults, thereby avoiding further expansion and deterioration of the faults. Unnecessary downtime and maintenance costs are reduced, and equipment utilization and production efficiency are improved. By timely discovering and troubleshooting, the risk of equipment failure can be reduced, ensuring the safe and stable operation of the production line. The accuracy and early warning capabilities of anomaly detection are significantly improved, resource utilization is optimized, system security is enhanced, and technological innovation and development are promoted. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings are incorporated into and constitute a part of this specification, illustrate embodiments consistent with the implementation of the present invention, and together with the description, serve to explain the principles of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the implementation of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0039] Figure 1 1 is a flow chart of a method for detecting abnormalities in a large die forging press, exemplified by an embodiment of the present application;

[0040] Figure 2 is a flow chart of a method for generating a correlation threshold matrix corresponding to each operation stage, exemplified in an embodiment of the present application;

[0041] Figure 3 It is a flow chart of a method for detecting abnormalities in a large die forging press, which is exemplified by another embodiment of the present application. DETAILED DESCRIPTION

[0042] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the present invention will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to provide a thorough understanding of the implementation of the example embodiments of the present invention.

[0043] In recent years, anomaly detection technology has been widely studied in fields such as network intrusion detection, fraud identification, and industrial damage monitoring. Its core goal is to identify unusual data values or patterns in data that do not conform to norms, rules, or given models. In time series anomaly detection, the task can be categorized into three types: point anomalies, subsequence anomalies, and pattern anomalies, depending on the detection target. Existing anomaly detection methods primarily include those based on statistical models (such as ARIMA and GARCH), clustering (such as k-means, EM, and SVM), similarity metrics, and rule-based constraints.

[0044] Statistical methods typically use a known sequence distribution and detect anomalies by maintaining a sliding window and calculating statistical feature indicators. This method is suitable for detecting discrete, sudden outliers, but is less effective at identifying persistent anomalous sequence intervals. Clustering-based methods identify outliers by quantifying the distance between anomalous points and normal point clusters. Their detection effectiveness depends on the quality of clustering, and the computational complexity of different clustering models varies. Similarity-based methods determine anomalies by calculating the similarity between standardized sequences, but are time-consuming. Rule-based methods use time series features to repair highly anomalous data, but are difficult to meet the needs of anomaly detection in sequences with variable patterns.

[0045] Despite the availability of numerous anomaly detection methods, practical applications still face numerous challenges. Accurately detecting anomalous data and patterns in single-dimensional time series with highly variable patterns is difficult. While some models have been proposed for multidimensional time series analysis, they have yet to fully explore the mechanisms underlying inter-series correlations. Furthermore, anomaly detection strategies leveraging sequence correlation information have yet to be established, resulting in a lack of efficient, reliable, and interpretable intelligent anomaly detection methods.

[0046] Especially in the context of industrial big data analysis, the amount of labeled data is limited, and the human and material resources required for labeling are high. Therefore, weakly supervised learning methods based on limited labels have become a research direction. However, existing methods still suffer from low detection accuracy and high computational costs when dealing with complex patterns and high-dimensional time series data.

[0047] To address these issues, researchers proposed a multidimensional time series anomaly detection method based on serial correlation analysis. This method leverages correlation knowledge from high-dimensional time series data to accurately identify abnormal data with complex patterns. Experimental results demonstrate that this method performs well in anomaly detection tasks for high-dimensional time series data, outperforming baseline algorithms based on statistical and machine learning models.

[0048] This application is based on a multidimensional time series anomaly detection method to solve the problem that the existing die forging press anomaly detection method fails to classify according to the different operating stages of the equipment, resulting in inaccurate correlation analysis. Figure 1 As shown, the embodiment of the present application provides a method for detecting abnormalities in a large die forging press, comprising the following steps:

[0049] S100: Collect sensor data from the die forging press as multidimensional time series training data, align timestamps, and process missing values.

