A fault detection method and system for underwater pelletizers

Abstract

This application relates to the field of data processing technology, and in particular to a fault detection method and system for underwater pelletizers. The method includes: acquiring preprocessed multi-dimensional equipment operation data; for any data point, constructing a condition change index reflecting changes in the operating conditions at the time of the data point; in response to the condition change index being greater than a preset change threshold, adjusting the initial distance between the current data point and other data points within a long-term data window based on the condition change index to obtain an optimal distance; and using a local outlier factor algorithm to detect anomalies in the data points based on the optimal distance. This application effectively reduces false alarms during the detection process of underwater pelletizers.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a fault detection method and system for underwater pelletizers. Background Technology

[0002] Underwater pelletizers are key equipment in the production of polymer materials (such as plastics and resins). The blades of an underwater pelletizer cut molten material into uniform pellets in an underwater environment. Due to the harsh working environment—constantly subjected to high temperature, high pressure, water erosion, and mechanical vibration—faults such as blade wear, bearing aging, and material blockage frequently occur, seriously affecting product quality and production safety. Therefore, real-time and effective fault detection of underwater pelletizers is crucial.

[0003] Currently, fault detection methods based on sensor data and anomaly detection algorithms have attracted attention. Among them, the Local Outlier Factor (LOF) algorithm is a classic unsupervised anomaly detection algorithm that identifies anomalies by comparing the density of data points with their neighborhoods. It can detect unknown faults without requiring fault samples, which has significant advantages. However, when processing real-time data streams acquired by sensors, the standard LOF algorithm requires calculation on all historical data, resulting in a huge computational burden that is difficult to meet the real-time requirements of industrial applications.

[0004] To address this issue, the industry has proposed an incremental LOF algorithm based on a sliding window. This method performs calculations only within a finite-sized sliding window, and only updates the LOF value of the affected data points when a new data point enters or an old data point leaves the window, thus significantly improving computational efficiency. However, existing incremental LOF algorithms typically use a fixed, pre-set window size. This exposes serious technical flaws in practical applications: the operating conditions of underwater pelletizers are dynamically changing, including multiple stages such as startup, stable operation, load adjustment, raw material switching, and shutdown. During these changes, naturally sparse regions are created, and the data points in these regions are not abnormal, leading to false alarms in traditional LOF detection algorithms. Summary of the Invention

[0005] To reduce false alarms in traditional underwater pelletizer detection methods, this application provides a fault detection method and system for underwater pelletizers.

[0006] Firstly, this application provides a fault detection method for an underwater pelletizer, employing the following technical solution: A fault detection method for an underwater pelletizer includes: acquiring preprocessed multi-dimensional equipment operation data; for any data point, constructing a condition change index reflecting changes in the operating conditions at the time the data point is located; in response to the condition change index being greater than a preset change threshold, adjusting the initial distance between the current data point and other data points within a long-term data window based on the condition change index to obtain an optimal distance; and using a local outlier factor algorithm to detect anomalies in the data points based on the optimal distance. The steps for constructing an operating condition change index that reflects changes in operating conditions at the time of the data point include: constructing a short-term observation window based on the data point within a preset long-term data window; analyzing the fluctuation degree of the data in the short-term observation window and the fluctuation degree of the data in the long-term data window to construct a relative fluctuation index; analyzing the correlation of data in each dimension within the short-term observation window and the correlation of data in each dimension within the long-term data window to construct a correlation change index; and using the ratio of the relative fluctuation index to the correlation change index as the operating condition change index.

[0007] For any data point, analyze the operating condition change index at the corresponding time. When the operating condition change index exceeds a certain threshold, the current operating condition is considered to have changed, resulting in sparse data point areas during the change. These data points are not truly outliers; therefore, in this case, adjust the initial distance between other data points and the current data point to obtain the optimal distance. Based on this optimal distance, use the local outlier factor algorithm to detect data points, reducing false positives.

[0008] The construction of the operating condition change index is primarily based on data changes over a short period. For these short-term changes, the degree of data fluctuation and the correlation between different data dimensions are used to reflect whether operating conditions have changed during the underwater pelletizer's operation. During normal equipment operation, the data across different dimensions maintain a certain correlation; that is, normal operating condition changes do not cause significant changes in the correlation between these dimensions. Therefore, the ratio of the relative fluctuation index to the correlation change index is used as the operating condition change index.

