Research and Development of Operational Technologies Using AI for Network Robustness and Business Automation

3.1 NOIM

We are investigating a network resource management technology based on a data model called Network Operation Injected Model (NOIM) for cross-network information utilization and analysis [2]. This technology expands the standardized data model and enables the integration of multi-layer and multi-domain network data by expressing the connections and relationships among network elements in a simple and unified format on the basis of “points” and “lines.”

When network configuration information is managed with separate data models and separate system databases that differ for each layer and service, it is necessary to analyze each layer and service separately and compile the results. By integrating network configuration information into a unified data model that does not depend on layers or services, this technology promotes network operation and system unification across different organizations and enables network data to be used as it is for data analysis across layers and services, which is expected to lead to quick recovery from failures and disasters and accurate understanding of the impact.

3.2 D-AINA

We are investigating Distributed Data Alignment with Intelligent Network Analysis (D-AINA), a technology for detecting and correcting inconsistencies in network information that is managed separately. In some cases, data related to originally identical data may be registered in multiple data with different notations. If the data are integrated with such inconsistencies, correct analysis cannot be conducted, and this becomes a major issue in the integration of network information with NOIM.

In the conventional technology, inconsistencies are often detected by comparison on the basis of the character string representation, and it is difficult to detect inconsistencies when the same element is registered separately with widely different character strings. Therefore, we focused on using “connection information,” which is a major feature of network information.

For example, by comparing the graph structure on the basis of the network topology in each layer with the conventional character string comparison for data in a network consisting of multiple layers and searching for the optimal correspondence relationship between layers, it becomes possible to detect that the same element is represented even if the data differs greatly as a character string. Therefore, the integration of network data is promoted by eliminating the notation difference between databases with high accuracy.

3.3 Konan

We are investigating Knowledge-based Autonomous Failure-event Analysis Technology for Networks (Konan), a rule-based fault point estimation technique, for speeding up fault isolation when a failure occurs [2]. With the advancement of communication technology, facilities, and services, failures are becoming more diverse and complex, and the need for technology to analyze and identify fault points and causes quickly and accurately is increasing. With this technique, by analyzing the alarm generated around the cause point at the time of failure, the alarm which characterizes the failure is automatically extracted from a small number of failure cases, and the rule that estimates the fault point and cause is generated. By using these rules during operation, the location and cause of a failure can be instantly estimated from the alarms that have been generated. Thus, alarm analysis for various and complex failures can be standardized, and rapid failure response can be achieved.

By adopting the NOIM data model for managing network information used for rule generation and estimation, we have succeeded in generating estimation rules that take into account the relation between alarms and failure causes across layers for failures in complex multi-layer networks.

4.1 Intent extraction technology

We are investigating an intent extraction technology for extracting network usage requests for capturing ambiguous intentions. We are developing a function that automatically extracts the user’s ambiguous service-use requests, which we call intents, from business information and other data and from interactions with the user. AI then outputs the communication requirements. By inputting these requirements into network control functions such as Cradio^® [3], a series of automatic controls from extracting intent to providing optimal network services becomes possible.

When extracting intent, this technology automatically collects external information on work schedules and user profiles and interviews about missing information in a simple dialogue. This is a significant advantage for users who do not have knowledge or skills in network configuration. Optimizing network quality can also improve quality of experience and reduce power consumption.

We aim to expand the scope of automation by applying this intent extraction technology to network operations and extract worker intent in various tasks.

4.2 ATOMN

We are investigating Autonomous Agent for Network Operation (ATOMN), which is a technology to achieve quick and accurate network operation with an autonomous AI agent.

For example, when restoring a network failure, operators may review the procedure and proceed with it while checking the status of network equipment. However, they may create the procedure by referring to various documents on the basis of the expertise of experienced operators, and when executing the procedure, they may encounter a situation that is different from the expected one, requiring a high level of skill, expertise, and a huge workload.

Therefore, by dividing, formulating, and layering the information necessary for creating such procedures in advance into a format that is easy for the AI agent to interpret, the agent can create an execution procedure, such as a command to be input and an expected response, while using the generative AI, and autonomously execute it while checking the status of network devices. When there is a mixture of operations that require human labor, such as the replacement of parts, the AI agent works with the workers to execute operations such as command input remotely. This makes it possible for the workers alone to execute operations that used to be carried out by multiple workers communicating and coordinating. By achieving autonomous execution of such complex procedures, maintenance work can be greatly reduced.

4.3 Task process generation technology

As a technology for automating non-routine tasks, we are investigating task process generation technology. Even in non-routine tasks that require customized responses for each customer, there is a strong demand for improving overall operational efficiency. However, non-routine tasks are challenging to automate because their processes are prone to change, making it difficult to create automation scenarios or standardize processes, hindering task handover and automation. This technology uses operation logs obtained from workers’ PCs and document information such as business procedure manuals and specifications to automatically generate work processes tailored to the current operational status using AI. This enables formalizing processes tailored to specific cases or situations, leading to business improvements through enhanced process understanding. By integrating it with automation technologies such as RPA, digital adoption platform, and intelligent process automation, automation of non-routine tasks is anticipated.

5.1 Human condition estimation technology

We are investigating human condition estimation technology as a means of preventing accidents by measuring biological and internal factors of humans and accurately assessing their work status. To eliminate accidents involving workers during on-site operations, it is necessary to prevent errors caused by workers’ cognition and judgment. However, in actual work environments, it is difficult to monitor cognitive states in real time due to factors such as differences in work environments and individual worker characteristics. This technology uses information on workers’ eye movements, such as gaze and pupil movement, to estimate human-concentration levels in real time and with high robustness. This technology enables real-time monitoring of workers’ concentration levels even in diverse situations, such as off-site work, where various conditions and individual differences among workers are present.

In the short term, we envision applications such as evaluating participants’ internal responses and cognitive characteristics in augmented reality (AR)-based hazard perception training. In the future, we aim to predict cognitive errors in real-world tasks in real time and prevent accidents, enabling safe, secure, and stable operations for everyone.

5.2 Holographic hand modeling technology

We are investigating holographic hand modeling technology to convey remote instructions intuitively. NTT’s operations involve a significant amount of manual work related to the maintenance of infrastructure facilities of various sizes. For less experienced workers, mastering the precise manual operations involving hands and tools within a short period is challenging, making remote guidance and support from an instructor highly effective. However, conveying such manual tasks using voice alone is difficult and can hinder workers’ understanding. This technology displays a holographic hand model replicating the movements of an instructor’s hands in real time on AR goggles, visually presenting the instructions and coordination points required for manual tasks. It enables the intuitive transmission of the intent behind actions and operations, allowing even inexperienced workers to perform intricate manual tasks effectively.