A Fusion Method and Fusion Device for a Network Digital Model

Abstract

The present invention discloses a method and a device for fusing network digital models, including: obtaining carriers included in a target service, topological relationships between the carriers, and service flows, and obtaining at least one alternative model corresponding to each carrier; for each of the carriers, obtaining the model capabilities of each of the alternative models, and selecting a suitable alternative model as the target model corresponding to the carrier according to the model capabilities; constructing a dependency relationship between the target models according to the service flows and the topological relationships, and constructing a message channel between the target models according to the dependency relationship; receiving interaction messages from other target models based on the message channel, and determining whether it is necessary to update the models and the model relationships according to the interaction messages.

Description

Technical Field

[0001] The present invention belongs to the field of digital models, and more specifically, relates to a method and device for fusing network digital models. Background Art

[0002] Full-element digitization is the development theme of the current era, which features data resources as key elements and information and communication networks as key transmission and exchange carriers. With the requirements of digitization for diverse and complex information connection and transmission, new network forms such as space-air-ground integrated communication, 5G, and deterministic networks have emerged continuously. The networking forms between networks are increasingly complex, the dependencies between network devices are also increasingly complex, and the network service load is increasing day by day. Based on traditional network design, deployment, and operation and maintenance management means, it is becoming increasingly powerless. Using digital means to support network design, deployment, and operation and maintenance is an important current technical idea, and the evolution towards network digitization is also an inevitable trend accompanying the development of full-element digitization.

[0003] For network digitization, it is necessary to perform digital modeling on each component of the network (devices, links, networks, services, etc.) and the production environment where they are located, and make the digital model accurately reflect the laws of physical changes of network facilities. Combining various data such as the production process and the real-time changes of network facilities themselves, using reasonable computing resources, the network operation behavior and state can be accurately characterized in real time. Network digitization can gradually evolve from the current technology means that rely heavily on manual and semi-automated methods to fully automated, intelligent, and autonomous technology means, and is expected to solve the problems of increasingly complex network forms and increasingly powerless network management means.

[0004] However, the increasing complexity of network forms also makes the dependencies between network components more complex, which in turn makes the process of fusing their corresponding digital models more complex and difficult to control. Especially during the fusion process, after an abnormality occurs, it is difficult to perform fault backtracking and positioning autonomously, thus hindering the development of network digitization.

[0005] Moreover, the current network digital modeling technology is still in its infancy, and the lack of accuracy is one of the main factors affecting the development of network digitization. The digital model needs to be continuously optimized and updated in a long-term evolution process. With the continuous update of the digital model, the complexity of data fusion in the production environment is further increased. Summary of the Invention

[0006] In view of the above defects or improvement requirements of the prior art, the present invention provides a method and device for fusing network digital models, aiming to support the iterative update of digital models, thereby solving the technical problems of low efficiency in model and data fusion and difficult fault positioning when fusion abnormalities occur in scenarios with multiple model levels and complex dependencies.

[0007] To achieve the above object, according to one aspect of the present invention, there is provided a method for fusing network digital models, including:

[0008] Obtain the carriers included in the target service, the topological relationship between the carriers, and the service flow direction, and obtain at least one alternative model corresponding to each carrier;

[0009] For each of the carriers, obtain the model capabilities of each of the alternative models, and select a suitable alternative model as the target model corresponding to the carrier according to the model capabilities;

[0010] Construct the dependency relationship between the target models according to the service flow direction and the topological relationship, and construct the message channels between the target models according to the dependency relationship;

[0011] Receive interaction messages from other target models based on the message channels, and determine whether the models and the model relationships need to be updated according to the interaction messages.

[0012] Further, the obtaining the model capabilities of each of the alternative models and selecting a suitable alternative model as the target model corresponding to the carrier according to the model capabilities includes:

[0013] Obtain the algorithm characterization capabilities of the algorithms corresponding to each alternative model, the model usage frequency within each time period, and the parameter list satisfaction degree;

[0014] Obtain the product of the algorithm characterization capabilities, the model usage frequency, and the parameter list satisfaction degree, and use this product as the model capability of the alternative model;

[0015] Sort in descending order according to the model capabilities, and start selecting from the one with the largest model capability until the target model is selected.

[0016] Further, the obtaining method of the model usage frequency is:

[0017] Obtain the number of times the model is used within the time period. Among them, each time the model is used, the number of times the model is used is incremented by one; each time the model is abandoned, the number of times the model is used is decremented by one; after each time period ends, the number of times the model is used is cleared and re-counted;

[0018] Calculate the ratio of the number of times the model is used to the time period, and use the value corresponding to the sum of this ratio and a random number as the model usage frequency.

[0019] Further, the obtaining method of the parameter list satisfaction degree is:

[0020] Obtain the performance metric indicators, and determine the minimum parameter set according to the performance metric indicators;

[0021] Obtain the probe components of each alternative model, and determine whether the detection accuracy of each probe component meets the preset accuracy. If it meets, add the parameter type corresponding to the probe component to the alternative model parameter list;

[0022] Determine whether all the parameter types included in the minimum parameter set exist in the alternative model parameter list;

[0023] If all the parameter types included in the minimum parameter set exist in the alternative model parameter list, the satisfaction degree of the parameter list is 1;

[0024] If at least one parameter type does not exist in the alternative model parameter list, the satisfaction degree of the parameter list is 0.

[0025] Further, constructing the dependency relationship between target models according to the service flow and the topological relationship, and constructing the message channel between target models according to the dependency relationship includes:

[0026] Construct the sequential dependency relationship between target models according to the service flow, and construct the nested dependency relationship between target models according to the topological relationship;

[0027] Set a unique model identifier for each target model;

[0028] For multiple target models with nested dependency relationships, the sub-model registers its model identifier into the parent model, and the parent model stores the received model identifier into the registration list; wherein, the parent model is the target model at the upper layer of the nested tree structure, and the sub-model is the model at the lower layer of the nested tree structure;

[0029] For multiple target models with sequential dependency relationships, the upstream model of the calculation registers its model identifier into the downstream model of the calculation, and the downstream model of the calculation stores the received model identifier into the registration list; wherein, the output of the upstream model of the calculation serves as the input of the downstream model of the calculation;

[0030] Construct the message channel between target models based on the registration list of each target model.