[0050] During operation, die forging presses generate a large amount of sensor data, such as pressure, temperature, and vibration sensors. This data is the foundation for anomaly detection. This step ensures the integrity and consistency of the training data, providing a reliable data source for subsequent analysis.

[0051] S200: Segment the training data into multiple segments through a sliding window, and classify the operating stages of the die forging press into five categories.

[0052] Sliding window technology can segment continuous data streams into multiple fixed-length segments for easier processing. The division of operating phases is based on the equipment's operational characteristics and process requirements. Converting continuous data streams into processable segments and classifying them by operating phase provides a foundation for subsequent correlation analysis and anomaly detection.

[0053] S300: Labeling the data segments of each operation stage, calculating the Pearson correlation coefficient matrix between the sensor variables of each operation stage, and generating a correlation threshold matrix corresponding to each operation stage.

[0054] The Pearson correlation coefficient measures the degree of linear correlation between two variables. By calculating the correlation coefficient matrix between sensor variables, we can reveal the correlations between variables. This step generates a correlation threshold matrix for each operational phase, which serves as a benchmark for subsequent anomaly detection. Labeling facilitates subsequent classification and identification.

[0055] S400: Training a neural network classification model based on the training data.

[0056] Neural network classification models can learn data features and classify them. Through training, the model can identify the operational phase of test data. This provides operational phase recognition for real-time test data, providing a foundation for subsequent correlation analysis and anomaly detection.

[0057] S500: Collects the detection data of the die forging press in real time and determines the current operation stage through the neural network classification model.

[0058] The detection data collected in real time is identified through a neural network classification model, which can determine the current operation stage of the equipment and provide accurate operation stage information for subsequent anomaly detection.

[0059] S600: Selecting a corresponding correlation threshold matrix according to the current operation stage, and calculating a real-time correlation coefficient matrix of the detection data.

[0060] By selecting the corresponding correlation threshold matrix according to the current operation stage and calculating the real-time correlation coefficient matrix, the correlation between variables in the current detection data can be revealed, and real-time correlation information can be provided for subsequent anomaly detection.

[0061] S700: Compare the real-time correlation coefficient matrix with the correlation threshold matrix. If the absolute value of the element in the real-time correlation coefficient matrix is lower than the absolute value of the element at the corresponding position in the correlation threshold matrix, and the abnormal warning triggering condition is met, the abnormal warning is triggered.

[0062] By comparing the real-time correlation coefficient matrix with the correlation threshold matrix, it is possible to determine whether the current detection data is abnormal. Abnormal data can be discovered and warned in a timely manner, providing timely information support for equipment maintenance and upkeep.

[0063] Existing methods for detecting anomalies in die forging presses fail to classify data based on its operating phases, resulting in inaccurate correlation analysis. This solution, through operational phase classification and correlation analysis, accurately identifies abnormal data with complex patterns. This improves the accuracy of anomaly detection. By classifying and correlating data based on operating phases, it can more accurately identify abnormal data. This avoids false alarms caused by not considering the operating phase. Timely detection and early warning of abnormal data facilitates equipment maintenance and improves equipment reliability and safety.

[0064] In some embodiments of the present application, the five operating stages are: shutdown stage, standby stage, no-load operation stage, pressure application stage, and pressure maintaining stage.

[0065] During the shutdown phase, the die forging press is completely stopped, and all sensor data remains static, with no significant fluctuations. This data is primarily used for baseline calibration, providing a reference standard for subsequent data anomaly detection.

[0066] During the standby phase, the die forging press is powered on but not performing any processing. Sensor data may fluctuate slightly due to equipment warm-up and system self-tests, but overall it remains relatively stable. This data can be used to identify the normal state of the machine, when it is powered on but not operating, and helps distinguish between equipment failure and standby mode.

[0067] During the no-load phase, the die forging press moves up and down, without actually processing the machine. Sensor data changes as the machine moves. This data can be used to assess the machine's operating status under no-load conditions, including wear and lubrication conditions of mechanical components.

