Hybrid model fault early warning method and system based on time series prediction and fuzzy logic

Abstract

The application provides a kind of hybrid model fault early warning method and system based on timing prediction and fuzzy logic, device state feature set is constructed by collecting the operation and maintenance data of equipment in continuous operation cycle, feature correlation analysis is carried out on device state feature set, device state potential vector reflecting the dynamic coupling relationship between parameters is generated, fuzzy rule matching is carried out by calling pre-trained fuzzy logic inference model, and fuzzy state set containing multi-dimensional fuzzy subset is generated;Fuzzy state set is processed by space-time correlation, and multi-dimensional decision cloud chart is generated, which is reconstructed in space by quantization weight distribution mechanism, and the optimized decision matrix is obtained;Based on the confidence distribution characteristics corresponding to each fault type, the fault early warning signal containing fault early warning level and fault location positioning information is generated.The application can improve the accuracy and timeliness of fault early warning, and provides reliable decision support for equipment operation and maintenance management.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, and in particular to a fault early warning method and system based on a hybrid model of time-series prediction and fuzzy logic. Background Technology

[0002] With the development of industrial intelligence, the analysis and processing of equipment operating status data can identify potential fault risks and issue early warning signals in advance, which is of great significance in ensuring the safe and stable operation of equipment and reducing maintenance costs. Currently, the common approach is to collect equipment operating parameters and analyze them by setting fixed thresholds or simple models to determine whether the equipment has fault risks. However, there are complex dynamic coupling relationships between parameters during equipment operation, and they are also affected by multiple environmental factors. Existing technologies are lacking in capturing the nonlinear correlations and dynamic evolution characteristics between parameters, and their ability to handle fuzziness and uncertainty in the operating state is limited, resulting in insufficient accuracy and timeliness of fault warnings. Therefore, improving the accuracy and adaptability of fault warnings has become a key research focus in current equipment operation and maintenance management. Summary of the Invention

[0003] In view of this, the present invention provides a hybrid model fault early warning method and system based on time-series prediction and fuzzy logic.

[0004] The technical solution of this invention is implemented as follows:

[0005] On one hand, embodiments of the present invention provide a hybrid model fault early warning method based on time-series prediction and fuzzy logic. The method includes: collecting operation and maintenance data of equipment during continuous operation cycles, constructing an equipment state feature set, wherein the equipment state feature set includes feature sequences of various operating parameters of the equipment at different monitoring times and corresponding environmental impact feature sequences; performing feature correlation analysis on the equipment state feature set to generate an equipment state potential vector reflecting the dynamic coupling relationship between parameters, wherein the equipment state potential vector includes trend evolution features and cross-influence features of the feature sequences; calling a pre-trained fuzzy logic inference model to perform fuzzy rule matching on the equipment state potential vector to generate a fuzzy state set containing multi-dimensional fuzzy subsets; performing spatiotemporal correlation processing on the fuzzy state set in conjunction with a preset domain fault knowledge base to generate a multi-dimensional decision cloud map with spatiotemporal correlation characteristics; performing spatial reconstruction operation on the multi-dimensional decision cloud map through a quantum weight allocation mechanism to obtain an optimized decision matrix; and generating a fault early warning signal containing fault early warning level and fault location information through a dynamic threshold determination mechanism based on the confidence distribution features corresponding to each fault type in the optimized decision matrix, and pushing the fault early warning signal to the equipment operation and maintenance management terminal.

[0006] On the other hand, embodiments of the present invention provide a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above-described method.

[0007] This invention provides a hybrid model fault early warning method based on time-series prediction and fuzzy logic. It constructs a device state feature set by collecting maintenance data from continuous equipment operation cycles, comprehensively capturing dynamic changes in equipment operating parameters and environmental influencing factors. By performing feature correlation analysis on the device state feature set to generate a device state potential vector, it achieves a quantitative representation of the dynamic coupling relationship between parameters, overcoming the limitations of traditional methods that analyze parameters in isolation. It calls a fuzzy logic inference model to perform fuzzy rule matching on the device state potential vector, effectively handling the fuzziness and uncertainty in the device's operating state and improving adaptability to complex operating conditions. Combining a domain fault knowledge base with spatiotemporal correlation processing of the fuzzy state set and generating an optimized decision matrix through a quantum weight allocation mechanism, it achieves deep integration of domain knowledge and data-driven approaches, enhancing the reliability and robustness of the decision-making process. Based on the optimized decision matrix, a fault early warning signal is generated through a dynamic threshold judgment mechanism, which can dynamically adjust the judgment criteria according to the actual operating state of the equipment, improving the accuracy and timeliness of fault early warning and providing reliable decision support for equipment operation and maintenance management. Attached Figure Description

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a hybrid model fault early warning method based on time-series prediction and fuzzy logic, provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] This invention provides a hybrid model fault early warning method based on time-series prediction and fuzzy logic, which can be executed by a processor of a computer system. The computer system can refer to a remote server, a laptop, tablet, desktop computer, or other device with data processing capabilities at the maintenance site.

[0011] Figure 1 This is a schematic diagram illustrating the implementation process of a hybrid model fault early warning method based on time-series prediction and fuzzy logic provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0012] Step S100: Collect the operation and maintenance data of the equipment during the continuous operation cycle and construct the equipment status feature set. The equipment status feature set includes the feature sequence of each operating parameter of the equipment at different monitoring times and the corresponding environmental impact feature sequence.

[0013] Equipment operation and maintenance data during a continuous operating cycle refers to various operation-related data generated during a complete, uninterrupted operating cycle. This includes operating parameters such as temperature, pressure, and rotational speed, as well as environmental influencing factors like ambient temperature, humidity, and air pressure. The equipment status characteristic set is a collection containing the characteristic sequences of each equipment operating parameter at different monitoring times. This sequence, formed by arranging each operating parameter according to different monitoring times, reflects the changes in that parameter over time. The corresponding environmental impact characteristic sequence is a sequence of environmental influencing factors associated with each operating parameter, arranged according to monitoring times, used to demonstrate the impact of environmental factors on equipment operation.

[0014] In practical applications, various sensors can be used to acquire operational and maintenance data. For example, for equipment temperature parameters, temperature sensors can be installed on key parts of the equipment and data can be collected at set time intervals to obtain the characteristic sequence of equipment temperature at different monitoring times. For ambient temperature, ambient temperature sensors can be installed in the environment where the equipment is located and data can also be collected at fixed time intervals to form a characteristic sequence of ambient temperature. By organizing and summarizing these collected data, a set of equipment status characteristics can be constructed.

[0015] Step S200: Perform feature correlation analysis on the equipment state feature set to generate an equipment state potential vector that reflects the dynamic coupling relationship between parameters. The equipment state potential vector includes the trend evolution features and cross-influence features of the feature sequence.

[0016] Feature correlation analysis refers to the in-depth study and analysis of the correlation relationships between various feature sequences in a set of equipment state features, in order to identify their interactions and influences. Dynamic coupling relationships between parameters refer to the interrelationships and mutual influences between various operating parameters of the equipment as they change over time. The equipment state potential vector is a vector that comprehensively reflects the equipment state; its trend evolution characteristics reflect the overall changing trend of the feature sequences over time, such as whether a certain operating parameter of the equipment shows an upward or downward trend; cross-influence characteristics reflect the degree of mutual influence between different feature sequences, such as how a change in one operating parameter affects other operating parameters.

[0017] As one implementation method, step S200 can be specifically implemented as the following steps S210~S260:

[0018] Step S210: Perform time series decomposition on each feature sequence in the equipment state feature set, extract the trend component and fluctuation component of each feature sequence at different time scales, and obtain the decomposed multi-scale feature sequence set.

[0019] Time series decomposition involves breaking down a feature series according to different time scales to separate its trend and fluctuation components. The trend component reflects the overall trend of the feature series over a longer time scale, such as whether a certain operating parameter of equipment continuously increases or decreases over a long period. The fluctuation component reflects the local fluctuation characteristics of the feature series over a shorter time scale, such as the small fluctuations in the operating parameter of equipment over a short period. A multi-scale feature series set is the collection obtained after decomposition, containing the trend and fluctuation components of each feature series at different time scales.

[0020] As one implementation method, step S210 can be specifically implemented as the following steps S211~S215:

[0021] Step S211: Initialize the time scale decomposition parameter set. The time scale decomposition parameter set includes multiple preset time window lengths and corresponding decomposition levels. The lengths of each time window are distributed in a geometric sequence.

[0022] The time-scale decomposition parameter set is a set of parameters used to guide the time series decomposition process. The time window length refers to the time range selected during time series decomposition; different time window lengths correspond to different time scales. The number of decomposition levels indicates the number of levels to decompose the feature sequence, with each level corresponding to a different time scale. Distributing the time window lengths in a geometric progression allows the decomposition process to more reasonably cover different time scales. During initialization, the time window length and number of decomposition levels need to be preset based on the equipment's operating characteristics and analytical requirements. For example, for equipment with a long operating cycle and slow changes, a larger preset time window length and more decomposition levels can be used. Assuming the preset initial time window length is 10 time units, the common ratio is 2, and the number of decomposition levels is 5, then the time window lengths would be 10, 20, 40, 80, and 160 time units respectively.

[0023] Step S212: For each feature sequence, the wavelet transform algorithm is used to perform multi-scale decomposition processing according to the length of each time window in the time scale decomposition parameter set, generating approximation coefficients and detail coefficients at different decomposition levels.

[0024] Wavelet transform is an algorithm used in signal processing and time series analysis, capable of decomposing a signal into wavelet components at different scales and locations. In this embodiment, it is applied to the multi-scale decomposition of feature sequences. For each feature sequence, wavelet transform is performed sequentially according to the preset time window lengths in the time scale decomposition parameter set. At different decomposition levels, corresponding approximation coefficients and detail coefficients are generated. The approximation coefficients reflect the overall trend information of the feature sequence at that time scale, while the detail coefficients reflect the local fluctuation information of the feature sequence at that time scale. For example, for a feature sequence of device temperature, when the time window length is 10 time units, the wavelet transform yields the first-level approximation coefficients and detail coefficients. The approximation coefficients describe the approximate temperature change trend of the feature sequence at a scale of 10 time units, while the detail coefficients show the small fluctuations in temperature at this time scale. As the time window length increases, higher-level decompositions are performed sequentially to obtain approximation coefficients and detail coefficients at different scales, thereby comprehensively analyzing the multi-scale characteristics of the feature sequence.

[0025] Step S213: Reconstruct the approximate coefficients under the same decomposition level to generate trend components at the corresponding time scale. The trend components reflect the overall change trend of the feature sequence at that time scale.

[0026] Reconstruction processing involves combining and processing the approximate coefficients obtained from decomposition to recover the trend information of the feature sequence at that time scale. The trend component, obtained through reconstruction, clearly demonstrates the overall trend of the feature sequence at that time scale.

