Engine fault early warning system and method based on multi-agent cooperation

Abstract

The invention relates to the technical field of engine fault diagnosis and predictive maintenance, in particular to an engine fault early warning system and method based on multi-agent collaboration, and the system comprises six collaboration agents: data perception and preprocessing, feature analysis, correlation analysis, time sequence prediction, fault classification and intelligent decision generation. According to the system, multi-source heterogeneous data are aligned through dynamic time warping, online updating of feature thresholds is achieved through EWMA, small sample fault classification is achieved by combining data-mechanism dual-drive correlation analysis and based on a transfer learning Transform model, and finally a priority decision is generated through a multi-level rule base. According to the invention, the early warning accuracy, the real-time performance and the system self-evolution capability are effectively improved.

Description

Technical Field

[0001] This application belongs to the field of fault early warning technology, and in particular relates to an engine fault early warning system and method based on multi-agent cooperation. Background Technology

[0002] As a core component of power systems in critical fields such as aviation, aerospace, shipbuilding, and energy, the operating status of engines directly affects the safety, reliability, and economy of the entire equipment system. Due to the extreme operating environment, complex internal structure, and variable operating conditions of engines, a failure can not only lead to equipment downtime and soaring maintenance costs, but also potentially cause catastrophic consequences in high-risk scenarios. Therefore, building an efficient, accurate, and real-time engine fault early warning and health management mechanism has become a core requirement in modern industrial intelligent operation and maintenance systems.

[0003] Currently, mainstream engine fault early warning technologies mainly include alarm systems based on fixed thresholds, modeling and analysis methods based on physical mechanisms, and data-driven diagnostic schemes based on single artificial intelligence models. Fixed threshold methods rely on manual experience to set parameter limits, making it difficult to adapt to the characteristics of engine performance degradation over time and dynamic changes in operating conditions. They generally suffer from problems such as response lag, high false alarm rates, and inability to identify early progressive faults. While physical mechanism models have good interpretability, their construction process is highly dependent on expert knowledge, resulting in high modeling complexity, large computational overhead, and difficulty in covering the entire operating range, aging states, and unknown fault modes, thus limiting their generalization ability. Traditional artificial intelligence methods typically adopt the idea of ​​"one model solving all problems," attempting to complete all tasks from data processing to fault classification through a single complex model. This not only presents challenges in training and strong dependence on labeled data but also lacks modular collaboration and system-level adaptive capabilities, performing poorly when facing multi-source heterogeneous data, small-sample fault scenarios, and real-time decision-making requirements.

[0004] More importantly, the aforementioned existing technologies generally lack multi-dimensional information fusion, cross-module collaborative reasoning, and closed-loop feedback optimization mechanisms, making it difficult to achieve the transformation from passive alarm to proactive predictive maintenance. Therefore, there is an urgent need for a new fault early warning architecture that can integrate multi-dimensional capabilities such as data perception, feature analysis, correlation modeling, time series prediction, and intelligent decision-making, and possess distributed collaboration, adaptive learning, and evolvable characteristics, in order to overcome the bottlenecks of existing technologies in terms of accuracy, real-time performance, robustness, and intelligence. Summary of the Invention

[0005] This application addresses the shortcomings of existing engine fault early warning technologies, such as fixed threshold alarm lag, poor generalization ability of physical models, and lack of collaborative mechanisms and self-evolution capabilities of single artificial intelligence models. Therefore, this invention adopts the following technical solution: This invention provides an engine fault early warning system based on multi-agent collaboration, belonging to the field of engine health management and predictive maintenance technology, and is particularly suitable for engine testing processes in high-reliability scenarios such as aviation, aerospace, shipbuilding, and energy. The system constructs a multi-agent system composed of multiple functionally collaborative and loosely coupled agents to achieve real-time perception, deep fusion, intelligent analysis, and closed-loop decision-making of multi-source heterogeneous engine data. Each agent undertakes tasks such as data perception and preprocessing, feature analysis, correlation analysis, time-series prediction, fault classification, and intelligent decision generation, and interacts with information and merges results through standardized interfaces to form a unified state assessment and maintenance recommendation output.