[0009] Optionally, the steps for constructing the relative volatility index include: obtaining a short-term volatility index that reflects data fluctuations in a short-term observation window and a long-term volatility index that reflects data fluctuations in a long-term data window, and using the absolute difference between the short-term volatility index and the long-term volatility index as the relative volatility index.

[0010] By calculating the absolute difference between short-term and long-term fluctuations, the degree of abnormality in current data fluctuations can be determined more accurately. If the instantaneous fluctuations are large but the long-term fluctuations are also consistently large, it indicates that this may be a normal high-fluctuation condition.

[0011] Optionally, the steps for obtaining the short-term volatility index, which reflects the data fluctuations in the short-term observation window, include: two adjacent data points in the short-term observation window form an adjacent point pair; the Euclidean distance between the two data points in each adjacent point pair is obtained; the mean of the Euclidean distances corresponding to each adjacent point pair is taken as the short-term volatility level; and the normalized result of the short-term volatility level is taken as the short-term volatility index.

[0012] By calculating the mean Euclidean distance between adjacent data points within a short observation window, it is possible to capture real-time changes in the data sequence with great sensitivity, providing a basis for constructing a highly sensitive operating condition change index.

[0013] Optionally, the steps for obtaining the long-term volatility index that reflects the data fluctuations in the long-term data window include: calculating the variance of each dimension corresponding to the data points in the long-term data window, taking the sum of the variances corresponding to each dimension as the long-term volatility level, and taking the normalized result of the long-term volatility level as the long-term volatility index.

[0014] By summing the variances of each data dimension within a long-term data window, a stable benchmark representing the overall historical volatility level is formed, providing a more stable and statistically significant reference background for judging the degree of abnormality of current short-term fluctuations.

[0015] Optionally, the steps for analyzing the correlation of data in each dimension within the short-term observation window and the correlation of data in each dimension within the long-term data window, and constructing a correlation change index, include: obtaining the corresponding covariance matrix based on the multidimensional data corresponding to each data point in the short-term observation window, using this covariance matrix as the short-term morphological matrix, and obtaining a short-term correlation matrix reflecting the correlation between data in each dimension based on the short-term morphological matrix; selecting data points in the long-term data window that are equal in number to the data points in the short-term observation window as representative data points, obtaining the covariance matrix of the representative data points as the long-term morphological matrix, obtaining a long-term correlation matrix based on the long-term morphological matrix, and constructing a correlation change index based on the difference between the short-term correlation matrix and the long-term correlation matrix.

[0016] A short-term correlation matrix is ​​constructed by using the covariance matrix corresponding to the data in the short-term observation window, and a long-term correlation matrix is ​​constructed by using the data in the long-term data window. By comparing the relationship between the two correlation matrices, it is possible to reflect whether the correlation between the current data in each dimension has changed compared to the correlation between the data in each dimension over a longer period of time.

[0017] Optionally, the Euclidean distance between the long-term correlation matrix and the short-term correlation matrix can be calculated, and the normalized result of the Euclidean distance can be used as the correlation change index.

[0018] Optionally, data points can be selected at equal intervals within a long-term data window as representative data points.

[0019] Optionally, the step of adjusting the initial distance between the current data point and other data points within the long-term data window based on the operating condition change index to obtain the optimal distance includes: for any data point in the long-term data window, obtaining its operating condition change index and the time interval between it and the current data point; obtaining the operating condition difference between the two data points based on the difference between the operating condition change index of the data point and the operating condition change index of the current data point; determining an adjustment coefficient based on the operating condition difference and the time interval; and using the product of the adjustment coefficient and the initial distance between the two data points as the optimal distance, wherein both the operating condition difference and the time interval are negatively correlated with the adjustment coefficient.

[0020] By correcting the distance between data points through time intervals and the operating condition change index between data points, this mechanism ensures that when calculating the local density of any data point, the points in its neighborhood are effective neighbors from similar operating conditions, thus fundamentally solving the problem of false alarms in the LOF algorithm caused by operating condition switching.

[0021] Optionally, the steps of using the Local Outlier Factor (LOF) algorithm to detect anomalies in data points include: calculating the anomaly score of the data points using the LOF algorithm, and confirming data anomalies in response to an anomaly score greater than a preset anomaly threshold.

