A data analysis method for bellows mechanical seal detection

Abstract

The present invention relates to the field of data processing technology, and more specifically, to a data analysis method for bellows mechanical seal detection, comprising: recording any historical moment as a target moment, and taking a sequence of pressure values ​​of the bellows mechanical seal at the target moment and several previous historical moments as a data feature vector at the target moment; determining a change eigenvalue of the data feature vector at the target moment; determining an imbalance degree of the data feature vector at the target moment; determining an abnormality degree of the data feature vector at the target moment; and based on the abnormality degree, implementing data analysis for bellows mechanical seal detection. By combining the standard deviation and kurtosis of pressure data with an imbalance degree analysis based on hierarchical clustering, the present invention no longer relies solely on single-point pressure values ​​and simple statistical indicators, and can more comprehensively and meticulously capture abnormal fluctuations in pressure data, thereby preventing abnormal fluctuations from being misclassified simply due to threshold misjudgment.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and more particularly to a data analysis method for bellows mechanical seal detection. Background Art

[0002] The bellows mechanical seal is composed of a dynamic ring, a static ring, a flange, a sleeve, a bellows and other components. It is a device used to achieve sealing between the rotating shaft and the body. It is often installed on the rotating equipment of the compressor and can effectively prevent the fluid in the equipment from leaking to the outside. However, as the compressor runs for a long time, the end faces of the dynamic and static rings of the bellows mechanical seal will wear and tear, resulting in a decrease in sealing performance, which in turn causes fluid leakage. Therefore, the bellows mechanical seal needs to be monitored in real time.

[0003] Chinese patent application publication number CN116933170A discloses a mechanical seal fault classification algorithm. This method classifies positive (healthy) and negative (abnormal) samples based on the following criteria: A sample is considered healthy and classified as positive if the pressure at a given moment, the pressure 50 hours after that moment, and the average pressure within the 50 hours after that moment are all above a specific threshold. A sample is considered abnormal and classified as negative if any of the pressure at a given moment, the pressure 50 hours after that moment, or the average pressure within the 50 hours after that moment are below a specific threshold.

[0004] However, the above patent application document only uses the pressure at a certain moment, the pressure 50 hours after that moment, and the average pressure within 50 hours after that moment as indicators for dividing positive and negative samples. If the pressure data fluctuates abnormally within 50 hours, and the pressure value at the 50th hour is just higher than a specific threshold, the abnormal data will be marked as a positive sample; at the same time, due to the influence of data noise, the pressure data of the mechanical seal in the normal state may be mistakenly judged as being lower than the threshold due to noise interference, causing the normal data to be marked as a negative sample, thereby affecting the accuracy of subsequent bellows mechanical seal detection. Summary of the Invention

[0005] In order to solve the problem that the existing technology uses three pressure indicators to divide positive and negative samples, and the pressure indicators have special states, abnormal data will be marked as positive samples under special conditions, and due to the influence of data noise, the pressure data under the normal state of the mechanical seal may be misjudged as being below the threshold due to noise interference, causing the normal data to be marked as negative samples, thereby affecting the accuracy of subsequent bellows mechanical seal detection, the present invention proposes a data analysis method for bellows mechanical seal detection, the method comprising the following steps:

[0006] Any historical moment is recorded as a target moment, and the sequence of pressure values ​​of the bellows mechanical seal at the target moment and several previous historical moments is used as the data feature vector of the target moment; the change eigenvalue of the data feature vector of the target moment is determined based on the standard deviation and kurtosis of all pressure values ​​in the data feature vector of the target moment and the previous historical moment; all two adjacent pressure values ​​in the data feature vector of the target moment are constructed as a data point to obtain multiple data points corresponding to the data feature vector of the target moment, all the data points are clustered using a hierarchical clustering algorithm to obtain multiple clusters, and a target line for the target moment is constructed; the imbalance degree of the data feature vector of the target moment is determined based on the number of data points located above the target line in all the clusters, the number of data points located below the target line in all the clusters, and the change eigenvalue; the degree of abnormality of the data feature vector of the target moment is determined based on the imbalance degree and the minimum and maximum values ​​of the merged distances between samples generated by the algorithm in each round of merging during the hierarchical clustering algorithm for clustering all the data points; and data analysis for bellows mechanical seal detection is implemented based on the degree of abnormality.

