A power grid intelligent customer service robot access response time anomaly detection method, system, device and storage medium

Abstract

The application discloses a kind of power grid intelligent customer service robot access response time anomaly detection method, system, equipment and storage medium, is related to power grid robot anomaly detection field, method includes: based on the dialogue log data of power grid intelligent customer service robot, generates input sample data set;Improved deep orthogonal hypersphere compression model is constructed, input sample data set is input into compression model, the abnormal score of input sample is calculated, the abnormal degree of input sample is evaluated and is marked;Based on the abnormal degree of input sample, the abnormal score of all abnormal samples is clustered, and the abnormal sample is corrected based on the clustering result;The abnormal sample after clustering correction is used as final detection result, to determine the abnormal condition of power grid intelligent customer service robot access response time;The application realizes high-precision anomaly detection by adaptive learning response time distribution, significantly reduces false alarm rate, to improve the operation and maintenance automation level and reliability of intelligent customer service system.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection for power grid robots, and in particular to a method, system, device, and storage medium for detecting anomalies in the access response time of a smart customer service robot for power grids. Background Technology

[0002] As a crucial tool for interaction between the power system and users, the response speed of intelligent customer service robots for power grids significantly impacts user experience. However, in actual operation, these robots may be affected by various factors, such as system failures, data transmission delays, and sudden surges in user requests, all of which can lead to abnormal robot response times.

[0003] Traditional smart power grid customer service systems typically rely on fixed rules or simple threshold judgments to detect abnormal access response times. This approach depends on human experience, has high configuration and maintenance costs, and requires different rules or thresholds to be set for different systems. However, different systems vary greatly, making it difficult to accurately define a rule or threshold, which can easily lead to false alarms. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention provides a method, system, device and storage medium for detecting abnormal access response time of a smart customer service robot for power grids.

[0005] Therefore, the technical problem solved by this invention is: how to more accurately and automatically detect anomalies in the access response time of the smart customer service robot for power grids, and effectively reduce the false alarm rate.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting abnormal response time of a smart customer service robot for power grids, comprising: Based on the dialogue log data of the smart customer service robot for power grids, an input sample dataset is generated; An improved deep orthogonal hypersphere compression model is constructed. The input sample dataset is input into the compression model. The compression model encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input sample, evaluates the degree of anomaly of the input sample, and labels it. Based on the degree of abnormality of the input samples, the abnormality scores of all abnormal samples are clustered, and the abnormal samples are corrected based on the clustering results. The cluster-corrected outlier samples are used as the final detection results to determine the abnormalities in the access response time of the smart customer service robot of the power grid.

[0007] As a preferred solution for detecting anomalies in the access response time of a smart customer service robot for power grids, the following is provided: The construction of the improved deep orthogonal hypersphere compression model involves inputting the input sample dataset into the compression model. The compression model encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input samples, evaluates the degree of anomaly of the input samples, and labels them, including: An improved deep orthogonal hypersphere compression model is constructed, which maps the input data into a continuous probabilistic latent representation through a variational inference-based encoder. The latent representation is orthogonalized and normalized through an orthogonal projection layer to eliminate feature correlation and construct a projected feature space that conforms to the hypersphere distribution assumption.

[0008] As a preferred solution for detecting anomalies in the access response time of a smart customer service robot for power grids, the following is provided: The construction of the improved deep orthogonal hypersphere compression model, which involves inputting the input sample dataset into the compression model, encoding and orthogonally transforming the data in the input sample dataset, calculating the anomaly score of the input samples, evaluating the anomaly degree of the input samples, and labeling them, further includes: During the training phase of the compressed model, the distance from the projected features to the center point of the space is used as a metric. An optimization algorithm is used to minimize the overall distance between normal samples and the center point to shrink the data distribution. Based on the quantiles of the distance distribution, a hypersphere decision radius that can cover a preset proportion of normal samples is determined.

