Method and device for processing work order, and electronic equipment

Abstract

The invention relates to the technical field of information technology and data processing, and provides a method for processing a work order, which comprises the following steps: acquiring historical work order data, and cleaning a work order text in the historical work order data to obtain work order text information; performing word segmentation and synonym normalization processing on the work order text information; converting the processed work order text information into a vector with a fixed dimension so as to construct a work order vector matrix; performing clustering analysis on the work order matrix to determine a cluster to which each work order belongs; calculating the center of mass of each cluster; and based on the centroid, determining the work order closest to the centroid from each cluster as a typical representative work order, and determining the work order whose distance to the centroid exceeds a preset threshold as a group-leaving work order. And furthermore, the work order data is classified and integrated, the overall distribution and main and main contradictions of the work orders are quickly mastered, a basis is provided for resource allocation and priority ranking, and the reply quality of the work orders is improved.

Description

Technical Field

[0001] This invention relates to the field of information technology and data processing, and in particular to a method, apparatus, and electronic device for processing work orders. Background Technology

[0002] The massive volume of work orders has been driven by the rapid development of information technology services. User problems, requests, or malfunctions are recorded in these work orders, which serve as important operational data resources for enterprises. Improving response speed, optimizing resource allocation, enhancing service quality, and identifying potential system problems all require enterprises to efficiently and accurately classify, analyze, and process work orders.

[0003] Initially, work order processing was dominated by human reading and experience-based judgment, which reduced efficiency and affected the stability of classification quality due to subjective human factors. The need for efficiency improvements led to the gradual adoption of rule-based automation methods in the industry, such as keyword matching for work order assignment and the implementation of decision tree models. These methods alleviated some of the manual burden, but their limited flexibility and high maintenance costs restricted their application, and they struggled to fully cover the diversity of work order descriptions and the continuous emergence of new problem types.

[0004] In recent years, new approaches have been offered by machine learning, particularly Natural Language Processing (NLP), for the automatic classification of work orders. Supervised text classification models, such as Support Vector Machines (SVM) or shallow neural networks, have been attempted in some solutions. However, these methods typically rely on large amounts of labeled, high-quality historical data as training sets, which incurs high costs and long development cycles for enterprises. Furthermore, the adaptability of these models is limited by changing business category systems.

[0005] Potential categories within work orders are being explored using unsupervised clustering algorithms such as K-Means. While labeled data doesn't require pre-preparation for this approach, the high-dimensional, sparse text features limit the performance of traditional clustering algorithms, making it difficult to effectively capture complex semantic relationships between texts. Furthermore, key issues such as accurately locating core work orders within the clustering results and effectively identifying outliers remain unresolved by current technologies. Emerging faults or complex issues may be foreshadowed by these outliers, requiring higher processing priority.

[0006] Therefore, existing work order processing methods have limitations in terms of efficiency, accuracy, and adaptability, which have constrained the improvement of operational service levels. It is necessary to classify and integrate work order data, quickly grasp the overall distribution and main contradictions of work orders, provide a basis for resource allocation and priority ranking, and improve the quality of work order responses, work order processing efficiency, and the uniformity of processing quality. Summary of the Invention

[0007] The purpose of this invention is to improve the efficiency of work order processing and the uniformity of processing quality.

[0008] To address the above objectives, in one embodiment of the present invention, a method for processing work orders to improve the quality of work order responses is proposed, comprising the following steps: S11: Obtain historical work order data and clean the work order text in the historical work order data to obtain work order text information. S12: Perform word segmentation and synonym normalization on the work order text information; S13: Convert the processed work order text information into a fixed-dimensional vector to construct a work order vector matrix; S14: Perform cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs; S15: Calculate the centroid of each cluster; S16: Based on the centroid, identify the work order closest to the centroid from each cluster as a typical representative work order, and identify work orders that are more than a preset threshold away from the centroid as outlier work orders.

[0009] Optionally, in some embodiments of the work order processing method, the step of calculating the centroid of each cluster includes: S151: Based on predefined comprehensive weight coefficients, perform weighted average calculation on the work order vectors within the cluster to generate a weighted centroid vector; The comprehensive weighting coefficient is dynamically calculated based on the processing time of the corresponding work order, semantic matching degree, and customer rating. S152: Based on the Mahalanobis distance between the weighted centroid vector and each vector within the cluster, the judgment thresholds for typical representative work orders and outlier work orders are recalibrated.

[0010] Optionally, in some embodiments of the work order processing method, the step of converting the processed work order text information into a fixed-dimensional vector to construct a work order vector matrix includes: S131: Use a pre-trained language model to perform deep semantic encoding on the work order text information after word segmentation and synonym normalization to generate a high-dimensional semantic vector; S132: Standardize the structured feature data of the work order to generate a numerical feature vector; S133: Concatenate and reduce the dimensionality of the high-dimensional semantic vector and the numerical feature vector to generate the final work order vector used to construct the work order vector matrix.

[0011] Optionally, in some embodiments of the work order processing method, the steps of performing word segmentation and synonym normalization on the work order text information include: S121: Perform word segmentation on the work order text information to generate an initial word segmentation sequence; S122: Perform entity recognition and semantic verification on terms in the initial word segmentation sequence based on the domain knowledge base; S123: Based on the pre-constructed thesaurus, the semantically validated terms are mapped to standard concept nodes to complete the thesaurus normalization process.

[0012] Optionally, in some embodiments of the work order processing method, the step of determining the work order closest to the centroid from each cluster as a typical representative work order based on the centroid includes: S161: Calculate the Mahalanobis distance from each work order vector within the cluster to the centroid; S162: Select the top K work orders with the smallest Mahalanobis distance as a set of candidate typical representative work orders; S163: Sort the work orders in the candidate typical representative work order set according to their comprehensive weight, and select the work order with the highest comprehensive weight as the final typical representative work order.

[0013] Optionally, in some embodiments of the work order processing method, the step of identifying work orders whose distance from the centroid exceeds a preset threshold as outlier work orders includes: S164: Calculate the mean and standard deviation of distances based on the distance distribution from all work orders to the centroid within the cluster; S165: Based on the distance mean and standard deviation, dynamically set the preset threshold to the distance mean plus N times the standard deviation; S166: Mark work orders whose distance exceeds the dynamically set preset threshold as critical outlier work orders; S167: Perform secondary clustering analysis on the key outlier work order set to distinguish emerging problem patterns from random noise.

[0014] Optionally, some embodiments of the work order processing method further include resource optimization allocation based on the distribution of work order clusters; this resource optimization allocation based on the distribution of work order clusters further includes the following steps: S171: Calculate the processing resource weights required for each cluster based on the size of each work order cluster and the complexity of the problems reflected by typical representative work orders. S172: Based on the weight of processing resources, dynamically adjust the matching relationship between support personnel and work order clusters; S173: Assign outlier sets to higher-level expert queues for specialized processing, and prioritize resource allocation to work order clusters containing emerging problem patterns.

[0015] Optionally, the work order processing method in some embodiments further includes establishing a feedback loop to continuously optimize the clustering model; the establishment of the feedback loop to continuously optimize the clustering model further includes the following steps: S181: Monitor the processing effect of typical representative work orders and collect feedback data from support personnel on the accuracy of cluster classification; S182: Based on feedback data and processing results, dynamically adjust the parameters of the synonym normalization rules and the vectorization model; S183: Re-execute cluster analysis based on the optimized parameters to generate an updated work order cluster distribution and typical representative work orders; S184: Compare and analyze the updated results with historical data, evaluate the optimization effect, and generate a model performance report.

[0016] Optionally, some embodiments of the work order processing method further include building a visual analysis interface; the building of the visual analysis interface further includes the following steps: S191: Generate a multidimensional pivot table based on the cluster distribution of work orders and extract key features of typical representative work orders; S192: A dimensionality reduction algorithm is used to project the work order vector matrix onto a two-dimensional space to generate a visual scatter plot; S193: Mark the locations of typical representative work orders and the distribution areas of outlier work orders in the scatter plot; S194: Clustering results, work order details, and resource allocation schemes are displayed in a linked manner through an interactive interface.

[0017] Optionally, in some embodiments of the work order processing method, the step of calculating the centroid of each cluster specifically includes: S154: Construct the Mahalanobis distance metric space based on the weighted covariance matrix of each work order vector within the cluster; S155: Solve for the centroid of the probability distribution of the cluster by minimizing the objective function of the weighted sum of squared Mahalanobis distances; S156: Generate a geometric space reference coordinate system to represent work order screening based on the normalized distance metric between the centroid of the probability distribution and each vector within the cluster. S157: Based on the statistical distribution characteristics under this reference coordinate system, dynamically optimize the significance level threshold for outlier determination.

[0018] Preferably, in some embodiments of the work order processing method, the step of determining outlier work orders based on centroid further includes: S168: Construct a chi-square distribution probability model of the Mahalanobis distance between samples within a cluster, and calculate its cumulative distribution function; S169: Determine the right-hand critical value θ_α based on the significance level α, and establish the first criterion for outlier determination based on statistical inference; S170: In parallel, calculate the interquartile range of the distance set and dynamically generate a second outlier determination criterion based on empirical distribution; S171: Integrate the judgment results of the first criterion and the second criterion to generate the final outlier work order identifier and verify its significance through hypothesis testing.

