In the context of Agriculture 4.0, advanced technologies such as the Internet of Things (IoT), artificial intelligence (AI), and big data analytics play a critical role in enhancing the eficiency and sustainability of farming operations. These innovations enable real-time monitoring and decision-making, improving the eficiency, sustainability, and productivity of agricultural systems. Central to Agriculture 4.0 is the deployment of sensors embedded in agricultural machinery, such as tractors, which continuously collect data on key operational metrics, including engine performance, fuel consumption, soil conditions, and equipment health. The efective analysis of such data is essential for predictive maintenance, as early detection of potential anomalies can prevent costly breakdowns and reduce downtime. However, finding real-world datasets containing examples of anomalies in agricultural machinery is highly challenging, making it dificult to develop and assess the efectiveness of anomaly detection models. Additionally, classical methods for anomaly generation, such as stochastic and adversarial approaches, may be dificult to apply given the intricate patterns and time dependency of these data. To address this gap, our work leverages Large Language Models (LLMs) and agentic workflows to generate realistic anomaly scenarios from agricultural data. Using a rule-based approach that combines prompt engineering techniques with a multiagent system, we create synthetic anomalies that can later be used to evaluate anomaly detection models. These models would then enable the timely identification of potential machinery failures, reducing maintenance costs, minimizing downtime, and significantly lowering the environmental impact by preventing ineficiencies such as increased fuel consumption from faulty equipment, reducing the need for replacement parts, and conserving energy and resources used in repairs.
(L. Sartori); 0000-0002-5076-6544 (P. Trunfio) IoT sensor networks are increasingly leveraged in Industry 4.0 and Smart Agriculture to enhance productivity and sustainability through advanced sensing, data fusion, and machine learning. In this context, anomaly detection techniques can be efectively applied for real-time monitoring of machinery and systems, preventing failures and optimizing operational eficiency [
CEUR
ceur-ws.org
and noisy datasets [
To address these limitations, we propose a novel rule-based anomaly generation approach that leverages the context-aware capabilities of Large Language Models (LLMs). Our methodology extends beyond a single LLM by employing LLM agents in a collaborative workflow, where each agent contributes specialized knowledge to produce the final synthetic anomalies. By incorporating LLM agents into the rule generation process, we enable a more informed, context-driven creation of anomalies that better reflect the spatiotemporal complexities of IoT sensor data. This hybrid approach combines the interpretability and simplicity of rule-based methods with the nuanced understanding and adaptability of LLM agents, resulting in a more eficient and realistic anomaly generation process suitable for testing detection algorithms in dynamic, real-world environments.
The main contributions of the paper can be summarized as follows: • We advance the application of LLM agents in Smart Agriculture, showing how such systems can cooperate within an agentic workflow to generate realistic synthetic anomalies. • The proposed method integrates a rule-based approach with the capabilities of LLMs, addressing the limitations of traditional methods in handling high-dimensional spatiotemporal IoT data. • Our approach enhances the testing of anomaly detection systems, leading to more reliable realtime monitoring and improved operational eficiency.
The remainder of the paper is organized as follows. In Section 2, we discuss related work in the field of anomaly generation, highlighting the main applications of LLMs to Smart Agriculture. Section 3 provides an in-depth description of the proposed approach showing its application to a real-world case study. Finally, Section 4 concludes the paper.
Large Language Models (LLMs) have recently gained significant traction due to their remarkable
natural language understanding and generation capabilities [
In this work, we explore how LLM-based agents can be synergistically leveraged in the field of smart agriculture to generate synthetic real-world anomalies. This task is critical for improving and evaluating the performance of anomaly detection systems. Several methodologies have been developed to generate synthetic anomalies that closely resemble real-world scenarios, enabling a robust assessment of detection algorithms. Major approaches in the literature leverage conditional generation approaches and Generative Adversarial Networks (GANs), in which two neural networks—a generator and a discriminator—compete with each other during the training process. Specifically, the discriminator tries to create realistic synthetic data, i.e., anomalous instances, while the discriminator tries to diferentiate between normal and anomalous data. This process leads to the generation of highly realistic anomalies that closely resemble actual outliers, making GANs particularly useful in testing the robustness of anomaly detection systems. As an example, Uzolas et al. leverage conditional GANs for the generation of realistic single-chromosome images following user-defined banding patterns [ 10], while Salem et al. [11] uses a Cycle-GAN to generate synthetic anomalous data from normal data for improving anomaly detection in imbalanced datasets. Zhang et al. [12] introduce DefectGAN, which generates anomaly samples by superimposing learned defect foregrounds onto a normal background, while Niu et al. propose SDGAN [13], which modifies defect-free images to introduce surface defects using a generator trained with cycle consistency loss on both normal and anomalous images. Duan et al. [14] introduce a few-shot defect image generation technique, producing structural anomalies from a limited set of defect samples. It enhances a pre-trained StyleGAN2 backbone by adding defect-aware residual blocks to manipulate features within learned defect masks.
