An automobile sensor real-time fault diagnosis method based on adaptive feature fusion

Abstract

The application discloses a kind of real-time fault diagnosis methods of automobile sensor based on adaptive feature fusion, it is related to the technical field of automobile electronics and fault diagnosis.The method extracts the physical mechanism feature and dynamic timing feature of sensor in parallel, map two kinds of features to unified dimension;Based on the sensitivity stability in physical mechanism feature and the scalar value of dynamic timing feature, the fusion weight is calculated by learnable Sigmoid function;The weight is used to weight fusion to obtain the fusion feature vector of two kinds of features;Finally, according to the fusion feature vector, the fault diagnosis result is output.The application also adapts the different channel selection strategy of different heterogeneous data set.The application solves the problem of insufficient feature expression, insufficient real-time, weak multi-working condition generalization ability in existing automobile sensor fault diagnosis, and the fault diagnosis accuracy reaches 92.1%, the inference delay is as low as 18ms, meets ISO26262ASIL-B on-board real-time diagnosis standard.

Description

Technical Field

[0001] This invention relates to the field of automotive electronics and fault diagnosis technology, and in particular to a real-time fault diagnosis method for automotive sensors based on adaptive feature fusion. Background Technology

[0002] With the rapid development of intelligent connected vehicles, Advanced Driver Assistance Systems (ADAS) have become the core for improving driving safety and comfort. As the fundamental sensing components of ADAS, the reliability of sensors directly determines the system's functional safety. According to field research data, sensor failures account for approximately 42% of electronic system failures in vehicles equipped with ADAS. Failures in powertrain-related sensors, such as the accelerator pedal position sensor and throttle position sensor, can easily lead to problems like unstable engine idling and acceleration jerking. If diagnosis is not timely, the fault may escalate, exacerbating safety risks.

[0003] Automotive sensor fault diagnosis requires strict real-time performance: according to the ISO 26262 functional safety standard, the maximum allowable latency for fault diagnosis at ASIL-B level (conventional ADAS functions) should not exceed 20ms, while higher safety levels or high dynamic response scenarios require latency requirements of 10ms. However, existing fault diagnosis methods have many shortcomings: traditional methods based on a single deep learning model (such as CNN and LSTM) either only extract spatial features while ignoring temporal information, or have limited temporal feature representation capabilities, resulting in low accuracy in identifying complex nonlinear faults; while methods combining physical models and deep learning improve diagnostic accuracy to some extent, their high model complexity and computational load make it difficult to meet the real-time requirements of automotive embedded platforms; existing multi-source feature fusion techniques mostly adopt fixed-weight fusion mechanisms, which cannot dynamically adjust the feature contribution ratio according to fault type and operating condition changes, making them prone to false alarms or missed detections in complex scenarios such as high-speed vibration, slow signal drift, and temperature fluctuations; most methods do not have adaptation strategies designed for heterogeneous datasets, resulting in poor generalization ability and making it difficult to directly apply to fault diagnosis of different vehicle models and sensor types. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a real-time fault diagnosis method for automotive sensors based on adaptive feature fusion. By constructing a dual-channel parallel extraction architecture of physical mechanism features and dynamic temporal features, and combining an adaptive weight allocation mechanism and a differentiated channel adaptation strategy, this invention solves the problems of insufficient feature expression, insufficient real-time performance, and weak generalization ability under multiple working conditions in existing technologies, thereby achieving high-precision, low-latency, and high-reliability diagnosis of automotive sensor faults.

[0005] The technical solution of this invention: A real-time fault diagnosis method for automotive sensors based on adaptive feature fusion, comprising: Step 1: Extract the physical mechanism features and dynamic temporal features of the sensor signals in parallel, and map the physical mechanism features and the dynamic temporal features to the same feature dimension respectively; Step 2: Based on the scalarized values ​​of sensitivity stability and dynamic temporal features extracted in Step 1, calculate the fusion weights using a learnable Sigmoid function. ; Step 3: Use the fusion weights calculated in Step 2 The physical mechanism features and dynamic temporal features are weighted and fused to obtain a fused feature vector. The calculation formula is: ; in, For physical mechanism feature vectors, It is a dynamic temporal feature vector; Step 4: Combine the fused feature vectors Output the fault diagnosis results, which include the fault type and the basis for diagnosis.

