Explainable operation and maintenance fault identification method and system based on quantitative business feature generation
By constructing a business feature labeling data set and a fault labeling data set, using deep neural networks and quantitative business feature vectors, combined with gain rate calculation, the problem of poor interpretability in operation and maintenance fault detection is solved, and the interpretability of operation and maintenance fault detection is realized and root cause analysis support.
Patent Information
- Application Number
- CN202411139840.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-20
AI Technical Summary
The prior art is difficult to achieve interpretability of fault results in operation and maintenance fault detection. The black box characteristics of the deep learning model make it difficult to interpret subsequent detection results and position the root cause, and the existing methods interpret the results complex and unintuitive.
By constructing a business feature annotation data set and a fault annotation data set, a machine learning model is used to generate quantitative business feature vectors, and a deep neural network failure detection model is trained, combined with gain rate calculation to achieve interpretability of failures.
It realizes good interpretability of operation and maintenance fault detection results, can attribute fault detection to specific business characteristics, and provides clear impact weights to support root cause analysis.
Smart Images

Figure CN118656649B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of fault detection technology, and in particular to an explainable operation and maintenance fault identification method and system generated based on quantitative business features. Background Art
[0002] In the field of AIOps, deep learning models have made certain progress in the field of fault detection. Deep models have good fitting and generalization capabilities, and have advantages over traditional models in detection accuracy. However, due to the black box characteristics of deep models, a series of tasks such as the interpretation of subsequent detection results and root cause location have been difficult. The operation and maintenance industry has particularly prominent requirements for the interpretability of fault alarms.
[0003] Patent document CN115550977A (application number: 202211136228.4) discloses a root cause location method and device for abnormal key performance indicators, including using a machine learning model to select a marked original data set, select multiple important performance indicators, and filter out the associated performance indicators that do not participate in the calculation of the key performance indicators. Based on the multiple cycles before and after the abnormal key performance indicators of the wireless access network occur, the associated performance indicator data, the marked key performance indicator data and the known root cause performance indicator data are constructed to construct a root cause location data set; the discrete associated performance indicator data in the root cause location data set is calculated for the cause difference to obtain the cause of the abnormal key performance indicator; the continuous associated performance indicator data in the root cause location data set is calculated for multiple correlations to obtain the root cause of the abnormal key performance indicator. The patent does not directly determine whether the current scene is a "fault", but uses machine learning technology to narrow the scope of the root cause to be checked when it is informed that the "fault" exists, which is different from the implementation function of the present invention. The patent is only attributed to some indicator dimensions, such as "throughput", and its subsequent interpretation is still difficult. This patent uses quantitative business features, such as "the probability of 'throughput rate-increase' is greater than 95%", and its interpretation results have accurately located specific business phenomena;
[0004] Patent document CN114792138A (application number: 202110104332.4) discloses a machine learning model interpretation method, device and storage medium, which relates to the field of artificial intelligence technology. The method includes: training a first machine learning model based on network performance data, obtaining feature weights corresponding to each feature contained in the network performance data, and selecting model features based on feature weights; training a second machine learning model based on model features, and obtaining a decision tree structure corresponding to the second machine learning model; determining a judgment selection path in the decision tree structure as an explanatory working path; the patent uses a relatively interpretable decision tree model to fit a model that is more difficult to interpret, so as to achieve model interpretation. However, the decision tree model is still difficult to interpret when the depth increases. The present invention uses the gain rate for interpretation, which is more intuitive. Summary of the invention
[0005] In view of the defects in the prior art, the purpose of the present invention is to provide an explainable operation and maintenance fault identification method and system based on quantitative business feature generation.
[0006] According to the present invention, an explainable operation and maintenance fault identification method based on quantitative business feature generation includes:
[0007] Step S1: construct a business feature annotation dataset, and construct and train a probability estimation model for each business feature based on the business feature annotation dataset;
[0008] Step S2: construct a fault annotation data set, and construct a quantized service feature vector of the fault annotation data set based on the trained probability estimation model of each service feature;
[0009] Step S3: construct a fault detection deep model, and train the fault detection deep model using the quantized service feature vector of the constructed fault annotation data set to obtain a trained fault detection deep model;
[0010] Step S4: Use the trained fault detection deep model to judge the fault probability of the business indicator data in the operation and maintenance scenario to be detected. When the business indicator data in the operation and maintenance scenario to be detected is abnormal, the abnormality is explained;
[0011] The probability estimation model of the business feature is to obtain the probability of the business feature existing in each scenario through a machine learning model;
[0012] The fault detection deep model uses a deep neural network to predict the probability of faults based on scenarios.