[0031] Further, receiving interaction messages from other target models based on the message channel, and determining whether to update the model and the model relationship according to the interaction messages includes:

[0032] Monitor the model efficiency of itself, and judge the model state of its own model according to the model efficiency;

[0033] If the own model is in a fault state, after receiving a request message from another target model, send a reply message of model fault to that model, and report the message of model fault to the parent model, and the parent model reorganizes the model again;

[0034] If the own model is in an idle state, after receiving a request message from another target model, the own model performs digital-analog fusion calculation according to the request message to obtain a digital-analog fusion body, sends the digital-analog fusion body to all models that have established a registration relationship with the own model, and determines whether it is necessary to update the model and the model relationship according to the digital-analog fusion body.

[0035] Further, the step of, after receiving a request message from another target model, the own model performs digital-analog fusion calculation according to the request message to obtain a digital-analog fusion body, sends the digital-analog fusion body to all models that have established a registration relationship with the own model, and determines whether it is necessary to update the model and the model relationship includes:

[0036] After receiving a request message from another target model, the own model performs a trial calculation according to the request message to obtain an instantiated value of the performance metric corresponding to the own model;

[0037] Judge whether the instantiated value of the performance metric satisfies the normal value range. If it satisfies the normal value range, form several digital-analog fusion bodies from several trial calculation results, and send the digital-analog fusion bodies to the downstream calculation models that have established a registration relationship with the own model in the order of first sequential dependence and then nested dependence;

[0038] Receive the calculation results of the digital-analog fusion body from the downstream calculation models. If all the calculation results are abnormal, the parent model corresponding to the own model reorganizes the model again, and reduces the number of times the model corresponding to the own model is used by one;

[0039] If at least one calculation result is normal, continue to notify the downstream calculation models of the digital-analog fusion body until the calculation of the model corresponding to the target service is completed.

[0040] Further, after judging whether the instantiated value of the performance metric satisfies the normal value range, it further includes:

[0041] If it does not satisfy the normal value range, initiate an abnormal calculation request to the upstream calculation model in the order of first nested dependence and then sequential dependence to determine the abnormal source.

[0042] Further, the step of monitoring the model efficiency of the own model and judging the model state of the own model according to the model efficiency includes:

[0043] Obtain the first ratio between the actual accuracy and the promised accuracy of the model, and the number of requests n within the time periodrq 1. The number n of response executions within a time period ex 2. The number n of normal responses within a time period rs and the number n of abnormal abandonments within a time period ab ;

[0044] For the number n of requests rq 2. The number n of response executions ex 3. The number n of normal responses rs and the number n of abnormal abandonments ab are summed to obtain a sum value, and the second ratio between the number n of normal responses rs and the sum value is calculated;

[0045] The product of the first ratio and the second ratio is used as the model efficiency;

[0046] If the model efficiency is not less than the set threshold, the model is in a busy state; if the model efficiency is less than the set threshold, the model is in an idle state; if the model efficiency is 0, the model is in a fault state.

[0047] According to another aspect of the present invention, there is provided a fusion device for a network digital model, including 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 after being executed by the processor, the instructions are used to complete the fusion method described in the first aspect.

[0048] Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the following beneficial effects are obtained:

[0049] (1) Support the dynamic update of digital models. The accuracy of the model's characterization of device behavior and status is continuously evolving. It is a long-term process from inaccurate to accurate to precise to high-precision representation; therefore, it is necessary to support dynamic selection and update during the use of digital models.

[0050] (2) An effective negotiation and interaction mechanism between models. Ensure operations such as pause, termination, and recovery and negotiation and interaction between models in scenarios such as model evolution and update, fault and exception, so as to ensure a higher credibility of digital models.

[0051] (3) The ability of autonomous negotiation and fusion. Through the autonomous negotiation of digital models, autonomous model selection and orchestration, and the ability of model efficiency monitoring and exception backtracking, support the efficient fusion of models and data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic flowchart of a fusion method for a network digital model provided by an embodiment of the present invention;

[0053] Figure 2 It is a schematic diagram of a multi-level nested dependency and multiple sequence dependency scenario in OCH provided by an embodiment of the present invention;

[0054] Figure 3 It is a schematic diagram of the relationship between an index, accuracy, and a network digital model provided by an embodiment of the present invention;

[0055] Figure 4 provided by an embodiment of the present invention Figure 1 Specific process schematic diagram of step 30 in;

[0056] Figure 5 It is a schematic diagram of the sequence dependency relationship of a network digital model provided by an embodiment of the present invention;

[0057] Figure 6 It is a schematic diagram of the nested dependency relationship of a network digital model provided by an embodiment of the present invention;

[0058] Figure 7 provided by an embodiment of the present invention Figure 1 Specific process schematic diagram of step 40 in;

[0059] Figure 8 It is the overall framework diagram of the network digital model and data fusion method provided by an embodiment of the present invention;

[0060] Figure 9 It is a schematic diagram of the efficiency monitoring of a network digital model provided by an embodiment of the present invention;

[0061] Figure 10 It is a schematic diagram of the complex dependency of a network digital model provided by an embodiment of the present invention;

[0062] Figure 11 It is a schematic diagram of a specific fusion example of a network digital model and data provided by an embodiment of the present invention;

[0063] Figure 12 It is a schematic diagram of the structure of a fusion device of a network digital model provided by an embodiment of the present invention. Specific Embodiments

[0064] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0065] Embodiment 1:

[0066] To solve the foregoing problems, this embodiment provides a method for fusing network digital models. Refer to Figure 1 , the fusion method includes:

[0067] Step 10: Obtain the carriers included in the target service, the topological relationship between the carriers, and the service flow direction, and obtain at least one alternative model corresponding to each carrier.

[0068] Among them, the target service can be understood as the specific service that the model needs to implement. The carrier refers to the physical entity or virtual entity required to implement the specific service, and the specific object referred to by the carrier depends on the target service.

[0069] Taking the OTN (Optical Transport Network) network as an example, the following levels of topological information are included in the OTN network: topological information of the OTS (Optical Transmission Section) layer, topological information of the OMS (Optical Multiplex Section) layer, topological information of the OCH (Optical Channel) layer, and topological information of the ODUk (Optical Channel Data Unit) layer. Among them, the OMS layer includes at least one OMS topology, and each OMS topology includes one or more OTS links. The OTS link is the physical layer topology between two optical devices. That is, the topology of the OMS layer is generated after the optical devices are networked and can be regarded as the physical layer topology.

[0070] Among them, the OCH layer topology is a virtual topology carried on the OMS layer and needs to be obtained through routing. The OCH layer includes at least one OCH topology, and the OCH topology includes one or more OMS topologies.

