Diesel generator set health detection method and system based on prediction residual

Abstract

This application provides a method and system for health detection of diesel generator sets based on predictive residuals. The method includes: an edge processing layer receiving and preprocessing the operating data of the diesel generator set in real time; an intelligent prediction layer constructing a time-series window sequence from the operating data and inputting it into a deep time-series prediction model to generate predicted operating trajectories for multiple future steps; then comparing it with the actual operating data at the corresponding time to construct a residual sequence, and feeding the residual sequence back to the edge processing layer; the edge processing layer selecting an appropriate anomaly judgment strategy based on the current operating status of the unit and performing statistical anomaly judgment on the residual sequence; and the intelligent prediction layer performing quantitative assessment and trend prediction of the generator set's health status based on the long-term accumulated residual sequence. This application can keenly capture minute deviations between operating parameters and predicted trajectories, detect gradual degradation in advance during the fault symptom stage, and effectively improve detection accuracy through multi-level collaboration and state adaptive judgment.

Description

Technical Field

[0001] This application relates to the field of energy monitoring technology, and in particular to a method and system for health monitoring of diesel generator sets based on predictive residuals. Background Technology

[0002] With the development of distributed energy and microgrid systems, diesel generator sets remain irreplaceable as crucial equipment for independent power supply and peak shaving, especially in military, port, mining, hospital, and data center applications. Because diesel engines are typical internal combustion engine-generator coupled systems, their output voltage and frequency are affected by multiple factors, including governor gain, stability parameters, mechanical losses, fuel supply status, and load disturbances. Over long-term operation, these devices often experience performance degradation, increased output fluctuations, and potential faults. Therefore, real-time assessment and early prediction of the health status of diesel generators are of significant engineering importance.

[0003] Diesel generator sets are widely used in scenarios such as data center backup power, marine power plants, mines, and emergency power supplies. Their operational quality is typically characterized by power quality indicators such as effective terminal voltage and frequency. In engineering, ECUs (Electronic Control Units, generator controllers) are commonly used to collect and protect the unit's status. However, under complex operating conditions such as start-up, sudden loading, weak grid fluctuations, and mechanical aging, V / f (Voltage / Frequency) can exhibit significant non-steady-state fluctuations. Relying solely on fixed threshold alarms can easily lead to false alarms or missed alarms. Furthermore, unit degradation often exhibits characteristics of both gradual and sporadic events. Traditional scheduled maintenance or reactive repairs struggle to capture early signs of degradation in a timely manner, leading the industry to gradually shift towards a "online monitoring + predictive maintenance" approach. In recent years, the industry has seen the emergence of digital operation and maintenance platforms that combine remote monitoring, predictive analytics, and expert support, forming a productized path for real-time data monitoring, trend analysis, and maintenance recommendations. For example, Cummins' PrevenTech emphasizes combining remote monitoring, predictive analytics, and expert support to reduce unplanned downtime and identify potential problems in advance.

[0004] From both an academic and engineering perspective, current diesel generator condition monitoring technologies can be categorized into three types: 1. Model-driven approaches emphasize mechanisms and empirical rules. A typical approach involves establishing a performance indicator system, weights, and scoring methods to output a health score and combine it with expert knowledge to locate faults. For example, CN112197973A proposes normalizing and modeling the correlation of collected parameters, establishing a multi-layered indicator system with weighted scores, and simultaneously using expert systems and machine learning to train a fault identification model, forming a diagnostic framework that combines health scores with machine recognition. This type of method offers strong interpretability, but it has limited sensitivity to nonlinear coupling under complex operating conditions, early characteristics of weak faults, and the indicator system / weights rely on experience.

[0005] 2. Data-driven approach: This route has developed rapidly in recent years. Its core is to use deep learning to learn normal behavior from time series data, classifying outputs or predicting residual / reconstruction errors for anomaly identification. In the diesel generator field, existing research has proposed AIoT (Artificial Intelligence of Things) systems for industrial scenarios, employing hybrid networks such as CNN (Convolutional Neural Network)-LSTM (Long Short-Term Memory) to identify and assist in the maintenance of generators under abnormal operating conditions, demonstrating an end-to-end approach encompassing sensor acquisition, edge / platform inference, and operational decision-making. In the diesel engine field, unsupervised anomaly detection based on Transformer has also emerged, emphasizing the ability to model complex temporal correlations. The advantage of this approach is its ability to learn complex patterns using historical data, making it more sensitive to early anomalies; the disadvantages are high requirements for data quality, temporal alignment, and real-time deployment, and it is prone to false alarms during non-steady-state phases such as startup, requiring engineered gating and robust preprocessing.

[0006] 3. A hybrid approach, attempting a compromise between interpretability, feasibility, and real-time capability: data cleaning, filtering, state gating, and security interlocking are performed at the edge, while heavier model inference and long-term evaluation are conducted at the upper-level side. The promotion of Industrial IoT and edge computing provides the architectural foundation for this: industry standards and case studies generally emphasize the collaboration between the end, edge, and cloud, improving predictive maintenance rates and reducing network dependence through edge-side online monitoring and AI analysis. In engineering, industrial gateways / vehicle communication gateways often extract fast variables on the order of 0.1–10 seconds by parsing CAN (Controller Area Network) / J1939 (SAE J1939, the CAN communication protocol standard for commercial vehicles / heavy vehicles) messages for event identification and maintenance alarms, reflecting the need for low-latency monitoring and event identification.

[0007] However, there are generally three engineering shortcomings: (1) It is difficult to support sub-second closed loop. (2) The false alarm rate is relatively high. (3) The alarms are mostly limited to platform display or manual maintenance suggestions. In contrast, industrial products emphasize connectivity, remote monitoring and trend analysis. For example, Cat Connect-type remote asset monitoring solutions focus on providing data access and remote management capabilities for power generation assets, but their public information usually does not specify the details of the implementation of residual closed loop protection.