[0013] Preferably, the step S1 adopts:
[0014] Step S1.1: Collect business indicator data in the operation and maintenance scenario, including: CPU usage, memory usage, disk read and write volume, network data transmission volume, network data reception volume, and business success rate; mark typical business features appearing in the business indicator data, including: stable, steep rise, steep drop, slow rise, slow drop, large fluctuation, peak value, and valley value, to obtain a business feature annotation data set;
[0015] Step S1.2: Select a non-repeated business label from the business feature labeling dataset and record the selected business label as current_label k ;
[0016] Step S1.3: Set the label to current_label k The sample label of is 1, and the labels of the remaining samples are 0, forming a training data set, denoted as sample_set k ;
[0017] Step S1.4: Build a business feature probability estimation model and use the training data set sample_set k Train the constructed business feature probability estimation model to obtain the trained business feature probability estimation model prob_model k ; Repeat triggering steps S1.2 to S1.4 until all business tags in the business feature annotation data set have constructed corresponding business feature probability estimation models.
[0018] Preferably, the step S1.1 adopts:
[0019] Step S1.1.1: All business indicators are organized into multiple time series and normalized;
[0020] Step S1.1.2: Divide the time series into samples using a sliding window method;
[0021] Step S1.1.3: Manually label each sample with business phenomena.
[0022] Preferably, the step S2 adopts:
[0023] Step S2.1: Annotate the operation and maintenance faults that occur based on the collected business indicator data in the operation and maintenance scenario, and construct an operation and maintenance fault training data sample;
[0024] Step S2.2: Each operation and maintenance fault training data sample contains all business indicator data within a time window, and each business indicator is estimated using the trained business feature probability estimation model; including: m business indicators and n business feature probability estimation models, converting the m business indicators of each operation and maintenance fault training data sample into a vector of length m×n, where each dimension is a numerical value with a clear business meaning, and recording the names of all dimensions to obtain a dimension information table.
[0025] Preferably, the step S2.1 adopts: aligning the business indicator data in all operation and maintenance scenarios according to time, and dividing them into samples using a sliding window method; marking the divided sample results, and marking 1 if the sample contains an operation and maintenance failure within the time range, otherwise it is marked as 0.
[0026] Preferably, the step S3 adopts: constructing a fault detection deep model, the input of the fault detection deep model is a quantized business feature vector of a fault annotation data set with a length of m×n, and the output of the fault detection deep model is the probability of whether there is a fault, 1 indicates that the probability of the input being a fault is 100%, and 0 indicates that the probability of the input being a fault is 0;
[0027] The fault detection deep model is trained using the quantized business feature vector of the fault annotation data set to obtain a trained fault detection deep model.
[0028] Preferably, the step S4 adopts:
[0029] Step S4.1: collect business indicator data in the operation and maintenance scenarios to be tested, align the business indicator data in all the operation and maintenance scenarios to be tested according to time, and divide them into samples using a sliding window method;
[0030] Step S4.2: Convert the business indicator data samples in the operation and maintenance scenario to be detected into quantitative business feature vectors, and input them into the trained fault detection deep model to obtain the probability of whether the current input is a fault, recorded as p_anomaly; when p_anomaly is greater than the preset value, it is considered that the sample corresponding to the current time window is abnormal;
[0031] Step S4.2: When there is an abnormality in the sample corresponding to the current time window, a fault explanation is performed on the sample corresponding to the current time window.
[0032] Preferably, the step S4.2 adopts:
[0033] Step S4.2.1: Failure probability is p_anomaly q The sample quantization business feature vector is: quant_vec q =[f1,f2,f3…f mxn-1 ,fmxn ];
[0034] Step S4.2.2: Select quant_vec q Replace any k dimensions in basis_vector with the corresponding values in basis_vector to form a new vector. The selection method of the current k dimensions is recorded as s. The result is the replaced vector, recorded as replace_vec q_k_s ; Wherein, basis_vector represents the base vector obtained by finding the median of the quantized service feature vectors of all samples bit by bit;
[0035] Step S4.2.3: replace_vec q_k_s The xth dimension of the k dimensions replaced by quant_vec q The value of the corresponding bit in the new vector is recorded as replace_recover_vec q_k_s_x ; Calculate the gain in estimated probability due to using the true value of the xth dimension:
[0036] lift q_k_s_x =replace_recover_vec q_k_s_x -replace_vec q_k_s ;
[0037] Repeat steps 4.2.2 to 4.2.3 to calculate the gain of the xth dimension when k is selected for any size and various choices of k, and calculate the sample sample q The average gain brought by the xth dimension is:
[0038] avg_lift q_x =
[0039] in, represents the number of dimensions to be randomly selected for replacement, Indicates the method number of the current k dimensions randomly selected, Indicates the number of business indicators, Indicates the number of business features, represents the number of ways to arbitrarily select i dimensions from the remaining dimensions while retaining the xth dimension;
[0040] Repeat steps 4.2.2 to 4.2.3 for sample q Calculate the average gain when x is each dimension, and then get the gain rate of each dimension:
[0041] lift_rate q_x =
[0042] Step 4.2.4: Sort each lift_rate in descending order q_x , get the maximum T dimensions, the current T dimensions are the fault detection deep model judgment samples sample q These are the T factors with the greatest abnormal impact.
