Leveraging GPT to Optimize Business Processes: A Case Study in Electrical Engineering | by Carlos Dias | Jul, 2024

In at present’s fast-paced enterprise atmosphere, effectivity is vital. As firms try to streamline their operations, synthetic intelligence, significantly giant language fashions like GPT (Generative Pre-trained Transformer), are rising as highly effective instruments for course of optimization. This text presents a real-world case research of how GPT was used to unravel a big enterprise downside in {an electrical} energy firm, leading to substantial time and value financial savings.

Within the electrical energy trade, upkeep of high-voltage gear is essential for guaranteeing dependable service. Our firm’s upkeep course of concerned a number of steps:

1. Discipline inspectors would study gear resembling energy relays and transformers, documenting any defects with photographs and detailed descriptions.
2. Electrical engineers would analyze these experiences to find out the basis trigger of every concern.
3. Based mostly on their evaluation, engineers would create upkeep plans for restore groups to execute.

Whereas this course of was thorough, it had a significant bottleneck: our engineers had been spending a mean of two hours of additional time day by day simply to maintain up with the inflow of defect experiences. This not solely led to elevated prices but additionally potential burnout amongst our useful engineering workers.

To construct a compelling case for change, we would have liked to quantify the influence of this inefficiency:

• Day by day additional time: 2 hours
• Weekly additional time: 10 hours (5 working days)
• Month-to-month additional time: 40 hours
• Yearly additional time: 480 hours (equal to twenty full working days)

Contemplating the common hourly fee for {an electrical} engineer in the USA ($40/hour), this additional time translated to potential annual financial savings of $19,200 per engineer.

To handle this problem, we turned to GPT, a state-of-the-art language mannequin identified for its means to know and generate human-like textual content. Right here’s an summary of our implementation course of:

1. Information Preparation: We gathered historic defect experiences that had been manually labeled by our engineers. This information was cleaned and reworked to make sure high quality enter for our mannequin.
2. Dataset Creation: The ready information was cut up into two units:
   – A coaching dataset for fine-tuning the GPT mannequin (Information from 2022).
   – A take a look at dataset for evaluating the mannequin’s efficiency (Information from 2023).
3. Mannequin High-quality-tuning: We fine-tuned a pre-trained GPT mannequin on our coaching dataset. This course of basically taught the mannequin to “suppose” like our skilled engineers when analyzing defect experiences.
4. Efficiency Analysis: Utilizing the take a look at dataset, we assessed the mannequin’s means to appropriately determine root causes of defects 84% of the time. This step was essential to make sure the mannequin’s reliability earlier than implementation.
5. Integration into Workflow: As soon as glad with the mannequin’s efficiency, we built-in it into our current workflow. Now, as a substitute of manually classifying each report, engineers overview and confirm the mannequin’s classifications, considerably lowering their workload.
6. Steady Enchancment: We carried out a suggestions loop the place engineers can flag any misclassifications. This information is used to periodically retrain the mannequin, guaranteeing its accuracy improves over time.

Now, let’s dive into the technical particulars of our implementation. We rapidly arrange a Jupyter atmosphere for prototyping and testing:

python -m venv venv_jupyter
supply venv_jupyter/bin/activate
pip set up jupyter
jupyter pocket book

Information Preparation

With the Jupyter pocket book able to go, we will import our historic dataset that might be used for fine-tuning:

Producing Coaching Information

We used Python to generate JSONL information for fine-tuning our mannequin. OpenAI’s API requires this particular format. Right here’s a snippet of our code:

# Generate the JSONL information for fine-tuning
jsonl_data = []
# Record of classification choices to incorporate within the immediate
classification_options = ["BATERIA DE CMD.", "CIRCUITO DE TRIP", "COM./ALIMENT. DISJ.",    "COM./ALIMENT. SECC.",    "CSC",    "FALHA COMUNICAÇÃO",    "IHM / SWITCH",    "MEDIÇÃO",    "OSCILOGRAFO",    "PAINEL",    "PROTEÇÃO TRAFO",    "REGULADOR DE TENSÃO",    "RELÉ",    "REMOTA / UAC",    "SINALIZ./ALARMES",    "TP/TC",    "VENTILAÇÃO",    "NO ROOT CAUSE FOUND"]
for _, row in defects_2022.iterrows():
entry = {
"messages": [
{
"role": "system", 
"content": f"You are a classification bot. 
Choose only one of the classifying options. 
If you cannot find a match, do not make up answers 
and instead choose 'NO ROOT CAUSE FOUND'. 
The classifying options are: 
{', '.join(classification_options)}"
},
{"role": "user", "content": row['ROOT CAUSE']},
{"position": "assistant", "content material": row['DEFECT DESCRIPTION']}
]
}
jsonl_data.append(json.dumps(entry))
# Save the info to a JSONL file
jsonl_filename = './fine_tuning_data.jsonl'
with open(jsonl_filename, 'w') as f:
for entry in jsonl_data:
f.write(entry + 'n')