[0019] Optionally, in some embodiments of the work order processing method, the step of performing cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs further includes: S141: Based on the kernel density estimation method, the work order vector matrix is ​​mapped to a high-dimensional probability distribution manifold; S142: Gradient ascent optimization is performed on the manifold using the mean drift algorithm, and the iteration converges to a local density maximum to determine the cluster mode; S143: Construct an attraction domain centered on each mode, and assign all work orders that converge to the same mode to the same cluster; S144: Based on the theory of persistent homology, calculate the topological persistent graph of the probability distribution manifold, identify transient connected components with lifetimes below a threshold, and mark their corresponding work orders as geometric outliers.

[0020] Preferably, in some embodiments of the work order processing method, the step of performing gradient ascent optimization on the manifold using the mean-shift algorithm and iteratively converging to a local density maximum point to determine the cluster mode further includes: S1421: Construct a gradient vector field on a probability distribution manifold based on Riemannian metrics, and compute the Lie derivative of the vector field along the integral curve; S1422: The dynamic system is numerically solved using the variable step size Runge-Kutta method, which accurately tracks the evolution trajectory of the work order vector on the manifold; S1423: When the norm of the Lie derivative is lower than the preset convergence threshold, the trajectory is determined to have converged to the steady-state attractor; S1424: Perform homotopy classification on all trajectories converging to the same attractor and establish the equivalence relation of work single clusters based on differential homeomorphism.

[0021] Optionally, in some embodiments of the work order processing method, the step of concatenating and reducing the dimensionality of the high-dimensional semantic vector and the numerical feature vector to generate the final work order vector for constructing the work order vector matrix further includes: S1331: Construct a deep feature distillation network based on an autoencoder, where the encoder is jointly optimized by minimizing the reconstruction error and maximizing the mutual information of latent features; S1332: Introduce orthogonal and sparsity constraints in the latent space to ensure that the dimensionality-reduced feature vectors maintain the manifold geometry and semantic relationships between modes; S1333: The backpropagation algorithm and adaptive moment estimation optimizer are used to iteratively update the weight parameters of the deep feature distillation network; S1334: Finally, through the forward propagation calculation of the encoder, the multimodal fusion vector is mapped to a low-dimensional dense vector representation.

[0022] Optionally, in some embodiments of the work order processing method, the step of constructing the Mahalanobis distance metric space based on the weighted covariance matrix of each work order vector within the cluster further includes: S1541: Perform eigenvalue decomposition on the weighted covariance matrix and extract its eigenvector matrix and eigenvalue diagonal matrix to characterize the data distribution principal axis and scaling factor of the cluster; S1542: Rotate the original work order vector using the eigenvector matrix and normalize the scale using the eigenvalue diagonal matrix to generate a whitened vector. S1543: Define the Mahalanobis distance metric based on the Euclidean distance between whitened vectors and establish an isotropic distribution space; S1544: By calculating the geometric distance between the weighted centroid vector and each whitening vector in this metric space, the work order can be accurately located and its similarity measured.

[0023] Optionally, some embodiments of the work order processing method further include a root cause analysis process based on association rule mining for key outlier work order sets or work order clusters containing emerging problem patterns: S161: Extract the structured metadata feature values ​​of key outlier work orders or work orders within work order clusters to form transaction item sets; S162: Perform frequent itemset mining on the transaction itemsets, calculate the support of each itemset, and filter out frequent itemsets whose support exceeds the preset minimum support threshold. S163: Generate association rules based on frequent itemsets, calculate the confidence level of each rule, and retain high-confidence rules whose confidence level exceeds the preset minimum confidence level threshold; S164: Associate high-confidence rules with the semantic topics of the corresponding work order clusters, identify the root cause feature combinations that lead to work order anomalies or frequent occurrences, and generate a root cause analysis report.

[0024] Optionally, in some embodiments of the work order processing method, after generating the root cause analysis report, the method further includes: S165: Based on the combination of root cause characteristics, generate corresponding preventive treatment strategy recommendations; S166: Push preventative handling strategy recommendations to the management terminal and provide decision support for subsequent work order processing.

[0025] Optionally, in some embodiments, step S14, which involves performing cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs, further includes: S141a: Mapping the work order vector matrix into a high-dimensional probability distribution manifold based on the kernel density estimation method; S142a: Gradient ascent optimization is performed on this manifold using the mean drift algorithm, and the cluster mode is determined by iterative convergence to a local density maximum point. S143a: Construct an attraction domain centered on each mode, and assign all work orders that converge to the same mode to the same cluster; S144a: Based on the theory of persistent homology, calculate the topological persistent graph of the probability distribution manifold, identify transient connected components with lifetimes below a threshold, and mark their corresponding work orders as geometric outliers.

[0026] Optionally, in some embodiments, step S142, which involves performing gradient ascent optimization on the manifold using the mean-shift algorithm and iteratively converging to a local density maxima to determine the cluster mode, further includes: S1421a: Construct a gradient vector field on a probability distribution manifold based on Riemannian metrics, and compute the Lie derivative of the vector field along the integral curve; S1422a: The dynamic system is numerically solved using the variable step size Runge-Kutta method, which accurately tracks the evolution trajectory of the work order vector on the manifold; S1423a: When the norm of the Lie derivative is lower than the preset convergence threshold, the trajectory is determined to have converged to the steady-state attractor; S1424a: Homotopy classification is performed on all trajectories converging to the same attractor, and the equivalence relation of work single clusters based on differential homeomorphism is established.

[0027] Optionally, in some embodiments, step S133, which concatenates and reduces the dimensionality of the high-dimensional semantic vector and the numerical feature vector to generate the final work order vector for constructing the work order vector matrix, further includes: S1331a: Construct a deep feature distillation network based on an autoencoder, whose encoder is jointly optimized by minimizing reconstruction error and maximizing mutual information of latent features; S1332a: Introduce orthogonal and sparsity constraints into the latent space to ensure that the dimensionality-reduced feature vectors maintain the manifold geometric structure and semantic associations between modes; S1333a: The backpropagation algorithm and adaptive moment estimation optimizer are used to iteratively update the weight parameters of the deep feature distillation network; S1334a: Finally, through the forward propagation calculation of the encoder, the multimodal fusion vector is mapped into a low-dimensional dense vector representation.

[0028] Optionally, in some embodiments, step S154, which constructs the Mahalanobis distance metric space based on the weighted covariance matrix of each work order vector within the cluster, further includes: S1541a: Perform eigenvalue decomposition on the weighted covariance matrix and extract its eigenvector matrix and eigenvalue diagonal matrix to characterize the data distribution principal axis and scaling factor of the cluster; S1542a: Rotate the original work order vector using the eigenvector matrix and normalize the scale using the eigenvalue diagonal matrix to generate a whitened vector. S1543a: Based on the Euclidean distance between whitened vectors, define the Mahalanobis distance metric and establish an isotropic distribution space; S1544a: By calculating the geometric distance between the weighted centroid vector and each whitening vector in this metric space, the work order can be accurately located and its similarity measured.

[0029] Optionally, in some embodiments, step S14 of performing cluster analysis on the work order vector matrix specifically involves performing a clustering operation based on density peaks on the work order vector matrix to divide it into multiple work order clusters; Step S15, which calculates the centroid of each cluster, specifically involves calculating the weighted centroid vector of each work order cluster based on predefined business weight coefficients. Step S16, which determines based on the centroid, specifically involves: using the Mahalanobis distance metric between the weighted centroid vector and the feature vectors of each work order, selecting typical representative work orders from each work order cluster and identifying outlier work orders.

[0030] Optionally, some embodiments of the method further include: performing a root cause analysis process based on association rule mining on a set of key outlier work orders or a cluster of work orders containing emerging problem patterns, the process including: S161: Extract the structured metadata feature values ​​of key outlier work orders or work orders within work order clusters to form transaction item sets; S162: Perform frequent itemset mining on the transaction itemsets, calculate the support of each itemset, and filter out frequent itemsets whose support exceeds the preset minimum support threshold. S163: Generate association rules based on frequent itemsets, calculate the confidence level of each rule, and retain high-confidence rules whose confidence level exceeds the preset minimum confidence level threshold; S164: Associate high-confidence rules with the semantic topics of the corresponding work order clusters, identify the root cause feature combinations that lead to work order anomalies or frequent occurrences, and generate a root cause analysis report.

[0031] Optionally, in some embodiments, after generating the root cause analysis report, the following steps are also included: S165: Based on the combination of root cause characteristics, generate corresponding preventive treatment strategy recommendations; S166: Push preventative handling strategy recommendations to the management terminal and provide decision support for subsequent work order processing.

[0032] This embodiment's work order processing method, by generating preventative action strategies based on root cause analysis and pushing them to the management terminal, achieves a complete closed loop from problem diagnosis to decision execution. This solution constructs an intelligent decision support system: Step S165 enables automatic strategy generation: based on the combination of root cause characteristics, corresponding preventative action strategy recommendations are generated. It should be noted that this step is based on the root cause analysis results obtained from association rule mining, and automatically generates targeted action plans for different combinations of root cause characteristics, such as specific measures like equipment replacement, system upgrades, or process optimization.

[0033] Step S166 completes strategy push and application: Preventative handling strategy recommendations are pushed to the management terminal, providing decision support for subsequent work order processing. This step ensures that the analysis results can be promptly translated into practical action, directly pushing the handling strategies to the operation interface of relevant management personnel through the system interface, providing real-time guidance for work order processing.

