Data fusion processing method for automobile engine exhaust temperature sensor

Abstract

The invention discloses an automobile engine exhaust temperature sensor data fusion processing method, which comprises the following steps: firstly, collecting engine exhaust temperature through a plurality of exhaust temperature sensors at different positions, and carrying out abnormal value correction on abnormal value data; secondly, decomposing the corrected data into a plurality of intrinsic mode functions through variational mode decomposition, dividing the intrinsic mode functions into components of different time scales based on center frequencies of mode components, and extracting a customized feature set; and then fusing sensor distance and time delay parameters by adopting an improved attention mechanism time convolutional network with physical constraints to realize multi-scale spatial-temporal feature weighted fusion. And finally, a final fusion temperature value is output through weighted pooling. According to the method, the precision and robustness of exhaust temperature monitoring are remarkably improved, the problem of fusion deviation caused by space-time asynchronism of multi-sensor data is solved, and the method is particularly suitable for an engine exhaust temperature state real-time monitoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile engine exhaust temperature detection, and particularly relates to a data fusion processing method for an automobile engine exhaust temperature sensor. BACKGROUND

[0002] In order to meet the strict emission standards of the sixth national standard, a three-way catalyst and a particle trap are added in the exhaust system of an automobile engine. The main function of the three-way catalyst is to convert the harmful gases such as carbon monoxide, hydrocarbons and nitrogen oxides in the exhaust gas into harmless carbon dioxide, water and nitrogen through oxidation and reduction under the action of the catalyst adjusted according to the exhaust temperature and the working state of the engine, so as to meet the emission standard requirements. The particle trap is mainly used for trapping particulate matters in the exhaust gas, and controls the combustion process of the particles according to the exhaust temperature to achieve the best combustion effect and reduce the particulate emission. Since the working of the three-way catalyst and the particle trap both need to rely on the exhaust temperature data, the accuracy and real-time performance of the exhaust temperature data will greatly affect the performance of the three-way catalyst and the particle trap.

[0003] In actual working conditions, the exhaust temperature data collected by a single exhaust temperature sensor is often disturbed by various factors and has obvious noise problems. These noises may be caused by the vibration of the engine during operation, the violent fluctuation of the exhaust gas flow, or the slight disturbance of the electronic elements of the sensor itself, so that the collected temperature values show irregular jumps and burrs. Secondly, the accuracy of the exhaust temperature data collected by a single exhaust temperature sensor is difficult to guarantee, for example: the exhaust temperature sensor may have calibration deviation after long-term use, resulting in fixed deviation between the detected value and the actual temperature; in high temperature environment, the sensitivity of the exhaust temperature sensor decreases, and it cannot accurately reflect the real temperature. Furthermore, the exhaust temperature data collected by a single exhaust temperature sensor is also unstable, and the temperature at the same position will fluctuate greatly in a short time due to the response delay of the exhaust temperature sensor, external electromagnetic interference and other factors, which brings trouble to temperature monitoring. Therefore, in order to obtain accurate exhaust temperature, multiple exhaust temperature sensors need to be arranged in the exhaust system, and these exhaust temperature sensors are respectively installed at both ends of the three-way catalyst and the particle trap. However, how to fuse and process the exhaust temperature data collected by these exhaust temperature sensors which may have noise interference, insufficient accuracy or lack of stability to obtain real-time and accurate exhaust temperature data of the entire exhaust system is a problem to be solved at present. SUMMARY

[0004] The application aims at solving the problem that the exhaust temperature data of the current automobile engine is difficult to obtain in real time and accurately, and provides an automobile engine exhaust temperature sensor data fusion processing method, which has the characteristics of high processing speed, good fusion effect and high accuracy.