[0043] According to the present invention, an explainable operation and maintenance fault identification system based on quantitative business feature generation is provided, comprising:
[0044] Module M1: Construct a business feature annotation dataset, and build and train a probability estimation model for each business feature based on the business feature annotation dataset;
[0045] Module M2: Construct a fault labeling dataset, and construct a quantitative service feature vector of the fault labeling dataset based on the trained probability estimation model of each service feature;
[0046] Module M3: construct a fault detection deep model, and use the quantized business feature vector of the constructed fault labeling data set to train the fault detection deep model to obtain a trained fault detection deep model;
[0047] Module M4: Use the trained fault detection deep model to judge the fault probability of the business indicator data in the operation and maintenance scenario to be detected. When the business indicator data in the operation and maintenance scenario to be detected is abnormal, the abnormality is explained;
[0048] The probability estimation model of the business feature is to obtain the probability of the business feature existing in each scenario through a machine learning model;
[0049] The fault detection deep model uses a deep neural network to predict the probability of faults based on scenarios.
[0050] Preferably, the module M1 adopts:
[0051] Module M1.1: Collect business indicator data in operation and maintenance scenarios, including: CPU usage, memory usage, disk read and write volume, network data transmission volume, network data reception volume, and business success rate; mark typical business features that appear in business indicator data, including: stable, steep rise, steep drop, slow rise, slow drop, large fluctuations, peak value, and trough, to obtain a business feature annotation data set;
[0052] Module M1.2: Select a business label from the business feature labeling dataset without duplication, and record the selected business label as current_label k ;
[0053] Module M1.3: Set label to current_labelk The sample label of is 1, and the labels of the remaining samples are 0, forming a training data set, denoted as sample_set k ;
[0054] Module M1.4: Construct a business feature probability estimation model and use the training data set sample_set k Train the constructed business feature probability estimation model to obtain the trained business feature probability estimation model prob_model k ; Repeat triggering modules M1.2 to M1.4 until all business tags in the business feature annotation data set have built corresponding business feature probability estimation models;
[0055] The module M2 adopts:
[0056] Module M2.1: Annotate the operation and maintenance faults that occur based on the collected business indicator data in the operation and maintenance scenarios, and construct operation and maintenance fault training data samples;
[0057] Module M2.2: Each operation and maintenance fault training data sample contains all business indicator data within a time window, and each business indicator is estimated using the trained business feature probability estimation model; including: m business indicators and n business feature probability estimation models, converting the m business indicators of each operation and maintenance fault training data sample into a vector of length m×n, where each dimension is a value with a clear business meaning, and recording the names of all dimensions to obtain a dimension information table;
[0058] The module M3 adopts: constructing a fault detection deep model, the input of which is a quantized service feature vector of a fault annotation data set of length m×n, and the output of which is a fault probability, 1 indicates that the probability of the input being a fault is 100%, and 0 indicates that the probability of the input being a fault is 0;
[0059] The fault detection deep model is trained using the quantized business feature vector of the fault annotation data set to obtain a trained fault detection deep model;
[0060] The module M4 adopts:
[0061] Module M4.1: Collect business indicator data in the operation and maintenance scenarios to be tested, align the business indicator data in all the operation and maintenance scenarios to be tested according to time, and divide them into samples using the sliding window method;
[0062] Module M4.2: Convert the business indicator data samples in the operation and maintenance scenario to be detected into quantitative business feature vectors, and input them into the trained fault detection deep model to obtain the probability of whether the current input is a fault, recorded as p_anomaly; when p_anomaly is greater than the preset value, it is considered that the sample corresponding to the current time window is abnormal;
[0063] Module M4.2: When there is an abnormality in the sample corresponding to the current time window, a fault explanation is performed on the sample corresponding to the current time window.