[0027] In practical implementation, the inverse wavelet transform algorithm can be used to reconstruct the approximate coefficients at the same decomposition level. For example, at the first decomposition level, the approximate coefficients obtained through wavelet transform are input into the inverse wavelet transform algorithm, and the trend component at the corresponding time scale is calculated. This trend component can intuitively reflect whether a certain operating parameter of the equipment is rising, falling, or remaining stable at that time scale. By reconstructing the approximate coefficients at different decomposition levels, trend components at different time scales can be obtained, thereby gaining a deeper understanding of the overall changes in the feature sequence at different time scales.

[0028] Step S214: Reconstruct the detail coefficients under the same decomposition level to generate the fluctuation component at the corresponding time scale. The fluctuation component reflects the local fluctuation characteristics of the feature sequence at that time scale.

[0029] Similar to step S213, this step involves reconstructing the detail coefficients. The detail coefficients contain local fluctuation information of the feature sequence at that time scale. This information is integrated through reconstruction to generate fluctuation components. These fluctuation components reflect the small-amplitude fluctuations of the feature sequence at that time scale.

[0030] Similarly, the inverse wavelet transform algorithm is used to reconstruct the detail coefficients at the same decomposition level. For example, at the second decomposition level, the obtained detail coefficients are input into the inverse wavelet transform algorithm, and the fluctuation component at the corresponding time scale is generated after calculation. This fluctuation component can reflect the local fluctuation characteristics of the equipment operating parameters at that time scale, such as whether there are periodic small fluctuations. By reconstructing the detail coefficients at different decomposition levels, the local fluctuation of the feature sequence at different time scales can be fully understood.

[0031] Step S215: Summarize the trend components and fluctuation components of all feature sequences at each time scale to obtain a multi-scale feature sequence set containing feature expressions at different time scales. Each element in the multi-scale feature sequence set contains a feature identifier, time scale parameters, trend components, and fluctuation components.

[0032] The aggregation process involves collecting and integrating the trend and fluctuation components of all feature sequences across various time scales. Since the trend and fluctuation components of different feature sequences may have different scales and ranges, they need to be standardized first to eliminate scale differences for easier subsequent analysis and processing. Standardization can be achieved using common methods, such as normalizing the data to ensure its values ​​fall within the range of [0, 1].

[0033] After standardization and aggregation, a multi-scale feature sequence set is obtained. Each element in this set contains a feature identifier to uniquely identify the feature sequence; a time scale parameter indicating the time scale corresponding to the trend and fluctuation components; and the trend and fluctuation components, reflecting the overall trend and local fluctuation of the feature sequence at that time scale, respectively. For example, for the temperature feature sequence of equipment, at a time scale of 20 time units, the corresponding elements include a temperature feature identifier, a time scale parameter of 20 time units, and the temperature trend and fluctuation components at that time scale. This multi-scale feature sequence set provides a comprehensive and unified data foundation for feature correlation analysis.

[0034] Step S220: Calculate the mutual information value between each feature sequence in the multi-scale feature sequence set, and generate a feature mutual information correlation matrix. The element values ​​in the feature mutual information correlation matrix represent the degree of nonlinear correlation between the corresponding two feature sequences.

[0035] Mutual information is an index used to measure the correlation between two random variables, capturing the nonlinear relationship between them. In this embodiment, the mutual information values ​​between each feature sequence in a multi-scale feature sequence set are calculated to determine the degree of nonlinear correlation between them. The feature mutual information correlation matrix is ​​a matrix where each element represents the mutual information value between two corresponding feature sequences, i.e., the degree of nonlinear correlation.

[0036] Mutual information values ​​can be calculated using methods such as kernel density estimation. For each pair of feature sequences in a multi-scale feature sequence set, kernel density estimation is first performed on them to obtain their respective probability density functions. Then, according to the definition of mutual information, their mutual information value is calculated. The mutual information values ​​between all feature sequences are combined to form a feature mutual information correlation matrix. For example, for the temperature and pressure feature sequences of a device, their mutual information value is calculated and filled into the corresponding position in the feature mutual information correlation matrix. Such a feature mutual information correlation matrix can intuitively show the nonlinear correlation between the feature sequences, providing a basis for clustering processing.

[0037] Step S230: Based on the feature mutual information correlation matrix, cluster the feature sequences using a community detection algorithm to generate a set of feature communities with cohesion. Each feature community set contains multiple feature sequences with correlation.

[0038] Community detection algorithms are used to divide nodes in a network into different communities. In this embodiment, the feature mutual information correlation matrix is ​​converted into a feature correlation network, and then the community detection algorithm is used to cluster the feature sequences. The feature community set is the set obtained after clustering. Each feature community contains multiple feature sequences with strong correlations and exhibits cohesion, meaning that the correlation between feature sequences within a community is strong, while the correlation between feature sequences in other communities is weak.

[0039] As one implementation method, step S230 can be specifically implemented as the following steps S231~S235:

[0040] Step S231: Convert the feature mutual information correlation matrix into a feature correlation network, where network nodes represent feature sequences and the weight values ​​of network edges represent the degree of nonlinear correlation between corresponding two feature sequences.

[0041] A feature association network (FAN) is a graph structure used to represent the relationships between feature sequences. The process of converting a feature mutual information matrix into a FAN involves mapping the element information from the matrix onto the graph structure. Network nodes correspond to feature sequences, with each node representing a single feature sequence. Network edges connect different nodes, and the weights of the edges represent the degree of non-linear correlation between the corresponding two feature sequences, i.e., the mutual information value.

[0042] For example, the element (i, j) in the feature mutual information association matrix represents the mutual information value between feature sequence i and feature sequence j. In the feature association network, node i and node j are connected by an edge, and the weight of the edge is set to this mutual information value. Through this transformation, the association relationship between feature sequences can be presented in an intuitive graph structure, which is convenient for subsequent community detection algorithms to process.

[0043] Step S232: Initialize the community allocation vector, randomly assign each feature sequence to a different initial community, and the number of initial communities is equal to the number of equal divisions of the total number of feature sequences.

[0044] The community assignment vector is a vector used to record the community to which each feature sequence belongs. During initialization, each feature sequence is randomly assigned to a different initial community. The number of initial communities is determined by equally dividing the total number of feature sequences. For example, if there are 100 feature sequences, and they are equally divided into 10 initial communities, then each initial community contains 10 feature sequences.

[0045] By randomly assigning features, an initial state is provided for community adjustment. This initial setting ensures that each feature sequence has an initial community affiliation. Subsequently, the communities will be adjusted based on the correlation between feature sequences to form a cohesive set of feature communities.

[0046] Step S233: Calculate the modularity contribution value of each feature sequence. The modularity contribution value is used to measure the change in the overall modularity of the network when the feature sequence is moved from the current community to other communities. Adjust the community affiliation of the feature sequences in descending order of modularity contribution value until the overall modularity of the network no longer increases or reaches the preset iteration limit.

[0047] Modularity is a metric used to measure the quality of network community partitioning, reflecting the difference between the strength of intra-community connections and the strength of random connections within a network. The modularity contribution value refers to the change in the overall modularity of the network when a feature sequence is moved from its current community to another community. By calculating the modularity contribution value of each feature sequence, it can be determined whether moving it to another community can improve the overall modularity of the network.

[0048] The modularity contribution value can be calculated using relevant formulas from modularity optimization algorithms. For each feature sequence, it is moved to other communities sequentially, and the difference in the overall modularity of the network before and after the move is calculated to obtain the modularity contribution value of that feature sequence. Then, the feature sequences are sorted in descending order of modularity contribution value, and the community affiliation of the feature sequences is adjusted accordingly. After each adjustment, the overall modularity of the network is recalculated. If the overall modularity of the network no longer increases, or if a preset iteration limit is reached, the adjustment is stopped. For example, if the preset iteration limit is 100 times, the adjustment is stopped even if the overall modularity of the network could still increase after the 100th iteration. Through this adjustment process, the community division of the feature sequences can be made more reasonable, improving the cohesion of the feature community set.

[0049] Step S234: Merge communities in the adjusted community structure, merging small communities with modularity contribution values ​​less than a preset threshold into adjacent communities to generate the final feature community set.

[0050] After the adjustments in step S233, some smaller communities may remain with relatively low modularity contributions. To further optimize the community structure, smaller communities with modularity contributions less than a preset threshold are merged into neighboring communities. Neighboring communities refer to communities that are closely connected to the smaller community in the feature association network.

[0051] The preset threshold is determined based on the actual situation and analysis requirements. For example, setting the preset threshold to 0.1 allows small communities with a modularity contribution value less than 0.1 to be merged into neighboring communities. During the merging process, the community allocation vector and relevant information of the feature association network need to be updated. Through this community merging operation, the number of communities can be reduced, the size and cohesion of each community can be increased, and a final set of feature communities can be generated.

[0052] Step S235: Calculate the internal average mutual information value and the inter-community average mutual information value of each feature community set to verify the aggregation characteristics of the feature community set, so that the internal average mutual information value and the inter-community average mutual information value exhibit a discriminative distribution.

[0053] The internal average mutual information value refers to the average mutual information value among the feature sequences within each community in a feature community set, reflecting the strength of the correlation between feature sequences within a community. The inter-community average mutual information value refers to the average mutual information value among feature sequences in different feature communities, reflecting the strength of the correlation between communities. By calculating these two values ​​and verifying whether they exhibit a discriminative distribution—that is, if the internal average mutual information value is significantly greater than the inter-community average mutual information value—we can determine whether the aggregation characteristics of the feature community set are good.

[0054] When calculating the internal average mutual information value, for each feature community, the mutual information value between all feature sequences within it is calculated and averaged. When calculating the inter-community average mutual information value, for every two distinct feature communities, the mutual information value between all feature sequences of them is calculated and averaged. Then, the internal average mutual information value and the inter-community average mutual information value are compared. If the internal average mutual information value is much greater than the inter-community average mutual information value, it indicates that the feature community set has good aggregation properties, that is, the feature sequences within a community are strongly correlated, while the correlation between communities is weak. For example, the calculated internal average mutual information value of the feature community set is 0.8, and the inter-community average mutual information value is 0.2, clearly showing a discriminative distribution, indicating that the feature community set has good aggregation properties.

[0055] Step S240: Perform principal component analysis on the feature sequences in each feature community set, and extract the principal component feature vectors that represent the overall trend of change of the feature sequences within the community. The dimension of the principal component feature vectors is the same as the number of feature community sets.

[0056] Principal component analysis (PCA) can transform multiple correlated variables into a few uncorrelated principal components. In this embodiment, PCA is performed on the feature sequences in each feature community set to extract principal component feature vectors that characterize the overall trend of change in the feature sequences within the community. The dimension of the principal component feature vectors is the same as the number of feature community sets, thus allowing for a smaller dimensionality to summarize the overall changes in each feature community.

[0057] Principal component analysis (PCA) first standardizes the feature sequences within each feature community set to ensure they have the same scale and range. Then, it calculates the covariance matrix of the feature sequences and solves for the eigenvalues ​​and eigenvectors. The eigenvectors with the same number of feature community sets are selected as principal components, and the feature sequences are projected onto these principal components to obtain the principal component eigenvectors. For example, with five feature community sets, the first five eigenvectors are selected as principal components, and the feature sequences from each feature community set are projected onto these five principal components to obtain the principal component eigenvectors for each feature community. These principal component eigenvectors effectively summarize the overall trend of feature sequences within each feature community, providing a concise and effective feature representation for feature fusion.