[0071] Among them, the ODUk layer topology is a virtual topology carried on the OCH layer and needs to be obtained through routing. The ODUk layer includes at least one ODUk topology, and each ODUk topology includes one or more OCH topologies. The ODUk topology is used to carry the CLIENT service.

[0072] Combined with Figure 2, for an OCH service, the bearers (virtual entities) corresponding to the OCH service include OCH links, OMS links, and OTS links. The bearers (physical entities) corresponding to the OCH service also include a transmitter, a multiplexer, a reconfigurable optical add-drop multiplexer (ROADM), an optical fiber, an optical amplifier (OA), a demultiplexer, and a receiver. Each bearer corresponds to a digital model, and one digital model is selected as the alternative model for the bearer.

[0073] For example, the OCH service includes bearers of virtual entities and bearers of physical entities, and it is necessary to obtain the alternative models of each bearer.

[0074] For the OCH service, the bearers (virtual entities) include OCH links, OMS links, and OTS links. For example, for the bearer - the OCH link, there are at least one OCH model corresponding to the OCH link, and one of them is selected as the alternative model of the OCH link.

[0075] The bearers (physical entities) corresponding to the OCH service also include a transmitter, a multiplexer, a reconfigurable optical add-drop multiplexer (ROADM), an optical fiber, an optical amplifier (OA), a demultiplexer, and a receiver. For example, for the bearer - the transmitter, there are at least one Tx model corresponding to the transmitter, and one of them is selected as the alternative model of the transmitter.

[0076] The topological relationships between the bearers are as follows:

[0077] The OCH service is a service carried on the OCH link. The OCH link is generally composed of a transmitter, a multiplexer, a reconfigurable optical add-drop multiplexer (ROADM), an optical fiber, an optical amplifier (OA), a demultiplexer, and a receiver. Among them, the reconfigurable optical add-drop multiplexer (ROADM), the optical fiber, and the optical amplifier (OA) constitute the OMS link, the optical fiber and the optical amplifier (OA) constitute the OTS link, and the pump light source and the EDFA erbium fiber constitute the optical amplifier.

[0078] The service flow direction is from the transmitter to the receiver.

[0079] Combined with Figure 3, select a suitable model according to the performance metric, the parameter types included in the parameter set, and the accuracy of the performance metric.

[0080] Among them, the performance metric is only a self-defined key output parameter with a tendency, which is a simplified description of the performance ability of the model or the physical entity corresponding to the model, and is also the basic condition for the digital model to perform behavior characterization and computational solution. However, to meet different index accuracy requirements, different mechanisms, perceptions, and calculations are used for support. During the specific support process, the parameter sets may be different. For example, when the EDFA amplifier uses the optical power calibration and fitting method to solve the amplification behavior, the parameter sets required for solving the amplification behavior using the optical power transmission equation will be different.

[0081] It can be understood that when the accuracy requirements of a certain performance metric of the model are different, the corresponding parameter sets are also different. For example, Figure 3 the parameter sets corresponding to accuracy 1 and accuracy 2 are different; the same accuracy will also correspond to different parameter sets. When the parameter sets are different, the models will also be different.

[0082] Selecting a suitable model means that the model can achieve the performance metric, can reach the accuracy of the performance metric, and can support the parameter types included in the parameter set.

[0083] Step 20: For each of the carriers, obtain the model capabilities of each of the alternative models, and select a suitable alternative model as the target model corresponding to the carrier according to the model capabilities.

[0084] Each carrier corresponds to at least one alternative model, and a suitable model is selected from the alternative models as the target model corresponding to the carrier according to needs.

[0085] Specifically, obtain the algorithm characterization ability of the algorithm corresponding to each alternative model, the model usage frequency within each time period, and the parameter list satisfaction; obtain the product of the algorithm characterization ability, the model usage frequency, and the parameter list satisfaction, and use this product as the model ability of the alternative model; sort in descending order of model ability, and start selecting from the one with the largest model ability until the target model is selected.

[0086] That is, the model ability can be obtained according to the following formula:

[0087] Model_ability = algorithm_ability × model_usage_frequency × parameter_sat

[0088] Among them, algorithm_ability is the algorithm characterization ability, model_usage_frequency is the model usage frequency within a time period, and parameter_sat is the parameter list satisfaction degree.

[0089] In this embodiment, the way to obtain the model usage frequency is as follows: obtain the number of times n that the model is used within the time period T. Among them, each time the model is used, the model usage times n is incremented by one; each time the model is abandoned, the model usage times n is decremented by one; after each time period ends, the model usage times is cleared and recounted. Among them, the model being abandoned means that when the model is selected, the model will be instantiated and participate in the calculation of the entire system. However, when the accuracy, actual calculation performance, etc. cannot meet the requirements of itself or other models, it needs to be abandoned and reselected. This is a necessary process in the continuous evolution of the model, following the natural elimination mechanism.

[0090] Calculate the ratio of the model usage times to the time period, and use the value corresponding to the sum of this ratio and the random number δ as the model usage frequency. In practical applications, there are situations where there are multiple usage frequencies with the same calculation result. By increasing the random number, the normal distribution characteristics of different models during long-term use can be increased, preventing the rigidity of model selection.

[0091] That is, the model usage frequency can be calculated according to the following formula: model_usage_frequency = n / T + δ. Among them, the random number δ carried in the model usage frequency can ensure that each model has the opportunity to be used.

[0092] In this embodiment, the way to obtain the parameter list satisfaction degree is as follows:

[0093] Obtain the performance metric indicators, and determine the minimum parameter set according to the performance metric indicators; obtain the probe components of each alternative model, and judge whether the detection accuracy of each probe component meets the preset accuracy. If it meets, add the parameter type corresponding to the probe component to the alternative model parameter list; judge whether all the parameter types included in the minimum parameter set exist in the alternative model parameter list; if all the parameter types included in the minimum parameter set exist in the alternative model parameter list, the parameter list satisfaction degree is 1; if at least one parameter type does not exist in the alternative model parameter list, the parameter list satisfaction degree is 0.

[0094] Among them, each alternative model can correspond to multiple algorithms. The alternative model selects a suitable algorithm from multiple algorithms as the algorithm of the alternative model. Each algorithm will correspond to a minimum parameter set when it is to be executed, and this minimum parameter set includes at least one parameter type;

[0095] For the algorithm to perform calculations properly, it must have the sampling data corresponding to these parameters, which is collected by the probe component. However, when the probe component included in the alternative model fails to work properly, the input requirements of the minimum parameter set cannot be met, and the corresponding algorithm cannot be used. It can be understood that when the satisfaction degree of the parameter list is 1, it means that the model can be selected; otherwise, it cannot be selected.

