A method and apparatus for calibrating a network digital model

Abstract

The application discloses a kind of correction method and correction device of network digital model, comprising: the input data of sampling data and / or calculation upstream model is calculated by the target algorithm selected by itself digital model to obtain calculation result;According to normal value set and abnormal value set, update reference interval, according to reference interval, the calculation result is identified;If calculation result falls outside the reference interval, the calculation result is suspicious data, request information with calculation result is sent to calculate downstream model, and the calculation result is identified by calculation downstream model, until the calculation result is identified as normal data or abnormal data;If calculation result is abnormal data, calculation result is added to abnormal value set;Adjust the fitting parameter of target algorithm, if after the adjustment of pre-set number of times, at least once normal value cannot be obtained, trigger itself digital model to reselect target algorithm.

Description

Technical Field

[0001] The present invention belongs to the field of digital networks, and more particularly, relates to a correction method and a correction device for a network digital model. Background Art

[0002] With the advancement of digitalization and informatization across all elements of society, communication networks, as fundamental elements for information connectivity and transmission, are also booming in this wave of digitalization. Mathematical modeling techniques (such as white-box modeling based on physical mechanisms and black-box modeling based on neural networks) enable the description of network-physical processes and changes. Sensor probe technology enables the real-time operational status of network facilities, enabling the precise characterization of the operational processes of network-physical devices.

[0003] However, while numerous digital models exist for network devices, and most demonstrate satisfactory accuracy in simulation or testing environments, these models face significant limitations in adaptability and generalization when applied to real-world scenarios requiring precise control, high reliability, and high real-time performance. This makes it difficult to guarantee model accuracy. Furthermore, network devices experience aging over time and due to environmental factors, and if the digital models cannot update their representations of these physical phenomena, the model's accuracy and generalization capabilities will decline. Furthermore, probe sensing can also degrade or malfunction, affecting their accuracy and, in turn, the model's representation accuracy. These factors ultimately lead to a decline in model credibility, hindering the rapid development of digital and autonomous networks. Summary of the Invention

[0004] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a correction method and correction device for a network digital model, the purpose of which is to combine the dynamic interval update method and the fitting parameter direction correction method to achieve the dynamic, rapid and autonomous correction capability of the model, thereby solving the technical problem that the digital model cannot be updated according to the physical equipment or environment, resulting in weak generalization ability of the digital model.

[0005] To achieve the above object, according to one aspect of the present invention, a method for correcting a network digital model is provided, comprising:

[0006] The target algorithm selected by the digital model itself is used to calculate the sampled data and / or the input data input by the upstream model to obtain the calculation results;

[0007] updating a reference interval according to a normal value set and an abnormal value set, and identifying a calculation result according to the reference interval;

[0008] If the calculation result falls outside the reference interval, the calculation result is suspicious data, request information carrying the calculation result is sent to a calculation downstream model, the calculation result is identified by the calculation downstream model until the calculation result is identified as normal data or abnormal data;

[0009] If the calculation result is abnormal data, the calculation result is added to an abnormal value set to update the reference interval according to the abnormal value set and the normal value set;

[0010] Adjust the fitting parameters of the target algorithm, and if at least one normal value cannot be obtained after a preset number of adjustments, trigger the self-digital model to reselect the target algorithm.

[0011] Further, the correction method further comprises:

[0012] If the calculation result falls within the reference interval, the calculation result is normal data, and the calculation result is added to a normal value set to update the reference interval according to the abnormal value set and the normal value set.

[0013] Further, the process of adjusting the fitting parameters of the target algorithm comprises:

[0014] According to the exploration efficiency, an adjustment step of the fitting parameters is selected, the fitting parameters are adjusted according to the adjustment step, and a first fitting parameter combination is obtained;

[0015] The first fitting parameter combination is updated into the target algorithm, and a trial result B is obtained by trial calculation using the updated target algorithm;

[0016] If the state of the trial result B is the same as that of the result A before the trial calculation, the fitting parameters are continuously adjusted according to the selected adjustment step until the state of the trial result B is opposite to that of the result A before the trial calculation;

[0017] The fitting parameter combination corresponding to the target algorithm in the previous trial calculation is adjusted using the reverse bisection method, a second fitting parameter combination is obtained, the second fitting parameter combination is updated into the target algorithm, and a trial result C is obtained by trial calculation using the updated target algorithm;

[0018] If the state of the trial result C is the same as that of the trial result B before the trial calculation, the fitting parameters are continuously adjusted according to the reverse bisection method until the state of the trial result C is opposite to that of the trial result B, and the fitting parameters are updated into the fitting parameter list in the algorithm description.

[0019] Further, the adjustment step of the fitting parameters according to the exploration efficiency comprises:

[0020] The number of explorations corresponding to the change of the trial calculation result from abnormal data to normal data is obtained. If the number of explorations is less than the set threshold, the adjustment step is reduced; if the number of explorations is greater than the set threshold, the adjustment step is increased.

[0021] Furthermore, the reverse dichotomy method is used to adjust the fitting parameters corresponding to the target algorithm in the previous trial calculation to obtain the second fitting parameter combination including:

[0022] If the adjustment step corresponding to the previous trial calculation is greater than 1, the second fitting parameter combination is half of the fitting parameters corresponding to the target algorithm in the previous trial calculation;

[0023] If the adjustment step corresponding to the previous trial calculation is less than 1, the second fitting parameter combination is twice the fitting parameters corresponding to the target algorithm in the previous trial calculation.

[0024] Furthermore, the outlier set includes a first outlier set and a second outlier set, data in the second outlier set are all greater than data in the first outlier set, and updating the reference interval according to the normal value set and the outlier set includes:

[0025] Determining a confidence interval based on a predetermined confidence level and n sample data, and determining a reference interval based on the confidence interval and a predetermined k-σ criterion;

[0026] Performing abnormality identification on the indicator data to be identified according to the reference interval, and selectively adding the corresponding indicator data to the normal value set, the first abnormal value set, or the second abnormal value set according to the identification result;

[0027] Update the mean based on the normal value set and variance σ, update the mean based on the first set of outliers and variance σ', update the mean based on the second set of outliers and variance σ”;

[0028] when When , the lower bound of the normal value is updated to when When the normal value upper bound is updated to update the reference interval.