[0008] Therefore, the development of existing technologies exhibits a clear trend: on the one hand, deep learning continues to be applied in the anomaly identification and predictive maintenance of diesel generators / engines, with technologies such as CNN-LSTM, Transformer, and AIoT platforms; on the other hand, engineering systems place greater emphasis on real-time performance, data quality assurance, and security interlocks at the edge, as well as collaborative model updates and long-term health management on the cloud / IPC (Industrial Personal Computer) side. However, current technologies have the following drawbacks: (1) Existing diesel generator condition monitoring is mostly based on threshold alarms or empirical rules, which makes it difficult to depict the trend of degradation under multiple working conditions such as start-up / stop / load. It often only alarms after the fault has occurred or the fluctuation is significant, which is insufficient in advance.

[0009] (2) Traditional diagnostic methods based on generator mechanism modeling require accurate mechanical-electric coupling parameters and operating condition identification. Field parameter drift, load uncertainty and sensor noise can lead to model mismatch. At the same time, there is a lack of unified quantitative indicators for long-term fluctuations / residuals, which makes it difficult to support operation and maintenance decisions and aging trend assessment. Summary of the Invention

[0010] To address the aforementioned problems in the prior art, this application provides a method and system for health detection of diesel generator sets based on predictive residuals, which can detect gradual degradation in advance and improve the detection accuracy of the model.

[0011] To achieve the above objectives, the technical solution adopted in this application is as follows: Firstly, this application provides a health detection method for diesel generator sets based on predicted residuals, applied to a health detection system including an edge processing layer and an intelligent prediction layer, comprising: The edge processing layer receives and preprocesses the operating data of the diesel generator set in real time, and uploads the preprocessed operating data to the intelligent prediction layer. The intelligent prediction layer constructs the received running data into a time-series window sequence and inputs it into the deep time-series prediction model to generate predicted running trajectories for multiple future steps. The intelligent prediction layer compares the predicted running trajectory with the actual running data at the corresponding time, constructs a residual sequence, and feeds the residual sequence back to the edge processing layer; The edge processing layer receives the residual sequence and selects an appropriate anomaly determination strategy according to the current unit operating status. It performs statistical anomaly determination on the residual sequence and sends a trend-level protection control command to the generator unit when the anomaly conditions are met multiple times in a row. The intelligent prediction layer calculates the statistical characteristics of the residuals based on the long-term accumulated residual sequence and constructs a health index to quantitatively assess and predict the health status of the generator set.

[0012] The beneficial effects of this application are as follows: The edge processing layer preprocesses and uploads operational data in real time; the intelligent prediction layer uses a deep time-series prediction model to generate future multi-step trajectories and compares them with actual values ​​to construct a residual sequence; the edge processing layer adaptively determines statistical anomalies and triggers protection commands based on the unit's operating status; and the intelligent prediction layer constructs a health index based on long-term residuals. This scheme can keenly detect minute deviations between operating parameters and predicted trajectories, identify gradual degradation in the early stages of fault symptoms, and effectively improve detection accuracy through multi-level collaboration and adaptive state determination.

[0013] Optionally, if the unit operating state includes at least startup state, steady state, and transition state, then the step of selecting an appropriate anomaly determination strategy based on the current unit operating state includes: The current operating status of the unit is determined based on at least one of the following: the rate of change of the unit's operating parameters, communication quality indicators, or generator controller status bits. If it is in the startup state, the residual-based anomaly detection is frozen, and only hardware threshold protection is performed; If the state is steady, then residual-based anomaly detection is enabled, and normal detection parameters are used. If it is a transition state, residual-based anomaly detection is enabled, but with relaxed detection parameters, including at least one of increasing the anomaly detection threshold or increasing the threshold for consecutive triggers.

[0014] As described above, this application introduces a three-state gating mechanism of startup state, steady state, and transition state, which identifies the unit's operating status in real time based on the rate of change of operating parameters, communication quality indicators, or generator controller status bits, and adopts differentiated judgment strategies under different states. This solves the problem of high false alarm rate under non-steady-state conditions mentioned in the background technology and further improves the reliability of detection.

[0015] Optionally, the step of determining statistical anomalies in the residual sequence includes: Calculate the mean μ and standard deviation σ of the residual series within the sliding window, where... ; ; In the formula, W is the length of the sliding window, and e i This refers to the i-th residual value within the sliding window; If the current residual e satisfies |e-μ|>kσ, it is marked as an outlier, where k is a preset threshold coefficient that is adaptively adjusted according to the unit's operating status. The condition of satisfying the abnormal condition multiple times consecutively is: M consecutive residual values ​​are marked as abnormal points, where M is a preset integer greater than 1.

[0016] Optionally, uploading the preprocessed running data to the intelligent prediction layer includes: the edge processing layer adding time sequence identification information to each frame of uploaded running data, the time sequence identification information including an incrementing sequence number and a running timestamp; The intelligent prediction layer constructs a time-series window sequence from the received runtime data, including: The intelligent prediction layer receives the preprocessed running data, determines whether the incrementing sequence number is continuous, and whether the running timestamp falls within a preset jitter buffer window. If both are true, the received running data is constructed into a time-series window sequence; otherwise, the running data is discarded and communication degradation processing is triggered. The preset jitter buffer window is set according to a preset jitter threshold. The communication degradation processing is used to instruct the edge processing layer to freeze the residual-based anomaly determination and retain only hardware threshold protection.

[0017] As described above, this application establishes a timing guarantee mechanism between the edge processing layer and the intelligent prediction layer. This mechanism adds an incrementing sequence number and a local timestamp to the uploaded data. The intelligent prediction layer determines data validity based on the sequence number continuity and whether the timestamp falls within a preset jitter buffer window. Only data that meets the timing consistency requirement is used for prediction and residual construction; otherwise, communication degradation is triggered and residual-based anomaly detection is frozen. This mechanism ensures accurate alignment between the predicted trajectory and the measured data, avoiding false alarms caused by communication problems.

