Abnormal oil consumption tracing system and method based on engine multi-sensor time sequence signals

Abstract

The application discloses an oil consumption anomaly tracing system and method based on engine multi-sensor time sequence signals, and the system comprises the following modules: a signal acquisition module, which acquires original time sequence signals of multiple sensors; a time sequence synchronization alignment module, which generates multi-channel time sequence data after synchronization alignment; a data binding module, which obtains working condition data from a vehicle bus and forms time sequence samples with working condition labels by slicing in fixed time windows; a feature extraction module, which extracts preset time domain features and frequency domain features and constructs a fusion feature vector; a diagnosis model module, which takes the fusion feature vector as input to perform reasoning and obtains multiple contribution degrees output in parallel; a filtering module, which performs working condition and attenuation filtering on the obtained multiple contribution degrees according to the working condition data; and an early warning pushing module, which pushes graded maintenance early warning information to a vehicle-mounted display and a mobile phone terminal according to the type and severity of the fault. The application has the beneficial effects of reducing the false positive rate and adapting to vehicle state degradation.

Description

Technical Field

[0001] This invention relates to the field of engine fault diagnosis technology, specifically to a fuel consumption anomaly tracing system and method based on engine multi-sensor time-series signals. Background Technology

[0002] With the continuous growth of car ownership and increasingly stringent emission regulations, rapid diagnosis of abnormally high engine oil consumption has become an important technical requirement in the vehicle maintenance field. Increased engine oil consumption can be caused by a variety of factors, including but not limited to: air-fuel mixture imbalance, valve timing misalignment, persistent knocking, and insufficient oil lubrication.

[0003] Currently, existing technologies for diagnosing abnormal engine fuel consumption mainly include the following categories:

[0004] (1) Threshold diagnosis method based on instantaneous sensor readings: The instantaneous values ​​of each sensor are read through the OBD-II interface and compared with the preset threshold to determine whether there is an over-limit. However, this method can only determine whether the current value of a single sensor deviates from the normal range. It cannot comprehensively analyze the temporal correlation between multiple sensors, nor can it distinguish the specific causes of increased fuel consumption. This makes maintenance personnel have to rely on experience to check possible faulty components one by one, resulting in low diagnostic efficiency and easy omission of coupled faults.

[0005] (2) Some existing technologies use single or a small number of sensor signals combined with machine learning models to detect specific faults; for example, Chinese patent CN117076964A discloses a diesel engine fault clustering diagnosis method based on LSTM network and artificial fish swarm algorithm, which uses time series data of a single sensor to cluster fault modes; the shortcomings of the above scheme are: it only relies on signals of a limited number of sensor types, lacks the ability to perform cross-domain time series fusion analysis of multiple different physical quantity sensors, and cannot trace multiple underlying causes of increased fuel consumption in parallel at the system level.

[0006] (3) Although the latest solutions also include multi-sensor fusion solutions, they lack external working condition interference filtering mechanisms and the AI ​​diagnostic models in existing solutions mostly rely on cloud training and updates. They cannot use the time-series labeled samples continuously accumulated in the daily operation of the vehicle to perform incremental self-learning of the model on the vehicle. As the vehicle's service life increases, the sensor characteristics will drift, resulting in a high false alarm rate and an inability to adapt to the degradation of the vehicle's condition. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a fuel consumption anomaly tracing system and method based on engine multi-sensor time-series signals, aiming to at least partially solve one of the technical problems in related existing technologies.

[0008] Firstly, this invention provides a fuel consumption anomaly tracing system based on engine multi-sensor time-series signals, comprising:

[0009] The signal acquisition module is used to simultaneously acquire raw time-series signals from multiple sensors;

[0010] The timing synchronization and alignment module is used to establish a unified time reference through a timing synchronization clock, interpolate and resample the original timing signals of various sensors with different sampling frequencies to achieve timing alignment and generate synchronized multi-channel timing data.

[0011] The data binding module is used to obtain operating condition data from the vehicle bus, bind the multi-channel time series data with the operating condition data according to the timestamp, and slice it according to a fixed time window to form a time series sample with operating condition tags.

[0012] The feature extraction module, deployed on the vehicle edge computing terminal, is used to extract preset time-domain features and frequency-domain features from each sensor signal in each time window and construct a fused feature vector.