[0006] This invention also provides an engine fault early warning method based on multi-agent cooperation, which includes the following steps: Step S1: The data perception and preprocessing agent receives multi-source heterogeneous raw data streams from the engine test environment, including sensor data and environmental parameters with different sampling rates, formats, and completeness. The agent first performs real-time quality assessment on the incoming raw data streams, determines whether the engine is in a stable operating condition based on its operating status, and adaptively fills in missing values ​​using linear interpolation or a regression model trained on history. Then, a dynamic time warping algorithm is used to time-align other sequences with the highest sampling rate sequence as the benchmark, constructing a dataset with a unified time benchmark.

[0007] The stable operating condition is determined as follows: engine speed fluctuation is less than 1% of the maximum speed and the duration exceeds 5 seconds. If this condition is met, linear interpolation is used to fill in the missing data; otherwise, a regression model trained based on historical normal data is used for predictive interpolation. The specific steps of the dynamic time warping algorithm include: designating the data with the highest sampling rate as the reference time series, and the remaining data as the series to be aligned; calculating the Euclidean distance between any two points in the reference series and the series to be aligned to form a distance matrix; finding the path with the minimum cumulative distance from the starting point to the ending point in the matrix; mapping the data points in the series to be aligned to the timestamps of the reference series according to the path; for timestamps without direct mapping points, a linear interpolation algorithm is used to calculate the corresponding values ​​based on the path trend to complete the time synchronization and alignment of all data.

[0008] Step S2: The feature analysis agent is used to extract local features and detect anomalies in the preprocessed time-series waveform data. The agent uses a sliding window method to segment the continuous data and calculates statistical features such as mean, standard deviation, peak-to-peak value, skewness, and kurtosis for each segment. The dynamic threshold range of each feature is initialized based on historical normal data. During system operation, the threshold is continuously updated using an exponentially weighted moving average algorithm for samples determined to be normal, achieving online self-learning. When the number of real-time feature values ​​exceeding the current dynamic threshold exceeds the set upper limit, it is determined to be an anomaly and an early warning is triggered.

[0009] Step S3: The association analysis agent is used to explore the multidimensional correlation between engine performance indicators and environmental operating parameters. The agent combines Pearson correlation coefficient and Spearman rank correlation coefficient for statistical analysis, generates a correlation strength matrix and visualizes it through a heat map. Furthermore, the key influencing factors discovered by data-driven discovery are input into the engine physical mechanism model, and its physical rationality is verified through simulation, thus constructing a "data-mechanism" dual-driven analysis framework.

[0010] The Pearson correlation coefficient is used to measure the degree of linear correlation between variables; the Spearman rank correlation coefficient is used to measure the monotonic nonlinear relationship between variables, and is obtained by calculating the Pearson correlation coefficient after sorting the original data; the physical mechanism model includes a thermodynamic cycle model, a fluid dynamics simulation module, and a component-level performance degradation model, which are used to verify whether the influencing factors identified by data-driven analysis conform to known physical laws.

[0011] Step S4: The time-series prediction agent is used to make high-precision predictions of future key parameter trends. To address the challenge of modeling long sequences due to high sampling rates, the agent first uses average pooling for downsampling. Then, the downsampled historical multivariate sequences are input into a deep learning model based on the Seq2Seq architecture. This model integrates an attention mechanism and applies a causal mask in the decoder to ensure that the prediction process does not leak future information.

[0012] The average pooling sliding window size is 20, meaning that the average of every 20 points of the original 0.01-second data is used to obtain a new sequence with a step size of 0.2 seconds. The Seq2Seq model's encoder and decoder both use LSTM or GRU units, with a hidden layer dimension of 128 and 4 layers. The attention mechanism uses additive attention or scaled dot product attention, and the decoder is only allowed to access information from time t and earlier when generating the output at step t. The model training configuration includes: an input sequence length of 200 (corresponding to 40 seconds of historical data), 50 prediction steps (corresponding to 10 seconds of future data), a loss function of mean absolute error (MAE), an optimizer of Adam, a learning rate of 0.0005, a batch size of 32, 20 training epochs, and a Dropout ratio of 0.4.

[0013] Step S5: The fault classification agent is used to identify and classify specific fault modes. The agent builds a classification model based on the Transformer encoder and uses the self-attention mechanism to capture global dependencies in long sequences. To address the problem of scarce engine fault samples, it is first pre-trained on a large public mechanical fault dataset, and then the classification head is fine-tuned based on a small amount of real engine fault data to achieve high-precision classification under small sample conditions.