[0007] The present invention combines the standard deviation, kurtosis and imbalance analysis of pressure data based on hierarchical clustering, and no longer relies solely on single-point pressure values ​​and simple statistical indicators. It can capture abnormal fluctuations in pressure data more comprehensively and meticulously, avoid abnormal fluctuations being misclassified simply due to threshold judgment errors, and reduce false positives and missed reports; by constructing data points for adjacent pressure values ​​and using cluster analysis, combined with change characteristics and imbalance, it can effectively filter out abnormal marks caused by noise interference, making pressure data under normal conditions more stable and less likely to be misjudged as abnormal; using the minimum and maximum values ​​of the merged distance in the hierarchical clustering process to quantitatively evaluate the degree of abnormality, provide more refined abnormality classification, and help to achieve early warning and accurate maintenance decisions; accurate abnormality detection results provide reliable data support for subsequent fault location, predictive maintenance and other links, promote safe and stable operation of equipment, reduce maintenance costs, and improve economic benefits.

[0008] Furthermore, the kurtosis is obtained by calculating the ratio of the fourth-order central moment to the fourth power of the standard deviation of all pressure values ​​in the data feature vector using a statistical method to obtain the kurtosis of all pressure values ​​in the data feature vector at the target moment.

[0009] Furthermore, the change characteristic value:

[0010] Where, For the The changing eigenvalue of the data eigenvector at each historical moment, For the The standard deviation of all pressure values ​​in the data feature vector at each historical moment, For the The standard deviation of all pressure values ​​in the data feature vector of the previous historical moment, For the The kurtosis of all pressure values ​​in the data feature vector at each historical moment, For the The kurtosis of all pressure values ​​in the data feature vector of the previous historical moment, is a hyperparameter, for Type curve function, is the absolute value symbol.

[0011] The present invention uses the relative changes of standard deviation and kurtosis to construct the variation characteristic value, which can comprehensively characterize the dispersion degree of pressure data and the dynamic change of peak characteristics over time, and capture the slight anomaly or trend change in pressure fluctuation; The function normalizes the standard deviation changes, making the changing characteristic values ​​respond to abnormal changes more smoothly and with nonlinear mapping, improving the ability to identify sudden changes in data while avoiding excessive response caused by noise; the standard deviation reflects the intensity of data fluctuations, and the kurtosis reflects the sharpness of the data tail distribution, reflecting the true trend of mechanical seal pressure data anomalies and reducing the error caused by a single indicator.

[0012] Furthermore, the target straight line is obtained by constructing a function straight line with equal horizontal and vertical coordinates in the cluster space, and using the function straight line as the target straight line at the current moment.

[0013] Furthermore, the imbalance degree satisfies:

[0014] Where, For the The imbalance degree of the data feature vector at each historical moment, For all clusters in the The number of data points above the target line at each historical moment, For all clusters in the The number of data points below the target line at each historical moment, For the The changing eigenvalue of the data eigenvector at each historical moment, For the The number of pressure values ​​in the data feature vector at each historical moment, is a custom parameter used to adjust the sensitivity of imbalance to differences in cluster distributions.

[0015] The imbalance index of the present invention compares the difference in the number of data points in the two areas above and below the target line, intuitively measuring the offset and asymmetry in the distribution of pressure data, helping to reveal potential abnormal patterns. The imbalance index introduces a variable characteristic value, so that the imbalance index not only reflects static distribution differences, but also dynamically responds to temporal changes in pressure data, improving sensitivity and accuracy to abnormal behavior. Custom parameters allow the imbalance index to adjust the response intensity of cluster distribution differences according to specific application requirements, adapting to different working conditions and noise levels, and realizing a more flexible and personalized anomaly detection strategy. By normalizing the difference by dividing by the total number of data points, the imbalance index avoids deviations caused by differences in data scale, ensuring the comparability and stability of the imbalance index between data feature vectors at different time points.