[0013] The diagnostic model module is used to take the fused feature vector as input, send it to the pre-trained lightweight LSTM model for inference, and obtain multiple contribution values ​​in parallel output.

[0014] The filtering module is used to perform working condition and attenuation filtering on the multiple contribution values ​​obtained based on the working condition data, and compare the filtered contribution values ​​with preset thresholds to distinguish between single hardware faults and multi-sensor coupling faults.

[0015] The early warning push module is used to push graded maintenance early warning information to the vehicle display and mobile phone based on the type and severity of the fault.

[0016] Preferably, the system further includes a self-learning module, which stores the confirmed diagnostic results as new labeled samples in the vehicle's local storage. When the accumulated new labeled samples reach a preset value, the lightweight LSTM model is incrementally trained at the vehicle's edge computing terminal.

[0017] Preferably, during timing alignment, for sensors with low sampling frequencies, cubic spline interpolation is used to upsample them so that their equivalent sampling frequency is aligned with the target frequency; for sensors with high sampling frequencies, they are first processed by a low-pass anti-aliasing filter and then downsampled to the target frequency to avoid spectral aliasing.

[0018] Preferably, the operating condition label is obtained through the following steps:

[0019] Multiple operating condition data are read synchronously through the vehicle's CAN bus interface, and the average value of each data point within each time window is calculated.

[0020] The obtained average value is compared with the corresponding set ratio threshold, and a corresponding label is generated for each time series window based on the comparison result.

[0021] Preferably, the time-domain and frequency-domain features are grouped and concatenated according to sensor type by the feature extraction module to form a fixed-dimensional fused feature vector, which is used as the input of the lightweight LSTM model.

[0022] Preferably, the lightweight LSTM model uses INT8 parameter quantization and includes a cascaded first LSTM layer and a second LSTM layer. A Dropout layer is provided between the first LSTM layer and the second LSTM layer, and the output is connected to multiple independent Sigmoid neurons, each corresponding to a different type of contribution.

[0023] Preferably, before fault differentiation, a congestion attenuation coefficient is calculated based on the average vehicle speed and start-stop frequency in the operating condition data, and a cold start attenuation coefficient is calculated based on the coolant temperature in the operating condition data; during filtering, each contribution is multiplied by the congestion attenuation coefficient and the cold start attenuation coefficient respectively to filter out normal fuel consumption fluctuations caused by road conditions and temperature.

[0024] Secondly, the present invention also provides a method for tracing fuel consumption anomalies based on engine multi-sensor time-series signals, applied to the fuel consumption anomaly tracing system based on engine multi-sensor time-series signals described in the first aspect, the method comprising:

[0025] Simultaneously acquire raw time-series signals from multiple sensors;

[0026] By establishing a unified time reference through a timing synchronization clock, the original timing signals of various sensors with different sampling frequencies are interpolated and resampled to achieve timing alignment and generate synchronized multi-channel timing data.

[0027] The system acquires operating condition data from the vehicle bus, binds the multi-channel time series data to the operating condition data by timestamp, and slices them by fixed time windows to form time series samples with operating condition labels.

[0028] For each sensor signal within each time window, preset time-domain features and frequency-domain features are extracted to construct a fused feature vector;

[0029] The fused feature vector is taken as input and fed into a pre-trained lightweight LSTM model for inference to obtain multiple contribution values ​​in parallel output.

[0030] Based on the operating condition data, the multiple contribution values ​​obtained are filtered by operating condition and attenuation. The filtered contribution values ​​are compared with preset thresholds to distinguish between single hardware faults and multi-sensor coupling faults.

[0031] Based on the type and severity of the fault, graded maintenance warning information is pushed to the vehicle display and mobile phone.

[0032] The beneficial effects of this invention are reflected in:

[0033] (1) By using a timing synchronization clock and interpolation alignment scheme, the problem of synchronous fusion of sensor signals with different sampling frequencies is solved, ensuring that the timing causal relationship learned by the LSTM model is based on the correctly aligned signal, which fundamentally improves the diagnostic accuracy.

[0034] (2) By introducing the corresponding attenuation coefficient, false alarms caused by external operating conditions are effectively filtered out, and the normal fuel consumption fluctuations caused by road conditions and temperature are filtered out, thus reducing the false alarm rate.