Besides GANs, Difusion Models (DMs) have been also leveraged for generating synthetic anomalies by perturbing normal patterns. DMs are a family of probabilistic generative models that progressively add noise to data and then learn to reverse this process to generate new samples. In the field of anomaly generation, Dai et al. present GRAD [15], an unsupervised anomaly detection framework using a difusion model called PatchDif to generate contrastive patterns by disrupting global structures while preserving local ones. GRAD also includes a self-supervised reweighting mechanism and a lightweight detector to eficiently identify anomalies. Hu et al. [ 16] propose a difusion-based few-shot anomaly generation model, leveraging the strong prior knowledge of a latent difusion model trained on large datasets to improve the realism of generated anomalies. Zhang et al. introduce RealNet [17], another difusion-based approach that relies on Strength-controllable Difusion Anomaly Synthesis (SDAS) to generate synthetic anomalies of varying strengths, mimicking real-world anomalies. RealNet also incorporates feature selection and residual detection methods to improve anomaly detection while managing computational cost, showing significant improvements on several benchmark datasets.
While these anomaly synthesis methods are efective, they depend on real defect images and cannot generate unseen types of anomalies. Furthermore, these methods are usually computationally intensive and often require extensive tuning to produce meaningful results. 3. Proposed Approach: Leveraging LLM Agents for Anomaly
Generation in Agricultural Machinery In this section, we provide a detailed description of the proposed approach aimed at generating realworld anomalies in multivariate sensor data from agricultural machinery, specifically tractors. We leverage an agentic workflow in which diferent LLM agents interact with each other to produce high-quality anomalous test data. The proposed methodology is articulated in two main phases: 1. Best LLM selection via zero-shot operational range generation — First, the best LLM must be selected from all those available, including GPT-4o and LLama3.1. For this purpose, CAN bus sensor data from tractors are analyzed to extract the operational ranges of the diferent variables considered. By comparing these real ranges with those generated by various Large Language Models (LLMs) through zero-shot prompting, we identify the LLM that exhibits the highest level of expertise in the domain of agriculture and tractor operations. 2. Anomaly generation through an agentic workflow — The methodology employs an agentic workflow to generate anomalies, which involves collaboration between two LLM-based agents: () the first agent generates anomaly rules based on insights from the selected LLM; () the second agent transforms the generated rules into executable Python code. This code applies the anomalies to the original non-anomalous data, efectively simulating real-world deviations and faults.
Finally, as the test anomalies are generated, they are used to assess the performance of deep learningbased anomaly detection models. Specifically, an LSTM-based autoencoder is trained on a dataset representing a work session of the tractor and then tested against the synthetic anomalies generated as described above. This approach mimics real-world processes of anomaly detection in agricultural machinery, allowing for an assessment of the efectiveness of the generated anomalies. 3.1. Best LLM selection via zero-shot operational range generation Figure 1 depicts the flowchart used in the first phase of the methodology, dedicated to selecting the LLM that exhibits the highest expertise in the agricultural domain, specifically regarding tractors and their sensor data. The selection process involves several LLMs, specifically GPT-4o, Llama 3.1 70B, Gemini Pro, and Mistral Large 2. Their efectiveness is measured by their ability to generate operational ranges for key tractor variables, which are then compared to the actual ranges extracted from tractor sensor data.
Real operational range extraction Sensor data
Operational range generation via Zero-shot Prompting Range evaluation
Jaccard Score
LLM selection
In the following yellow box, we report the prompt used for querying the diferent LLMs to generate operational ranges of variables. Each model is provided with a prompt containing the variable name, its unit of measurement, and a description. Generation is performed through zero-shot prompting, which means that the prompt used to interact with the model does not include any example or demonstration.
