Fault criterion mining method and device based on K-means clustering

Abstract

The invention discloses a fault criterion mining method based on K-means clustering, and belongs to the technical field of fault diagnosis, and the method comprises the steps: generating a clustering model through a K-means method; respectively solving a data point contour coefficient and a clustering contour coefficient; an optimal K value is autonomously selected based on contour coefficient evaluation; determining a plurality of generalized parameters with reference significance for fault diagnosis, and counting positive-abnormal scene triggering rates of the generalized parameters to obtain a rule mining result; when the fault judgment rule is edited, an interaction panel is adopted to display a fault diagnosis result and a threshold graphical representation corresponding to the adjusted rule in real time, so that rule formulation and revision are facilitated. According to the method, two fault diagnosis methods of data driving and knowledge driving are fused, so that the stability and the real-time performance of diagnosis are ensured, and the workload of manually compiling and maintaining rules is reduced.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis and processing, and in particular to a fault criterion mining method and device based on K-means clustering. Background Art

[0002] Fault detection methods for complex equipment like aircraft are primarily categorized into two types: data-driven and rule-driven. Data-driven methods analyze existing operational or simulation data, using machine learning or neural networks to identify fault conditions within the target data. Rule-driven methods, on the other hand, analyze fault mechanisms and patterns, then determine fault conditions within the target data by developing a series of expert rules to determine if a fault exists.

[0003] Both methods have their advantages and disadvantages: the data-driven method can deal with unknown fault types. However, on the one hand, for online fault detection scenarios with multiple parameters, high concurrency, and strong real-time requirements, it requires a certain amount of computing power from the computer; on the other hand, the output results of the neural network structure are not stable enough and have low interpretability, which is equivalent to a "black box" for operators and difficult to maintain. The rule-driven method has good interpretability. However, for complex fault detection objects, the rules written may be extremely complex, so that writing rules often requires a large number of professionals to spend a lot of time; not only that, the written rule system is also difficult to deal with unknown fault modes. This makes the rule-driven method difficult to apply to systems that lack historical fault data and have unknown specific fault modes.

[0004] In the prior art, Patent 1 CN110259648A discloses a method for diagnosing wind turbine blade faults based on optimized K-means clustering, and Patent 2 CN110196165A discloses a method for diagnosing rotating machinery ball bearing faults based on K-means clustering and evidence reasoning. Patents 1 and 2 use K-means to achieve data-driven anomaly detection or fault classification. They are data analysis methods designed based on normal data or multi-type fault data. In essence, they directly interpret fault data through data-driven methods. This patent is mainly designed for writing manual expert rules. Through K-means clustering of known normal data samples and single-type fault data, it provides expert rules mined from data in specific fault scenarios, thereby assisting manual experts in further modifying the rules to form specific rules describing the fault.

[0005] Therefore, in view of the current situation in which data-driven methods in automated system fault diagnosis are insufficiently interpretable and manual rule writing is extremely complex, it is necessary to propose a fault diagnosis model that combines the advantages of the two existing methods, while ensuring stability and real-time performance while avoiding a large amount of manual work. Summary of the Invention

[0006] The present invention aims to solve the problems of insufficient interpretability of data-driven methods and extremely complex manual rule writing in automated system fault diagnosis. A fault criterion mining method and device based on K-means clustering are proposed. By combining the advantages of two types of existing methods, a fault diagnosis model can avoid a large amount of manual work while ensuring stability and real-time performance.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution of the present invention is as follows:

[0008] A fault criterion mining method based on K-means clustering, comprising:

[0009] Input the normal timing data and abnormal timing data of the target device;

[0010] Performing K-means clustering on the normal time series data and the abnormal time series data respectively, thereby extracting the numerical distribution range of the normal time series data and the abnormal time series data;

[0011] Based on the principal component analysis method of singular value decomposition, multiple principal component characteristic parameters with the most reference value for fault diagnosis are determined from the input data. Then, the positive-abnormal scene triggering rate of each principal component characteristic parameter is counted, and the rule mining results are obtained based on the statistical results.