[0009] As a preferred solution for detecting anomalies in the access response time of a smart customer service robot for power grids, the following is provided: The construction of the improved deep orthogonal hypersphere compression model, which involves inputting the input sample dataset into the compression model, encoding and orthogonally transforming the data in the input sample dataset, calculating the anomaly score of the input samples, evaluating the anomaly degree of the input samples, and labeling them, further includes: During the inference phase of the compressed model, the distance from the projected features of a new sample to the center point is calculated. An anomaly score is obtained by comparing the square of the distance with the square of the decision radius. Based on the anomaly score, the sample is labeled as either abnormal or normal.

[0010] The beneficial effects of this preferred technical solution are: it provides a clear and efficient anomaly scoring and judgment mechanism. By calculating the squared difference between the projection distance and the decision radius as the anomaly score, its physical meaning is intuitive and the calculation is simple, making the model reasoning process fast and interpretable. It can directly output a quantitative assessment of the degree of anomaly and clear binary classification labels, which is convenient for integration into automated operation and maintenance systems for real-time monitoring and alarms.

[0011] As a preferred solution for detecting anomalies in the access response time of a smart customer service robot for power grids, the following is provided: The process of clustering the anomaly scores of all anomalous samples based on the anomaly degree of the input samples, and then correcting the anomalous samples based on the clustering results, includes: For the set of abnormal samples initially detected, the optimal number of clusters is automatically determined based on the preset clustering quality evaluation index, and the abnormal samples are divided into multiple clusters according to the number of clusters.

[0012] The beneficial effects of this preferred technical solution are as follows: Preliminary anomaly results are subdivided through cluster analysis, avoiding the conflation of different types of anomalies or potential misjudgments. The optimal number of clusters is automatically determined using cluster quality evaluation indicators, overcoming the blindness of manually pre-setting the number of clusters. This makes the cluster structure more closely match the distribution characteristics of the data itself, providing a reliable data grouping foundation for subsequent accurate identification and correction of misjudged samples.

[0013] As a preferred solution for detecting anomalies in the access response time of a smart customer service robot for power grids, the following is provided: The step of clustering the anomaly scores of all anomalous samples based on the anomaly degree of the input samples, and correcting the anomalous samples based on the clustering results, further includes: Calculate the average anomaly score of samples within each cluster, and identify the cluster with the lowest average anomaly score as the cluster of normal samples that were misjudged. All samples in the misclassified normal sample cluster are relabeled as normal, thus completing the correction of the abnormal detection results.

[0014] As a preferred solution for detecting anomalies in the access response time of a smart customer service robot for power grids, the following is provided: The dialogue log data from the smart customer service robot based on the power grid, used to generate the input sample dataset, includes: Obtain structured dialogue log data from the smart customer service robot for the power grid, wherein the log data contains multiple parameters associated with the operation of each service; The log data is grouped and associated according to the session identifier, and the access response time parameter corresponding to the initial service request in each session is extracted. The extracted response time parameter is cleaned to handle missing values ​​and normalized to generate an input sample dataset.

[0015] Secondly, the present invention provides a system for detecting abnormal response time of a smart customer service robot for power grids, comprising: The sample dataset construction module is used to generate input sample datasets based on the dialogue log data of the smart customer service robot for the power grid. The intelligent anomaly scoring module is used to construct an improved deep orthogonal hypersphere compression model. The input sample dataset is input into the compression model, which encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input sample, evaluates the degree of anomaly of the input sample, and labels it. The clustering correction module is used to cluster the anomaly scores of all abnormal samples based on the anomaly degree of the input samples, and to correct the abnormal samples based on the clustering results. The anomaly decision and output module is used to determine the anomalies in the access response time of the smart customer service robot of the power grid by taking the cluster-corrected anomaly samples as the final detection results.

[0016] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for detecting abnormal access response time of the smart customer service robot of the power grid are implemented.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of a method for detecting anomalies in the access response time of a smart customer service robot for power grids.