[0016] Furthermore, the abnormality degree satisfies:

[0017] Where, For the The abnormality of the data feature vector at each historical moment, For the The imbalance degree of the data feature vector at each historical moment, In the process of clustering all the data points by the hierarchical clustering algorithm, The minimum value of the merge distance when merging samples generated by the algorithm during round merging. In the process of clustering all the data points by the hierarchical clustering algorithm, The maximum value of the merge distance when merging samples generated by the algorithm during round merging. is the preset number of rounds of calculation when merging all the data points in the hierarchical clustering algorithm. is a hyperparameter.

[0018] The abnormality degree of the present invention combines the imbalance degree and the relative proportion of the sample merging distance in the hierarchical clustering process, taking into account the distribution imbalance of the internal points of the data feature vector and integrating the similarity measurement in the clustering process, making the abnormality judgment more scientific and comprehensive; by introducing the ratio of the minimum and maximum values ​​in each round of merging distance as an adjustment factor, it effectively reflects the density and boundary changes between samples, helps to distinguish between tightly distributed normal data and sparsely distributed abnormal data, and reduces the probability of misjudgment; the abnormality degree index comprehensively considers the imbalance degree and clustering distance dynamics, effectively filters noise interference, accurately reflects the real abnormal signal, and improves the response speed and reliability of the detection system to abnormalities; by accurately quantifying the abnormality degree, it is convenient to set a reasonable threshold, realize timely and accurate bellows mechanical seal operation abnormality warning and maintenance decision, and reduce equipment failure risk and maintenance cost.

[0019] Furthermore, the data analysis for bellows mechanical seal detection is implemented, including:

[0020] In response to the abnormality degree being not less than a preset abnormality threshold, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as an abnormal state; otherwise, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as a normal state. Based on the data feature vectors of all historical moments that have been marked, the KNN algorithm is used to judge the state of the mechanical seal corresponding to the data feature vector at the current moment through voting method to complete the data analysis for bellows mechanical seal detection.

[0021] The present invention determines the degree of abnormality by setting an abnormality threshold, realizes automatic labeling of the status of the bellows mechanical seal, eliminates the subjectivity of manual experience judgment, and improves the objectivity and consistency of the detection results; uses the data feature vectors marked at all historical moments as the training basis of the KNN algorithm, and combines the voting method to make the current state judgment, effectively integrating time series information, and enhancing the robustness and accuracy of the judgment; the KNN algorithm is based on the state voting of neighboring samples, can adapt to the complex structure of the data feature space, cope with diverse and nonlinear abnormal patterns, and avoid errors caused by single time point judgment; combined with continuously updated historical data and the KNN model, it can continuously optimize and adjust the detection accuracy, adapt to changes in the mechanical seal use environment, and ensure long-term stable operation.

[0022] Furthermore, the data analysis for bellows mechanical seal detection is implemented, including:

[0023] In response to the abnormality degree being not less than a preset abnormality threshold, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as an abnormal state; otherwise, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as a normal state. The neural network model is trained by completing the data feature vectors of all historical moments that have been marked, and a neural network model for judging the bellows mechanical seal state corresponding to the data feature vector at the current moment is obtained, thereby completing the data analysis for bellows mechanical seal detection.

[0024] The present invention uses data feature vectors from all historical moments to train a neural network model through automatic labeling based on the degree of abnormality, which can fully tap into the complex nonlinear relationships in the data and achieve smarter and more accurate identification of the status of bellows mechanical seals. By training the neural network with a large amount of labeled data, the model can adapt to different working environments and pressure fluctuation patterns, improve the ability to identify unknown abnormalities and complex fault types, and reduce the need for manual parameter adjustment. Accurate and continuous status identification provides a scientific basis for equipment health assessment and maintenance decisions, effectively reducing failure risks and operating and maintenance costs, and improving equipment reliability and service life.

[0025] Furthermore, the neural network model adopts an LSTM model.

[0026] Furthermore, the neural network model adopts a GRU model.

