When teams look for files using visual similarity, spoken content, or contextual semantics, old-style search tools fall short because they rely on manual tags or simple keyword indexing, which fail to understand the meaning of this content. Instead of just scanning filenames or descriptions, semantic and multimodal search systems turn text, images, video, and audio into a shared semantic space that enables retrieval based on meaning rather than exact metadata matches.

The Challenge:

A leading retail/e-commerce company is drowning in digital files – countless product images, raw video clips, unfinished ads, audio tracks piling up daily. This flood of data refused to slow down. Managing it became nearly impossible. Files piled higher every week. The search took forever.

Hours slipped away as video editors dug through folders, not timelines. Marketing teams lost momentum searching for past campaign assets rather than planning new launches. Old visuals got rebuilt again and again – just because nobody could track them down fast enough. Time meant for real tasks bled into endless searches across cluttered drives.

The media library contained an estimated 25–30% duplicate assets. Multiple outdated or unapproved versions mixed with new ones. Team members guessed where things might be. Some assets vanished entirely. Others got reused by accident. Time slipped away on busywork instead of real tasks. Mistakes crept into live campaigns. Frustration grew behind closed doors. The core issues were:

• Search by meaning wasn’t possible
• Duplicate content and low discovery
• Slow workflows and high operational costs
• Lack of multimodal search support

Put simply, the team wanted a smarter method for organizing digital files – one that understood meaning in images, audio, and text instead of relying only on labels – so finding and using old material became faster during projects.

The Wish List:

So the company set clear goals meant to make a real difference both technically and commercially.

• Contextual search without manual tagging.
• Ability to search media using text, image, or audio inputs.
• Faster indexing of large volumes of video and image data.
• A cost-efficient alternative to metadata-heavy or LLM-centric solutions.
• Seamless integration with the existing MAM/DAM platform.
• Measurable ROI, driving faster discovery, lower content creation costs.

The Solution:

One step ahead, the team brought Gyrus AI Semantic Media Search and integrated it into their media/digital asset management setup. Mostly behind the scenes, it works by understanding content deeply before delivering results.

1. Contextual search, no tagging needed – Editors could now just type simple queries like “product unboxing close-up” or “model wearing blue jacket” and instantly find the scene they were looking for.
2. 80% faster processing speed – An hour of video gets indexed in ~ 5 minutes by an RTX 3090/4060.
3. Up to 10× more cost-effective – Our solution was able to deliver the most cost savings when compared to metadata-heavy or LLM-based solutions.
4. Compact multimodal model – It is optimized to process video, audio, and images while staying lightweight and efficient.
5. Flexible deployment – Able to run on-prem aligning with enterprise requirements.

The Results:

After integrating Gyrus AI Semantic Media Search into its existing Media Asset Management platform, the retail/e-commerce company observed the following measurable outcomes:

Now operations run faster because the media library has become simpler, cost-effective, and one piece feeds many tasks. Savings add up when files get reused instead of remade each time. Workflows feel smoother since assets load more quickly across online stores. The whole setup adapts easily as needs shift.

The post How Semantic Media Search Helped a Retail Company Create Marketing Assets Faster. appeared first on Gyrus AI | Blog | Insights on AI & Intelligent Media Search, In-scene Ad Placement, Automated Video Anonymization Technologies.

But broadcasters these days are not satisfied with just storage anymore. They are demanding speed, intelligence, and cost efficiency. They have to find events of interest in video files based on context and not just titles or manual tags. This is the point where traditional metadata search falls short and contextual AI search proves its value.

The Challenge: 

A European broadcaster was evaluating new MAM platforms. Their biggest frustration they faced was in the search part:

• Manual tagging was slow and inconsistent, and the editors wasted hours tagging the footage or searching for moments based on incomplete metadata.
• LLM searches were not as affordable as they thought – Attempts were made to implement actually working solutions based on large language models, but the cost was too high to scale.
• Workflow delays – To find the right clip, their team often had to scrub through the entire footage, relying mostly on luck and sometimes spending hours just to locate a single scene.

In short, this broadcaster wanted a system/solution that:

• Not only organizes the media library but also makes finding relevant clips fast and hassle-free.
• Would search for scenes contextually, without any tags or metadata.
• Fast, affordable, and flexible (either on the cloud or on-prem).

The Solution:

A Media/Digital Asset Manager bidding for this customer integrated Gyrus AI Semantic Media Search into their Media Asset Management platform, delivering advanced search capabilities. Here’s what stood out:

• Contextual search, no tagging needed – Editors could now just type simple queries like “goal celebration” or “sunset cityscape” and instantly find the scene they were looking for.
• 80% faster processing speed – An hour of video gets indexed in ~ 5 minutes by an RTX 3090/4060.

