保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。

Revisit consent button We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data.

Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features.

Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc.

Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors.

Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns.

Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle SearchToggle Mobile Navigation Publish AI, ML & data-science insights to a global community of data professionals. Sign inSubmit an Article LatestEditor's PicksNewsletterShipAI

Write For TDS LinkedInXToggle Search Data Science

NumPy for Absolute Beginners: A Project-Based Approach to Data Analysis

Build a high-performance sensor data pipeline from scratch and unlock the true speed of Python's scientific computing core Ibrahim Salami November 4, 202513 min read To learn NumPy, I’ve been running a series where I build mini projects. I’ve built a Personal Habit and Weather Analysis project. But I haven’t really gotten the chance to explore the full power and capability of NumPy. I want to try to understand why NumPy is so useful in data analysis. To wrap up this series, I’m going to be showcasing this in real time. I’ll be using a fictional client or company to make things interactive. In this case, our client is going to beEnviroTech Dynamics , a global operator of industrial sensor networks. Currently, EnviroTech relies on outdated, loop-based Python scripts to process over 1 million sensor readings daily. This process is agonizingly slow, delaying critical maintenance decisions and impacting operational efficiency. They need a modern, high-performance solution. I’ve been tasked with creating a NumPy-based proof-of-concept to demonstrate how to turbocharge their data pipeline. Learn this step by step with the interactive AI and Data Scientist roadmap.

The Dataset: Simulated Sensor Readings

To prove the concept, I’ll be working with a large, simulated dataset generated using NumPy‘s random module, featuring entries with the following key arrays:

Project Objectives

The core goal is to provide four clear, vectorised solutions to EnviroTech’s data challenges, demonstrating speed and power. So I’ll be showcasing all of these:

By the end of this article, you should be able to get a full grasp of NumPy and its usefulness in data analysis.

Objective 1: Performance and Efficiency Benchmark

First, we need a massive dataset to make the speed difference obvious. I’ll be using the 1,000,000 temperature readings we planned earlier. python

import numpy as np
# Set the size of our data
NUM_READINGS = 1_000_000

# Generate the Temperature array (1 million random floating-point numbers)
# We use a seed so the results are the same every time you run the code
np.random.seed(42)
mean_temp = 45.0
std_dev_temp = 12.0
temperature_data = np.random.normal(loc=mean_temp, scale=std_dev_temp, size=NUM_READINGS)

print(f”Data array size: {temperature_data.size} elements”)
print(f”First 5 temperatures: {temperature_data[:5]}”)

Output: text

Data array size: 1000000 elements
First 5 temperatures: [50.96056984 43.34082839 52.77226246 63.27635828 42.1901595 ]

Now that we have our records. Let’s check out the effectiveness of NumPy. Assuming we wanted to calculate the average of all these elements using a standard Python loop, it’ll go something like this. python

# Function using a standard Python loop
def calculate_mean_loop(data):
total = 0
count = 0
for value in data:
total += value
count += 1
return total / count

# Let’s run it once to make sure it works
loop_mean = calculate_mean_loop(temperature_data)
print(f”Mean (Loop method): {loop_mean:.4f}”)

There’s nothing wrong with this method. But it’s quite slow, because the computer has to process each number one by one, constantly moving between the Python interpreter and the CPU. To truly showcase the speed, I’ll be using the%timeit command. This runs the code hundreds of times to provide a reliable average execution time. python

# Time the standard Python loop (will be slow)
print(“ — — Timing the Python Loop — -”)
%timeit -n 10 -r 5 calculate_mean_loop(temperature_data)

Output markup

--- Timing the Python Loop ---244 ms ± 51.5 ms per loop (mean ± std. dev. of 5 runs, 10 loops each)

Using the -n 10, I’m basically running the code in the loop 10 times (to get a stable average), and using the -r 5, the whole process will be repeated 5 times (for even more stability). Now, let’s compare this with NumPy vectorisation. And by vectorisation, it means the entire operation (average in this case) will be performed on the entire array at once, using highly optimised C code in the background.