[0027] The present invention has the following beneficial effects:

[0028] (1) Breaking through the limitations of traditional methods that rely only on pressure at a single moment, pressure after a fixed time period, and average value, a multi-dimensional abnormality assessment system is constructed by integrating the standard deviation and kurtosis (reflecting the severity of data fluctuations and distribution form) of the data feature vector, the imbalance degree of the data points (reflecting the stability of the pressure change trend), and the extreme value of the merged distance in the clustering process (characterizing the clustering characteristics of the data points). This system can effectively capture abnormal pressure fluctuations, avoid misjudgments caused by threshold judgments at a single time point, and significantly reduce the probability of abnormal data being mislabeled as normal and normal data being mislabeled as abnormal.

[0029] (2) By constructing adjacent pressure values ​​into data points and performing hierarchical clustering, a target straight line is generated to analyze the imbalance of the upper and lower distribution of data points. This allows for accurate identification of short-term fluctuation trends and long-term change patterns in pressure data. For example, for abnormal fluctuations that occur within 50 hours and then recover, traditional methods may misjudge them as normal because the endpoint pressure meets the standard. However, the present invention can capture such potential anomalies through the coordinated analysis of the imbalance and the change characteristic value, thereby improving sensitivity to complex pressure fluctuations.

[0030] (3) Statistical features such as standard deviation and kurtosis are introduced to quantify the discreteness and distribution characteristics of the data, and the distance parameter in the clustering process is combined to filter out noise interference. For the temporary deviation of the pressure value caused by noise under normal conditions, the present invention can effectively distinguish between noise interference and real anomalies by evaluating the aggregation and trend consistency of the overall data, reduce the impact of data noise on the detection results, and ensure the accuracy of state division.

[0031] (4) More accurate abnormality analysis results can provide data support for the status assessment of bellows mechanical seals, helping operation and maintenance personnel to promptly detect potential faults (such as seal wear, pressure leakage, etc.), formulate maintenance strategies in advance, avoid equipment shutdowns or safety accidents caused by missed inspections, reduce operation and maintenance costs, and extend equipment service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of the steps of a data analysis method for bellows mechanical seal detection according to an embodiment of the present invention.

[0033] Figure 2 It is a schematic diagram of the distribution of steady pressure data of a data analysis method for bellows mechanical seal detection according to an embodiment of the present invention.

[0034] Figure 3 It is a distribution diagram of data points and corresponding target lines after clustering of steady pressure data in a data analysis method for bellows mechanical seal detection according to an embodiment of the present invention.

[0035] Figure 4 It is a schematic diagram of the distribution of pressure drop data of a data analysis method for bellows mechanical seal detection according to an embodiment of the present invention.

[0036] Figure 5 It is a distribution diagram of data points and corresponding target straight lines after clustering of falling pressure data in a data analysis method for bellows mechanical seal detection according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. The described embodiments are part of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of the present invention.

[0038] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] See also Figure 1 , which shows a flowchart of a data analysis method for bellows mechanical seal detection provided by one embodiment of the present invention, the method comprising the following steps:

[0040] S01: Record any historical moment as the target moment, and use the sequence of pressure values ​​of the bellows mechanical seal at the target moment and several previous historical moments as the data feature vector of the target moment.

[0041] It should be noted that the pressure data at the bellows mechanical seal is collected through a pressure transmitter, the size of the time window is set to K, and the sequence consisting of the pressure data (pressure values) at the target moment and K-1 historical moments before the target moment, a total of K moments, is used as the data feature vector at the target moment.

[0042] Implementers can set the value of K and the acquisition frequency based on specific implementation conditions. For example, if K = 100 and the acquisition frequency is 10 times / s, the first 100 moments collected in history will only serve as the data set for subsequent moments and will not be included in the following calculations.

[0043] S02: Determine the change eigenvalue of the data eigenvector at the target moment.

[0044] It should be noted that, regardless of whether the mechanical seal is in a normal state or an abnormal state, the pressure data in each time period (the target moment and several previous historical moments) will have certain changes due to noise. When the mechanical seal is in different states, the pressure data will also have different changes. In order to further analyze the change of pressure data and better judge the state of the mechanical seal, this step obtains the change characteristic value of the data characteristic vector at each historical moment based on the data characteristic vector at each historical moment.

[0045] The change eigenvalue of the data eigenvalue at the target moment is determined based on the standard deviation and kurtosis of all pressure values ​​in the data eigenvalue vector at the target moment and the previous historical moment of the target moment.