[0006] Preferably, the extraction of physical mechanism features in step one includes: calculating the sensitivity stability of the sensor signal. Angle mapping signal fluctuation amplitude and noise level The sensitivity stability The calculation formula is: ; in, The number of sampling points. For the first The actual output voltage of the time sensor. This is the voltage estimate after filtering by a 5th-order FIR low-pass filter; The noise level The calculation formula is: ; And extract noise level Simultaneously calculate the ratio of the energy in the 20-50Hz frequency band to the total energy. When the ratio is greater than 0.6, the current signal is determined to be a road interference signal and is not included in the fault diagnosis basis.

[0007] Preferably, the low-pass filter is implemented using a 5th-order FIR filter with a cutoff frequency of 10Hz. The group delay of this filter is 2 sampling points, and it is processed in parallel with the filter through a sliding window.

[0008] Preferably, the extraction of dynamic temporal features in step one is achieved using a bidirectional long short-term memory network. The input of the bidirectional long short-term memory network is a sensor signal sequence with a length of 50 points, and its hidden layer dimension is 64. Training samples are generated using a sliding window with a step size of 5, and the sliding window and model inference are executed in parallel.

[0009] Preferably, in step two, the fusion weights The calculation formula is: ; in, For dynamic time-series feature vectors, The scalarization function for time series features is calculated using the mean of all elements in the time series feature vector. The learning parameters are updated synchronously through the backpropagation algorithm during the offline training phase. 10% of heterogeneous working condition data is incorporated into the training process to improve generalization ability.

[0010] Preferably, the method further includes step five: dynamically adjusting the fusion strategy of the physical mechanism features and the dynamic temporal features based on the completeness of the sensor parameter fields and the signal stability index of the input data; specifically, enabling dual-channel full feature fusion for datasets containing complete physical parameters, extracting only general physical mechanism features for datasets with incomplete physical parameters, and enabling only the temporal channel for datasets without physical parameters.

[0011] Preferably, before outputting the fault diagnosis result in step four, a timing consistency verification step is also included: a sliding window of length 5 is used to vote on the continuously output diagnosis results. The fault type is only output when more than half of the judgment results in the window are consistent; otherwise, the diagnosis result of the previous moment is maintained.

[0012] Preferably, the dynamic temporal features extracted by the bidirectional long short-term memory network include the average signal change rate. Forward average absolute fluctuation and Fourier transform main frequency The average signal change rate The calculation formula is: ; in, For a fixed sampling time interval, when the average signal change rate of 5 consecutive sampling points is greater than 0.5V / s and the throttle opening change rate is greater than... At that time, the fault threshold for the average signal change rate is adjusted to three times the normal threshold.

[0013] Compared with existing technologies, the beneficial effects of this invention are: 1. This invention significantly improves fault diagnosis accuracy by constructing a dual-channel parallel extraction architecture for physical mechanism features and dynamic temporal features, and introducing a learnable adaptive weight allocation mechanism to dynamically adjust the fusion ratio of the two types of features. Experimental verification shows that the diagnostic accuracy of this invention reaches 92.1% on a publicly available automotive sensor fault dataset, which is 4.6 percentage points higher than the traditional LSTM model. Simultaneously, the lightweight network design and parallel processing strategy result in a single inference latency as low as 18ms, fully meeting the 20ms real-time constraint required by the ISO 26262 ASIL-B standard. The false alarm rate is below 8.7% under various operating conditions, and further reduced to 5.3% after optimization, balancing the engineering requirements of high accuracy and low latency. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of the real-time fault diagnosis method for automotive sensors based on adaptive feature fusion, as described in this invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Example 1 See attached document Figure 1 This embodiment discloses a real-time fault diagnosis method for automotive sensors based on adaptive feature fusion, including the following steps: Step S1: Input raw sensor data; Specifically, the raw voltage signals output by the vehicle's sensors are collected, with a sampling frequency set to 100Hz to meet the bandwidth requirements of the vehicle's sensors. The raw signals undergo preprocessing, including removing DC components, outlier removal, and standardization. Global standardization (mean 0, variance 1) is used during offline training, while online diagnostics uses standardization within a 50-point sliding window to accommodate signal drift caused by changes in operating conditions.

[0017] Step S2: Dual-channel parallel feature extraction; Its dual-channel feature extraction architecture extracts physical mechanism features and dynamic temporal features in parallel. Both types of features are ultimately mapped to 64-dimensional feature vectors, including: Step S21: Physical mechanism feature extraction channel. Based on the working principle of the sensor and the physical characteristics of the fault, the following three types of core physical mechanism features are extracted: Sensitivity stability Its stability reflects the stability of the sensor's output signal as the input changes, and is a key indicator for judging sensor aging and signal drift. Sensitivity stability The calculation formula is: ; in, The number of sampling points. For the first The actual output voltage of the time sensor. The voltage estimate is the value after filtering by a 5th-order FIR low-pass filter; where the voltage estimate is... The delay was obtained using a 5th-order FIR low-pass filter with a cutoff frequency of 10Hz. The group delay of this filter is 2 sampling points, and it is processed in parallel with the filter through a sliding window, without adding any additional system delay.