[0058] Step S250: Standardize the principal component feature vectors corresponding to all feature community sets to eliminate dimensional differences and generate a standardized principal component feature vector set; concatenate the standardized principal component feature vector set to generate a trend evolution feature containing the dynamic evolution law of each feature community, and calculate the cross-influence coefficient matrix between different feature community sets as the cross-influence feature.

[0059] Standardization aims to eliminate dimensional differences between principal component eigenvectors corresponding to different feature community sets, giving them a uniform scale and range. Common standardization methods can be used, such as normalizing the principal component eigenvectors to ensure their values ​​are within the range of [0, 1]. The standardized principal component eigenvector sets are then concatenated, that is, they are combined sequentially to form a new vector. This vector represents the trend evolution feature containing the dynamic evolutionary patterns of each feature community.

[0060] The cross-influence coefficient matrix measures the degree of mutual influence between different feature community sets. It can be calculated using correlation analysis. For every two distinct feature community sets, the correlation coefficient between them is calculated, and these coefficients are combined to form the cross-influence coefficient matrix. For example, for feature community set A and feature community set B, the correlation coefficient between them is calculated, and this coefficient is filled into the corresponding position in the cross-influence coefficient matrix. The resulting cross-influence coefficient matrix is ​​the cross-influence feature, which, together with the trend evolution feature, provides important information for the subsequent generation of equipment state potential vectors.

[0061] Step S260: The trend evolution features and cross-influence features are fused to generate a device state potential vector that represents the dynamic coupling relationship between parameters. The dimension of the device state potential vector is equal to the sum of the dimension of the trend evolution features and the dimension of the cross-influence features.

[0062] Before fusing trend evolution features and cross-influence features, they need to be standardized to eliminate dimensional differences. Standardization can be performed using the same method as in step S250, such as normalization. After standardization, the trend evolution features and cross-influence features are then fused in a specific way, for example, by connecting them sequentially.

[0063] The equipment state potential vector is the fused result, with a dimension equal to the sum of the dimensions of the trend evolution features and the cross-influence features. This vector comprehensively characterizes the dynamic coupling relationships between various equipment parameters, providing a comprehensive representation of equipment state for fault early warning analysis. For example, if the dimension of the trend evolution features is 10 and the dimension of the cross-influence features is 5, then the dimension of the equipment state potential vector is 15. Through this fusion process, the trend evolution features and cross-influence features are integrated into a unified vector, more effectively reflecting the complex relationships between equipment parameters.

[0064] Step S300: Call the pre-trained fuzzy logic reasoning model to perform fuzzy rule matching on the device state potential vector to generate a fuzzy state set containing multi-dimensional fuzzy subsets.

[0065] The pre-trained fuzzy logic reasoning model is a model that has been trained beforehand. It includes a fuzzification layer, a rule base module, and a defuzzification layer, and can perform fuzzy rule matching and reasoning based on the input device state potential vector. Fuzzy rule matching refers to comparing the device state potential vector with the preset fuzzy rules in the model to find the matching rule. The fuzzy state set is a set generated after reasoning, which contains multi-dimensional fuzzy subsets, each corresponding to a fuzzy description of a device operating state category.

[0066] As one implementation method, step S300 can be specifically implemented as the following steps S310~S360:

[0067] Step S310: Input the device state potential vector into the fuzzification layer of the fuzzy logic reasoning model, and perform fuzzification processing on each dimension feature of the device state potential vector through a preset set of membership functions to generate the corresponding fuzzy membership vector.

[0068] The fuzzification layer is a component of the fuzzy logic inference model, its function being to convert precise input data into fuzzy data. The preset membership function set is a group of functions used to describe the degree to which an element belongs to different fuzzy sets; possible membership function types include triangular membership functions, trapezoidal membership functions, and Gaussian membership functions, etc. Fuzzification processing refers to calculating the membership degree value of each dimension of the equipment state potential vector to each preset fuzzy set based on the preset membership function set. The fuzzy membership vector is the vector obtained after fuzzification processing, where each element represents the membership degree value of one dimension of the equipment state potential vector to each preset fuzzy set.

[0069] As one implementation method, step S310 can be specifically implemented as the following steps S311~S315:

[0070] Step S311: Analyze the fuzzification layer configuration parameters of the fuzzy logic reasoning model, and extract the membership function types and parameter sets corresponding to the features of each dimension of the device state potential vector. The membership function types include triangular membership function, trapezoidal membership function and Gaussian membership function.

[0071] The fuzzification layer configuration parameters are the settings for the fuzzification layer in the fuzzy logic inference model. These parameters include the membership function types and parameter sets corresponding to each dimension of the device state potential vector. Parsing the fuzzification layer configuration parameters involves reading this information from the model. For example, for the first dimension of the device state potential vector, the parsed parameter set is a triangular membership function with three key points (a, b, c). This parsing operation provides the specific membership functions and parameter basis for fuzzification processing.

[0072] Step S312: For each dimension feature of the device state potential vector, calculate the membership value of the dimension feature to each preset fuzzy set according to the corresponding membership function type and parameter set, and generate a dimension membership vector. The length of the dimension membership vector is equal to the number of preset fuzzy sets corresponding to the dimension feature.

[0073] For each dimension of the device state potential vector, based on the membership function type and parameter set obtained in step S311, the membership value of that dimension belonging to each preset fuzzy set is calculated using the corresponding membership function calculation formula. For example, if the membership function type is a triangular membership function and the parameter set is (a, b, c), for a specific dimension feature value x, its membership value belonging to each preset fuzzy set is calculated according to the formula of the triangular membership function.

[0074] The membership values ​​of each dimensional feature belonging to each of the preset fuzzy sets are combined to form a dimensional membership vector. The length of the dimensional membership vector is equal to the number of preset fuzzy sets corresponding to that dimensional feature. For example, if a certain dimensional feature corresponds to 3 preset fuzzy sets, the generated dimensional membership vector will be 3 in length, and the elements in the vector will represent the membership values ​​of that dimensional feature belonging to these 3 preset fuzzy sets.

[0075] Step S313: Normalize the dimension membership vectors so that the sum of all element values ​​in each dimension membership vector is 1, thus generating a standardized dimension membership vector.

[0076] Normalization is performed to ensure the comparability and reasonableness of elements in the dimension membership vector, guaranteeing that the sum of the membership values ​​of each dimension feature belonging to each preset fuzzy set is 1. Common normalization methods can be used, such as dividing each element in the dimension membership vector by the sum of all elements in the vector.

[0077] For example, for a membership vector of dimension (0.2, 0.3, 0.5), the sum of all its elements is 1, so no normalization is needed. However, if the membership vector of dimension is (0.1, 0.2, 0.3), the sum of all its elements is 0.6. Dividing each element by 0.6 yields a normalized membership vector of dimension (0.1 / 0.6, 0.2 / 0.6, 0.3 / 0.6). This normalization process generates a normalized membership vector that better meets the requirements of fuzzy logic, providing accurate membership information for processing.

[0078] Step S314: Concatenate the standardized dimensional membership vectors corresponding to all dimensional features according to the dimensional order of the device state potential vector to generate a fuzzy membership vector containing fuzzy information of all dimensional features.

[0079] The standardized membership vectors corresponding to each dimensional feature are concatenated sequentially according to the dimensional order of the device state potential vector to form a new vector. This vector is the fuzzy membership vector containing fuzzy information from all dimensional features. For example, the device state potential vector has three dimensional features, with corresponding standardized membership vectors of (0.2, 0.3, 0.5), (0.1, 0.4, 0.5), and (0.3, 0.3, 0.4). Concatenating these in dimensional order yields the fuzzy membership vector (0.2, 0.3, 0.5, 0.1, 0.4, 0.5, 0.3, 0.3, 0.4). This fuzzy membership vector comprehensively contains fuzzy information from all dimensional features of the device state potential vector, providing input data for rule matching.

[0080] Step S315: Calculate the fuzzy entropy value of the fuzzy membership vector, evaluate the effectiveness of the current fuzzification process, and when the fuzzy entropy value is lower than the preset threshold, adjust the membership function parameter set and perform fuzzification process again until the fuzzy entropy value is greater than the preset threshold.

[0081] Fuzzy entropy is an indicator used to measure the uncertainty of fuzzy sets. In this embodiment, calculating the fuzzy entropy value of the fuzzy membership vector can evaluate the effectiveness of the current fuzzification process. If the fuzzy entropy value is low, it indicates that the result of the fuzzification process is not fuzzy enough, and there may be information loss or inaccuracies.

[0082] The fuzzy entropy value can be calculated using the fuzzy entropy calculation formula. For each element in the fuzzy membership vector, the corresponding entropy value is calculated based on its membership value. Then, the entropy values ​​of all elements are added together to obtain the fuzzy entropy value of the fuzzy membership vector. The preset threshold is determined based on the actual situation and analysis requirements. When the fuzzy entropy value is lower than the preset threshold, the membership function parameter set needs to be adjusted, for example, modifying the parameters (a, b, c) of the triangular membership function. Then, the fuzzification process in steps S311-S314 is repeated until the fuzzy entropy value is greater than the preset threshold. Through this evaluation and adjustment process, the effectiveness of the fuzzification process can be ensured, and the accuracy of fuzzy logic reasoning can be improved.

[0083] Step S320: Extract the linguistic variable identifiers corresponding to each dimension of the fuzzy membership vector, and construct a set of fuzzy feature descriptors containing linguistic variable names and membership values.

[0084] Linguistic variable identifiers are linguistic terms used to describe the operating status of equipment, such as "high," "medium," and "low." In a fuzzy membership vector, each dimension of a feature corresponds to a set of membership values, and also to a corresponding linguistic variable identifier. These linguistic variable identifiers are extracted and combined with their corresponding membership values ​​to form a set of fuzzy feature descriptors.

[0085] For example, for a single dimension feature in a fuzzy membership vector, with corresponding linguistic variable labels "high," "medium," and "low," and membership values ​​of 0.2, 0.3, and 0.5 respectively, the constructed fuzzy feature descriptors would be ("high," 0.2), ("medium," 0.3), and ("low," 0.5). Combining the fuzzy feature descriptors corresponding to all dimensions forms a fuzzy feature descriptor set. This set of fuzzy feature descriptors can intuitively display the fuzzy descriptive information of each dimension feature of the device state potential vector, providing clear input for rule matching.

[0086] Step S330: Call the rule base module of the fuzzy logic reasoning model, perform rule matching operation based on the fuzzy feature descriptor set, and filter out the preset fuzzy rule set with the highest matching degree with the current fuzzy feature descriptor set.

[0087] The rule base module stores preset fuzzy rules in the fuzzy logic reasoning model. The rule matching operation compares the set of fuzzy feature descriptors with the preset fuzzy rules in the rule base module to find the rule with the highest matching degree. The preset fuzzy rule set is a set of rules obtained after filtering, containing rules with the highest matching degree to the current set of fuzzy feature descriptors.

