Overview of SR-CNN algorithm in Azure Anomaly Detector
Author: Tony Xing (@XingGuodong), AI Platform, C + AI In the last blog “Introducing Azure Anomaly Detector API”, I didn't provide enough details on one of the algorithms. As the algorithm paper was in the publishing process. The paper was accepted by KDD 2019 for oral presentation later, and this blog serves as an overview of the SR-CNN algorithm and for more details user can always read the paper. By the way, we have a 2-minute video here. Problem definition Before we go into details, let us revisit the problem definition of time series anomaly detection. Challenges For any time-series anomaly detection system that is operating in production with a large scale, there are quite a few challenges, especially on the three areas below: 1. Lack of labels - As you can imagine, with signals generated from clients, services, and sensors every second, the huge amount of volume makes it infeasible to manually label the data. 2. Generalization - With real-world data, there are so many different types of time series with different characteristics, which make it hard to generalize and find a silver bullet to solve all the problems. Some examples can be found in the figure below. 3. Efficiency - For any online anomaly detection system, efficiency is one of the key challenges. The system is expected to have low compute cost and low latency for serving. Inspiration In the computer vision domain, there is this concept called “visual saliency detection”. Saliency is what "stands out" in a photo or scene, enabling our eye-brain to quickly focus on the most important regions, as shown in figures below. Fig. Original image Fig. The salient part of the original image When we look at the time series chart, the most dominant and stand-out part is the anomalies. This similarity is where we got the inspiration and it turned out to generate great results. Algorithm Our solution then borrowed Spectral Residual (SR) from the visual saliency detection domain, then apply CNN on the results produced by the SR model As you can see from the algorithm architecture, after SR transformation, the transformed result magnifies the anomalies and the resulting signal is easier to generalize, therefore it provides us a way to training CNN with synthetic data. Spectral Residual The spectral residual algorithm consists of three major steps: Fourier Transform to get the log amplitude spectrum Calculation of spectral residual Inverse Fourier Transform that transforms the sequence back to the spatial domain Benefits SR is unsupervised, efficient, and has good generality. The problem becomes much easier based on the output of the SR model. We can train CNN on the SR output using fully synthetic data with simple synthetic rule Randomly select several points in the saliency map and calculate the injection value to replace the original point. Result We have performed online and offline experimentation, it outperformed state-of-the-arts consistently on open datasets and internal production datasets.
Smart Auditing: Leveraging Azure AI Agents to Transform Financial Oversight
In today's data-driven business environment, audit teams often spend weeks poring over logs and databases to verify spending and billing information. This time-consuming process is ripe for automation. But is there a way to implement AI solutions without getting lost in complex technical frameworks? While tools like LangChain, Semantic Kernel, and AutoGen offer powerful AI agent capabilities, sometimes you need a straightforward solution that just works. So, what's the answer for teams seeking simplicity without sacrificing effectiveness? This tutorial will show you how to use Azure AI Agent Service to build an AI agent that can directly access your Postgres database to streamline audit workflows. No complex chains or graphs required, just a practical solution to get your audit process automated quickly. The Auditing Challenge: It's the month end, and your audit team is drowning in spreadsheets. As auditors reviewing financial data across multiple SaaS tenants, you're tasked with verifying billing accuracy by tracking usage metrics like API calls, storage consumption, and user sessions in Postgres databases. Each tenant generates thousands of transactions daily, and traditionally, this verification process consumes weeks of your team's valuable time. Typically, teams spend weeks: Manually extracting data from multiple database tables. Cross-referencing usage with invoices. Investigating anomalies through tedious log analysis. Compiling findings into comprehensive reports. With an AI-powered audit agent, you can automate these tasks and transform the process. Your AI assistant can: Pull relevant