[0012] The rule mining results, fault diagnosis status and threshold diagram information are displayed in an editable interactive panel.

[0013] Furthermore, the rule content in the interactive panel is an editable part. When the operator edits the fault determination rule, the interactive panel displays the fault diagnosis result and threshold graphic information corresponding to the adjusted rule in real time.

[0014] Furthermore, when performing K-means clustering on the normal time series data and abnormal time series data, the range in which the optimal K value may exist is first manually specified, and then all possible clustering models are constructed within the interval, and the K value with the best clustering effect is determined based on the silhouette coefficient of each model.

[0015] Furthermore, the K-means clustering model is constructed as follows: First, K data points are randomly selected in space as the centers of K clusters. Then, the remaining data points are assigned to the corresponding clusters according to their distance from each center. The mean of all points in each cluster is calculated, and the K mean points are used as the new K cluster centers. The assignment process is repeated, and this process is repeated until all clusters remain unchanged. At this point, the clustering model is constructed.

[0016] Furthermore, the silhouette coefficient of each model can be expressed as the average of the silhouette coefficients of all data points in the model, that is:

[0017]

[0018] Where s(i) is the silhouette coefficient of a single data point.

[0019] Furthermore, the silhouette coefficient of a single data point is calculated as follows:

[0020]

[0021] Where i represents a data point; a(i) represents the average distance between point i and the rest of the points in the cluster. If there are no other points, a(i) = 0. The average distance between point i and the points in K-1 other clusters is calculated, and the minimum value of the K-1 average distances is b(i).

[0022] Furthermore, the statistical analysis of the positive-abnormal scene triggering rate of each principal component characteristic parameter includes:

[0023] First, based on the mixed Gaussian distribution hypothesis, the distribution center and radius of each cluster distribution of each principal component characteristic parameter in the clustering model of normal time series data and abnormal time series data are determined. The upper and lower limit thresholds of the distribution of each cluster of normal time series data and abnormal time series data are determined based on the distribution center and radius data; then the ratio of healthy data in normal time series data and abnormal time series data within the above threshold range is examined respectively.

[0024] Furthermore, obtaining rule mining results based on statistical results includes:

[0025] If the ratio of healthy data within the threshold range of each cluster of normal time series data and abnormal time series data meets the fault tolerance requirement, the rule corresponding to the principal component characteristic parameter under the cluster distribution is retained; otherwise, the corresponding rule is not retained.

[0026] The present invention also proposes a fault criterion mining device based on K-means clustering, comprising:

[0027] a data input module configured to input normal time series data and abnormal time series data of a target device;

[0028] A clustering model configuration module is configured to perform K-means clustering on the normal time series data and the abnormal time series data respectively, thereby extracting the numerical distribution range of the normal time series data and the abnormal time series data;

[0029] The fault mining module is configured to determine multiple principal component characteristic parameters that are most meaningful for fault diagnosis from the input data based on the principal component analysis method of singular value decomposition, then calculate the positive-abnormal scenario triggering rate of each principal component characteristic parameter, and obtain rule mining results based on the statistical results;

[0030] The interactive rule revision module is configured to display rule mining results, fault diagnosis status and threshold diagram information in an editable interactive panel.

[0031] Furthermore, the interactive panel includes a rule editing unit and an instant display unit. The rule editing unit is used by the operator to edit the fault judgment rules. The instant display unit is configured to display the fault diagnosis results and threshold graphic information corresponding to the adjusted rules in real time according to the editing content of the rule editing unit.

[0032] Furthermore, the fault mining module is configured to count the positive-abnormal scene triggering rate of each principal component characteristic parameter in the following manner: first, based on the mixed Gaussian distribution hypothesis, determine the distribution center and radius of each cluster distribution of each principal component characteristic parameter in the normal time series data and abnormal time series data clustering model, and determine the upper and lower limit thresholds of each cluster distribution of normal time series data and abnormal time series data based on the distribution center and radius data; then examine the ratio of healthy data in normal time series data and abnormal time series data within the above-mentioned threshold range.