[0018] Angle mapping signal fluctuation amplitude It characterizes the fluctuation characteristics of the sensor's angle-signal conversion relationship and is applicable to angle-type sensors such as accelerator pedal position sensors and throttle position sensors. Its calculation formula is: ; in, For the first i The sensor input angle at any given time. The average angle over the sampling period. This is the angle-to-signal conversion function for the sensor.

[0019] noise level This reflects the noise intensity of the sensor's output signal and is used to distinguish between sensor aging noise and external interference noise. When calculating the noise level, a Fast Fourier Transform is simultaneously performed on the signal to calculate the ratio of energy in the 20-50Hz frequency band to the total energy. When this ratio is greater than 0.6, the current noise is determined to be mainly caused by road vibration and is not included in the fault diagnosis criteria, effectively reducing the false alarm rate. Noise Level The calculation formula is: ; The extracted 3D physical mechanism feature vector is mapped to a 64D feature vector through a linear layer (fully connected layer). This ensures that the output dimension is consistent with that of the dynamic temporal feature channel, thus guaranteeing the efficiency of subsequent fusion.

[0020] Step S22: Dynamic temporal feature extraction channel. An improved bidirectional long short-term memory network (BiLSTM) is used to extract the dynamic temporal features of the sensor signal. BiLSTM can simultaneously capture the forward and backward temporal dependencies of the signal, and has a stronger temporal feature representation capability than traditional LSTM.

[0021] The network configuration is as follows: the BiLSTM hidden layer dimension is set to 64, and the input is a sensor signal sequence of length 50 points. Training and inference samples are generated using a sliding window with a step size of 5. Data preprocessing and model inference are performed in parallel, and the diagnostic results are updated at an interval of 10ms, which meets the real-time requirements.

[0022] The core temporal features extracted by the BiLSTM network include: Average signal change rate It reflects the rate of change of the signal over time. An operating condition identification factor is introduced to avoid false alarms in high-dynamic operating conditions. The formula is: ; in, With a fixed sampling time interval (10ms), when the average rate of change of the signal at 5 consecutive sampling points is greater than 0.5V / s and the rate of change of the throttle opening is greater than 5° / s, the fault threshold is adjusted to 3 times the normal threshold to avoid misjudgment.

[0023] Forward Average Absolute Fluctuation This characterizes the local fluctuation properties of a signal and is used to identify faults such as intermittent poor contact. Its calculation formula is: ; Among them, when hour, .

[0024] Fourier transform frequency The core frequency domain features of the signal are extracted to identify periodic fluctuation-type faults. The calculation formula is as follows: ; in, This represents the Discrete Fourier Transform operator. It is a time-domain voltage signal sequence of length N.

[0025] Step S3: Adaptive feature fusion, which dynamically adjusts the fusion ratio of physical mechanism features and temporal features according to the fault type. "Adaptive" specifically refers to the process of dynamically calculating and allocating fusion weights based on the physical mechanism features and dynamic temporal features of the sensor signal using a learnable weight allocation function. This process automatically adjusts the contribution ratio of physical mechanism features and temporal features in the final diagnostic decision for different fault types (such as signal drift, periodic fluctuations, etc.). The specific process is as follows: S31: Weighting coefficient calculation, the calculation method is as follows: ; in, For dynamic time-series feature vectors, The scalarization function for time series features is calculated using the mean of all elements in the time series feature vector. The learning parameters are updated synchronously through the backpropagation algorithm during the offline training phase. 10% of heterogeneous working condition data is incorporated into the training process to improve generalization ability.

[0026] S32: Feature fusion, calculated as follows: ; in, To fuse feature vectors, For physical mechanism feature vectors, This is a dynamic temporal feature vector.

[0027] Fusion weights The value range is [0,1]: when When the value is close to 1, the model mainly relies on physical mechanism characteristics for diagnosis, and is suitable for slowly varying faults such as signal drift and sensitivity decay; when... When the value is close to 0, the model mainly relies on time-series characteristics for diagnosis, and is suitable for rapidly changing faults such as periodic fluctuations and intermittent poor contact.

[0028] Step S4: Fault diagnosis and output, which involves fusing feature vectors. The input fault classifier consists of two fully connected layers and a Softmax activation function. The first fully connected layer maps the 64-dimensional fused features to 32-dimensional features, and the second fully connected layer maps the 32-dimensional features to the output dimension of the number of fault categories. The probability distribution of each fault category is obtained through the Softmax activation function.