[0018] Optionally, the preprocessing includes: A filtering algorithm is used to calculate the position parameter and scale parameter within a local window of length w. It is then determined whether the current value deviates from the position parameter by more than the product of the threshold coefficient and the scale parameter. If so, the current value is replaced with the position parameter or the previous valid value. A low-pass filter is used to smooth the running data after anomaly processing.

[0019] Optionally, the filtering algorithm is a Hampel filter, the location parameter is the median m, and the scale parameter is the median absolute deviation MAD, the calculation formulas of which are respectively; m = median(x); MAD = median(|xm|); In the formula, x is the current value; The low-pass filter is a first-order infinite impulse response filter, and its recursive formula is: y[n] = αx[n] + (1-α)y[n-1]; In the formula, α is the smoothing coefficient, x[n] is the current input value, and y[n] is the current output value.

[0020] Optionally, the deep temporal prediction model is a temporal prediction network, whose input is a temporal window sequence of length L, and whose output is the predicted trajectory for the next H steps. The sampling period is Δ, and the prediction output is expressed as: .

[0021] Optionally, the time-series prediction network is a dual-branch hybrid network, comprising: The first approach uses a bidirectional long short-term memory network to encode the temporal window sequence forward and backward, extracting local dynamic features; The second branch performs a linear mapping of the temporal window sequence and superimposes position encoding before inputting it into the Transformer encoder, which extracts globally relevant features through a multi-head self-attention mechanism. The feature fusion layer concatenates the output features of the two branches, maps them through a fully connected layer and a non-linear activation function, and generates a predicted trajectory for the next multiple steps.

[0022] As described above, this application employs a dual-branch hybrid temporal prediction network. The first branch extracts local dynamic features, while the second branch extracts globally relevant features through a multi-head self-attention mechanism. The resulting fusion outputs a multi-step predicted trajectory. This structure combines sensitivity to short-term inertial changes with the ability to model long-range coupling of multiple variables, improving the stability and generalization ability of multi-step prediction. Without relying on training with enumerated working condition data, it provides a more accurate baseline trajectory for residual anomaly determination, thereby enhancing the model's detection accuracy for early degradation features.

[0023] Optionally, the calculation of the statistical characteristics of the residuals and the construction of a health index to quantitatively assess and predict the trend of the generator set's health status includes: In the long time window T L Calculate the standard deviation σ of the voltage residuals respectively. V (t) and the standard deviation of the frequency residual σ f (t), and divided by the pre-established baseline standard deviation σ V ′ and σ f ′, thus obtaining the normalized fluctuation ratio R V (t) and R f (t): R V (t)=σ V (t) / σ V ′; R f (t)=σ f (t) / σ f ′; Count the number of outliers N within the same window abn Given the total number of samples N, calculate the anomaly density D(t): D(t) = N abn / N; The overall degradation S(t) is obtained by weighted summation of voltage fluctuation ratio, frequency fluctuation ratio, and anomaly density: S(t) = (R V (t)-1)+(R f (t)-1)+γD(t); In the formula, γ is a preset weighting coefficient; The overall degradation is converted into a health index HI(t) using an exponential mapping function: HI(t)=min(100,max(0,100·exp(-λ·S(t)))); In the formula, λ is the sensitivity coefficient; In multiple long-term time windows T L The combined trend time window T H Calculate the rate of change k of the health index HI (t): k HI (t)=(HI(t)-HI(tT H )) / T H ; When the health index is lower than a preset maintenance threshold, or when the rate of change of the health index exceeds a preset rate threshold, the restriction strategy is adjusted.

[0024] Secondly, this application provides a diesel generator set health monitoring system based on predicted residuals, including: The edge processing layer is used to receive and preprocess the operating data of the diesel generator set in real time, and upload the preprocessed operating data to the intelligent prediction layer; it is used to receive the residual sequence, select an appropriate anomaly judgment strategy according to the current unit operating status, perform statistical anomaly judgment on the residual sequence, and issue a trend-level protection control command to the generator set when the anomaly conditions are met multiple times in a row. The intelligent prediction layer is used to construct a time-series window sequence from the received operating data and input it into a deep time-series prediction model to generate predicted operating trajectories for multiple future steps; it is used to compare the predicted operating trajectory with the actual operating data at the corresponding time to construct a residual sequence and feed the residual sequence back to the edge processing layer; it is used to calculate the statistical characteristics of the residuals and construct a health index based on the long-term accumulated residual sequence to quantitatively assess and predict the trend of the generator set's health status.

[0025] The second aspect provides a diesel generator set health detection system based on predicted residuals, referring to the relevant description of the diesel generator set health detection method based on predicted residuals provided in the first aspect. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the main process of the diesel generator set health detection method based on predicted residuals according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the diesel generator set health monitoring system based on predictive residuals according to an embodiment of this application. Figure 3 This is a schematic diagram of the overall process of the diesel generator set health detection method based on predicted residuals according to an embodiment of this application; Figure 4 This is a schematic diagram illustrating the processing of the deep temporal prediction model involved in the embodiments of this application; Figure 5 This is a gating logic diagram of the unit operating status involved in the embodiments of this application. Detailed Implementation

[0027] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0028] The embodiments of this application are applied to scenarios requiring health monitoring of diesel generator sets, such as military applications, ports, mines, hospitals, and data centers. Existing technologies often only issue alarms after a fault has occurred or significant fluctuations have taken place, and their model detection accuracy is insufficient.