[0064] Compared with the prior art, the present invention has the following beneficial effects:
[0065] 1. The present invention converts basic indicator data into "probability of meeting specific business characteristics", that is, "quantified business characteristics". For example, the probability that the current disk read and write indicators meet the business characteristic of "slowly rising" makes the feature dimension of the model input have good interpretability, which meets the requirements of high interpretability in operation and maintenance scenarios;
[0066] 2. Based on interpretable business features, the deep neural model is used as a tool for subsequent probability estimation, while also combining the good generalization and fitting capabilities of the deep neural network model;
[0067] 3. The present invention calculates the gain of each dimension of the vector of the input model, attributes the factors of each detection as "fault" to the specific feature dimension, and explains the reasons for each detection as "fault" for the fault detection deep model based on the interpretable feature dimension, which facilitates the root cause analysis of the fault, making full use of the advantage of the high readiness rate of the deep model in "fault" detection and making up for the shortcomings of the deep model in interpretability;
[0068] 4. The present invention not only retains the powerful fitting and generalization capabilities of the deep model, but also relies on quantitative business features and fault model interpretation to interpret the results of the deep model. It can attribute fault detection to a feature with clear business significance and give a quantitative value for the impact weight of the feature, thereby achieving good interpretability and providing good support for root cause location. For example, the business feature "the probability that the current scenario meets the 'steep increase in disk read and write rates' is 95%" contributes 30% to the judgment of "fault". BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0070] Figure 1 The flowchart of the explainable operation and maintenance fault identification method generated based on quantitative business features. DETAILED DESCRIPTION
[0071] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0072] Example 1
[0073] According to the present invention, an explainable operation and maintenance fault identification method based on quantitative business feature generation is provided. Figure 1 As shown, including:
[0074] Step 1: Build a quantitative business feature generation module. The input of this step is the operation and maintenance indicator data sample and the corresponding annotation, and the output is a series of basic business feature probability estimation models.
[0075] Step 2: Build a deep model for operation and maintenance fault detection. This step is based on the quantitative business feature generation module, with the input being the fault indicator data set to be trained and the output being the deep model for operation and maintenance fault detection.
[0076] Step 3: Operation and maintenance fault detection and model judgment as fault explanation. This step is based on the quantitative business feature generation module and the operation and maintenance fault detection deep model. The input is the indicator data to be detected, and the output is the judgment of whether it is a fault. If it is a fault, the various factors and corresponding weights that give the judgment are given at the same time;
[0077] Among them, the business feature probability estimation model can select various machine learning models with output probabilities ranging from 0 to 1, such as traditional logistic regression models. By using the business feature probability estimation model, the probability of the existence of business features can be estimated for each scenario later. For example, for a certain scenario, the probability of the existence of a business feature "steep rise" is 0.3, the probability of the existence of a business feature "large fluctuation" is 0.81, etc.
[0078] The fault detection deep model is constructed by a deep neural network; the input of the neural network is a vector of length m×n, where m is the number of business indicators and n is the number of business features; the output of the neural network is a number from 0 to 1, indicating the probability of a fault. The model achieves the purpose of estimating the probability of a fault based on the scenario.
[0079] The step 1 comprises the following steps:
[0080] Step 1.1: Construct a business feature annotation data set. Collect business indicator data (in the form of multiple independent time series) in the operation and maintenance scenario, and annotate the typical business features that appear on the business indicators. These indicators include: CPU occupancy, memory occupancy, disk read and write volume, network data sent, network received data, business success rate, etc. The data aggregation units of these indicators are, for example, seconds, minutes, etc. Typical business features include: stable, steep rise, steep decline, slow rise, slow decline, large fluctuations, reaching peak, reaching bottom, etc.
[0081] The following is the construction process of the labeled dataset:
[0082] Step 1.1.1: Normalize multiple time series consisting of all business indicators, for example, by using the quantile method or the maximum and minimum value method;
[0083] Step 1.1.2: Divide the time series into samples using a sliding window method, such as a window length of 10 minutes and a sliding step of 1 minute;
[0084] Step 1.1.3: Manually annotate each sample with business phenomena, such as stable, steep rise, steep decline, slow rise, slow decline, large fluctuation, peak, bottom, etc. There are no less than 100 annotated samples for each business phenomenon, and all annotated data together constitute the "business feature annotated data set".
[0085] Step 1.2: Build a business feature probability estimation model. Based on the "business feature annotation data set", build a probability estimation model for each business feature. The probability estimated by each model is the quantified business feature. The process of building a business feature probability estimation model is as follows:
[0086] Step 1.2.1: Select a business label from the "Business Feature Label Dataset" without duplication, and record the selected business label as current_label k ;
[0087] Step 1.2.2: Set the label to current_label k The sample label of is 1, and the labels of the remaining samples are 0, forming a training data set, denoted as sample_set k ;
[0088] Step 1.2.3: Select a simple machine learning probability estimation model (the model requires the probability of the output sample belonging to the business feature, the value is a floating point number between 0 and 1, such as logistic regression) for the data set sample_set k Perform training, and the trained model is recorded as prob_model k ;
[0089] Step 1.2.4: Repeat steps 1.2.1 to 1.2.3 until all business tags in the "business feature annotation dataset" have built corresponding profile estimation models;
[0090] The step 2 comprises the following steps:
[0091] Step 2.1: Construct a fault annotation dataset. Collect business indicator data in the operation and maintenance scenario (the collected business indicator data category is the same as that in step 1.1), annotate the operation and maintenance faults that occur, and construct operation and maintenance fault training data samples. The sample construction process is as follows:
[0092] Step 2.1.1: Align the business indicator data in all operation and maintenance scenarios by time, and divide them into samples using the sliding window method. The window size and step size need to be the same as those in step 1.1.2.