[0022] Secondly, this application provides a fault detection system for an underwater pelletizer, employing the following technical solution: A fault detection system for an underwater pelletizer includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a fault detection method for an underwater pelletizer as described above is implemented.

[0023] The beneficial effect is that the above-mentioned fault detection method for underwater pelletizers is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. Thus, a system can be made based on the memory and processor, which is convenient to use.

[0024] This application has the following technical effects: This application constructs an operating condition change index. This index can comprehensively assess short-term fluctuations and long-term structural changes in equipment operating data, thereby accurately capturing normal operating condition transitions. Based on this index, it is determined whether an operating condition transition has occurred during equipment operation. During an operating condition transition, the distance between data points is dynamically adjusted to effectively shield normal operating condition changes from interference with anomaly detection. Attached Figure Description

[0025] Figure 1 This is a flowchart of a fault detection method for an underwater pelletizer according to this application.

[0026] Figure 2This is a flowchart of step S2 of a fault detection method for an underwater pelletizer according to this application. Detailed Implementation

[0027] This application discloses a fault detection method for an underwater pelletizer. By analyzing the differences between short-term data changes and long-term data states from multiple dimensions, a working condition change index reflecting changes in working conditions is constructed. When working conditions change, the optimal distance is obtained by adjusting the distance between the data point and the current data point, and finally, the abnormal score is corrected to reduce false alarms.

[0028] Reference Figure 1 A fault detection method for an underwater pelletizer includes steps S1-S3.

[0029] S1: Obtain preprocessed multi-dimensional device operation data.

[0030] In this embodiment, multiple sensors are first deployed at key locations of the underwater pelletizer to comprehensively capture its operating status. Specifically, the deployed sensors and the data they collect include: current and voltage sensors at the cutter head drive motor to collect real-time current and voltage data, reflecting the drive load; a torque sensor installed on the cutter head spindle to collect cutter head resistance data during pelletizing, directly correlated with cutting force and material state; temperature sensors at the cooling water inlet and outlet to monitor cooling water temperature changes, indirectly reflecting heat accumulation within the pelletizing chamber; and vibration and acoustic sensors installed on the equipment body to collect vibration and noise signals during operation, characterizing the health of the mechanical structure.

[0031] It is understandable that the raw signals acquired by the aforementioned sensors are continuous time-series signals. To facilitate subsequent digital processing, the system preprocesses these raw signals. For example, the preprocessing steps include: filtering with a bandpass filter to remove frequency band noise unrelated to the device state; using wavelet thresholding for noise reduction to smooth the signal and retain key transient features; and finally, using a max-min normalization method to normalize the data of different dimensions, mapping them to the [0,1] interval to eliminate scale differences between different features.

[0032] S2: For any data point, construct a working condition change index that reflects the change in working conditions at the time of the data point.

[0033] In order to capture signals of changes in the operating conditions of underwater pelletizers, this paper analyzes current short-term data changes and constructs an operating condition change index that reflects the current changes in operating conditions.

[0034] Reference Figure 2Step S2 includes steps S21-S22.

[0035] S21: Construct a short-term observation window based on the data point in the preset long-term data window, analyze the fluctuation degree of the data in the short-term observation window and the fluctuation degree of the data in the long-term data window, and construct a relative fluctuation index.

[0036] In this step, first define a size as A short-term observation window, which contains only the most recent m data points. This is a fixed value smaller than the long-term data window size. Based on this short-term observation window, an instantaneous fluctuation index reflecting the changes in the current data over a short period of time is constructed.

[0037] The long-term data window is a sliding window in the traditional incremental LOF algorithm, which slides as the data is updated. The incremental LOF algorithm is a conventional technique in this field and will not be elaborated further here. The long-term data window can contain data over a relatively long period, such as 10 hours of data, while the short-term observation window contains data within a shorter time interval, such as data within a time interval of 5 to 10 minutes.

[0038] After the short-term observation window is constructed, the data changes within the window can be analyzed to reflect the current level of data fluctuation. In this embodiment, two adjacent data points within the short-term observation window are considered as adjacent point pairs. The Euclidean distance between the two data points in each adjacent point pair is obtained. The mean of the Euclidean distances for each adjacent point pair is taken as the short-term fluctuation level, and the normalized result of the short-term fluctuation level is taken as the short-term fluctuation index.