[0029] Furthermore, determining the confidence interval according to a predetermined confidence level and n sample data, and determining the reference interval according to the confidence interval and a predetermined k-σ criterion includes:

[0030] Get the mean x and variance σ of n sample data, according to the mean Calculate the confidence interval using the variance σ, the confidence constant z, and the number of sample data n

[0031] determining an initial reference interval according to the confidence interval and a predetermined k-σ criterion

[0032] wherein in the subsequent process, when any one of the normal value set, the first abnormal value set or the second abnormal value set is changed, the mean and variance of the corresponding set are updated to update the reference interval.

[0033] Further, the selectively adding the corresponding index data into the normal value set, the first abnormal value set or the second abnormal value set according to the identification result comprises:

[0034] if the index data is less than the index data is added to the first abnormal value set; if the index data is greater than the index data is added to the second abnormal value set, and if the index data falls into the initial value interval, the index data is added to the normal value set.

[0035] Further, the correction method further comprises:

[0036] if the abnormal situation still cannot be eliminated by modifying the fitting parameters and reselecting the algorithm, all the alternative models corresponding to the carrier are obtained, and the algorithm characterization ability of the algorithm corresponding to each alternative model, the model use frequency in each time period and the parameter list satisfaction degree are obtained;

[0037] the product of the algorithm characterization ability, the model use frequency and the parameter list satisfaction degree is obtained, and the product is taken as the model ability of the alternative model;

[0038] the model abilities are sorted in descending order, and the selection is started from the largest model ability until a suitable model is selected.

[0039] According to another aspect of the present application, a correction device of a network digital model is provided, comprising at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions executable by the at least one processor, and the instructions are used to complete the correction method of the network digital model according to the first aspect after being executed by the processor.

[0040] Overall, compared with the prior art, the above technical solutions conceived by the present application have the following beneficial effects: the correction method of the embodiments of the present application can realize the dynamic and rapid self-correction ability of the model by combining the dynamic interval updating method, and the automatic correction and adaptation ability of the model under the change of itself and the environment is improved by combining the model correction method of the reference interval updating, thereby improving the generalization performance of the digital model. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 This is a flow chart of a method for correcting a network digital model provided by an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the overall framework of the correction of a network digital model provided by the present invention;

[0043] Figure 3 Schematic diagram of the optical network equipment through which the OCH channel passes provided by an embodiment of the present invention;

[0044] Figure 4 The embodiment of the present invention provides Figure 3 Schematic diagram of the relationship between the digital models of physical devices such as amplifiers and optical fibers corresponding to the OCH channel and the optical transmission section, optical multiplexing section, and optical channel models;

[0045] Figure 5 The embodiment of the present invention provides Figure 1 Specific flow diagram of step 40;

[0046] Figure 6 Schematic diagram of the updating process of the normal data set and the abnormal data set of the network digital model provided by the embodiment of the present invention;

[0047] Figure 7 Schematic diagram of the steps for adjusting the network digital model fitting parameters provided by an embodiment of the present invention;

[0048] Figure 8 Schematic diagram of the reference interval updating steps of the network digital model provided by an embodiment of the present invention;

[0049] Figure 9 Schematic diagram of reverse trial and error of a network digital model provided by an embodiment of the present invention;

[0050] Figure 10 Schematic diagram of autonomous negotiation correction of a network digital model provided by an embodiment of the present invention;

[0051] Figure 11 The figure is a schematic structural diagram of a correction device for a network digital model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0053] Example 1:

[0054] To solve the above problems, this embodiment provides a method for correcting a network digital model. Figure 1 , the correction method comprises the following steps:

[0055] Step 10: Calculate the sampled data and / or the input data input by the upstream model using the target algorithm selected by the digital model itself to obtain the calculation results.

[0056] In this embodiment, the data source of the self-digital model includes at least one of sampling data collected by the self-probe component and input data input by the upstream calculation model.

[0057] Among them, each model can correspond to multiple algorithms, and the model can select a suitable algorithm from multiple algorithms as the target algorithm of the model (the target algorithm can be understood as the currently used algorithm);

[0058] For models that support dynamic data collection, each model corresponds to at least one probe component, and the sampled data is the data dynamically collected by the probe component. Of course, in other embodiments, some models may not have a probe component. In this case, the data source of the digital model includes the input data input by the upstream model, rather than the sampled data collected by the probe component itself.

[0059] Before explaining the calculation of the upstream model, we first introduce the dependencies between models and how to establish the interaction channels between models. The details are as follows:

[0060] There are dependencies between different models, which include order dependencies and nested dependencies. In actual application scenarios, the dependency between target models can be constructed based on the business flow and the topological relationship, and a model identifier can be set for each model. The computing upstream model that is dependent on its own model is determined based on the dependency relationship, and the model identifier of its own model is registered in the computing upstream model to construct a message channel between target models. Different models send interaction messages through the message channel, where the interaction messages include request messages, response messages and notification messages.

[0061] Specifically, when order dependency is determined based on the service flow relationship of the bearer, models 1 to n are sequential and independent of each other, and do not nest within each other. The input of model n is provided by the output of model n-1. After model n processes all input data, it can notify model n-1 of the calculation results and status. The input of model n-1 is provided by the output of model n-2. After model n-1 processes all input data, it can notify model n-2 of the calculation results and status. Similarly, the input of model 2 is provided by the output of model 1. After model 2 processes all input data, it can notify model 1 of the calculation results and status.

[0062] When a dependency is nested, a nested relationship is formed between the model and its sub-models. The model's input is provided by the outputs of multiple sub-models. When organizing models, the model and sub-models establish a message exchange relationship, optionally carrying requirements such as model capability and algorithm capabilities with the request message to the sub-model.