[0018] The beneficial effects of this invention are as follows: This invention employs an improved deep orthogonal hypersphere contraction algorithm, which automatically learns the response time distribution characteristics under different business scenarios through a data-driven approach. It eliminates the need for manually setting thresholds and can adaptively identify different types of anomalies, reducing manual intervention and improving the intelligence and accuracy of detection. An anomaly score clustering strategy is used; after initial anomaly detection, samples with high anomaly scores are clustered for optimization, allowing for better differentiation between truly anomalous and normal data. This effectively reduces the false alarm rate, ensuring the accuracy and reliability of anomaly detection, thereby improving system stability. Furthermore, replacing the autoencoder in the original deep orthogonal hypersphere contraction algorithm with a variational autoencoder, and using latent variable probability modeling based on the variational autoencoder, improves model robustness and generalization ability, making the latent representation of data more continuous and stable. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is an overall flowchart of a method for detecting abnormal access response time of a smart customer service robot for power grids, provided by the present invention.

[0021] Figure 2 This is a structural diagram of an improved deep orthogonal hypersphere contraction algorithm for detecting abnormal access response time of a smart customer service robot for power grids, provided by this invention.

[0022] Figure 3 This is an orthogonal projection layer schematic diagram of a method for detecting abnormal access response time of a smart customer service robot for power grids provided by the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for detecting abnormal access response time of a smart customer service robot for power grids, including: S1: Generate an input sample dataset based on the dialogue log data of the smart customer service robot for the power grid; S2: Construct an improved deep orthogonal hypersphere compression model, input the input sample dataset into the compression model, the compression model encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input sample, evaluates the degree of anomaly of the input sample and labels it; S3: Based on the degree of abnormality of the input samples, cluster the abnormality scores of all abnormal samples, and correct the abnormal samples based on the clustering results; S4: Use the cluster-corrected outlier samples as the final detection result to determine the abnormality of the power grid intelligent customer service robot access response time.

[0025] It should be noted that through steps S1-S4, a complete and adaptive intelligent anomaly detection system was constructed. This system automatically learns the complex distribution characteristics of normal response times using a data-driven approach. Its core lies in utilizing an improved deep learning model (deep orthogonal hypersphere compression model) to establish a compact "hypersphere" boundary for normal patterns, and combining this with a clustering post-processing strategy to optimize and correct the preliminary results. This achieves high-precision, low-false-report automated identification of anomalies in the response time of the power grid intelligent customer service robot without the need for preset thresholds, significantly improving the intelligence and reliability of system operation and maintenance.

[0026] Example 2, refer to Figures 1-3 As an embodiment of the present invention, based on the previous embodiment, a method for detecting abnormal access response time of a smart customer service robot for power grids is provided, comprising: In this embodiment, the step S1 above, which generates the input sample dataset based on the dialogue log data of the power grid intelligent customer service robot, includes: The dialogue log data of the power grid intelligent customer service robot includes the operational status of each service, which includes: timestamp, log level, log components, request type, status code, and response time.

[0027] Specifically, timestamps are used to identify the exact time each interaction occurs; log levels include debug, information, warning, and fault, used to record the severity of log messages; log components are used to identify which part of the customer service system generated the log; request types are used to describe the nature of the user request; status codes include HTTP status codes or other business status codes, used to indicate the processing result of the request; and response time is the time taken for each API response.

[0028] After obtaining the dialogue log data from the power grid intelligent customer service robot, the data is preprocessed, including: Log content is grouped according to session ID, all log entries in the same session are associated, and the access response time parameter values ​​of each service during the first round of dialogue of the robot in each session are extracted. Parse the log content and extract the service access response time parameters corresponding to each session ID; The extracted response time parameters are processed for missing values ​​and normalized.

[0029] It should be noted that grouping log entries by session ID ensures that all log entries within the same session can be correctly associated. Simultaneously, only the response time parameter values ​​for each service during the first round of dialogue with the robot in a single session are retained, avoiding discrepancies between sessions that could lead to varying service calls and difficulties in data aggregation.

[0030] In another possible implementation, regular expressions can be used to extract the access response time parameters for each service corresponding to each session ID from the logs. When handling missing values ​​and normalizing the extracted response time parameters, forward padding can be used to fill in missing values, i.e., filling missing values ​​with the previous valid value. If the beginning of a column of data contains missing values, backward padding is used to supplement them, ensuring data integrity and consistency. The access response time is normalized using a min-max normalization method. Where X is the original data, and These represent the minimum and maximum values ​​of the feature, respectively.