• Up to 10× more cost-effective – Our solution was able to deliver the most cost savings when compared to metadata-heavy or LLM-based solutions.
• Compact multimodal model – It is optimized to process video, audio, and images while staying lightweight and efficient.
• Flexible deployment – Able to run on-prem or in the cloud, depending on broadcaster needs.

Key Technologies Behind It.

Our Semantic Media Search features foundation multimodal models, similar in lineage with CLIP (Contrastive Language-Image Pre-training), CLAP (Contrastive Language-Audio Pretraining), and advanced video-language encoders:

• Extract the features from video, audio, and text.
• Convert them into semantic embeddings (digital fingerprints of meaning).
• Store embeddings in a vector database for really fast retrieval.

This way, the queries like “black car entering the scene” return the clips very relevant to such a scene, even if there is no actual metadata describing those clips.

The Benefits for the Broadcaster

After testing the Gyrus AI’s Semantic Media Search enabled MAM, the broadcaster immediately saw the difference and the impact was clear:

Result:

The broadcaster, after testing the solution, decided to migrate to a new MAM platform that had integrated Gyrus AI’s Semantic Media Search feature. Therefore, two big challenges were met at once:

On one hand, the broadcaster gained a cost-effective, AI-powered search solution; on the other, the MAM provider differentiated its platform with intelligence that competitors lacked.

On the broadcaster’s side, it meant faster turnarounds, lower operational costs, and a reliable system that scaled without adding technical complexity.  For the MAM player, it meant adding the large enterprise customer that is always looking for differentiated value – a MAM that really has intelligent contextual search and media management that is future-ready.

Future Outlook.

As multimodal AI continues to evolve, semantic search will also expand to deliver:

• Personalized Search: Results tailored to the project context or user history.
• Deeper Insights: Automated clustering, thematic mapping, and trend analysis of archives.
• Predictive Recommendations: Suggestions of content based on cultural context and storytelling patterns.

Conclusion

The use of MAM + Semantic Media Search made search and retrieval operations nearly 8× faster, while delivering a solution that was 10× more cost-effective than regular LLM-based or manual metadata approaches. The system was able to deliver real-time speed and scale without sacrificing accuracy.

This case highlights that the future of Media Asset Management is AI-empowered contextual intelligence-solutions that can flexibly be deployed on-prem or on the cloud, accordingly adapted to broadcaster needs, and capable of technically accommodating an exponential growth in content demands.

The post How Gyrus AI Search Turned a Regular MAM into a Smart Solution – and Helped Win Over a Broadcaster. appeared first on Gyrus AI | Blog | Insights on AI & Intelligent Media Search, In-scene Ad Placement, Automated Video Anonymization Technologies.

Industry Challenge:

Many automotive companies with the vision of achieving self-driven cars often struggle with the time-consuming and labor-intensive process of anonymizing training data. This task is very important when handling personal information, particularly in areas requiring conformity to the GDPR and other data protection laws, including the need to anonymize PII (Personally Identifiable Information) to comply with GDPR. When it is done manually, it becomes tiresome and time-consuming when blurring the license plates of vehicles or faces of people, and it can be so incorrect with inconsistent outcomes.

Client’s Specific Problem:

The client’s team was manually blurring license plates and faces in their video datasets, a process that was:

1. Extremely time-consuming: Taking up to 1,000 hours per month

2. Prone to human error.

3. Inconsistent across different team members.

4. A bottleneck in their development pipeline: Delays extending project timelines by 30%

This manual process significantly slowed down their whole processes and also hampered the sharing of data to their global teams because of the GDPR regulations.

Our Solution:

We proposed our state-of-the-art Video Anonymization AI model, leveraging advanced AI and ML techniques. Key features of our solution include:

1. Automatic Detection and Blurring: Identifies and blurs license plates and faces with ~98% accuracy.

2. Scalability: Processes large volumes of video data quickly, reducing processing times by 85%.

3. Consistency: Ensures uniform anonymization across all processed videos, eliminating human inconsistencies with 100% uniformity.

4. GDPR Compliance: By automatically anonymizing sensitive data, the solution helps ensure compliance with GDPR and other privacy regulations.

5. Lightweight & Easy Integration: Easily integrates with existing systems, runs on lighter GPUs and CPUs, and can be deployed on cloud, on-premise, or edge devices.

Implementation:

Gyrus team worked closely with the client to deploy our AI model on their own servers, utilizing their GPUs within their premises.

1. Dataset Collection: The client provided their sample video datasets for training and enhancing the model.

2. Data Preparation: In the next stage, the customer data is cleaned, enriched, and made ready for data science. At this stage, the data quality is measured and all the data enhancement techniques are implemented.