[0063] For multiple target models with nested dependencies, the child model registers its model identifier with the parent model, and the parent model stores the received model identifier in the registration list; the parent model is the target model at the upper level of the nested tree structure, and the child model is the model at the lower level of the nested tree structure;

[0064] For multiple target models with sequential dependencies, the upstream computing model registers its model identifier with the downstream computing model, and the downstream computing model stores the received model identifier in the registration list; wherein the output of the upstream computing model serves as the input of the downstream computing model;

[0065] The upstream computing model and the downstream computing model exchange messages, such as through the TCP protocol, and the upstream computing model registers its model identifier to the downstream computing model.

[0066] After building the dependencies and interaction channels between models in the aforementioned manner, the models will be able to interact with messages and data.

[0067] In case of sequential dependency, the starting side of the business is the upstream side, and the ending side of the business is the downstream side. In case of nested dependency, the parent model side is the upper layer, and the child model side is the lower layer.

[0068] Based on this, we define upstream and downstream computations: upstream is the upstream in an ordered dependency or the lower level in a nested dependency, and downstream is the downstream in an ordered dependency or the upper level in a nested dependency. The upstream computation model is the model corresponding to the upstream computation, and the downstream computation model is the model corresponding to the downstream computation.

[0069] by Figure 3For example, the order dependency is shown. The business flow is from transmitter to receiver. Relative to the multiplexer, the model corresponding to the transmitter is the computational upstream model of the model corresponding to the multiplexer, and the model corresponding to the multiplexer is the computational downstream model of the model corresponding to the transmitter.

[0070] by Figure 4 For example, nested dependencies are shown for the optical multiplexing segment model. OMS In terms of optical multiplexing section model OMS Depends on the optical transmission segment model OTS1 and optical transmission segment model OTS2 , Optical Multiplexing Section Model OMS Located in the upper layer of nested dependencies, the optical transmission segment model Model OTS1 and optical transmission segment model OTS2 Located at the lower level of nested dependencies, the optical multiplexing segment model Model OMS Model for the optical transmission segment OTS1 and optical transmission segment model OTS2 Calculation downstream model, optical transmission segment model Model OTS1 and optical transmission segment model OTS2 Optical multiplexing section model OMS The computational upstream model.

[0071] The calculation result is the instantiation value of the performance metric corresponding to the model. For example, if the performance metric of the model is OSNR (Optical Signal to Noise Ratio), the calculation result is the instantiation value of the OSNR of the model.

[0072] Step 20: Update the reference interval according to the normal value set and the abnormal value set, and identify the calculation result according to the reference interval.

[0073] In actual application scenarios, since physical equipment will gradually age over time and in response to changes in the environment during its self-operation, if the digital model cannot update its depiction of these physical phenomena, the model's accuracy and generalization capabilities will be further reduced. In terms of probe perception, the probe components themselves may also degrade or fail, affecting the accuracy of the probes, which in turn directly affects the reference interval of the model's depiction accuracy. Therefore, the reference interval needs to be continuously and dynamically updated. In order to adapt to the actual application of physical equipment and probe components, it is necessary to use a large amount of data information, including real-time perception measurement data of network status, calibration data at different time periods, and historical data from various stages of network operation, to update the reference interval, so as to improve the model's application generalization capabilities in ever-changing environments and scenarios and enhance the model's credibility.

[0074] In the embodiment, the sampling data is real-time perception data of the probe on the network state, the labeling data (normal, suspicious or abnormal) of the calculation result based on the reference interval is the labeling data of different time periods, and the historical data of each stage of network operation and a large amount of data information are the historical normal data set and the abnormal data set. Therefore, the normal value set and the abnormal value set are constantly replenished, and after the normal value set and the abnormal value set change, the reference interval may also change.

[0075] The current calculation result of the model is labeled in combination with the reference interval, so as to add the calculation result to the corresponding normal or abnormal value set according to the identification result of the calculation result, and dynamically update the reference interval, so that the reference interval can change with the change of the environment, the device and the application scenario, support the normal interval update of the physical device in the slow change process and the rationality verification of the calculation result.

[0076] If the calculation result falls outside the reference interval, the calculation result is suspicious data, and request information carrying the calculation result is sent to a calculation downstream model, and the calculation result is identified by the calculation downstream model until the calculation result is identified as normal data or abnormal data.

[0077] If the calculation result falls outside the reference interval, the calculation result is suspicious data, and request information carrying the calculation result is sent to a calculation downstream model, and the calculation result is identified by the calculation downstream model until the calculation result is identified as normal data or abnormal data.

[0078] If the calculation result falls within the reference interval, the calculation result is normal data, and the calculation result is added to the normal value set, so as to update the reference interval according to the abnormal value set and the normal value set.

[0079] For the model, if the calculation result can be identified as normal data by itself, the reference interval is updated according to the calculation result, and the specific updating process is described below. If the calculation result cannot be identified as normal data by itself, the calculation result is temporarily marked as suspicious data (which means that the model cannot determine whether the data is normal or abnormal), and the downstream calculation model is used for judgment.

[0080] In actual application scenarios, when the model itself cannot determine whether the data is abnormal or normal, the downstream calculation model is used for identification until the topmost layer is reached. Therefore, the topmost layer can be determined by simple judgment or even artificial judgment.

[0081] If the identification result of the calculation downstream model is normal data, the calculation result is corrected to normal data, and the calculation result is added to the normal value set, so as to update the reference interval according to the abnormal value set and the normal value set;

[0082] If the recognition result of the downstream computing model is still suspicious data, the calculation continues to the downstream computing model until a model that can be judged is reached, or until the end-to-end model makes a judgment.

[0083] To enable direct feedback of end-to-end decision results to this model, a message channel can be established between this model and the end-to-end model, achieving one-hop communication. Alternatively, the request-response fusion process, an iterative recursive process, can be used, though it is less convenient and direct.

[0084] Step 40: If the calculation result is abnormal data, the calculation result is added to the abnormal value set, so as to update the reference interval according to the abnormal value set and the normal value set.

[0085] If the downstream calculation model returns abnormal data as a result, the result is added to the outlier set, thereby updating the reference interval based on the outlier set and the normal value set. This indicates that the model's algorithm may not be suitable for the current environment and needs to be adjusted. See step 50 for details.