[0034] This method establishes an automated process from root cause analysis to action implementation through the close integration of the two steps described above. The strategy generation step transforms data analysis results into actionable business instructions, ensuring that the action measures accurately correspond to the root causes; the strategy delivery step achieves seamless integration from analysis results to the execution level, ensuring that action recommendations are implemented in a timely manner. This end-to-end processing mechanism effectively solves the technical problem of the disconnect between analysis results and action actions in traditional work order management, improving the timeliness and effectiveness of problem handling.

[0035] By leveraging technological solutions, a complete intelligent closed loop for work order management has been achieved: by automatically converting root cause analysis results into preventative action strategies, the time delays and information loss associated with manually interpreting analysis reports and formulating measures, as required by traditional methods, are avoided. Simultaneously, by directly connecting to the system, action strategies are pushed to management terminals, ensuring that decision-making recommendations reach the relevant personnel immediately, significantly improving the timeliness of problem response and the accuracy of action measures. This fully automated mechanism, from problem discovery to root cause analysis to action implementation, significantly improves the efficiency and quality of work order processing, realizing a shift from a reactive to a proactive prevention-oriented operation and maintenance model.

[0036] In other embodiments of this application, an electronic device for work order processing is also proposed, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform steps in the method of any other embodiment.

[0037] In other embodiments of this application, an apparatus for processing work orders is also provided, comprising: The historical data acquisition unit is used to acquire historical work order data and clean the work order text in the historical work order data to obtain work order text information. The normalization unit is used to perform word segmentation and synonym normalization on work order text information; The work order vector matrix construction unit is used to convert the processed work order text information into a fixed-dimensional vector to construct the work order vector matrix. The work order cluster determination unit is used to perform cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs; Centroid calculation unit, used to calculate the centroid of each cluster; The outlier work order determination unit is used to determine the work order closest to the centroid from each cluster as a typical representative work order based on the centroid, and to determine the work orders that are more than a preset threshold away from the centroid as outlier work orders.

[0038] [Technical Effects]: Some implementations of the work order processing methods utilize unsupervised machine learning to automate and structure historical work orders, thereby identifying common and abnormal issues. This approach constructs a complete and logically rigorous technology chain: Steps S11-S13 complete the data preparation, which forms the foundation of the entire scheme. First, step S11 extracts and cleans valuable textual information from the raw data. Then, in step S12, the text is standardized through word segmentation and synonym normalization, transforming human language into standardized semantic units that can be processed by computers. Finally, in step S13, vectorization technology is used to map the textual information to a high-dimensional numerical space, transforming unstructured work orders into structured work order vector matrices, providing input for subsequent mathematical calculations.

[0039] Steps S14-S15 constitute the core analysis phase: this is the engine of the technical solution. The cluster analysis in step S14 acts as an "unsupervised classifier" at this stage. Without pre-labeling, it automatically discovers the inherent similarities between work orders, dividing them into different "problem groups" (clusters). The centroid calculation in step S15 is an abstract generalization of each group, finding its geometric center and providing a benchmark for the next step of measurement.

[0040] The output of step S16 is the final manifestation of the solution's value. Using distance metric, a mathematical tool, the abstract clustering results are transformed into concrete, actionable business insights: typical representative work orders (benchmark cases for training and analysis) and outlier work orders (abnormal issues requiring priority attention and handling).

[0041] These steps in this method replace the inefficient traditional model of relying on manual experience to read and classify work orders with technological means. It automatically and quickly extracts core knowledge (typical problems) and risk points (outliers) from massive amounts of work orders, directly achieving the goals of "helping managers quickly grasp the overall distribution and main contradictions of work orders" and "providing a basis for resource allocation," ultimately aiming to "improve the quality of work order responses." The essence of this technical solution is to transform a vague management problem into a clear, computable machine learning problem, and provide an end-to-end technical solution, thereby improving: work order processing efficiency and the uniformity of processing quality.

[0042] In summary, the work order processing method in this embodiment significantly improves the quality and accuracy of work order clustering analysis by introducing a dynamic weighting mechanism and adaptive distance metric. Based on the basic clustering results, the scheme first dynamically calculates a comprehensive weight coefficient according to business indicators such as work order processing time, semantic matching degree, and customer rating, thereby generating a weighted centroid vector. This step overcomes the sensitivity of simple arithmetic means to outliers, making the centroid more biased towards efficiently processed and high-quality work orders, thus more accurately representing the core features of each type of problem. Subsequently, the scheme constructs a Mahalanobis distance metric space based on the weighted covariance matrix. This space eliminates the correlation and dimensional differences between features, making the distance calculation more reflective of the true position of the work order in the statistical distribution. Finally, this metric space is used to recalibrate the judgment thresholds for typical representatives and outliers, achieving accurate identification of core and abnormal work orders. The above steps are interconnected: the weighted centroid provides a more robust benchmark for distance measurement, while the scientific distance measurement provides a theoretical basis for threshold calibration. Together, they solve the problem of inaccurate identification caused by data distribution skew or feature correlation in the basic scheme, and improve the robustness and practicality of the overall scheme.

[0043] In other embodiments, by introducing multivariate statistical methods to optimize the centroid calculation and distance metric of work order clusters, the distortion in distance calculation caused by the correlation and inconsistent dimensions of feature dimensions is effectively solved. The solution constructs a refined centroid calculation and spatial construction process: Step S154 establishes a scientific metric benchmark: First, a Mahalanobis distance metric space is constructed based on the weighted covariance matrix of each work order vector within the cluster. It should be noted that this calculation considers the business weight of each work order, accurately quantifying the dispersion of work order data within the cluster across various feature dimensions and their interrelationships through the weighted covariance matrix, thereby constructing a metric space capable of automatically adjusting the contribution of different feature dimensions.

[0044] Step S155 achieves precise centroid location: The centroid of the cluster's probability distribution is solved by minimizing the objective function of the weighted sum of squared Mahalanobis distances. From a statistical perspective, this point is the optimal center point of the probability distribution of the work order cluster, exhibiting stronger robustness than the arithmetic mean centroid in the presence of outliers.

[0045] Steps S156-S157 above complete the benchmark calibration: a geometrical coordinate system is generated based on the normalized distance metric between the centroid of the probability distribution and each vector within the cluster, and the outlier determination threshold is dynamically optimized based on the statistical distribution characteristics under this coordinate system. By analyzing the statistical distribution of the normalized distance metric, the outlier determination threshold is dynamically adjusted, making the determination mechanism more adaptive.

[0046] The above method, through the synergistic effect of the aforementioned steps, elevates centroid calculation and distance measurement from a purely geometric level to a statistical level. The weighted covariance matrix ensures the accuracy of the data distribution representation, the Mahalanobis distance metric space provides a scientific benchmark for distance calculation, and the dynamically optimized threshold enables refined outlier identification. Together, they solve the technical problems of centroid representation distortion and outlier misidentification caused by data distribution skewness, feature correlation, or inconsistencies in dimensions.

[0047] In another embodiment, by establishing a dual decision-making mechanism combining statistical inference and empirical analysis, the reliability of outlier detection is significantly improved. This scheme constructs a robust outlier determination system: Through steps S168-S169, a theoretical judgment criterion is established: a chi-square distribution probability model of the Mahalanobis distance of samples within a cluster is constructed and its cumulative distribution function is calculated. Based on the significance level, the right-hand critical value is determined, establishing a first outlier judgment criterion based on statistical inference. This criterion is based on the theoretical foundation of the multivariate normal distribution, providing a mathematical and statistical basis for outlier judgment.

[0048] Step S170 establishes an empirical judgment criterion: the interquartile range of the distance set is calculated in parallel, and a second criterion based on the empirical distribution is dynamically generated. This criterion can adapt to the actual shape of the data distribution and avoid the risk of misjudgment caused by the deviation between theoretical assumptions and reality.

[0049] Step S171 achieves decision fusion: the judgment results of the first criterion and the second criterion are fused and their significance is verified by hypothesis testing to generate the final outlier work order identifier. This collaborative mechanism, which combines theoretical models with empirical rules and is assisted by dual verification and statistical validation, ensures the robustness of the judgment results.

[0050] This method constructs a decision system with dual verification. The first criterion provides theoretical rigor, the second criterion ensures practical adaptability, and the final fusion and verification steps ensure high confidence in the decision results. It effectively overcomes the limitations that may exist in a single judgment model and solves the technical problems of missed and misjudged outliers caused by invalid model assumptions or unreasonable threshold settings.

[0051] In summary, the two embodiments described above jointly construct an outlier detection and verification system based on rigorous statistical theory, greatly enhancing the reliability and automation level of work order analysis results. In the basic scheme, by calculating the chi-square distribution probability model of the Mahalanobis distance of samples within a cluster, a first criterion for outlier determination based on significance testing is established. This criterion defines "low-probability events" as outliers from a statistical theory perspective, providing a solid mathematical foundation for the determination process. In the improved scheme of the preferred embodiment, instead of relying on a single theoretical model in isolation, the interquartile range of the distance set is calculated in parallel, generating a second criterion based on empirical distribution. This effectively avoids the risk of misjudgment that may arise when theoretical assumptions do not perfectly match the actual data distribution. Finally, by fusing the determination results of the two criteria and verifying them through hypothesis testing, the final outlier work order identifier is generated. This collaborative mechanism of "combining theoretical models with empirical rules, and assisting with dual verification and statistical validation" completely solves the pain points of the basic scheme, such as the strong subjectivity and poor universality of preset thresholds, achieving automation, objectivity, and high accuracy in outlier detection.