[0096] Taking the OSNR (Optical Signal to Noise Ratio) metric as an example, the OSNR metrics supported by the OCH model include the power flatness of multiple channels. The accuracy requirements for this power flatness can be ≤|±0.8|, ≤|±1|, ≤|±1.2|, ≤|±2|,...; the OSNR flatness of multiple channels. The accuracy requirements for this OSNR flatness can be ≤|±1|, ≤|±1.5|, ≤|±2|,...; the OSNR error margin of multiple channels. The accuracy requirements for this OSNR flatness can be ≤0.8dB, ≤1.0dB, ≤1.2dB, ≤1.5dB,...; a total of 1 to M, where M refers to the number of performance metric indicators that can characterize OSNR.

[0097] Taking the OSNR error margin of multiple channels as the performance metric indicator and the accuracy of the performance metric indicator being ≤1.0dB as an example, it is determined that its minimum parameter set includes the input single-wave power, output single-wave power, input total power, output total power, output single-wave OSNR value, and OSNR error. The purpose of this minimum parameter set is to determine whether the satisfaction degree of the parameter list of the alternative model parameter list is 1 or 0.

[0098] When the detection accuracy of the probe component included in the alternative model does not meet the preset accuracy, it means that the probe component cannot be used properly, and the parameter type corresponding to this probe component is not added to the alternative model parameter list; when the detection accuracy of the probe component included in the alternative model meets the preset accuracy, it means that the probe component can be used properly, and the parameter type corresponding to this probe component is added to the alternative model parameter list. It can be understood that the alternative model parameter list is the parameter type that the alternative model can support.

[0099] That is, the satisfaction degree of the parameter list can be calculated in the following way:

[0100]

[0101] Among them, parameter_sat is the satisfaction degree of the parameter list, min{parameters} is the minimum parameter set of an algorithm in the alternative model, and parameter_list is the alternative model parameter list, which includes the measurement parameters corresponding to several sensors.

[0102] Step 30: Construct the dependency relationship between target models according to the service flow direction and the topological relationship, and construct the message channel between target models according to the dependency relationship.

[0103] Construct the dependency relationship between target models according to the service flow direction and the topological relationship, set a model identifier for each model, determine the upstream calculation models that have dependencies with the own model according to the dependency relationship, and register the model identifier of the own model into the upstream calculation models, so as to construct the message channel between target models. Different models send interaction messages through this message channel, where the interaction messages include request messages, response messages, and notification messages.

[0104] Among them, the format of the interaction message is as follows:

[0105] The format of the request message is: [request source identifier: request target identifier: metric accuracy: computing performance];

[0106] The format of the response message is: [response source identifier: response target identifier: completed / failed];

[0107] The format of the notification message is: [notification source identifier: notified target: notification data].

[0108] In an optional embodiment, as Figure 4 shown, step 30 specifically includes the following steps:

[0109] Step 301: Construct the sequential dependency relationship between target models according to the service flow direction, and construct the nested dependency relationship between target models according to the topological relationship;

[0110] There are two relationships between models: sequential dependency relationship (as Figure 5 shown); nested dependency relationship (as Figure 6 shown). Taking Figure 2 as an example, the nested dependency relationship means that the OSNR of OCH OCH depends on the OSNR of several OMSs OMS , and there is a parent-child relationship between them, thus forming a nested dependency.

[0111] When it is order-dependent, it is determined according to the business flow relationship of the carrier. Therefore, there is a sequential relationship among models 1 to n, and they are not subordinate to each other and not nested. The input of model n is provided by the output of model n - 1. After model n processes all the input data through calculation, it can notify the calculation results and status to model n - 1; the input of model n - 1 is provided by the output of model n - 2. After model n - 1 processes all the input data through calculation, it can notify the calculation results and status to model n - 2; and so on. The input of model 2 is provided by the output of model 1. After model 2 processes all the input data through calculation, it can notify the calculation results and status to model 1.

[0112] When it is nested-dependent, a nested relationship is formed between the model and the sub-models. At this time, the input of the model is jointly provided by the outputs of multiple sub-models. When organizing the model, a message interaction relationship is established between the model and the sub-models, and it is optional to carry requirements such as model ability (Model_ability) and algorithm ability to the sub-models along with the request message.

[0113] Step 302: Set a unique model identifier for each target model;

[0114] Among them, the model identifier can be the address number of each model. For example, the model identifier can be an IP address.

[0115] Step 303: For multiple target models with nested dependency relationships, the sub-model registers its model identifier with the parent model, and the parent model stores the received model identifier in the registration list; among them, the parent model is the target model at the upper layer of the nested tree structure, and the sub-model is the model at the lower layer of the nested tree structure;

[0116] Among them, the sub-model and the parent model perform message interaction. For example, using the TCP protocol, the sub-model registers its model identifier with the parent model.

[0117] Step 304: For multiple target models with order dependency relationships, the upstream model in calculation registers its model identifier with the downstream model in calculation, and the downstream model in calculation stores the received model identifier in the registration list; among them, the output of the upstream model in calculation serves as the input of the downstream model in calculation;

[0118] Among them, when the models are order-dependent, the starting side of the business is the upstream, and the terminating side of the business is the downstream. When the models are nested-dependent, the parent model side is the upper layer, and the sub-model side is the lower layer, and there is no upstream and downstream relationship.

[0119] However, for the sake of convenience in description, the upstream calculation and the downstream calculation are defined. Upstream calculation: the upstream in the order dependence and the lower layer in the nested dependence. Downstream calculation: the downstream in the order dependence and the upper layer in the nested dependence. The model corresponding to the upstream calculation is called the upstream calculation model, and the model corresponding to the downstream calculation is called the downstream calculation model.

[0120] Among them, the upstream calculation model and the downstream calculation model perform message interaction through the TCP protocol, and the upstream calculation model registers its model identifier to the downstream calculation model.

[0121] For example, for the model m of OCH, its sub-models with nested relationships include models u, w, n, z, and t, and models m, u, w, n, z, and t all have IP addresses.

[0122] Models u, w, n, z, and t respectively perform message interaction with model m using the IP address and the TCP protocol, register the relevant IP address information on model m, and form a registration list on model m; when model m needs to send a notification message, it uses the registration list to send the notification message to each registered model u, w, n, z, and t.

[0123] For models u, w, n, z, and t, their relationship is an order relationship. Therefore, the IP address of model u is registered to model w, and the IP address of model w is registered to model n. Similarly, models n, z, and t are all registered using a similar method.