[0093] Step 2.1.2: Label the sample results divided in step 2.1. If the time range of the sample contains an operation and maintenance failure, it is marked as 1, otherwise it is marked as 0;
[0094] Step 2.2: Construct the quantitative business feature vector of the fault labeling data set. For the samples generated in step 2.1, each sample contains the data of all business indicator data in a time window. Each business indicator is estimated using the business feature probability estimation model constructed in step 1. If each sample has a total of m business indicators, step 1 generates a total of n business feature probability estimation models. Then, for each fault labeling sample, the m business indicators can be converted into a vector of length m×n, where each dimension is a value with a clear business meaning. For example, for dimension j, its name (meaning) can be the probability of "disk read and write volume-steep increase", and for dimension k, its name (meaning) can be the probability of "network data volume-slow decrease". The names of all dimensions are recorded to obtain a dimension information table, recorded as feature_name_dict, which is used in the subsequent fault detection interpretation process. The median of the quantitative business feature vectors of all samples is calculated bit by bit to obtain the base vector, recorded as basis_vector, which will be used for fault model interpretation. For example, for basis_vector, the jth median can be 0.032, and the kth median can be 0.112;
[0095] The dimension information table is:
[0096]
[0097] Step 2.3: Train a deep fault detection model based on quantitative service features. Construct a deep neural network classification model. The input of the model is a vector of length m×n. The output of the model is the probability of whether it is a fault. 1 means that the probability of the input being a fault is 100%, and 0 means that the probability of the input being a fault is 0. Use the data generated in step 2.2 to train the above model and obtain a deep fault detection model.
[0098] The step 3 comprises the following steps:
[0099] Step 3.1: Implement the detection of operation and maintenance failures. The specific process is as follows:
[0100] Step 3.1.1: Collect the business indicator data of the operation and maintenance scenarios to be tested (the collected business indicator data categories are the same as those in step 1.1), align the business indicator data of all operation and maintenance scenarios by time, and divide them into samples using the sliding window method. The window size and step size need to be the same as those in step 1.1.2;
[0101] Step 3.1.2: Convert the data in step 3.1.1 into a quantitative business feature vector in the same way as step 2.2;
[0102] Step 3.1.3: Input the "quantized service feature vector" output from step 3.1.1 into the "fault detection depth model" (generated by step 2.3) to obtain the probability of whether the input is a fault, recorded as p_anomaly. If p_anomaly>anomaly_threshold (the value is a decimal between 0 and 1), it is determined that the sample corresponding to the time window is abnormal;
[0103] Step 3.2: For the samples that are judged to be abnormal in step 3.1, give an explanation of the failure determined by the deep model in the operation and maintenance scenario. q is abnormal, and the probability of it being a fault is p_anomaly q The quantized service feature vector of this sample is:
[0104] quant_vec q = [f1, f2, f3… f mxn-1 ,f mxn ]
[0105] The explanation process for determining it as an abnormality in the operation and maintenance scenario is as follows:
[0106] Step 3.2.1: Select quant_vec q Replace any k dimensions in basis_vector with the corresponding values in basis_vector to form a new vector. The selection method of these k dimensions is denoted as s. The result is the replaced vector, denoted as replace_vecq_k_s ;
[0107] Step 3.2.2: replace_vec q_k_s The k-th dimension in the replaced dimension is replaced by quant_vec q The value of the corresponding bit in the new vector is recorded as replace_recover_vec q_k_s_x The gain in estimated probability due to using the true value of the xth dimension can be calculated:
[0108] lift q_k_s_x =replace_recover_vec q_k_s_x -replace_vec q_k_s ;
[0109] Step 3.2.3: Repeat steps 3.2.1 to 3.2.2 to calculate the gain of the xth dimension for various choices of k and calculate the sample q The average gain brought by the xth dimension is:
[0110] avg_lift q_x =
[0111] Step 3.2.4: Repeat steps 3.2.1 to 3.2.3 for sample q Calculate the average gain when x is each dimension, and then get the gain rate of each dimension:
[0112] lift_rate q_x =
[0113] Step 3.2.5: Sort each lift_rate in descending order q_x , get the maximum T dimensions, these T dimensions are the model judgment sample q The T factors with the greatest abnormal impact, the meaning of each factor can be found from the dimension information table (feature_name_dict, constructed in step 2.2), so as to obtain the judgment and explanation of the current fault. For example, the largest T dimensions include dimension j, lift_rate q_j The value of dimension j is 0.3, the value of dimension j is 0.95, and the name of dimension j is "disk read / write volume-steep increase". Based on the above, we can get the explanation: the estimated probability that the current sample meets the "disk read / write volume-steep increase" is 95%, which is very important for judging sample q The contribution to operation and maintenance failures is 30%.