[0068] During the pressure phase, when the die forging press is either applying or releasing pressure, sensor data can fluctuate significantly, particularly for sensors related to pressure and displacement. This data can be used to identify critical issues during the process, such as pressure anomalies and displacement deviations.

[0069] During the holding phase, the die forging press maintains constant pressure, and sensor data should remain relatively stable, fluctuating only slightly. This data can be used to test the stability and durability of the pressure control system, as well as the performance of the equipment under extended periods of constant pressure.

[0070] This embodiment, by finely categorizing the die forging press's operating stages, can more accurately identify abnormal data at different stages. This refined categorization of operating stages allows anomaly detection to more closely reflect the equipment's actual operating status, improving detection accuracy. This avoids false positives and missed negatives caused by not considering stage differences, reducing unnecessary maintenance costs and downtime. Furthermore, correlation analysis based on the characteristics of different stages enhances the pertinence and accuracy of anomaly detection.

[0071] In some embodiments of the present application, the length of the sliding window is dynamically adjusted according to the working cycle of the device, and the window length is k time points, and k≥5.

[0072] Sliding windows are a common technique used in time series data analysis. They capture local features of the data by moving a fixed-length window across the data sequence. This window acts like a "sliding" filter, continuously moving as the data sequence progresses, enabling continuous monitoring and analysis of the data.

[0073] The primary function of a sliding window is to extract local features from time series data, facilitating subsequent analysis and processing. By setting windows of varying lengths, data changes at different time scales can be captured. Its functions include data smoothing: averaging or weighted averaging the data within the window can reduce random fluctuations and improve data stability; feature extraction: The data within the window can be used to construct feature vectors for subsequent machine learning model training or anomaly detection; and trend analysis: By comparing data changes within different windows, the overall data trend, such as upward, downward, or stable, can be analyzed.

[0074] Because different devices have varying operating cycles, a fixed-length window may not accurately capture local data features. In this example, dynamically adjusting the window length k ≥ 5 can better adapt to the device's operating cycle and improve analysis accuracy. A window length that is too short may result in information loss and inability to capture complete features, while a window length that is too long may introduce unnecessary noise. Dynamically adjusting the window length k ≥ 5 can ensure information integrity while reducing noise interference.

[0075] In actual operation, a basic window length k_base is first determined based on the typical working cycle of the equipment and data analysis requirements, and k_base ≥ 5 to ensure that the window contains enough data points.

[0076] Monitor the device's operating status in real time, including running, stopped, and standby. Calculate the device's duty cycle T based on changes in the device's operating status. Dynamically adjust the window length k based on the duty cycle T. For example, set a scaling factor α such that k = α * T and k ≥ k_base. Adjusting the value of α further allows for control over the flexibility of the window length.

[0077] Initialize a window of length k at the beginning of the data sequence. As the data sequence progresses, the window is continuously moved, one data point at a time, and the data within the window is updated. At each window position, the characteristics of the data within the window (such as the mean and standard deviation) are calculated for subsequent analysis and processing.

[0078] Through the above practical operation process, this embodiment can realize the dynamic adjustment of the sliding window length, improve the accuracy and flexibility of time series analysis, and provide strong support for the abnormality detection of large die forging presses.

[0079] In some embodiments of the present application, reference Figure 2 As shown, step S300 of generating a correlation threshold matrix corresponding to each operation stage includes:

[0080] S310: retain the variable pairs with absolute values of correlation coefficients greater than or equal to 0.9, and set the rest to zero.

[0081] By setting the absolute value threshold of the correlation coefficient (such as ≥0.9), we can screen out strongly correlated variable pairs and filter out weakly correlated variable pairs, thereby reducing the interference of noise on the analysis.

[0082] S320: Multiply the retained correlation coefficient by 0.85 as the abnormality determination threshold.

[0083] Multiply the retained correlation coefficient by a factor less than 1 (e.g., 0.85) to serve as the threshold for anomaly detection. This can improve the sensitivity of anomaly detection and promptly capture unusual changes in variable relationships. If the real-time correlation coefficient falls below this threshold, it may indicate an unusual change in the relationship between the variables, requiring further investigation.