[0046] Specifically, the kurtosis is obtained by calculating the ratio of the fourth-order central moment to the fourth power of the standard deviation of all pressure values ​​in the data eigenvector using a statistical method to obtain the kurtosis of all pressure values ​​in the data eigenvector at the target moment.

[0047] Specifically, the change characteristic value:

[0048] ;

[0049] Where, For the The changing eigenvalue of the data eigenvector at each historical moment, For the The standard deviation of all pressure values ​​in the data feature vector at each historical moment, For the The standard deviation of all pressure values ​​in the data feature vector of the previous historical moment, For the The kurtosis of all pressure values ​​in the data feature vector at each historical moment, For the The kurtosis of all pressure values ​​in the data feature vector of the previous historical moment, is a hyperparameter, for Type curve function, is the absolute value symbol.

[0050] Implementers can set hyperparameters according to specific implementation conditions, for example, 0.001. The existence of hyperparameters is to prevent , which makes the formula meaningless.

[0051] in, The larger the The more discrete the pressure data distribution is within the data feature vector at each historical moment, the The greater the change in the pressure data within the data feature vector at each historical moment, the The larger the change eigenvalue of the data eigenvector at a historical moment, The smaller the The more concentrated the pressure data distribution is in the data feature vector of the historical moment, the The smaller the change in the pressure data within the data feature vector at each historical moment, the The smaller the change eigenvalue of the data eigenvector at a historical moment. The larger the The greater the difference between the peak value of the pressure data distribution in the data feature vector at a historical moment and the peak value of the pressure data distribution in the data feature vector at the previous historical moment, the greater the difference between the peak value of the pressure data distribution in the data feature vector at the previous historical moment. The more likely a historical moment is to have a drastic change and a new peak, the more likely The larger the change eigenvalue of the data eigenvector at a historical moment, The smaller the The smaller the difference between the peak value of the pressure data distribution in the characteristic vector of the data at the previous historical moment and the peak value of the pressure data distribution in the characteristic vector of the data at the previous historical moment, the smaller the difference between the peak value of the pressure data distribution in the characteristic vector of the data at the previous historical moment, which means that the distribution shape is The more likely it is that no change has occurred at a historical moment, the The smaller the change eigenvalue of the data eigenvector at a historical moment.

[0052] For example: For the convenience of calculation, assume , No. The data feature vector of a historical moment is , No. The data feature vector of a historical moment is , then the calculation is as follows:

[0053] 、 , substitute , substitute ;

[0054] 、 , substitute ;

[0055] Substitute into the formula .

[0056] S03: Determine the imbalance degree of the data feature vector at the target moment.

[0057] It should be noted that when the mechanical seal is operating under different conditions, the pressure data will show different trends of change. For example, when the mechanical seal is in normal condition, the pressure data will show a steady change overall, while when the mechanical seal is leaking, the pressure data will show a gradual downward trend. Therefore, the mechanical seal status can be analyzed by the changing trend of the pressure data. Therefore, this step obtains the imbalance degree of the data feature vector at each historical moment based on the changing eigenvalue of the data feature vector at each historical moment and the data feature vector.

[0058] All two adjacent pressure values ​​in the data feature vector at the target time are constructed as a data point, and multiple data points corresponding to the data feature vector at the target time are obtained (for example, The data feature vector of a historical moment is , then The data point corresponding to the data feature vector at each historical moment is ,Right now , the number of data points is 4), a hierarchical clustering algorithm is used to cluster all the data points to obtain multiple clusters, and a target straight line at the target moment is constructed. The imbalance degree of the data feature vector at the target moment is determined according to the number of data points above the target straight line in all the clusters, the number of data points below the target straight line in all the clusters, and the change eigenvalue.

[0059] Specifically, the target straight line is obtained by constructing a function straight line with equal horizontal and vertical coordinates in the cluster space, and using the function straight line as the target straight line at the current moment.

[0060] Specifically, the imbalance degree satisfies:

[0061] ;

[0062] Where, For the The imbalance degree of the data feature vector at each historical moment, For all clusters in the The number of data points above the target line at each historical moment, For all clusters in the The number of data points below the target line at each historical moment, For the The changing eigenvalue of the data eigenvector at each historical moment, For the The number of pressure values ​​in the data feature vector at each historical moment, is a custom parameter used to adjust the sensitivity of imbalance to differences in cluster distributions.