[0086] Step 50: Adjust the fitting parameters of the target algorithm. If at least one normal value cannot be obtained after a preset number of adjustments, the self-digital model is triggered to reselect the target algorithm.

[0087] In a preferred embodiment, since part of the data source corresponding to the calculation result comes from the probe component, it is necessary to eliminate the abnormality caused by the abnormality of the probe component. Therefore, if the calculation result is abnormal data, first detect whether the probe component is abnormal. Specifically, it can be determined whether the probe component is abnormal by detecting whether the sampling data is within the nominal range; if the probe component is abnormal, the input parameters provided to the corresponding algorithm are abnormal, and the current algorithm cannot run normally, then it is impossible to calculate normal data according to the current algorithm, and the self-digital model is triggered to reselect the target algorithm to restore the model as soon as possible; if the probe component is normal, first correct the fitting parameters of the target algorithm. If the calculation result is still abnormal data after correction, reselect the algorithm.

[0088] Of course, in actual application scenarios, you can also replace the probe to ensure the normal operation of the model, but this method takes a long time and has poor real-time performance. You can first trigger the reselection of the target algorithm to ensure the operation of the model, and then replace the probe.

[0089] In this embodiment, the algorithm is adjusted in two main cases, as follows:

[0090] (1) First, adjust the fitting parameters of the target algorithm. If the calculation result is normal data after a preset number of adjustments, the corresponding fitting parameters are updated to the target algorithm.

[0091] (2) If at least one normal value cannot be obtained after a preset number of adjustments, it indicates that the self-digital model is triggered to reselect the target algorithm.

[0092] For example, for the algorithm, y = a1·x1+a2·x2+b, x1 and x2 are parameters, which are updated through input methods such as sampling, and y is updated through function solution; a1, a2, and b are fitting parameters, which generally do not change under specific accuracy and application environment. However, when the application scenario and accuracy are changed, it is necessary to adjust a1, a2, and b to achieve the purpose of fine-tuning the function equation to adapt to the new environment and accuracy requirements.

[0093] If the abnormality cannot be eliminated by correcting the fitting parameters and reselecting the algorithm, it is necessary to reselect the model. Specifically, each carrier corresponds to at least one candidate model, and a suitable model is selected from the candidate models as the target model corresponding to the carrier as needed. Specifically, the algorithm characterization capability, model usage frequency in each time period, and parameter list satisfaction of the algorithm corresponding to each candidate model are obtained; the product of the algorithm characterization capability, the model usage frequency, and the parameter list satisfaction is obtained, and this product is used as the model capability of the candidate model; the models are sorted in descending order of capability, starting with the model with the largest capability, until a suitable model is selected. The model usage frequency is obtained by obtaining the number of model uses n in the time period T, wherein each time the model is used, the number of model uses n is increased by one; each time the model is abandoned, the number of model uses n is reduced by one; after each time period, the number of model uses is reset and counted again; the ratio of the number of model uses to the time period is calculated, and the value corresponding to the ratio plus the random number δ is used as the model usage frequency. That is, the model usage frequency can be calculated according to the following formula: model_usage_frequency = n / T + δ, where the random number δ carried in the model usage frequency can ensure that each model has the opportunity to be used. The parameter list satisfaction is obtained in the following manner: obtaining the performance metric to be achieved and determining the minimum parameter set based on the performance metric; obtaining the probe component of each candidate model and determining whether the detection accuracy of each probe component meets the preset accuracy. If so, adding the parameter type corresponding to the probe component to the measurement parameter list; determining whether all parameter types included in the minimum parameter set exist in the measurement parameter list; if all parameter types included in the minimum parameter set exist in the measurement parameter list, the parameter list satisfaction is 1; if at least one parameter type does not exist in the measurement parameter list, the parameter list satisfaction is 0.

[0094] That is, in this embodiment, the correction of the model includes reselecting the algorithm, updating the fitting parameters of the algorithm, and reselecting the model. The corresponding correction method can be selected according to the actual situation.

[0095] In general, the calibration method of this embodiment can be roughly divided into the following three steps:

[0096] 1. Determine the normal value reference interval. Using the current historical data for calculation, first find the normal value reference interval of the model in the current environment and state, which can support normal value determination and classification. When abnormalities are suspicious, it is necessary to combine end-to-end determination as normal or abnormal to achieve positive abnormal value classification;

[0097] 2. Actively explore normal or outliers. Maintaining the model structure and algorithm, adjust each scenario data set and use positive outlier exploration methods to actively explore the upper and lower boundaries of the normal reference interval (this may require combining end-to-end closed-loop judgment). Utilize active exploration data and production data (real-world and operational data) to update the new normal reference interval.

[0098] 3. Fine-tune the fitting parameters and calibrate the model. Based on the model's real-time status and environment, use the newly generated reference interval and real-time data to fine-tune the fitting parameters to ensure the calculation results are within the newly generated reference interval. Alternatively, you can run the model several times with real-time data and then calculate the mean of the fitting parameters.

[0099] The following combination Figures 5 to 7 Explain how to adjust the algorithm's fitting parameters, including the following steps:

[0100] Step 401: selecting an adjustment step of the fitting parameters according to the exploration efficiency, and adjusting the fitting parameters according to the adjustment step to obtain a first fitting parameter combination;

[0101] Specifically, the number of explorations corresponding to the change of the trial calculation result from abnormal data to normal data is obtained. If the number of explorations is less than a set threshold, the adjustment step is reduced; if the number of explorations is greater than the set threshold, the adjustment step is increased.

[0102] Step 402: updating the first fitting parameter combination to the target algorithm, and performing a trial calculation using the updated target algorithm to obtain a trial calculation result B;

[0103] Step 403: If the trial calculation result B is the same as the result A before the trial calculation, continue adjusting the fitting parameters according to the selected adjustment step until the trial calculation result B is the opposite of the result A before the trial calculation;

[0104] If the trial calculation result B is the same as the result A before the trial calculation, the process returns to step 401 and updates the first fitting parameter combination. If the trial calculation result B is different from the result A before the trial calculation, the process jumps to step 404.

