Incremental frequent item set mining method and cache method for unmanned underwater vehicle

Abstract

The application provides an incremental frequent item set mining method and a cache method for an unmanned underwater vehicle, and the method comprises the following steps: acquiring data of multiple sensors of the unmanned underwater vehicle, and merging initial frequent item sets of all sensor data as a global candidate item set list; respectively calculating edge distribution of each sensor data and a correlation coefficient between the sensor data, and constructing a joint distribution of different sensor data according to the edge distribution and the correlation coefficient; generating a frequent item set for new sensor data, calculating a similarity between the new frequent item set and all initial frequent item sets, if the similarity is greater than a set similarity threshold, then the new frequent item set is merged into the corresponding initial frequent item set, otherwise the new frequent item set is added to the global candidate item set list as a new item set; and according to the updated global candidate item set list, dynamically adjusting parameters of the joint distribution and support degrees of each frequent item set, so as to update the global candidate item set list.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of data mining, and particularly relates to an incremental frequent item set mining method and a cache strategy for an unmanned underwater vehicle. BACKGROUND

[0002] In the era of big data, the processing demand of massive dynamic data is rising, which is particularly evident in complex system platforms such as unmanned underwater vehicles (UUVs). Various sensors on these platforms generate a large amount of data that continues to grow. Traditional data mining methods have many drawbacks when dealing with such large-scale dynamic data, such as excessive consumption of computing resources and insufficient memory supply. In particular, in the frequent item set updating link of incremental mining, repeated scanning operations on the original data are often required, directly leading to extremely low computational efficiency.

[0003] Currently, the related field has introduced frequent item set mining based on multi-scale theory. This method has the advantage of being able to analyze data from different levels and different granularities of business, and can complete the incremental updating of frequent item sets with relatively small computational cost, providing a new way of thinking and approach for big data processing.

[0004] However, the existing technology still has some deficiencies. The existing frequent item set mining method is difficult to efficiently and cost-effectively mine incremental frequent item sets in large-scale dynamic data processing and parallel computing environments. For UUV platforms, the data processing resources are limited, and when processing large-scale dynamic data, traditional incremental frequent item set mining methods require repeated scanning of the entire data set, resulting in large consumption of computing resources and low efficiency. In underwater tasks, the continuous growth of sensor data further exacerbates this problem. At the same time, the traditional frequent item set mining method has high computational complexity when updating frequent item sets, relying on multiple scans of the data set and frequent adjustments of the tree structure. In real-time data analysis and fault prediction scenarios for UUVs, it is difficult to reduce computational cost and improve computational efficiency while ensuring accuracy, especially in distributed computing environments, where efficient data processing and mining are urgent technical problems to be solved. SUMMARY

[0005] The present application proposes an incremental frequent item set mining method and cache strategy for an unmanned underwater vehicle, solving the problem of large computational resource consumption and low efficiency of existing frequent item set mining methods.

[0006] To solve the above technical problems, the present application provides an incremental frequent item set mining method for an unmanned underwater vehicle, comprising the following steps:

[0007] Step S1: obtaining data of multiple sensors of the unmanned underwater vehicle, performing frequent item set mining on data of all sensors to generate initial frequent item sets of each sensor, and merging all initial frequent item sets as a global candidate item set list;

[0008] Step S2: calculating an edge distribution of data of each sensor and a correlation coefficient between data, and constructing a joint distribution of data according to the edge distribution and the correlation coefficient;

[0009] Step S3: updating the global candidate item set list: generating frequent item sets for new data of sensors of the unmanned underwater vehicle in the navigation process, calculating a similarity between the new frequent item sets and all initial frequent item sets, if the similarity is greater than a set similarity threshold, merging the new frequent item sets into the corresponding initial frequent item sets, otherwise, adding the new frequent item sets as new item sets to the global candidate item set list;

[0010] Step S4: dynamically adjusting parameters of the joint distribution according to the updated global candidate item set list, updating support of the frequent item sets according to the updated joint distribution, and thus updating the global candidate item set list.