As a seasoned expert in New Holland T7 165 S tractors, we seek your expertise in diagnosing various operational variables retrieved from the CAN bus of the tractor. You are provided with a list of variables, each with its name, unit of measurement, and description. These variables are listed according to the following format: - <var_name> (<unit>): <description>. Your task is to generate the operational range of each variable, which jointly takes into account the diferent activities performed by the tractor, i.e. idling, moving, plowing, and turning.
Format your output as follows: - <var_name>: <operational_range> (<unit>) - … - CAN1.LFE1.EngineFuelRate (l/h): Amount of fuel consumed by the engine per unit of time. - CAN1.EFLP1.EngineOilPressure1 (kPa): Gage pressure of oil in the engine lubrication system as provided by the oil pump.
- …
Table 1 presents the operational ranges generated by each LLM for the various key variables associated with tractor sensor data. Each row of the table details the ranges produced by the evaluated models for a given variable, while the final column provides the actual ranges extracted from the sensor data. This comparative analysis highlights the discrepancies and alignments between the internal knowledge of LLMs and real-world data, which are crucial for determining the most efective LLM for the subsequent phases of the methodology.
To quantitatively assess the accuracy of LLM-generated ranges, we compared them with ground truth values derived from tractor sensor data by introducing a continuous version of the Jaccard index that quantifies the similarity between two ranges. Given two intervals [ 1, 1] and [ 2, 2], where 1 and 1 represent the lower and upper bounds of the first interval, and 2 and 2 represent the bounds of the F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26
) − min( 1, 2) be the union of the two intervals (i.e., total covered range length). max(0, min( 1, 2
) − max( 1, 2)) be the intersection of the intervals, which is calculated based on the overlap between the intervals. = 0 if the intervals do not overlap. Otherwise, represents the length of the overlapping interval.
Then, the Jaccard similarity for intervals can be expressed as ([ 1, 1], [ 2, 2]) =
, with ∈ [
Jaccard score: 0.68
0.12 0.81 Jaccard score: 0.42 0.46
0.04 0.58 Jaccard score:
0.51 0.50 0.38
ℳ represents the percentage of features where ℳ achieved a higher Jaccard score compared to ℳ .
For each variable, the average Jaccard similarity score was calculated across all comparisons between the real and generated ranges. The LLM with the highest average Jaccard score was selected as the most appropriate model for generating anomaly rules in the subsequent steps of the proposed methodology. Figure 2 illustrates the win rates of the evaluated LLMs alongside the average Jaccard interval scores achieved by each model. The plot shows that GPT-4o consistently outperforms all other models and demonstrates good accuracy in generating intervals that closely resemble the actual operational ranges extracted from tractor sensor data, confirming its suitability as the chosen model. 3.2. Anomaly generation through an agentic workflow Once the most appropriate LLM is selected, the anomaly generation process is performed through an agentic workflow, as illustrated in Figure 3.
LLM selection
Agent 1 - Expert Farmer
Prompt chaining Generation of realworld anomaly
instances Generation of a set of rules for each anomaly instance
Python script generation Agent 2 - Expert Developer
Zero-shot prompting • Expert farmer : Its role is to generate realistic cases of anomalies in the form of rules that can be applied to test data, resulting in anomalous test instances. • Expert developer : Its role is to convert the set of rules generated by the expert farmer into a runnable Python script, which can be executed, via tool use, on the test dataset to produce a structured set of anomalous test instances for benchmarking anomaly detection methods.
In the following sections, the prompts used to query the LLM-based agents are shown, along with the generated output. 3.2.1. Expert farmer Agent — Anomaly generation via prompt chaining In this step, the prompt chaining technique is employed to generate meaningful anomaly instances, as indicated by the green-colored boxes. Using prompt chaining, a sequence of prompts generates complex outputs by linking multiple tasks together. Initially, the first agent (i.e., the expert farmer ) generates a set of significant anomaly cases across various activities, such as plowing, moving, turning, and idle operations. These anomalies are then used to create rules that modify the operational ranges of the variables, thereby generating anomalies. For each anomaly, a corresponding rule is created that specifies its duration and how the operational ranges are altered to simulate the anomaly within the data. These rules are then passed to the second agent (i.e., the expert developer ) for further processing.
As a seasoned expert in New Holland T7 165 S tractors, we seek your expertise in diagnosing various operational variables retrieved from the CAN bus of the tractor. You are provided with a list of variables, each with its name, operational range, unit of measurement and description. These variables are listed according to the following format: - <var_name> (<operational_range> <unit>): <description>. Your task is to generate instances of significant anomalies based on the activity performed by the tractor, i.e., “plowing,” “moving,” “turning,” “starting,” and “idling”. Each anomaly instance must include: - a description of the anomaly instance - the list of variables involved in the anomaly instance - the activity performed by the tractor when the anomaly shows up.