[0052] In another embodiment of the work order processing method, a fundamental innovation in the work order clustering analysis paradigm is achieved by introducing differential geometry and topological data analysis theory. The scheme constructs a clustering analysis framework based on manifold learning: Step S141 achieves data space transformation: the work order vector matrix is ​​mapped to a high-dimensional probability distribution manifold based on the kernel density estimation method. This transforms the discrete point cloud data analysis problem into the analysis problem of a continuous geometric object, providing a richer mathematical foundation for subsequent analysis.

[0053] Steps S142-S143 complete mode discovery and cluster partitioning: Gradient ascent optimization is performed on the probability distribution manifold using the mean-shift algorithm, iteratively converging to a local density maxima to determine cluster modes; an attraction domain is constructed centered on each mode, and all work orders converging to the same mode are assigned to the same cluster. This method can automatically discover clusters of arbitrary shapes without pre-setting the number of clusters.

[0054] Step S144 performs topological outlier detection: Based on the persistent homology theory, the topological persistent graph of the probability distribution manifold is calculated, and transient connected components with lifetimes below a threshold are identified and their corresponding work orders are marked as geometric outliers. Work orders that may not have low local density but are abnormal in the overall topological structure are discovered at the topological level.

[0055] This method transforms point clouds into continuous manifolds and analyzes their geometric gradients and topological persistence, achieving a unified framework for clustering and outlier detection. Kernel density estimation provides a continuous analytical foundation, the mean-shift algorithm enables precise modality localization and the discovery of clusters of arbitrary shapes, while persistent homology theory reveals deep anomalous structures at the topological level, solving the technical problems of traditional clustering algorithms being sensitive to preset parameters and unable to identify complex manifold structures and topological outliers.

[0056] In another embodiment of the work order processing method, the accuracy and theoretical depth of cluster analysis are further improved by transforming the mean drift process into a precise solution problem of a differential dynamical system. The scheme constructs a trajectory tracking system based on dynamical system theory: Step S1421a constructs the dynamical system model: a gradient vector field is constructed on the probability distribution manifold based on Riemannian metrics, and the Lie derivative of this vector field along the integral curve is calculated. This step provides a key mathematical tool for analyzing the stability of the dynamical system and lays the foundation for scientifically determining convergence.

[0057] Step S1422a achieves precise trajectory tracking: The variable-step Runge-Kutta method is used to numerically solve the dynamic system, accurately tracking the evolution trajectory of the work order vector on the manifold. This method can adaptively adjust the integration step size according to the curvature of the solution curve, improving computational efficiency while ensuring accuracy.

[0058] Steps S1423a-S1424a complete convergence determination and cluster partitioning: when the norm of the Lie derivative is lower than the preset convergence threshold, the trajectory is determined to have converged to the steady-state attractor; homotopy classification is performed on all trajectories that converge to the same attractor, and a cluster equivalence relation based on differential homeomorphism is established. This provides the deepest mathematical explanation for cluster partitioning from the perspective of dynamical system stability.

[0059] This method elevates the clustering process from a heuristic iterative algorithm to a rigorously mathematically analyzeable and controllable dynamical system problem by introducing the theory of differential dynamical systems and high-precision numerical calculation methods. Riemannian metrics and gradient vector fields provide accurate geometric constraints for trajectory evolution, the variable-step-size Runge-Kutta method guarantees the numerical accuracy of trajectory solutions, the convergence criterion based on Lie derivatives provides rigorous mathematical meaning for modality determination, and homotopy classification ensures the rationality of cluster partitioning from a topological perspective. This systematically solves the technical problems of traditional mean-shift algorithms, such as convergence determination relying on experience, insufficient trajectory tracking accuracy, and a lack of rigorous mathematical interpretation.

[0060] In summary, the above embodiments represent a profound transformation in the essence of cluster analysis: introducing computational geometry and topology into the field of work order analysis, breaking through the inherent limitations of traditional clustering algorithms. Firstly, in the optional scheme, the discrete work order vector matrix is ​​reconstructed into a continuous probability distribution manifold through kernel density estimation, transforming the clustering problem into finding modal points with density peaks on this manifold. Gradient ascent optimization is performed using the mean-shift algorithm, automatically locating modal points and assigning clusters based on their attraction domains. This method can naturally identify clusters of arbitrary shapes without pre-setting the number of clusters. In the preferred scheme, the mean-shift process is further interpreted as the evolution problem of a differential dynamical system, using the Runge-Kutta method to precisely solve the trajectory, and employing the Lie derivative... In this embodiment, the technical solution achieves in-depth diagnosis from problem identification to root cause analysis by performing association rule mining on key outlier work order sets or emerging problem pattern work order clusters. This solution constructs a root cause analysis system based on data mining technology: In another embodiment, step S161 completes data preprocessing: for key outlier work orders or work orders within work order clusters, their structured metadata feature values ​​are extracted to form transaction itemsets. It should be noted that this step converts the metadata features of each work order into a standard transaction format suitable for association rule mining, providing standardized data input for subsequent analysis.

[0061] Step S162 implements frequent pattern discovery: Frequent itemset mining is performed on the transaction itemsets, calculating the support of each itemset, and filtering out frequent itemsets whose support exceeds a preset minimum support threshold. This step uses algorithms such as FP-Growth to identify frequently occurring feature combination patterns in problem work orders and filter out irrelevant combinations that occur only occasionally.

[0062] Step S163 completes the association rule generation: association rules are generated based on frequent itemsets, and the confidence score of each rule is calculated. High-confidence rules with confidence scores exceeding a preset minimum confidence threshold are retained. The confidence score measures the credibility of the rules to ensure that the discovered associations are statistically significant.

[0063] Step S164 achieves root cause localization: High-confidence rules are mapped to the semantic themes of the corresponding work order clusters to identify the root cause feature combinations leading to work order anomalies or frequent occurrences, and a root cause analysis report is generated. This step combines data mining results with semantic analysis to form a multi-dimensional root cause judgment, improving the accuracy and interpretability of the analysis conclusions.

[0064] In this way, the method establishes a complete analytical chain from data preparation and pattern discovery to root cause localization through the organic combination of the above steps. Transaction itemset construction ensures the standardization of data format, frequent itemset mining discovers potential feature combination patterns, association rule generation filters out reliable causal relationships, and the final association mapping realizes the transformation from data patterns to business insights. This progressive analytical approach effectively solves the technical problem of traditional work order analysis's difficulty in discovering deep-seated, multi-dimensional root causes, providing a scientific basis for problem prevention and precise policy implementation.

[0065] The technical effectiveness of the solution lies in its deep integration of unsupervised clustering and association rule mining techniques. By mining the hidden feature associations in outlier work orders and emerging problem clusters, it can discover complex root cause combinations that are difficult to spot intuitively, such as the failure modes of specific equipment models under specific conditions. This deep analysis capability significantly improves the intelligence level of work order management, shifting the handling strategy from passive response to proactive prevention, ultimately reducing the problem recurrence rate. Attached Figure Description

[0066] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a work order processing and cluster analysis method in an embodiment of this application; Figure 2 This is a flowchart of the work order clustering trajectory tracking and convergence determination based on differential dynamical systems in a work order processing method according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a work order processing device according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a work order processing device in an embodiment of this application. Detailed Implementation

[0067] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0068] like Figure 1 As shown, one embodiment of this application relates to a specific application scenario for improving the quality and efficiency of government hotline work order processing. In this scenario, the method clusters all work orders to identify their core issues and outliers, thereby discovering the most prominent types of problems. This helps managers quickly grasp the overall distribution and main contradictions of work orders, providing a basis for resource allocation and prioritization, thus improving the quality of work order responses. The core steps of the method include: Data collection and preprocessing: Historical work order data is exported from the government hotline system, and key fields of the work orders (such as work order title, problem description, processing record, etc.) are extracted. The work order text is then cleaned (irrelevant characters, HTML tags, stop words, etc.) to obtain the work order text information. Subsequently, the work order text information is processed by word segmentation and synonym normalization.

[0069] Work order vectorization: Convert the processed work order text information into a fixed-dimensional vector and construct a work order vector matrix.

[0070] Cluster analysis: Perform cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs, calculate the centroid of each cluster, and identify typical representative work orders and outliers based on the centroids.

[0071] In this scenario, the municipal service hotline center receives a large number of complaints, inquiries, and requests for assistance from citizens daily. These work order documents are presented with diverse descriptions and colloquial language, covering a wide range of departments. Inefficiency and error-proneness stem from the reliance on manual experience in classification and dispatching. The method described in this application utilizes in-depth analysis of historical work orders to achieve automation, identifying high-frequency problem types and unusual cases, thereby optimizing resource allocation and improving the accuracy and efficiency of problem handling.

[0072] This process Figure 1 The technical solution demonstrates a sequential data processing and analysis workflow. The output of each step is used as the input for the next step, ultimately generating valuable business insights from the raw data.

[0073] First, in step S11, historical work order data is obtained, and the work order text in the historical work order data is cleaned to obtain work order text information.

[0074] Specifically, recent work order data was exported from the government hotline system, including fields such as work order title, detailed description, acceptance time, and processing department. The implementation steps are detailed below: Relevant characters and noise removal are achieved through text cleaning of the work orders. This includes filtering HTML tags, special symbols, emoticons, and irrelevant line breaks and spaces; converting full-width characters to half-width characters; and performing standardized text encoding, such as using UTF-8 to avoid garbled characters. Furthermore, invalid work orders with blank content or containing only meaningless characters need to be identified and removed. The standardized plain text information obtained after these cleaning operations lays the foundation for subsequent deep text processing.