[0011] Preferably, the generating initial frequent item sets of each sensor in step S1 comprises the following steps: performing frequent item set mining on data of all sensors to generate candidate frequent item sets of each sensor, calculating support of all candidate frequent item sets, and taking candidate frequent item sets with support greater than a set threshold as initial frequent item sets of each sensor.

[0012] Preferably, an expression for calculating the support is:

[0013] ;

[0014] In the formula, is support of the candidate frequent item set ; is the number of occurrences of the candidate frequent item set ; is the total number of transactions.

[0015] Preferably, an expression for calculating the edge distribution of data of the sensor in step S2 is:

[0016]

[0017] ;

[0018] ;

[0019] In the formula, is the edge distribution of the sensor data ; is the edge probability density of the sensor data ; is the number of sampling samples of the sensor data ; is the kernel function is the observation value of the i-th sensor data point.

[0020] Preferably, the expression for calculating the correlation coefficient between the data in step S2 is:

[0021] ;

[0022] ;

[0023] ;

[0024] In the above formula, is the Copula function , is the edge distribution value of the sensor data x, y is the correlation coefficient between the sensor data x, y is the bivariate standard normal distribution function of ; is the inverse function of the standard normal distribution is the dependence parameter of the Copula function.

[0025] Preferably, when constructing the joint distribution of the data in step S2, the parameters of the joint distribution are estimated by the historical data of the sensor, and the expression for estimating the parameters of the joint distribution is:

[0026] ;

[0027] ;

[0028] In the above formula, is the maximum likelihood estimation function , is the parameter of the Copula function is the sample number of the historical sensor data is the Copula function , is the edge distribution of the historical sensor data .

[0029] Preferably, the expression for constructing the joint distribution of the data according to the edge distribution and the correlation coefficient in step S2 is:

[0030] ;

[0031] wherein, Pxyis the joint distribution of sensor data x, y; , Pxis the marginal distribution of sensor data x, y, respectively; rxyis the correlation coefficient between sensor data x, y; , Pxis the marginal probability density of sensor data x, y, respectively.

[0032] Preferably, the expression for calculating the similarity between frequent itemsets in step S3 is:

[0033] ;

[0034] wherein, is the Jaccard similarity; , represents the sensor data; represents the frequent itemsets in the sensor data D.

[0035] Preferably, the expression for dynamically adjusting the parameters of the joint distribution according to the updated global candidate itemset list in step S4 is:

[0036] ;

[0037] wherein, rxyis the correlation coefficient between updated sensor data x, y; represents the value of the parameter when the objective function reaches the maximum value; is the newly added sensor data; , Pxis the marginal distribution of sensor data x, y, respectively; rxyis the correlation coefficient between sensor data x, y.

[0038] A caching strategy suitable for the incremental frequent itemset mining method of the unmanned underwater vehicle described above, comprising the following steps:

[0039] Step S1: Convert the sensor data into a resilient distributed dataset (RDD) using a distributed computing framework, and store all RDDs in memory;

[0040] Step S2: Select a portion of the sensor data for testing, record the number of dependencies N and the calculation time T for each RDD, and calculate the generation value of each RDD; store all generation values in a one-dimensional array C, and sort all generation values;

[0041] Step S3: test all sensor data, when RDD is used, update the dependent times N of the RDD N=N-1, eliminate the RDD with dependent times 0;

[0042] Step S4: when the memory is insufficient, eliminate the RDD with the minimum generation value in the cache according to the sorting result and update C.

[0043] The beneficial effects of the present application at least include:

[0044] 1. After generating frequent item sets for the new data, the similarity is judged to determine whether to merge or add the global candidate item set list, avoiding frequent large-scale calculation on the entire data set, reducing the consumption of computing resources, and improving the real-time performance and efficiency of data processing;

[0045] 2. According to the updated global candidate item set list, the joint distribution parameters are dynamically adjusted, and the support of the frequent item set is updated, realizing the timely updating and optimization of the data mining model, and ensuring the rapid response to data changes;