[0114] The present invention also provides an explainable operation and maintenance fault identification system generated based on quantitative business features. The explainable operation and maintenance fault identification system generated based on quantitative business features can be realized by executing the process steps of the explainable operation and maintenance fault identification method generated based on quantitative business features, that is, those skilled in the art can understand the explainable operation and maintenance fault identification method generated based on quantitative business features as a preferred implementation mode of the explainable operation and maintenance fault identification system generated based on quantitative business features.
[0115] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.
[0116] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.
Claims
1. A method for identifying explainable operation and maintenance faults based on the generation of quantitative business features, characterized in that: include: Step S1: construct a business feature annotation dataset, and construct and train a probability estimation model for each business feature based on the business feature annotation dataset; Step S2: construct a fault annotation data set, and construct a quantized service feature vector of the fault annotation data set based on the trained probability estimation model of each service feature; Step S3: construct a fault detection deep model, and train the fault detection deep model using the quantized service feature vector of the constructed fault annotation data set to obtain a trained fault detection deep model; Step S4: Use the trained fault detection deep model to judge the fault probability of the business indicator data in the operation and maintenance scenario to be detected. When the business indicator data in the operation and maintenance scenario to be detected is abnormal, the abnormality is explained; The construction of the business feature annotation data set is based on the collected business indicator data in the operation and maintenance scenario, and the typical business features appearing in the business indicator data are annotated to obtain the business feature annotation data set; The fault annotation data set is constructed based on the collected business indicator data in the operation and maintenance scenario, and the operation and maintenance faults that occur are annotated to obtain the fault annotation data set; The probability estimation model of the business feature is to obtain the probability of the business feature existing in each scenario through a machine learning model; The fault detection deep model is to use a deep neural network to predict the probability of a fault based on the scenario; The step S4 adopts: Step S4.1: collect business indicator data in the operation and maintenance scenarios to be tested, align the business indicator data in all the operation and maintenance scenarios to be tested according to time, and divide them into samples using a sliding window method; Step S4.2: Convert the business indicator data samples in the operation and maintenance scenario to be detected into quantitative business feature vectors, and input them into the trained fault detection deep model to obtain the probability of whether the current input is a fault, recorded as p_anomaly; when p_anomaly is greater than the preset value, it is considered that the sample corresponding to the current time window is abnormal; Step S4.3: When there is an abnormality in the sample corresponding to the current time window, a fault explanation is performed on the sample corresponding to the current time window; The step S4.2 adopts: Step S4.2.1: Failure probability is p_anomaly q The sample quantization business feature vector is: quant_vec q =[f1,f2,f3…f mxn-1 ,f mxn ]; Step S4.2.2: Select quant_vec q Replace any k dimensions in basis_vector with the corresponding values in basis_vector to form a new vector. The selection method of the current k dimensions is recorded as s. The result is the replaced vector, recorded as replace_vec q_k_s ; Wherein, basis_vector represents the base vector obtained by finding the median of the quantized service feature vectors of all samples bit by bit; Step S4.2.3: replace_vec q_k_s The xth dimension of the k dimensions replaced by quant_vec q The value of the corresponding bit in the new vector is recorded as replace_recover_vec q_k_s_x ; Calculate the gain in estimated probability due to using the true value of the xth dimension: lift q_k_s_x =replace_recover_vec q_k_s_x -replace_vec q_k_s ; Repeat steps 4.2.2 to 4.2.3 to calculate the gain of the xth dimension when k is selected for any size and various choices of k, and calculate the sample sample q The average gain brought by the xth dimension is: avg_lift q_x = in, represents the number of dimensions to be randomly selected for replacement, Indicates the method number of the current k dimensions randomly selected, Indicates the number of business indicators, Indicates the number of business features, represents the number of ways to arbitrarily select i dimensions from the remaining dimensions while retaining the xth dimension; Repeat steps 4.2.2 to 4.2.3 for sample q Calculate the average gain when x is each dimension, and then get the gain rate of each dimension: lift_rate q_x = Step 4.2.4: Sort each lift_rate in descending order q_x , get the maximum T dimensions, the current T dimensions are the fault detection deep model judgment samples sample q These are the T factors with the greatest abnormal impact.