[0075] Subsequently, in step S12, the work order text information is processed by word segmentation and synonym normalization.

[0076] The detailed implementation steps are as follows: The cleaned work order text is processed by a word segmentation algorithm that combines dictionary and statistical models. Entities and professional terms specific to the government domain are identified by the word segmentation algorithm, such as "noise disturbance," "street vending," and "medical insurance reimbursement." After the initial word segmentation sequence is generated, the terms in the sequence undergo entity recognition and semantic verification through a pre-built government domain knowledge base, ensuring the accuracy of the word segmentation results.

[0077] Furthermore, unified standard concept nodes are mapped to semantically similar or identical words through a pre-constructed thesaurus. For example, expressions such as "streetlight not working," "lighting malfunction," and "dark lights" are uniformly normalized to "streetlight malfunction"; "sewer blockage," "sewage overflow," and "poor drainage" are uniformly normalized to "drainage blockage." The ambiguity caused by the diversity of natural language expressions is effectively eliminated through normalization, and standardized business terminology is elevated from colloquial descriptions.

[0078] In step S13, the processed work order text information is converted into a fixed-dimensional vector to construct a work order vector matrix.

[0079] The detailed implementation steps are as follows: After word segmentation and normalization, the work order text undergoes deep semantic encoding through a pre-trained language model, generating a high-dimensional semantic feature vector. This vector captures the deep semantic information of the work order text. Simultaneously, structured feature data of the work order is extracted, such as its urgency, historical processing time, and the number of times a citizen has submitted the work order repeatedly. Standardization and normalization of these structured features are then performed, and numerical feature vectors are obtained. The high-dimensional semantic feature vector and the numerical feature vector are concatenated to form a multimodal fusion vector.

[0080] Then, in step S14, cluster analysis is performed on the work order vector matrix to determine the cluster to which each work order belongs.

[0081] Specifically, a density peak-based clustering algorithm is used to analyze the work order vector matrix. This algorithm automatically determines the number of clusters without pre-setting and can discover clusters of arbitrary shapes. By calculating the local density and distance of each data point, density peak points are identified as cluster centers. Each work order vector is then assigned to the cluster represented by its nearest neighbor cluster center, thus completing the automatic classification of work orders.

[0082] For example, the analysis results may show that work orders are divided into several major categories such as "urban appearance and environment", "road traffic", "social security" and "public safety". Each major category may be further subdivided into several subcategories, such as "garbage collection", "illegal construction" and "greening maintenance" under the "urban appearance and environment" category. This step realizes the automated and refined classification of massive work orders.

[0083] Further, in step S15, the centroid of each cluster is calculated.

[0084] For each work order cluster obtained from cluster analysis, its centroid is calculated to represent the core features of the cluster. The centroid calculation is not a simple arithmetic average, but a weighted average. A comprehensive weight is assigned to each work order vector within the cluster. This weight is dynamically calculated based on business indicators such as processing time, semantic matching degree, and final citizen rating. Work orders with shorter processing times, higher semantic matching degrees, and higher citizen ratings have greater weights. A weighted average is then calculated for all work order vectors within the cluster based on these weights to obtain a weighted centroid vector. Geometrically, the weighted centroid vector represents the "center" of the cluster; business-wise, it tends to represent typical work orders that are processed efficiently and with high quality, thus making it more representative.

[0085] In step S16, based on the centroid, the work order closest to the centroid in each cluster is identified as a typical representative work order, and work orders that are more than a preset threshold away from the centroid are identified as outlier work orders.

[0086] Specifically, firstly, based on the weighted centroid vectors obtained above, the Mahalanobis distance from each work order vector within the cluster to the centroid is calculated. The Mahalanobis distance takes into account the correlation between work order features and reflects the true distribution of work orders in the feature space better than the Euclidean distance. The top K work orders with the smallest Mahalanobis distances are selected as a set of candidate representative work orders.

[0087] Subsequently, the work orders in the candidate set are sorted based on their overall weight, and the work order with the highest overall weight is selected as the final representative work order of the cluster. This work order represents the optimal processing example for this type of problem.

[0088] In parallel, based on the distance distribution of all work orders within a cluster to the centroid, the mean and standard deviation of the distances are calculated, and a preset threshold is dynamically set to the mean distance plus N times the standard deviation. Any work order whose distance to the centroid exceeds the dynamic threshold is marked as a critical outlier. These outliers may be complex or emerging issues that require priority handling by an expert team.

[0089] In summary, through the above process, the government hotline management department can obtain a clear analysis report: the report shows the overall classification of current work orders; typical representative work orders for each type of problem, which can be used as excellent cases for training customer service and handling personnel; and a list of all outlier work orders that require close attention. Based on this, managers can allocate resources more scientifically, such as quickly promoting the handling experience of typical work orders, or prioritizing the allocation of outlier work orders to experienced expert teams, thereby systematically improving the efficiency, accuracy, and consistency of work order processing.

[0090] The above is a first embodiment of a work order processing method provided by the present application. The following is a second embodiment of a work order processing method provided by the present application.

[0091] It is understandable that high-dimensional, highly correlated work order features are often susceptible to interference from factors such as skewed data distribution and inconsistent dimensions when processed by traditional clustering methods, leading to centroid distortion and misjudgment of outliers. This embodiment effectively overcomes these technical problems by constructing a sophisticated computational process.

[0092] Specifically, this embodiment includes the following steps: Step S141: Map the work order vector matrix to a high-dimensional probability distribution manifold based on the kernel density estimation method.

[0093] It should be noted that the transformation from discrete work order vector data points to continuous geometric objects is analyzed in this step. The work order vector matrix is ​​subjected to kernel density estimation using a multivariate Gaussian kernel function, and the probability density value at each location is calculated, thus constructing a continuous probability distribution manifold. The clustering and sparsity of work orders in the feature space are intuitively reflected by the elevation of the "terrain" on the manifold. For example, for the mixed text and numerical features in government work orders, the bandwidth matrix of the Gaussian kernel function can be dynamically adjusted to optimize the smoothness and resolution of the density estimation.

[0094] Step S142: Perform gradient ascent optimization on the manifold using the mean drift algorithm, iteratively converging to a local density maximum point to determine the cluster mode.

[0095] Specifically, after the probability distribution manifold is constructed, gradient ascent optimization is performed on it using the mean-shift algorithm. Starting from a randomly initialized seed point, the algorithm iteratively moves along the probability density gradient until it converges to a local density maximum (i.e., a cluster mode). Understandably, the method can automatically discover clusters of arbitrary shapes without pre-setting the number of clusters. For example, to handle plateau regions on the manifold, a momentum term can be introduced to accelerate convergence and suppress oscillations.

[0096] Step S143: Construct an attraction domain centered on each mode, and assign all work orders that converge to the same mode to the same cluster.

[0097] After the modal points of each cluster are determined, each mode is used as the center to calculate its geodesic distance matrix to all other points on the manifold, and the entire manifold space is divided into several attraction domains using the Voronoi diagram partitioning algorithm. Work points belonging to the same cluster are considered to be work points within each attraction domain. For work points falling on the boundaries of multiple attraction domains, soft assignment can be performed based on the ratio of their geodesic distances to each modal point, and the membership weights are recorded.

[0098] Step S144: Calculate the topological persistent graph of the probability distribution manifold based on the persistent homology theory, identify transient connected components with lifetimes below the threshold, and mark their corresponding work orders as geometric outliers.

[0099] It should be noted that work orders that may have high local density but are abnormal in the overall topology are identified at the topology level in this step. By constructing a hierarchical filtering structure of a probability distribution manifold, the connected components of the level set at each density threshold are calculated, and a persistent topology graph is plotted. Transient connected components with a lifetime shorter than a preset threshold (e.g., average processing time) are identified, and their corresponding work order points are marked as geometric outliers. These outliers may correspond to emerging problem patterns or deep anomalies and should be prioritized.

[0100] Step S154: Construct the Mahalanobis distance metric space based on the weighted covariance matrix of each work order vector within the cluster.

[0101] After cluster partitioning, to accurately calculate centroids and measure distances, the Mahalanobis distance metric space is first constructed based on the weighted covariance matrix of each work order vector within the cluster. It should be noted that the business weight of each work order (such as processing time, semantic matching degree, and customer rating) is considered in the calculation of the weighted covariance matrix, thereby more accurately quantifying the dispersion and correlation of feature dimensions. The contribution of different feature dimensions can be automatically adjusted by the metric space to eliminate the influence of units and correlations.

[0102] Step S155: Solve for the centroid of the probability distribution of the cluster by minimizing the objective function of the weighted sum of squared Mahalanobis distances.

[0103] Based on the aforementioned metric space, the centroid of the probability distribution of the cluster is solved by minimizing the objective function of the weighted sum of squared Mahalanobis distances. Statistically, the centroid is the optimal center point (mathematical expectation) of the probability distribution of the work order cluster. Compared with the arithmetic mean centroid, it has stronger robustness to outliers, and the core characteristics of the cluster can be more accurately represented.

[0104] Step S156: Generate a geometric space reference coordinate system for representing work order screening based on the normalized distance metric between the centroid of the probability distribution and each vector within the cluster.

[0105] After obtaining the robust centroid, the Mahalanobis distance from each work order vector within the cluster to the centroid is calculated, and this set of distances is normalized (e.g., mapped to the [0,1] interval). The normalized distance metric system constitutes a new, standardized geometric space reference coordinate system, and a unified scale benchmark is provided for subsequent selection of typical representative work orders and outlier identification.