[0046] 3. By calculating the marginal distribution of each sensor data and the correlation coefficient between the sensor data, the joint distribution of different sensor data is constructed, fully considering the correlation and mutual influence between the sensor data, which is helpful to more comprehensively and accurately understand the information reflected by each sensor data of the unmanned underwater vehicle, and to mine the complex relationships and patterns hidden behind the data, thereby providing a basis for more accurate data analysis and decision support. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 It is a method flow diagram of the embodiment of the present application;

[0048] Figure 2 It is an RDD conversion flowchart of the embodiment of the present application;

[0049] Figure 3 It is a cache strategy flowchart of the embodiment of the present application;

[0050] Figure 4 It is a comparison chart of Copula functions under different parameters in the embodiment of the present application;

[0051] Figure 5 It is a performance comparison diagram of the cache strategy and the LRU strategy of the embodiment of the present application. DETAILED DESCRIPTION

[0052] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0053] As shown in the drawings, the embodiments of the present application provide an incremental frequent item set mining method of an unmanned underwater vehicle, comprising the following steps: Figure 1

[0054] Step S1: Obtain data of multiple sensors of the unmanned underwater vehicle, perform frequent item set mining on data of all sensors, generate initial frequent item sets of each sensor, and combine all initial frequent item sets as a global candidate item set list.

[0055] Specifically, the original data set of the sensor is read from the UUV distributed storage system through a text file TextFile, and the sensor data set is divided into multiple scale levels based on the multi-scale theory.

[0056] For example, the sensor data can be divided into the following scale data sets by taking 1 hour as the time scale and each sensor as the division criterion:

[0057] Scale data set 1: sonar data in the past 1 hour.

[0058] Scale data set 2: navigation data in the past 1 hour.

[0059] Scale data set 3: power system data in the past 1 hour.

[0060] As shown in the drawings, Figure 2 ​The RDD transformation flowchart is shown, which shows the process of transforming the unmanned underwater vehicle sensor data into a resilient distributed dataset RDD and used for frequent item set mining, including the following steps: reading the original sensor dataset from the distributed storage system of the unmanned underwater vehicle through the text file TextFile; based on the multi-scale theory, the dataset is divided into multiple benchmark scale datasets according to the sensor type. Then generate the corresponding resilient distributed dataset for each benchmark scale dataset, denoted as the transaction resilient distributed dataset TransactionRDD, and store it in the memory. The FP-Growth algorithm is used to mine the frequent item sets of each TransactionRDD, and the candidate frequent item sets of each benchmark scale dataset are obtained. The support of these candidate frequent item sets is calculated, and the candidate frequent item sets with support greater than the set threshold are selected as the initial frequent item sets of each benchmark scale, and the benchmark scale dataset frequent item set resilient distributed dataset BSDSFrequentItemRDD is generated. Finally, by merging the mining results of different benchmark scales, the global candidate item set list can_list is formed, denoted as CandidataItemRDD, which lays the foundation for subsequent incremental update and global frequent item set screening.

[0061] The FP-Growth algorithm is used to mine the frequent item sets of each scale level sensor dataset. The support of each frequent item set is measured using Jaccard similarity, and the frequent item sets with support greater than the set threshold are used as the initial frequent item sets of each scale. The similarity between the frequent item sets of each scale is calculated to prepare for subsequent incremental update. The expression for calculating the support of the frequent item set is:

[0062] ;

[0063] In the formula, is the support of the candidate frequent item set ; is the number of occurrences of the candidate frequent item set ; is the total number of transactions.

[0064] Step S2: Calculate the marginal distribution of the data of each sensor and the correlation coefficient between the data, respectively, and construct the joint distribution of the data according to the marginal distribution and the correlation coefficient.

[0065] Specifically, based on the Copula theory, a binary joint probability density function model of sensor data is constructed to separate the marginal distribution and the dependence structure of the sensor data. The specific steps are as follows:

[0066] Step S21: Calculate the marginal probability density function of sonar, navigation and other sensor data through kernel density estimation (KDE):

[0067] ;

[0068] ;

[0069] In the above formula, is the edge probability density of the sensor data ; is the number of sampling samples of the sensor data is the bandwidth of the kernel density estimation is the kernel function is the observation value of the i-th sensor data point.