2. The explainable operation and maintenance fault identification method based on quantified business feature generation according to claim 1 is characterized in that: The step S1 adopts: Step S1.1: Collect business indicator data in the operation and maintenance scenario, including: CPU usage, memory usage, disk read and write volume, network data transmission volume, network data reception volume, and business success rate; mark typical business features appearing in the business indicator data, including: stable, steep rise, steep drop, slow rise, slow drop, large fluctuation, peak value, and valley value, to obtain a business feature annotation data set; Step S1.2: Select a non-repeated business label from the business feature labeling dataset and record the selected business label as current_label k ; Step S1.3: Set the label to current_label k The sample label of is 1, and the labels of the remaining samples are 0, forming a training data set, denoted as sample_set k ; Step S1.4: Build a business feature probability estimation model and use the training data set sample_set k Train the constructed business feature probability estimation model to obtain the trained business feature probability estimation model prob_model k ; Repeat triggering steps S1.2 to S1.4 until all business tags in the business feature annotation data set have constructed corresponding business feature probability estimation models.
3. The explainable operation and maintenance fault identification method based on quantified business feature generation according to claim 2 is characterized in that: The step S1.1 adopts: Step S1.1.1: All business indicators are organized into multiple time series and normalized; Step S1.1.2: Divide the time series into samples using a sliding window method; Step S1.1.3: Manually label each sample with business phenomena.
4. The explainable operation and maintenance fault identification method based on quantified business feature generation according to claim 1 is characterized in that: The step S2 adopts: Step S2.1: Annotate the operation and maintenance faults that occur based on the collected business indicator data in the operation and maintenance scenario, and construct an operation and maintenance fault training data sample; Step S2.2: Each operation and maintenance fault training data sample contains all business indicator data within a time window, and each business indicator is estimated using the trained business feature probability estimation model; including: m business indicators and n business feature probability estimation models, converting the m business indicators of each operation and maintenance fault training data sample into a vector of length m×n, where each dimension is a numerical value with a clear business meaning, and recording the names of all dimensions to obtain a dimension information table.
5. The explainable operation and maintenance fault identification method based on quantified business feature generation according to claim 4 is characterized in that: The step S2.1 adopts: aligning the business indicator data in all operation and maintenance scenarios according to time, and dividing them into samples using the sliding window method; marking the divided sample results, and marking 1 if the time range of the sample contains an operation and maintenance failure, otherwise it is marked as 0.
6. The explainable operation and maintenance fault identification method based on quantified business feature generation according to claim 1 is characterized in that: The step S3 adopts: constructing a fault detection deep model, the input of the fault detection deep model is a quantized business feature vector of a fault annotation data set with a length of m×n, and the output of the fault detection deep model is the probability of whether there is a fault, 1 indicates that the probability of the input being a fault is 100%, and 0 indicates that the probability of the input being a fault is 0; The fault detection deep model is trained using the quantized business feature vector of the fault annotation data set to obtain a trained fault detection deep model.
7. An explainable operation and maintenance fault identification system based on quantitative business feature generation, characterized in that: include: Module M1: Construct a business feature annotation dataset, and build and train a probability estimation model for each business feature based on the business feature annotation dataset; Module M2: Construct a fault labeling dataset, and construct a quantitative service feature vector of the fault labeling dataset based on the trained probability estimation model of each service feature; Module M3: construct a fault detection deep model, and use the quantized business feature vector of the constructed fault labeling data set to train the fault detection deep model to obtain a trained fault detection deep model; Module M4: Use the trained fault detection deep model to judge the fault probability of the business indicator data in the operation and maintenance scenario to be detected. When the business indicator data in the operation and maintenance scenario to be detected is abnormal, the abnormality is explained; The construction of the business feature annotation data set is based on the collected business indicator data in the operation and maintenance scenario, and the typical business features appearing in the business indicator data are annotated to obtain the business feature annotation data set; The fault annotation data set is constructed based on the collected business indicator data in the operation and maintenance scenario, and the operation and maintenance faults that occur are annotated to obtain the fault annotation data set; The probability estimation model of the business feature is to obtain the probability of the business feature existing in each scenario through a machine learning model; The fault detection deep model is to use a deep neural network to predict the probability of a fault based on the scenario; The module M4 adopts: Module M4.1: Collect business indicator data in the