Here’s how the average will be calculated using NumPy python

# Using the built-in NumPy mean function
def calculate_mean_numpy(data):
return np.mean(data)
# Let’s run it once to make sure it works
numpy_mean = calculate_mean_numpy(temperature_data)
print(f”Mean (NumPy method): {numpy_mean:.4f}”)

Output: markup

Mean (NumPy method): 44.9808

Now let’s time it. python

# Time the NumPy vectorized function (will be fast)
print(“ — — Timing the NumPy Vectorization — -”)
%timeit -n 10 -r 5 calculate_mean_numpy(temperature_data)

Output: markup

--- Timing the NumPy Vectorization ---1.49 ms ± 114 μs per loop (mean ± std. dev. of 5 runs, 10 loops each)

Now, that’s a huge difference. That’s like almost non-existent. That’s the power of vectorisation. Let's present this speed difference to the client:

“We compared two methods for performing the same calculation on one million temperature readings — a traditional Python for-loop and a NumPy vectorized operation. The difference was dramatic: The pure Python loop took about 244 milliseconds per run while the NumPy version completed the same task in just 1.49 milliseconds. That’s roughly a 160× speed improvement.”

Objective 2: Foundational Statistical Baseline

Another cool feature NumPy offers is the ability to perform basic to advanced statistics — this way, you can get a good overview of what’s going on in your dataset. It offers operations like:

Now that we’ve managed to provide an alternative and efficient solution to retrieve and perform summaries and calculations on their huge dataset, we can start playing around with it. We already managed to generate our simulated temperature data. Let’s do the same for pressure. Calculating pressure is a great way to demonstrate the ability of NumPy to handle multiple massive arrays in no time at all. For our client, it also allows me to showcase a health check on their industrial systems. Also, temperature and pressure are often related. A sudden pressure drop might be the cause of a spike in temperature, or vice versa. Calculating baselines for both allows us to see if they are drifting together or independently

python

# Generate the Pressure array (Uniform distribution between 100.0 and 500.0)
np.random.seed(43) # Use a different seed for a new dataset
pressure_data = np.random.uniform(low=100.0, high=500.0, size=1_000_000)
print(“Data arrays ready.”)

Output: markup

Data arrays ready.

Alright, let’s begin our calculations. python

print(“\n — — Temperature Statistics — -”)
# 1. Mean and Median
temp_mean = np.mean(temperature_data)
temp_median = np.median(temperature_data)

# 2. Standard Deviation
temp_std = np.std(temperature_data)

# 3. Percentiles (Defining the 90% Normal Range)
temp_p5 = np.percentile(temperature_data, 5) # 5th percentile
temp_p95 = np.percentile(temperature_data, 95) # 95th percentile

# Formating our results
print(f”Mean (Average): {temp_mean:.2f}°C”)
print(f”Median (Middle): {temp_median:.2f}°C”)
print(f”Std. Deviation (Spread): {temp_std:.2f}°C”)
print(f”90% Normal Range: {temp_p5:.2f}°C to {temp_p95:.2f}°C”)

Here’s the output: markup

--- Temperature Statistics ---Mean (Average): 44.98°CMedian (Middle): 44.99°CStd. Deviation (Spread): 12.00°C90% Normal Range: 25.24°C to 64.71°C

So to explain what you’re seeing here The Mean (Average): 44.98°C basically gives us a central point around which most readings are expected to fall. This is pretty cool because we don’t have to scan through the entire large dataset. With this number, I’ve gotten a pretty good idea of where our temperature readings usually fall. The Median (Middle): 44.99°C is quite identical to the mean if you notice. This tells us that there aren’t extreme outliers dragging the average too high or too low. The standard deviation of 12°C means the temperatures vary quite a bit from the average. Basically, some days are much hotter or cooler than others. A lower value (say 3°C or 4°C) would have suggested more consistency, but 12°C indicates a highly variable pattern. For the percentile , it basically means most days hover between 25°C and 65°C,
If I were to present this to the client, I could put it like this:

“On average, the system (or environment) maintains a temperature around 45°C , which serves as a reliable baseline for typical operating or environmental conditions. A deviation of 12°C indicates that temperature levels fluctuate significantly around the average. To put it simply, the readings are not very stable. Lastly, 90% of all readings fall between 25°C and 65°C. This gives a realistic picture of what “normal” looks like, helping you define acceptable thresholds for alerts or maintenance. To improve performance or reliability, we could identify the causes of high fluctuations (e.g., external heat sources, ventilation patterns, system load).” Let’s calculate for pressure also. python

print(“\n — — Pressure Statistics — -”)
# Calculate all 5 measures for Pressure
pressure_stats = {
“Mean”: np.mean(pressure_data),
“Median”: np.median(pressure_data),
“Std. Dev”: np.std(pressure_data),
“5th %tile”: np.percentile(pressure_data, 5),
“95th %tile”: np.percentile(pressure_data, 95),
}
for label, value in pressure_stats.items():
print(f”{label:<12}: {value:.2f} kPa”)

To improve our codebase, I’m storing all the calculations performed in a dictionary called pressure stats, and I’m simply looping over the key-value pairs. Here’s the output: markup

--- Pressure Statistics ---Mean : 300.09 kPaMedian : 300.04 kPaStd. Dev : 115.47 kPa5th %tile : 120.11 kPa95th %tile : 480.09 kPa

If I were to present this to the client. It’d go something like this:

“Our pressure readings average around 300 kilopascals , and the median — the middle value — is almost the same. That tells us the pressure distribution is quite balanced overall. However, the standard deviation is about 115 kPa , which means there’s a lot of variation between readings. In other words, some readings are much higher or lower than the typical 300 kPa level.
Looking at the percentiles , 90% of our readings fall between 120 and 480 kPa. That’s a wide range, suggesting that pressure conditions are not stable — possibly fluctuating between low and high states during operation. So while the average looks fine, the variability could point to inconsistent performance or environmental factors affecting the system.”

Objective 3: Critical Anomaly Identification

One of my favourite features of NumPy is the ability to quickly identify and filter out anomalies in your dataset. To demonstrate this, our fictional client, EnviroTech Dynamics, provided us with another helpful array that contains system status codes. This tells us how the machine is consistently operating. It’s simply a range of codes (0–3).

They receive millions of readings per day, and our job is to find every machine that’s both in a critical state and running dangerously hot.
Doing this manually, or even with a loop, would take ages. This is where Boolean Indexing (masking) comes in. It lets us filter huge datasets in milliseconds by applying logical conditions directly to arrays, without loops. Earlier, we generated our temperature and pressure data. Let’s do the same for the status codes. python

# Reusing 'temperature_data' from earlier
import numpy as np

np.random.seed(42) # For reproducibility

status_codes = np.random.choice(
a=[0, 1, 2, 3],
size=len(temperature_data),
p=[0.85, 0.10, 0.03, 0.02] # 0=Normal, 1=Warning, 2=Critical, 3=Offline
)

# Let’s preview our data
print(status_codes[:5])

Output: markup

[0 2 0 0 0]

Each temperature reading now has a matching status code. This allows us to pinpoint which sensors report problems and how severe they are. Next, we’ll need some sort of threshold or anomaly criteria. In most scenarios, anything above mean + 3 × standard deviation is considered a severe outlier, the kind of reading you don’t want in your system. To compute that python

temp_mean = np.mean(temperature_data)
temp_std = np.std(temperature_data)
SEVERITY_THRESHOLD = temp_mean + (3 * temp_std)
print(f”Severe Outlier Threshold: {SEVERITY_THRESHOLD:.2f}°C”)

Output: markup

Severe Outlier Threshold: 80.99°C

Next, we’ll create two filters (masks) to isolate data that meets our conditions. One for readings where the system status is Critical (code 2) and another for readings where the temperature exceeds the threshold. python