[0124] Step 305: Construct a message channel between target models based on the registration list of each target model.

[0125] In this embodiment, messages are sent, replied to, or notified according to the identifier in the interaction message and the model identifier in the registration list.

[0126] Step 40: Receive interaction messages from other target models based on the message channel, and determine whether the model and the model relationship need to be updated according to the interaction messages.

[0127] Among them, model update includes two processes: (1) algorithm update; (2) reorganize the model; if the calculation result in the interaction message is abnormal, the algorithm is updated first. If the calculation result can be corrected to normal after the algorithm update, there is no need to reorganize the model; if the calculation result still cannot be corrected to normal after the algorithm update, the model needs to be reorganized.

[0128] In step 40, it specifically includes: receiving interaction messages from other target models based on the message channel, parsing the interaction messages to obtain calculation results. If the calculation results are abnormal, reselect the algorithm, and perform trial calculations according to the reselected algorithm to obtain trial calculation results. Form multiple digital-analog fusion bodies from the trial calculation results, and send the digital-analog fusion bodies to the downstream calculation model. The downstream calculation model performs calculations based on the digital-analog fusion bodies to obtain new calculation results. If all the new calculation results are abnormal, reorganize the model. If there is one new calculation result that is normal, use the corresponding algorithm as the algorithm of the model. Among them, if the model is reorganized, it is necessary to re-register the model as described in detail above.

[0129] Specifically, in combination with Figures 7 to 9 , in step 40, it specifically includes:

[0130] Step 401: Monitor the model efficiency of its own model, and judge the model state of its own model according to the model efficiency;

[0131] When its own model receives requests from other models, monitor the model efficiency of its own model, and judge the model state of its own model according to the model efficiency to give corresponding responses.

[0132] In this embodiment, the acquisition method of the model efficiency is:

[0133] Obtain the first ratio between the actual accuracy and the promised accuracy of the model, the number of requests n within the time period rq , the number of response executions n within the time period ex , the number of normal responses n within the time period rs and the number of abnormal abandons n within the time period ab ; Sum the number of requests n rq , the number of response executions n ex , the number of normal responses n rs and the number of abnormal abandons n ab to obtain a sum value, and calculate the second ratio between the number of normal responses n rs and the sum value; Take the product of the first ratio and the second ratio as the model efficiency.

[0134] Among them, the actual accuracy can be understood as the accuracy during the actual operation of the model; the promised accuracy can be understood as the accuracy that the model promises to other models that it can achieve.

[0135] The number of requests n rq refers to the number of request messages obtained, but these messages have not been processed yet;

[0136] The number of response executions n exRefers to the number of messages being processed by the model but not yet completed;

[0137] Number of normal responses n rs Refers to the number of messages that have been processed by the model and can be normally responded to;

[0138] Number of abnormal abandons n ab Refers to the number of messages that the model abandons due to various reasons and cannot process.

[0139] That is, the model efficiency calculation formula:

[0140]

[0141] If the model efficiency is not less than the set threshold, the model is in a busy state; if the model efficiency is less than the set threshold, the model is in an idle state; if the model efficiency is 0, the model is in a fault state.

[0142] If its own model is in a busy state, it sends a waiting response message to other models.

[0143] Step 402: If its own model is in a fault state, after receiving the request message from other target models, it sends a model fault response message to this model and reports the model fault message to the parent model, and the parent model reorganizes the model again;

[0144] Step 403: If its own model is in an idle state, after receiving the request message from other target models, its own model performs digital-analog fusion calculation according to the request message to obtain a digital-analog fusion body, sends the digital-analog fusion body to all models that have established a registration relationship with its own model, and determines whether it is necessary to update the model and the model relationship according to the digital-analog fusion body.

[0145] Specifically, after receiving the request message from other target models, its own model performs a trial calculation according to the request message to obtain an instantiated value of the performance metric corresponding to its own model; determines whether the instantiated value of the performance metric satisfies the normal value range. Specifically, it can be based on a reference interval to determine whether the instantiated value of the performance metric satisfies the normal value range. If it satisfies, it means that the trial calculation result of this model is normal and can continue to notify other models of the digital-analog fusion body formed by the several trial calculation results; if it does not satisfy, it means that the trial calculation result of this model is abnormal and an abnormal calculation request needs to be sent to the upstream model of the calculation to eliminate the abnormal source before notifying other models of the corresponding digital-analog fusion body, otherwise it will affect the calculation of other models.

[0146] Specifically, if the instantiated value of the performance metric meets the normal value range, several trial calculation results are formed into a digital-analog fusion body, and the digital-analog fusion body is sent to the downstream calculation model that has established a registration relationship with its own model in the order of first order dependence and then nested dependence. And receive the calculation results of the digital-analog fusion body from the downstream calculation model. If all the calculation results are abnormal, it means that the own model cannot meet the requirements of the parent model. The parent model corresponding to the own model reorganizes the model again and reduces the number of times the model corresponding to the own model is used by one; if at least one calculation result is normal, continue to notify the downstream calculation model of the digital-analog fusion body until the calculation of the model corresponding to the target service is completed.

[0147] If the instantiated value of the performance metric does not meet the normal value range, an abnormal calculation request is initiated to the upstream calculation model in the order of first nested dependence and then order dependence to determine the abnormal source. After receiving the abnormal calculation request, the upstream calculation model re-obtains the sampling data through the probe component, recalculates according to the re-obtained sampling data, and determines whether the new calculation result is abnormal. If it is abnormal, it detects whether the probe component or the behavior component is faulty. If the new calculation result is normal, the new calculation result is sent to the upstream calculation model again so that the upstream calculation model can determine the abnormal source according to the new calculation result. This process can be understood as a backtracking process to facilitate the determination of the abnormal source.

[0148] For example, for the OSNR performance metric of the EDFA model, there are two alternative models with different precisions, M1EDFA and M2EDFA. When the model capabilities of M1EDFA and M2EDFA are 1.0 and 2.0 respectively, obviously M2EDFA is more suitable as the selected model. Therefore, M2EDFA is selected as the target model and forms a message interaction channel with its dependent parent model OTS. M2EDFA starts monitoring the model efficiency and calculates its own efficiency according to step 401. If the efficiency threshold is 0.7, but the current calculation efficiency value is 0.65, the efficiency is low and a reply message indicating that the efficiency is low and waiting is required to be sent to the parent model. The parent model can choose to wait or re-select a new model according to the actual situation.