[0070] Wherein is used to control the smoothness of the estimated curve, the greater the value, the smoother the estimated curve, but the details may be lost. The integral mean square error (MISE) can be minimized by cross-validation, or adjusted adaptively according to the sensor sampling frequency in the UUV scene. The kernel function is used to weight the adjacent data points, and the probability density distribution of the sensor variable is estimated through historical data .

[0071] Step S22: Select the type of Copula function according to the dependence characteristics between variables, in the embodiment of the application, the Gaussian Copula is used to capture the general linear correlation of the sensor data, and the Clayton Copula is used to capture the asymmetric dependence in the extreme state. The expression for calculating the correlation coefficient between the sensor data by using the Copula function is:

[0072] ;

[0073] ;

[0074] ;

[0075] In the above formula, is the Copula function , is the probability distribution value obtained by converting the sensor data x and y through the edge distribution function is the correlation coefficient between the sensor data x and y , describing the linear correlation between the sensor data is the bivariate standard normal distribution function of ; is the inverse function of the standard normal distribution, which converts the edge distribution value , into a standard normal variable is the dependence parameter of the Clayton Copula function.

[0076] wherein, The greater the value is, the stronger the lower tail dependence between two sensor data modeled by the Clayton Copula function is. For example, the linkage of sensors at extreme low values. If the sonar signal strength and the power system temperature are highly correlated at low values, the Clayton Copula is selected and a larger .

[0077] Step S23: estimating the parameters of the Copula function by the historical sensor data The parameters of the Copula function are estimated by the maximum likelihood estimation, specifically:

[0078] ;

[0079] ;

[0080] In the above formula, is the maximum likelihood estimation function; , is the parameter of the Copula function; is the sample number of the historical sensor data; is the Copula function; , is the marginal distribution of the historical sensor data .

[0081] Step S3: updating the global candidate item set list: generating the frequent item sets for the new data of the sensors of the unmanned underwater vehicle during the navigation process, calculating the similarity of the new frequent item sets and all initial frequent item sets, if the similarity is greater than the set similarity threshold, merging the new frequent item sets into the corresponding initial frequent item sets, otherwise, adding the new frequent item sets as new item sets to the global candidate item set list.

[0082] Specifically, when the item set appearing in the new data set intersects with the existing item set, the support of the frequent item set is likely to exceed the critical threshold due to the change of the incremental sample, or the historical support of the frequent item set is close to the threshold, the critical item set may become a frequent item set or a non-frequent item set after updating, a specific method needs to be taken to handle it. For example, maintain the support boundary zone item set index table, or add the item set appearing in the incremental data with a frequency exceeding the early warning threshold to the recalculation queue, and then label the affected item set in the global candidate item set list to enter the recalculation process. In this process, the FP-growth algorithm is used to obtain the frequent item set of different groups of new data, and the similarity between the new data and the original data set of the reference scale is calculated. The frequent item set is compared with the initial frequent item set BSi of the i-th reference scale data set in the global candidate item set list, the intersection and difference set are calculated, and the global candidate item set list is updated.

[0083] For unknown count estimation, the frequent item set is updated by comparing the new data set (D2) and the original data set (D1), and the update is based on the calculation of similarity. When comparing D1 and D2, set operations are involved, and the formulas used include intersection D1∩D2, difference set D1−D2 and D2−D1.

[0084] Step S4: Dynamically adjusting the parameters of the joint distribution according to the updated global candidate item set list, updating the support of the frequent item set according to the updated joint distribution, and thereby updating the global candidate item set list.

[0085] Specifically, after the sensor data set of the UUV is updated, the new data The marginal distribution of the sensor data is updated, and the parameters of the Copula function are corrected based on the updated marginal distribution:

[0086] ;

[0087] In the formula, is the updated correlation coefficient between the sensor data x and y; represents the value of the parameter that makes the objective function reach the maximum value; is the original correlation coefficient between the sensor data x and y.