3. Model Development: Gyrus customized the video anonymization AI model using transfer learning, starting with prebuilt models tailored to the client’s specific needs.

4. Quality Assurance: The model underwent explainability checks, bias evaluations, and differential privacy enhancements to ensure high accuracy and GDPR compliance.

5. Deployment: The AI model was deployed on the client’s in-house servers, leveraging their existing GPU infrastructure for efficient processing.

6. Continuous Improvement: The model was regularly updated with new data and features to maintain accuracy, consistency, and reproducibility over time.

This approach ensured the client’s data was processed efficiently, securely, and in full compliance with GDPR and other privacy regulations.

The Impact:

After implementing our AI-powered video anonymization model, the client experienced significant improvements in their development process:

1. Time Efficiency: The automated process reduced the time required for video anonymization by over 90%, allowing faster iteration in model training.

2. Improved Data Quality: Consistent and accurate anonymization significantly improved the quality of training datasets.

3. Enhanced Collaboration: GDPR-compliant datasets could be easily shared among global teams, fostering better collaboration and faster development cycles.

4. Cost Reduction: Significant cutting down of the amount of manual labor required, reducing labor costs by 70%.

5. Scalability: Enabled the client to process and use 5x-10x more training data, leading to more credible autonomous driving models.

Conclusion:

Thus, by solving their major concerns and offering an effective, fully automated solution, we not only helped them meet the privacy regulations but also contributed to enhancing their operational efficiency and data quality upgrades, and significantly reducing the overall project cost.

This case study highlights the potential of AI-driven solutions in the automation of data processing channels particularly in industries that value data confidentiality.

For other businesses that may encounter the same challenges while looking for ways on how to protect client’s data, Gyrus provides them with a sound method of protecting data while at the same time benefiting from useful data.

Keep in touch with us to find out how we can assist your organization in getting past data protection challenges and reaching your objectives. Write to us at info@gyrus.ai for more information.

The post How Gyrus AI Helped a Top European Automotive OEM to cut Video Anonymization Time by 90%. appeared first on Gyrus AI | Blog | Insights on AI & Intelligent Media Search, In-scene Ad Placement, Automated Video Anonymization Technologies.

Introduction

Enterprises often face challenges when it comes to managing their warehouses and ensuring the smooth transfer of products. With an enterprise growing big, the challenges of managing inventory also rise. This results in enterprises spending a large amount on managing the inventory leading to large expenses.

However, in today’s digital era, enterprises are adopting technology for all critical operations to accelerate their digital transformation journey. As per PwC, AI is gradually transforming the corporate world. It can contribute up to $15.7 trillion to the global economy by 2030. Here are some of the areas where enterprises face challenges in inventory management and implementing proper ML/AI algorithms can have a significant impact to the bottomline.

Challenges of Traditional Inventory Management

Today inventory management is a semi-automatic process. This means the decision-making process remains with the humans as they derive insights from loads of datasheets. Rule-based traditional models are complex to maintain. With enterprises growing in size, the amount of data generated also multiplies. This pressurizes the human staff to put in more effort and time, leading to tedious, time-consuming efforts. This leads to endless process of data management. 

1. Data explosion – Data is an important resource to understand product performance and buyer behavior. However, it has suddenly become a menace difficult to control. If decades ago, getting accurate data was a problem, today it is its abundance that is creating chaos. The process of controlling and analyzing data has become tedious. This has put a lot of pressure on teams to simplify data management. While automated systems can help track and store data; processing requires a different set of skills and efforts from the human staff to process all of it.
2. Data tracking – Tracking items in the inventory is increasingly becoming tedious, let alone deriving insights from it. A slight miss in tracking the incoming and outgoing items can hugely impact the product’s performance in the market. This results in demand fluctuations, customer losses and eventually impacting the turnover negatively.
3. Business planning – Inventory management plays an important role in business planning and strategizing. Only accurate data can help plan a product roadmap or promotional activity, leading to successful results. This requires real-time collection and processing of all data, which drives accurate decisions on time.
4. Operational costs – Huge inventories mean more products, more markets, more delivery, more space, and more staff. This often increases the operational costs of maintaining the inventory taking the share from other essential expenses.

How can AI help optimize inventory management?

Enterprises have always leveraged technology to gain a competitive edge. From automating repetitive processes to implementing cognitive abilities; enterprises across the world have come a long way in terms of conducting business. While automation has become quite commonplace, it is artificial intelligence and machine learning that are gaining traction. This may be due to their ability to optimize processes and provide actionable insights, which further accelerate businesses. As per a recent 451 Research Report, Accelerating AI with Data Management; 68 percent of respondents reported that they are either already using ML capabilities; or plan to within the next three years.