[0035] (3) Through continuous self-learning at the vehicle end, the lightweight LSTM model can be continuously optimized throughout the entire life cycle of the vehicle, adapting to the personalized characteristics of each vehicle and the state degradation after long-term use. Attached Figure Description

[0036] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0037] Figure 1 This is a schematic diagram of a fuel consumption anomaly tracing system based on engine multi-sensor time-series signals provided in an embodiment of the present invention;

[0038] Figure 2 This is a flowchart of a fuel consumption anomaly tracing method based on engine multi-sensor timing signals provided in an embodiment of the present invention. Detailed Implementation

[0039] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0040] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0041] Throughout this specification, references to "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination.

[0042] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by those skilled in the art to which this invention pertains.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a fuel consumption anomaly tracing system based on engine multi-sensor time-series signals, including:

[0045] The signal acquisition module is used to simultaneously acquire raw time-series signals from multiple sensors;

[0046] The timing synchronization and alignment module is used to establish a unified time reference through a timing synchronization clock, interpolate and resample the original timing signals of various sensors with different sampling frequencies to achieve timing alignment and generate synchronized multi-channel timing data.

[0047] The data binding module is used to obtain operating condition data from the vehicle bus, bind the multi-channel time series data with the operating condition data according to the timestamp, and slice it according to a fixed time window to form a time series sample with operating condition tags.

[0048] The feature extraction module, deployed on the vehicle edge computing terminal, is used to extract preset time-domain features and frequency-domain features from each sensor signal in each time window and construct a fused feature vector.

[0049] The diagnostic model module is used to take the fused feature vector as input, send it to the pre-trained lightweight LSTM model for inference, and obtain multiple contribution values ​​in parallel output.

[0050] The filtering module is used to perform working condition and attenuation filtering on the multiple contribution values ​​obtained based on the working condition data, and compare the filtered contribution values ​​with preset thresholds to distinguish between single hardware faults and multi-sensor coupling faults.

[0051] The early warning push module is used to push graded maintenance early warning information to the vehicle display and mobile phone based on the type and severity of the fault.

[0052] In this embodiment, multiple sensors employ various physical quantities, including an oxygen sensor, a crankshaft position sensor, a camshaft position sensor, a knock sensor, and an oil pressure sensor. These multiple sensors cover the status information of the three core systems of engine combustion, valve timing, and lubrication, forming a complete signal chain for diagnosing the causes of increased fuel consumption.

[0053] Since the five sensors belong to different types of physical quantity sensors, and their sampling frequencies span from Hz to kHz, direct fusion will inevitably lead to timing misalignment. This invention achieves timing synchronization and alignment through the following steps:

[0054] Step A1: Establish a unified timing synchronization clock reference. Using the crankshaft angle signal from the crankshaft position sensor as the reference clock source, a unified timestamp sequence is generated through the hardware timer of the engine control unit (ECU); using the TDC reference pulse from the crankshaft position sensor as the synchronization trigger signal, a unified time reference is established.

[0055] Step A2: Low-speed sensor interpolation upsampling. For oxygen sensor signals (1~10Hz) and oil pressure sensor signals (10~100Hz) with low sampling frequencies, cubic spline interpolation is used to upsample them so that their equivalent sampling frequencies reach the target frequencies they are aligned with.

[0056] Step A3: High-speed sensor downsampling and anti-aliasing. For knock sensor signals with high sampling frequencies (10~50kHz), the signals are first processed by a low-pass anti-aliasing filter and then downsampled to the target frequency to avoid spectral aliasing.

[0057] Furthermore, the system also includes a self-learning module, which stores the confirmed diagnostic results as new labeled samples in the vehicle's local storage. When the accumulated new labeled samples reach a preset value, the lightweight LSTM model is incrementally trained at the vehicle's edge computing terminal.

[0058] During application, the following data are synchronously read through the vehicle's CAN bus interface: instantaneous fuel consumption data, coolant temperature data, vehicle speed data, and engine speed data;

[0059] The synchronized time series data is sliced ​​into fixed 10-second windows, and each window forms a time series data matrix with 5 channels × 10,000 sampling points.

[0060] The operating condition label is obtained through the following steps:

[0061] Multiple operating condition data are read synchronously through the vehicle's CAN bus interface, and the average value of each data point within each time window is calculated.