[0063] Implementers can set the custom parameter to a value not less than 1, for example, 1, based on specific implementation circumstances.

[0064] Among them, since the two parameters of the data point are the pressure data at a historical moment and the pressure data at the next historical moment, the data points above the target line represent an increase in data, and the data points below the target line represent a decrease in data. Usually, when the pressure data is stable, due to the existence of noise, the number of data points on both sides of the target line should be similar. When the pressure data shows a downward trend, the number of data points below the target line will be more than the number of data points above the target line, such as Figure 2 、 Figure 3 、 Figure 4 as well as Figure 5 As shown in Figure 2, the number of data points on both sides of the target straight line corresponding to the steady pressure data is similar, while the number of data points below the target straight line corresponding to the decreasing pressure data is greater. Calculate the imbalance, The larger the value, the greater the difference between the number of data points on both sides of the target line, and the more unbalanced the number of data points on both sides of the target line. The greater the imbalance of the data feature vector at a historical moment; The smaller the value, the smaller the difference between the number of data points on both sides of the target line, and the more balanced the number of data points on both sides of the target line. The smaller the imbalance of the data feature vector at each historical moment. When the pressure data is stable, there will occasionally be a few data points that deviate from the target line due to random noise. If the imbalance is calculated based only on the number of data points on both sides of the target line, it may cause some stable data feature vectors to be too large. Therefore, by changing the eigenvalue Participate in the calculation of imbalance, The larger the value is, the greater the distribution change of pressure data in the feature vector of adjacent time data is. The collection of pressure data is more likely to be interfered by the external environment, thus generating a certain amount of noise, making the number of data points on both sides of the target line unbalanced when the pressure data is stable. The smaller it is, the smaller the distribution change of pressure data in the characteristic vector of adjacent time data is, the more stable the change of pressure data is, and the imbalance of the number of data points on both sides of the target straight line is more likely to be caused by the actual rise or fall of pressure; therefore, by right Make a downward correction. The bigger it is, right The greater the correction strength, the smaller the influence of noise data on the imbalance calculation. The smaller the time, right The smaller the correction strength, the less likely it is to be over-corrected, which may lead to the inability to distinguish the distribution of data points on both sides of the target line.

[0065] Exemplary: , , No. The data feature vector of a historical moment is , the corresponding data points are , the target straight line has equal horizontal and vertical coordinates, that is , then there is 1 data point above the target line, that is, There are 3 data points below the target line, that is, data points;

[0066] Substitute into the formula .

[0067] S04: Determine the abnormality level of the data feature vector at the target moment.

[0068] It's important to note that the hierarchical clustering algorithm's clustering process exhibits distinct phased characteristics: in the first round of clustering, the algorithm considers the two closest data points as the initial samples and merges them into a new sample. In subsequent phases, the algorithm continues to follow the same logic for merging samples. If the sample to be merged contains multiple data points, the center point of the sample is used as the reference point for distance measurement, and the closest samples are continuously merged into new samples, ultimately forming multiple clusters. Therefore, the merged distances at different stages of the clustering process can reflect the evolution of the similarity between the data. When the signal is stable, the variation between adjacent data points is small, the similarity between the resulting data points is high, and the merged distances for hierarchical clustering are small and grow slowly. However, when the signal exhibits abnormal fluctuations or trend changes (such as a continuous decrease in pressure), the variation between adjacent data points increases, causing the distances between some data points and surrounding samples to rapidly increase. Therefore, this step determines the degree of anomaly based on the imbalance of the data feature vector and the distances during the hierarchical clustering process.

[0069] The abnormality degree of the data feature vector at the target moment is determined according to the imbalance degree and the minimum and maximum values ​​of the merged distances when the samples generated by the algorithm are merged in each round of merging in the process of clustering all the data points by the hierarchical clustering algorithm.