[0149] For the target model M2EDFA, after establishing a message channel with its parent model OTS (other target model), when M2EDFA receives a request message from the parent model OTS, M2EDFA will collect probe sensing and other input data to obtain its output power and OSNR value, and get a trial calculation result. When the trial calculation result is normal, the trial calculation result will be sent to the parent model OTS through the message channel, and the parent model OTS will perform calculations based on this trial calculation result to obtain the corresponding calculation result. The parent model OTS can judge whether the calculation result is abnormal according to the normal range. If it is abnormal, M2EDFA needs to recalculate, and then the parent model OTS will send a request for recalculation to the target model M2EDFA again.

[0150] After receiving the request for recalculation, the target model M2EDFA combines the latest sensing and performs reverse detection for possible abnormalities. If an abnormality is found, the abnormality information will be sent to the parent model, and the calculation will be stopped. The parent model OTS will reselect a new model.

[0151] When performing normal fusion, it is divided into the following three cases:

[0152] (1) When the model has only sequential dependencies, calculate from left to right;

[0153] (2) When the model has only nested dependencies, calculate from bottom to top;

[0154] (3) When the model has both sequential dependencies and nested dependencies, sequential dependencies take precedence over nested dependencies; first complete the left-to-right calculation, then complete the bottom-to-top calculation, and finally complete its own calculation by combining the two calculation results.

[0155] When fusion is abnormal, backtracking calculation is required to determine the source of the abnormality, which is divided into the following three cases:

[0156] (1) When the model has only sequential dependencies, retry from right to left.

[0157] For example, in Figure 10 , the model M0 is the parent model of models M1, M4, and M5; the model M1 is the parent model of models M2 and M3; at the same time, there is a nested dependency relationship between M1.subM1 and M1.subM2 and M1, and there is no direct relationship with M2 and M3, so there is no need to establish a notification message relationship. There is a sequential relationship between M2 and M3; there is a sequential dependency relationship between M1, M4, and M5;

[0158] The model M0 depends on models M1, M4, and M5; the model M1 depends on models M2 and M3, that is, only when both M2 and M3 have completed calculations can the model M1 perform fusion calculations; M3 depends on the output after M2 has completed calculations as its input. M1 depends on M1.subM1 and M1.subM2.

[0159] For example, a message channel is established among M1, M4, and M5. After the models M1, M4, and M5 perform calculations in sequence and an anomaly is found, reverse recalculation is required. In this process, M5 needs to update the dynamic sampling data, perform calculations based on the sampling data, and determine whether the new calculation result is abnormal; if it is abnormal, M5 issues an anomaly notice to M4. M4 updates the dynamic sampling data, M4 performs calculations based on the sampling data, and then sends the calculation result to M5. M5 updates the dynamic sampling data, recalculates again, and determines whether it is abnormal. If it is abnormal, M5 notifies M4 of the anomaly, and M4 notifies M1 of the anomaly. Similarly, update the dynamic data from M1, M4, and M5, perform calculations in sequence, and determine whether it is abnormal.

[0160] (2) When the model has only nested dependencies, retry from top to bottom. For example, for M1 and M1.SubM1, M1.SubM2, each retry process is carried out in accordance with the foregoing process.

[0161] (3) When the model has both sequential dependencies and nested dependencies, nested dependencies take precedence over sequential dependencies. Carry out the foregoing processes (2) and (3) respectively according to the priorities.

[0162] That is, in the fusion process, it is divided into two major situations: normal fusion and abnormal fusion. Each major situation contains three minor situations. The data fusion directions of normal fusion and abnormal fusion are the same, but the operation directions are opposite.

[0163] The network digital model of this embodiment has at least the following effects:

[0164] (1) Support the dynamic update of the digital model. The accuracy of the model's description of device behavior and status is continuously evolving. It is a long-term process from inaccurate to accurate to precise to high-precision representation; therefore, it is necessary to support dynamic selection and update during the use of the digital model.

[0165] (2) An effective negotiation and interaction mechanism between models. Ensure operations such as pause, termination, and recovery and negotiation and interaction between models in scenarios such as model evolution and update, faults, and anomalies, so as to ensure a higher credibility of the digital model.

[0166] (3) The ability of autonomous negotiation and fusion. Through the autonomous negotiation of the digital model, autonomous model selection and orchestration, model efficiency monitoring, and anomaly backtracking capabilities, support the efficient fusion of the model and data.

[0167] Embodiment 2:

[0168] To facilitate the understanding of the fusion method in the foregoing embodiment, the implementation process of the fusion method is described below in combination with a specific example:

[0169] Take the OSNR index of the optical channel OCH in the long-haul optical fiber transmission scenario as an example. The optical channel OCH generally consists of a transmitter (Tx), a multiplexer (Mux), a ROADM, an optical fiber (Fiber), an optical amplifier (OA), a demultiplexer (Demux), and a receiver (Rx). The schematic diagram is as Figure 7 shown.

[0170] I. Determine the target model corresponding to each carrier

[0171] 1.1. There are 1, 2,..., M OSNR indexes supported by the OCH model. According to the model selection method, the OCH model m is finally selected. The OCH model m is composed of multiple segments spliced together, and it is necessary to expand and organize the transceiver model, multiplexer (Mux), demultiplexer (Demux), optical multiplex section (OMS) model, etc.

[0172] For example, the OSNR indexes supported by the OCH model include the power flatness of multiple channels ≤|±0.8|, ≤|±1|, ≤|±1.2|, ≤|±2|,...; the OSNR flatness of multiple channels ≤|±1|, ≤|±1.5|, ≤|±2|,...; the OSNR error margin of multiple channels ≤0.8dB, ≤1.0dB, ≤1.2dB, ≤1.5dB,...; a total of 1 to M.

[0173] When selecting the indexes, if only the OSNR error margin of multiple channels is selected as ≤1.0dB, determine that its parameter list includes the single-channel power of input / output, the total power of input / output, the OSNR value of the output single channel, and the OSNR error. The algorithms include the linear fitting algorithm based on calibration, the Maxwell equation algorithm based on linear and nonlinear effects, the AI fitting algorithm, etc.

[0174] The probe components required for the digital model include OPM, temperature measuring instrument, barometric pressure measuring instrument, etc. In some scenarios, if the temperature measuring instrument is not available or not working properly, it is necessary to consider excluding several algorithms that utilize these dynamic data and the models that utilize these algorithms.