[0029] Therefore, in various embodiments of this application, a health detection system comprising an edge processing layer and an intelligent prediction layer is applied. The edge processing layer receives and preprocesses the operating data of the diesel generator set in real time, and uploads the preprocessed operating data to the intelligent prediction layer. The intelligent prediction layer constructs a time-series window sequence from the received operating data and inputs it into a deep time-series prediction model to generate predicted operating trajectories for multiple future steps. The intelligent prediction layer compares the predicted operating trajectory with the actual operating data at the corresponding time, constructs a residual sequence, and feeds the residual sequence back to the edge processing layer. The edge processing layer receives the residual sequence and selects an appropriate anomaly judgment strategy based on the current operating status of the generator set, performs statistical anomaly judgment on the residual sequence, and issues a trend-level protection control command to the generator set when the anomaly conditions are met multiple times consecutively. Based on the long-term accumulated residual sequence, the intelligent prediction layer calculates the statistical characteristics of the residuals and constructs a health index to quantitatively assess and predict the health status of the generator set. Thus, by capturing small deviations between operating parameters and predicted trajectories, gradual degradation can be detected in advance during the fault symptom stage. At the same time, through multi-level collaboration and state adaptive judgment, the detection accuracy is effectively improved.

[0030] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0031] This application provides a health detection method for diesel generator sets based on predicted residuals, applied to a health detection system including an edge processing layer and an intelligent prediction layer, such as... Figure 2 As shown, the health monitoring system also includes a power generation unit layer. These layers form a bidirectional data channel via an industrial communication interface, thus constructing an integrated system architecture for real-time monitoring, predictive analysis, and closed-loop control. This system uses the real-time operating status of the diesel generator set as the monitoring object. Through the synergy of the real-time processing capabilities of the edge processing layer and the model prediction capabilities of the intelligent prediction layer, it achieves trend identification of diesel generator operating anomalies and a comprehensive assessment of operational health. Specifically: (1) Description of the generator unit layer structure: The diesel generator set is the power source of the system and is used to output electrical energy; the sensor group is set in the key parts of the diesel generator to collect operating parameters such as voltage, current, frequency, and torque, and output analog signals to the generator controller (ECU); the generator controller collects, organizes and encapsulates the sensor signals in a timing manner, converts them into standardized digital communication data frames, and sends them to the edge processing layer through a serial communication interface. The communication method can be RS485, CAN or industrial Ethernet, etc., and the serial interface can be UART (Universal Asynchronous Receiver / Transmitter) / RS485 (Recommended Standard 485) / CAN, etc. The communication rate is configurable, for example, the baud rate of UART is 9600 bit / s.

[0032] This layer is used to complete the basic signal acquisition function of the diesel generator.

[0033] (2) Edge processing layer structure description: The microcontroller is the core of local real-time computing and processing. In this embodiment, the STM32 high-performance series microcontroller is selected.

[0034] (3) Description of the intelligent prediction layer structure: The industrial control computer (IPC) receives real-time data from the edge processing layer to complete in-depth predictive analysis and long-term health assessment. The cloud server is optional and can be deployed using public cloud, private cloud or cloud architecture.

[0035] Therefore, as Figure 1 and Figure 3 As shown, the method includes: Step S1: The edge processing layer receives and preprocesses the operating data of the diesel generator set in real time, and uploads the preprocessed operating data to the intelligent prediction layer.

[0036] In this embodiment, the preprocessing in step S1 includes: Step S11: Using a filtering algorithm, calculate the position parameter and scale parameter within a local window of length w, and determine whether the current value deviates from the product of the position parameter and the scale parameter. If so, replace the current value with the position parameter or the previous valid value.

[0037] The filtering algorithm is the Hampel filter, the location parameter is the median m, and the scale parameter is the median absolute deviation MAD. The calculation formulas are as follows: m = median(x); MAD = median(|xm|); In the formula, x is the current value.

[0038] Therefore, the edge processing layer eliminates occasional bad pixels by performing Hampel outlier removal, thereby suppressing communication spikes.

[0039] In other embodiments, Hampel filtering can also be replaced by median filtering, quantile filtering, etc.

[0040] Step S12: Use a low-pass filter to smooth the running data after anomaly processing.

[0041] The low-pass filter is a first-order infinite impulse response filter, and its recursive formula is: y[n] = αx[n] + (1-α)y[n-1]; In the formula, α is the smoothing coefficient, x[n] is the current input value, and y[n] is the current output value.

[0042] Therefore, this embodiment performs IIR (Infinite Impulse Response) digital filtering on the input sequence to reduce high-frequency noise without significantly weakening the main trend of the startup process, thereby suppressing interference such as communication jitter and measurement noise.

[0043] In one embodiment, uploading the preprocessed runtime data to the intelligent prediction layer includes: the edge processing layer adding time sequence identification information to each frame of uploaded runtime data, the time sequence identification information including an incrementing sequence number and a runtime timestamp.

[0044] Specifically, to avoid relying on ECU protocol modifications, it is preferable to include an incrementing sequence number Seq and a source timestamp T in each frame of runtime data. rx It is used for packet loss detection, latency / jitter estimation, and cross-end timing alignment. Specifically, when the sampling period Δ < 500ms, the incrementing sequence number Seq increases by 1 per period and supports rollback.

[0045] Step S2: The intelligent prediction layer constructs the received running data into a time-series window sequence and inputs it into the deep time-series prediction model to generate the predicted running trajectory for multiple future steps.

[0046] In one embodiment, step S2, in which the intelligent prediction layer constructs a time-series window sequence from the received runtime data, includes: The intelligent prediction layer receives the preprocessed running data, determines whether the incrementing sequence number is continuous and whether the running timestamp falls within the preset jitter buffer window. If both are true, the received running data is constructed into a time-series window sequence; otherwise, the running data is discarded and communication degradation processing is triggered. The preset jitter buffer window is set according to the preset jitter threshold. Communication degradation processing is used to instruct the edge processing layer to freeze the residual-based anomaly judgment and retain only hardware threshold protection.

[0047] Specifically, to support sub-second closed-loop operation with Δ < 500ms, this application establishes a timing guarantee between the STM32 and IPC using a time base combined with jitter buffering: the STM32 generates a message data frame record for each sampling point, including sequence number, timestamp, and data, while simultaneously counting the number of packet losses N. lost One-way delay τ and jitter J: When packet loss occurs or the delay exceeds the watchdog timeout, the edge processing layer enters degradation protection and triggers interlocking restrictions, thereby performing local fast protection and real-time preprocessing of the collected data. The IPC side is aligned with a fixed sampling frequency, and J is set in the circular buffer. max Jitter buffer, only when the sample timestamp falls within [tJ max ,t+J max Only enter the window when the time is right; otherwise, freeze the trend alarm count and enter the communication degradation strategy, retaining only the hard threshold / communication watchdog interlock.