[0070] Specifically, the abnormality degree satisfies:

[0071] ;

[0072] Where, For the The abnormality of the data feature vector at each historical moment, For the The imbalance degree of the data feature vector at each historical moment, In the process of clustering all the data points by the hierarchical clustering algorithm, The minimum value of the merge distance when merging samples generated by the algorithm during round merging. In the process of clustering all the data points by the hierarchical clustering algorithm, The maximum value of the merge distance when merging samples generated by the algorithm during round merging. is the preset number of rounds of calculation when merging all the data points in the hierarchical clustering algorithm. Hyperparameters exist to prevent extreme cases (where the data is completely stationary) , which makes the formula meaningless.

[0073] Implementers can set the number of rounds involved in the calculation based on specific implementation circumstances, for example, the first two rounds.

[0074] in, The larger the value is, the more unstable the pressure data in the data feature vector is, and the more likely the mechanical seal is to be in an abnormal state. The greater the abnormality of the data feature vector at a historical moment; The smaller the value, the more stable the pressure data in the data feature vector is, and the more likely the mechanical seal is to be in a normal state. The smaller the abnormality of the data feature vector at each historical moment. In the hierarchical clustering process, under normal conditions, the data points corresponding to the data feature vector should be relatively balanced, the distance difference between each category is small, and the ratio of the minimum to the maximum value of the combined distance in the clustering process is close to 1; when there are abnormal fluctuations, trend changes or mutation interference in the signal, the distance distribution between samples becomes highly unbalanced, and the ratio decreases significantly; therefore, The larger it is, the more uniform the data points corresponding to the data feature vector are, and the smaller the abnormality of the data feature vector is. The smaller it is, the more uneven the data points corresponding to the data feature vector are, and the greater the abnormality of the data feature vector is.

[0075] Exemplary: The hierarchical clustering process (Euclidean distance, centroid method) is as follows;

[0076] No. The distance matrix of round 1 of the data points corresponding to the data feature vectors of historical moments is shown in Table 1:

[0077] Table 1

[0078]

[0079] Round 1: Calculate the Euclidean distance between all data points and merge the data points with the smallest Euclidean distance to obtain multiple new merged samples. Due to the small number of data points, only P1, P2, and P3 are merged in this round to obtain the new sample C1. Therefore, the maximum and minimum values ​​of the merged distance in this round are both 0.1414.

[0080] No. The distance matrix of round 2 of the data points corresponding to the data feature vectors of historical moments is shown in Table 2:

[0081] Table 2

[0082]

[0083] Round 2: At this time, there are only samples C1 and P4. Since the hierarchical clustering algorithm will eventually merge into one cluster, C1 and P4 are merged in this round to obtain the final cluster. Therefore, the maximum and minimum values ​​of the merge distance during this round are both 0.7667.

[0084] but ;

[0085] Substitute into the formula .

[0086] S05: Based on the abnormality level, perform data analysis for bellows mechanical seal detection.

[0087] Specifically, the data analysis for bellows mechanical seal detection includes:

[0088] In response to the abnormality degree being not less than a preset abnormality threshold, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as an abnormal state; otherwise, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as a normal state. Based on the data feature vectors of all historical moments that have been marked, the KNN algorithm is used to judge the state of the mechanical seal corresponding to the data feature vector at the current moment through voting method to complete the data analysis for bellows mechanical seal detection.

[0089] Implementers can set abnormality thresholds based on specific implementation situations. For example, when 50 mechanical seals are manually determined to be in abnormal states, the average value of the abnormality degree of the corresponding data feature vector is used.

[0090] In another embodiment, the data analysis for bellows mechanical seal detection is implemented, including:

[0091] In response to the abnormality degree being not less than a preset abnormality threshold, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as an abnormal state; otherwise, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as a normal state. The neural network model is trained by completing the data feature vectors of all historical moments that have been marked, and a neural network model for judging the bellows mechanical seal state corresponding to the data feature vector at the current moment is obtained, thereby completing the data analysis for bellows mechanical seal detection.