[0088] The joint distribution of the sensor data is updated using the updated correlation coefficient, and the support of the frequent item set is recalculated.

[0089] Through the above process, only the changed parameters need to be updated, avoiding repeated scanning of the full data set. The updated parameters in this process are shown in Table 1.

[0090] Table 1. Updated parameters

[0091]

[0092] In the embodiment of the present application, in order to meet the real-time data mining needs of the UUV system, distributed computing framework (Spark) is used for parallel computing, and low-cost second execution cache strategy (LCDERP) is applied to optimize the use of computing resources. The specific steps are as follows:

[0093] S1: Use TextFile to read the sensor data set of the UUV from the distributed file system HDFS, convert each benchmark scale data set into RDD, mark it as TransactionRDD, and store it in the memory.

[0094] S2: According to the grouping strategy of concept hierarchy, distribute TransactionRDD to different groups, and the number of groups is determined by the number of benchmark scale data sets. Each partition runs a benchmark scale data set.

[0095] S3: Each group uses the FP-growth method to obtain the frequent item set of the benchmark scale data set in this group, and marks it as the elastic distributed data set of the benchmark scale data set frequent item set BSDSFrequentItemRDD.

[0096] S4: Merge the mining results of different groups to obtain the global candidate item set list can_list (<m, itemsets, i, count, sum>), mark it as CandidataItemRDD, where m represents the hierarchical identifier of the benchmark scale, that is, the sensor type; itemsets represent the frequent item set collection, that is, the set of items mined from the sensor data that meet the support threshold; i represents the index of the benchmark scale data set to which the frequent item set belongs, used to locate the source data set of the item set; count represents the number of occurrences of the frequent item set in the corresponding benchmark scale data set; sum represents the cumulative number of occurrences of the frequent item set in the global range.

[0097] S5: Estimate and update can_list for unknown item sets.

[0098] S6: Set the support threshold min_sup, and filter the frequent item sets with support greater than min_sup and store them in the frequent item set list FI_list (<itemset, count>), mark it as the target scale frequent item set elastic distributed data set TSFrequentItemRDD, and itemset represents the frequent item set.

[0099] S7: Convert the newly added different benchmark scale data sets into RDD and distribute them to the corresponding groups.

[0100] S8: the FP-growth method is used to obtain the frequent item set of each group of new data, the similarity between the new data and the benchmark scale original data set is calculated after the frequent item set of the benchmark scale data set is obtained, the frequent item set is compared with the corresponding BSi in the can_list, the intersection difference set is calculated and the can_list is updated.

[0101] S9: according to the updated sum value in the can_list, the frequent item set meeting the target scale is filtered out.

[0102] Through the above method, the Spark framework is used to parallelize the processing of each benchmark scale data set, so that the calculation time and memory consumption can be reduced, thereby ensuring the calculation efficiency of the UUV in a resource-limited environment.

[0103] The low-cost secondary execution cache strategy LCDERP of the embodiment of the application can effectively avoid re-computing intermediate results when the memory is insufficient, thereby improving the calculation efficiency, as shown in the following formula: Figure 3 The method comprises the following steps:

[0104] S1: select a part of all sensor data for testing, record the dependency times of each RDD and the calculation time , calculate the generation value of each RDD , store all the generation values of the RDDs in a one-dimensional array , and sort all the generation values;

[0105] S3: test all sensor data, when the RDD is used, update the dependency times of the RDD , and eliminate the RDDs with the dependency times of 0;

[0106] S4: when the memory is insufficient, eliminate the RDD with the minimum generation value in the cache according to the sorting result and update .

[0107] In order to verify the effectiveness of the method provided by the embodiment of the application, comparative experiments are carried out based on UUV sensor data:

[0108] Experimental design: 100,000 transaction data are processed in 5 batches, and the traditional FP-Growth and the method of the embodiment of the application are compared.

[0109] Results: the single batch processing time is reduced from 120s to 40s, reduced by 67%; the total processing time is reduced from 600s to 220s, reduced by 63%. The parallel speedup ratio is close to linear, which can effectively meet the real-time demand of the UUV.