[0175] 1.2. There are 1, 2,..., N OSNR indexes supported by the OMS model. According to the model selection method, the OMS model n is finally selected. The OMS model n is also composed of multiple segments spliced together, and it is necessary to further expand and organize through the optical cross-connect (ROADM) model and the optical transmission section (OTS).

[0176] Similarly, the OMS model indexes include the errors of OSNR under 5 spans, 10 spans, and 15 spans ≤0.8dB, ≤1.0dB, ≤1.2dB, ≤1.5dB,...; a total of 1 to N.

[0177] Similarly, for a certain model of OMS, its parameter list includes input / single-wave input power, total input-output power, output single-wave OSNR value, and OSNR error. The algorithms include a calibration-based linear fitting algorithm, a Maxwell equation algorithm based on linear and nonlinear effects, an AI fitting algorithm, and so on.

[0178] The probe components required for the digital model include OPM, temperature measuring instrument, barometric pressure measuring instrument, etc. In some scenarios, if the temperature measuring instrument is unavailable or not working properly, then it is necessary to consider excluding several algorithms that utilize these dynamic data and the models that use these algorithms.

[0179] 1.3. Similarly, the OTS model supports the selection of the OTS model p from among the OSNR metrics 1, 2,..., P. The OTS model p consists of a long-haul optical fiber + amplifier (Pump\EDFA). The optical amplifier (OA) model supports the selection of the OA model q from among the OSNR metrics 1, 2,..., Q. The OTS model q requires the use of pump light wavelength parameters and power parameters of the EDFA module. The long-haul optical fiber model selects the Fiber model r from among its 1, 2,..., R models. The EDFA amplification model selects the EDFA model s from among its 1, 2,..., S models.

[0180] 1.4. Similarly, for the transceiver, ROADM, Mux, and Demux, the corresponding models, namely the Rx model t, Tx model u, ROADM model v, Mux model w, and Demux model z, can be selected.

[0181] II. Establish the dependency relationships between models and monitor real-time efficiency

[0182] 2.1. The OCH model m depends on the Tx model u, Mux model w, OMS model n, Demux model z, and Rx model t, which is a nested relationship; from the perspective of the OCH transmission process, there is an order dependency relationship among the Tx model u, Mux model w, OMS model n, Demux model z, and Rx model t. A message channel is constructed through the interaction modules of each model, including the registration and processing of request, response, and notification messages, to calculate the efficiency of each model at regular intervals.

[0183] 2.2. The OMS model m depends on the OTS1 model p1 and the OTS2 model p2, showing a nested relationship. The OTS1 model p1 and the OTS2 model p2 have an order-dependent relationship. The OTS1 model p1 depends on the OA1 model q1 and the Fiber1 model r1, presenting a nested relationship. The OA1 model q1 and the Fiber1 model r1 have an order-dependent relationship. The OTS2 model p2 is similar and will not be elaborated here. The OA1 model q1 depends on the EDFA1 model s1, and the OA2 model q2 depends on the EDFA2 model s2, showing a nested dependency relationship. Message channels are constructed through each model interaction module, including the registration and processing of request, response, and notification messages. The efficiency of each model is calculated at regular intervals.

[0184] 2.3. According to the model efficiency monitoring method, when reorganizing the model, a request for reselecting the model is sent to the parent model of its nested relationship, and a notification of suspending calculation is sent to other relevant models simultaneously.

[0185] III. Model Negotiation and Fusion

[0186] Since the process has similarities, only the detailed steps of the fusion of the EDFA1 model s1 and the OA model q1 will be elaborated here. Combining Figure 11 , the model negotiation and fusion process is as follows:

[0187] 3.1. When the input light of the EDFA1 module changes (such as the input optical power becoming significantly stronger or weaker), the EDFA1 model s1 completes a calculation based on the optical sampling data in combination with the selected algorithm, and verifies that its OSNR calculation result meets the normal value range. If it meets the normal value range, step 3.2 is executed.

[0188] 3.2. The EDFA1 model s1 constructs a digital-analog fusion body s1_F and notifies the digital-analog fusion body s1_F to the OA1 model q1.

[0189] 3.3. After the OA1 model q1 receives the notification message and the notification of the digital-analog fusion body s1_F, the data such as OSNR in the digital-analog fusion body s1_F is updated to the OA1 model q1, and calculations are performed according to the algorithm selected by itself. After the OA1 model q1 completes the calculation, a determination is made. If the calculation result is abnormal, it jumps to step 3.4. If it is not abnormal, it jumps to step 3.8.

[0190] 3.4. The OA model q1 sends a new calculation request to the EDFA1 model s1 and marks the calculation request as an OSNR parameter abnormality.

[0191] 3.5. After receiving the calculation request, the EDFA1 model s1 selects an algorithm for trial calculation. And verifies that its OSNR calculation result meets the normal value range, if it meets the normal value range.

[0192] 3.6. The EDFA1 model s1 forms several (≥1) new digital model fusion bodies {s1_F’, s1_F”,...} etc. from a number of trial calculation results, and sends the digital model fusion bodies to the OA model q1.

[0193] 3.7. After the OA model q1 receives the notification message and the digital model fusion bodies {s1_F’, s1_F”,...}, data such as the OSNR in s1_F’ is updated to the OA model q1, and calculations are performed according to the algorithm selected by itself. After q1 finishes the calculation, a determination is made. If all are still abnormal, it jumps to step 3.9. If at least one is not abnormal, it jumps to step 3.8.

[0194] 3.8. Notify the digital model fusion body of the OA1 model q1 to the OTS1 model.

[0195] 3.9. The OA1 model q1 needs to reselect a new model q1’ for the EDFA1 model s1, reduce the usage frequency of the OA1 model q1 by one, and send a notification to the EDFA1 model s1 to pause the calculation.

[0196] 3.10. And so on, finally complete the calculation of the OCH model, and send out the relevant digital model fusion bodies through the notification message.

[0197] Embodiment 3:

[0198] Based on the fusion method provided in the above embodiment, the present invention further provides a fusion device that can be used to implement the above method, as Figure 12 shown, which is a schematic diagram of the device architecture of an embodiment of the present invention. The fusion device of this embodiment includes one or more processors 21 and a memory 22. Among them, Figure 12 One processor 21 is taken as an example.

[0199] The processor 21 and the memory 22 can be connected through a bus or other means, Figure 12 Taking connection through a bus as an example.

[0200] The memory 22, as a non-volatile computer-readable storage medium for a fusion method, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the fusion method in Embodiment 1. The processor 21 executes various functional applications and data processing of the device by running the non-volatile software programs, instructions, and modules stored in the memory 22, that is, implements the fusion method of the foregoing embodiment.