[0084] The correlation threshold matrix is a matrix used to describe the strength of correlations between variables in multivariate time series data analysis. Each element in the matrix represents the correlation coefficient between two variables, typically ranging from -1 to 1. A larger absolute value of the correlation coefficient indicates a stronger linear relationship between the two variables; a smaller absolute value of the correlation coefficient indicates a weaker linear relationship between the two variables.

[0085] In this embodiment, the correlation threshold matrix primarily quantifies the correlation between variables, providing a basis for subsequent anomaly detection and analysis. By setting thresholds, we can filter out strongly correlated variable pairs and ignore weakly correlated ones, thereby simplifying the problem and improving the accuracy of the analysis.

[0086] In some embodiments of the present application, the neural network classification model is a multilayer perceptron, comprising an input layer and an output layer. The input layer of the neural network classification model has the same dimensions as the number of sensors in the die forging press, and the output layer has a dimension of five, corresponding to the five operating stages.

[0087] A neural network classification model is a mathematical model based on artificial neurons that simulates the structure and function of neural networks in the human brain. It can classify input data and output classification results. A multilayer perceptron is a type of neural network classification model. It consists of multiple fully connected layers, with each hidden layer containing multiple neurons connected by weights. Multilayer perceptrons can learn complex nonlinear relationships and are suitable for classification and regression tasks.

[0088] In this embodiment, the input layer dimension matches the number of sensors, meaning each sensor corresponds to a neuron or set of features in the input layer. This design allows the neural network to directly receive and process raw sensor data without requiring additional feature selection or dimensionality reduction. Each sensor data point contains information about the corresponding operational phase, which is then used for subsequent classification or regression tasks. By matching the input layer dimension to the number of sensors, we ensure that all sensor data is fully fed into the neural network, preserving the data's integrity and original features.

[0089] Furthermore, by ensuring that the input layer can receive and process data from all sensors, the neural network can learn more about the inherent patterns and characteristics of the data, thereby improving the model's accuracy and generalization capabilities. In practical applications, data processing is often a complex and time-consuming process. By matching the input layer dimensions with the number of sensors, the data processing process can be simplified, reducing the workload of data preprocessing and feature engineering, thereby improving the efficiency and responsiveness of the entire system.

[0090] In combination with the large-scale die forging press abnormality detection scenario of this application, when the multi-layer perceptron is used for classification tasks, such as judging the operating stage of the equipment based on sensor data, the input layer dimension matches the number of sensors. By directly inputting the data of each sensor into the input layer of the neural network, the neural network can learn the complex relationship between these parameters and accurately judge the operating stage of the equipment.

[0091] In some embodiments of the present application, missing value processing includes: filling the missing sensor data with linear interpolation, or directly setting it to zero.

[0092] Missing values in sensor data may be caused by device failure, data transmission errors, or inconsistent data collection intervals. If left unaddressed, these missing values can negatively impact subsequent data analysis and model training, leading to inaccurate results or reduced model performance. Filling or removing missing values ensures data integrity and consistency, thereby improving data quality. Missing value processing also reduces noise during model training, improving model accuracy and generalization. The processed data is more complete and reliable, making it suitable for subsequent data analysis and mining tasks.

[0093] In this example, two methods are used to handle missing values. Linear interpolation can estimate the approximate range of missing values based on adjacent known data points, thereby maintaining data continuity. This method is suitable for situations where data changes relatively smoothly and can better restore the true data trend. Linear interpolation is more effective when there are few missing values and adjacent data points are close together.

[0094] Direct zeroing is simple and easy to implement, requires no additional computing resources, and is fast. It's suitable for situations where the number of missing values is small and their impact on the overall data is minimal. Direct zeroing is more suitable and efficient when the dataset contains a large number of zero values or when zero values have a specific meaning (such as indicating an inactive sensor or a downtime device).