[0033] Furthermore, the fault mining module is configured to obtain rule mining results in the following manner, including:

[0034] If the ratio of healthy data within the threshold range of each cluster of normal time series data and abnormal time series data meets the fault tolerance requirement, the rule corresponding to the principal component characteristic parameter under the cluster distribution is retained; otherwise, the corresponding rule is not retained.

[0035] In summary, the present invention has the following advantages:

[0036] 1. The present invention integrates data-driven and knowledge-driven fault diagnosis methods to ensure the stability and real-time performance of diagnosis, and reduces the workload of manual rule writing and maintenance.

[0037] 2. This invention maps existing data into a high-dimensional linear space for cluster analysis, thereby mining fault edge criteria and implementing rule writing, providing a basis for building a fault diagnosis system with strong interpretability and high degree of automation;

[0038] 3. The present invention realizes adaptive clustering configuration based on boundary fuzziness through the K-means algorithm, and realizes fault judgment rule mining based on the statistics of positive-abnormal scene triggering rate based on the principal component analysis method. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A logical diagram of the adaptive configuration of the clustering model in the present invention;

[0040] Figure 2 A schematic diagram of an interaction panel used in the interaction rule revision step of the present invention;

[0041] Figure 3This is the fault criterion mining process for the wheel brake device fault scenario in Example 1 of the present invention;

[0042] Figure 4 The normal time series data and abnormal time series data set of the flight parameters in Example 1 of the present invention;

[0043] Figure 5 Schematic diagram of the adaptive K-means clustering method based on silhouette coefficient evaluation in Example 1 of the present invention;

[0044] Figure 6 For the general Figure 4 The data in the figure is redrawn according to the K-means clustering principle;

[0045] Figure 7 This is a schematic diagram showing the interactive panel before expert adjustment;

[0046] Figure 8 Schematic diagram of the interaction panel display adjusted for experts. DETAILED DESCRIPTION

[0047] In order to more clearly illustrate the present invention, the present invention is further described below in conjunction with preferred embodiments and drawings. Those skilled in the art should understand that the following specific description is illustrative rather than restrictive and should not be used to limit the scope of protection of the present invention.

[0048] The present invention provides a fault criterion mining method based on K-means clustering, comprising the following steps:

[0049] Step 1: Data Acquisition

[0050] Analyze relevant parameters that can reflect the target device failure scenario, and obtain normal time series data and abnormal time series data of the aforementioned relevant parameters.

[0051] Step 2: Configure the clustering model

[0052] K-means clustering is performed on normal time series data and abnormal time series data respectively. Normal time series data represents the parameter distribution range of normal scenarios, while abnormal time series data represents the parameter distribution range of abnormal scenarios. This step uses the adaptive K-means clustering method to mine multi-segment thresholds for multi-parameter coupling based on given implementation data; and by applying the adaptive K-means clustering method to normal time series data and abnormal time series data respectively, the distribution ranges of normal time series data and abnormal time series data can be extracted.

[0053] In the K-means algorithm, each piece of data is treated as a point in a high-dimensional linear space. All data is ultimately divided into a finite number of clusters, ensuring that points within each cluster are as close as possible and points between different clusters are as far apart as possible. The algorithm works as follows: First, K data points are randomly selected from the space as the centers of the K clusters. The remaining data points are then assigned to the corresponding clusters based on their distance from each center. The mean of all points in each cluster is calculated, and the K mean points are used as the new centers of the K clusters. This assignment process is repeated until all clusters remain unchanged, completing the clustering model.