[0014] The Transformer encoder comprises a multi-head self-attention layer and a feedforward neural network layer, with the input sequence being normalized multivariate temporal features. The pre-training phase employs a cross-entropy loss function, initializing model parameters on a public dataset. The fine-tuning phase freezes all weights of the Transformer backbone network, training only the last added fully connected classification head, with the fine-tuning loss function also using cross-entropy. The fault categories include normal state, valve fault, turbopump fault, and pipeline rupture fault. The model output generates probability distributions for each category using a Softmax function, selecting the category with the highest probability as the final classification result and its confidence level.

[0015] Step S6: The intelligent decision-making agent acts as the system's decision-making hub, responsible for integrating the structured outputs of the four analytical agents to form a unified state vector, and generating executable maintenance suggestions based on the built-in multi-level rule base; the state vector includes fault type, prediction trend, abnormal features, related factors and comprehensive health; the rule base includes at least three types of rules: (1) fault emergency rule layer, which directly triggers the corresponding emergency response plan based on the fault classification results; (2) predictive maintenance rule layer, which generates preventive maintenance suggestions based on the time-series prediction results; (3) abnormal handling rule layer, which generates targeted handling plans based on the abnormal patterns identified by feature analysis and correlation analysis.

[0016] Compared with existing technologies, the beneficial effects of this application's engine fault early warning system and method based on multi-agent collaboration are as follows: By constructing a multi-agent collaborative architecture composed of a data perception and preprocessing agent, a feature analysis agent, a correlation analysis agent, a time series prediction agent, a fault classification agent, and an intelligent decision generation agent, a closed-loop processing of multi-source heterogeneous engine data is achieved; among them, the dynamic threshold mechanism and EWMA algorithm enable feature analysis to have online self-learning capabilities; DTW time alignment and average pooling solve the problems of multi-rate data fusion and long sequence modeling; "data The "data-mechanism" dual-driven correlation analysis enhances the credibility and interpretability of the conclusions; the small-sample fault classification based on transfer learning improves the practicality of the model in real industrial scenarios; the multi-level rule base-driven intelligent decision-making mechanism ensures the operability and safety of the output suggestions; the loosely coupled design of each agent ensures the maintainability and evolvability of the system; in summary, this invention effectively overcomes the defects of existing technologies such as slow response, high false alarm and false negative rates, weak model generalization ability, and lack of system-level collaborative intelligence, and significantly improves the accuracy, real-time performance, interpretability, and system adaptability of engine fault early warning. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the control method in an embodiment of the present invention; Figure 2 This is a block diagram illustrating the principle of the control method in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the timing prediction agent in an embodiment of the present invention. Figure 4 This is a block diagram of the control system in an embodiment of the present invention. Detailed Implementation

[0018] Example 1: To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0019] Please refer to Figures 1-4 This embodiment is an engine fault early warning method based on multi-agent cooperation, which includes the following steps: S1. Multi-source heterogeneous data perception and preprocessing: Receives multi-source heterogeneous data from the engine test environment, processes missing values ​​based on mathematical statistics and machine learning algorithms, performs data alignment using a dynamic time adjustment algorithm, and outputs a standardized time-series dataset. The multi-source heterogeneous data includes environmental data and engine test data. By preprocessing the multi-source heterogeneous data, data quality is improved, enabling the method in this embodiment to adapt to different engine models, sensors, sampling rates, and data formats, achieving a unified data standard system.

[0020] The multi-source heterogeneous data sensing and preprocessing specifically includes the following steps: Step S11: Perform real-time quality assessment on the incoming raw data stream: Use mathematical statistics and machine learning methods to automatically identify and mark missing data in the data, and fill in the missing data.

[0021] The methods for filling in missing data can be either using a regression model trained on historical data for data prediction and imputation, or directly using linear interpolation.

[0022] Specifically, the engine's stability is determined using speed sensor data. The criteria are: a stable state is defined as an engine speed fluctuation less than a fixed threshold for a duration exceeding a certain time threshold; otherwise, it is considered an unstable state. If the engine is in a stable state, linear interpolation is used to fill in missing data; if it is in an unstable state, a regression model trained on historical data is used for data prediction and interpolation.

[0023] For example, the criteria for judging the stable state of an engine are: if the engine speed fluctuation is less than 1% of the maximum speed and the duration is greater than 5 seconds, it is judged to be in a stable state.