[0062] The obtained average value is compared with the corresponding set ratio threshold. Based on the comparison result, a corresponding label is generated for each time series window. For example, if the average fuel consumption exceeds the preset ratio threshold of the benchmark fuel consumption (such as 15%), the time series window is marked as abnormal fuel consumption; otherwise, it is marked as normal fuel consumption.

[0063] Furthermore, to reduce the input dimension and computational complexity of the model and enable its deployment on edge computing devices with limited onboard computing power, the feature extraction module groups and concatenates the time-domain and frequency-domain features according to sensor type to form a fixed-dimensional fused feature vector, which serves as the input to the lightweight LSTM model.

[0064] Specifically, for each sensor signal within each time window, eight time-domain features and four frequency-domain features are extracted, resulting in a fixed feature vector of 5×12=60 dimensions. This feature extraction process employs pipelined parallel processing, with feature extraction for the five sensors executed independently and in parallel.

[0065] In this embodiment, the lightweight LSTM model uses INT8 parameter quantization and includes a cascaded first LSTM layer and a second LSTM layer. A Dropout layer is provided between the first LSTM layer and the second LSTM layer, and the output is connected to multiple independent Sigmoid neurons, each corresponding to a different type of contribution.

[0066] Among them, the model forward inference outputs four contribution values ​​in the range of [0,1], which correspond to air-fuel mixture imbalance, valve timing offset, continuous knocking and insufficient oil lubrication, respectively.

[0067] Model parameter quantization techniques (such as INT8 quantization) are used to compress floating-point model parameters into fixed-point parameters, reducing storage and computational overhead. The LSTM time step T is dynamically adjusted according to actual operating conditions. Under urban congestion conditions, T can be shortened for rapid response, while under high-speed cruising conditions, T can be lengthened to smooth false alarms (T is the number of feature frames within the time window).

[0068] Parallel four-channel output: The output layer of the model uses four independent Sigmoid neurons to output the contribution of four inducing factors in parallel; by learning the nonlinear mapping relationship between the four inducing factors and five sensor signals in a unified temporal feature space at the same time, the coupling information between multi-sensor signals is fully utilized, which improves the comprehensiveness and accuracy of diagnosis.

[0069] If only one of the four contribution categories exceeds the preset single fault threshold, and the contribution of the other three categories is below the coupling exclusion threshold, it is determined to be a single hardware fault, and the corresponding faulty component repair suggestion is output. If at least two contribution categories exceed the coupling judgment threshold at the same time, it is determined to be a multi-sensor coupling fault, and the coupling fault type and related component repair suggestion are output.

[0070] When applying the application, a tiered push strategy is adopted based on the fault type and severity:

[0071] Minor anomalies: Individual contribution rates are between 0.5 and 0.7. The owner will be notified via message on the mobile app only. It is recommended to check the relevant parts during the next maintenance.

[0072] Moderate abnormality: A single contribution exceeds 0.7, or two contributions are between 0.5 and 0.7, and a warning is pushed through the vehicle's dashboard indicator light and the mobile app. It is recommended to have the vehicle checked at a repair shop as soon as possible.

[0073] Serious anomaly: A single contribution score exceeds 0.85, or three or more contribution scores exceed 0.5, and are simultaneously accompanied by audible and visual alarms on the vehicle dashboard and strong reminders pushed to the mobile app. It is recommended to stop the vehicle immediately for inspection.

[0074] The push notification includes: fault type (single / coupled), specific contribution value, a list of recommended parts to be checked, and a suggested repair priority ranking.

[0075] Furthermore, before fault differentiation, a congestion attenuation coefficient is calculated based on the average vehicle speed and start-stop frequency in the operating condition data, and a cold start attenuation coefficient is calculated based on the coolant temperature in the operating condition data. During filtering, each contribution is multiplied by the congestion attenuation coefficient and the cold start attenuation coefficient to filter out normal fuel consumption fluctuations caused by road conditions and temperature.

[0076] For example, if CAN speed data indicates that the vehicle is in urban congestion conditions (average speed below 15km / h and frequent starts and stops) within the analysis window, or if CAN temperature data indicates that the vehicle is in a cold start phase (coolant temperature below 40°C), then the four types of contribution are multiplied by the corresponding operating condition attenuation coefficients to filter out normal fuel consumption fluctuations caused by road conditions and temperature, thereby reducing the false alarm rate.