[0088] As one implementation method, step S330 can be specifically implemented as the following steps S331~S336:

[0089] Step S331: Parse the fuzzy feature descriptor set and extract a feature-membership pair list containing linguistic variable names and membership values. Each element in the feature-membership pair list contains a linguistic variable name and its corresponding membership value.

[0090] Parsing a fuzzy feature descriptor set involves breaking it down into feature-membership pairs. For example, for the fuzzy feature descriptor set (("high", 0.2), ("medium", 0.3), ("low", 0.5)), the parsed list of feature-membership pairs is [("high", 0.2), ("medium", 0.3), ("low", 0.5)]. This list of feature-membership pairs provides specific matching elements for rule matching.

[0091] Step S332: Traverse all preset fuzzy rules in the rule base module, and for each preset fuzzy rule, extract the list of rule-feature pairs contained in its preconditions.

[0092] The rule base module stores a large number of preset fuzzy rules, each consisting of a premise and a conclusion. It iterates through all rules in the rule base module, and for each rule, extracts the rule-feature pair list contained in its premise section. Each element in the rule-feature pair list contains a linguistic variable name and its corresponding rule condition.

[0093] For example, a pre-defined fuzzy rule is "If the temperature is high and the pressure is medium, then the equipment is in a dangerous state." Its prerequisite part contains a rule-feature pair list of [("temperature", "high"), ("pressure", "medium")]. Through such traversal and extraction operations, the prerequisite information for the rule is provided for the matching degree calculation.

[0094] Step S333: Calculate the matching degree value between the feature-membership pair list and the rule-feature pair list of each preset fuzzy rule. The matching degree value is obtained by calculating the sum of the products of the membership values ​​corresponding to the same linguistic variable names in the two lists.

[0095] For the feature-membership pair list and the rule-feature pair list for each preset fuzzy rule, find the elements with the same linguistic variable name, multiply their corresponding membership values, and then add all the products to get the matching degree value.

[0096] Step S334: Sort all preset fuzzy rules in descending order according to the matching degree value to generate a rule matching degree sorting table.

[0097] All preset fuzzy rules are sorted in descending order of their matching scores to form a rule matching score ranking table. For example, if there are three preset fuzzy rules with matching scores of 0.37, 0.25, and 0.45, the sorted rule matching score ranking table would be [(Rule 3, 0.45), (Rule 1, 0.37), (Rule 2, 0.25)]. This rule matching score ranking table visually displays the degree of matching between each preset fuzzy rule and the feature-membership pair list, facilitating filtering operations.

[0098] Step S335: Select preset fuzzy rules with matching degree values ​​greater than the preset matching threshold from the rule matching degree sorting table to generate a preliminary matching rule set.

[0099] The preset matching threshold is determined based on the actual situation and analysis needs. From the rule matching degree ranking table, preset fuzzy rules with matching degree values ​​greater than the preset matching threshold are selected and combined to form a preliminary matching rule set. For example, if the preset matching threshold is 0.3, rules 3 and 1 with matching degree values ​​greater than 0.3 are selected from the rule matching degree ranking table [(Rule 3, 0.45), (Rule 1, 0.37), (Rule 2, 0.25)] to generate the preliminary matching rule set {Rule 3, Rule 1}. This preliminary matching rule set contains rules with high matching degrees to the feature-membership pair list.

[0100] Step S336: Perform rule conflict identification on the preliminary matching rule set. When there are preset fuzzy rules with contradictory rule conclusions, retain the preset fuzzy rule with the higher matching degree value and generate the final preset fuzzy rule set.

[0101] Rule conflict identification refers to checking whether there are contradictory conclusions among the initial matching rule set. For example, if rule A concludes "the equipment status is normal" and rule B concludes "the equipment status is dangerous", then these two rules conflict.

[0102] When a rule conflict is detected, the matching degree values ​​of the conflicting rules are compared, and the rule with the higher matching degree value is retained. For example, if rule A has a matching degree value of 0.37 and rule B has a matching degree value of 0.45, and their conclusions contradict each other, then rule B is retained. The rules that have undergone conflict identification and filtering are combined to generate the final preset fuzzy rule set. This rule conflict identification and filtering process ensures the consistency and accuracy of the final preset fuzzy rule set.

[0103] Step S340: Calculate the trigger intensity of each fuzzy rule in the preset fuzzy rule set to generate a rule trigger intensity vector. The dimension of the rule trigger intensity vector is the same as the number of rules in the preset fuzzy rule set.

[0104] Trigger strength refers to the degree to which each fuzzy rule is triggered under the current input. For each fuzzy rule in the preset fuzzy rule set, the trigger strength is calculated based on its preconditions and corresponding membership values. The rule trigger strength vector is a vector formed by combining the trigger strengths of all rules, and its dimension is the same as the number of rules in the preset fuzzy rule set.

[0105] There are several methods for calculating trigger strength, such as taking the minimum membership value of all linguistic variables in the rule's preconditions (i.e., taking the minimum value). For example, if a predefined fuzzy rule's preconditions are "high temperature and medium pressure," with "temperature" having a membership value of 0.2 and "pressure" having a membership value of 0.3, then the trigger strength of this rule is 0.2. The trigger strengths of all rules in the predefined fuzzy rule set are combined to form a rule trigger strength vector. For example, if the predefined fuzzy rule set contains three rules with trigger strengths of 0.2, 0.3, and 0.5 respectively, then the rule trigger strength vector is (0.2, 0.3, 0.5).

[0106] Step S350: Perform weighted aggregation processing on the preset fuzzy rule set according to the value of each element in the rule trigger intensity vector to generate an intermediate fuzzy decision vector containing multiple fuzzy decision variables.

[0107] Weighted aggregation processing involves weighting the rule conclusions from a pre-defined set of fuzzy rules according to the element values ​​in the rule trigger strength vector. For each fuzzy decision variable, the conclusions related to that variable from all rules are summed in weight according to their trigger strength to obtain the weighted value of that variable.

[0108] For example, a preset fuzzy rule set contains three rules with a rule trigger strength vector of (0.2, 0.3, 0.5) and rule conclusions of "Equipment status is normal (0.8)", "Equipment status is dangerous (0.6)", and "Equipment status is normal (0.9)" (the values ​​in parentheses are membership degrees). For the fuzzy decision variable "Equipment status", its weighted value is 0.2 × 0.8 + 0.3 × 0.6 + 0.5 × 0.9 = 0.79. Combining the weighted values ​​of all fuzzy decision variables forms an intermediate fuzzy decision vector. This intermediate fuzzy decision vector comprehensively considers the trigger strength and conclusion of each rule in the preset fuzzy rule set, providing an intermediate result for defuzzification.

[0109] Step S360: Input the intermediate fuzzy decision vector into the defuzzification layer of the fuzzy logic reasoning model, and perform defuzzification processing on the intermediate fuzzy decision vector using the centroid method to generate a fuzzy state set containing multi-dimensional fuzzy subsets. Each fuzzy subset corresponds to a fuzzy description of a device operating state category.

[0110] The defuzzification layer is the final step in the fuzzy logic reasoning model. Its function is to convert the intermediate fuzzy decision vectors into specific descriptions of equipment operating states. The final output value is determined by calculating the centroid position of the intermediate fuzzy decision vectors using the centroid method.

[0111] The intermediate fuzzy decision vector is input into the defuzzification layer. Using the centroid method, the centroid position of each fuzzy decision variable is calculated, yielding a specific numerical value. These values ​​are combined to form a fuzzy state set containing multi-dimensional fuzzy subsets. Each fuzzy subset corresponds to a fuzzy description of a device's operating state category, such as "device status is normal" or "device status is dangerous." Through this defuzzification process, the fuzzy intermediate decision results are transformed into specific descriptions of device operating states, providing a clear input for spatiotemporal correlation processing.

[0112] Step S400: Combine the preset domain fault knowledge base to perform spatiotemporal correlation processing on the fuzzy state set to generate a multidimensional decision cloud map with spatiotemporal correlation characteristics. Perform spatial reconstruction operation on the multidimensional decision cloud map through a quantum weight allocation mechanism to obtain an optimized decision matrix.

[0113] The pre-defined domain fault knowledge base is a database storing knowledge related to equipment faults, including fault types, fault characteristics, and fault evolution paths. Spatiotemporal correlation processing combines the fuzzy state set with information from the domain fault knowledge base, considering both temporal and spatial factors to identify the relationships between them. The multidimensional decision cloud map, generated after spatiotemporal correlation processing, possesses spatiotemporal correlation characteristics and can intuitively display the temporal and spatial distribution of equipment fault states. The quantized weight allocation mechanism is a method for allocating weights; by using this mechanism to spatially reconstruct the multidimensional decision cloud map, the decision matrix can be optimized to more accurately reflect the equipment fault situation.

[0114] As one implementation method, step S400 can be specifically implemented as the following steps S410~S460:

[0115] Step S410: Parse the preset domain fault knowledge base, extract a set of fault cases containing fault type identifiers, fault feature vectors and fault evolution paths, and construct a fault case feature index table.

[0116] Parsing the pre-defined domain fault knowledge base involves reading relevant information from the knowledge base and extracting a set of fault cases containing fault type identifiers, fault feature vectors, and fault evolution paths. Fault type identifiers are used to uniquely identify different fault types, such as "short circuit fault" and "overload fault." Fault feature vectors are vectors describing the characteristics of the fault, containing feature information about the fault in various aspects. The fault evolution path refers to the description of the process from fault occurrence to development, including the fault's time sequence and feature changes.

[0117] Constructing a fault case feature index table involves organizing and indexing information from a set of fault cases to facilitate rapid retrieval and querying. The fault case feature index table can employ a hash table or other index structures, using the fault type identifier as the key and the corresponding fault feature vector and fault evolution path as values. This parsing and indexing operation provides the informational foundation of fault cases for similarity calculation and spatiotemporal alignment.

[0118] In one implementation, step S410 can be specifically implemented as the following steps S411~S416:

[0119] Step S411: Connect to the preset domain fault knowledge base database, perform a structured query operation, and extract the basic information of fault cases stored in the relational data table. The basic information of fault cases includes the unique identifier of the fault case, the fault type identifier, the fault occurrence timestamp, and the fault handling result.

[0120] Connecting to the pre-defined domain fault knowledge base database is achieved through a database connection interface. Structured query operations use SQL statements to query the database and extract basic fault case information from relational data tables. A unique fault case identifier is used to uniquely identify each fault case, a fault type identifier is used to distinguish different fault types, a fault occurrence timestamp records the specific time the fault occurred, and the fault handling result records the handling status of the fault.

[0121] Step S412: Parse the unstructured data associated with the basic information of the fault case, extract the text information containing the description of the fault development process and the time series data of the fault characteristic change curve, and generate the fault evolution path descriptor.