[0048] In this embodiment, the deep temporal prediction model is a temporal prediction network. Its input is a temporal window sequence of length L, which includes the effective value of voltage, frequency, and can be expanded to include fast variables such as torque / speed and AVR gain. The output is the predicted trajectory for the next H steps, with a sampling period of Δ. The prediction output is expressed as: .

[0049] Specifically, the deep temporal prediction model is DG-Net (Diesel Generator-Net, a hybrid temporal prediction network for diesel generators). For example... Figure 4 As shown, the time series prediction network is a dual-branch hybrid network, including: (1) The first branch uses a bidirectional long short-term memory network to encode the temporal window sequence in both forward and backward directions to extract local dynamic features.

[0050] Specifically, Bi-LSTM is used to encode the time window sequence in both forward and backward directions to extract local dynamic features of short-term inertial changes such as startup recovery and voltage regulation stabilization, making it more sensitive to short-term inertial changes such as startup / recovery.

[0051] (2) The second branch performs linear mapping on the temporal window sequence and superimposes position encoding, then inputs it to the Transformer encoder to extract global relevant features through a multi-head self-attention mechanism.

[0052] Specifically, the input temporal window sequence is linearly mapped to obtain an embedding vector, which is then superimposed with positional encoding. Subsequently, a Transformer encoder block consisting of a multi-head self-attention network and a feedforward network is used to extract global correlation and cross-variable coupling features, making it more sensitive to multivariable coupling and long correlations.

[0053] (3) Feature fusion layer: The output features of the two branches are spliced ​​together and mapped through a fully connected layer and a nonlinear activation function to generate the predicted running trajectory for multiple future steps.

[0054] Specifically, the output feature vectors of the two are concatenated (Concat), and then subjected to fully connected mapping (Linear) and nonlinear activation (ReLU) combined with Dropout to suppress overfitting, to obtain the fused feature output predicting the trajectory of the future time domain.

[0055] Therefore, the intelligent prediction layer IPC sets the STM32's running data as a time-series window sequence as model input, and calls the hybrid deep learning prediction model to output short-term voltage and frequency trajectory predictions for the next H steps in a rolling manner. , And integrate them into a two-dimensional vector. The relationship is as follows: ; To align with the goal of trend prediction and early warning, a multi-step prediction sequence is preferred over a single-point prediction; that is, the output sequence is: .

[0056] in, The prediction step size is consistent with the sampling period, and H is the prediction time domain length. Considering the typical dynamic time scale of diesel generator voltage / frequency regulation, and taking into account the preprocessing, communication transmission, and IPC inference latency on the STM32 side, the parameter ranges are as follows: Sampling period =100~500ms, input window length L=5~20 points, prediction steps H=4~10 steps. The prediction adopts a rolling update method: the prediction for the next H steps is updated every time a new sampling point is reached, ensuring continuous coverage of the time axis without gaps. When multiple interval prediction results overlap, the strategy of using the latest prediction to overwrite the old prediction is preferred to ensure responsiveness to the latest observations. In practical engineering, the response speed of unit speed / voltage regulation and network jitter J can be considered. max and action requirements Adjust L and H.

[0057] In other embodiments, the short-term forecasting model can be replaced with time series models such as TCN, Informer, PatchTST, or a Transformer-only / LSTM-only structure can be used, but the closed-loop framework of forecasting, residuals, marginal trend interlocking and long-term HI (Health Index) can still be maintained.

[0058] Step S3: The intelligent prediction layer compares the predicted running trajectory with the actual running data at the corresponding time, constructs a residual sequence, and feeds the residual sequence back to the edge processing layer.

[0059] Therefore, when the actual measured value Y(t) arrives over time, IPC aligns the predicted trajectory with the measured trajectory at the corresponding time according to the timestamp / serial number, constructing a residual sequence of the difference between the predicted and actual values: ; Afterwards, the IPC sends the residual sequence back to the STM32 in the form of residual return frames containing sequence number, timestamp and data, and writes the residuals to the historical residual log library to provide a data basis for long-term health assessment.

[0060] Step S4: The edge processing layer receives the residual sequence and selects an appropriate anomaly judgment strategy according to the current unit operating status. It performs statistical anomaly judgment on the residual sequence. When the anomaly conditions are met multiple times in a row, it issues a trend-level protection control command to the generator unit.

[0061] In one embodiment, to balance false alarm suppression and fault detectability under strong non-steady-state operating conditions such as startup, sudden load grid connection / de-station, etc., such as Figure 5 As shown, this application defines the system's unit operating state as three states, namely: (1) Startup state S start In the initial stage of startup or before the waveform has converged, only hard threshold / communication watchdog interlock is executed, and the residual 3σ count is frozen.

[0062] (2) Steady-state S steady The condition is met that the rate of change falls below a threshold and remains stable for a period T. stable Then, DG-Net residuals are allowed to participate in the decision-making process.

[0063] (3) Transition state S trans : Triggered by events such as |dV / dt| or |df / dt| exceeding the threshold, abnormal Seq interval, timeout delay, or ECU status bit change, maintaining the transition period T hold During this period, the threshold coefficient is increased or a higher 3σ outlier counting threshold is adopted to avoid false triggering by short-term fluctuations; when the rate of change falls back and continues for T_back, it exits to a steady state. The transition period T is... hold It lasts for 2-10 seconds.

[0064] Therefore, as Figure 5 As shown, step S4 involves selecting an appropriate anomaly detection strategy based on the current unit operating status, including: The current operating status of the unit is determined based on at least one of the following: the rate of change of the unit's operating parameters, communication quality indicators, or generator controller status bits. If it is in the startup state, the residual-based anomaly detection is frozen, and only hardware threshold protection is performed; If the state is steady, then residual-based anomaly detection is enabled, and normal detection parameters are used. If it is a transition state, residual-based anomaly detection is enabled, but with relaxed detection parameters, including at least one of increasing the anomaly detection threshold or increasing the threshold for consecutive triggers.