Format your output as follows: - <instance_name>: <description> - variables involved: - <var_name> - … - <activity_performed> - … Input variables: - CAN1.LFE1.EngineFuelRate (0 - 29.35 l/h): Amount of fuel consumed by the engine per unit of time. - CAN1.EFLP1.EngineOilPressure1 (96 - 536 kPa): Gage pressure of oil in the engine lubrication system as provided by the oil pump.
- … Based on: () the generated anomaly instances, () the descriptions, () the activities performed, and ( ) the operational range of the involved variables, generate a set of rules for each anomaly instance describing how each variable involved varies numerically. Also, specify the overall duration of the anomaly for each instance. Consider that the session in which the anomalies will be applied lasts approximately 2 hours, with observations recorded at a frequency of 1 Hz.
Format your output as follows: - <instance_name> (<activity_performed>): - … • The anomaly name, which concisely describes the issue. • The performed activity during which the anomaly occurs. • An issue description that provides useful details on how the anomaly afects the normal operation of the tractor. • The duration of the anomaly. • The variables afected.
• The associated rules specifying how each variable deviates from its expected range over time. ID Anom. Name Activity Issue description
Dur. Involved features
Fuel The tractor shows unusually high fuel 1 Consumption Plowing consumption during operation, despite
Spike consistent speed and load. Instantaneous fuel economy drops sharply, and the fuel rate is well above normal.
10 min
CAN1.LFE1.EngineInstantaneousFuelEconomy CAN1.LFE1.EngineFuelRate
CAN1.EEC2.EnginePercentLoadAtCurrentSpeed Ifnrocmreaasensortmoaabloravneg8e0%of 30-50%. 2 3 4
Overheating Engine Torque Instability Battery Voltage
Drop
Anomaly instances generated by GPT-4o. Each instance includes a description and a set of associated features. 3.2.2. Expert developer Agent — Python script generation and application of rules to test data The second agent, acting as a Python programming expert, is prompted to transform the anomaly rules, generated by the expert farmer LLM agent, into an executable Python script.
As an expert Python developer, we seek your assistance in code scripting. You are provided with a set of rules for diferent anomaly instances that describe how each variable involved varies numerically, along with the overall duration of the anomaly. Anomaly instances are listed according to the following format: - <instance_name> (<activity_performed>): - <var_name>: <rule_description> Based on this information, generate a Python function that applies a given anomaly instance to a time series of sensor data. The code must adhere to the following requirements: - all anomaly instances are handled; - random values are used instead of fixed anomalous values; - the input dataframe is read from a csv given as input; the start time and the anomaly to be applied are given as input; - output the required function without any example usage.
Input anomaly instances: - Fuel Consumption Spike (Plowing): - 10 minutes
In this case, as shown in the blue-colored box, zero-shot prompting is employed, wherein the agent generates a Python script based on the provided anomaly rules without any prior examples or specific training data. The script is designed to take the clean, non-anomalous test dataset as input and apply the anomalies according to the rules generated by the first agent. The generated script is executed to create four distinct datasets by applying the anomalies to the test dataset for each possible activity.
Through this agentic workflow, the entire process of anomaly rule generation and application can be automated, providing a robust method for simulating consistent anomalous behaviors. This, in turn, supports the evaluation of anomaly detection models, by providing realistic and domain-specific anomalies that accurately reflects potential issues that could arise in real-world operations. 3.3. Auto-encoder evaluation on synthetic test anomalies Here we show how the previously generated anomalous test datasets can be efectively leveraged to assess the efectiveness of a deep learning-based anomaly detection model. In particular, for each possible activity, including plowing, moving, turning, or idle, an LSTM autoencoder is trained on a normal working session, encompassing non-anomalous data from CAN bus sensors (see Figure 4).
The LSTM autoencoder works by reconstructing the input time series. A large reconstruction error suggests that the input data may deviate from normal patterns, indicating an anomaly. The detection performance of each autoencoder is measured using the Area Under the Receiver Operating Characteristic Curve (AUC) score. It ranges from 0 to 1, where a score of 1 indicates perfect separation between anomalies and normal data, while 0.5 suggests that the model is equivalent to random guessing.