[0122] Unstructured data associated with basic information about fault cases may include documents, logs, charts, etc., containing detailed descriptions of the fault development process and changes in fault characteristics. This unstructured data is parsed using text mining and data processing techniques to extract textual information describing the fault development process and time-series data of fault characteristic change curves. For example, for a fault report document, natural language processing techniques are used to extract descriptions of the fault development process, such as "the fault initially manifested as a temperature increase, then a pressure decrease, and finally equipment shutdown." For charts of fault characteristic change curves, image processing and data extraction techniques are used to convert the curves into time-series data. The extracted textual information and time-series data are combined to generate a fault evolution path descriptor. Such a fault evolution path descriptor can comprehensively describe the fault development process, providing detailed fault information for spatiotemporal alignment.

[0123] Step S413: Perform feature extraction processing on the time series data in the fault evolution path descriptor to generate a fault feature vector that represents the entire process of fault occurrence and development. The dimension of the fault feature vector is the same as the dimension of the equipment state potential vector.

[0124] Feature extraction involves extracting key information that characterizes the fault from the time-series data in the fault evolution path descriptor. For ease of subsequent comparison and analysis, the generated fault feature vector has the same dimension as the equipment state potential vector.

[0125] Common feature extraction methods, such as principal component analysis (PCA) and wavelet transform, can be employed. For temperature time-series data in the fault evolution path descriptor, PCA is used to extract its main features, which are then combined to form part of the fault feature vector. For other types of time-series data, feature extraction is performed similarly, ultimately generating a fault feature vector representing the entire process from fault occurrence to development. This feature extraction process transforms the fault's time-series data into a representative fault feature vector, providing an effective feature representation for similarity calculation.

[0126] Step S414: Associate and store the unique identifier of the fault case, the fault type identifier, the fault feature vector, and the fault evolution path descriptor to generate a fault case set containing complete fault case information.

[0127] The unique fault case identifier and fault type identifier extracted in step S411, the fault feature vector generated in step S413, and the fault evolution path descriptor generated in step S412 are associated and stored in a data structure to form a fault case set containing complete fault case information. For example, a dictionary data structure can be used, with the unique fault case identifier as the key and the fault type identifier, fault feature vector, and fault evolution path descriptor as values ​​stored in the dictionary. Such a fault case set can comprehensively record the detailed information of each fault case, providing rich data resources for fault analysis and early warning.

[0128] Step S415: Perform principal component analysis on all fault feature vectors in the fault case set, extract principal component features that represent the main changes in fault feature vectors, and construct the fault case feature space.

[0129] Principal component analysis (PCA) is a method for data dimensionality reduction and feature extraction, capable of transforming multiple correlated variables into a few uncorrelated principal components. PCA is performed on all fault feature vectors in a fault case set to calculate their covariance matrix, and then the eigenvalues ​​and eigenvectors of the covariance matrix are determined. The top few eigenvectors with larger eigenvalues ​​are selected as principal components, and the fault feature vectors are projected onto these principal components to obtain the principal component features.

[0130] The principal component features of all failure cases are combined to construct a failure case feature space. This feature space is a low-dimensional space where each point represents a principal component feature of a failure case. Through principal component analysis and feature space construction, the dimensionality of failure feature vectors can be reduced, improving the efficiency of failure case retrieval and providing an effective feature space for failure case matching.

[0131] Step S416: Map each fault case in the fault case set to the fault case feature space to generate a fault case feature index table containing the fault case coordinates and fault type identifiers, which is used to quickly retrieve similar fault cases.

[0132] The fault feature vector of each fault case in the fault case set is projected into the fault case feature space to obtain its coordinate position in the feature space. The coordinate position of the fault case and the fault type identifier are combined to generate a fault case feature index table.

[0133] For example, for a fault case, if its fault feature vector is projected onto the fault case feature space at coordinates (x, y, z), and the fault type is identified as "short circuit fault," then it will be recorded in the fault case feature index table as ((x, y, z), "short circuit fault"). Through such mapping and indexing operations, fault cases can be located in the feature space, facilitating the rapid retrieval of similar fault cases. When it is necessary to find fault cases similar to the current fault state, it is only necessary to search for cases with close coordinate positions in the fault case feature index table.

[0134] Step S420: Calculate the commonality measure between the fuzzy state set and the fault feature vectors of each fault case in the fault case feature index table, and generate a fault similarity vector. The dimension of the fault similarity vector is the same as the number of fault cases in the set.

[0135] Commonality metrics are indicators used to measure the similarity between a fuzzy state set and the fault feature vectors of fault cases. Common similarity calculation methods, such as Euclidean distance and cosine similarity, can be used to calculate commonality metrics.

[0136] For each fault feature vector in the fuzzy state set and the fault case feature index table, calculate the commonality metric between them, for example, using Euclidean distance to calculate similarity. Combine the commonality metrics of all fault cases to form a fault similarity vector. The dimension of the fault similarity vector is the same as the number of fault cases in the set, and each element in the vector represents the similarity between the fuzzy state set and the corresponding fault case. Through this calculation, the degree of similarity between the fuzzy state set and each fault case can be quantified, providing a basis for selecting similar fault cases.

[0137] Step S430: Based on the fault similarity vector, select the preset number of fault cases with the highest similarity value to generate a set of similar fault cases. The set of similar fault cases includes fault type identifiers, historical fault feature evolution sequences, and fault handling solutions.

[0138] Based on the similarity values ​​in the fault similarity vector, fault cases are sorted, and a preset number of fault cases with the highest similarity values ​​are selected. The preset number is determined based on the actual situation and analysis needs; for example, selecting the 5 fault cases with the highest similarity values.

[0139] By combining the fault type identifiers, historical fault feature evolution sequences, and fault handling solutions of the selected fault cases, a set of similar fault cases is generated. For example, a set of similar fault cases might include ("short circuit fault", [temperature rise sequence, pressure drop sequence], "fuse replacement"), ("overload fault", [current increase sequence, equipment heating sequence], "load reduction")). Such a set of similar fault cases can provide fault case information similar to the current fuzzy state set, providing a reference for spatiotemporal alignment and fault handling.

[0140] Step S440: Spatiotemporally align the historical fault feature evolution sequence in the set of similar fault cases with the current fuzzy state set to generate a spatiotemporal correlation matrix containing time dimension and feature dimension. The rows of the spatiotemporal correlation matrix represent the time step and the columns represent the feature dimension.

[0141] Spatiotemporal alignment involves matching and aligning the evolutionary sequences of historical fault features in a set of similar fault cases with the current fuzzy state set in time and space to identify their correspondences. The generated spatiotemporal correlation matrix can demonstrate the strength of the correlation between the historical fault feature evolutionary sequences and the current fuzzy state set in both time and feature dimensions.

[0142] In one implementation, step S440 can be specifically implemented as the following steps S441~S446:

[0143] Step S441: Extract the historical fault feature evolution sequence corresponding to each fault case in the similar fault case set, and determine the number of time steps and the number of feature dimensions of the historical fault feature evolution sequence.

[0144] The historical fault feature evolution sequence corresponding to each fault case is extracted from a set of similar fault cases. These sequences record the changes in various features during the fault's development. The number of time steps in the historical fault feature evolution sequence is determined, i.e., the number of time points in the sequence; the number of feature dimensions is also determined, i.e., the number of features contained in each time point in the sequence. For example, a historical fault feature evolution sequence might be [[Temperature 1, Pressure 1], [Temperature 2, Pressure 2], [Temperature 3, Pressure 3]], with 3 time steps and 2 feature dimensions. This extraction and determination process provides basic information about the historical fault feature evolution sequence for time alignment and feature difference calculation.

[0145] Step S442: Convert the current fuzzy state set into the current fault feature vector sequence, and determine the time step interval of the current fault feature vector sequence according to the equipment status monitoring frequency.

[0146] The current fuzzy state set is converted into a current fault feature vector sequence, that is, the fuzzy description corresponding to the fuzzy state set is converted into specific feature vectors. The time step interval of the current fault feature vector sequence is determined based on the equipment status monitoring frequency. The equipment status monitoring frequency refers to the frequency of equipment status data acquisition; for example, if data is acquired once per second, the time step interval is 1 second. For example, if the fuzzy description corresponding to the current fuzzy state set is "high temperature, medium pressure," it is converted into a current fault feature vector sequence [temperature value, pressure value]. Assuming the equipment status monitoring frequency is once per minute, the time step interval of the current fault feature vector sequence is 1 minute. Through this conversion and determination operation, the time information of the current fault feature vector sequence is provided for time alignment.

[0147] Step S443: Based on the dynamic time warping algorithm, perform time alignment processing on the historical fault feature evolution sequence and the current fault feature vector sequence to generate a time alignment mapping relationship. The time alignment mapping relationship is used to indicate the correspondence between time points in the two sequences.

[0148] Dynamic time warping is an algorithm used to solve time series alignment problems. It can find the optimal alignment between two series without requiring strict alignment of their time steps.

[0149] The historical fault feature evolution sequence and the current fault feature vector sequence are input into a dynamic time warping algorithm. By calculating the distance between elements in the sequences and performing path planning, the optimal alignment path between the two sequences is found. The generated time alignment mapping relationship records the correspondence between time points in the two sequences. For example, the time alignment mapping relationship might be ((historical sequence time point 1, current sequence time point 2), (historical sequence time point 2, current sequence time point 3)). Through this time alignment processing, the historical fault feature evolution sequence and the current fault feature vector sequence can be matched in time, providing an accurate time correspondence for feature difference calculation.

[0150] Step S444: Based on the time alignment mapping relationship, convert the historical fault feature evolution sequence and the current fault feature vector sequence into aligned sequence pairs with the same number of time steps.

[0151] Based on the time alignment mapping relationship generated in step S443, the historical fault feature evolution sequence and the current fault feature vector sequence are adjusted to have the same number of time steps. For time points in the historical fault feature evolution sequence and the current fault feature vector sequence that do not match, interpolation or deletion methods can be used for processing.

[0152] For example, based on the time alignment mapping relationship ((historical sequence time point 1, current sequence time point 2), (historical sequence time point 2, current sequence time point 3)), the historical fault feature evolution sequence and the current fault feature vector sequence are adjusted to have the same number of time steps. Through this transformation operation, aligned sequence pairs with the same number of time steps are obtained, providing a unified time basis for feature difference calculation.

[0153] Step S445: For each time step in the aligned sequence pair, calculate the feature difference between the historical fault feature vector and the current fault feature vector, and generate the feature difference matrix.

[0154] For each time step in the aligned sequence pair, calculate the degree of difference between the historical fault feature vector and the current fault feature vector. Common methods for calculating the degree of difference can be used, such as Euclidean distance, Manhattan distance, etc., without specific limitations. Combine the feature difference values ​​of each time step to form a feature difference matrix. The rows of the feature difference matrix represent the time step, and the columns represent the feature dimensions.

[0155] By calculating the degree of difference in different feature dimensions at each time step, we can clearly understand the deviation of historical fault feature evolution from current fault features in various aspects.

[0156] Step S446: The feature difference matrix is ​​fused with the historical fault feature evolution sequence and the current fault feature vector sequence in the alignment sequence pair to generate a spatiotemporal correlation matrix containing time dimension and feature dimension. Each element value of the spatiotemporal correlation matrix represents the fault feature correlation strength under the corresponding time step and feature dimension.