usage data directly from your database Identify billing anomalies like unexpected usage spikes Generate natural language explanations of findings Create audit reports that highlight key concerns For example, when reviewing a tenant's invoice, your audit agent can query the database for relevant usage patterns, summarize anomalies, and offer explanations: "Tenant_456 experienced a 145% increase in API usage on April 30th, which explains the billing increase. This spike falls outside normal usage patterns and warrants further investigation." Let’s build an AI agent that connects to your Postgres database and transforms your audit process from manual effort to automated intelligence. Prerequisites: Before we start building our audit agent, you'll need: An Azure subscription (Create one for free). The Azure AI Developer RBAC role assigned to your account. Python 3.11.x installed on your development machine. OR You can also use GitHub Codespaces, which will automatically install all dependencies for you. You’ll need to create a GitHub account first if you don’t already have one. Setting Up Your Database: For this tutorial, we'll use Neon Serverless Postgres as our database. It's a fully managed, cloud-native Postgres solution that's free to start, scales automatically, and works excellently for AI agents that need to query data on demand. Creating a Neon Database on Azure: Open the Neon Resource page on the Azure portal Fill out the form with the required fields and deploy your database After creation, navigate to the Neon Serverless Postgres Organization service Click on the Portal URL to access the Neon Console Click "New Project" Choose an Azure region Name your project (e.g., "Audit Agent Database") Click "Create Project" Once your project is successfully created, copy the Neon connection string from the Connection Details widget on the Neon Dashboard. It will look like this: postgresql://[user]:[password]@[neon_hostname]/[dbname]?sslmode=require Note: Keep this connection string saved; we'll need it shortly. Creating an AI Foundry Project on Azure: Next, we'll set up the AI infrastructure to power our audit agent: Create a new hub and project in the Azure AI Foundry portal by following the guide. Deploy a model like GPT-4o to use with your agent. Make note of your Project connection string and Model Deployment name. You can find your connection string in the overview section of your project in the Azure AI Foundry portal, under Project details > Project connection string. Once you have all three values on hand: Neon connection string, Project connection string, and Model Deployment Name, you are ready to set up the Python project to create an Agent. All the code and sample data are available in this GitHub repository. You can clone or download the project. Project Environment Setup: Create a .env file with your credentials: PROJECT_CONNECTION_STRING="<Your AI Foundry connection string> "AZURE_OPENAI_DEPLOYMENT_NAME="gpt4o" NEON_DB_CONNECTION_STRING="<Your Neon connection string>" Create and activate a virtual environment: python -m venv .venv source .venv/bin/activate # on macOS/Linux .venv\Scripts\activate # on Windows Install required Python libraries: pip install -r requirements.txt Example requirements.txt: Pandas python-dotenv sqlalchemy psycopg2-binary azure-ai-projects ==1.0.0b7 azure-identity Load Sample Billing Usage Data: We will use a mock dataset for tenant usage, including computed percent change in API calls and storage usage in GB: tenant_id date api_calls storage_gb tenant_456 2025-04-01 1000 25.0 tenant_456 2025-03-31 950 24.8 tenant_456 2025-03-30 2200 26.0 Run python load_usage_data.py Python script to create and populate the usage_data table in your Neon Serverless Postgres instance: # load_usage_data.py file import os from dotenv import load_dotenv from sqlalchemy import ( create_engine, MetaData, Table, Column, String, Date, Integer, Numeric, ) # Load environment variables from .env load_dotenv() # Load connection string from environment variable NEON_DB_URL = os.getenv("NEON_DB_CONNECTION_STRING") engine = create_engine(NEON_DB_URL) # Define metadata and table schema metadata = MetaData() usage_data = Table( "usage_data", metadata, Column("tenant_id", String, primary_key=True), Column("date", Date, primary_key=True), Column("api_calls", Integer), Column("storage_gb", Numeric), ) # Create table with engine.begin() as conn: metadata.create_all(conn) # Insert mock data conn.execute( usage_data.insert(), [ { "tenant_id": "tenant_456", "date": "2025-03-27", "api_calls": 870, "storage_gb": 23.9, }, { "tenant_id": "tenant_456", "date": "2025-03-28", "api_calls": 880, "storage_gb": 24.0, }, { "tenant_id": "tenant_456", "date": "2025-03-29", "api_calls": 