At Gyrus AI, we believe focused AI implementation and leveraging machine learning capabilities can help optimize inventory management; just like other business operations. Let’s look at what enterprises can achieve in inventory management with AI & ML.

AI provides visibility into inventory

Clear visibility into product performance and markets, helps devise sales and promotional strategies. Along with the uniformity and level of granularity of the data; the accuracy of strategy also depends on the timeliness of product data. Visibility into product data can avoid overstocking and understocking.

Inventory can derive greater value from adopting AI according to McKinsey’s global survey. Almost 76% of the survey respondents working at supply chain companies have reported moderate to significant value from deploying AI. AI enables timely, accurate insights and visibility into sales, order, and inventory data through a robust data pipe. This avoids a lot of cacophony in the market and helps deliver good quality data on time. Not only is AI capable of providing insights that were previously unavailable; but it can also take into account various aspects that in some way influence the demand. 

AI empowers data mining

Machine Learning capabilities enable AI systems to learn from the data patterns and predict demand for the product. For example, it can help plan inventory in certain geographical locations, as per the seasonal demands. Analyzing the situation can help suggest to overstock or replace the stock with relevant products on the shelf. This helps reduce the pressure on the sales and marketing research teams. 

AI improves stock management

Managing stocks well can result in improved customer satisfaction and a sense of fulfillment. A slight deviation in this can lead to shortages or delays; adversely affecting the demand for the particular product, negatively impacting the revenue. Moreover, AI helps stock right with its ability to analyze customer behavior patterns with the help of big data. This minimizes the risk of mismanagement of stocks resulting in improved customer satisfaction. Insights derived from data mining also helps plan the transportation of perishable goods from factories to retail outlets in time; reducing wastage and impacting sales. 

Why are enterprises still skeptical about the use of AI & ML?

Though leveraging AI and ML capabilities have proven beneficial in inventory management; there are possibilities of setbacks while implementing AI tools. However, being more in number, the advantages often shadow the setbacks. Nevertheless, it is important to make the team aware of the complications, that may arise during the implementation of AI. Awareness of potential risks and challenges can help reduce or alleviate the risk to a great extent. 

There are mainly two areas, where AI implementation and running can turn complicated. Firstly, data presentation. For an AI system to work successfully, you must present it with high quality data. It is with this data, that AI analyzes and derives actionable insights. However, enterprises often find it a task to gather data and feed the systems.

Secondly, implementing AI into existing systems can be time-consuming. Data and processes often load legacy systems; making it difficult for AI systems to sit atop them and integrate. However, these are just a few measures of caution and should lead you to a successful AI implementation. Just ensure to prepare the staff well beforehand.

Bottom-line

There’s no doubt, artificial intelligence and machine learning are making great strides in transforming inventory management. It is empowering enterprises to handle both physical tasks and complex data-related tasks; and providing complete control over their inventory while adding value to their business planning and consumer behavior predictions. 

According to research, by 2021, inventory management software in the retail sector will see a CAGR of 8.3%. Another research forecasts that the global enterprise cloud services market would grow at a CAGR of 23.83% by 2020. Well! We can already see the trend of applying AI to inventory management picking up pace. Various enterprises leverage AI-powered inventory management tools to their benefit.

AI is a great boon to larger and medium enterprises. Especially those who deal with large inventory and aren’t able to control and manage their inventory data. By handling enormous data and repetitive processes; AI lets human workers focus on more strategic and important tasks at hand. In a few years from now, we will see enterprises leveraging transformative technologies to optimize inventory management.

はじめに

企業は、倉庫管理と製品の円滑な輸送確保に関する課題にしばしば直面します。 企業が成長し大きくなるにつれ、在庫管理の課題も増えています。 その結果、企業は大きな出費につながる在庫管理に多額の費用を費やしています。

ただし、今日のデジタル時代では、企業はすべての重要な運用にテクノロジーを採用して、デジタル変革の旅を加速しています。 PwCによると、AIの導入により、は徐々に企業の世界を変革されていくとしています。 2030年までに世界経済に最大15.7兆ドルの貢献をする可能性があります。このブログでは、企業が在庫管理で課題に直面し、適切なML/AIアルゴリズムを実装することが収益に大きな影響を与える可能性がある分野をいくつか紹介します。

従来の在庫管理の課題

今日、在庫管理は半自動プロセスです。これは、大量のデータシートから洞察を引き出すため、意思決定プロセスは人間に委ねられていることを意味します。ルールベースの従来のモデルは、維持するのが複雑です。企業の規模が大きくなるにつれて、生成されるデータ量も倍増します。これにより、人間のスタッフはより多くの労力と時間を費やすようになり、退屈で時間のかかる作業につながります。 これは、データ管理の無限のプロセスにつながります。