operation and maintenance scenarios to be tested, align the business indicator data in all the operation and maintenance scenarios to be tested according to time, and divide them into samples using the sliding window method; Module M4.2: Convert the business indicator data samples in the operation and maintenance scenario to be detected into quantitative business feature vectors, and input them into the trained fault detection deep model to obtain the probability of whether the current input is a fault, recorded as p_anomaly; when p_anomaly is greater than the preset value, it is considered that the sample corresponding to the current time window is abnormal; Module M4.3: When there is an abnormality in the sample corresponding to the current time window, a fault explanation is performed on the sample corresponding to the current time window; The module M4.2 adopts: Module M4.2.1: Failure probability p_anomaly q The sample quantization business feature vector is: quant_vec q =[f1,f2,f3…f mxn-1 ,f mxn ]; Module M4.2.2: Select quant_vec q Replace any k dimensions in basis_vector with the corresponding values in basis_vector to form a new vector. The selection method of the current k dimensions is recorded as s. The result is the replaced vector, recorded as replace_vec q_k_s ; Wherein, basis_vector represents the base vector obtained by finding the median of the quantized service feature vectors of all samples bit by bit; Module M4.2.3: replace_vec q_k_s The xth dimension of the k dimensions replaced by quant_vec q The value of the corresponding bit in the new vector is recorded as replace_recover_vec q_k_s_x ; Calculate the gain in estimated probability due to using the true value of the xth dimension: lift q_k_s_x =replace_recover_vec q_k_s_x -replace_vec q_k_s ; Repeat steps 4.2.2 to 4.2.3 to calculate the gain of the xth dimension when k is selected for any size and various choices of k, and calculate the sample sample q The average gain brought by the xth dimension is: avg_lift q_x = in, represents the number of dimensions to be randomly selected for replacement, Indicates the method number of the current k dimensions randomly selected, Indicates the number of business indicators, Indicates the number of business features, represents the number of ways to arbitrarily select i dimensions from the remaining dimensions while retaining the xth dimension; Repeat steps 4.2.2 to 4.2.3 for sample q Calculate the average gain when x is each dimension, and then get the gain rate of each dimension: lift_rate q_x = Step 4.2.4: Sort each lift_rate in descending order q_x , get the maximum T dimensions, the current T dimensions are the fault detection deep model judgment samples sample q These are the T factors with the greatest abnormal impact.
8. The explainable operation and maintenance fault identification system based on quantified business feature generation according to claim 7 is characterized in that: The module M1 adopts: Module M1.1: Collect business indicator data in operation and maintenance scenarios, including: CPU usage, memory usage, disk read and write volume, network data transmission volume, network data reception volume, and business success rate; mark typical business features that appear in business indicator data, including: stable, steep rise, steep drop, slow rise, slow drop, large fluctuations, peak value, and trough, to obtain a business feature annotation data set; Module M1.2: Select a business label from the business feature labeling dataset without duplication, and record the selected business label as current_label k ; Module M1.3: Set label to current_label k The sample label of is 1, and the labels of the remaining samples are 0, forming a training data set, denoted as sample_set k ; Module M1.4: Construct a business feature probability estimation model and use the training data set sample_set k Train the constructed business feature probability estimation model to obtain the trained business feature probability estimation model prob_model k ; Repeat triggering modules M1.2 to M1.4 until all business tags in the business feature annotation data set have built corresponding business feature probability estimation models; The module M2 adopts: Module M2.1: Annotate the operation and maintenance faults that occur based on the collected business indicator data in the operation and maintenance scenarios, and construct operation and maintenance fault training data samples; Module M2.2: Each operation and maintenance fault training data sample contains all business indicator data within a time window, and each business indicator is estimated using the trained business feature probability estimation model; including: m business indicators and n business feature probability estimation models, converting the m business indicators of each operation and maintenance fault training data sample into a vector of length m×n, where each dimension is a value with a clear business meaning, and recording the names of all dimensions to obtain a dimension information table; The module M3 adopts: constructing a fault detection deep model, the input of which is a quantized service feature vector of a fault annotation data set of length m×n, and the output of which is a fault probability, 1 indicates that the probability of the input being a fault is 100%, and 0 indicates that the probability of the input being a fault is 0; The fault detection deep model is trained using the quantized business feature vector of the fault annotation data set to obtain a trained fault detection deep model.
Citation Information
Patent Citations
Machine learning model interpretation method and device and storage medium
CN114792138A
Root cause positioning method and device for key performance index abnormity
CN115550977A
System fault diagnosis intelligent operation and maintenance method and system based on machine learning
CN109800127A
Operation and maintenance fault chain labeling system and method based on self-evolution network knowledge base
CN116094955A
Remote fault detection method and detection device for energy storage charging equipment
CN117591863A