900, "storage_gb": 24.5, }, { "tenant_id": "tenant_456", "date": "2025-03-30", "api_calls": 2200, "storage_gb": 26.0, }, { "tenant_id": "tenant_456", "date": "2025-03-31", "api_calls": 950, "storage_gb": 24.8, }, { "tenant_id": "tenant_456", "date": "2025-04-01", "api_calls": 1000, "storage_gb": 25.0, }, ], ) print("✅ usage_data table created and mock data inserted.") Create a Postgres Tool for the Agent: Next, we configure an AI agent tool to retrieve data from Postgres. The Python script billing_agent_tools.py contains: The function billing_anomaly_summary() that: Pulls usage data from Neon. Computes % change in api_calls. Flags anomalies with a threshold of > 1.5x change. Exports user_functions list for the Azure AI Agent to use. You do not need to run it separately. # billing_agent_tools.py file import os import json import pandas as pd from sqlalchemy import create_engine from dotenv import load_dotenv # Load environment variables load_dotenv() # Set up the database engine NEON_DB_URL = os.getenv("NEON_DB_CONNECTION_STRING") db_engine = create_engine(NEON_DB_URL) # Define the billing anomaly detection function def billing_anomaly_summary( tenant_id: str, start_date: str = "2025-03-27", end_date: str = "2025-04-01", limit: int = 10, ) -> str: """ Fetches recent usage data for a SaaS tenant and detects potential billing anomalies. :param tenant_id: The tenant ID to analyze. :type tenant_id: str :param start_date: Start date for the usage window. :type start_date: str :param end_date: End date for the usage window. :type end_date: str :param limit: Maximum number of records to return. :type limit: int :return: A JSON string with usage records and anomaly flags. :rtype: str """ query = """ SELECT date, api_calls, storage_gb FROM usage_data WHERE tenant_id = %s AND date BETWEEN %s AND %s ORDER BY date DESC LIMIT %s; """ df = pd.read_sql(query, db_engine, params=(tenant_id, start_date, end_date, limit)) if df.empty: return json.dumps( {"message": "No usage data found for this tenant in the specified range."} ) df.sort_values("date", inplace=True) df["pct_change_api"] = df["api_calls"].pct_change() df["anomaly"] = df["pct_change_api"].abs() > 1.5 return df.to_json(orient="records") # Register this in a list to be used by FunctionTool user_functions = [billing_anomaly_summary] Create and Configure the AI Agent: Now we'll set up the AI agent and integrate it with our Neon Postgres tool using the Azure AI Agent Service SDK. The Python script does the following: Creates the agent Instantiates an AI agent using the selected model (gpt-4o, for example), adds tool access, and sets instructions that tell the agent how to behave (e.g., “You are a helpful SaaS assistant…”). Creates a conversation thread A thread is started to hold a conversation between the user and the agent. Posts a user message Sends a question like “Why did my billing spike for tenant_456 this week?” to the agent. Processes the request The agent reads the message, determines that it should use the custom tool to retrieve usage data, and processes the query. Displays the response Prints the response from the agent with a natural language explanation based on the tool’s output. # billing_anomaly_agent.py import os from datetime import datetime from azure.ai.projects import AIProjectClient from azure.identity import DefaultAzureCredential from azure.ai.projects.models import FunctionTool, ToolSet from dotenv import load_dotenv from pprint import pprint from billing_agent_tools import user_functions # Custom tool function module # Load environment variables from .env file load_dotenv() # Create an Azure AI Project Client project_client = AIProjectClient.from_connection_string( credential=DefaultAzureCredential(), conn_str=os.environ["PROJECT_CONNECTION_STRING"], ) # Initialize toolset with our user-defined functions functions = FunctionTool(user_functions) toolset = ToolSet() toolset.add(functions) # Create the agent agent = project_client.agents.create_agent( model=os.environ["AZURE_OPENAI_DEPLOYMENT_NAME"], name=f"billing-anomaly-agent-{datetime.now().strftime('%Y%m%d%H%M')}", description="Billing Anomaly Detection Agent", instructions=f""" You are a helpful SaaS financial assistant that retrieves and explains billing anomalies using usage data. The current date is {datetime.now().strftime("%Y-%m-%d")}. """, toolset=toolset, ) print(f"Created agent, ID: {agent.id}") # Create a communication thread thread = project_client.agents.create_thread() print(f"Created thread, ID: {thread.id}") # Post a message to the agent thread message = project_client.agents.create_message( thread_id=thread.id, role="user", content="Why did my billing spike for tenant_456 this week?", ) print(f"Created message, ID: {message.id}") # Run the agent and process the query run = project_client.agents.create_and_process_run( thread_id=thread.id, agent_id=agent.id ) print(f"Run finished with status: {run.status}") if run.status == "failed": print(f"Run failed: {run.last_error}") # Fetch and display the messages messages = project_client.agents.list_messages(thread_id=thread.id) print("Messages:") pprint(messages["data"][0]["content"][0]["text"]["value"]) # Optional cleanup: # project_client.agents.delete_agent(agent.id) # print("Deleted agent") Run the agent: To run the agent, run the following command python billing_anomaly_agent.py Snippet of output from agent: Using the Azure AI Foundry Agent Playground: After running your agent using the Azure AI Agent SDK, it is saved within your Azure AI Foundry project. You can now experiment with it using the Agent Playground. To try it out: Go to the Agents section in your Azure AI Foundry workspace. Find your billing anomaly agent in the list and click to open it. Use the playground interface to test different financial or billing-related questions, such as: “Did tenant_456 exceed their API usage quota this month?” “Explain recent storage usage changes for tenant_456.” This is a great way to validate your agent's behavior without writing more code. Summary: You’ve now created a working AI agent that talks to your Postgres database, all using: A simple Python function Azure AI Agent Service A Neon Serverless Postgres backend This approach is beginner-friendly, lightweight, and practical for real-world use. Want to go further? You can: Add more tools to the agent Integrate with vector search (e.g., detect anomaly reasons from logs using embeddings) Resources: Introduction to Azure AI Agent Service Develop an AI agent with Azure AI Agent Service Getting Started with Azure AI Agent Service Neon on Azure Build AI Agents with Azure AI Agent Service and Neon Multi-Agent AI Solution with Neon, Langchain, AutoGen and Azure OpenAI Azure AI Foundry GitHub Discussions That's it, folks! But the best part? You can become part of a thriving community of learners and builders by joining the Microsoft Learn Student Ambassadors Community. Connect with like-minded individuals, explore hands-on projects, and stay updated with the latest in cloud and AI. 💬 Join the community on Discord here and explore more benefits on the Microsoft Learn Student Hub.Building a digital guide dog for railway passengers with impaired vision
Catching your train on time can be challenging under the best of circumstances. Trains typically only stop for a few minutes, leaving little room for mistakes. For example, at Munich Main station around 240 express trains and 510 regional trains leave from 28 platforms per day. Some trains can also be quite long, up to 346 meters (1,135 ft) for express ICE trains. It is extremely important to quickly find the correct platform and platform section, and then the door closest to a reserved seat needs to be located. This already challenging adventure becomes even more so, if a vision impairment forces a customer to rely exclusively on auditory or tactile feedback. When traveling autonomously, without assistance, it is common practice to walk along the outside of a train, continuously tapping it with a white cane, to discover opened and closed doors (figure 1). While this works in principle, this practice has limitations, both in terms of speed and reliability. We therefore partnered with DB Systel GmbH, the digital partner for all Deutsche Bahn Group companies, to build the Digital Guide Dog. This is a feasibility study based on an AI-powered smartphone application that uses computer vision, auditory and haptic feedback to guide customers to the correct platform section and train car door. In this blog post, we are sharing some of the details and unique challenges that we experienced while the AI model behind this application.
Building Image Classifiers made easy with Azure Custom Vision
In our previous blog, we outlined that Supervised machine learning (ML) models need labeled data, but majority of the data collected in the raw format lacks labels. So, the first step before building a ML model would be to get the raw data labeled by domain experts. To do so, we outlined how Doccano is an easy tool for collaborative text annotation. However, not all data that gets collected is in text format, many a times we end up with a bunch of images but the end goal is again to build a Supervised ML model using them. Like stated previously, the first step would be to tag these images with specific labels. Image tagging as well as building and even deploying either a multi-class or a multi-label classifier can be done in a few simple steps using Azure Custom Vision.