[0095] In some embodiments of the present application, the abnormal warning triggering conditions are:

[0096] If the absolute value of the element in the real-time correlation coefficient matrix is lower than the absolute value of the element at the corresponding position in the correlation threshold matrix for three consecutive time points.

[0097] This condition can capture subtle shifts in data correlation trends, helping to identify potential issues early. By monitoring data at continuous time points, it increases the reliability of early warnings and reduces false positives caused by accidental factors. This condition is suitable for scenarios where data correlations are relatively stable and abnormal changes take a long time to manifest. This condition is suitable for scenarios where high early warning accuracy is required and false positives are desired.

[0098] Or, at a single point in time, the absolute value of the element at the corresponding position in the correlation threshold matrix is more than 30% lower. This condition can quickly respond to significant changes in correlation between data and is very effective for quickly identifying anomalies. By setting a magnitude threshold, you can more accurately capture abnormal changes and improve the sensitivity of early warnings. This condition is suitable for situations where data correlations change rapidly and a quick response to anomalies is required, as well as for scenarios where data is sensitive to changes and the sensitivity of early warnings is desired.

[0099] By combining the two aforementioned anomaly warning trigger conditions—consistent time points and amplitude changes—we can more accurately identify abnormal changes in data, reducing false positives and missed alerts. For rapidly changing data, setting amplitude thresholds can quickly trigger warnings, improving system response speed. Accurate warnings enable more efficient resource allocation, focusing on addressing real issues and improving resource utilization efficiency.

[0100] In some embodiments of the present application, the multidimensional time series data includes signals collected by pressure sensors, displacement sensors, temperature sensors, and vibration sensors.

[0101] Multidimensional time series data refers to a collection of data collected by multiple different types of sensors at consecutive time points. This data not only changes over time but also contains information from multiple dimensions at each time point. In the embodiments of this application, the multidimensional time series data specifically includes signals collected by pressure, displacement, temperature, and vibration sensors.

[0102] Pressure sensor signals are used to monitor pressure changes within a device or system, assess the device's operating status, predict failures, and ensure safe operation. Displacement sensor signals provide information on the position and motion of a device or component, helping to analyze parameters such as the device's motion trajectory, speed, and acceleration, thereby determining the device's operating status and performance. Temperature sensor signals reflect temperature changes within the device or environment to monitor the device's thermal state, prevent overheating, and optimize energy consumption. Vibration sensor signals capture vibration information from a device or component, which can be used to identify mechanical failures, assess the device's health, and optimize its operating parameters.

[0103] In this embodiment, by integrating information from multiple dimensions, a more comprehensive understanding of the equipment's operating status can be achieved, improving monitoring accuracy and reliability. Furthermore, abnormal changes in multidimensional time series data can be identified earlier, thereby enhancing the timeliness and accuracy of early warnings. The analysis results based on multidimensional time series data can provide more scientific decision support for equipment maintenance, optimization, and upgrades. Furthermore, by monitoring and analyzing multidimensional time series data in real time, potential problems can be promptly identified and addressed, improving system stability and reliability.

[0104] In some embodiments of the present application, the calculation formula of the Pearson correlation coefficient matrix is:

[0105] ;

[0106] The Pearson correlation matrix is a statistical tool used to measure the degree of linear correlation between multiple variables. In multidimensional time series data analysis, the Pearson correlation matrix can reveal the correlations and trends between different sensor variables. By calculating the Pearson correlation coefficients between each sensor variable, a matrix is constructed in which each element represents the degree of linear correlation between two variables.

[0107] in, represents the Pearson correlation coefficient between variables x and y; Represent the data values of the two sensor variables at data point i respectively; Represent the means of the two variable data respectively; n represents the number of data points.

[0108] Specifically, are the observed values of the two sensor variables at data point i. In practical applications, these values are usually real-time data collected by sensors. These two symbols represent the means of variables x and y, respectively, which is the average of all observations. Calculating the mean is a fundamental step in data analysis, helping to eliminate bias in the data and making the correlation coefficient more accurate. i represents the number of observations. In continuous time series data, i is usually equal to the length of the time series. is the final calculated Pearson correlation coefficient, which ranges from -1 to 1. When it is close to 1, it means there is a strong positive correlation between x and y; when When it is close to -1, it indicates a strong negative correlation; when When it is close to 0, it means there is almost no linear correlation between the two variables.