[0039] Specifically, the formula for calculating the short-term volatility index can be expressed as: In the formula, Indicates time The short-term fluctuation index (i.e., the sampling time of the current data point); Indicates time The corresponding short-time observation window The Euclidean distance between two data points in a pair of adjacent points; This indicates the number of data points in the short-term observation window; It is a linear normalization function.

[0040] This represents the local change between two adjacent data points. By statistically analyzing the local changes of all adjacent data points, we can reflect the short-term fluctuation of the data within the short-term observation window. Normalizing the short-term fluctuation level yields the short-term fluctuation index, which is convenient for subsequent calculations.

[0041] After the short-term volatility index is calculated, the long-term volatility index of the corresponding long-term data window at the current moment is calculated to reflect the long-term trend of the data.

[0042] In this embodiment, the variance of each dimension corresponding to the data points in the long-term data window is calculated, the sum of the variances corresponding to each dimension is taken as the long-term volatility, and the normalized result of the long-term volatility is taken as the long-term volatility index.

[0043] Finally, the absolute difference between the short-term volatility index and the long-term volatility index is used as the relative volatility index.

[0044] The long-term volatility index represents the amplitude of data fluctuations over a long period. If the short-term volatility index is large at a certain moment, but its corresponding long-term volatility index is also large, it indicates that although the local data at that moment has large fluctuations, it is not abnormal compared to the normal fluctuations over a long period, and the possibility of changes in operating conditions is relatively small. Conversely, if the short-term volatility index is small at a certain moment, but its corresponding long-term volatility index is much smaller, then although the short-term volatility index is small at that moment, its corresponding long-term volatility index is even smaller, and the possibility of changes in operating conditions at that moment is relatively greater.

[0045] S22: Analyze the correlation of data in each dimension within the short-term observation window and the correlation of data in each dimension within the long-term data window, construct a correlation change index, and use the ratio of the relative fluctuation index to the correlation change index as the operating condition change index.

[0046] Changes in operating conditions are not only reflected in the fluctuation range of individual characteristic values, but more profoundly in the alteration of the relationships between these characteristics. For example, under normal operating conditions, the resistance experienced by the cutter shaft is correlated with the hardness of the material being cut; changes in the material's hardness will cause specific changes in the resistance experienced by the cutter shaft. Therefore, this step aims to quantify the correlation structure between data characteristics in the short term.

[0047] Specifically, firstly, based on the multidimensional data corresponding to each data point in the short-term observation window, the corresponding covariance matrix is ​​obtained, and this covariance matrix is ​​used as the short-term morphological matrix. Based on the short-term morphological matrix, a short-term correlation matrix reflecting the correlation between data in each dimension is obtained. Based on step S1, in this embodiment, each data point includes data in multiple dimensions, and each short-term observation window includes multiple data points. The multi-dimensional data of multiple data points constitute a data matrix. The covariance matrix of this data matrix is ​​obtained. The construction of the covariance matrix is ​​a conventional technique in this field and will not be described in detail here. In the covariance matrix, the off-diagonal elements represent the local covariance (i.e., correlation) between different features. Therefore, this covariance matrix is ​​used as the short-term morphological matrix of the data point at the corresponding time.

[0048] Construct the corresponding short-term correlation matrix based on the short-term morphological matrix.

[0049] For any element in the correlation matrix, its calculation formula can be expressed as: In the formula, Represents the first element in the correlation matrix. line, number The elements of the column represent characteristics. and characteristics The correlation between them; The covariance matrix (i.e., short-time morphological matrix) of the data in the short-time observation window represents the first... line, number The elements of the column represent characteristics. and characteristics Covariance between them; The covariance matrix (i.e., short-time morphological matrix) of the data in the short-time observation window represents the first... line, number Column elements; The covariance matrix (i.e., short-time morphological matrix) of the data in the short-time observation window represents the first... line, number The elements of the column.

[0050] Then, select data points in the long-term data window that are equal in number to the data points in the short-term observation window as representative data points, obtain the covariance matrix of the representative data points as the long-term morphological matrix, obtain the long-term correlation matrix based on the long-term morphological matrix, and construct the correlation change index based on the difference between the short-term correlation matrix and the long-term correlation matrix.

[0051] The long-term data window contains a large number of data points. To facilitate subsequent calculations and reduce computational load, representative data points are selected at equal intervals within the long-term data window, with the number of selected data points being equal to the number of data points in the short-term observation window.