[0106] Step S157: Based on the statistical distribution characteristics under the reference coordinate system, dynamically optimize the significance level threshold for outlier determination.

[0107] Understandably, the density of data distribution varies across different work order clusters, making outlier identification with a fixed threshold inapplicable. This step analyzes the statistical distribution of normalized distance metrics (such as mean and standard deviation) and dynamically adjusts the outlier identification threshold based on statistical principles (such as the three sigma rule) or preset confidence levels. For example, a stricter threshold is set for clusters with concentrated distributions, while the threshold is relaxed for clusters with dispersed distributions, thereby improving the adaptability and accuracy of outlier identification.

[0108] Step S168: Construct a chi-square distribution probability model of the Mahalanobis distance of samples within the cluster, and calculate its cumulative distribution function.

[0109] To establish a theory-driven outlier determination criterion, a chi-square distribution probability model of the Mahalanobis distance between samples within a cluster is first constructed. According to the theory of multivariate normal distribution, the chi-square distribution is governed by the square of the Mahalanobis distance, and its degrees of freedom are equal to the feature dimension. Based on this theory, the cumulative distribution function of the squared distance values ​​is calculated, describing the probability of a specific distance value occurring under the assumption that the data follows a normal distribution.

[0110] Step S169: Determine the right-hand critical value θ_α based on the significance level α, and establish the first criterion for outlier determination based on statistical inference.

[0111] A relatively small significance level α (e.g., 0.05) is chosen, and the corresponding right-hand critical value θ_α is determined according to the chi-square distribution table or its inverse cumulative distribution function. If the squared Mahalanobis distance of a work order is greater than θ_α, it is considered a low-probability event, highly unlikely to occur under the null hypothesis, and is thus identified as an outlier. This criterion provides a solid mathematical statistical foundation for outlier identification.

[0112] Step S170: In parallel, calculate the interquartile range of the distance set and dynamically generate a second outlier determination criterion based on the empirical distribution.

[0113] To compensate for the inadequacy of theoretical models that may fail due to deviations from ideal assumptions in actual distributions, a nonparametric empirical method is employed in parallel in this step. Specifically, the interquartile range (IQR) of all Mahalanobis distance values ​​is calculated, and a decision threshold is dynamically generated based on this (e.g., upper bound = Q³ + k × IQR, where k is typically 1.5 or 3). Any work order with a distance value exceeding this upper bound is classified as an outlier. The distribution assumptions are not dependent on this criterion, exhibiting strong practical adaptability.

[0114] Step S171: Integrate the judgment results of the first criterion and the second criterion to generate the final outlier work order identifier and verify its significance through hypothesis testing.

[0115] A decision fusion mechanism is designed in this step. For example, it requires that a work order must be identified as an outlier by both criteria simultaneously before it is finally marked as an outlier work order, in order to reduce false alarms. Furthermore, for the marked outlier work orders, hypothesis testing methods such as the Mann-Whitney U test can be used to verify the distribution differences of their key features with the main cluster work orders, providing additional statistical support for the judgment results.

[0116] Preferably, for key outlier sets of work orders or work order clusters containing emerging problem patterns, the above embodiments further include a systematic root cause analysis process, the purpose of which is to: mine deep-seated causal patterns from the data and provide a basis for proactive intervention decisions. This process can be specifically implemented through the following steps: Step S161: Extract its structured metadata feature values ​​to form a transaction item set.

[0117] This step is the data preparation stage for association rule mining.

[0118] Specifically, for all work orders within a cluster of identified key outliers or specific problem patterns, the system extracts relevant metadata features from their structured information. These features may include, but are not limited to: work order source channel, problem location, device model, timestamp, associated user identifier, and initial problem classification code. All feature values ​​for each work order are combined into a set, called a "transaction." The set of transactions for all analyzed work orders constitutes a "transaction itemset" for subsequent mining. This step transforms the unstructured work order classification problem into a standardized data format suitable for frequent pattern mining.

[0119] Step S162: Perform frequent itemset mining and filter itemsets whose support exceeds a preset threshold.

[0120] After constructing the transaction itemsets, this step helps discover frequently occurring feature combinations: Frequent itemset mining algorithms such as FP-Growth or Apriori are used to calculate the frequency of all possible feature combinations (i.e., "itemsets") appearing in the entire transaction itemset; this frequency is called "support." The system presets a minimum support threshold and filters out all itemsets with support higher than this threshold, called "frequent itemsets." For example, if the minimum support is set to 5%, then an itemset such as {"device model = A", "problem category = communication interruption"} must appear simultaneously in more than 5% of outlier work orders to be retained. This step filters out random feature combinations and focuses on prevalent patterns.

[0121] Step S163: Generate association rules and calculate confidence scores, retaining rules with high confidence scores.

[0122] Here, based on the frequent itemsets obtained in the previous step, this step further generates "association rules" with causal implications. A rule typically takes the form of "antecedent → consequent," such as "device model = A → problem classification = communication interruption." For each possible rule, its "confidence" is calculated, which is the conditional probability that the consequent feature will also occur given that the antecedent feature has occurred.

[0123] Preferably, the system presets a minimum confidence threshold, retaining only rules with a confidence level exceeding this threshold; these are called "high-confidence rules." Higher confidence indicates a more reliable causal relationship described by the rule. This step is crucial for extracting strong associations from frequent co-occurrence patterns.

[0124] Step S164: Associate and map the rules with the semantic topics of the work order cluster, identify the root cause feature combination, and generate a root cause analysis report.

[0125] The purpose of this step is to interpret and integrate the high-confidence rules discovered at the business level.

[0126] The software system executing the work order processing method maps the consequent or antecedent features of each rule to the semantic topics (such as "device hardware failure," "network configuration error," etc.) derived from text analysis of the corresponding work order cluster. Through this cross-validation and semantic fusion, the system identifies the root cause feature combinations that lead to the anomalies or frequent occurrences of this type of work order. For example, it might be found that the combination of "device model = A" and "firmware version = V1.0" is the root cause of the "communication interruption" problem.

[0127] Next, a structured root cause analysis report is automatically generated by the system, clearly listing key information such as the identified root cause combinations, the corresponding association rules and their confidence levels, and the scope of affected work orders.

[0128] Step S165: Based on the combination of root cause characteristics, generate corresponding preventive treatment strategy recommendations.

[0129] After identifying the root cause, the transformation from analysis to action is achieved through this step.

[0130] A strategy knowledge base is pre-configured within the system or dynamically generated through algorithms. Different root cause characteristics are mapped to specific action recommendations through the strategy knowledge base. For example, for the root cause "Device Model = A & Firmware Version = V1.0", preventative action strategy recommendations such as "Initiate firmware upgrade program for all users using device model A with firmware version V1.0" and "Suspend the entry of this model of device into the inventory during the procurement process" are automatically generated by the system. The decision support capability of the solution is reflected in this step.

[0131] Step S166: Push the preventive action strategy recommendations to the management terminal to provide decision support for subsequent work order processing.

[0132] Finally, the implementation of the analysis results is ensured by step S166: the generated preventative action strategy recommendations are automatically pushed to the work terminals or decision support system interfaces of relevant managers through the system interface. This timely guidance not only provides immediate guidance for the processing of current work orders, but more importantly, it enables proactive preventative measures to be taken by management based on data insights. This reduces the recurrence of similar problems at the source, achieving a closed loop in work order management from passive response to proactive optimization.

[0133] Further, refer to Figure 2 As shown, in the mean-shift clustering process, in order to improve its theoretical rigor and computational accuracy, a set of optional refinement steps is provided by this embodiment. The iterative process is incorporated into the framework of differential dynamical systems for precise analysis and control.

[0134] Step S1421: Construct a gradient vector field on the probability distribution manifold based on the Riemannian metric, and calculate the Lie derivative of the vector field along the integral curve.

[0135] A precise dynamic model is established through this step. First, based on the probability distribution manifold determined by the kernel density estimate, its intrinsic Riemannian metric tensor field is utilized, and the gradient vector field characterizing the direction of the steepest ascent of the probability density is calculated. Subsequently, to further analyze the evolution characteristics of the vector field along its own integral curve (i.e., the mean drift trajectory), the "Lie derivative" of the vector field along the integral curve needs to be calculated. The Lie derivative is a key concept in differential geometry; it quantifies the rate of change of the vector field along its own direction, and its calculation involves the covariant derivative of the metric tensor and the Lie bracket operations of the vector field components.

[0136] Step S1422: The dynamic system is numerically solved using the variable step size Runge-Kutta method to accurately track the evolution trajectory of the work order vector on the manifold.

[0137] After establishing the dynamic system model, the motion trajectory of the work vector is simulated with high precision through this step: the gradient vector field is expressed as a system of first-order ordinary differential equations, and a high-order method in numerical analysis—the variable-step Runge-Kutta method—is used for solving it. The method (such as the fourth-order Runge-Kutta method) possesses minimal local truncation error, providing a high-precision numerical solution. The rate of change of the trajectory curvature or Lie derivative norm can be dynamically adjusted according to its "variable-step" characteristic: the step size is automatically reduced in regions of rapid change to ensure stability, and increased in regions of gentle change to improve efficiency. Accurate and adaptive tracking of the work vector's evolving trajectory on the manifold is thus achieved.

[0138] Step S1423: When the norm of the Lie derivative is lower than the preset convergence threshold, it is determined that the trajectory has converged to the steady-state attractor.