[0092] Implementers can select a neural network model based on specific implementation circumstances, such as an LSTM model or a GRU model.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A data analysis method for bellows mechanical seal detection, characterized in that: include: Record any historical moment as the target moment, and take the sequence of pressure values ​​of the bellows mechanical seal at the target moment and several previous historical moments as the data feature vector of the target moment; Determine the change eigenvalue of the data eigenvalue at the target moment based on the standard deviation and kurtosis of all pressure values ​​in the data eigenvalue at the target moment and the previous historical moment; constructing all two adjacent pressure values ​​in the data feature vector at the target moment as one data point to obtain multiple data points corresponding to the data feature vector at the target moment, clustering all the data points using a hierarchical clustering algorithm to obtain multiple clusters, and constructing a target straight line at the target moment, and determining the imbalance degree of the data feature vector at the target moment based on the number of data points above the target straight line in all the clusters, the number of data points below the target straight line in all the clusters, and the change characteristic value; Determine the abnormality of the data feature vector at the target moment according to the imbalance degree and the minimum and maximum values ​​of the merged distances between samples generated by the algorithm in each round of merging in the process of clustering all the data points using the hierarchical clustering algorithm; Based on the abnormality degree, data analysis for bellows mechanical seal detection is implemented.

2. A data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The kurtosis is obtained by calculating the ratio of the fourth-order central moment to the fourth power of the standard deviation of all pressure values ​​in the data eigenvector using a statistical method to obtain the kurtosis of all pressure values ​​in the data eigenvector at the target moment.

3. The data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The changing characteristic value: ； Where, For the The changing eigenvalue of the data eigenvector at each historical moment, For the The standard deviation of all pressure values ​​in the data feature vector at each historical moment, For the The standard deviation of all pressure values ​​in the data feature vector of the previous historical moment, For the The kurtosis of all pressure values ​​in the data feature vector at each historical moment, For the The kurtosis of all pressure values ​​in the data feature vector of the previous historical moment, is a hyperparameter, for Type curve function, is the absolute value symbol.

4. The data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The target straight line is obtained by constructing a function straight line with equal horizontal and vertical coordinates in the cluster space, and using the function straight line as the target straight line at the current moment.

5. The data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The imbalance degree satisfies: ； Where, For the The imbalance degree of the data feature vector at each historical moment, For all clusters in the The number of data points above the target line at each historical moment, For all clusters in the The number of data points below the target line at each historical moment, For the The changing eigenvalue of the data eigenvector at each historical moment, For the The number of pressure values ​​in the data feature vector at each historical moment, is a custom parameter used to adjust the sensitivity of imbalance to differences in cluster distributions.

6. The data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The degree of abnormality meets the following requirements: ； Where, For the The abnormality of the data feature vector at each historical moment, For the The imbalance degree of the data feature vector at each historical moment, In the process of clustering all the data points by the hierarchical clustering algorithm, The minimum value of the merge distance when merging samples generated by the algorithm during round merging. In the process of clustering all the data points by the hierarchical clustering algorithm, The maximum value of the merge distance when merging samples generated by the algorithm during round merging. is the preset number of rounds of calculation when merging all the data points in the hierarchical clustering algorithm. is a hyperparameter.

7. The data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The data analysis for bellows mechanical seal detection is implemented, including: In response to the abnormality degree being not less than a preset abnormality threshold, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as an abnormal state; otherwise, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as a normal state. Based on the data feature vectors of all historical moments that have been marked, the KNN algorithm is used to judge the state of the mechanical seal corresponding to the data feature vector at the current moment through voting method to complete the data analysis for bellows mechanical seal detection.

8. The data analysis method for bellows mechanical seal detection according to claim 1, characterized in that: The data analysis for bellows mechanical seal detection is implemented, including: In response to the abnormality degree being not less than a preset abnormality threshold, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as an abnormal state; otherwise, the bellows mechanical seal state corresponding to the data feature vector at the target moment is marked as a normal state. The neural network model is trained by completing the data feature vectors of all historical moments that have been marked, and a neural network model for judging the bellows mechanical seal state corresponding to the data feature vector at the current moment is obtained, thereby completing the data analysis for bellows mechanical seal detection.

9. The data analysis method for bellows mechanical seal detection according to claim 8, characterized in that: The neural network model adopts the LSTM model.

10. The data analysis method for bellows mechanical seal detection according to claim 8, characterized in that: The neural network model adopts the GRU model.

Images