[0077] In application, an elastic weight consolidation method is used at the vehicle edge computing end to incrementally update the LSTM model parameters. During incremental training, a large update constraint is imposed on model parameters that are important in historical tasks (parameter importance is estimated through the Fisher information matrix) to prevent catastrophic forgetting; thus ensuring that the model can adapt to sensor characteristic drift and engine condition degradation that occur with the increase of vehicle service life.

[0078] Feature extraction and model inference are both completed at the vehicle edge computing terminal, without relying on cloud network connection. This reduces diagnostic latency, meets the real-time requirements of vehicles, and protects the privacy and security of vehicle operation data.

[0079] The above technical solution has the following beneficial effects:

[0080] (1) By using a timing synchronization clock and interpolation alignment scheme, the problem of synchronous fusion of sensor signals with different sampling frequencies is solved, ensuring that the timing causal relationship learned by the LSTM model is based on the correctly aligned signal, which fundamentally improves the diagnostic accuracy.

[0081] (2) By introducing the corresponding attenuation coefficient, false alarms caused by external operating conditions are effectively filtered out, and the normal fuel consumption fluctuations caused by road conditions and temperature are filtered out, thus reducing the false alarm rate.

[0082] (3) Through continuous self-learning at the vehicle end, the lightweight LSTM model can be continuously optimized throughout the entire life cycle of the vehicle, adapting to the personalized characteristics of each vehicle and the state degradation after long-term use.

[0083] Example 2

[0084] Based on the same inventive concept, and referring to Figure 2 This invention also provides a method for tracing fuel consumption anomalies based on engine multi-sensor time-series signals, applied to the aforementioned fuel consumption anomaly tracing system based on engine multi-sensor time-series signals. The method includes:

[0085] S1: Simultaneously acquire raw time-series signals from multiple sensors;

[0086] S2: Establish a unified time reference through a timing synchronization clock, interpolate and resample the original timing signals of various sensors with different sampling frequencies to achieve timing alignment and generate synchronized multi-channel timing data;

[0087] S3: Obtain operating condition data from the vehicle bus, bind the multi-channel timing data with the operating condition data according to timestamps, and slice them according to fixed time windows to form timing samples with operating condition labels;

[0088] S4: Extract preset time-domain and frequency-domain features for each sensor signal within each time window, and construct a fused feature vector;

[0089] S5: The fused feature vector is input and fed into the pre-trained lightweight LSTM model for inference to obtain multiple contributions in parallel output.

[0090] S6: Based on the operating condition data, perform operating condition and attenuation filtering on the multiple contribution values ​​obtained, and compare the filtered contribution values ​​with preset thresholds to distinguish between single hardware faults and multi-sensor coupling faults.

[0091] S7: Push graded maintenance warning information to the vehicle display and mobile phone based on the type and severity of the fault.

[0092] Furthermore, based on the above technical solution, the method also includes:

[0093] The confirmed diagnostic results are stored as new labeled samples in the vehicle's local storage. When the accumulated new labeled samples reach a preset value, the lightweight LSTM model is incrementally trained at the vehicle's edge computing terminal.

[0094] It should be noted that for a more detailed description of the workflow and implementation of the method embodiments, please refer to the aforementioned system embodiments section, which will not be repeated here.

[0095] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can also be implemented in other ways. The method embodiments described above are merely illustrative. It should also be noted that in some alternative implementations, the functions marked in the boxes may occur in a different order than those shown in the accompanying drawings.

[0096] Furthermore, the functional modules in the various embodiments of this invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. The integrated modules can be implemented in hardware or as software functional modules. When in use, each module will only collect and store user information with the user's full authorization and in compliance with relevant laws and regulations, protecting the security and privacy of user data and strictly prohibiting unauthorized access; data processing will be conducted within the scope stipulated by law and will not exceed the purpose and scope authorized by the user; at the same time, users have the rights to access, correct, delete, restrict, and refuse processing of their personal data; and must strictly comply with applicable laws and regulations and conduct compliance reviews.