[0054] K-means clustering relies on the selection of the hyperparameter K (i.e., the number of clusters). This step is based on the silhouette coefficient evaluation to autonomously select the optimal K value, thereby improving the adaptive ability of clustering. The silhouette coefficient is a number in the interval [-1,1]. The better the clustering effect, the larger its value. In the specific implementation, the range of possible optimal K values is first manually specified, and then all possible clustering models are constructed within this interval. The K value with the best clustering effect is determined based on the silhouette coefficient of each model. For reference, Figure 1 shown.

[0055] Specifically, the silhouette coefficient of the clustering model is solved as follows:

[0056] For a clustering model with m data points, its silhouette coefficient can be expressed as the average of the silhouette coefficients of all data points, that is:

[0057]

[0058] Where s(i) is the silhouette coefficient of a single data point, which is calculated as follows:

[0059]

[0060] Where i represents a data point; a(i) represents the average distance between point i and the rest of the points in the cluster. If there are no other points, a(i) = 0. The average distance between point i and the points in K-1 other clusters is calculated, and the minimum value of the K-1 average distances is b(i).

[0061] Step 3: Fault rule mining

[0062] First, using principal component analysis, we identify several generalized parameters (i.e., principal component characteristic parameters) that are most valuable for fault diagnosis from the relevant parameters that reflect the fault status of the target equipment. We then examine these generalized parameters and calculate their positive-abnormal scenario triggering rates to obtain rule mining results.

[0063] a. Determine the generalized parameters

[0064] Principal component analysis compresses information from a high-dimensional space into a low-dimensional space. For a large amount of information in a high-dimensional space, the directions with the largest data variance are considered the primary information. In practice, principal component analysis employs singular value decomposition (SVD) to obtain generalized parameters whose parameters satisfy certain mathematical relationships. This step simplifies a large number of actual parameters into a small number of generalized parameters within a domain of orthogonal parameters in linear space, reducing unnecessary computational effort for subsequent rule determination and optimizing performance.

[0065] b. Rule determination

[0066] First, based on the Gaussian mixture distribution hypothesis, the distribution centers and radii of each cluster distribution of each set of generalized parameters in the normal and faulty data sets are determined. The upper and lower thresholds for each cluster distribution of normal and faulty data are determined based on the distribution centers and radii. Next, the ratio of healthy data within the threshold range is examined in both normal and faulty data. If the ratio meets the fault tolerance requirement, the threshold (i.e., rule) corresponding to the generalized parameter in that cluster distribution is retained; otherwise, the corresponding threshold (i.e., rule) is discarded.

[0067] Preferably, in the present invention, the fault tolerance requirement is that the proportion of the cluster fault data belonging to the normal data distribution range does not exceed 5%.

[0068] Step 4: Modify the interaction rules

[0069] This step relies on Figure 2 The operator can view rule-related information, fault diagnosis status, and threshold and parameter trend information on the panel.

[0070] Among them, the rule content is the editable part. When the operator edits the fault judgment rule, the interactive panel displays the fault diagnosis results and threshold diagrams corresponding to the adjusted rule in real time to facilitate the formulation and revision of the rule.

[0071] Example 1

[0072] The following example illustrates the application of the fault criterion mining method based on K-means clustering proposed in the present invention in a wheel brake device failure scenario.

[0073] The wheel brake failure scenario is mainly reflected in the two flight parameters of wheel brake amount setting and wheel hydraulic brake pressure. Once the wheel brake failure occurs, the wheel brake amount setting and the wheel hydraulic brake pressure at the relevant time will no longer be coordinated, such as Figure 4 When anomaly detection detects an anomaly in the flight parameter timing of a fault scenario, the normal timing and the abnormal timing can be separated according to the interpretation result of the anomaly detection.

[0074] Refer to Figure 3 The fault criterion mining process for the wheel brake device fault scenario shown in the figure is based on the K-means clustering method as follows:

[0075] Step 1: Clustering model configuration

[0076] K-means clustering is performed on normal time series data and abnormal time series data respectively, and the optimal K value is selected based on the silhouette coefficient evaluation, thereby improving the adaptive ability of clustering. In this embodiment, the adaptive K-means clustering method based on silhouette coefficient evaluation is as follows: Figure 5 shown.