[0024] Step S12: For sensor data with different sampling rates, the Dynamic Time Warping (DTW) algorithm is used for data synchronization and alignment; a data queue with a unified time reference is constructed to ensure the consistency of multi-source data in the time dimension, providing a foundation for subsequent correlation analysis. Specifically, this includes: Step S121: Designate the data with the highest sampling rate as the baseline time series. Other data that needs alignment are treated as the sequence to be aligned. , Where n is the number of samples in the baseline time series X, and m is the number of samples in the sequence Y to be aligned; Step S122: Calculate the baseline time series With the sequence to be aligned Any two points and The Euclidean distance between them forms a Distance matrix ; Step S123: In the distance matrix In the middle, find a way from arrive path This results in the cumulative distance along the path. Minimum; Step S124: Align the sequence to be aligned Data points in the dataset are mapped to the baseline time series. On the timestamp; for the base time series In the path For timestamps without direct mapping points, a linear interpolation algorithm is used, based on the path. The corresponding mapping trend is calculated. Sequence values ​​are used to synchronize and align all data over time, forming a standard dataset sample. .

[0025] Step S2: Waveform Feature Extraction and Dynamic Threshold Analysis: Using methods such as sliding window and dynamic thresholding, key features such as peaks and troughs in engine test waveform data are automatically extracted and analyzed. Based on historical data, a data threshold exceeding standard is automatically constructed, and the threshold is dynamically updated upon receiving new data, enabling online learning and adjustment of the judgment mechanism, significantly reducing false alarm and false negative rates. Specifically, this includes: Step S21: Data Segmentation: Using the sliding window method, the standard dataset samples Z from step S124 are converted into fixed-length data segments. This process preserves temporal locality characteristics. The sliding window length is set according to the engine's main rotational frequency cycle; for example, when the main frequency is 100Hz, the window length is set to 0.02 seconds. The sliding step size is equal to the window length to ensure complete coverage of all data. Through this step, the continuous time-series data stream is converted into a series of fixed-length data segments.

[0026] Step S22: Feature Extraction: For each data segment, calculate its basic statistical features, including the mean. Standard deviation Peak-to-peak value (i.e., the difference between the maximum and minimum values), skewness and kurtosis .

[0027] Step S23: Dynamic Threshold Construction and Anomaly Detection: A dynamic threshold anomaly detection method is used to independently construct and update thresholds for each feature extracted in step S22. The threshold range is initialized based on historical normal data, and the threshold statistics are dynamically updated for subsequent data judged as normal, thereby achieving adaptive learning. Anomaly detection is achieved by comparing real-time feature values ​​with their corresponding current dynamic thresholds. Specifically, this includes: Step S231: Threshold Initialization: In the initial stage of system deployment, collect historical normal data for a period of time, extract its features, and use the extracted features as the initial threshold. For example, the initial threshold is calculated based on the historical data of the first 10 normal test runs, and the average value of each feature is taken. Add or subtract three standard deviations As the upper and lower limits, the initial threshold range is set to .

[0028] Step S232: Threshold Dynamic Update: After the system is put into online operation, the feature values ​​calculated for data segments that are not judged as abnormal are used to dynamically update the threshold statistics. The Exponential Weighted Moving Average (EWMA) algorithm is used for updating, taking into account both historical memory and the latest changes.

[0029] Taking the mean as an example, the initial threshold for the mean. ,in, It is the forgetting factor ( ), This is the current normal sample value.

[0030] Step S233: Anomaly detection: Based on features extracted from real-time data (mean) Standard deviation Peak-to-peak value skewness and kurtosis The data segment is compared with the latest threshold. If the number of features exceeding the threshold exceeds the upper limit, the data segment is determined to be in an abnormal state, and an alert is triggered.