[0005] To solve the above problems, the application is realized by the following technical scheme:

[0006] An automobile engine exhaust temperature sensor data fusion processing method, comprising the following steps:

[0007] Step 1, using two or more exhaust temperature sensors installed at different positions of the automobile engine, collecting the exhaust temperature of the automobile engine within a period of time;

[0008] Step 2, calculating the average exhaust temperature of all temperature sensors at each sampling point, and screening out the abnormal exhaust temperature of the temperature sensor at the corresponding sampling point and performing calibration;

[0009] Step 3, using the variational mode decomposition algorithm to perform variational mode decomposition on the exhaust temperature collected by each exhaust temperature sensor, to obtain two or more IMF components and their corresponding center frequencies of each exhaust temperature sensor;

[0010] Step 4, distributing each IMF component of each exhaust temperature sensor to the corresponding time scale according to its corresponding center frequency, and extracting features in the time window corresponding to each time scale to obtain a multi-scale feature set of each exhaust temperature sensor;

[0011] Step 5, sending each time scale feature set of the multi-scale feature set of each exhaust temperature sensor into a time convolution network respectively to obtain the temperature features of each exhaust temperature sensor at different time scales;

[0012] Step 6, sending the temperature features of each exhaust temperature sensor at different time scales into a spatio-temporal attention mechanism network to obtain temperature enhanced features of each exhaust temperature sensor at different time scales;

[0013] Step 7, using a linear interpolation function to align the dimensions of the temperature enhanced features of the exhaust temperature sensor at different time scales to obtain temperature aligned features of each exhaust temperature sensor at different time scales;

[0014] Step 8, dynamically assigning time scale weights according to the importance of different time scale features, and using the time scale weights to weight and sum the temperature aligned features of each exhaust temperature sensor at different time scales to obtain temperature fusion features of each exhaust temperature sensor;

[0015] Step 9: Perform average pooling to integrate the temperature fusion features of all exhaust temperature sensors to obtain multi-sensor temperature integration features, and map the multi-sensor temperature integration features to the final exhaust temperature value.

[0016] In step 1 above, the sampling frequency of the exhaust temperature sensor is an integer multiple of 1 Hz and greater than or equal to 1 Hz, and the sampling duration is an integer multiple of 10 minutes and greater than or equal to 10 minutes.

[0017] In step 2 above, the specific process of screening out the abnormal exhaust temperatures of each temperature sensor at the corresponding sampling point and calibrating them is as follows:

[0018] When the temperature sensor detects the exhaust temperature x at sampling point t m (t) and the average exhaust temperature at sampling point t If the difference is within the preset allowable range, then the exhaust temperature x is retained. m (t);

[0019] When the temperature sensor detects the exhaust temperature x at sampling point t m (t) and the average exhaust temperature at sampling point t If the temperature exceeds the preset allowable range, the average exhaust temperature at the previous sampling point t-1 will be used. Replace this exhaust temperature x m (t);

[0020] In the above, m = 1, 2, ..., M, where M is the number of exhaust temperature sensors, and t = 1, 2, ..., T, where T is the sampling duration.

[0021] The specific process of step 4 above is as follows:

[0022] Step 4.1: Convert the individual IMF components u of each exhaust temperature sensor m,k According to its corresponding center frequency ω m,k Assigned to the corresponding time scale, that is:

[0023] When ω m,k When the frequency is ≥1Hz, the IMF component of the exhaust temperature sensor will be u m,k Assign to a second scale; within the second scale, use a time window of size 1 second to assign the IMF component u. m,k The time windows are divided into segments, with the number of segments being T. s ;

[0024] When 1 / 600Hz < ω m,k When <1Hz, the IMF component u of the exhaust temperature sensor will be... m,k Assign to a minute scale; within the minute scale, use a time window with a window size of 1 minute for the IMF component u. m,kperforming the division, the number of time windows is T min ;

[0025] When ω m,k ≤ 1 / 600 Hz, the IMF component u m,k of the exhaust temperature sensor is assigned to the 10-minute scale; in the 10-minute scale, the IMF component u m,k is divided using time windows with a window size of 10 minutes, the number of time windows is T 10min ;

[0026] Step 4.2, for each individual IMF component u m,k of the exhaust temperature sensor, according to the time scale to which it is assigned, the features in the corresponding time window are extracted;