[0052] Subsequently, a covariance matrix is ​​constructed based on representative data points as a long-term morphological matrix. Then, a long-term correlation matrix is ​​constructed based on the long-term morphological matrix. The construction of the long-term correlation matrix is ​​similar to that of the short-term correlation matrix, and will not be elaborated here.

[0053] A correlation change index is constructed based on the difference between the short-term correlation matrix and the long-term correlation matrix.

[0054] Specifically, the correlation change index can be calculated using the following formula: In the formula, Indicates time The correlation change index; Represents the short-term correlation matrix; Represents the long-term correlation matrix; The dimension of the data points is used to normalize the difference between the two correlation matrices; This represents the Frobenius norm between two matrices.

[0055] Finally, the ratio of the relative volatility index to the correlation change index is used as the operating condition change index.

[0056] In calculating the operating condition change index, if only the pelletizer load change causes all sensor readings to increase or decrease proportionally, then the covariance will change, but the correlation coefficient matrix will remain the same. Maintain stability. It will be very low. However, if bearing wear occurs, causing a change in the relationship between vibration and current, then... It will deviate significantly , The increase in size precisely indicates a change in the system's internal physical model.

[0057] S3: In response to the operating condition change index being greater than a preset change threshold, adjust the initial distance between the current data point and other data points within the long-term data window based on the operating condition change index to obtain the optimal distance; based on the optimal distance, use the local outlier factor algorithm to detect anomalies in the data points.

[0058] The magnitude of the operating condition change index reflects the probability of a change in operating conditions at the current moment. During periods without change, the operating conditions are stable and without fluctuations, and the anomaly score calculated by the LOF algorithm is relatively accurate in this case, requiring no adjustment. However, when the operating conditions change, the data changes, creating coefficient ranges that lead to inaccurate anomaly score calculations. Therefore, this embodiment sets a change threshold. When the operating condition change index of a data point exceeds the preset threshold, the distance between each pair of data points is adjusted to obtain the optimal distance, thereby achieving the effect of adjusting the LOF anomaly score.

[0059] Specifically, for any data point in the long-term data window, its operating condition change index and the time interval between it and the current data point are obtained. The operating condition difference between the two data points is obtained based on the difference between the operating condition change index of the data point and the operating condition change index of the current data point. An adjustment coefficient is determined based on the operating condition difference and the time interval. The product of the adjustment coefficient and the initial distance between the two data points is taken as the optimal distance. The operating condition difference and the time interval are both negatively correlated with the adjustment coefficient.

[0060] In one embodiment, the formula for calculating the optimal distance can be expressed as: In the formula, Represents data points in the long-term data window The optimal distance between the current data point and the current data point; Indicates the sampling time of the current data point. Indicates the first in the long-term data window The sampling time of each data point; This represents the index of changes in operating conditions at the current data point at that time. Indicates the first in the long-term data window The operating condition change index for each data point Indicates the first in the long-term data window The initial distance between each data point and the current data point, i.e., the Euclidean distance between the two data points; Indicates the maximum time interval, used to normalize molecules; Represented by natural constant An exponential function with base 0.

[0061] In another embodiment, the formula for calculating the optimal distance can be expressed as: In the formula, Represents data points in the long-term data window The optimal distance between the current data point and the current data point; Indicates the sampling time of the current data point. Indicates the first in the long-term data window The sampling time of each data point; This represents the index of changes in operating conditions at the current data point at that time. Indicates the first in the long-term data window The operating condition change index for each data point Indicates the first in the long-term data window The initial distance between each data point and the current data point, i.e., the Euclidean distance between the two data points; Indicates the maximum time interval, used to normalize molecules; Represented by natural constant An exponential function with base 0.

[0062] In the formula, after a change in operating conditions occurs, data collected a long time ago contributes less to determining whether the current data point is abnormal. Therefore, the initial distance is initially adjusted based on the time interval between two data points. Simultaneously, during the change in operating conditions, the data changes from one steady state to another, which can be divided into three stages: the first operating condition stage, the buffer change stage, and the second operating condition stage. By comparing the operating condition change index between two data points, it can be reflected whether the two data points are in the same stage. Data points in the same stage better reflect the current state of the data point. Therefore, the initial distance is adjusted again based on the difference in the operating condition change index between the two data points to improve the accuracy and robustness of the optimal distance calculation.