[0105] Step 404: Using the reverse dichotomy method, adjust the fitting parameters corresponding to the target algorithm in the previous trial calculation to obtain a second fitting parameter combination, update the second fitting parameter combination to the target algorithm, and perform a trial calculation using the updated target algorithm to obtain a trial calculation result C;

[0106] Among them, if the adjustment step corresponding to the previous trial calculation is greater than 1 (enlargement), the second fitting parameter combination is half of the fitting parameters corresponding to the target algorithm in the previous trial calculation; if the adjustment step corresponding to the previous trial calculation is less than 1 (reduction), the second fitting parameter combination is twice the fitting parameters corresponding to the target algorithm in the previous trial calculation.

[0107] Step 405: If the trial result C is the same as the trial result B before the trial calculation, continue to adjust the fitting parameters according to the reverse dichotomy method until the trial result C is opposite to the trial result B, and update the fitting parameters to the fitting parameter list in the algorithm description.

[0108] If the trial calculation result C is the same as the trial calculation result B before the trial calculation, then jump to step 404. Here, the trial calculation result status before and after the trial calculation is the same means: if the data before the calculation is abnormal, then the data after the calculation is abnormal; if the data before the calculation is normal, then the data after the calculation is normal. The calculation result status before and after the trial calculation is opposite means: if the data before the calculation is abnormal, then the data after the calculation is normal; if the data before the calculation is normal, then the data after the calculation is abnormal.

[0109] Steps 401 to 403 are forward fine-tuning, and steps 404 to 405 are reverse fine-tuning. Forward fine-tuning and reverse fine-tuning are understood as follows: forward fine-tuning, solving from the initial normal result to the abnormal result; or solving from the initial abnormal result to the normal result; reverse fine-tuning: solving from the abnormal result of forward fine-tuning to the normal result; or solving from the normal result of forward fine-tuning to the abnormal result.

[0110] In this embodiment, "adjusting the fitting parameters of the target algorithm" has two functions: 1. If the calculation result is abnormal data, the target algorithm is updated by "adjusting the fitting parameters of the target algorithm"; 2. Active exploration is performed using the fitting parameter combinations generated in this process to enrich the data set and obtain a better new reference interval based on the mean and variance.

[0111] The following combination Figure 8 Explain how the reference interval is updated:

[0112] In this embodiment, the outlier set includes a first outlier set (corresponding to Figure 8 The parameter abnormal statistical distribution on the left) and the second abnormal value set (corresponding to Figure 8 The parameter abnormal statistical distribution on the right side), the data in the second abnormal value set are all greater than the data in the first abnormal value set, the normal value set (corresponding to Figure 8 The parameter abnormal statistical distribution in the middle) and the reference interval for the abnormal value set update include:

[0113] Firstly, a confidence interval is determined according to a predetermined confidence level and n sample data, and a reference interval is determined according to the confidence interval and a predetermined k-σ criterion, the reference interval being an initial reference interval which is constantly updated.

[0114] wherein the confidence level is a constant at a given confidence level and can be obtained by table lookup. For example, when the confidence level is 95%, the Z value is 1.96.

[0115] wherein the k value in the k-σ criterion can be 3, i.e. 3σ criterion.

[0116] Specifically, the mean value and the variance σ of the n sample data are obtained. The confidence interval is calculated according to the mean value

[0117] The initial reference interval is determined according to the confidence interval and a predetermined k-σ criterion.

[0118] wherein in the subsequent process, when any one of the normal value set, the first abnormal value set or the second abnormal value set is changed, the mean value and the variance of the corresponding set are updated to update the reference interval.

[0119] Secondly, abnormal identification is performed on the to-be-identified index data according to the reference interval, and the corresponding index data is selectively added to the normal value set, the first abnormal value set or the second abnormal value set according to the identification result.

[0120] Specifically, if the index data is less than the index data is added to the first abnormal value set; if the index data is greater than the index data is added to the second abnormal value set, and if the index data falls within the initial value interval, the index data is added to the normal value set.

[0121] Then, the mean value and the variance σ are updated according to the normal value set, the mean value and the variance σ' are updated according to the first abnormal value set, and the mean value and the variance σ" are updated according to the second abnormal value set.

[0122] The calculation and updating method of the mean value and the variance is as follows:

[0123] x1, x2,..., x n n samples, the mean value and the variance of which are and When a new sample x n+1 is added, the updating method is as follows:

[0124]

[0125]

[0126] when When , the lower bound of the normal value is updated to when When the normal value upper bound is updated To update the reference interval. Different from the existing technology, this embodiment has at least the following advantages: it supports dynamic updates and rapid judgment of the normal or abnormality of the model on key indicators; it supports the update of the reference interval of a single parameter, the reference interval of a dual-parameter combination, and the reference interval of a complex combination of multiple parameters; compared with the existing technology, it can use each normal calculation to conduct trial and error to find the abnormal values ​​of the model parameters and supplement the abnormal value set; at the same time, it can support the dynamic correction and update of the algorithm fitting parameters in a fixed scenario. Compared with the existing technology, it can combine the dynamic interval update method and the fitting parameter direction correction method to realize the dynamic, rapid and autonomous correction capability of the model.

[0127] Example 2:

[0128] Different from Example 1, this embodiment also provides a method for updating the reference interval. In this embodiment, the reference interval can be updated in real time through active exploration methods and fitting parameter adjustment methods. Before introducing these two methods, it should be noted that the data sources for model calculation include: real-time dynamic data: real-time sampling data of probe components; real-time calculation data of the upstream model (i.e., input data of Example 1); non-real-time dynamic data, such as static configuration data, quasi-static configuration data, etc.

[0129] The normal value reference interval is only for the calculation output results of this model. The normal value reference interval is not constructed for the source data. Each model needs to be responsible for its own calculation results (output).

[0130] The basic basis of active exploration is to keep the model organizational structure unchanged, keep the selected algorithm unchanged, try to keep the fitting parameters unchanged, and mainly fine-tune the input source data.