[0109] By calculating the correlation coefficient, the embodiment of the present application can clearly reveal the linear correlation between different sensor variables, which helps to understand the interactions and changing trends within the system. When a system fails, it is often accompanied by changes in the correlation between multiple sensor variables. By monitoring the changes in the correlation coefficient matrix, anomalies can be discovered in a timely manner, providing strong support for fault warning. By analyzing the correlation coefficient matrix, key variables that have a greater impact on system performance can be identified, and then the system configuration and parameter settings can be optimized to improve system performance and stability. In multidimensional time series data, there may be a large amount of redundant information. By calculating the correlation coefficient matrix, highly correlated variables can be identified, thereby performing data dimensionality reduction and feature extraction, and simplifying the subsequent data analysis process.

[0110] In some embodiments of the present application, reference Figure 3 As shown, the method further comprises the steps of:

[0111] S800: Regularly add newly collected detection data to the training data to retrain the neural network classification model and the correlation threshold matrix.

[0112] Over time, the operating state of the system may change, causing characteristics such as the distribution, range, or correlation of the detection data to shift. Neural network classification models and correlation threshold matrices are trained based on historical data and are capable of capturing regularities and patterns in the data. However, when data characteristics change, the original model may not accurately reflect the new data state, resulting in a decrease in the accuracy of predictions or analysis results.

[0113] Therefore, regularly adding new data to the training set and retraining the model allows the model to adapt to changes in data characteristics and maintain the accuracy of its predictions and analysis. This process is similar to the human process of continuous learning and optimization, which helps to continuously improve the performance and reliability of the system.

[0114] This embodiment, through step S800, ensures that the neural network classification model and correlation threshold matrix always accurately reflect the true state of the data by regularly updating the training set and retraining the model, thereby improving the accuracy of prediction and analysis. As system operating conditions change, regular model updates help improve the system's adaptability, enabling it to better cope with various complex situations. Through continuous learning and optimization, false positives and false negatives caused by model inaccuracies can be reduced, thereby optimizing resource utilization and reducing maintenance costs. This helps improve the overall performance and reliability of the system.

[0115] From the above content, it can be seen that the overall operation steps of the large die forging press abnormality detection method are:

[0116] Large die forging presses are monitored in real time using various sensors, including pressure, displacement, temperature, and vibration, to collect various data during operation. The collected data is cleaned to remove noise, outliers, and other useless information, and the data is standardized.

[0117] Key features are extracted from preprocessed data, and correlations between different features are analyzed using methods such as the Pearson correlation coefficient matrix. Historical data is then used to build anomaly detection models such as neural network classification models or correlation threshold matrices. This training model is then able to accurately identify normal operating and abnormal conditions of equipment.

[0118] Real-time data is fed into the trained model for real-time monitoring. When the model detects an abnormal condition, it immediately triggers an early warning mechanism, alerting the operator. Based on the warning information, the operator diagnoses the equipment fault and takes appropriate measures to correct the problem. Newly collected test data is regularly added to the training set and the model is retrained to maintain its accuracy and effectiveness. Based on actual operation and feedback, the model is optimized and improved to enhance its performance and reliability.

[0119] In summary, the large die forging press anomaly detection method provided by this application can accurately identify the abnormal state of the equipment and reduce the false alarm and missed alarm rates by using multiple types of sensors and advanced anomaly detection models. Through real-time monitoring and early warning mechanisms, potential equipment failures can be discovered in a timely manner, providing operators with sufficient time to diagnose and troubleshoot the faults, avoiding further expansion and deterioration of the faults. Reduce unnecessary downtime and maintenance costs, and improve equipment utilization and production efficiency. By timely discovering and troubleshooting, the risk of equipment failure can be reduced, ensuring the safe and stable operation of the production line. Significantly improve the accuracy and early warning capabilities of anomaly detection, optimize resource utilization, enhance system security, and promote technological innovation and development.