[0031] Step S3: Multi-dimensional Data Correlation Analysis: Statistical analysis methods are used to analyze the linear and nonlinear relationships between engine test data and environmental operating condition data, providing explanations and optimization suggestions based on mechanistic models. Specifically, this includes: Step S31: Calculate the linear correlation strength between engine performance indicators (such as thrust, fuel consumption, turbine outlet temperature, etc.) and environmental operating parameters (such as intake air temperature, atmospheric pressure, fuel temperature, etc.) using the Pearson correlation coefficient method; generate a correlation coefficient matrix and visualize it through a heat map to identify key environmental factors with significant linear effects (such as a strong nonlinear negative correlation between fuel temperature and turbine speed). Step S32: Use the Spearman method to calculate the strength and direction of the monotonic nonlinear correlation between engine performance indicators and environmental operating condition parameters, and analyze the complex nonlinear interactions between variables; Step S33: Combine the data-driven analysis results with the engine physical mechanism model; based on the key influencing factors output by the correlation analysis, conduct simulation verification through the mechanism model to provide a physical explanation for the data analysis results; establish a "data-mechanism" dual-driven analysis framework to enhance the credibility and interpretability of the conclusions.

[0032] Step S4: Key Parameter Timing Prediction: Based on an attention mechanism-based timing prediction model architecture, a deep learning model suitable for engine timing prediction tasks is designed and trained. The trained model is deployed to the experimental system to perform rolling predictions on real-time acquired key engine parameters (such as turbine speed and combustion chamber pressure), and outputs structured analysis results including prediction trends, over-limit times, and prediction confidence. Specifically, this includes: Step S41: When using a deep learning model for time series prediction, the model's memory length (i.e., effective receptive field) for processing long sequence data is limited. If the original data sampling rate is extremely high, resulting in excessively long individual sequences, model training will be extremely difficult, and the model will struggle to learn long-term dependencies. Therefore, it is necessary to downsample the original data. Here, the average pooling method is used to downsample the original time series data. Specifically, a sliding window of N is set, and the mean of every N samples is taken as the sample value for that window.

[0033] For example: the original time series data step size is 0.01s. The original data is divided into groups of 20 (i.e., sliding window N=20), which changes the data step size from 0.01s to 0.2s. Then, the data of each group of 20 is averaged (the average is taken as a point).

[0034] Step S42: Input the downsampled historical test data into the time series prediction model for training to capture complex time dependencies. Specifically, this includes: Step S421: Input and Output Construction: Construct the training samples for the model from the downsampled historical test data. For a given multivariate time series dataset... ,in This represents a vector containing all sensor data at time t. A historical sequence is selected. As input to the encoder, a sequence immediately following the encoder input is selected. As the target output of the decoder.

[0035] Step S422: Model architecture design: The basic model is Seq2Seq. In the decoder, attention masking is used to ensure that when predicting the t-th time step, only information before the t-th time step (including the t-th time step) can be accessed, thus preventing information leakage.

[0036] Step S423: Model Training: Using the constructed large number of training samples The input model is used for iterative training. The hyperparameters are configured as follows during the model training process: Loss function: Mean Absolute Error (MAE); Optimizer: Adam; Input sequence length: 200 (i.e., input the first 40 seconds of data); Prediction steps: 50 (i.e., predicting data for the next 10 seconds); Batch size: 32; Training epochs: 20; Learning rate: 0.0005; Hidden layer size: 128; Number of model layers: 4; Dropout: 0.4.

[0037] Step S5: Fault Classification: A fault classification agent is used to identify and classify fault patterns using a self-attention-based classification model combined with transfer learning, with particular emphasis on optimizing recognition performance under small sample sizes. Specifically, this includes: Step S51: Based on historical engine test data Construct a sample library for fault classification, in which The feature vector sequence of the i-th sample ( For sequence length, (for feature dimensions) For its corresponding one-hot encoded fault label, This represents the number of fault categories.

[0038] Normalize the features: ; in, These are the original eigenvalues. and These are the mean and standard deviation of the feature across all training samples, respectively. These are the normalized eigenvalues.

[0039] The failure modes include several known categories such as normal state, valve failure, turbine pump failure, and pipeline rupture failure.

[0040] Step S52: Employ a Transformer encoder based on a self-attention mechanism as the core classification model to capture features of long sequence data. Given an input sequence... The computational process of its self-attention mechanism is as follows: Step S521: Linear projection: , , ,in This is a learnable weight matrix.

[0041] Step S522: Scaling the dot product attention: , The feature dimension of the projected query, key, and value vector is a configurable hyperparameter.

[0042] Step S523: Multi-head attention: Concatenate the outputs of multiple self-attention heads and project them again. .