[0031] It should be noted that data aggregation integrates scattered data into a unified format, facilitating subsequent processing. Log parsing techniques are used to extract key parameter values ​​from the dialogue logs. Data cleaning, including missing value handling and data normalization, ensures data integrity and consistency, providing high-quality input for anomaly detection algorithms.

[0032] In this embodiment, the improved deep orthogonal hypersphere compression model constructed in step S2 above, the input sample dataset is input into the compression model, the compression model encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input samples, evaluates the anomaly degree of the input samples, and labels them, including: It should be noted that the core objective of constructing and applying the improved deep orthogonal hypersphere compression model is to learn a compact "hypersphere" decision boundary for the data distribution of normal samples. Similar to the idea of ​​OC SVM (One-Class Support Vector Machine) learning a segmenting hyperplane, this step aims to use deep learning to wrap most of the normal samples (e.g., 95%) in the training set within a hypersphere, while samples outside the sphere are considered anomalous.

[0033] Specifically, an improved deep orthogonal hypersphere compression model is constructed, employing a variational autoencoder (VAE) as the encoding layer. This VAE learns probabilistic latent variables from the input data to generate a more robust and continuous latent representation, enhancing the model's generalization ability. Subsequently, an orthogonal projection layer is introduced after this latent representation. This layer orthogonalsizes the latent representation using singular value decomposition (SVD), eliminating correlations between features and achieving standardization. This constructs a feature space that conforms to the ideal "standard hypersphere" assumption, resolving the boundary distortion problem that may be caused by the original feature distribution.

[0034] Specifically, the orthogonal projection layer is defined as: in, This represents the latent feature matrix of the current batch (b represents the number of samples, and k represents the number of features). This is the feature representation after orthogonal projection. for An identity matrix is ​​a dimensional matrix representing the pairwise orthogonal and standardized feature dimensions after projection.

[0035] Singular value decomposition is represented as: It is a singular value diagonal matrix. It is a right singular vector matrix; Indicates taking the first A submatrix composed of columns of right singular vectors. diag This indicates that the corresponding preceding text should be retained. A singular value.

[0036] Secondly, the model is trained to shrink the hypersphere. After inputting the input sample dataset into this compressed model, the Euclidean distance from the orthogonally projected feature vectors to the center of the hypersphere (the mean of the features of all training samples in the projection space) is used as the basis for anomaly measurement. The training process drives the projection points of normal samples to cluster towards the center of the hypersphere by minimizing the following objective function: in, For the first The feature vector obtained after passing each sample through an orthogonal projection layer As the new center point, The coefficients of the L2 regularization term, For the network parameters that need to be trained, The set of weights for L2 regularization.

[0037] It should be noted that the first term in the objective function is to make these data points as close to the center as possible in the projected space. To achieve the "shrinkage" of the hypersphere, the second step is to adjust the network parameters. Regular expression penalty, coefficient The weights are used to balance these two factors. This optimization can be viewed as approximating the center of a hypersphere in the projection space and simultaneously regularizing the network during training.

[0038] Next, the decision boundary radius is determined. After model training is complete, the hypersphere decision radius is determined using the quantile method based on the set of distances from the training samples to the center of the sphere in the projection space. This radius is expressed as: in, Let the radius be the hypersphere. For samples in the projection space To the center distance, The set of distances for all training samples. Let be the quantile hyperparameter corresponding to the hypersphere decision boundary; in this embodiment, the quantile hyperparameter corresponding to the hypersphere decision boundary is preferably set to 0.95.

[0039] Finally, anomaly scoring and determination are performed. During the model inference (or testing) phase, for a new sample, its feature vector is first obtained through the trained VAE and orthogonal projection layer, and its distance to the center of the sphere is calculated. The anomaly score is calculated as follows: in, For the first Individual sample anomaly scores; like If the value is outside the hypersphere, the sample is labeled as an anomalous sample with a label of 1; if... If the value is 0, it indicates that the sample is located inside the hypersphere, and the sample is then labeled as a normal sample with a label of 0.