1. データの急増 – データは、製品の性能と購入者の行動を理解するための重要な情報源です。しかし、突然、制御が困難な脅威になりました。数十年前は正確なデータを取得することが問題でしたが、今日では、そのデータ豊富さにより問題が生み出されており、データの制御と分析のプロセスは面倒になりました。これにより、データ管理を簡素化するようAI導入チームに大きなプレッシャーがかかりました。自動化されたシステムは、データの追跡と保存に役立ちますが、すべてを処理するためには、人間のスタッフとは異なる一連のスキルと努力が必要です。
2. データのトラッキング – 在庫内のアイテムの追跡は、そこから洞察を引き出すことは言うまでもなく、ますます面倒になっています。入荷と出庫の追跡にわずかなミスがあると、市場での製品の性能に大きな影響を与える可能性があります。これにより、需要の変動、顧客の損失が発生し、最終的に売上高に悪影響を及ぼします。
3. 事業計画 – 在庫管理は、事業計画と戦略化において重要な役割を果たします。正確なデータのみが製品ロードマップまたは、販促活動の計画に役立ち、成功する結果につながります。これには、すべてのデータのリアルタイムの収集と処理が必要であり、時間通りに正確な意思決定を行うことができます。
4. 運用コスト – 膨大な在庫は、より多くの製品、より多くの市場、より多くの配送、より多くのスペース、およびより多くのスタッフが必要になることを意味します。これにより、他の重要な費用からシェアを奪って在庫を維持するための運用コストが増加する事例がよくあります。

AIは在庫管理の最適化にどのように役立ちますか？

企業は常にテクノロジーを活用して競争力を獲得してきました。反復プロセスの自動化から認知能力の実装まで。 世界中の企業は、ビジネスを行う上で長い道のりを歩んできました。自動化はごく当たり前になっていますが、注目を集めているのは人工知能と機械学習です。これは、プロセスを最適化し、実用的な洞察を提供する能力が原因である可能性があり、ビジネスをさらに加速させます。最近の451Research Reportによると、AIを活用したデータ管理を加速するかとの質問に対して、 回答者の68％が、すでにML機能を使用しているまたは今後3年以内に使用する計画と報告しています。

Gyrus AIでは、フォーカスを絞ったAIの実装と機械学習機能の活用が、他の事業と同様に、在庫管理の最適化に役立つと考えています。 MLとAIを活用した在庫管理で企業が何を達成できるかを見てみましょう。

AIは在庫の可視性を提供します

製品性能と市場に対する明確な可視性は、販売および販売促進戦略の考案に役立ちます。データの均一性と粒度のレベルとともに。 戦略の正確さは、製品データの適時性にも依存します。製品データを可視化することで、在庫過剰や在庫不足を回避することができます。

マッキンゼーのグローバル調査によると、在庫はAIを採用することで、より大きな価値を引き出すことができると報告されています。サプライチェーン企業で働く調査回答者の約76％が、AI導入による重要な価値が提供されるとあると報告しています。AIは、堅牢なデータパイプを通じて、販売、注文、在庫データに対するタイムリーで正確な洞察と可視性を実現します。これにより、市場での多くの不協和音が回避され、高品質のデータを時間どおりに配信できます。AIは、以前は利用できなかった洞察を提供できるだけでなく、 しかし、それはまた、何らかの形で需要に影響を与えるさまざまな側面を考慮に入れることができます。

AIはデータマイニングを強化します

AIシステムは、機械学習機能により、データパターンから学習し、製品需要を予測できます。 たとえば、季節的な需要に応じて、特定の地理的な場所で在庫を計画するのに役立ちます。 状況を分析することは、過剰在庫にするか、棚にある関連製品と交換することを提案するのに役立ちます。 これにより、営業およびマーケティング調査チームへのプレッシャーを軽減できます。

AIは在庫管理を改善します

在庫をうまく管理することで、顧客満足度と充実感を高めることができます。これに僅かなずれがあると、不足や遅延が発生する可能性があり、 特定の製品需要に悪影響を及ぼし、収益に悪影響を及ぼします。さらに、AIは、ビッグデータの助けを借りて顧客の行動パターンを分析する機能により、在庫を正しく確保することに役立ちます。
これにより、在庫管理ミスのリスクが最小限に抑えられ、顧客満足度が向上します。データマイニングから得られた洞察は、工場から小売店への生鮮食品の輸送を時間内に計画することで無駄を減らし、売上向上に役立ちます。

 