[0027] For the IMF component u m,k assigned to the second scale, 4 time-domain features of each time window i s , namely the mean variance skewness and peak amplitude are calculated using the exhaust temperature of all sampling points in each time window i m,k , and the second feature F m,k,s of the IMF component u m,k is obtained;

[0028] For the IMF component u m,k assigned to the minute scale, 5 time-domain features of each time window i min , namely the mean variance energy skewness and kurtosis and 2 derived features, namely the mean correlation coefficient of adjacent moments and the mean sequence trend factor are calculated using the exhaust temperature of all sampling points in each time window i m,k , and the minute feature F m,k,min of the IMF component u m,k is obtained;

[0029] For the IMF component u m,k assigned to the 10-minute scale, 4 time-domain features of each time window i 10min , namely the mean variance energy and kurtosis and 2 derived long-term trend features, namely the long-term standard deviation and the sequence cumulative sum and obtain the IMF component u m,k of the 10-minute feature F m,k,10min ;

[0030] Step 4.3, for each exhaust temperature sensor, the second feature F m,k,s of all IMF components allocated to the second scale is classified into the second feature set F m,s , the minute feature F m,k,min of all IMF components allocated to the minute scale is classified into the minute feature set F m,min , the 10-minute feature F m,k,10min of all IMF components allocated to the 10-minute scale is classified into the 10-minute feature set F m,10min , thereby obtaining the multi-scale feature set F m of each exhaust temperature sensor = {F m,s , F m,min , F m,10min};

[0031] The above m = 1, 2, …, M, M is the number of exhaust temperature sensors; k = 1, 2, …, K, K is the number of IMF components of each exhaust temperature sensor; i s = 1, 2, …, T s , T s is the number of second scale time windows, i min = 1, 2, …, T min , T min is the number of minute scale time windows, i 10min = 1, 2, …, T 10min , T 10min is the number of 10-minute scale time windows.

[0032] The specific process of the above step 6 is as follows:

[0033] Step 6.1, construct the correlation feature of the current exhaust temperature sensor m and other exhaust temperature sensors m ′

[0034]

[0035] Step 6.2, perform dimension compression on the correlation feature of the current exhaust temperature sensor m and other exhaust temperature sensors m' using 1x1 convolution, to obtain the dimension compression feature of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0036] Step 6.3, classify the dimension compression feature of the current exhaust temperature sensor m and other exhaust temperature sensors m' into the dimension compression feature set F ​​into a tanh function activation layer to obtain the temporal attention score function of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0037]

[0038] Step 6.4, the temporal attention score function of the current exhaust temperature sensor m and other exhaust temperature sensors m' into a softmax function activation layer to obtain the temporal attention weight of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0039]

[0040] Step 6.5, the temporal attention weight of the current exhaust temperature sensor m and other exhaust temperature sensors m' and the temporal delay feature of the current exhaust temperature sensor m and other exhaust temperature sensors m' into a weighted fusion layer to obtain the temporal attention feature of the current exhaust temperature sensor m

[0041]

[0042] Step 6.6, the relative distance feature of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0043]

[0044] Step 6.7, the relative distance feature of the current exhaust temperature sensor m and other exhaust temperature sensors m' ′ into a tanh function activation layer to obtain the spatial attention score function of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0045]

[0046] Step 6.8, the spatial attention score function of the current exhaust temperature sensor m and other exhaust temperature sensors m' into a softmax function activation layer to obtain the spatial attention weight of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0047]

[0048] ​Step 6.9, spatial attention weight of current exhaust temperature sensor m and other exhaust temperature sensors m' and relative distance feature of current exhaust temperature sensor m and other exhaust temperature sensors m' are fed into the weighted fusion layer to obtain the spatial attention feature of current exhaust temperature sensor m

[0049]

[0050] Step 6.10, dimension compression of spatial attention feature of current exhaust temperature sensor m by using 1x1 convolution to obtain temperature enhancement feature of current exhaust temperature sensor m and other exhaust temperature sensors m'