[0097] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A fuel consumption anomaly tracing system based on engine multi-sensor time-series signals, characterized in that, include: The signal acquisition module is used to simultaneously acquire raw time-series signals from multiple sensors; The timing synchronization and alignment module is used to establish a unified time reference through a timing synchronization clock, interpolate and resample the original timing signals of various sensors with different sampling frequencies to achieve timing alignment, and generate synchronized multi-channel timing data. The data binding module is used to obtain operating condition data from the vehicle bus, bind the multi-channel time series data with the operating condition data according to the timestamp, and slice it according to a fixed time window to form a time series sample with operating condition tags. The feature extraction module, deployed on the vehicle edge computing terminal, is used to extract preset time-domain features and frequency-domain features from each sensor signal in each time window and construct a fused feature vector. The diagnostic model module is used to take the fused feature vector as input, send it to the pre-trained lightweight LSTM model for inference, and obtain multiple contribution values ​​in parallel output. The filtering module is used to perform working condition and attenuation filtering on the multiple contribution values ​​obtained based on the working condition data, and compare the filtered contribution values ​​with preset thresholds to distinguish between single hardware faults and multi-sensor coupling faults. The early warning push module is used to push graded maintenance early warning information to the vehicle display and mobile phone based on the type and severity of the fault.

2. The system according to claim 1, characterized in that, The system also includes a self-learning module, which stores the confirmed diagnostic results as new labeled samples in the vehicle's local storage. When the accumulated new labeled samples reach a preset value, the lightweight LSTM model is incrementally trained at the vehicle's edge computing terminal.

3. The system according to claim 1 or 2, characterized in that, During timing alignment, for sensors with low sampling frequencies, cubic spline interpolation is used to upsample them so that their equivalent sampling frequency is aligned with the target frequency. For sensors with high sampling frequencies, they are first processed by a low-pass anti-aliasing filter and then downsampled to the target frequency to avoid spectral aliasing.

4. The system according to claim 3, characterized in that, The operating condition label is obtained through the following steps: Multiple operating condition data are read synchronously through the vehicle's CAN bus interface, and the average value of each data point within each time window is calculated. The obtained average value is compared with the corresponding set ratio threshold, and a corresponding label is generated for each time series window based on the comparison result.

5. The system according to claim 2, characterized in that, The feature extraction module groups and concatenates the time-domain and frequency-domain features according to sensor type to form a fixed-dimensional fused feature vector, which serves as the input to the lightweight LSTM model.

6. The system according to claim 5, characterized in that, The lightweight LSTM model uses INT8 parameter quantization and includes a cascaded first LSTM layer and a second LSTM layer. A Dropout layer is provided between the first LSTM layer and the second LSTM layer. The output is connected to multiple independent Sigmoid neurons, each corresponding to a different type of contribution.

7. The system according to claim 1, characterized in that, Before fault differentiation, the congestion attenuation coefficient is calculated based on the average vehicle speed and start-stop frequency in the operating condition data, and the cold start attenuation coefficient is calculated based on the coolant temperature in the operating condition data. During filtering, each contribution is multiplied by the congestion attenuation coefficient and the cold start attenuation coefficient to filter out normal fuel consumption fluctuations caused by road conditions and temperature.

8. A method for tracing the source of fuel consumption anomalies based on time-series signals from multiple engine sensors, characterized in that, The method applied to the fuel consumption anomaly tracing system based on engine multi-sensor time-series signals as described in claim 1 includes: Simultaneously acquire raw time-series signals from multiple sensors; By establishing a unified time reference through a timing synchronization clock, the original timing signals of various sensors with different sampling frequencies are interpolated and resampled to achieve timing alignment and generate synchronized multi-channel timing data. The system acquires operating condition data from the vehicle bus, binds the multi-channel time series data to the operating condition data by timestamp, and slices them by fixed time windows to form time series samples with operating condition labels. For each sensor signal within each time window, preset time-domain features and frequency-domain features are extracted to construct a fused feature vector; The fused feature vector is taken as input and fed into a pre-trained lightweight LSTM model for inference to obtain multiple contribution values ​​in parallel output. Based on the operating condition data, the multiple contribution values ​​obtained are filtered by operating condition and attenuation. The filtered contribution values ​​are compared with preset thresholds to distinguish between single hardware faults and multi-sensor coupling faults. Based on the type and severity of the fault, graded maintenance warning information is pushed to the vehicle display and mobile phone.

9. The method according to claim 8, characterized in that, The method further includes: The confirmed diagnostic results are stored as new labeled samples in the vehicle's local storage. When the accumulated new labeled samples reach a preset value, the lightweight LSTM model is incrementally trained at the vehicle's edge computing terminal.

Images