The output might be one thing like this:

We used OpenAI’s API to fine-tune our mannequin (an online interface can also be avaliable of their web site!) After producing an API key and importing our JSONL file, we initiated the fine-tuning course of:

Nice! Now, we’ve a mannequin custom-fit to our use case.

Evaluating Mannequin Efficiency

Lets use information from 2023 to evaluate its efficiency. We begin by loading it in, similar to we did earlier than:

We carried out a script to iterate by means of our dataset and request completions from our fine-tuned mannequin. A easy comparability is made, checking if the return from the API matches the manually labeled information:

responses = []
correct_classification_count = 0
for _, row in defects_2023.iterrows():
entry = {
"messages": [
{
"role": "system",
"content": f"You are a classification bot.
Choose only one of the classifying options.
If you cannot find a match, do not make up answers
and instead choose 'NO ROOT CAUSE FOUND'.
The classifying options are: 
{', '.join(classification_options)}"
},
{"role": "user", "content": row['DEFECT DESCRIPTION']}
]
}
completion = openai.ChatCompletion.create(
mannequin="ft:gpt-3.5-turbo-0613:private::8Bx1qNTB",
messages=entry["messages"]
)
ai_classification = completion.decisions[0].message.content material
responses.append(ai_classification)
# Examine if the AI's classification matches the unique "DEFECT DESCRIPTION"
is_correct = ai_classification == row['DEFECT DESCRIPTION']
if is_correct:
correct_classification_count += 1
# Add the AI's classification and verification column to the dataframe
defects_2023['MODEL CLASSIFICATION'] = responses
defects_2023['VERIFY MATCH'] = defects_2023['MODEL CLASSIFICATION'] == defects_2023['DEFECT DESCRIPTION']
# Calculate the accuracy of the AI
accuracy = correct_classification_count / len(anomaly_history_2023) * 100
print(f"Mannequin's accuracy: {accuracy:.2f}%")Outcomes and Enterprise Affect

Our easy analysis yielded a powerful accuracy fee of 84.51%! That is ok for a primary attempt.

Mannequin's accuracy: 84.51%

We are able to save our outcomes to a csv file to additional course of:

defects_2023.to_csv('./defects_2023_machine_classified.csv')

The implementation and refinement of this GPT-based resolution yielded spectacular outcomes:

1. Time Financial savings: Engineers’ additional time was decreased by roughly 80%, releasing up useful time for extra advanced duties.
2. Value Discount: The corporate saved an estimated $15,000 per engineer yearly in additional time prices.
3. Improved Accuracy: The consistency of classifications improved, resulting in extra standardized upkeep planning.
4. Scalability: The system can deal with elevated workloads with out proportional will increase in human effort.

This venture highlighted a number of key takeaways:

1. AI as a Collaboration Device: GPT wasn’t used to switch engineers however to enhance their capabilities, permitting them to deal with higher-value duties.
2. Information High quality is Essential: The success of the mannequin closely trusted the standard and amount of historic information used for coaching.
3. Steady Studying: Implementing a suggestions loop for steady enchancment was very important for long-term success.

Wanting forward, we’re exploring methods to develop this resolution to different areas of our operations, resembling predictive upkeep and useful resource allocation.

The profitable implementation of GPT in our defect classification course of demonstrates the potential of AI to unravel real-world enterprise issues. By strategically making use of these applied sciences, firms can’t solely scale back prices but additionally empower their workforce to deal with extra useful, artistic duties.

As we proceed to navigate the AI revolution, it’s clear that essentially the most profitable organizations might be these that may successfully mix human experience with synthetic intelligence capabilities. This case research serves as a testomony to the transformative energy of this method.