[0040] It should be noted that the data is encoded using a VAE encoder, and a continuous and robust representation is generated by learning the latent distribution of the data; an orthogonal projection layer is used to perform orthogonal transformation on the data to ensure that the data distribution conforms to the standard hypersphere assumption; finally, the anomaly score of the sample is calculated based on the hypersphere radius to evaluate its degree of anomaly, which can obtain a more compact hypersphere boundary that fits the assumption, and significantly improve the accuracy of anomaly detection.

[0041] Furthermore, such as Figure 2 The diagram shows the improved Deep Orthogonal Hypersphere Contraction (DOHSC) algorithm architecture. The model utilizes a representation learning module (VAE encoder) to extract features from the input data, mapping high-dimensional data to a low-dimensional latent space. An orthogonal projection layer is added after the latent representation, using singular value decomposition to ensure that the projected features satisfy orthogonality and normalization requirements, overcoming the hypersphere shape deviation problem caused by inconsistent feature variances and correlations. Next, the model optimizes the objective function to make the projected representation of normal samples as close as possible to a preset center, constructing an approximately standard hypersphere decision boundary. Finally, 2D trend plots and 3D detection results visualization visually demonstrate the evolution of data distribution and the final anomaly detection effect.

[0042] Furthermore, after the input sample data passes through the VAE encoder, parameters describing the distribution of latent variables are generated. Through reparameterization techniques, latent variables are sampled from this distribution, ensuring effective backpropagation during model training. Compared to traditional autoencoders, the VAE encoding layer generates probabilistic latent variables and utilizes KL divergence regularization to achieve a smooth and continuous distribution in the latent space, thus better capturing data uncertainty, improving the robustness and accuracy of anomaly detection, and providing a more stable and continuous feature representation for subsequent orthogonal projection and anomaly detection.

[0043] like Figure 3As shown, simply minimizing distance often leads to the learned decision boundary exhibiting a hyperellipsoidal or even more irregular shape, which deviates from the hypersphere assumption. To address this issue, an orthogonal projection layer is added after the latent representation to standardize and decorrelate the features, ensuring that the projected features satisfy orthogonality. Orthogonal projection is achieved through singular value decomposition, effectively adjusting the data distribution to better conform to the standard hypersphere assumption.

[0044] In this embodiment, step S3 above, which involves clustering the anomaly scores of all abnormal samples based on the degree of anomaly of the input samples, and then correcting the abnormal samples based on the clustering results, includes: The silhouette coefficient is used as the evaluation index. Within the preset range of the number of clusters, the K value that maximizes the silhouette coefficient is selected as the optimal number of clusters. Initialize K cluster centers, calculate the Euclidean distance from each outlier sample to each cluster center, and assign it to the nearest cluster; recalculate the mean of all sample points in each cluster as the new cluster center, until the cluster centers no longer change or the preset maximum number of iterations is reached; Based on the clustering results, the average anomaly score corresponding to the centroid of each cluster is calculated, and the cluster C with the lowest anomaly score is identified as a misjudged cluster composed of normal samples at the edge. The labels of all samples in cluster C are relabeled from abnormal to normal, completing the post-processing correction of the abnormality detection results.

[0045] It should be noted that in high-dimensional space, normal data may be distributed on or around the surface of the hypersphere, and therefore easily misclassified as anomalous samples. To reduce misclassification, label transformation is performed based on the clustering results. The anomalous score of each cluster centroid is calculated, and the cluster C with the lowest anomalous score is found. The samples in cluster C are then converted from anomalous to normal, meaning that these samples may be misclassified as normal data.

[0046] In this embodiment, in step S4 above, the cluster-corrected abnormal samples are used as the final detection result to determine the abnormal situations in the access response time of the power grid intelligent customer service robot, including: Based on the corrected anomaly classification results, samples labeled 1 are taken as the final anomaly samples.