[0077] The silhouette coefficient of the clustering model is solved as follows:

[0078] For a clustering model with m data points, its silhouette coefficient can be expressed as the average of the silhouette coefficients of all data points, that is:

[0079]

[0080] Where s(i) is the silhouette coefficient of a single data point, which is calculated as follows:

[0081]

[0082] Where i represents a data point; a(i) represents the average distance between point i and the rest of the points in the cluster. If there are no other points, a(i) = 0. The average distance between point i and the points in K-1 other clusters is calculated, and the minimum value of the K-1 average distances is b(i).

[0083] For example, Figure 4 The normal time series and abnormal time series data shown are redrawn according to the K-means clustering principle, and the following can be obtained: Figure 6 The effect shown. Figure 6 As can be seen in the figure, although abnormal time series are relatively far apart from normal time series, the anomaly detection algorithm itself may have false alarms due to its own limitations, causing some overlap in the distribution of positive abnormal time series. If only threshold mining is performed on abnormal time series, the false alarm areas corresponding to the overlapping parts will affect the accuracy of the judgment in subsequent use.

[0084] The adaptive K-means clustering method can be used to mine multi-segment thresholds of multi-parameter coupling based on given implementation data, and applying the adaptive K-means clustering method to normal data and abnormal data respectively can extract the distribution range of normal data and the distribution range of abnormal data respectively.

[0085] Step 2: Fault rule mining

[0086] First, remove the numerical distribution range in the abnormal timing data that overlaps too much with the normal timing data (i.e., the range exceeds 5%), and the difference set ideally corresponds to the numerical distribution range of the abnormal timing scenario.

[0087] As Figure 6 shown, the figure shows removing the part of the abnormal timing data distribution that overlaps significantly with the normal timing data distribution, thereby obtaining the upper and lower limits of the threshold of the abnormal timing data shown by the dashed line (there may be multiple segments), such as [wheel hydraulic brake pressure low, wheel hydraulic brake pressure upp, wheel brake quantity given low, wheel brake quantity given upp], where upp and low represent the upper and lower limits of the threshold respectively. Further, according to the template of the K-means method, the upper and lower limits of the threshold can be expressed as [(region 1 parameter 1 low < parameter 1 < region 1 parameter 1 upp and region 2 parameter 2 low < parameter 2 < region 2 parameter 2 upp) and …] and …, and filling the threshold according to the aforementioned template can obtain the corresponding rule. Then, according to Figure 5 the mined rule is: (wheel hydraulic brake pressure low < wheel hydraulic brake pressure < wheel hydraulic brake pressure upp) and (wheel brake quantity given low < wheel brake quantity given < wheel brake quantity given upp).

[0088] For the rule mining involving more parameters, the threshold template of a single parameter may have a large redundancy. Therefore, it is necessary to further simplify it using the PCA method to obtain the principal component characteristic parameters. For example, (3 < A < 4 and 3 < B < 4) or … or (50 < A < 51 and 50 < B < 51) can be simplified to -0.001 < A - B < 0.001, that is, the principal component characteristic parameter “A - B” reflecting a certain numerical relationship between the two is obtained from the threshold template of the single parameters A and B.

[0089] Step 3. Interactive rule correction

[0090] Let experts make further modifications according to the results through the interactive panel.

[0091] For example, Figure 7 as the display of the interactive panel before the expert's adjustment, it can be seen from the figure that there is a certain amount of noise in the wheel hydraulic brake pressure. At this time, the expert can perform sliding window smoothing through the mean function Mean((wheel hydraulic brake pressure, 10)), and see that the smoothed timing has good discrimination in the interactive panel, such as Figure 8The display of the interactive panel after expert adjustment is shown. The positive and abnormal data can be distinguished with 0.07 as the demarcation line. Therefore, the expert rule has been changed from "wheel hydraulic brake pressure low < wheel hydraulic brake pressure < wheel hydraulic brake pressure upp" to "wheel hydraulic brake pressure low < Mean(wheel hydraulic brake pressure, 10) < wheel hydraulic brake pressure upp".