Figure 5 presents the ROC curves for the four anomalous instances considered during the anomaly generation process. Each curve illustrates the model’s ability to distinguish between anomalous and non-anomalous data across a diverse set of potential scenarios. Specifically, two cases (figure 5b and 5d) achieve perfect classification (AUC = 1.00), while the other two cases (figure 5a and 5c) show strong (AUC = 0.90) and moderate (AUC = 0.76) performance, respectively. These results suggest that the model is highly efective in detecting anomalies, with some variability depending on the specific type of anomaly and the amount of training data from sensors. Furthermore, the ability to generate activity-specific test data facilitates a more granular analysis of model performance, providing insights into how diferent types of anomalies might be detected in real-world deployments. 1.0 tae0.8 R ive0.6 it s Po0.4 e rTu0.2 AUC Score = 0.90 0.00.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate 1.0 tae0.8 R ive0.6 it s Po0.4 e rTu0.2 AUC Score = 1.00 0.00.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate 1.0 tae0.8 R ive0.6 it s Po0.4 e rTu0.2 AUC Score = 0.76 0.00.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate 1.0 tae0.8 R ive0.6 it s Po0.4 e rTu0.2 AUC Score = 1.00 0.00.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate (a) Fuel Consumption Spike (b) Overheating Engine (c) Torque Instability (d) Battery Voltage Drop
In this work, we advance the application of LLM agents in Smart Agriculture by proposing a rulebased approach for the automatic generation of synthetic anomalies in agricultural machinery. By generating realistic, domain-specific anomalies, the system creates a rich dataset that accurately reflects potential issues that could arise in real-world operations. This enables efective evaluation of anomaly detection models and allows researchers and developers to test their algorithms against a variety of plausible scenarios. The generated datasets support thorough benchmarking, helping to identify the strengths and weaknesses of diferent anomaly detection methods. Moreover, the ability to generate diverse datasets tailored to specific activities—such as plowing, moving, turning, and idling—facilitates more granular analysis of model performance. This can lead to insights into how diferent types of anomalies might afect operational eficiency, safety, and tractor maintenance. Ultimately, the proposed methodology fosters an iterative feedback loop, where the performance of anomaly detection models can be continuously improved based on simulated data. This enhances their robustness and reliability in realworld applications, ensuring eficient utilization of agricultural resources and paving the way for more sustainable agricultural practices. Future work will focus on integrating domain-specific knowledge through agentic RAG (Retrieval-Augmented Generation), further improving context awareness of the system and enabling LLMs to better comprehend complex scenarios.
This work has been funded by the project “AGRITECH: National Research Centre for Agricultural Technologies” - CUP CN00000022, of the National Recovery and Resilience Plan (PNRR) financed by the European Union “Next Generation EU”, and by the “FAIR – Future Artificial Intelligence Research” project - CUP H23C22000860006. comprehensive review from an extension standpoint on artificial intelligence and machine learning, Indian Research Journal of Extension Education 24 (2024) 108–123. [10] L. Uzolas, J. Rico, P. Coupé, J. C. SanMiguel, G. Cserey, Deep anomaly generation: An image translation approach of synthesizing abnormal banded chromosome images, IEEE Access 10 (2022) 59090–59098. [11] M. Salem, S. Taheri, J. S. Yuan, Anomaly generation using generative adversarial networks in host-based intrusion detection, in: 2018 9th IEEE Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON), IEEE, 2018, pp. 683–687. [12] G. Zhang, K. Cui, T.-Y. Hung, S. Lu, Defect-gan: High-fidelity defect synthesis for automated defect inspection, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 2524–2534. [13] S. Niu, B. Li, X. Wang, H. Lin, Defect image sample generation with gan for improving defect recognition, IEEE Transactions on Automation Science and Engineering 17 (2020) 1611–1622. [14] Y. Duan, Y. Hong, L. Niu, L. Zhang, Few-shot defect image generation via defect-aware feature manipulation, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 2023, pp. 571–578. [15] S. Dai, Y. Wu, X. Li, X. Xue, Generating and reweighting dense contrastive patterns for unsupervised anomaly detection, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2024, pp. 1454–1462. [16] T. Hu, J. Zhang, R. Yi, Y. Du, X. Chen, L. Liu, Y. Wang, C. Wang, Anomalydifusion: Few-shot anomaly image generation with difusion model, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2024, pp. 8526–8534. [17] X. Zhang, M. Xu, X. Zhou, Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16699–16708.