なぜ、企業はMLとAIの使用にまだ懐疑的なのでしょうか？

MLとAIの機能を活用することは、在庫管理に有益であることが証明されていますが、 AIツールの実装中に挫折する可能性があります。 ただし、数が多いほど、利点はしばしば挫折を隠します。 それでも、AI実装中に発生する可能性のある複雑さをチームに認識させることが重要です。 潜在的なリスクと課題を認識することで、リスクを大幅に軽減または軽減できます。

AIの実装と運用が複雑になる可能性がある主に2つの領域があります。 第一に、データの提示です。 AIシステムが正常に機能するには、高品質のデータを提示する必要があります。 AIが分析し、実用的な洞察を導き出すのは、このデータです。 ただし、多くの企業は、データを収集し、システムにデータをフィードすることが課題であると感じています。

第二に、既存システムにAIを実装するには時間がかかる可能性があります。 多くの場合、データとプロセスはレガシーシステムに負担を与えるため、AIシステムをレガシーシステム上に統合する上で困難にしています。
ただし、これらは注意点は、ほんの一部であり、AI実装を成功させることが出来るはずです。事前にスタッフの準備を十分に行なってください。

結論

機械学習と人工知能の導入は、間違いなく、在庫管理の変革において大きな進歩を遂げています。 これにより、企業は物理的なタスクと複雑なデータ関連のタスクの両方を処理できるようになります。 事業計画と消費者行動の予測に付加価値を与えながら、在庫を完全に管理できます。

民間調査会社によると、2021年までに、小売部門の在庫管理ソフトウェアのCAGRは、8.3％になると報告しています。別の調査報告では、世界のエンタープライズ向けクラウドサービス市場は2020年までに23.83％のCAGRで成長すると予測しています。AIを在庫管理に適用する傾向はすでに加速しています。さまざまな企業が、AIを活用した在庫管理ツールを活用して利益を上げています。

AIは、中小企業にとって大きな恩恵をもたらします。
特に、大規模な在庫を扱い、在庫データの制と管理ができない人とってより大きなメリットがあります。
AIの導入により膨大なデータと反復プロセスを処理させることで、人間の労働者は目前のより戦略的で重要なタスクに集中することができます。今後数年以内に、企業が革新的なテクノロジーを活用して在庫管理を最適化するのを目にするでしょう。

The post Transform your Inventory management. Embrace AI/ML. appeared first on Gyrus AI | Blog | Insights on AI & Intelligent Media Search, In-scene Ad Placement, Automated Video Anonymization Technologies.

Highlights

Industrial adoption of the Internet of Things (IoT) and digitization has provided great visibility into the several processes in a Factory or a Manufacturing Unit.  The Analytics on top of the monitored data helped managers and executives improve productivity by manually monitoring each of the metrics. 

Moreover, a large volume of data is generated from the various IIoT (Industrial IoT) sensors in the field, factory, etc., and this data can be used to learn patterns using advanced Machine Learning (ML) and Artificial Intelligence (AI) algorithms. 

Gyrus has developed several such ML/AI algorithms for Industrial IoT that directly improve productivity. Gyrus models are adapts using the Customer data from such sensors and the custom model is integrated into Global Operations giving the below mentioned (Figure-1) top-line results. The rest of the paper talks about the different models developed, sensors used and the specific target achieved. 

Company Profile

The customer is a large industrial conglomerate with over $4B+ in revenue, 10K+ employees. The Customer has several factories, warehouses, and logistics operations to produce, store and deliver to their customers worldwide. They operate on very tight margins and its protection is of paramount importance. In the first phase (this paper covers that phase), the IIoT project is implemented in a single department of the company. 

Challenges 

The Customer has existing systems in place at the factory floors, Global Operations center and at the warehouses. These Hardware and Software systems are a no-touch as they are operational and the current staff is well versed with them. Integrating new hardware is a big challenge as it involves installation and maintenance. For the Software systems as well, there are several databases and existing software modules that are customization and being in use case. Any new software has to integrate with those modules. 

The overarching goal of predicting failures, improving efficiency, and making an impact on the TCO from the top management helped to a good extent when certain existing processes had to change. 

Solution

To address the challenges mentioned above, the installation of new sensors set up without disturbing the existing flow. Certainly, Gyrus worked with a Systems Integrator who installs and maintains equipment for the Customer to install 

• Wireless industrial IoT Modules
  • Sensors – Vibration, Humidity, Temperature, Gyro, Accelerometer
• Wireless Controller with Connectivity to cloud

The usability of a single module with all the sensors populated with wireless connectivity helps to reduce the number of SKUs to install. Creating a parallel path for data collection and processing, without changing any of the existing hardware and machinery. It helps in the integration part to be completed fast and with the least friction.