[0139] The convergence criterion is provided via step S1423: a convergence threshold ε based on machine precision is set. At each step of the numerical integration, the norm of the current Lie derivative is calculated in real time. The direction and intensity of the vector field along the trajectory have almost stopped changing, implying that the Lie derivative norm approaches zero, and the system tends to stabilize.

[0140] Therefore, when the norm of the Lie derivative is detected to be below the preset threshold ε for several consecutive iterations, it can be determined that the trajectory has converged to a local extremum, i.e., a "steady-state attractor". Compared with the traditional empirical judgment method that relies on position changes, this criterion based on the stability theory of dynamical systems is more mathematically rigorous.

[0141] Step S1424: Perform homotopy classification on all trajectories converging to the same attractor and establish the equivalence relation of work-line clusters based on differential homeomorphism.

[0142] Here, after all trajectories converge, the cluster partitioning is refined and distilled from a topological perspective through step S1424. All work order vector trajectories converging to the same steady-state attractor are collected, and their parameterized curve representations are obtained. Subsequently, the "homotopy theory" from algebraic topology is applied, and the continuous deformation relationship between these curves is analyzed. Two trajectories are considered "homotopically equivalent" when they can be obtained by continuously deforming each other without departing from the manifold. Based on this equivalence relationship, all trajectories are partitioned into different homotopy classes, each corresponding to a "potentially attractive basin" in a dynamic sense. Finally, all work orders belonging to the same potential attractive basin are assigned to the same final cluster, and the work order cluster equivalence relationship based on "differential homeomorphism" (a smooth and invertible transformation) is thus established. The consistency of members within the cluster is ensured from the deep structure of the topology; the cluster partitioning result is based not only on the convergence endpoint but also on the topological properties of the convergence path.

[0143] The above is a second embodiment of a work order processing method provided in this application. The following is a block diagram of a work order processing device provided in this application. Please refer to [link / reference]. Figure 4 As shown, the device for work order processing in this embodiment is based on unsupervised machine learning technology. Through the collaborative work of multiple functional units, it achieves automated clustering analysis and outlier identification of historical work order data, thereby improving the efficiency and quality uniformity of work order processing. The device is particularly suitable for scenarios such as government hotlines and IT operations and maintenance where work order texts are diverse and semantically complex.

[0144] Specifically, the device includes the following units: The historical data acquisition unit 11 is used to respond to the work order analysis request, acquire the historical work order data to be analyzed, and clean the work order text in the historical work order data to obtain the work order text information.

[0145] Normalization unit 12 is used to perform word segmentation and synonym normalization on work order text information.

[0146] The work order vector matrix construction unit 13 is used to convert the processed work order text information into a fixed-dimensional vector to construct the work order vector matrix.

[0147] The work order cluster determination unit 14 is used to perform cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs.

[0148] Centroid calculation unit 15 is used to calculate the centroid of each cluster.

[0149] The outlier work order determination unit 16 is used to determine, based on the centroid, the work order closest to the centroid in each cluster as a typical representative work order, and to determine the work orders that are more than a preset threshold away from the centroid as outlier work orders.

[0150] Furthermore, the historical data acquisition unit 11 cleans the work order text in the historical work order data, specifically including: Export historical work order data within a specified time period from the work order management system. The data includes fields such as work order title, detailed description, acceptance time, processing record, and completing department. Filter HTML tags, special characters, emojis, and irrelevant line breaks and spaces; convert all full-width characters to half-width characters. The text is standardized and encoded; and invalid work orders with blank content or containing only meaningless characters are identified and removed to obtain standardized work order text information.

[0151] Furthermore, the normalization unit 12 performs word segmentation and synonym normalization processing on the work order text information, specifically including: The work order text information is segmented into words to generate an initial segmentation sequence; Entity recognition and semantic verification of terms in the initial word segmentation sequence are performed based on a pre-built domain knowledge base; Based on the pre-constructed thesaurus, semantically validated terms are mapped to unified standard concept nodes to complete the thesaurus normalization process, thereby eliminating ambiguity caused by the diversity of natural language expressions.

[0152] Furthermore, the work order vector matrix construction unit 13 converts the processed work order text information into a fixed-dimensional vector to construct a work order vector matrix, specifically including: A pre-trained language model is used to perform deep semantic encoding on the work order text information after word segmentation and synonym normalization to generate a high-dimensional semantic vector. The structured feature data of the work order is standardized to generate a numerical feature vector; The high-dimensional semantic vector and the numerical feature vector are concatenated and their dimensions reduced to generate the final work order vector used to construct the work order vector matrix.

[0153] Furthermore, the work order cluster determination unit 14 performs cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs, specifically including: A density peak-based clustering algorithm is used to perform cluster analysis on the work order vector matrix. By calculating the local density and distance of each data point, the density peak point is identified as the cluster center, and each work order vector is assigned to the cluster represented by its nearest neighbor cluster center to determine the cluster to which each work order belongs.

[0154] Furthermore, the centroid calculation unit 15 calculates the centroid of each cluster, specifically including: Based on a predefined comprehensive weight coefficient, a weighted average is calculated on the work order vectors within the cluster to generate a weighted centroid vector; where the comprehensive weight coefficient is dynamically calculated from the processing time, semantic matching degree, and customer rating of the corresponding work order.

[0155] Furthermore, the outlier work order determination unit 16, based on the centroid, identifies the work order closest to the centroid from each cluster as a typical representative work order, and determines work orders whose distance from the centroid exceeds a preset threshold as outlier work orders, specifically including: Calculate the Mahalanobis distance from each work vector within the cluster to the centroid; The top K work orders with the smallest distance from Maharanobis were selected as a set of candidate typical representative work orders. The work orders in the candidate typical representative work order set were sorted according to their comprehensive weight, and the work order with the highest comprehensive weight was selected as the final typical representative work order. Based on the distance distribution of all work orders within the cluster to the centroid, the mean and standard deviation of the distance are calculated. A preset threshold is dynamically set as the mean distance plus N times the standard deviation, and work orders whose distance exceeds the dynamically set preset threshold are marked as critical outlier work orders.

[0156] Furthermore, the work order cluster determination unit 14 performs cluster analysis on the work order vector matrix, which also includes: mapping the work order vector matrix to a high-dimensional probability distribution manifold based on the kernel density estimation method; performing gradient ascent optimization on the manifold through the mean drift algorithm, iteratively converging to a local density maximum point to determine the cluster mode; constructing an attraction domain centered on each mode, and assigning all work orders that converge to the same mode to the same cluster; calculating the topological persistent graph of the probability distribution manifold based on the persistent homology theory, identifying transient connected components with lifetimes below a threshold and marking their corresponding work orders as geometric outliers.

[0157] Furthermore, the centroid calculation unit 15 calculates the centroid of each cluster, and also includes: Based on the weighted covariance matrix of each work order vector within the cluster, a Mahalanobis distance metric space is constructed. The centroid of the probability distribution of the cluster is solved by minimizing the objective function of the weighted sum of squared Mahalanobis distances. Based on the normalized distance metric between the centroid of the probability distribution and each vector within the cluster, a geometric space reference coordinate system is generated to represent the work order screening. Based on the statistical distribution characteristics under this reference coordinate system, the significance level threshold for outlier determination is dynamically optimized.

[0158] Furthermore, the outlier work order determination unit 16, which determines outlier work orders based on the centroid, also includes: Construct a chi-square distribution probability model of the Mahalanobis distance between samples within a cluster, and calculate its cumulative distribution function; The right-hand critical value θ_α is determined based on the significance level α, and a first criterion for outlier determination based on statistical inference is established. The interquartile range of the distance set is calculated in parallel and a second outlier determination criterion based on empirical distribution is dynamically generated. By integrating the judgment results of the first and second criteria, the final outlier work order identifier is generated and its significance is verified by hypothesis testing.

[0159] Furthermore, the device also includes a resource optimization and allocation unit for optimizing resource allocation based on the distribution of work order clusters, specifically: Calculate the required processing resource weights for each cluster based on the size of each work order cluster and the complexity of the problems reflected by typical representative work orders. Based on the weight of processing resources, the matching relationship between support personnel and work order clusters is dynamically adjusted; Outlier sets are assigned to high-level expert queues for specialized processing, and resources are prioritized for processing clusters of work orders containing emerging problem patterns.

[0160] Furthermore, the device also includes a feedback loop unit for establishing a feedback loop to continuously optimize the clustering model, specifically: Monitor the processing effect of typical representative work orders and collect feedback data from support personnel on the accuracy of cluster classification; based on the feedback data and processing results, dynamically adjust the synonym normalization rules and vectorization model parameters; re-execute cluster analysis according to the optimized parameters to generate updated work order cluster distribution and typical representative work orders; compare and analyze the updated results with historical data, evaluate the optimization effect, and generate a model performance report.

[0161] Furthermore, the device also includes a visualization analysis unit for constructing a visualization analysis interface. Specifically, it generates a multidimensional pivot table based on the distribution of work order clusters and extracts key features of typical representative work orders; it uses a dimensionality reduction algorithm to project the work order vector matrix onto a two-dimensional space to generate a visualization scatter plot; it marks the locations of typical representative work orders and the distribution areas of outlier work orders in the scatter plot; and it displays the clustering results, work order details, and resource allocation schemes through an interactive interface.

[0162] The device in this embodiment, through the organic integration of the aforementioned units, constructs a complete technology chain from data acquisition, text processing, vectorization, cluster analysis to centroid calculation and outlier identification. The device can automatically extract typical representative work orders and outlier work orders from massive amounts of work orders, providing reliable technical means for resource optimization, problem prevention, and decision support in work order management, significantly improving the efficiency, accuracy, and uniformity of work order processing.