[0092] The above are only the preferred embodiments of the present invention, and do not impose any formal limitations on the present invention. Any simple modification or equivalent change made to the above embodiments based on the technical essence of the present invention falls within the protection scope of the present invention.

[0093] Embodiment 2

[0094] Based on the same inventive concept, an embodiment of the present invention further provides a fault criterion mining device based on K-means clustering, including a data input module, a clustering model configuration module, a fault mining module, and an interactive rule revision module.

[0095] The data input module is configured to input the normal time series data and abnormal time series data of the target device;

[0096] The clustering model configuration module is configured to perform K-means clustering on the normal time series data and abnormal time series data respectively, so as to extract the numerical distribution ranges of the normal time series data and abnormal time series data;

[0097] The fault mining module is configured to determine multiple principal component feature parameters most significant for fault diagnosis from the input data according to the principal component analysis method of singular value decomposition, and then count the positive-abnormal scenario trigger rates of each principal component feature parameter, and obtain the rule mining result according to the statistical result;

[0098] The interactive rule revision module is configured to display the rule mining result, the fault diagnosis status, and the threshold diagram information on an editable interactive panel.

[0099] Further, the interactive panel includes a rule editing unit and an instant display unit. The rule editing unit is used for an operator to edit the fault determination rule, and the instant display unit is configured to display the fault diagnosis result and the threshold diagram information corresponding to the adjusted rule in real time according to the editing content of the rule editing unit.

[0100] Furthermore, the fault mining module is configured to count the positive-abnormal scene triggering rate of each principal component characteristic parameter in the following manner: first, based on the mixed Gaussian distribution hypothesis, determine the distribution center and radius of each cluster distribution of each principal component characteristic parameter in the normal time series data and abnormal time series data clustering model, and determine the upper and lower limit thresholds of each cluster distribution of normal time series data and abnormal time series data based on the distribution center and radius data; then examine the ratio of healthy data in normal time series data and abnormal time series data within the above-mentioned threshold range.

[0101] Furthermore, the fault mining module is configured to obtain rule mining results in the following manner, including: if the ratio of healthy data within the threshold range of each cluster of normal time series data and abnormal time series data meets the fault tolerance requirement, then the rule corresponding to the principal component characteristic parameter under the cluster distribution is retained; otherwise, the corresponding rule is not retained.

Claims

1. A fault criterion mining method based on K-means clustering, characterized in that: include: Input the normal timing data and abnormal timing data of the target device; Performing K-means clustering on the normal time series data and the abnormal time series data respectively, thereby extracting the numerical distribution range of the normal time series data and the abnormal time series data; Based on the principal component analysis method of singular value decomposition, multiple principal component characteristic parameters with the most reference value for fault diagnosis are determined from the input data. Then, the positive-abnormal scene triggering rate of each principal component characteristic parameter is counted, and the rule mining results are obtained based on the statistical results. The rule mining results, fault diagnosis status and threshold diagram information are displayed in an editable interactive panel.

2. The fault criterion mining method based on K-means clustering according to claim 1, characterized in that: The rule content in the interactive panel is an editable part. When the operator edits the fault judgment rule, the interactive panel displays the fault diagnosis results and threshold graphic information corresponding to the adjusted rule in real time.

3. The fault criterion mining method based on K-means clustering according to claim 1, characterized in that: When performing K-means clustering on the normal time series data and the abnormal time series data, the range in which the optimal K value may exist is first manually specified, and then all possible clustering models are constructed within the interval, and the K value with the best clustering effect is determined based on the silhouette coefficient of each model.