# Mask 1 — Readings where system status = Critical (code 2)
critical_status_mask = (status_codes == 2)

# Mask 2 — Readings where temperature exceeds threshold
high_temp_outlier_mask = (temperature_data > SEVERITY_THRESHOLD)

print(f”Critical status readings: {critical_status_mask.sum()}”)
print(f”High-temp outliers: {high_temp_outlier_mask.sum()}”)

Here’s what's going on behind the scenes. NumPy creates two arrays filled with True or False. Every True marks a reading that satisfies the condition. True will be represented as 1, and False will be represented as 0. Summing them quickly counts how many match. Here’s the output: markup

Critical status readings: 30178High-temp outliers: 1333

Let’s combine both anomalies before printing our final result. We want readings that are both critical and too hot. NumPy allows us to filter on multiple conditions using logical operators. In this case, we’ll be using the AND function represented as &. python

# Combine both conditions with a logical AND
critical_anomaly_mask = critical_status_mask & high_temp_outlier_mask

# Extract actual temperatures of those anomalies
extracted_anomalies = temperature_data[critical_anomaly_mask]
anomaly_count = critical_anomaly_mask.sum()

print(“\n — — Final Results — -”)
print(f”Total Critical Anomalies: {anomaly_count}”)
print(f”Sample Temperatures: {extracted_anomalies[:5]}”)

Output: markup

--- Final Results ---Total Critical Anomalies: 34Sample Temperatures: [81.9465697 81.11047892 82.23841531 86.65859372 81.146086 ]

Let’s present this to the client

“After analyzing one million temperature readings, our system detected 34 critical anomalies — readings that were both flagged as ‘critical status’ by the machine and exceeded the high-temperature threshold. The first few of these readings fall between 81°C and 86°C , which is well above our normal operating range of around 45°C. This suggests that a small number of sensors are reporting dangerous spikes , possibly indicating overheating or sensor malfunction.
In other words, while 99.99% of our data looks stable, these 34 points represent the exact spots where we should focus maintenance or investigate further.”

Let’s visualise this real quick with matplotlib When I first plotted the results, I expected to see a cluster of red bars showing my critical anomalies. But there were none. At first, I thought something was wrong, but then it clicked. Out of one million readings, only 34 were critical. That’s the beauty of Boolean masking: it detects what your eyes can’t. Even when the anomalies hide deep within millions of normal values, NumPy flags them in milliseconds.

Objective 4: Data Cleaning and Imputation

Lastly, NumPy allows you to get rid of inconsistencies and data that doesn’t make sense. You might have come across the concept of data cleaning in data analysis. In Python, NumPy and Pandas are often used to streamline this activity. To demonstrate this, our status_codes contain entries with a value of 3 (Faulty/Missing). If we use these faulty temperature readings in our overall analysis, they will skew our results. The solution is to replace the faulty readings with a statistically sound estimated value. The first step is to figure out what value we should use to replace the bad data. The median is always a great choice because, unlike the mean, it is less affected by extreme values. python

# TASK: Identify the mask for ‘Valid’ data (where status_codes is NOT 3 — Faulty/Missing).
valid_data_mask = (status_codes != 3)

# TASK: Calculate the median temperature ONLY for the Valid data points. This is our imputation value.
valid_median_temp = np.median(temperature_data[valid_data_mask])
print(f”Median of all valid readings: {valid_median_temp:.2f}°C”)

Output: markup

Median of all valid readings: 44.99°C

Now, we’ll perform some conditional replacement using the powerful np.where() function. Here’s a typical structure of the function. np.where(Condition, Value_if_True, Value_if_False) In our case:

python

# TASK: Implement the conditional replacement using np.where().
cleaned_temperature_data = np.where(
status_codes == 3, # CONDITION: Is the reading faulty?
valid_median_temp, # VALUE_IF_TRUE: Replace with the calculated median.
temperature_data # VALUE_IF_FALSE: Keep the original temperature value.
)

# TASK: Print the total number of replaced values.
imputed_count = (status_codes == 3).sum()
print(f”Total Faulty readings imputed: {imputed_count}”)