[0043] Step S53: To address the issue of limited sample size, the Transformer model is pre-trained based on a large, publicly available time-series dataset of mechanical faults, using the cross-entropy loss function: , in, The category to which the sample is predicted by the model The probability, It is a publicly available dataset.

[0044] Step S54: Based on a small dataset of engine faults accumulated from historical tests By freezing all weights of the Transformer encoder Only the last added classification head is trained. We fine-tune the pre-trained Transformer model to achieve fast convergence and excellent performance under small sample conditions. The loss function for fine-tuning is: .

[0045] Step S55: Process the preprocessed multi-source heterogeneous data from step S1 The input is fed into a pre-trained fault classification model for forward propagation inference; the model's output layer generates probability distributions for each fault mode using the Softmax function. ;

[0046] in, The classification header corresponds to the category. The system then selects the category with the highest probability value as the final classification result and outputs the label and corresponding confidence level of the failure mode, forming a structured analysis result.

[0047] Step S6: Intelligent Decision Generation: The intelligent decision generation agent, acting as the decision-making center of the entire multi-agent system, is responsible for generating specific and actionable maintenance recommendations based on the outputs of other analytical agents and pre-defined decision rules. This includes: Step S61: The decision-generating agent continuously listens to and receives the structured output results from each analysis agent in steps S2 to S5, and performs time alignment on this multi-source information to form a unified structured state vector. The structured state vector includes the following fields: fault diagnosis (including fault type, confidence level, and timestamp), time series prediction (including key parameters, prediction trend, over-limit time, and prediction confidence level), feature analysis (including anomaly type, anomaly level, and key features), correlation analysis (including main influencing factors, correlation strength, and correlation type), and comprehensive health (calculated based on the weighted average of scores from each subsystem).

[0048] Step S62: The intelligent decision-making agent internally constructs and maintains a multi-level decision rule base, which contains three types of rules: The first layer consists of emergency response rules, which directly generate response plans based on the diagnostic results of the fault classification agent. For example, if the fault type is valve fault and the confidence level is >0.8, the machine should be stopped immediately for inspection.

[0049] The second layer consists of predictive maintenance rules, which combine the prediction results of the time-series predictive agent to generate preventative maintenance recommendations. For example, if the turbine speed is predicted to exceed the limit within 15 minutes and the prediction confidence is >0.8, then it is recommended to reduce the load.

[0050] The third layer consists of anomaly handling rules. Based on the anomaly results from feature analysis and correlation analysis, specific handling plans are generated. For example, if the main frequency amplitude drops by more than 30% and shows a strong nonlinear negative correlation with fuel temperature, the fuel supply strategy is adjusted and monitoring is strengthened.

[0051] Step S63: The intelligent decision-making agent integrates the solutions generated after matching all rules and outputs them in order of priority (emergency > prediction > anomaly). For example, "Immediately stop the engine, check the fuel regulating valve, and record the complete test data for model iteration."

[0052] The state vector data structure is in JSON format and includes fields such as "fault diagnosis", "time series prediction", "feature analysis", "correlation analysis" and "overall health".

[0053] An example of the emergency response rule is: "If the fault type is valve fault and the confidence level is greater than 0.8, then immediately stop the machine for inspection." An example of the predictive maintenance rule is: "If the turbine speed is predicted to exceed the limit within 15 minutes and the prediction confidence is greater than 0.8, it is recommended to operate at reduced load." An example of the abnormal handling rule is: "If the main frequency amplitude drops by more than 30% and has a strong nonlinear negative correlation with fuel temperature, then adjust the fuel supply strategy and strengthen monitoring." The priority ranking logic is as follows: emergency response rules have the highest priority, followed by predictive maintenance rules, and finally anomaly handling rules.

[0054] Throughout the system's operation, the various agents communicate loosely via message queues (such as Kafka) or RESTful APIs, supporting independent deployment, upgrades, and replacements. For example, when a new sensor is added, only the parsing module of the data perception and preprocessing agent needs to be updated; when a new fault mode is added, only the sample library of the fault classification agent needs to be expanded and the classification head fine-tuned. This modular design gives the system excellent scalability and engineering feasibility, making it suitable for high-reliability scenarios such as aero-engine ground testing, online monitoring of ship propulsion systems, and remote maintenance of gas turbines. Through the aforementioned multi-agent collaborative mechanism, this invention achieves a closed-loop process from raw data perception to intelligent decision output, significantly improving the accuracy, real-time performance, interpretability, and system adaptability of engine fault early warning.

