The world’s first sound-based AI Multi-Agent system
for Advanced Quality Control and Predictive Maintenance

Quality control and predictive maintenance on the edge.

iDetect-4.0 is an AI operating system that offers audio based advanced quality control and predictive maintenance capabilities through AI-based sound analysis. iDetect-4.0 is designed to run seamlessly on Edge industrial PCs, offering advanced capabilities in maintaining deep learning AI models, managing data of interest, and supporting quality processes.
(request technical documentation: support@bondzai.io)
Download datasheet

Each operating mode (quality, maintenance) is powered by a dual-headed deep learning AI (DavinSy). One creates deep learning models used in production, and the other continuously improves model performance based on corrections made by the operator and validation by the quality engineers.

Ultimately, iDetect-4.0 empowers autonomous Machine Learning Operations (MLOPs), eliminating the need for additional infrastructure while significantly enhancing manufacturing processes.

iDetect-4.0 interfaces


iDetect-4.0 offers a bi-directional communication with the PLC in charge of controlling the mechanical processes of the machines. From the machine, iDetect-4.0 gets the start and stop events, the different custom attributes (meta-data) and the process overall status (errors and exceptions). To the machine, iDetect-4.0 sends the result of quality control inference, up to the PLC to determine which action to take regarding those results. Exchanges take the shape of a registry sent over a socket.


iDetect-4.0 offers 3 different views for 3 different roles:

  • A process monitoring view for the operator who gets the possibility to correct or enrich inference predictions,
  • A model management view for the quality expert, who can create, qualify and deploy new models in a live fashion,
  • A configuration view for the administrator, chow can modify look and feel of the UI, and the cleaning and backup mechanism.


iDetect-4.0 supports up to 4 microphones at 48Khz to listen to the machine. Reference implementation by Optimaxes relies on an Advantech -USB 5801.


All data and processes are local and autonomous. This is also true for the MLOPS features. The only effort requested for IT is the periodic backup of the system. iDetect-4.0 makes sure to keep the storage size under control, and thus regularly backup and clean its data (records, models, inference results). The backed files are stored in a shared folder, up to IT to monitor and archive those files).


iDetect-4.0 relies heavily on DavinSy agents’ collaborative possibilities. Each machine mode is associated to a specific agent, a central agent is also responsible of qualifying the different created models following the quality policy.  Inside each agent several successive models allows for an agile and explainable detection of issues and anomalies.

iDetect-4.0 Functional description

application services layer


Several complex mechanisms are at work in iDetect-4.0, several real-time events are to be processed transparently, the supervisor (called super-agent in iDectect-4.0) is in charge of the orchestration of it making sure signals are processed by the correct agent, results are gotten in time and models are deployed without hindering the production cycle.


Receive the audio signal through a socket, store it on the local storage, create a record in the database and forward it to the agent through the gateway.


Receive through a table format the different registers provided to iDetect-4.0 and store those values in the database, along with the audio signal, See further for a detailed format of this table.


iDetect-4.0 comes with an off-the-shelf UI that can be used directly. Parts of this UI can be configured through Maestro. The UI relies on a web socket API. If you want to write your own UI, refer to the API Documentation.


iDetect-4.0 provides a high-level API to build UI for operators. We will describe, later, the functionalities provided by this API.


iDetect-4.0 runs its own database management. It stores the history of captured records along with the inference status, the corrections, the meta data coming from automation. It stores also, the different models and their status, as well as the local configuration of the iDetect-4.0 instance.


Depending on administrator’s configuration, iDetect-4.0 will run cleaning and archiving tasks both on audio records and database content. Configuration and usage will be detailed later.


When a signal is received, it is forwarded to intended agents (depending on the source of the signal and the selected mode) that will classify it and gives a result indicating the quality of the produced item. The result is sent back and stored into the database. DavinSy agents rely on their built-in communication capabilities to send events to the super-agent and refresh the UI.


A Specific DavinSy Agent is dedicated to the qualification process for new models. On demand from a user with the quality role, a new model can be generated gathering a test dataset following the settings of the quality policy. As background task several inferences are run on each record of the test dataset and a quality score is computed. This new model if satisfying can be deployed without stopping the production.


As described before users with different roles will interact with iDetect-4.0: the operators, the people in charge of quality and people in charge of administration. Those different roles get different rights and access to different functions. As for now, iDetect-4.0 handles a simple login/password scheme.

iDetect-4.0 Implementation

application services layer

iDetect-4.0 is an AI operating system that offers audio based advanced quality control and predictive maintenance capabilities through AI-based sound analysis. iDetect-4.0 is designed to run seamlessly on Edge industrial PCs, offering advanced capabilities in maintaining deep learning AI models, managing data of interest, and supporting quality processes.

 6 classes of service coexist in iDetect-4.0 system:

  1. The super-agent acting as the orchestrator of the system,
  2. The agents responsible for model generation and model execution,
  3. The gateway which routes the different messages between the services,
  4. The User Interface back-end which serves the Web UI used by users,
  5. The CRON micro-service in charge of backup and cleaning of the data.
  6. The database manager which runs a Postgres server.

In addition, 3 extra services run outside the docker, directly on the host:

  1. The Audio driver, responsible for capturing signal coming from the sound card, If you want to implement such a component for your own sound card, details are to be found in the API Documentation.
  2. The communication service, making the bridge between the automaton and, iDetect-4.0, If you need to write such a component for a particular protocol not yet supported by iDetect-4.0, the protocol is details in the API Documentation.
  3. The licence manager is responsible for checking that agent licence grant it the right to run on this specific hardware

When a persistent storage is needed (for raw records, database or agents files) a volume is created to allow an update of the container containing the code without losing the data.

    DavinSy initials implemented characteristics

    Feature Value Description
    Max. number of sources 4 iDetect-40 can process up to 4 different microphone streams
    Records size <60s, 48KHz High definition capture
    Attributes per record <=10 Attributes are sent by PLC and stored in iDetect-4.0 DB.
    Max. number of modes 4 A mode is a change in the production process whose differences implies to create a separated model.
    Max. number of agents 5 One agent per mode + one agent for qualification.
    Time to generate a new model < 5 minutes Model generation time depends on the number of records that where corrected.
    Inference time < 100ms Real time results
    Accuracy > 90% Accuracy depends on quality of training dataset.
    Min. CPU Core I5 Low-end PCs without GPU are okay
    Min. RAM 12 Gb Mainly needed by docker
    Host OS Windows 10 Can work on Mac and Linux but need porting of audio and PLC connectors
    Meet Bondzai