[0110] Abnormality detection: Inject 5% sensor abnormal data, the recall rate of the method of the embodiment of the application is increased to 95%, the false positive rate is reduced to 7%, and the traditional method is only 82% and 18% respectively.

[0111] After dynamically adjusting the edge distribution and the Copula parameter, the parameter updating time is only 30% of the traditional method, and the affected item set recalculation ratio is reduced by 70%, avoiding global scanning.

[0112] As shown in Figure 4 the parameter of the Clayton Copula function Figure 4 The comparison chart of the relationship between the low-value abnormality linkage recall rate and the parameter of the Clayton Copula function. The low-value abnormality linkage is an abnormal state in which two or more sensor data appear extremely low values at the same time, and the recall rate is a proportion of the abnormal mode being successfully identified. As can be seen from , when the parameter of the Clayton Copula function is set to 2.5, the recall rate of the low-value abnormality linkage is increased by 25% compared with other parameter values or the traditional method, which indicates that the parameter value can more accurately capture the correlation of the sensor data at extremely low values, and the sensitivity to the abnormal state of the underwater vehicle is strengthened. The result verifies the modeling advantage of the Clayton Copula function on the asymmetric lower tail dependence, and the identification ability of the extreme low-value linkage mode can be optimized by adjusting the parameter .

[0113] As shown in Figure 5 , after the LCDERP strategy is used, the RDD elimination efficiency is increased by 50% when the memory is insufficient, the cost calculation time is <0.1ms, and the measured cache hit rate is increased by 25% compared with the LRU strategy.

[0114] Embodiments of the application aim at the problem of limited data processing resources of UUV, and propose a UUV-oriented incremental frequent item set mining method. By combining a multi-scale incremental frequent item set mining algorithm and parallel computing technology, the use of computing resources is optimized, and the data mining efficiency is significantly improved under the condition of limited data processing resources of UUV. First, the multi-scale theory is introduced to divide the UUV sensor data set into layers, and a joint distribution model with separated edge and dependence is constructed based on the Copula theory. The incremental data fusion is realized by dynamically updating the edge distribution and the Copula parameter, avoiding repeated scanning of the full data set and reducing the consumption of computing resources. On this basis, the parallel computing framework Spark is combined for distributed data processing, effectively improving the speed and efficiency of UUV data mining. Through reasonable resource scheduling and memory management strategy, the memory use efficiency in the incremental frequent item set mining process is ensured, and the data redundancy and computing overhead are reduced.

[0115] To solve the problem of computational complexity in large-scale data mining of UUV, an optimization method combining incremental frequent itemset mining algorithm and low-cost second execution cache strategy (LCDERP) is proposed. This method divides and processes the UUV sensor data set in parallel, and uses a distributed computing framework to distribute different benchmark data sets to different computing nodes for parallel processing, thereby speeding up the mining of frequent itemsets. At the same time, combined with the low-cost cache strategy, the low-cost RDD is cached in the memory in priority, avoiding repeated calculation and data loss, further improving the performance and efficiency of the algorithm, meeting the needs of real-time data analysis and fault prediction of UUV.

[0116] The technical features of the above embodiments can be combined in any manner. To make the description simple, not all possible combinations of the technical features in the above embodiments are described, only the preferred embodiments of the present application are expressed, which are described in more detail and in detail, but it should not be understood as a limitation on the scope of the present application. As long as the combination of these technical features does not exist, it should be considered as the scope of the present application.

[0117] It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of the present application should be subject to the appended claims.