[0029] The output diagnostic results include: fault type (such as normal, signal drift, sensitivity decay, periodic fluctuation, intermittent poor contact, etc.), fault probability, and the contribution of each feature. The feature contribution is obtained by calculating the gradient of each feature with respect to the final classification result, providing a clear basis for fault diagnosis.

[0030] Step S5: Differentiated channel adaptation, which dynamically adjusts the fusion strategy of the physical mechanism features and the dynamic temporal features based on the completeness of the sensor parameter fields and the signal stability index of the input data; specifically, dual-channel full feature fusion is enabled for datasets with complete physical parameters, only general physical mechanism features are extracted for datasets with incomplete physical parameters, and only the temporal channel is enabled for datasets without physical parameters. This can be further divided into: S51: The feature completeness detection module calculates the completeness of parameter fields (such as whether it contains physical parameters such as angle and rotation speed) and signal stability indicators (such as signal variance and fluctuation amplitude) of the input sensor data in real time. S52: When the input data is detected to contain complete physical parameters and the signal is stable, the dual-channel full-feature fusion mode is enabled to fully utilize physical mechanisms and temporal features to improve diagnostic accuracy. S53: When incomplete physical parameters of the input data are detected, the physical mechanism feature channel only extracts general features such as sensitivity stability and noise level for fusion. S54: When the input data is found to have no physical parameters, the physical mechanism feature channel is automatically turned off, and only the dynamic timing feature channel is enabled for fault diagnosis.

[0031] The working principle of the method based on the above embodiments is as follows: physical mechanism features and dynamic temporal features are extracted in parallel through dual channels, and the fusion ratio of the two types of features is dynamically adjusted according to the fault type using an adaptive weight allocation mechanism to achieve high-precision fault diagnosis; at the same time, a sliding window and model inference parallel execution strategy is adopted to ensure real-time requirements.

[0032] Example 2 Based on the above embodiments, this embodiment provides experimental verification data.

[0033] Experimental Data: The performance of the method is validated using three publicly available and reproducible datasets, including: 1. SAE J1939 Automotive Sensor Fault Dataset: Contains 1500 samples, covering 8 common sensor faults, such as signal drift, sensitivity decay, periodic fluctuations, etc., collected in a constant temperature and humidity laboratory environment. 2. NASA Bearing Dataset Extended Version: Contains 1500 samples, covering 12 bearing failure conditions, providing only the raw vibration time-series signals; 3. UCI Sensor Fault Dataset: 1200 samples, containing 7 types of automotive sensor faults.

[0034] Dataset preprocessing: Global standardization (mean 0, variance 1) is used in the offline training stage, and standardization within a sliding window of 50 sampling points is used in the online diagnosis stage; each dataset is divided into training set, validation set and test set in a 7:2:1 ratio, and multi-sensor signals are fused at the feature level to form a unified input.

[0035] Experimental configuration: The experimental platform consisted of an NVIDIA RTX 3090 GPU and an Intel Xeon E5-2686v4 CPU; the embedded platform was an NVIDIA Jetson Orin NX. The software environment was Ubuntu 22.04, Python 3.9.16, and PyTorch 2.1.0. Experimental results: The complete model achieved an accuracy of 92.1% with an inference latency of 18ms on the RTX 3090; and an accuracy of 91.7% with an inference latency of 22ms on the Jetson Orin NX, meeting the ISO 26262 ASIL-B standard.

[0036] Experimental Results and Verification: Performance Comparison Experiment: The performance comparison between the method of this invention and existing methods on the SAEJ1939 dataset is shown in the table below: Table 1:

[0037] The results show that the accuracy difference between the method of the present invention and that of a single BiLSTM channel is p=0.023<0.05, which is statistically significant.

[0038] Robustness verification under different operating conditions: The diagnostic performance under different operating conditions is shown in the table below. Table 2:

[0039] The results show that the method of the present invention maintains stable performance under complex road conditions and high-speed vibration interference, and has a significant effect on suppressing temperature fluctuations.

[0040] The ablation experiments on the SAEJ1939 dataset are shown in the table below: Table 3:

[0041] Ablation experiments verified the effectiveness of introducing physical mechanism features and adopting an adaptive fusion mechanism. Lightweight optimization reduced the inference latency from 23ms to 18ms while only decreasing the accuracy by 1.4 percentage points, achieving an engineering balance between accuracy and real-time performance.