[0157] Before fusion, the feature dissimilarity matrix must first be normalized. Normalization aims to eliminate differences in the numerical ranges of different elements in the feature dissimilarity matrix, making each element comparable. A common normalization method is to subtract the minimum value from each element in the feature dissimilarity matrix and then divide by the difference between the maximum and minimum values. Simultaneously, the historical fault feature evolution sequences and the current fault feature vector sequences in the aligned sequence pairs are also standardized. Z-score standardization can be used for standardization. After standardization, the historical fault feature evolution sequences and the current fault feature vector sequences have the same scale and range. Then, the normalized dissimilarity matrix is ​​weighted and fused with the standardized historical fault feature evolution sequences and the current fault feature vector sequences.

[0158] Step S450: Convert the spatiotemporal correlation matrix into a multidimensional decision cloud map. Each cloud droplet in the multidimensional decision cloud map represents the fault characteristic state under the corresponding time step and feature dimension. The position coordinates of the cloud droplet are determined by the row index and column index of the spatiotemporal correlation matrix, and the size of the cloud droplet is determined by the size of the feature value at the corresponding position.

[0159] Multidimensional decision cloud maps are visualization tools that intuitively display spatiotemporal correlation information. When converting a spatiotemporal correlation matrix into a multidimensional decision cloud map, the location coordinates of cloud droplets are first determined based on the row and column indices of the spatiotemporal correlation matrix. The row index corresponds to the time step, and the column index corresponds to the feature dimension, thus allowing the location of each cloud droplet to be determined in multidimensional space.

[0160] For example, for element s in the spatiotemporal correlation matrix 32 If the row index is 3 and the column index is 2, then the corresponding cloud droplet's position coordinates in the multidimensional decision cloud map are (3, 2). The size of the cloud droplet is determined by the magnitude of the eigenvalue at the corresponding position, i.e., the value of that element in the spatiotemporal correlation matrix. If s 32 A larger value indicates a stronger correlation between the fault features at the third time step and the second feature dimension, resulting in larger cloud droplets; conversely, if s 32 A smaller value results in a smaller cloud droplet. Through this transformation, the multidimensional decision cloud map can intuitively display the fault characteristic state under different time steps and feature dimensions, providing a visual data foundation for spatial reconstruction operations.

[0161] Step S460: Calculate the spatial weights of the cloud droplets in the multidimensional decision cloud map using a quantum weight allocation mechanism to generate a cloud droplet weight matrix. Perform a spatial reconstruction operation on the multidimensional decision cloud map based on the cloud droplet weight matrix to obtain an optimized decision matrix. The row dimension of the optimized decision matrix is ​​equal to the number of fault types, and the column dimension is equal to the number of feature dimensions.

[0162] The quantized weight allocation mechanism is a weight allocation method based on quantum theory that can take into account the quantum properties of cloud droplets in multidimensional space and assign reasonable spatial weights to cloud droplets.

[0163] As one implementation method, step S460 can be specifically implemented as the following steps S461~S467:

[0164] Step S461: Initialize the set of quantized weight allocation parameters, which includes the qubit configuration, measurement basis vectors, and weight quantization interval.

[0165] Qubit configuration refers to the number and state settings of qubits used in a quantum system, determining the dimension and representation of the quantum state space. Measurement basis vectors are the basis vectors used to measure quantum states; different measurement basis vectors will affect the measurement results. Weight quantization interval refers to the size of the interval when quantizing the weights of a cloud droplet, determining the degree of weight discretization.

[0166] During initialization, these parameters need to be determined based on the characteristics of the multidimensional decision cloud and the analytical requirements. For example, for a multidimensional decision cloud with high dimensionality and complex structure, a larger number of qubits can be set, along with appropriate measurement basis vectors and weight quantization intervals. Assuming the number of qubits is set to 3, the dimension of the quantum state space is 2^32. 3 = 8. The measurement basis vectors can be selected as orthogonal basis vectors, and the weight quantization interval can be set according to the approximate range of cloud droplet weights, such as 0.1. This initialization operation provides a clear parameter basis for quantum state mapping and weight calculation.

[0167] Step S462: Map each droplet in the multidimensional decision cloud map to the quantum state space to generate the corresponding quantum state vector. The dimension of the quantum state vector is equal to the dimension corresponding to the quantum bit configuration.

[0168] Based on the qubit configuration initialized in step S461, each droplet in the multidimensional decision cloud map is mapped to the quantum state space. The quantum state space is a high-dimensional complex vector space, and each quantum state vector can be represented as a linear combination of qubits.

[0169] For example, with a qubit configuration of 3, the quantum state space has a dimension of 8. The feature information of a cloud droplet in a multidimensional decision cloud is encoded and converted into an 8-dimensional quantum state vector. A predefined encoding method can be used, such as binary encoding the cloud droplet's position coordinates and size information, and then mapping it to a vector in the quantum state space. Through this mapping operation, the classical information of the cloud droplet is converted into quantum state information, providing a quantum representation for quantum measurement and weight calculation.

[0170] Step S463: Perform quantum measurement operations on the quantum state vector to obtain the cloud droplet quantum measurement results, and calculate the initial weight value of each cloud droplet based on the cloud droplet quantum measurement results and the weight quantization interval.

[0171] Quantum measurement operations involve measuring a quantum state vector to obtain a classical measurement result. According to the principles of quantum mechanics, the measurement result is random but follows a certain probability distribution. In this embodiment, the measurement basis vector set in step S461 is used to measure the quantum state vector.

[0172] For example, for an 8-dimensional quantum state vector, measurement using orthogonal basis vectors yields an 8-bit binary measurement result. Based on the measurement result and the weight quantization interval, the initial weight value for each cloud droplet is calculated. The measurement result can be converted into a numerical value, then discretized according to the weight quantization interval to obtain the initial weight value. For example, if the measurement result corresponds to a numerical value of 0.35 and the weight quantization interval is 0.1, the initial weight value can be taken as 0.4. Through this quantum measurement and weight calculation operation, an initial spatial weight is assigned to each cloud droplet.

[0173] Step S464: Based on the spatial position of cloud droplets in the multidimensional decision cloud map, calculate the spatial distance between adjacent cloud droplets and generate a cloud droplet spatial distance matrix.

[0174] The spatial position of a cloud droplet in a multidimensional decision cloud map is determined by its coordinates, and the spatial distance between adjacent cloud droplets can be calculated based on these coordinates. Common distance metrics, such as Euclidean distance and Manhattan distance, can be used to calculate spatial distances.

[0175] The spatial distances between all adjacent cloud droplets are calculated and combined to form a cloud droplet spatial distance matrix. This cloud droplet spatial distance matrix is ​​a symmetric matrix whose elements d... ij This represents the spatial distance between cloud droplets i and j. Through this calculation, the relative position information of the cloud droplets in multidimensional space is obtained, providing a foundation for spatial smoothing processing.

[0176] Step S465: Perform spatial smoothing on the initial weight values ​​based on the cloud droplet spatial distance matrix to generate a cloud droplet weight matrix that considers spatial correlation. The dimension of the cloud droplet weight matrix is ​​the same as the dimension of the multidimensional decision cloud map.

[0177] Spatial smoothing aims to make the weight values ​​of adjacent cloud droplets more even and reasonable, taking into account the spatial correlation between cloud droplets. A distance-based weighted averaging method can be used to smooth the initial weight values.

[0178] For each cloud droplet i, its initial weight is w. i Based on the spatial distance matrix D of cloud droplets, calculate the weighted average weight value of its neighboring cloud droplets. Set a smoothing radius r, and only consider neighboring cloud droplets whose distance from cloud droplet i is less than r. For a neighboring cloud droplet j, its weight value is w. j The distance is d ij Weighting coefficients can be used (when d) ij When the sum is not equal to 0, it is weighted. Then the weight value of cloud droplet i after smoothing is... The smoothed weight values ​​of all cloud droplets are combined to form a cloud droplet weight matrix. The dimension of the cloud droplet weight matrix is ​​the same as the dimension of the multidimensional decision cloud map, thus taking into account the spatial correlation of cloud droplets in multidimensional space and making the weight allocation more reasonable.

[0179] Step S466: Standardize the cloud droplet feature values ​​in the multidimensional decision cloud map to generate a cloud droplet feature value standardization matrix. Multiply the cloud droplet weight matrix with the cloud droplet feature value standardization matrix of the multidimensional decision cloud map element by element to generate a weighted cloud map feature matrix.

[0180] To ensure the comparability and reasonableness of cloud droplet eigenvalues, the cloud droplet eigenvalues ​​in the multidimensional decision cloud map are first standardized. Standardization can employ common methods such as Z-score standardization or normalization. For example, using Z-score standardization, for each element f in the cloud droplet eigenvalue matrix F... ij Calculate its standardized value , where µ is the mean of the cloud droplet eigenvalue matrix, and σ is the standard deviation of the cloud droplet eigenvalue matrix. After standardization, the standardized cloud droplet eigenvalue matrix is ​​obtained. .

[0181] Then, the cloud droplet weight matrix W and the cloud droplet eigenvalue normalization matrix are combined. Perform element-wise multiplication. For each element w in the cloud droplet weight matrix W... ij and cloud droplet eigenvalue normalization matrix elements in Multiplying them yields the elements in the weighted contour map feature matrix M. This operation integrates the weight and feature information of cloud droplets, highlighting the features of important cloud droplets and providing more valuable data for dimensionality reduction.

[0182] Step S467: Perform principal component analysis to reduce the dimensionality of the weighted cloud map feature matrix, extract the low-dimensional feature matrix representing the main fault decision information, and generate the optimized decision matrix. The rows of the optimized decision matrix represent different fault types, and the columns represent the feature dimensions after dimensionality reduction.

[0183] Principal component analysis (PCA) can transform high-dimensional data into a low-dimensional representation while preserving the main information of the data. To perform PCA on the feature matrix of a weighted contour map, we first calculate its covariance matrix, and then solve for the eigenvalues ​​and eigenvectors of the covariance matrix.

[0184] The top k eigenvectors with the largest eigenvalues ​​are selected as principal components. The weighted contour map feature matrix is ​​projected onto these principal components to obtain a dimensionality-reduced low-dimensional feature matrix. The value of k is determined based on the actual situation and analysis requirements. The dimensionality-reduced low-dimensional feature matrix is ​​then rearranged so that rows represent different fault types and columns represent the dimensionality of the reduced features, generating an optimized decision matrix. The optimized decision matrix can more concisely represent the main fault decision information, removing redundant information and providing effective data support for the generation of fault early warning signals.

[0185] Step S500: Based on the confidence distribution characteristics of each fault type in the optimized decision matrix, a fault warning signal containing fault warning level and fault location information is generated through a dynamic threshold determination mechanism, and the fault warning signal is pushed to the equipment operation and maintenance management terminal.

[0186] The confidence distribution characteristics of each fault type in the optimized decision matrix reflect the probability and distribution of each fault type. The dynamic threshold determination mechanism is a method that dynamically adjusts thresholds based on the current operating status of the equipment and historical data. This mechanism allows for a more accurate determination of whether a fault warning signal needs to be issued.