Time series data is collected from the sensors and the ML/AI algorithms are run on that data. The data is collected from existing sources as well when required and available. For Global operation software, make use of existing software to make API calls for all the models. It also generates alerts with severity levels for predictive maintenance. And also provided with custom dashboards based on the raw data and the analysis by the ML/AI engines.

Algorithms Implemented  

Implementation of ML/AI algorithms for improving overall TCO with the IoT hardware as follows.

Predictive Maintenance

With the various sensors including the Vibration sensors, providing continuous time-series data, the algorithm looks for conditions that could foretell a failure. At a base level, In order to predict any failure, anomalies located to detect in the sensor data. Over of period of time, the development of data annotation set in place, as the base data set from gyrus does not cover all the specific cases for the customer.

Remote Asset Management

The goal of asset tracking is to easily locate and monitor key assets and track their usage. It also helps in optimizing the supply chain logistics. Monitoring of active usage of the equipment is for the proper use also and reported usage is corroborated for billing. Machine Learning techniques are proven to be invaluable as an absolute match of values is not possible.

Supply Chain and Inventory Management

Gyrus Inventory Management model is used for supply chain decisions. The data from Asset Management, Weather, Seasonality, Demand Variability, Supplier Variability, Macro-economic Production, Macro-economic Consumption, Inventory levels, Sales Demand, Lead Times, even more, are in use as features to predict what to order and at the order levels.

Optimizing Manufacturing yield

This is a standard optimization model that adapts to the customer use case. Moreover, the model governs equipment usage for optimal output, Raw material input to yield optimal results. As a result, this model prescribes (Prescriptive) the usage times, and the raw material quality and quantity metrics.

Output Quality Management

Development of an output quality prediction model from the various inputs of the learning process of features affecting the output quality. The output quality prediction model predicts if the quality metrics fall below the threshold before the manufacturing and QA process. It also presents the causal features and the expected thresholds suggesting the specific areas to improve. 

Please refer to the complete IoT ML/AI models from Gyrus here.

Results

Retooling parts of a Factory from both hardware and software perspective is a major challenge especially protecting the existing process in place. Setting up a parallel path for the new sensors helped the process. It’s challenging, but doable by working closely with the Customer Team in assessing the integration points and by implementing APIs for ML/AI models. 

Furthermore, The results exceeded the initially set goals and the customer plans to do the next phases of the project.

The models predicted any maintenance required for the machinery with more than 99% accuracy. At times when the time was not sufficient to call in maintenance before the breakdown, just switching off the system helped save repair costs. 

Moreover, the Inventory Management model saved costs by effectively placing orders and quantities. Certainly, this is a direct impact on the efficiency of the operation. Similarly, Yield and Quality improvement models had an uplift with the ML/AI models. 

Above all, the whole process set in place a framework for future extensions and to have the customer embrace Industry 4.0 methodology.

ハイライト

モノのインターネット (IoT)の産業分野への導入とデジタル化により、工場または製造ユニットのいくつかのプロセスに対する優れた可視性が提供されています。 監視されたデータに対する分析を行うことで、マネージャと経営幹部は、各指標を手動で監視することで生産性を向上させることができるようになりました。

さらに、フィールドや工場などのさまざまな産業IoT (Industrial IoT)センサから大量のデータが生成され、このデータを使用して、高度な機械学習(ML)および人工知能(AI)アルゴリズムを使用してパターンを学習できます。

Gyrusは、生産性を直接向上させる産業用IoT向けのML/AIアルゴリズムを開発しました。 Gyrusモデルは、そのようなセンサからの顧客データを使用して適応され、カスタムモデルはグローバルオペレーションに統合され、以下(図1)に示すトップラインの結果を提供します。 残りのペーパーでは、開発したさまざまなモデル、使用されたセンサ、および達成された特定の目標について説明します。

顧客会社概要

顧客は、40億ドル以上の収益、1万人以上の従業員を抱える大規模な産業コングロマリットです。 お客様には、世界中のお客様に生産、保管、配送するためのいくつかの工場、倉庫、およびロジスティクス業務があります。 それらは非常に狭いマージンで動作し、その保護は最も重要です。 最初のフェーズ（このペーパーではそのフェーズについて説明します）では、インダストリアルIoTプロジェクトは、会社の一つの事業部門に実装されます。

課題

お客様は、工場のフロア、グローバルオペレーションセンター、および倉庫に既存のシステムを導入しています。 これらのハードウェアおよびソフトウェアシステムは、運用可能であり、現在のスタッフはそれらに精通しているため、手間がかかりません。 新しいハードウェアの統合は、インストールとメンテナンスを伴うため、大きな課題です。 ソフトウェアシステムについても、カスタマイズされて使用されている、複数のデータベースと既存のソフトウェアモジュールがあります。 新しいソフトウェアは、これらのモジュールと統合する必要があります。