[0201] The memory 22 may include high-speed random access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 22 optionally includes a memory remotely disposed relative to the processor 21, and these remote memories can be connected to the processor 21 through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0202] The program instructions / modules are stored in the memory 22 and, when executed by the one or more processors 21, perform the fusion method in the above embodiments. For example, perform each of the steps described above Figure 2 shown.

[0203] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium may include: a read-only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, an optical disc, or the like.

[0204] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fusing network digital models, characterized in that, it includes: Obtain the carriers included in the target service, the topological relationship between the carriers, and the service flow direction, and obtain at least one alternative model corresponding to each carrier; For each of the carriers, obtain the model capabilities of each of the alternative models, and select a suitable alternative model as the target model corresponding to the carrier according to the model capabilities; Construct the dependency relationship between the target models according to the service flow direction and the topological relationship, and construct the message channels between the target models according to the dependency relationship; Receive interaction messages from other target models based on the message channels, and determine whether to update the model and the model relationship according to the interaction messages; including: monitoring the model efficiency of itself, and judging the model state of its own model according to the model efficiency; If its own model is in a fault state, after receiving a request message from another target model, send a response message of model fault to this model, and report the message of model fault to the parent model, and the parent model reorganizes the model again; if its own model is in an idle state, after receiving a request message from another target model, its own model performs digital-analog fusion calculation according to the request message to obtain a digital-analog fusion body, sends the digital-analog fusion body to all models that have established a registration relationship with its own model, and determines whether to update the model and the model relationship according to the digital-analog fusion body.

2. The fusion method according to claim 1, characterized in that, The obtaining of the model capabilities of each of the alternative models and selecting a suitable alternative model as the target model corresponding to the carrier according to the model capabilities includes: Obtain the algorithm characterization capabilities of the algorithms corresponding to each alternative model, the model usage frequency within each time period, and the parameter list satisfaction degree; Obtain the product of the algorithm characterization capabilities, the model usage frequency, and the parameter list satisfaction degree, and use this product as the model capability of the alternative model; Sort in descending order of model capabilities, and start selecting from the one with the largest model capability until the target model is selected.

3. The fusion method according to claim 2, characterized in that, The obtaining method of the model usage frequency is: Obtain the number of times the model is used within the time period. Among them, each time the model is used, the number of times the model is used is incremented by one; each time the model is abandoned, the number of times the model is used is decremented by one; after each time period ends, the number of times the model is used is cleared and re-counted; Calculate the ratio of the number of times the model is used to the time period, and use the value corresponding to adding a random number to this ratio as the model usage frequency.

4. The fusion method according to claim 2, characterized in that, The obtaining method of the parameter list satisfaction degree is: Obtain performance metric indicators, and determine the minimum parameter set according to the performance metric indicators; Obtain the probe components of each alternative model, and judge whether the detection accuracy of each probe component meets the preset accuracy. If it meets, add the parameter type corresponding to this probe component to the alternative model parameter list; Judge whether all the parameter types included in the minimum parameter set exist in the alternative model parameter list; If all the parameter types included in the minimum parameter set exist in the alternative model parameter list, the satisfaction degree of the parameter list is 1; If at least one parameter type does not exist in the alternative model parameter list, the satisfaction degree of the parameter list is 0.

5. The fusion method according to claim 1, characterized in that the constructing of the dependency relationship between target models according to the service flow direction and the topological relationship, and the constructing of the message channel between target models according to the dependency relationship includes: constructing the sequential dependency relationship between target models according to the service flow direction, and constructing the nested dependency relationship between target models according to the topological relationship; setting a unique model identifier for each target model; for multiple target models with nested dependency relationships, the sub-model registers its model identifier to the parent model, and the parent model stores the received model identifier into the registration list; wherein, the parent model is the target model at the upper layer of the nested tree structure, and the sub-model is the model at the lower layer of the nested tree structure; for multiple target models with sequential dependency relationships, the computing upstream model registers its model identifier to the computing downstream model, and the computing downstream model stores the received model identifier into the registration list; wherein, the output of the computing upstream model is used as the input of the computing downstream model; constructing the message channel between target models based on the registration list of each target model.

6. The fusion method according to claim 1, characterized in that after receiving the request message from other target models, the own model performs digital-analog fusion calculation according to the request message to obtain a digital-analog fusion body, sends the digital-analog fusion body to all models that have established a registration relationship with the own model, and determines whether to update the model and the model relationship according to the digital-analog fusion body includes: after receiving the request message from other target models, the own model performs a trial calculation according to the request message to obtain the instantiated value of the performance metric corresponding to the own model; judging whether the instantiated value of the performance metric satisfies the normal value range, if it satisfies the normal value range, forming several digital-analog fusion bodies from several trial calculation results, and sending the digital-analog fusion bodies to the computing downstream models that have established a registration relationship with the own model in the order of first sequential dependency and then nested dependency; receiving the calculation results of the digital-analog fusion body from the computing downstream models, if all the calculation results are abnormal, the parent model corresponding to the own model reorganizes the model again, and reduces the number of times the model corresponding to the own model is used by one; if at least one calculation result is normal, continue to notify the digital-analog fusion body to the computing downstream models until the calculation of the models corresponding to the target service is completed.

7. The fusion method according to claim 6, characterized in that after judging whether the instantiated value of the performance metric satisfies the normal value range, it further includes: if it does not satisfy the normal value range, an abnormal calculation request is sent to the computing upstream model in the order of first nested dependency and then sequential dependency to determine the abnormal source.

8. The fusion method according to claim 1, characterized in that Monitoring the model efficiency of itself and judging the model state of its own model according to the model efficiency includes: Obtain the first ratio between the actual accuracy and the promised accuracy of the model, the number of requests n within the time period rq , the number of response executions n within the time period ex , the number of normal responses n within the time period rs and the number of abnormal abandonments n within the time period ab ; For the number of requests n rq , the number of response executions n ex , the number of normal responses n rs and the number of abnormal abandons n ab perform summation to obtain a summation value, and calculate the second ratio between the number of normal responses n rs and the summation value; Taking the product of the first ratio and the second ratio as the model efficiency; If the model efficiency is not less than the set threshold, the model is in a busy state; if the model efficiency is less than the set threshold, the model is in an idle state; if the model efficiency is 0, the model is in a fault state.

9. A fusion device for a network digital model Characterized in that It includes 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 after the instructions are executed by the processor, they are used to complete the fusion method according to any one of claims 1-8.

Images