[0187] As one implementation method, step S500 can be specifically implemented as the following steps S510~S570:

[0188] Step S510: Extract the fault type identifier and its confidence distribution features corresponding to each row in the optimization decision matrix, and generate a fault type-confidence mapping table. Each entry in the fault type-confidence mapping table contains a fault type identifier and a corresponding confidence vector.

[0189] Each row of the optimized decision matrix corresponds to a fault type. The fault type identifier and corresponding confidence distribution feature for each row are extracted from the matrix. The confidence distribution feature can be represented by a vector, where each element represents the confidence value of the fault type under different feature dimensions. For example, the first row of the optimized decision matrix corresponds to the fault type "short circuit fault," with a confidence distribution feature of (0.2, 0.3, 0.5). This is recorded in the fault type-confidence mapping table as ("short circuit fault," (0.2, 0.3, 0.5)). All fault types and their corresponding confidence vectors are combined to form the fault type-confidence mapping table. This mapping table clearly shows the confidence distribution of each fault type, providing a data foundation for peak detection and ranking.

[0190] Step S510: Extract the fault type identifier and its confidence distribution features corresponding to each row in the optimization decision matrix, and generate a fault type-confidence mapping table. Each entry in the fault type-confidence mapping table contains a fault type identifier and a corresponding confidence vector.

[0191] Each row of the optimization decision matrix represents a fault type. By traversing each row, the identifier of each fault type and its corresponding confidence distribution feature can be accurately obtained. For example, in an optimization decision matrix containing multiple fault types, a certain row might correspond to "motor overheating fault," whose confidence distribution feature is a vector. The elements in the vector represent the confidence values ​​of this fault type under different feature dimensions. These feature dimensions can be parameters closely related to the equipment's operating state, such as temperature, current, and speed. Suppose that for "motor overheating fault," its confidence vector is (0.2, 0.6, 0.2), which may mean that the confidence of this fault occurring in the three feature dimensions of temperature, current, and speed are 0.2, 0.6, and 0.2, respectively. Combining the identifier of each fault type and its corresponding confidence vector into an entry, all these entries together constitute a fault type-confidence mapping table. This mapping table acts like an index, facilitating quick searching and analysis of the confidence status of each fault type. By establishing such a mapping table, the information in the optimization decision matrix can be structured and organized, providing a clear data foundation for fault analysis and early warning.

[0192] Step S520: Perform peak detection processing on each confidence vector in the fault type-confidence mapping table to determine the maximum confidence value and its location index corresponding to each fault type.

[0193] Peak detection is a method for identifying local maxima in a data sequence. In this embodiment, the goal is to find the maximum confidence value and its index in each confidence vector. For each fault type's corresponding confidence vector, a suitable peak detection algorithm is used. For example, a simple iterative comparison method can be employed, starting with the first element of the confidence vector and comparing it sequentially with subsequent elements, recording the current maximum value and its index. Assuming a fault type's confidence vector is (0.1, 0.3, 0.5, 0.2), through iterative comparison, the maximum value is found to be 0.5, with an index of 2 (indexes start counting from 0). This maximum confidence value represents the highest probability of this fault type occurring in a certain feature dimension, while the index indicates the corresponding feature dimension. By performing peak detection on all confidence vectors in the fault type-confidence mapping table, a comprehensive understanding can be obtained of which feature dimension has the highest confidence for each fault type. This is crucial for subsequently determining the main characteristics and priority of the fault.

[0194] Step S530: Sort all fault types based on the maximum confidence value to generate a fault type priority sorting table. The fault type priority sorting table is arranged in descending order of the maximum confidence value.

[0195] After obtaining the maximum confidence value for each fault type, all fault types need to be sorted according to these values ​​to determine their priority. The purpose of sorting is to prioritize fault types with a higher probability of occurrence when issuing fault warnings. Common sorting algorithms, such as bubble sort and quicksort, can be used. Arrange all fault types in descending order of their maximum confidence values ​​to form a fault type priority ranking table.

[0196] Step S540: Invoke the dynamic threshold determination mechanism, calculate the dynamic threshold for fault warning under the current operating conditions based on the current operating stage of the equipment and historical fault statistics. The dynamic threshold for fault warning is dynamically adjusted with the equipment operating time and the cumulative number of faults.

[0197] The dynamic threshold determination mechanism is an intelligent threshold determination method that considers the actual operating conditions and historical data of the equipment, enabling more accurate judgment of whether a fault has occurred. First, it collects relevant information about the current operating phase of the equipment, including parameters such as cumulative operating time, the last maintenance time, and the current load rate. These parameters reflect the current operating status and condition of the equipment. Simultaneously, it queries the equipment's historical fault statistics database to obtain historical fault records for the same equipment model, statistically analyzes the probability of fault occurrence under different standardized operating phases and load rates, and generates a fault probability distribution matrix. This matrix shows the likelihood of fault occurrence under different operating conditions. Based on the set of equipment operating status parameters and the fault probability distribution matrix, a Bayesian inference method is used to calculate the prior probability of fault occurrence under the current operating condition. Bayesian inference is a probability-based inference method that uses prior knowledge and new observation data to update the probability estimate. Based on the calculated prior probability of fault occurrence, combined with preset constraints on the false alarm rate and missed alarm rate of fault warnings, a threshold optimization objective function is constructed. The optimization objective of this objective function is to minimize the overall warning cost, that is, to determine a suitable fault warning threshold while minimizing false alarms and missed alarms. The objective function for threshold optimization is solved using the Particle Swarm Optimization (PSO) algorithm to obtain the initial fault warning threshold under the current operating conditions. PSO is a swarm intelligence-based optimization algorithm that finds the optimal solution by simulating the collective behavior of flocks of birds or schools of fish. Finally, the initial fault warning threshold is dynamically adjusted based on the cumulative number of equipment failures. When the cumulative number of failures exceeds a preset threshold, it indicates that the equipment's reliability may be decreasing, and the fault warning threshold should be appropriately lowered to detect faults earlier. When the cumulative fault-free operating time exceeds a preset threshold, it indicates that the equipment is operating well, and the fault warning threshold can be appropriately increased to reduce unnecessary warnings. The adjusted fault warning threshold is compared with the preset upper and lower limits to ensure that the fault warning threshold is within a reasonable range, ultimately generating the dynamic fault warning threshold for the current operating conditions. This dynamic threshold can be adjusted in real time based on the actual operating conditions and historical data of the equipment, improving the accuracy and reliability of fault warnings.

[0198] Step S550: Compare the maximum confidence value of each fault type in the fault type priority sorting table with the fault warning dynamic threshold, filter out the fault types with the maximum confidence value greater than the fault warning dynamic threshold, and generate a set of potential fault types.

[0199] After obtaining the fault type priority ranking table and the fault warning dynamic threshold, the two need to be compared to determine which fault types are potential faults. For each fault type in the fault type priority ranking table, its maximum confidence value is compared with the fault warning dynamic threshold. If the maximum confidence value of a fault type is greater than the fault warning dynamic threshold, it indicates that the fault type is more likely to occur and has potential fault risk, and it is included in the potential fault type set.

[0200] Step S560: For each fault type in the potential fault type set, determine the corresponding fault warning level and fault location information based on its confidence vector and the feature distribution in the optimization decision matrix.

[0201] For each fault type in the potential fault type set, the fault warning level and fault location information are determined by comprehensively considering its confidence vector and the feature distribution in the optimized decision matrix. The confidence vector reflects the confidence level of the fault type across different feature dimensions, while the feature distribution in the optimized decision matrix shows the correlation between each feature dimension and the fault type. First, the fault warning level is determined based on preset rules and standards, combined with the maximum confidence value in the confidence vector. For example, the maximum confidence value can be divided into different intervals, each corresponding to a fault warning level, such as high, medium, and low. If the maximum confidence value is greater than 0.8, it is determined as a high-level warning; if it is between 0.5 and 0.8, it is a medium-level warning; and if it is less than 0.5, it is a low-level warning. Then, based on the location index of the maximum confidence value in the confidence vector, combined with the feature dimension information in the optimized decision matrix, the fault location information is determined. Assuming the feature dimension corresponding to the location index of the maximum confidence value is the temperature of a component of the equipment, the fault location can be located at that component.

[0202] Step S570: Integrate fault warning level and fault location information to generate a fault warning signal that includes fault type identifier, warning level, location coordinates and confidence value.

[0203] After determining the fault warning level and fault location information for each potential fault type, this information is fused with the fault type identifier and corresponding confidence value to generate a complete fault warning signal. For each potential fault type, its fault type identifier, warning level, location coordinates (determined by the fault location information), and corresponding maximum confidence value are combined. For example, for a "short circuit fault," with a high warning level, location coordinates at a circuit node of the equipment, and a maximum confidence value of 0.8, the generated fault warning signal can be represented as ("short circuit fault," "high-level warning," (x, y), 0.8). The fault warning signals for all potential fault types are combined to form the final fault warning signal set. This set contains comprehensive fault information, clearly informing the equipment operation and maintenance management terminal about the type, severity, location, and probability of the fault. Finally, the fault warning signal is pushed to the equipment operation and maintenance management terminal, enabling maintenance personnel to promptly understand the equipment fault status and take appropriate measures to handle it.

[0204] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention, such as Euclidean distance algorithm, cosine distance algorithm, conflict resolution algorithm, etc., can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solution of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set the threshold based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select the activation function, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0205] Figure 2 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 2 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0206] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) of the processor 1001 and various modules in the computer system 1000. It can be implemented by flash memory or random access memory (RAM).

[0207] The processor 1001 executes the program to implement any of the steps of the hybrid model fault warning method based on timing prediction and fuzzy logic mentioned above. The processor 1001 typically controls the overall operation of the computer system 1000.

[0208] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault early warning method based on a hybrid model of time-series prediction and fuzzy logic, characterized in that, The method includes: Collect operation and maintenance data of the equipment during continuous operation cycles, and construct an equipment status feature set. The equipment status feature set includes the feature sequence of each operating parameter of the equipment at different monitoring times and the corresponding environmental impact feature sequence. A feature correlation analysis is performed on the equipment state feature set to generate an equipment state potential vector that reflects the dynamic coupling relationship between parameters. The equipment state potential vector includes the trend evolution features and cross-influence features of the feature sequence. The pre-trained fuzzy logic reasoning model is invoked to perform fuzzy rule matching on the device state potential vector, generating a fuzzy state set containing multi-dimensional fuzzy subsets; The fuzzy state set is spatiotemporally correlated with a pre-defined domain fault knowledge base to generate a multidimensional decision cloud map with spatiotemporal correlation characteristics. The multidimensional decision cloud map is then spatially reconstructed using a quantum weight allocation mechanism to obtain an optimized decision matrix. Based on the confidence distribution characteristics of each fault type in the optimization decision matrix, a fault warning signal containing fault warning level and fault location information is generated through a dynamic threshold determination mechanism, and the fault warning signal is pushed to the equipment operation and maintenance management terminal. The process involves performing feature correlation analysis on the equipment state feature set to generate an equipment state potential vector reflecting the dynamic coupling relationship between parameters. This device state potential vector includes trend evolution features and cross-influence features of the feature sequence, including: The time series decomposition process is performed on each feature sequence in the device state feature set to extract the trend component and fluctuation component of each feature sequence at different time scales, resulting in a set of decomposed multi-scale feature sequences. Calculate the mutual information value between each feature sequence in the multi-scale feature sequence set, and generate a feature mutual information correlation matrix. The element values ​​in the feature mutual information correlation matrix represent the degree of nonlinear correlation between two corresponding feature sequences. Based on the feature mutual information correlation matrix, the feature sequences are clustered to generate a set of feature communities with cohesion. Each feature community set contains multiple feature sequences with correlation. Principal component analysis is performed on the feature sequences in each feature community set to extract principal component feature vectors that characterize the overall trend of change of feature sequences within the community. The dimension of the principal component feature vectors is the same as the number of feature community sets. The principal component feature vectors corresponding to all feature community sets are standardized to eliminate dimensional differences and generate a standardized principal component feature vector set. The standardized principal component feature vector set is then concatenated to generate a trend evolution feature containing the dynamic evolution law of each feature community. At the same time, the cross-influence coefficient matrix between different feature community sets is calculated as the cross-influence feature. The trend evolution feature and the cross-influence feature are fused to generate a device state potential vector that represents the dynamic coupling relationship between parameters. The dimension of the device state potential vector is equal to the sum of the dimension of the trend evolution feature and the dimension of the cross-influence feature.