障害を予測し、効率を改善するという、トップマネジメントからTCOに影響を与えるという包括的な目標は、特定の既存のプロセスを変更する必要がある場合に大変役立ちました。

ソリューション

上記の課題に対処するために、既存の流れを妨げることなく新しいセンサーを設置しました。 確かに、Gyrusは、お客様がインストールする機器をインストールおよび保守するシステムインテグレータと協力しました。

ワイヤレス産業用IoTモジュール
センサー–振動、湿度、温度、ジャイロ、加速度計
クラウドに接続できるワイヤレスコントローラ

すべてのセンサにワイヤレス接続が組み込まれた単一のモジュールの使いやすさは、インストールするSKUの数を減らすのに役立ちます。 既存のハードウェアや機械を変更することなく、データの収集と処理のための並列パスを作成します。 これは、統合部分を迅速かつ最小限の摩擦で完了するのに役立ちます。

時系列データはセンサから収集され、ML/AIアルゴリズムがそのデータに対して実行されます。 データは、必要に応じて利用可能な場合、既存のソースからも収集されます。 グローバルオペレーションソフトウェアの場合は、既存のソフトウェアを利用して、すべてのモデルのAPI呼び出しを行います。 また、予知保全のための重大度レベルのアラートを生成します。 また、生データとML/AIエンジンによる分析に基づいたカスタムダッシュボードも提供されます。

実装されたアルゴリズム

IoTハードウェアを使用して全体的なTCOを改善するためのML/AIアルゴリズムを次のように実装しました。

予知保全

振動センサを含むさまざまなセンサを使用して、連続的な時系列データを提供し、アルゴリズムは障害を予測できる条件を探します。 基本レベルでは、障害を予測するために、センサ データで検出するために特定された異常。 脳回からの基本データセットが顧客の特定のケースすべてを網羅しているわけではないため、一定期間にわたって、データ注釈セットの開発が実施されました。

リモート資産管理

資産追跡の目標は、主要な資産を簡単に見つけて監視し、それらの使用状況を追跡することです。 また、サプライチェーンのロジスティクスの最適化にも役立ちます。 機器のアクティブな使用状況の監視も適切な使用のためであり、報告された使用状況は請求のために裏付けられます。 値を完全に一致させることは不可能であるため、機械学習手法は非常に貴重であることが証明されています。

サプライチェーンと在庫管理

Gyrus 在庫管理モデルは、サプライチェーンの決定に使用されます。 資産管理、天候、季節性、需要変動性、サプライヤーの変動性、マクロ経済生産、マクロ経済消費、在庫レベル、販売需要、リードタイムなどのデータは、何を注文するか注文レベルを予測する機能として使用されています。

製造歩留まりの最適化

これは、お客様のユースケースに適応する標準の最適化モデルです。 さらに、このモデルは、最適な出力のための機器の使用、最適な結果を生み出すための原材料の入力を管理します。 その結果、このモデルは(規範的な)使用時間、および原材料の品質と量のメトリックを規定します。

出力品質管理

出力品質に影響を与える特徴量の学習プロセスのさまざまな入力からの出力品質予測モデルの開発。 出力品質予測モデルは、製造およびQAプロセスの前に品質メトリックがしきい値を下回るかどうかを予測します。 また、原因となる特徴と、改善すべき特定の領域を示唆する予想されるしきい値も示します

GyrusのIoT ML/AIモデルを参照してください。

結果

ハードウェアとソフトウェアの両方の観点から、工場の部品を改造することは、特に既存のプロセスを保護するための大きな課題です。 新しいセンサの並列パスを設定すると、プロセスが役立ちました。 やりがいがありますが、統合ポイントを評価する際にお客様チームと緊密に連携し、ML/AIモデルのAPIを実装することで、実行可能です。

さらに、結果は当初設定された目標を上回り、顧客はプロジェクトの次のフェーズを実行することを計画しています。

このモデルは、99％を超える精度で機械に必要なメンテナンスを予測しました。 故障前にメンテナンスを依頼する時間が足りなかったときは、システムの電源を切るだけで修理費を節約できました。

さらに、在庫管理モデルは、効果的に注文と数量を出すことでコストを節約しました。 確かに、これは操作の効率に直接影響します。 同様に、歩留まりと品質の改善モデルも、ML/AIモデルとともに上昇しました。

とりわけ、プロセス全体は、将来の拡張のためのフレームワークを設定し、顧客にインダストリー4.0の方法論を受け入れさせるためのものです。

The post Industrial IoT ML/AI models for TCO improvement appeared first on Gyrus AI | Blog | Insights on AI & Intelligent Media Search, In-scene Ad Placement, Automated Video Anonymization Technologies.