[0047] This invention addresses the problem of inaccurate anomaly identification due to manually set rules or thresholds. It proposes an automated detection method based on an improved DOHSC algorithm, enhancing detection accuracy and intelligence. Traditional methods rely on manually setting fixed rules or thresholds, but the applicable thresholds may vary across different business scenarios, leading to significant false positives and false negatives. This invention employs an improved DOHSC algorithm, automatically learning response time distribution characteristics under different business scenarios through a data-driven approach, eliminating the need for manual threshold setting. This method can adaptively identify different types of anomalies, reducing manual intervention and improving the intelligence and accuracy of detection.

[0048] To address the issue of sparse data distribution enclosed by hyperspheres in high-dimensional data spaces, which may lead to false positives, an anomaly score clustering optimization strategy is proposed to reduce the false positive rate. Traditional hypersphere methods in high-dimensional spaces may result in some normal data being far from the center point, leading to misclassification as anomalies and causing false positives. This invention employs an anomaly score clustering strategy. After initial anomaly detection, samples with high anomaly scores are clustered for optimization, better distinguishing truly anomalous data from normal data. This strategy effectively reduces the false positive rate, ensuring the accuracy and reliability of anomaly detection, thereby improving system stability.

[0049] To address the issue of insufficient latent representation optimization in the original DOHSC algorithm's AE encoder during data encoding, this paper proposes replacing the AE encoder with a VAE encoder. Based on the latent variable probability modeling of VAE, the robustness and generalization ability of the model are improved, making the latent representation of the data more continuous and stable.

[0050] Example 3: The above is an illustrative scheme of a method for detecting anomalies in the access response time of a smart customer service robot for power grids according to this embodiment. It should be noted that the technical solution of a system for detecting anomalies in the access response time of a smart customer service robot for power grids and the technical solution of the method for detecting anomalies in the access response time of a smart customer service robot for power grids described above belong to the same concept. Details not described in detail in the technical solution of the system for detecting anomalies in the access response time of a smart customer service robot for power grids in this embodiment can be found in the description of the technical solution of the method for detecting anomalies in the access response time of a smart customer service robot for power grids described above.

[0051] This embodiment also provides a power grid intelligent customer service robot access response time anomaly detection system, including: The sample dataset construction module is used to generate input sample datasets based on the dialogue log data of the smart customer service robot for the power grid. The intelligent anomaly scoring module is used to construct an improved deep orthogonal hypersphere compression model. The input sample dataset is input into the compression model, which encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input sample, evaluates the degree of anomaly of the input sample, and labels it. The clustering correction module is used to cluster the anomaly scores of all abnormal samples based on the anomaly degree of the input samples, and to correct the abnormal samples based on the clustering results. The anomaly decision and output module is used to determine the anomalies in the access response time of the smart customer service robot of the power grid by taking the cluster-corrected anomaly samples as the final detection results.

[0052] This embodiment also provides an electronic device applicable to a method for detecting abnormal response time of a smart customer service robot for power grids, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement the abnormal response time detection method for the smart customer service robot access in the above embodiments.

[0053] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a method for detecting abnormal access response time of a smart customer service robot for power grids as proposed in the above embodiment.

[0054] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for detecting abnormal access response time of a smart customer service robot for power grid proposed in the above embodiment. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting abnormal response time of a smart customer service robot for power grids, characterized in that, include: Based on the dialogue log data of the smart customer service robot for power grids, an input sample dataset is generated; An improved deep orthogonal hypersphere compression model is constructed. The input sample dataset is input into the compression model. The compression model encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input sample, evaluates the degree of anomaly of the input sample, and labels it. Based on the degree of abnormality of the input samples, the abnormality scores of all abnormal samples are clustered, and the abnormal samples are corrected based on the clustering results. The cluster-corrected outlier samples are used as the final detection results to determine the abnormalities in the access response time of the smart customer service robot of the power grid.