4. A fault criterion mining method based on K-means clustering according to claim 1 or 3, characterized in that: The K-means clustering model is constructed as follows: First, K data points are randomly selected in space as the centers of K clusters. The remaining data points are then assigned to the corresponding clusters according to their distance from each center. The mean of all points in each cluster is calculated, and the K mean points are used as the new centers of the K clusters. The assignment process is repeated until all clusters remain unchanged, and the clustering model is constructed.

5. The fault criterion mining method based on K-means clustering according to claim 3, characterized in that: The silhouette coefficient of each model can be expressed as the average of the silhouette coefficients of all data points in the model, that is: Where s(i) is the silhouette coefficient of a single data point.

6. The fault criterion mining method based on K-means clustering according to claim 5, characterized in that: The silhouette coefficient of a single data point is calculated as follows: Where i represents a data point; a(i) represents the average distance between point i and the rest of the points in the cluster. If there are no other points, a(i) = 0. The average distance between point i and the points in K-1 other clusters is calculated, and the minimum value of the K-1 average distances is b(i).

7. The fault criterion mining method based on K-means clustering according to claim 1, characterized in that: The statistical analysis of the positive-abnormal scene triggering rate of each principal component characteristic parameter includes: First, based on the mixed Gaussian distribution hypothesis, the distribution center and radius of each cluster distribution of each principal component characteristic parameter in the clustering model of normal time series data and abnormal time series data are determined. The upper and lower limit thresholds of the distribution of each cluster of normal time series data and abnormal time series data are determined based on the distribution center and radius data; then the ratio of healthy data in normal time series data and abnormal time series data within the above threshold range is examined respectively.

8. The fault criterion mining method based on K-means clustering according to claim 7, characterized in that: The rule mining results obtained according to the statistical results include: If the ratio of healthy data within the threshold range of each cluster of normal time series data and abnormal time series data meets the fault tolerance requirement, the rule corresponding to the principal component characteristic parameter under the cluster distribution is retained; otherwise, the corresponding rule is not retained.

9. A fault criterion mining device based on K-means clustering, characterized in that: include: a data input module configured to input normal time series data and abnormal time series data of a target device; A clustering model configuration module is configured to perform K-means clustering on the normal time series data and the abnormal time series data respectively, thereby extracting the numerical distribution range of the normal time series data and the abnormal time series data; The fault mining module is configured to determine multiple principal component characteristic parameters that are most meaningful for fault diagnosis from the input data based on the principal component analysis method of singular value decomposition, then calculate the positive-abnormal scenario triggering rate of each principal component characteristic parameter, and obtain rule mining results based on the statistical results; The interactive rule revision module is configured to display rule mining results, fault diagnosis status and threshold diagram information in an editable interactive panel.

10. The fault criterion mining device based on K-means clustering according to claim 9, characterized in that: The interactive panel includes a rule editing unit and an instant display unit. The rule editing unit is used by operators to edit fault judgment rules. The instant display unit is configured to display the fault diagnosis results and threshold graphic information corresponding to the adjusted rules in real time according to the editing content of the rule editing unit.

11. The fault criterion mining device based on K-means clustering according to claim 9, characterized in that: The fault mining module is configured to count the positive-abnormal scene triggering rate of each principal component characteristic parameter in the following manner: first, based on the mixed Gaussian distribution hypothesis, determine the distribution center and radius of each cluster distribution of each principal component characteristic parameter in the normal time series data and abnormal time series data clustering model, and determine the upper and lower limit thresholds of each cluster distribution of normal time series data and abnormal time series data based on the distribution center and radius data; then, examine the ratio of healthy data appearing in normal time series data and abnormal time series data within the aforementioned threshold range.

12. The fault criterion mining method based on K-means clustering according to claim 11, characterized in that: The fault mining module is configured to obtain rule mining results in the following manner, including: If the ratio of healthy data within the threshold range of each cluster of normal time series data and abnormal time series data meets the fault tolerance requirement, the rule corresponding to the principal component characteristic parameter under the cluster distribution is retained; otherwise, the corresponding rule is not retained.

Images