Claims

1. An incremental frequent itemset mining method for unmanned underwater vehicles, characterized in that, Includes the following steps: Step S1: Acquire data from multiple sensors of the unmanned underwater vehicle, perform frequent itemset mining on the data from all sensors, generate initial frequent itemsets for each sensor, and merge all initial frequent itemsets as a global candidate itemset list; Step S2: Calculate the edge distribution of data from each sensor and the correlation coefficient between the data, and construct the joint distribution of the data based on the edge distribution and correlation coefficient; Step S3: Update the global candidate itemset list: Generate frequent itemsets for new sensor data from the unmanned underwater vehicle during navigation, calculate the similarity between the new frequent itemsets and all initial frequent itemsets. If the similarity is greater than a set similarity threshold, merge the new frequent itemset into the corresponding initial frequent itemset; otherwise, add the new frequent itemset as a new itemset to the global candidate itemset list. The expression for calculating the similarity between frequent itemsets is: ； In the formula, Jaccard similarity; , Represents sensor data; Represents the frequent itemsets in sensor data D; Step S4: Dynamically adjust the parameters of the joint distribution based on the updated global candidate set list: ； In the formula, The correlation coefficient between the updated sensor data x and y; The parameter represents the parameter that maximizes the objective function. The possible values ​​of ; For newly added sensor data; , These represent the edge distributions of sensor data x and y, respectively. The correlation coefficient between sensor data x and y; The support of frequent itemsets is updated based on the updated joint distribution, thereby updating the global candidate itemset list; The expression for the support is: ； In the formula, For candidate frequent itemsets Support level; For candidate frequent itemsets The number of times it appears; This represents the total number of transactions.

2. The incremental frequent itemset mining method for unmanned underwater vehicles according to claim 1, characterized in that: The step S1 of generating the initial frequent itemsets for each sensor includes the following steps: performing frequent itemset mining on the data of all sensors to generate candidate frequent itemsets for each sensor, calculating the support of all candidate frequent itemsets, and using the candidate frequent itemsets with support greater than a set threshold as the initial frequent itemsets for each sensor.

3. The incremental frequent itemset mining method for unmanned underwater vehicles according to claim 1, characterized in that: The expression for calculating the edge distribution of the sensor data in step S2 is: ； ； ； In the formula, For sensor data The marginal distribution; For sensor data The marginal probability density; This refers to the number of samples collected from the sensor data. The bandwidth for kernel density estimation; For kernel functions; Let be the observed value of the i-th sensor data point.

4. The incremental frequent itemset mining method for unmanned underwater vehicles according to claim 1, characterized in that: The expression for calculating the correlation coefficient between the data in step S2 is: ； ； ； In the above formula, For Copula functions; , These are the edge distribution values ​​of the sensor data x and y; The correlation coefficient between sensor data x and y; for The bivariate standard normal distribution function; It is the inverse function of the standard normal distribution; These are the dependent parameters of the Copula function.

5. The incremental frequent itemset mining method for unmanned underwater vehicles according to claim 1, characterized in that: In step S2, when constructing the joint distribution of the data, the parameters of the joint distribution are estimated using historical data from the sensors. The expression for estimating the parameters of the joint distribution is as follows: ； ； In the above formula, Let be the maximum likelihood estimation function; , These are the parameters of the Copula function; This represents the number of samples in the historical sensor data. For Copula functions; , For historical sensor data The marginal distribution.

6. The incremental frequent itemset mining method for unmanned underwater vehicles according to claim 1, characterized in that: In step S2, the expression for constructing the joint distribution of the data based on the marginal distribution and correlation coefficient is as follows: ； In the formula, The joint distribution of sensor data x and y; , These represent the edge distributions of sensor data x and y, respectively. The correlation coefficient between sensor data x and y; , These are the marginal probability densities of sensor data x and y, respectively.

7. A caching method applicable to the incremental frequent itemset mining method for unmanned underwater vehicles as described in any one of claims 1-6, characterized in that, Includes the following steps: Step S1: Use a distributed computing framework to convert sensor data into a resilient distributed dataset RDD, and store all RDDs in memory; Step S2: Select a portion of all sensor data for testing, record the number of times each RDD is depended on (N) and the computation time (T), and calculate the cost of each RDD. Store the values ​​of all RDDs in a one-dimensional array C and sort all the values. Step S3: Test all sensor data. When an RDD is used, update the number of times the RDD is depended on, N=N-1, and discard RDDs with a number of times they are depended on, 0. Step S4: When memory is insufficient, evict the RDD with the lowest generation value in the cache according to the sorting result and update C.

Images