[0163] Figure 3This is a block diagram illustrating a work order processing apparatus according to an exemplary embodiment. For example, the apparatus may be a mobile phone, computer, digital broadcasting terminal, messaging device, game console, tablet device, medical device, fitness equipment, personal digital assistant, etc.

[0164] Reference Figure 3 The device 300 may include one or more of the following components: processing component 301, memory 302.

[0165] The following explanations of the main terms used in this patent application are provided to aid in understanding the technical solution of this invention. Some terms are further illustrated with concise formulas.

[0166] 1. Work order vector matrix This refers to the set of numerical representations obtained after processing multiple work order text data using vectorization techniques. Each work order is converted into a vector of fixed dimensions, and the set of all vectors arranged in rows or columns forms a matrix, denoted as X, where X ∈ R^(N×D), N represents the number of work orders, and D represents the dimension of the vectors. This matrix serves as the input data basis for subsequent clustering algorithms.

[0167] 2. Center of mass The geometric center of a work order cluster in the feature space can be regarded as the core feature representative of that type of work order.

[0168] 3. Weighted covariance matrix A statistic used to describe the variance of each feature dimension in a weighted dataset and the covariance relationship between different feature dimensions. Its calculation considers not only the distribution of data points but also assigns a weight w_i to each data point.

[0169] 4. Distance from Maharanobis A distance metric based on the data distribution pattern, used to calculate the distance between a data point and a distribution (usually defined by the centroid and covariance matrix). It considers the correlation between features and performs scale normalization, reflecting the statistical characteristics of the data better than Euclidean distance.

[0170] 5. Chi-square distribution A continuous probability distribution. In the context of this application, if a data vector x originates from a multivariate normal distribution with mean μ and covariance matrix Σ, then the square of its Mahalanobis distance follows a chi-square distribution with degrees of freedom D (feature dimension). This property provides a theoretical basis for identifying outliers based on statistical significance.

[0171] 6. Probability Distribution Manifold A mathematical concept referring to the space defined by a continuous and smooth probability density function p(x) reconstructed from a discrete set of work order vector data points {x_i} using nonparametric methods such as kernel density estimation. Each point on this "manifold" corresponds to a probability density value, and the elevation of its "terrain" intuitively reflects the clustering and sparsity of work orders in the feature space.

[0172] 7. Li Derivative In differential geometry, this concept measures the rate of change of one vector field along the direction of another. In this application, it is used to measure the change of the gradient vector field ξ on a probability distribution manifold along its own integral curve (i.e., the mean drift trajectory), denoted as L_ξ ξ. The magnitude of its norm ‖L_ξ ξ‖ is used to scientifically determine whether the mean drift iteration process has converged to a stationary point (attractor).

[0173] 8. Deep Feature Distillation Network This refers to a deep learning model based on an autoencoder architecture, used to extract low-dimensional, information-rich core feature representations from high-dimensional multimodal fusion vectors. Its goal is to retain the most important and discriminative information from the input data, while filtering out redundancy and noise, by jointly optimizing objective functions such as reconstruction error (e.g., mean squared error, MSE) and mutual information, much like a "distillation" process.

[0174] 9. Mutual Information Mutual information (I(X; Z)) is a concept in information theory used to measure the degree of interdependence between two random variables. For input data X and its latent representation Z, mutual information I(X; Z) measures how much information gain X gains knowing Z. In neural network training, maximizing I(X; Z) encourages the encoder to learn feature representations that are highly relevant to and informative of the input.

Claims

1. A method for processing work orders, characterized in that, Includes the following steps: Obtain historical work order data and clean the work order text in the historical work order data to obtain work order text information; The work order text information is processed by word segmentation and synonym normalization; The processed work order text information is converted into a fixed-dimensional vector to construct a work order vector matrix; Perform cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs; Calculate the centroid of each of the clusters; Based on the centroid, the work order closest to the centroid in each cluster is determined as a typical representative work order, and work orders that are more than a preset threshold away from the centroid are determined as outlier work orders.

2. The method according to claim 1, characterized in that, The step of calculating the centroid of each of the clusters includes: Based on a predefined comprehensive weight coefficient, a weighted average is calculated on the work order vectors within the cluster to generate a weighted centroid vector; wherein, the comprehensive weight coefficient is dynamically calculated from the processing time, semantic matching degree, and customer rating of the corresponding work order. Based on the Mahalanobis distance between the weighted centroid vector and each vector within the cluster, the judgment thresholds for the typical representative work order and the outlier work order are recalibrated.

3. The method according to claim 1, characterized in that, The step of converting the processed work order text information into a fixed-dimensional vector to construct a work order vector matrix includes: A pre-trained language model is used to perform deep semantic encoding on the work order text information after word segmentation and synonym normalization to generate a high-dimensional semantic vector. The structured feature data of the work order is standardized to generate a numerical feature vector; The high-dimensional semantic vector and the numerical feature vector are concatenated and their dimensions reduced to generate the final work order vector used to construct the work order vector matrix.

4. The method according to claim 1, characterized in that, The steps of performing word segmentation and synonym normalization on the work order text information include: The work order text information is segmented to generate an initial segmentation sequence; Entity recognition and semantic verification are performed on the terms in the initial word segmentation sequence based on the domain knowledge base; Based on the pre-constructed thesaurus, semantically validated terms are mapped to standard concept nodes, thus completing the thesaurus normalization process.

5. The method according to claim 1, characterized in that, The step of determining the work order closest to the centroid from each of the clusters as a typical representative work order based on the centroid includes: Calculate the Mahalanobis distance from each work vector within the cluster to the centroid; The top K work orders with the smallest Mahalanobis distance are selected as the candidate typical representative work order set; Based on the comprehensive weight of each work order in the candidate typical representative work order set, the work order with the highest comprehensive weight is selected as the final typical representative work order.

6. The method according to claim 1, characterized in that, The step of identifying work orders that are more than a preset threshold away from the centroid as outlier work orders includes: Based on the distance distribution from all work orders within the cluster to the centroid, calculate the mean and standard deviation of the distances; Based on the mean and standard deviation of the distance, the preset threshold is dynamically set to the mean distance plus N times the standard deviation; Work orders whose distance exceeds the preset threshold set by the dynamic system are marked as critical outlier work orders; A secondary clustering analysis was performed on the set of key outliers to distinguish emerging problem patterns from random noise.

7. The method according to claim 1, characterized in that, The method further includes: Resource optimization allocation is performed based on the distribution of the aforementioned work order clusters; The resource optimization allocation based on the distribution of the work order cluster includes: Based on the size of each work order cluster and the problem complexity reflected by the typical representative work order, calculate the processing resource weight required for each cluster; Based on the processing resource weights, the matching relationship between support personnel and work order clusters is dynamically adjusted; The outlier work order set is assigned to a high-level expert queue for specialized processing, and resources are prioritized for processing work order clusters containing emerging problem patterns.

8. The method according to claim 1, characterized in that, The method further includes: establishing a feedback loop to continuously optimize the clustering model; and... The establishment of a feedback loop to continuously optimize the clustering model includes: Monitor the processing effect of the typical representative work orders and collect feedback data from support personnel on the accuracy of cluster classification; Based on the feedback data and processing results, the synonym normalization rule and the parameters of the vectorization model are dynamically adjusted. Based on the optimized parameters, the cluster analysis is re-executed to generate an updated work order cluster distribution and typical representative work orders; The updated results are compared and analyzed with historical data to evaluate the optimization effect and generate a model performance report.

9. The method according to claim 1, characterized in that, The method also includes building a visual analysis interface: The construction of the visual analysis interface includes: A multidimensional pivot table is generated based on the work order cluster distribution, and key features of the typical representative work orders are extracted. The work order vector matrix is ​​projected onto a two-dimensional space using a dimensionality reduction algorithm to generate a visual scatter plot. The locations of the typical representative work orders and the distribution areas of the outlier work orders are marked in the scatter plot. The clustering results, work order details, and resource allocation schemes are displayed in a linked manner through an interactive interface.

10. The method according to claim 1, characterized in that, The step of calculating the centroid of each of the clusters further includes: Based on the weighted covariance matrix of each work order vector within the cluster, a Mahalanobis distance metric space is constructed. The centroid of the probability distribution of the cluster is solved by minimizing the objective function of the weighted sum of squared Mahalanobis distances. Based on the normalized distance metric between the centroid of the probability distribution and each vector within the cluster, a geometric space reference coordinate system is generated to represent the work order screening. Based on the statistical distribution characteristics under this reference coordinate system, the significance level threshold for outlier determination is dynamically optimized.

11. An electronic device for processing work orders, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method as described in any one of claims 1-10.

12. An apparatus for processing work orders, characterized in that, include: The historical data acquisition unit is used to acquire historical work order data and clean the work order text in the historical work order data to obtain work order text information. The normalization unit is used to perform word segmentation and synonym normalization processing on the work order text information; The work order vector matrix construction unit is used to convert the processed work order text information into a fixed-dimensional vector to construct a work order vector matrix. The work order cluster determination unit is used to perform cluster analysis on the work order vector matrix to determine the cluster to which each work order belongs; A centroid calculation unit is used to calculate the centroid of each of the clusters; The outlier work order determination unit is used to determine, based on the centroid, the work order closest to the centroid in each cluster as a typical representative work order, and to determine the work orders that are more than a preset threshold away from the centroid as outlier work orders.

Images