[0131] During active exploration, the downstream calculation of the current model is updated in real time, but the upstream calculation data is not updated in real time. Instead, it is adjusted using fine-tuning methods. Specifically, the specific process of the positive outlier active exploration stage is as follows:

[0132] Fine-tune the source data, mainly calculate the upstream data, and all the history data up to the current time; calculate the upstream data, and through the vector or matrix to fine-tune. Then combine the calculation downstream and end-to-end model to determine the data calculated by the model as normal or abnormal, and mark the data as generated data (pseudo data), and add it to the normal and abnormal data set.

[0133] Update the fine-tune fitting parameter combination (as a supplement to the fine-tune source data method), real-time calculation of upstream data, and combine the calculation downstream and end-to-end model to determine the data calculated by the model as normal or abnormal.

[0134] Fine-tune the source data and update the fine-tune fitting parameter combination. After adjustment, only the calculation data is retained to the corresponding normal / abnormal value set. The data after fine-tuning the source data and the data after fine-tuning the fitting parameters are not retained.

[0135] Thus, the production environment data and the pseudo data generated by fine-tuning are obtained, and the production data and the generated data are used to update the normal value reference interval.

[0136] The differences between the active exploration method and the fitting parameter adjustment method are as follows:

[0137] Active exploration method: mainly through adjusting the input parameter value (auxiliary adjustment fitting parameter), to update the normal value set and the abnormal value set, and then update the normal value reference interval. When adjusting, keep the calculation upstream unchanged (i.e. the input data of the calculation upstream model remains unchanged); through the probe component of the model to obtain sampling data, the calculation downstream is updated in real time, the adjusted parameter value is not saved, only the normal value (generated data) and abnormal value (generated data) explored are saved, and added to the corresponding set, and the normal value reference interval is updated.

[0138] Fitting parameter adjustment method: need to update the output of the calculation upstream as the input of the model, then only adjust the fitting parameter, for normal value and abnormal value only for determination, and save the adjusted fitting parameter adjustment value, but the normal abnormal value after determination is not updated to the corresponding set.

[0139] The generated data obtained by the active exploration of normal abnormal value method is mainly used to enrich the data in the normal data set and the abnormal data set, and then update the reference interval.

[0140] It should be noted that in the active exploration stage, the supplement of the trusted data is the main target. The updating of the fine-tuned fitting parameter combination is mainly to use the latest fine-tuned fitting parameter combination that has been generated (a new fitting parameter combination will be generated after the fine-tuned fitting parameter is completed), and the purpose of doing so is to ensure the smooth continuity of the algorithm in the continuously changing scene and the credibility of the generated data. Therefore, the generated data here needs to be retained.

[0141] In the fitting parameter adjustment stage, the fine-tuned fitting parameter is the main target, so after the fitting parameter is adjusted here, there will be a lot of low-trust data, which all need to be discarded. The simplest method is to directly discard all the calculation result data in the adjustment of the fitting parameter process, and no longer add them to the normal / abnormal value set.

[0142] Therefore, in the active exploration stage, no new fitting parameter combination is generated; in the fine-tuned fitting parameter stage, no new trusted pseudo data is generated.

[0143] The following takes the fitting parameter adjustment method as an example to explain and describe:

[0144] When the calculation result is abnormal data,

[0145] (1) According to the exploration efficiency, the adjustment step of the fitting parameter is selected;

[0146] (2) The fitting parameter is adjusted according to the adjustment step to obtain a first fitting parameter combination; the first fitting parameter combination is updated to the target algorithm, and the trial calculation result is obtained by using the updated target algorithm for trial calculation;

[0147] (3) If the trial calculation result is normal data, record the calculation result, and jump to step (4). Otherwise, jump to step (1) to continue adjusting the fitting parameter according to the selected adjustment step until the trial calculation result is normal data;

[0148] (4) The fitting parameter corresponding to the target algorithm in the previous trial calculation is adjusted by using the reverse bisection method to obtain a second fitting parameter combination, and the second fitting parameter combination is updated to the target algorithm. The trial calculation result is obtained by using the updated target algorithm for trial calculation;

[0149] (5) Keep the model structure relationship unchanged and perform trial calculation;

[0150] (6) If the trial calculation result is still abnormal data, record the value and the fitting parameter combination, otherwise jump to step (4);

[0151] (7) Active exploration can be performed based on the existing fitting parameter combination.

[0152] When the calculation result is abnormal data,

[0153] (1) Selecting the adjustment step of fitting parameters according to the exploration efficiency;

[0154] (2) Adjusting the fitting parameters according to the adjustment step to obtain a first fitting parameter combination; updating the first fitting parameter combination to the target algorithm, and performing a trial calculation using the updated target algorithm to obtain a trial calculation result;

[0155] (3) If the trial calculation result is abnormal data, record the calculation result and jump to step (4). Otherwise, jump to step (1) to continue adjusting the fitting parameters according to the selected adjustment step until the trial calculation result is normal data;

[0156] (4) Adjusting the fitting parameters corresponding to the target algorithm in the previous trial calculation using the reverse bisection method to obtain a second fitting parameter combination, updating the second fitting parameter combination to the target algorithm, and performing a trial calculation using the updated target algorithm to obtain a trial calculation result;

[0157] (5) Keeping the model structure relationship unchanged, performing a trial calculation

[0158] (6) If the trial calculation result is still normal data, record the value and the fitting parameter combination, otherwise jump to step (4);

[0159] (7) Active exploration can be performed based on the existing fitting parameter combination.

[0160] The foregoing process is based on the existing reference interval to determine normal or abnormal. However, this normal or abnormal needs to be continuously explored, and a new reference interval based on the average value and variance can be obtained by continuously enriching the data set in the foregoing manner.

[0161] Example 3:

[0162] Unlike Example 1, in combination with Figure 9 and Figure 10 , the present embodiment also provides a process of inter-model autonomous negotiation correction for implementing step 40 in Example 1:

[0163] (1) Abnormal suspicious determination of the calculation result x of the model M0. When the abnormal rate is significantly increased or the normal interval is temporarily unavailable, it is determined and marked as abnormal suspicious; otherwise, it is directly determined according to the reference interval.