[0055] This embodiment also provides an engine fault early warning system based on multi-agent collaboration. The entire system consists of six core agents, namely, a data perception and preprocessing agent, a feature analysis agent, a correlation analysis agent, a time-series prediction agent, a fault classification agent, and an intelligent decision generation agent.

[0056] The data perception and preprocessing agent is mainly used to receive multi-source heterogeneous data from the engine, and to perform data quality assessment, missing value processing and time synchronization alignment, and output a standardized time series dataset.

[0057] The feature analysis agent is mainly used to extract waveform features from the standardized time-series data, and to construct and dynamically update feature thresholds based on historical data, so as to realize online detection and early warning of abnormal states.

[0058] The correlation analysis agent is mainly used to analyze the linear and nonlinear correlation between engine performance data and environmental condition data, and to verify and interpret the analysis results in conjunction with the engine mechanism model. The temporal prediction agent is based on a sequence-to-sequence deep learning model with an attention mechanism to predict the future state of key engine parameters. The fault classification agent is based on a pre-trained and fine-tuned model with a self-attention mechanism to identify and classify fault modes in engine status. The intelligent decision-making agent serves as the decision-making center of the system. It receives and integrates the output results of the feature analysis agent, the correlation analysis agent, the time series prediction agent, and the fault classification agent. Based on a preset multi-level decision rule base, it generates and outputs an operable maintenance decision scheme.

[0059] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0060] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for early warning of engine faults based on multi-agent cooperation, characterized in that, Includes the following steps: Step S1: Receive multi-source heterogeneous raw data from the engine, perform missing value imputation and time synchronization alignment processing, and output a standardized time series dataset with a unified time reference. Step S2: Extract waveform features from the standardized time-series dataset, and construct and dynamically update feature thresholds based on historical data to perform anomaly detection; Step S3: Analyze the linear and nonlinear correlations between engine performance data and environmental operating condition data, and verify and interpret the analysis results in conjunction with the engine mechanism model; Step S4: A sequence-to-sequence deep learning model based on an attention mechanism is used to predict the future state of key engine parameters in a temporal sequence. Step S5: Based on the self-attention mechanism, a pre-trained and fine-tuned model is used to identify and classify the engine status fault modes. Step S6: Receive and integrate the output results of steps S2 to S5, and generate and output an operable maintenance decision scheme based on the preset multi-level decision rule base.

2. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, The missing value filling process in step S1 includes: determining whether the engine is in a stable state based on the data from the engine speed sensor; if it is in a stable state, using linear interpolation to fill the missing data; if it is in an unstable state, calling a regression model trained based on historical data to perform data prediction interpolation.

3. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, The time synchronization and alignment process in step S1 includes: using a dynamic time warping algorithm to synchronize and align the multi-source heterogeneous original data, taking the data sequence with the highest sampling rate as the reference time sequence, and the remaining sequences as the sequences to be aligned, calculating the Euclidean distance matrix between the sequences to be aligned and the reference time sequence, and finding the path with the minimum cumulative distance, mapping all data to a unified time reference.

4. The engine fault early warning method based on multi-agent cooperation as described in claim 3, characterized in that, The time synchronization alignment process in step S1 includes: Step S11: Designate the data with the highest sampling rate as the baseline time series X, and designate the other data that need to be aligned as the sequence Y to be aligned; Step S12: Calculate the Euclidean distance between any two points in the reference time series X and the sequence Y to be aligned, forming the Euclidean distance matrix L; Step S13: In the Euclidean distance matrix L, find a regular path W that has the shortest cumulative distance from the starting point to the ending point; Step S14: Map the data points in the sequence Y to be aligned to the timestamps of the reference time series X according to the normalization path W. For the timestamps in the reference time series X that have no direct mapping points on the normalization path W, perform linear interpolation according to the mapping trend of the normalization path W to complete the time synchronization alignment.

5. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, Step S2 includes: Step S21: Using the sliding window method, the continuous time-series data stream is converted into data segments of fixed length, preserving the temporal locality characteristics; Step S22: Calculate the mean, standard deviation, peak-to-peak value, skewness, and kurtosis of each data segment as basic statistical characteristics; Step S23: Use the dynamic threshold anomaly detection method to construct and update the dynamic threshold for each statistical feature, compare the real-time statistical feature value with its corresponding current dynamic threshold, and make anomaly judgment.

6. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, Step S23 includes: Step S231, Threshold initialization: Collect historical normal data and calculate the initial statistical value of each basic statistical feature; Step S232: Update the threshold center value of each basic statistical feature using the exponentially weighted moving average algorithm; Step S233: Calculate each statistical feature based on the real-time data segment and compare it with the latest dynamic threshold range one by one. If the number of real-time statistical features that exceed the current dynamic threshold range exceeds the upper limit, the data segment is determined to be in an abnormal state and an early warning is triggered.

7. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, Step S3 includes: Step S31, Correlation Analysis: The Pearson correlation coefficient method is used to calculate the linear correlation strength between engine performance indicators and environmental operating condition parameters; Step S32: Calculate the monotonic nonlinear correlation strength between engine performance indicators and environmental operating condition parameters using the Spearman rank correlation coefficient method; Step S33: Based on the results of steps S31 and S32, generate a correlation coefficient matrix and visualize it through a heatmap to identify key influencing factors with significant linear effects. S34. Input key influencing factors into the physical mechanism model for simulation verification. Based on the simulation verification data, determine whether the correlation found conforms to known physical laws, construct a "data-mechanism" dual-drive analysis framework, and output structured analysis results.

8. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, Step S5 includes: Step S51: Based on historical engine test data, construct a sample library containing multiple known fault mode categories; standardize the feature data in the sample library so that the mean of each feature is 0 and the standard deviation is 1. Step S52: Based on a large public time series dataset of mechanical faults, pre-train a classification model with a Transformer encoder as its core; the pre-training is optimized using the cross-entropy loss function to enable the model to learn general fault feature representations; Step S53: Based on a small sample dataset of engine historical faults, freeze the Transformer encoder parameters pre-trained in step S52, train only the classification head parameters newly added to the model output layer, fine-tune the model, and obtain a fault classification model adapted to the target engine. Step S54: Input the real-time collected and preprocessed multi-source heterogeneous engine data into the fault classification model fine-tuned in step S53 for forward inference. The model outputs the probability distribution of each fault category through the Softmax function, selects the category with the highest probability value as the final fault classification result, and outputs the corresponding confidence level.

9. The engine fault early warning method based on multi-agent cooperation as described in claim 1, characterized in that, Step S6 includes: Step S61: Receive the structured output results of feature analysis, correlation analysis, time series prediction and fault classification corresponding to steps S2 to S5; perform time alignment and information integration on each structured output result to form a unified structured state vector containing multi-dimensional engine health status information; Step S62: Based on a preset multi-level decision rule base, match the structured state vector with the rules in the rule base; the multi-level decision rule base includes at least: The fault emergency rules layer directly triggers the corresponding emergency response plan based on the fault classification results. The predictive maintenance rules layer generates preventative maintenance recommendations based on time-series prediction results. The anomaly handling rules layer generates targeted handling solutions based on the anomaly patterns identified by feature analysis and correlation analysis; Step S63: Combine and prioritize the multiple candidate decision schemes generated by matching in step S62, and generate and output the final executable and maintainable decision scheme.

10. An engine fault early warning system based on multi-agent cooperation, characterized in that, include: The data sensing and preprocessing agent receives multi-source heterogeneous raw data from the engine, performs missing value filling and time synchronization alignment, and outputs a standardized time-series dataset with a unified time reference. The feature analysis agent extracts waveform features from the standardized time-series dataset and constructs and dynamically updates feature thresholds based on historical data to perform anomaly detection. The correlation analysis agent analyzes the linear and nonlinear correlations between engine performance data and environmental condition data, and verifies and interprets the analysis results in conjunction with the engine mechanism model. A time-series predictive agent, based on a sequence-to-sequence deep learning model with an attention mechanism, predicts the future state of key engine parameters in a time sequence. The fault classification agent, based on a pre-trained and fine-tuned model with a self-attention mechanism, performs fault mode recognition and classification of engine status. The intelligent decision generation agent receives and integrates the output results of the feature analysis agent, the correlation analysis agent, the time series prediction agent, and the fault classification agent, and generates and outputs an operable maintenance decision scheme based on a preset multi-level decision rule base.

Images