[0065] The aforementioned gating is performed locally and in real time on the STM32, without requiring additional operating condition labels to be provided to the IPC as model input. The processed data is encapsulated into CAN / RS485 messages by the STM32 according to the communication protocol and uploaded to the IPC of the intelligent prediction layer in real time. The windowed input is automatically constructed by the IPC at the receiving end through a circular buffer, eliminating the need for full window processing by the STM32 and thus reducing the communication burden.

[0066] In this process, after receiving the residual data, the STM32 first verifies whether the Seq continuity and time delay τ satisfy the closed-loop constraint. If a timeout occurs or the residual is invalid, the trend determination is downgraded, and only the hard threshold / interlock protection is retained. Subsequently, a 3σ anomaly detection is performed on the residual sequence, online estimation of the residual distribution is performed based on a sliding statistical window, and statistical anomaly detection of the residual sequence is performed in step S4, including: Calculate the mean μ and standard deviation σ of the residual series within the sliding window, where... ; ; In the formula, W is the length of the sliding window, and e i This represents the i-th residual value within the sliding window; If the current residual e satisfies |e-μ|>kσ, it is marked as an outlier, where k is a preset threshold coefficient that is adaptively adjusted according to the unit's operating status. The condition for an anomaly to be met multiple times consecutively is: M consecutive residual values ​​are marked as an anomaly, where M is a preset integer greater than 1.

[0067] Where k is 3, the trend-level protection action is triggered only when the 3σ abnormal condition is met M times consecutively. Its trend-level protection action employs a deterministic interlocking mechanism: the STM32 outputs alarm, load reduction, and shutdown commands according to preset priorities, and confirms that the ECU has received and executed them via heartbeat / response frames; simultaneously, transient faults such as overvoltage / undervoltage, overfrequency / underfrequency, emergency stop, and grid connection failure are handled in parallel by the ECU or STM32's hard threshold and interlocking mechanism, thus forming a dual-layer safety detection system of transient-level hard protection and predictive residual trend protection.

[0068] Step S5: The intelligent prediction layer calculates the statistical characteristics of the residuals based on the long-term accumulated residual sequence and constructs a health index to quantitatively assess and predict the health status of the generator set.

[0069] In this embodiment, the IPC or optional cloud server uses a statistical window T. L and with an update cycle TU Rolling updates. At each update time t, the system retrieves the most recent time T. L residual samples e(tT) within the range L ), ..., e(t), calculate the mean residuals μ of voltage and frequency respectively. L With standard deviation σ L The former characterizes whether there is a systematic bias in the residuals, while the latter characterizes the intensity of residual fluctuations and serves as a primary indicator for assessing health. Their calculation formulas are the same as before. In this embodiment, the statistical window T... L The preferred interval is 10 minutes, with an update cycle of T. U The optimal time for rolling updates is 1 minute.

[0070] Therefore, step S5 calculates the statistical characteristics of the residuals and constructs a health index to quantitatively assess and predict the trend of the generator set's health status, including: Step S51, within the long-term time window T L Calculate the standard deviation σ of the voltage residuals respectively. V (t) and the standard deviation of the frequency residual σ f (t), and divided by the pre-established baseline standard deviation σ V ′ and σ f ′, thus obtaining the normalized fluctuation ratio R V (t) and R f (t): R V (t)=σ V (t) / σ V ′; R f (t)=σ f (t) / σ f ′.

[0071] When the unit's operating status is close to the baseline value, R V (t)≈1、R f (t)≈1; When the unit degrades or the abnormal trend intensifies, R V (t)>1、R f If (t)>1, both will increase.

[0072] Step S52: Count the number of outliers N within the same window. abn Given the total number of samples N, calculate the anomaly density D(t): D(t) = N abn / N.

[0073] The function takes the value 1 when the short-term residual is judged to be abnormal at time t; otherwise, it takes the value 0. Step S53: Weight the voltage fluctuation ratio, frequency fluctuation ratio, and anomaly density to obtain the comprehensive degradation S(t): S(t) = (R V (t)-1)+(R f (t)-1)+γD(t); In the formula, γ is a preset weighting coefficient.

[0074] Step S54: Convert the overall degradation amount into a health index HI(t) using an exponential mapping function: HI(t)=min(100,max(0,100·exp(-λ·S(t)))); In the formula, λ is the sensitivity coefficient.

[0075] In this embodiment, to ensure the stability of the output range, the system performs amplitude limiting processing on the calculation results, with a maximum value of 100. This can be achieved through comparison operations to avoid the health output exceeding the limit under abnormal circumstances.

[0076] Step S55: In multiple long-term time windows T L The combined trend time window T H Calculate the rate of change k of the health index HI (t): k HI (t)=(HI(t)-HI(tT H )) / T H .

[0077] Step S56: When the health index is lower than the preset maintenance threshold, or the rate of change of the health index exceeds the preset rate threshold, adjust the restriction strategy.

[0078] The adjustment of restriction policies includes system output maintenance warnings, maintenance prompts, and so on.

[0079] In summary, this application achieves improved accuracy in the early detection and assessment of gradual degradation of diesel generator sets by constructing a multi-level closed-loop framework that coordinates the edge processing layer and the intelligent prediction layer. Specific beneficial effects are as follows: 1. The edge processing layer preprocesses and uploads operational data in real time. The intelligent prediction layer uses a deep time-series prediction model to generate future multi-step trajectories and compares them with actual values ​​to construct residual sequences. The edge processing layer adaptively determines statistical anomalies and triggers protection commands based on the unit's operating status. The intelligent prediction layer constructs a health index based on long-term residuals. This multi-level collaborative architecture can keenly capture minute deviations between operating parameters and predicted trajectories, detect gradual degradation in advance during the fault symptom stage, and effectively improve detection accuracy through state adaptive judgment, overcoming the shortcomings of insufficient lead time in traditional threshold alarm methods.