2. The method for detecting abnormal access response time of a smart customer service robot for power grids as described in claim 1, characterized in that, The construction of the improved deep orthogonal hypersphere compression model involves inputting the input sample dataset into the compression model. The compression model encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input samples, evaluates the degree of anomaly of the input samples, and labels them, including: An improved deep orthogonal hypersphere compression model is constructed, which maps the input data into a continuous probabilistic latent representation through a variational inference-based encoder. The latent representation is orthogonalized and normalized through an orthogonal projection layer to eliminate feature correlation and construct a projected feature space that conforms to the hypersphere distribution assumption.

3. The method for detecting abnormal response time of a smart customer service robot for power grids as described in claim 2, characterized in that, The construction of the improved deep orthogonal hypersphere compression model, which involves inputting the input sample dataset into the compression model, encoding and orthogonally transforming the data in the input sample dataset, calculating the anomaly score of the input samples, evaluating the anomaly degree of the input samples, and labeling them, further includes: During the training phase of the compressed model, the distance from the projected features to the center point of the space is used as a metric. An optimization algorithm is used to minimize the overall distance between normal samples and the center point to shrink the data distribution. Based on the quantiles of the distance distribution, a hypersphere decision radius that can cover a preset proportion of normal samples is determined.

4. A method for detecting abnormal access response time of a smart customer service robot for power grids as described in any one of claims 1-3, characterized in that, The construction of the improved deep orthogonal hypersphere compression model, which involves inputting the input sample dataset into the compression model, encoding and orthogonally transforming the data in the input sample dataset, calculating the anomaly score of the input samples, evaluating the anomaly degree of the input samples, and labeling them, further includes: During the inference phase of the compressed model, the distance from the projected features of a new sample to the center point is calculated. An anomaly score is obtained by comparing the square of the distance with the square of the decision radius. Based on the anomaly score, the sample is labeled as either abnormal or normal.

5. The method for detecting abnormal response time of a smart customer service robot for power grids as described in claim 4, characterized in that, The process of clustering the anomaly scores of all anomalous samples based on the anomaly degree of the input samples, and then correcting the anomalous samples based on the clustering results, includes: For the set of abnormal samples initially detected, the optimal number of clusters is automatically determined based on the preset clustering quality evaluation index, and the abnormal samples are divided into multiple clusters according to the number of clusters.

6. The method for detecting abnormal access response time of a smart customer service robot for power grids as described in claim 5, characterized in that, The step of clustering the anomaly scores of all anomalous samples based on the anomaly degree of the input samples, and correcting the anomalous samples based on the clustering results, further includes: Calculate the average anomaly score of samples within each cluster, and identify the cluster with the lowest average anomaly score as the cluster of normal samples that were misjudged. All samples in the misclassified normal sample cluster are relabeled as normal, thus completing the correction of the abnormal detection results.

7. The method for detecting abnormal response time of a smart customer service robot for power grids as described in claim 1, characterized in that, The dialogue log data from the smart customer service robot based on the power grid, used to generate the input sample dataset, includes: Obtain structured dialogue log data from the smart customer service robot for the power grid, wherein the log data contains multiple parameters associated with the operation of each service; The log data is grouped and associated according to the session identifier, and the access response time parameter corresponding to the initial service request in each session is extracted. The extracted response time parameter is cleaned to handle missing values ​​and normalized to generate an input sample dataset.

8. A power grid intelligent customer service robot access response time anomaly detection system, using the method described in any one of claims 1 to 7, characterized in that, include: The sample dataset construction module is used to generate input sample datasets based on the dialogue log data of the smart customer service robot for the power grid. The intelligent anomaly scoring module is used to construct an improved deep orthogonal hypersphere compression model. The input sample dataset is input into the compression model, which encodes and orthogonally transforms the data in the input sample dataset, calculates the anomaly score of the input sample, evaluates the degree of anomaly of the input sample, and labels it. Clustering correction module, used for clustering based on input samples The degree of abnormality is determined, and the abnormality scores of all abnormal samples are clustered. The abnormal samples are then corrected based on the clustering results. The anomaly decision and output module is used to determine the anomalies in the access response time of the smart customer service robot of the power grid by taking the cluster-corrected anomaly samples as the final detection results.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images