[0120] Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the disclosure of the specification and examples. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in this disclosure.

Claims

1. A large die forging press abnormality detection method, characterized in that: The following steps are involved: Collect sensor data from the die forging press as multidimensional time series training data, align timestamps, and handle missing values; Segmenting the training data into a plurality of segments through a sliding window, and classifying the operating phases of the die forging press into five categories; Labeling the data segments of each operation stage, calculating the Pearson correlation coefficient matrix between the sensor variables of each operation stage, and generating a correlation threshold matrix corresponding to each operation stage; training a neural network classification model based on the training data; collecting detection data of the die forging press in real time and determining the current operation stage through the neural network classification model; Selecting a corresponding correlation threshold matrix according to the current operation stage, and calculating a real-time correlation coefficient matrix of the detection data; Comparing the real-time correlation coefficient matrix with the correlation threshold matrix, if the absolute value of the element in the real-time correlation coefficient matrix is lower than the absolute value of the element at the corresponding position in the correlation threshold matrix, and the abnormal warning triggering condition is met, the abnormal warning is triggered.

2. The large die forging press abnormality detection method according to claim 1, characterized in that: The five types of operating stages are: shutdown stage, standby stage, no-load operation stage, pressure application stage, and pressure maintenance stage; During the shutdown phase, the sensor data is in a static state; During the standby phase, the die forging press is powered on but does not perform any processing; During the no-load operation phase, the die forging press performs an ascending action, or the die forging press performs a descending action; During the pressure application phase, the die forging press is in a pressurized state, or the die forging press is in a pressure relief state; During the pressure holding stage, the die forging press maintains a constant pressure.

3. The large die forging press abnormality detection method according to claim 1, characterized in that: The length of the sliding window is dynamically adjusted according to the working cycle of the equipment. The window length is k time points, and k≥5.

4. The large die forging press abnormality detection method according to claim 1, characterized in that: The step of generating a correlation threshold matrix corresponding to each of the operating stages comprises: Variable pairs with absolute values of correlation coefficients greater than or equal to 0.9 were retained, and the rest were set to zero; The retained correlation coefficient was multiplied by 0.85 as the abnormality judgment threshold.

5. The large die forging press abnormality detection method according to claim 1, characterized in that: The neural network classification model is a multi-layer perceptron, and the neural network classification model includes: an input layer and an output layer; The dimension of the input layer of the neural network classification model is the same as the number of sensors of the die forging press; The output layer dimension of the neural network classification model is 5, corresponding to the five types of operating stages.

6. The large die forging press abnormality detection method according to claim 1, characterized in that: The missing value processing includes: filling the missing sensor data with a linear interpolation method, or directly setting the missing sensor data to zero.

7. The large die forging press abnormality detection method according to claim 1, characterized in that: The abnormal warning triggering conditions are: If the absolute value of the element in the real-time correlation coefficient matrix is lower than the absolute value of the element at the corresponding position in the correlation threshold matrix for three consecutive time points; Or the magnitude of the absolute value of the element at the corresponding position in the correlation threshold matrix at a single time point is greater than 30%.

8. The large die forging press abnormality detection method according to claim 1, characterized in that: The sensor data includes signal data collected by pressure sensors, displacement sensors, temperature sensors and vibration sensors.

9. The large die forging press abnormality detection method according to claim 1, characterized in that: The calculation formula of the Pearson correlation coefficient matrix is: ； in, are the data values of the two sensor variables at data point i, is the mean of the two variable data; represents the Pearson correlation coefficient between variables x and y; i represents the number of data points.

10. The large die forging press abnormality detection method according to claim 1, characterized in that: The method further comprises the steps of: The newly collected detection data is regularly added to the training data to retrain the neural network classification model and the correlation threshold matrix.

Images