[0080] 2. By introducing a three-state gating mechanism (start-up, steady-state, and transition state), the unit's operating status is identified in real time based on the rate of change of operating parameters, communication quality indicators, or generator controller status bits. Differentiated judgment strategies are employed for each state: in the start-up state, residual-based anomaly judgment is frozen, and only hardware threshold protection is implemented; in the transition state, judgment parameters are relaxed; and in the steady-state state, statistical anomaly judgment is based on residuals. This strategy effectively suppresses false alarms under strong non-steady-state conditions such as start-up and sudden loading, while maintaining sensitivity to early degradation trends in the steady state. This solves the problem of high false alarm rates under non-steady-state conditions and further improves detection reliability.

[0081] 3. By establishing a time-series guarantee mechanism between the edge processing layer and the intelligent prediction layer, incremental sequence numbers and local timestamps are added to uploaded data. The intelligent prediction layer determines data validity based on the continuity of the sequence numbers and whether the timestamp falls within a preset jitter buffer window. Only data that meets the time-series consistency requirements is used for prediction and residual construction; otherwise, communication degradation processing is triggered and residual-based anomaly detection is frozen. This mechanism ensures accurate alignment between the predicted trajectory and the measured data, avoiding false alarms due to excessive residuals caused by communication packet loss and latency jitter, providing engineering assurance for the reliable application of deep learning models in real industrial environments.

[0082] 4. A dual-branch hybrid temporal prediction network, DG-Net, is adopted. The first branch uses a bidirectional long short-term memory network to extract local dynamic features, while the second branch uses a Transformer encoder with a multi-head self-attention mechanism to extract globally relevant features. The resulting fusion outputs the predicted trajectory for future multi-step predictions. This structure combines sensitivity to short-term inertial changes with the ability to model long-range coupling of multiple variables, improving the stability and generalization ability of multi-step predictions. Without relying on training with enumerated working condition data, it provides a more accurate baseline trajectory for residual anomaly determination, thereby enhancing the model's detection accuracy for early degradation features.

[0083] 5. A health index is constructed based on long-term accumulated residual sequences. By calculating the normalized fluctuation ratio and anomaly density of voltage and frequency residuals, a comprehensive degradation amount is obtained and mapped to a health index of 0-100. Simultaneously, the rate of change of the health index is calculated on a trend window. This quantitative assessment method can intuitively reflect the evolution trend of the unit's health status. When the health index falls below the maintenance threshold or the rate of decline is too rapid, a maintenance warning is output, providing a quantifiable basis for predictive maintenance decisions and overcoming the shortcomings of traditional methods that lack long-term quantitative indicators.

[0084] In one embodiment, this application also provides a diesel generator set health monitoring system based on predicted residuals, comprising: The edge processing layer is used to receive and preprocess the operating data of the diesel generator set in real time, and upload the preprocessed operating data to the intelligent prediction layer; it is used to receive the residual sequence, select the appropriate anomaly judgment strategy according to the current unit operating status, perform statistical anomaly judgment on the residual sequence, and issue trend-level protection control commands to the generator set when the anomaly conditions are met multiple times in a row. The intelligent prediction layer is used to construct a time-series window sequence from the received operating data and input it into the deep time-series prediction model to generate predicted operating trajectories for multiple future steps; it is used to compare the predicted operating trajectory with the actual operating data at the corresponding time to construct a residual sequence and feed the residual sequence back to the edge processing layer; it is used to calculate the statistical characteristics of the residuals and construct a health index based on the long-term accumulated residual sequence to quantitatively assess and predict the trend of generator unit health status.

[0085] This embodiment of the diesel generator set health monitoring system based on predictive residuals only shows a block diagram of a portion of the structure related to the present application solution, and does not constitute a limitation on the device on which the present application solution is applied. Specific devices may include more or fewer components than shown in the figures, or combinations of certain components, or different component arrangements, such as power supplies, input / output interfaces, etc. Furthermore, the diesel generator set health monitoring system based on predictive residuals of this embodiment can operate on an operating system stored in memory, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0086] Furthermore, the specific descriptions of the technical effects and steps of the diesel generator set health detection system based on prediction residuals in the above embodiments are all based on the relevant descriptions of the embodiments of the diesel generator set health detection method based on prediction residuals.

[0087] Since the systems / devices described in the above embodiments of this application are systems / devices used to implement the methods of the above embodiments of this application, those skilled in the art can understand the specific structure and modifications of the system / devices based on the methods described in the above embodiments of this application, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of this application fall within the scope of protection of this application.

[0088] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0090] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0091] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0092] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0093] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, then this application should also include these modifications and variations.

Claims

1. A health monitoring method for diesel generator sets based on predictive residuals, applied to a health monitoring system including an edge processing layer and an intelligent prediction layer, characterized in that, include: The edge processing layer receives and preprocesses the operating data of the diesel generator set in real time, and uploads the preprocessed operating data to the intelligent prediction layer. The intelligent prediction layer constructs the received running data into a time-series window sequence and inputs it into the deep time-series prediction model to generate predicted running trajectories for multiple future steps. The intelligent prediction layer compares the predicted running trajectory with the actual running data at the corresponding time, constructs a residual sequence, and feeds the residual sequence back to the edge processing layer; The edge processing layer receives the residual sequence and selects an appropriate anomaly determination strategy according to the current unit operating status. It performs statistical anomaly determination on the residual sequence and sends a trend-level protection control command to the generator unit when the anomaly conditions are met multiple times in a row. The intelligent prediction layer calculates the statistical characteristics of the residuals based on the long-term accumulated residual sequence and constructs a health index to quantitatively assess and predict the health status of the generator set.

2. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 1, characterized in that, The unit operating states include at least startup state, steady state, and transition state. Therefore, the step of selecting an appropriate anomaly detection strategy based on the current unit operating state includes: The current operating status of the unit is determined based on at least one of the following: the rate of change of the unit's operating parameters, communication quality indicators, or generator controller status bits. If it is in the startup state, the residual-based anomaly detection is frozen, and only hardware threshold protection is performed; If the state is steady, then residual-based anomaly detection is enabled, and normal detection parameters are used. If it is a transition state, residual-based anomaly detection is enabled, but with relaxed detection parameters, including at least one of increasing the anomaly detection threshold or increasing the threshold for consecutive triggers.

3. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 2, characterized in that, The statistical anomaly determination of the residual sequence includes: Calculate the mean μ and standard deviation σ of the residual series within the sliding window, where... ； ； In the formula, W is the length of the sliding window, and e i This refers to the i-th residual value within the sliding window; If the current residual e satisfies |e-μ|>kσ, it is marked as an outlier, where k is a preset threshold coefficient that is adaptively adjusted according to the unit's operating status. The condition of satisfying the abnormal condition multiple times consecutively is: M consecutive residual values ​​are marked as abnormal points, where M is a preset integer greater than 1.

4. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 1, characterized in that, Uploading the preprocessed running data to the intelligent prediction layer includes: the edge processing layer adding time sequence identification information to each frame of uploaded running data, the time sequence identification information including an incrementing sequence number and a running timestamp; The intelligent prediction layer constructs a time-series window sequence from the received runtime data, including: The intelligent prediction layer receives the preprocessed running data, determines whether the incrementing sequence number is continuous, and whether the running timestamp falls within a preset jitter buffer window. If both are true, the received running data is constructed into a time-series window sequence; otherwise, the running data is discarded and communication degradation processing is triggered. The preset jitter buffer window is set according to a preset jitter threshold. The communication degradation processing is used to instruct the edge processing layer to freeze the residual-based anomaly determination and retain only hardware threshold protection.

5. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 1, characterized in that, The preprocessing includes: A filtering algorithm is used to calculate the position parameter and scale parameter within a local window of length w. It is then determined whether the current value deviates from the position parameter by more than the product of the threshold coefficient and the scale parameter. If so, the current value is replaced with the position parameter or the previous valid value. A low-pass filter is used to smooth the running data after anomaly processing.

6. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 5, characterized in that, The filtering algorithm is the Hampel filter, the location parameter is the median m, the scale parameter is the median absolute deviation MAD, and their calculation formulas are as follows: m = median(x); MAD = median(|xm|); In the formula, x is the current value; The low-pass filter is a first-order infinite impulse response filter, and its recursive formula is: y[n] = αx[n] + (1-α)y[n-1]; In the formula, α is the smoothing coefficient, x[n] is the current input value, and y[n] is the current output value.

7. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 1, characterized in that, The deep temporal prediction model is a temporal prediction network. Its input is a temporal window sequence of length L, and its output is the predicted trajectory for the next H steps. The sampling period is Δ, and the prediction output is expressed as: .

8. The method for health monitoring of diesel generator sets based on predicted residuals according to claim 7, characterized in that, The time-series prediction network is a dual-branch hybrid network, comprising: The first approach uses a bidirectional long short-term memory network to encode the temporal window sequence forward and backward, extracting local dynamic features; The second branch performs a linear mapping of the temporal window sequence and superimposes position encoding before inputting it into the Transformer encoder, which extracts globally relevant features through a multi-head self-attention mechanism. The feature fusion layer concatenates the output features of the two branches, maps them through a fully connected layer and a non-linear activation function, and generates a predicted trajectory for the next multiple steps.

9. The method for health monitoring of diesel generator sets based on predicted residuals according to any one of claims 1 to 8, characterized in that, The calculation of the statistical characteristics of the residuals and the construction of a health index are used to quantitatively assess and predict the trend of the generator set's health status, including: In the long time window T L Calculate the standard deviation σ of the voltage residuals respectively. V (t) and the standard deviation of the frequency residual σ f (t), and divided by the pre-established baseline standard deviation σ V ′ and σ f ′, thus obtaining the normalized fluctuation ratio R V (t) and R f (t): R V (t)=σ V (t) / s V ′; R f (t)=σ f (t) / s f ′; Count the number of outliers N within the same window abn Given the total number of samples N, calculate the anomaly density D(t): D(t)=N abn / N; The overall degradation S(t) is obtained by weighted summation of voltage fluctuation ratio, frequency fluctuation ratio, and anomaly density: S(t)=(R V (t)-1)+(R f (t)-1)+γD(t); In the formula, γ is a preset weighting coefficient; The overall degradation is converted into a health index HI(t) using an exponential mapping function: HI(t)=min(100,max(0,100·exp(-λ·S(t)))); In the formula, λ is the sensitivity coefficient; In multiple long-term time windows T L The combined trend time window T H Calculate the rate of change k of the health index HI (t): k HI (t)=(HI(t)-HI(t-T H )) / T H ; When the health index is lower than a preset maintenance threshold, or when the rate of change of the health index exceeds a preset rate threshold, the restriction strategy is adjusted.

10. A diesel generator set health monitoring system based on predicted residuals, characterized in that, include: The edge processing layer is used to receive and preprocess the operating data of the diesel generator set in real time, and upload the preprocessed operating data to the intelligent prediction layer. Used to receive the residual sequence, select an appropriate anomaly determination strategy according to the current unit operating status, perform statistical anomaly determination on the residual sequence, and issue a trend-level protection control command to the generator unit when the anomaly conditions are met multiple times in a row. The intelligent prediction layer is used to construct a time-series window sequence from the received running data and input it into the deep time-series prediction model to generate predicted running trajectories for multiple future steps; it is used to compare the predicted running trajectory with the actual running data at the corresponding time, construct a residual sequence, and feed the residual sequence back to the edge processing layer; This is used to calculate the statistical characteristics of residuals based on long-term accumulated residual sequences and construct a health index to quantitatively assess and predict the health status of generator sets.

Images