[0063] After obtaining the optimal distance between each data point and the current data point, the LOF algorithm is used to obtain the optimal anomaly score based on the optimal distance. If the optimal anomaly score is greater than a preset threshold, the data point is determined to be an anomaly.

[0064] This application also discloses a fault detection system for an underwater pelletizer, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a fault detection method for an underwater pelletizer according to this application.

[0065] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0066] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A fault detection method for an underwater pelletizer, characterized in that, include: Acquire preprocessed multi-dimensional equipment operation data; for any data point, construct an operating condition change index that reflects the change in operating conditions at the time of the data point. In response to a condition change index exceeding a preset threshold, the initial distance between the current data point and other data points within a long-term data window is adjusted based on the condition change index to obtain the optimal distance. Based on this optimal distance, a local outlier factor algorithm is used to detect anomalies in the data points. The step of constructing a condition change index reflecting changes in the condition at the time the data point is located includes: constructing a short-term observation window based on the data point within a preset long-term data window; analyzing the fluctuation levels of the data in the short-term observation window and the long-term data window to construct a relative fluctuation index; analyzing the correlation between data in each dimension within the short-term observation window and the long-term data window to construct a correlation change index; and using the ratio of the relative fluctuation index to the correlation change index as the condition change index. The step of obtaining the optimal distance includes: for any data point in the long-term data window, obtaining its condition change index and the time interval between it and the current data point; obtaining the condition difference between the two data points based on the difference between the condition change index of the data point and the condition change index of the current data point; determining an adjustment coefficient based on the condition difference and the time interval; and using the product of the adjustment coefficient and the initial distance between the two data points as the optimal distance.

2. The fault detection method for an underwater pelletizer according to claim 1, characterized in that, The steps to construct a relative volatility index include: obtaining a short-term volatility index that reflects data fluctuations in a short-term observation window and a long-term volatility index that reflects data fluctuations in a long-term data window, and using the absolute difference between the short-term volatility index and the long-term volatility index as the relative volatility index.

3. A fault detection method for an underwater pelletizer according to claim 2, characterized in that, The steps to obtain the short-term volatility index, which reflects the data fluctuations in a short-term observation window, include: two adjacent data points in the short-term observation window form an adjacent point pair; the Euclidean distance between the two data points in each adjacent point pair is obtained; the mean of the Euclidean distances corresponding to each adjacent point pair is taken as the short-term volatility level; and the normalized result of the short-term volatility level is taken as the short-term volatility index.

4. A fault detection method for an underwater pelletizer according to claim 2, characterized in that, The steps to obtain the long-term volatility index, which reflects the data fluctuations in a long-term data window, include: calculating the variance of each dimension corresponding to the data points in the long-term data window, taking the sum of the variances corresponding to each dimension as the long-term volatility level, and taking the normalized result of the long-term volatility level as the long-term volatility index.

5. A fault detection method for an underwater pelletizer according to claim 1, characterized in that, The steps for analyzing the correlation of data across dimensions in a short-term observation window and a long-term data window, and constructing a correlation change index, include: obtaining the corresponding covariance matrix based on the multidimensional data corresponding to each data point in the short-term observation window; using this covariance matrix as the short-term morphological matrix; and obtaining a short-term correlation matrix reflecting the correlation between data across dimensions based on the short-term morphological matrix. In the long-term data window, selecting data points of equal number to those in the short-term observation window as representative data points; obtaining the covariance matrix of these representative data points as the long-term morphological matrix; obtaining a long-term correlation matrix based on the long-term morphological matrix; and constructing a correlation change index based on the difference between the short-term correlation matrix and the long-term correlation matrix.

6. A fault detection method for an underwater pelletizer according to claim 5, characterized in that, Calculate the Euclidean distance between the long-term correlation matrix and the short-term correlation matrix, and use the normalized result of the Euclidean distance as the correlation change index.

7. A fault detection method for an underwater pelletizer according to claim 5, characterized in that, Data points are selected at equal intervals within a long-term data window as representative data points.

8. A fault detection method for an underwater pelletizer according to claim 1, characterized in that, The steps for anomaly detection of data points using the Local Outlier Factor (LOF) algorithm include: calculating the anomaly score of the data point using the LOF algorithm, and confirming the data anomaly in response to the anomaly score being greater than a preset anomaly threshold.

9. A fault detection system for an underwater pelletizer, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a fault detection method for an underwater pelletizer according to any one of claims 1-8.

Images