Output: markup

Total Faulty readings imputed: 20102

I didn’t expect the missing values to be this much. It probably affected our reading above in some way. Good thing, we managed to replace them in seconds. Now, let’s verify the fix by checking the median for both the original and cleaned data python

# TASK: Print the change in the overall mean or median to show the impact of the cleaning.
print(f”\nOriginal Median: {np.median(temperature_data):.2f}°C”)
print(f”Cleaned Median: {np.median(cleaned_temperature_data):.2f}°C”)

Output: markup

Original Median: 44.99°CCleaned Median: 44.99°C

In this case, even after cleaning over 20,000 faulty records, the median temperature remained steady at 44.99°C, indicating that the dataset is statistically sound and balanced. Let’s present this to the client:

“Out of one million temperature readings, 20,102 were marked as faulty (status code = 3). Instead of removing these faulty records, we replaced them with the median temperature value (≈ 45°C) — a standard data-cleaning approach that keeps the dataset consistent without distorting the trend.
Interestingly, the median temperature remained unchanged (44.99°C) before and after cleaning. That’s a good sign: it means the faulty readings didn’t skew the dataset, and the replacement didn’t alter the overall data distribution.”

Conclusion

And there we go! We initiated this project to address a critical issue for EnviroTech Dynamics : the need for faster, loop-free data analysis. The power of NumPy arrays and vectorisation allowed us to fix the problem and future-proof their analytical pipeline. NumPy ndarray is the silent engine of the entire Python data science ecosystem. Every major library, like Pandas , scikit-learn , TensorFlow , and PyTorch, uses NumPy arrays at its core for fast numerical computation. By mastering NumPy, you’ve built a powerful analytical foundation. The next logical step for me is to move from single arrays to structured analysis with the Pandas library, which organises NumPy arrays into tables (DataFrames) for even easier labelling and manipulation. Thanks for reading! Feel free to connect with me: Medium LinkedIn Twitter YouTube

Register Now to Continue Reading

Unlock Towards Data Science with a Free Account

All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account

Written By

Ibrahim Salami

See all from Ibrahim Salami Deep DivesBig DataData AnalyticsNumpyNumpy ArrayPython Share This Article

Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. Write for TDS

Related Articles

My Medium Journey as a Data Scientist: 6 Months, 18 Articles, and 3,000 Followers

Data Science Real numbers, earnings, and data-driven growth strategy for Medium writers Yu Dong November 11, 20249 min read

Transfer Learning using Pre-Trained AlexNet Model and Fashion-MNIST

Data Science Building CNN classifiers in Apache MxNet Gluon framework to classify articles of clothing Andrea Yoss December 16, 202010 min read

Vectorization in Python

Artificial Intelligence PyTrix#1: Speeding up our Python Code Kurtis Pykes May 7, 20204 min read

Add a Customized Rasa Chatbot for Your Facebook Page

NLP A simple walkthrough for setting up a Rasa chatbot for your Facebook page Ikshura Tachintha July 27, 20217 min read

3 Most Asked Python Interview Questions for Data Scientists

Data Science Top python interview questions for data scientists Alexandros Zenonos, PhD August 16, 20204 min read

Extreme Value Theory in a Nutshell with Various Applications

Coding In the beginning when statistics was discovered to answer questions related to gambling win chances in the 18th century, normal... Hanan Ahmed February 19, 20218 min read

Data Science: Write Robust Python With Static Typing

Artificial Intelligence Updated for Python 3.11 Nikos Kafritsas July 27, 20227 min read

From Static to Dynamic Skills: A Different Model for Agent Knowledge

Agentic AI Why the skill-inflation panic is aimed at the wrong thing, and what it costs to make agent knowledge a build artifact instead of a file. Tomer Mesika September 14, 202614 min read

How to Negotiate Your Salary as a Data Scientist

Data Science And how much I made my first year Haden Pelletier October 27, 20246 min read

Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our NewsletterWrite For TDS•About•Advertise•Privacy Policy•Terms of Use © Insight Media Group, LLC 2026•The New Stack•roadmap.sh•Creator Network

Recommended Articles

Close