[0042] The fault diagnosis example uses signal drift fault as an example, and the diagnostic criteria are shown in the table below: Table 4:

[0043] The diagnosis was a signal drift fault, which was reliable and verified the interpretability of the method of the present invention.

[0044] In summary, the fault diagnosis method based on adaptive feature fusion provided by this invention solves the core contradiction between accuracy and real-time performance through a specific technical architecture. Experimental data shows that compared to the basic LSTM model, which is the closest to existing technology, the diagnostic accuracy of this invention is significantly improved, while simultaneously reducing inference latency to a level that meets automotive functional safety standards.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automobile sensor real-time fault diagnosis method based on adaptive feature fusion, characterized in that, include: Step 1: Extract the physical mechanism features and dynamic temporal features of the sensor signals in parallel, and map the physical mechanism features and the dynamic temporal features to the same feature dimension respectively; Step 2: Based on the scalarized values ​​of sensitivity stability and dynamic temporal features extracted in Step 1, calculate the fusion weights using a learnable Sigmoid function. ; Step 3: Use the fusion weights calculated in Step 2 The physical mechanism features and dynamic temporal features are weighted and fused to obtain a fused feature vector. The calculation formula is: ； wherein, is a physical mechanism feature vector, is a dynamic timing feature vector; Step 4: Combine the fused feature vectors Output the fault diagnosis results, which include the fault type and the basis for diagnosis.

2. The real-time fault diagnosis method for vehicle sensors based on adaptive feature fusion according to claim 1, characterized in that, The extraction of physical mechanism features in step one includes: calculating the sensitivity and stability of the sensor signal. Angle mapping signal fluctuation amplitude and noise level The sensitivity stability The calculation formula is: ； in, The number of sampling points. For the first The actual output voltage of the time sensor. This is the voltage estimate after filtering by a 5th-order FIR low-pass filter; The noise level The calculation formula is: ； and extracting noise level When the ratio of the signal energy in the 20-50 Hz band to the total energy is greater than 0.6, the current signal is determined to be a road disturbance signal and is not included in the basis for fault diagnosis.

3. The real-time fault diagnosis method for vehicle sensors based on adaptive feature fusion according to claim 2, characterized in that, The low-pass filter is implemented using a 5th-order FIR filter with a cutoff frequency of 10Hz. The group delay of this filter is 2 sampling points, and it is processed in parallel with the filter through a sliding window.

4. The real-time fault diagnosis method for automotive sensors based on adaptive feature fusion according to claim 1, characterized in that, The extraction of dynamic temporal features in step one is achieved using a bidirectional long short-term memory network. The input of the bidirectional long short-term memory network is a sensor signal sequence with a length of 50 points. Its hidden layer dimension is 64. Training samples are generated using a sliding window with a step size of 5, and the sliding window and model inference are executed in parallel.

5. The real-time fault diagnosis method for automotive sensors based on adaptive feature fusion according to claim 1, characterized in that, The fusion weight in the step two The calculation formula is: ； in, For dynamic time-series feature vectors, The scalarization function for time series features is calculated using the mean of all elements in the time series feature vector. The learning parameters are updated synchronously through the backpropagation algorithm during the offline training phase. 10% of heterogeneous working condition data is incorporated into the training process to improve generalization ability.

6. The real-time fault diagnosis method for automotive sensors based on adaptive feature fusion according to claim 1, characterized in that, The method also includes step five: dynamically adjusting the fusion strategy of the physical mechanism features and the dynamic temporal features based on the completeness of the sensor parameter fields and the signal stability index of the input data; specifically, dual-channel full feature fusion is enabled for datasets with complete physical parameters, only general physical mechanism features are extracted for datasets with incomplete physical parameters, and only the temporal channel is enabled for datasets without physical parameters.

7. The real-time fault diagnosis method for automotive sensors based on adaptive feature fusion according to claim 1, characterized in that, Before outputting the fault diagnosis result in step four, a timing consistency verification step is also included: a sliding window of length 5 is used to vote on the continuously output diagnosis results. The fault type is only output when more than half of the judgment results in the window are consistent; otherwise, the diagnosis result of the previous moment is maintained.

8. The real-time fault diagnosis method for automotive sensors based on adaptive feature fusion according to claim 1, characterized in that, The dynamic temporal features extracted by the bidirectional long short-term memory network include the average signal change rate. Forward average absolute fluctuation and Fourier transform main frequency The average signal change rate The calculation formula is: ； in, For a fixed sampling time interval, when the average signal change rate of 5 consecutive sampling points is greater than 0.5V / s and the throttle opening change rate is greater than... At that time, the fault threshold for the average signal change rate is adjusted to three times the normal threshold.

Images