[0164] (2) Determining whether the probe data relied on by the algorithm currently selected by the model M0 is normal. When all the probe data are normal, jump to step (3), otherwise, adjust the bias of the probe. If it cannot be adjusted, the model reselects the algorithm, and the model notifies the dependent model to suspend calculation.

[0165] (3) Forward calculation. Initiate request by model M0, carry the calculation result x to participate in calculation, until the model Ma that can directly judge whether it is normal value or abnormal value. The message relationship is established between model Ma and M0. When the calculation result y of Ma is normal value, it is directly returned to model M0 through message, and the calculation result x is updated to normal. When the calculation result y of Ma is abnormal value, the calculation result x is updated to abnormal. If it is still abnormal suspicious, jump to step (4).

[0166] (4) Reverse trial and error. According to the fitting parameter adjustment method, adjust the fitting parameters of the dependent model {M2, M1,...} of M0, and perform trial calculation according to the fitting parameter adjustment method. The calculation result is y'. When the calculation result y' is normal value, it is directly returned to model M0 through message, and the calculation result x is updated to normal. When the calculation result y' is abnormal value, the calculation result x is updated to abnormal.

[0167] (5) When the fitting parameters are adjusted, the normal or abnormal judgment cannot be obtained, the calculation is abandoned, the fitting parameter combination fine tuning process fails, and it is not necessary to update to the algorithm description.

[0168] (6) The fitting parameter combination after fine tuning is updated to the algorithm according to the algorithm description.

[0169] Embodiment 4:

[0170] Based on the foregoing embodiments, this embodiment also provides a specific example based on a specific application scenario:

[0171] In this embodiment, the confidence degree is selected as 95%, k in the k-σ criterion is 3, that is, 3σ criterion. Define the normal data as 1, the abnormal data as 2, and the suspicious data as 0. When the abnormal rate per unit time is greater than or equal to 10%, it indicates that the precision is obviously decreased, and the model needs to be corrected.

[0172] Combined with Figure 3 and Figure 4 , the amplifier model OA1 in an OCH optical channel is taken as an example.

[0173] The index of the amplifier model OA1 is output optical signal-to-noise ratio (OSNR), the amplifier model OA1 is dependent on the optical fiber model EDFA1, and is dependent on the optical transmission section model OTS1. The organization of the amplifier model OA1 is kept unchanged, the selected algorithm of the amplifier model OA1 is kept unchanged, but the dynamic parameters can be updated and the fitting parameters can be adjusted.

[0174] The algorithm A1 of the amplifier model OA1 depends on the optical power meter (OPM), and the amplifier model OA1 depends on the fiber model EDFA1. The algorithm A1 of the optical amplifier model OA1 has fitting parameters {a1, a2, a3}. The algorithm B1 of the fiber model EDFA1 has fitting parameters {b1, b2}, as shown in the figure. Figure 7 、 8 shown.

[0175] The amplifier model OA1 is calibrated as follows:

[0176] (1) The optical power sampling data collected by the probe component OPM is input into the amplifier model OA1. The amplifier model OA1 calculates the optical signal-to-noise ratio based on the optical power sampling data and algorithm A1 to determine whether the optical signal-to-noise ratio calculation result is normal. According to the interval update method, the optical power sampling data is added to the normal set or abnormal set, and the normal value interval or abnormal value interval is updated. For example, after 1000 calculations, if the abnormality rate of the calculation result exceeds 10%, the optical power sampling data is marked as suspicious data.

[0177] (2) First, check the probe component OPM of the amplifier OA1 to determine whether the optical power sampling data exceeds its nominal range. If so, it is determined that the OPM has failed and the algorithm needs to be reselected. Otherwise, jump to step (3).

[0178] (3) Maintaining the current model and model relationship unchanged, determine whether the optical amplifier model OA1 is dependent on other models. If so, continue to pass the relevant data to the optical transmission segment model OTS1, the optical multiplexing segment model OMS1, and the optical channel model OCH until it is possible to directly determine whether the calculation result is normal or abnormal. If it is normal, mark the suspicious data participating in this model calculation as normal, and update the reference interval according to the reference interval update method.

[0179] (4) When the optical channel model OCH cannot directly determine the calculation result as normal or abnormal, that is, when it is still abnormal and suspicious, according to the model selection algorithm and the determined organizational structure, reverse calculation is performed based on the fitting parameters {a1, a2, a3} of the optical amplifier model OA1 and the fitting parameters {b1, b2} of the optical fiber model EDFA1 according to the fitting parameter adjustment method until all data marked as suspicious are marked as normal data after the new fitting parameters are adjusted, and the reference interval is updated. The latest fitting parameter values ​​are also updated to the current algorithm.

[0180] (5) If, after adjusting the fitting parameters, a clear judgment of normality or abnormality cannot be obtained, the model needs to select a new algorithm. If, after adjusting all available algorithms, the calculation requirements are still not met, the calculation is abandoned. Jump to step (6).

[0181] (6) The physical device corresponding to the model is marked as abnormal and suspicious, and a notification is issued.

[0182] (7) Whether the physical device is truly abnormal is manually confirmed, or determined by an end-to-end closed-loop method. If it is abnormal, the corresponding physical device needs to be updated. If it is not abnormal, the model is marked as needing to update the fitting parameters or algorithm.

[0183] Example 5:

[0184] Based on the network digital model correction method provided in the above embodiment, the present invention also provides a device that can be used to implement the above method, such as Figure 11 FIG2 is a schematic diagram of the device architecture of an embodiment of the present invention. The device of this embodiment includes one or more processors 21 and a memory 22. Figure 11 A processor 21 is taken as an example.

[0185] The processor 21 and the memory 22 may be connected via a bus or other means. Figure 11 The bus connection is taken as an example.