2. The method according to claim 1, characterized in that, The process involves performing time series decomposition on each feature sequence in the device state feature set, extracting the trend and fluctuation components of each feature sequence at different time scales, and obtaining a set of decomposed multi-scale feature sequences, including: Initialize the time scale decomposition parameter set, which includes multiple preset time window lengths and corresponding decomposition levels, with each time window length distributed in a geometric sequence. For each feature sequence, multi-scale decomposition is performed sequentially according to the length of each time window in the time scale decomposition parameter set to generate approximation coefficients and detail coefficients at different decomposition levels. The approximate coefficients under the same decomposition level are reconstructed to generate trend components at the corresponding time scale. The trend components reflect the overall change trend of the feature sequence at that time scale. The detail coefficients under the same decomposition level are reconstructed to generate the fluctuation component at the corresponding time scale. The fluctuation component reflects the local fluctuation characteristics of the feature sequence at that time scale. The trend components and fluctuation components of all feature sequences at various time scales are summarized to obtain a multi-scale feature sequence set containing feature expressions at different time scales. Each element in the multi-scale feature sequence set contains a feature identifier, a time scale parameter, a trend component, and a fluctuation component.

3. The method according to claim 1, characterized in that, The feature sequences are clustered based on the feature mutual information correlation matrix to generate a set of cohesive feature communities. Each feature community set contains multiple correlated feature sequences, including: The feature mutual information correlation matrix is ​​converted into a feature correlation network, where network nodes represent feature sequences and the weight values ​​of network edges represent the degree of nonlinear correlation between two corresponding feature sequences. Initialize the community assignment vector and randomly assign each feature sequence to a different initial community. The number of initial communities is equal to the number of equal divisions of the total number of feature sequences. Calculate the modularity contribution value for each feature sequence. The modularity contribution value is used to measure the change in the overall modularity of the network when the feature sequence is moved from the current community to other communities. Adjust the community affiliation of the feature sequences in descending order of modularity contribution value until the overall modularity of the network no longer increases or reaches the preset iteration limit. The adjusted community structure is then merged, and small communities with a modularity contribution value less than a preset threshold are merged into adjacent communities to generate the final set of feature communities. Calculate the internal average mutual information value and the inter-community average mutual information value for each feature community set to verify the aggregation characteristics of the feature community sets, so that the internal average mutual information value and the inter-community average mutual information value exhibit a discriminative distribution.

4. The method according to claim 1, characterized in that, The process involves calling a pre-trained fuzzy logic inference model to perform fuzzy rule matching on the device state potential vector, generating a fuzzy state set containing multi-dimensional fuzzy subsets, including: The device state potential vector is input into the fuzzification layer of the fuzzy logic reasoning model. The fuzzification of each dimension feature of the device state potential vector is performed through a preset set of membership functions to generate the corresponding fuzzy membership vector. Extract the linguistic variable identifiers corresponding to each dimension of the fuzzy membership vector, and construct a set of fuzzy feature descriptors containing linguistic variable names and membership values; The rule base module of the fuzzy logic reasoning model is invoked to perform rule matching operations based on the fuzzy feature descriptor set, and to filter out the preset fuzzy rule set with the highest matching degree with the current fuzzy feature descriptor set; For each fuzzy rule in the preset fuzzy rule set, the trigger intensity is calculated to generate a rule trigger intensity vector. The dimension of the rule trigger intensity vector is the same as the number of rules in the preset fuzzy rule set. The preset fuzzy rule set is weighted and aggregated according to the values ​​of each element in the rule-triggered intensity vector to generate an intermediate fuzzy decision vector containing multiple fuzzy decision variables. The intermediate fuzzy decision vector is input into the defuzzification layer of the fuzzy logic reasoning model to perform defuzzification processing on the intermediate fuzzy decision vector, generating a fuzzy state set containing multi-dimensional fuzzy subsets, each fuzzy subset corresponding to a fuzzy description of a device operating state category.

5. The method according to claim 4, characterized in that, The step of inputting the device state potential vector into the fuzzification layer of the fuzzy logic reasoning model, and fuzzifying each dimension of the device state potential vector using a preset set of membership functions to generate a corresponding fuzzy membership vector includes: The fuzzification layer configuration parameters of the fuzzy logic reasoning model are analyzed, and the membership function types and parameter sets corresponding to the features of each dimension of the device state potential vector are extracted. For each dimension feature of the device state potential vector, the membership degree value of the dimension feature to each preset fuzzy set is calculated according to the corresponding membership function type and parameter set, and a dimension membership vector is generated. The length of the dimension membership vector is equal to the number of preset fuzzy sets corresponding to the dimension feature. The dimensional membership vectors are normalized so that the sum of all element values ​​in each dimensional membership vector is 1, thus generating standardized dimensional membership vectors. The standardized dimensional membership vectors corresponding to all dimensional features are concatenated according to the dimensional order of the device state potential vector to generate a fuzzy membership vector containing fuzzy information of all dimensional features. Calculate the fuzzy entropy value of the fuzzy membership vector, evaluate the effectiveness of the current fuzzification process, and when the fuzzy entropy value is lower than a preset threshold, adjust the membership function parameter set and perform fuzzification process again until the fuzzy entropy value is greater than the preset threshold.

6. The method according to claim 4, characterized in that, The rule base module that invokes the fuzzy logic reasoning model performs rule matching operations based on the fuzzy feature descriptor set, filtering out a preset fuzzy rule set with the highest matching degree to the current fuzzy feature descriptor set, including: The fuzzy feature descriptor set is parsed, and a feature-membership pair list containing linguistic variable names and membership values ​​is extracted. Each element in the feature-membership pair list contains a linguistic variable name and its corresponding membership value. Iterate through all the preset fuzzy rules in the rule base module, and for each preset fuzzy rule, extract the list of rule-feature pairs contained in its precondition part; Calculate the matching degree value between the feature-membership pair list and the rule-feature pair list of each preset fuzzy rule. The matching degree value is obtained by calculating the sum of the products of the membership values ​​corresponding to the same linguistic variable names in the two lists. Based on the matching degree value, all preset fuzzy rules are sorted in descending order to generate a rule matching degree sorting table; Select preset fuzzy rules with matching scores greater than a preset matching threshold from the rule matching score sorting table to generate a preliminary matching rule set; The preliminary matching rule set is subjected to rule conflict identification. When there are preset fuzzy rules with contradictory rule conclusions, the preset fuzzy rule with the higher matching degree value is retained to generate the final preset fuzzy rule set.

7. The method according to claim 1, characterized in that, The process involves combining a pre-defined domain fault knowledge base with spatiotemporal correlation processing of the fuzzy state set to generate a multidimensional decision cloud map with spatiotemporal correlation characteristics. A quantized weight allocation mechanism is then used to spatially reconstruct the multidimensional decision cloud map to obtain an optimized decision matrix, including: The preset domain fault knowledge base is parsed to extract a set of fault cases containing fault type identifiers, fault feature vectors, and fault evolution paths, and a fault case feature index table is constructed. Calculate the commonality metric between the fuzzy state set and the fault feature vectors of each fault case in the fault case feature index table, and generate a fault similarity vector. The dimension of the fault similarity vector is the same as the number of fault cases in the set. Based on the fault similarity vector, a preset number of fault cases with the highest similarity values ​​are selected to generate a set of similar fault cases. The set of similar fault cases includes fault type identifiers, historical fault feature evolution sequences, and fault handling solutions. The historical fault feature evolution sequence in the set of similar fault cases is spatiotemporally aligned with the current fuzzy state set to generate a spatiotemporal correlation matrix containing time dimension and feature dimension. The rows of the spatiotemporal correlation matrix represent the time step and the columns represent the feature dimension. The spatiotemporal correlation matrix is ​​converted into a multidimensional decision cloud map. Each cloud droplet in the multidimensional decision cloud map represents the fault characteristic state under the corresponding time step and feature dimension. The position coordinates of the cloud droplet are determined by the row index and column index of the spatiotemporal correlation matrix, and the size of the cloud droplet is determined by the size of the feature value at the corresponding position. The cloud droplets in the multidimensional decision cloud map are spatially weighted using a quantum weighting allocation mechanism to generate a cloud droplet weight matrix. Based on the cloud droplet weight matrix, the multidimensional decision cloud map is spatially reconstructed to obtain an optimized decision matrix. The row dimension of the optimized decision matrix is ​​equal to the number of fault types, and the column dimension is equal to the number of feature dimensions.

8. The method according to claim 7, characterized in that, The process involves parsing the preset domain fault knowledge base, extracting a set of fault cases containing fault type identifiers, fault feature vectors, and fault evolution paths, and constructing a fault case feature index table, including: Connect to the preset domain fault knowledge base database, perform structured query operations, and extract basic information of fault cases stored in relational data tables. The basic information of fault cases includes a unique identifier of the fault case, a fault type identifier, a fault occurrence timestamp, and a fault handling result. The unstructured data associated with the basic information of the fault cases is parsed, and time-series data containing text information describing the fault development process and fault characteristic change curves are extracted to generate fault evolution path descriptors. Feature extraction processing is performed on the time-series data in the fault evolution path descriptor to generate a fault feature vector that represents the entire process of fault occurrence and development. The dimension of the fault feature vector is the same as the dimension of the equipment state potential vector. The unique identifier of the fault case, the fault type identifier, the fault feature vector and the fault evolution path descriptor are associated and stored to generate a fault case set containing complete fault case information. Principal component analysis is performed on all fault feature vectors in the fault case set to extract principal component features that represent the main changes in fault feature vectors and construct a fault case feature space. Each fault case in the fault case set is mapped to the fault case feature space to generate a fault case feature index table containing the fault case coordinates and fault type identifiers, which is used to quickly retrieve similar fault cases.

9. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images