[0186] The memory 22, as a non-volatile computer-readable storage medium for the network digital model correction method, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the network digital model correction method in Example 1. The processor 21 executes the non-volatile software programs, instructions, and modules stored in the memory 22 to execute various functional applications and data processing of the device, thereby implementing the network digital model correction method of the embodiment.

[0187] The memory 22 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 22 may optionally include a memory remotely located relative to the processor 21, and such remote memory may be connected to the processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0188] The program instructions / modules are stored in the memory 22 , and when executed by the one or more processors 21 , the method for correcting the network digital model in the above embodiment is executed.

[0189] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a disk or an optical disk, etc.

[0190] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for correcting a network digital model, characterized in that: include: The target algorithm selected by the digital model is used to calculate the sampled data to obtain the calculation results; the sampled data is the real-time perception data of the network status by the probe; The reference interval is updated according to the normal value set and the abnormal value set, and the calculation result is identified according to the reference interval; the abnormal value set includes a first abnormal value set and a second abnormal value set, and the data in the second abnormal value set are all greater than the data in the first abnormal value set. The updating of the reference interval according to the normal value set and the abnormal value set includes: determining the confidence interval according to a predetermined confidence level and n sample data, and determining the reference interval according to the confidence interval and a predetermined k-σ criterion; performing abnormal identification on the indicator data to be identified according to the reference interval, and selectively adding the corresponding indicator data to the normal value set, the first abnormal value set or the second abnormal value set according to the identification result; updating the mean according to the normal value set and variance , update the mean according to the first set of outliers and variance , update the mean according to the second set of outliers and variance ;when When , the lower bound of the normal value is updated to ;when When the normal value upper bound is updated , to update the reference interval; If the calculation result falls outside the reference interval, the calculation result is suspicious data, and a request message carrying the calculation result is sent to the downstream calculation model, which identifies the calculation result until it is identified as normal data or abnormal data; If the calculation result is abnormal data, the calculation result is added to the abnormal value set, so as to update the reference interval according to the abnormal value set and the normal value set; Adjust the fitting parameters of the target algorithm. If, after a preset number of adjustments, at least one normal value cannot be obtained, the self-digital model is triggered to reselect the target algorithm.

2. The calibration method according to claim 1, wherein: The correction method further comprises: If the calculation result falls within the reference interval, the calculation result is normal data, and the calculation result is added to the normal value set to update the reference interval according to the abnormal value set and the normal value set.

3. The calibration method according to claim 2, wherein: The process of adjusting the fitting parameters of the target algorithm is as follows: Selecting an adjustment step of the fitting parameters according to the exploration efficiency, and adjusting the fitting parameters according to the adjustment step to obtain a first fitting parameter combination; The first fitting parameter combination is updated to the target algorithm, and a trial calculation is performed using the updated target algorithm to obtain a trial calculation result B; If the trial result B is the same as the result A before the trial calculation, continue adjusting the fitting parameters according to the selected adjustment step until the trial result B is the opposite of the result A before the trial calculation; The reverse dichotomy method is used to adjust the fitting parameter combination corresponding to the target algorithm in the previous trial calculation to obtain a second fitting parameter combination, the second fitting parameter combination is updated to the target algorithm, and the updated target algorithm is used to perform a trial calculation to obtain a trial calculation result C; If the trial result C is the same as the trial result B before the trial calculation, continue to adjust the fitting parameters according to the reverse dichotomy method until the trial result C is opposite to the trial result B, and update the fitting parameters to the fitting parameter list in the algorithm description.

4. The calibration method according to claim 3, wherein: The adjustment step of the fitting parameters selected according to the exploration efficiency includes: The number of explorations corresponding to the change of the trial calculation result from abnormal data to normal data is obtained. If the number of explorations is less than the set threshold, the adjustment step is reduced; if the number of explorations is greater than the set threshold, the adjustment step is increased.

5. The calibration method according to claim 3, wherein: The reverse dichotomy method is used to adjust the fitting parameters corresponding to the target algorithm in the previous trial calculation to obtain the second fitting parameter combination, which includes: If the adjustment step corresponding to the previous trial calculation is greater than 1, the second fitting parameter combination is half of the fitting parameters corresponding to the target algorithm in the previous trial calculation; If the adjustment step corresponding to the previous trial calculation is less than 1, the second fitting parameter combination is twice the fitting parameters corresponding to the target algorithm in the previous trial calculation.

6. The calibration method according to claim 1, wherein: Determining the confidence interval according to a predetermined confidence level and n sample data, and determining the reference interval according to the confidence interval and a predetermined k-σ criterion includes: Get the mean of n sample data and variance , according to the mean ,variance , confidence constant z and sample data number n to calculate the confidence interval ; Determine the initial reference interval based on the confidence interval and the predetermined k-σ criterion ; In the subsequent process, when any one of the normal value set, the first abnormal value set or the second abnormal value set changes, the mean and variance of the corresponding set are updated to update the reference interval.

7. The calibration method according to claim 1, wherein: The selectively adding the corresponding indicator data to the normal value set, the first abnormal value set or the second abnormal value set according to the recognition result includes: If the indicator data is less than , then add the indicator data to the first abnormal value set; if the indicator data is greater than , then add the indicator data to the second abnormal value set; if the indicator data falls within the initial value interval, then add the indicator data to the normal value set.

8. The calibration method according to any one of claims 1 to 7, wherein: The correction method further comprises: If the abnormality cannot be eliminated by modifying the fitting parameters and reselecting the algorithm, then all candidate models corresponding to the carrier are obtained, and the algorithm characterization capability, model usage frequency in each time period, and parameter list satisfaction of the algorithm corresponding to each candidate model are obtained; Obtaining the product of the algorithm characterization capability, the model usage frequency, and the parameter list satisfaction, and using the product as the model capability of the candidate model; Sort the models in descending order of capability, starting with the one with the greatest capability, until a suitable model is selected.

9. A correction device for a network digital model, characterized in that: The method comprises at least one processor and a memory, wherein the at least one processor and the memory are connected via a data bus, and the memory stores instructions that can be executed by the at least one processor, and after being executed by the processor, the instructions are used to complete the correction method of the network digital model described in any one of claims 1-8.

Images