[0051] In the above formula, is the correlation feature of current exhaust temperature sensor m and other exhaust temperature sensors m', and concat(*) represents concatenation, is the temperature feature of current exhaust temperature sensor m, is the temperature feature of other exhaust temperature sensors m'; m, m' = 1, 2, …, M, and M is the number of exhaust temperature sensors; is the time attention score function of current exhaust temperature sensor m and other exhaust temperature sensors m', and tanh(*) represents the tanh function, tanh(*) m,m′ is the signal transmission time delay of current exhaust temperature sensor m and other exhaust temperature sensors m', and τ m,m′ = d m,m′ / v, d m,m′ is the relative distance of current exhaust temperature sensor m and other exhaust temperature sensors m', and v is the exhaust flow rate; W time1 is the score output weight matrix, U time1 is the feature transformation weight matrix, V time1 is the time delay weight matrix; is the time attention weight of current exhaust temperature sensor m and other exhaust temperature sensors m'; is the time attention feature of current exhaust temperature sensor m, is the time attention weight of current exhaust temperature sensor m and other exhaust temperature sensors m'; is the relative distance feature of current exhaust temperature sensor m and other exhaust temperature sensors m', is the time attention feature of other exhaust temperature sensors m'; is the spatial attention score function of current exhaust temperature sensor m and other exhaust temperature sensors m', and W sp1for spatial attention weight matrix, U sp1 for spatial attention input weight matrix, V sp for spatial attention distance weight matrix for spatial attention weight of current exhaust temperature sensor m and other exhaust temperature sensors m'; for spatial attention feature of current exhaust temperature sensor m.

[0052] Compared with the prior art, the present application has the following characteristics:

[0053] 1. The present application adopts a variational mode decomposition algorithm to separate the sensor temperature signal into different frequencies, and accurately maps it to different time scales according to the center frequency, captures transient high-frequency fluctuations on the second scale, reflects medium-term trends on the minute scale, and represents long-term stability on the 10-minute scale, combined with a time convolution network, solving the problem of poor adaptability of traditional methods to complex temperature dynamics.

[0054] 2. The present application adopts a dynamic threshold calibration strategy, which uses the average temperature at the current time or the previous time as a reference to replace abnormal data with a deviation of more than 15%, avoiding measurement distortion caused by single sensor failure. Compared with traditional filtering algorithms, this method does not rely on prior noise models and is more robust to random errors of the sensor array. Combined with the decomposition ability of VMD for nonlinear signals, the original temperature sequence is decomposed into multiple IMF components, effectively separating noise and real signals, and solving the problem of relying on artificial optimization of filtering parameters in the prior art.

[0055] 3. The present application utilizes the parallel convolution operation characteristics of the time convolution network, compared with the traditional Bayesian fusion algorithm, the present application realizes batch data processing through dilated causal convolution and residual connection, meeting the real-time sampling demand of 10Hz of the automobile engine.

[0056] 4. The present application introduces an attention mechanism containing a time lag factor τ and a sensor distance parameter d after the time convolution network, the model can automatically learn, the weight of the sensor with closer distance is higher, and more attention is allocated to the transient temperature mutation point in the time sequence, without manually presetting the importance of the sensor.

[0057] 5. The present application completely abandons the dependence on sensor noise distribution, prior confidence and other parameters in traditional methods, the model can automatically weaken the weight of abnormal sensors through multi-scale feature fusion, while traditional methods need to manually recalibrate parameters, improving the reliability of the system under complex working conditions, without the need for regular maintenance of algorithm parameters. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 It is a flowchart of a data fusion processing method for an automobile engine exhaust temperature sensor.

[0059] Figure 2 Structure diagram of the spatiotemporal attention mechanism network.

[0060] Figure 3 Total temperature data fusion comparison chart of exhaust temperature multi-sensor.

[0061] Figure 4 Total temperature data fusion chart of exhaust temperature multi-sensor.

[0062] Figure 5 10-minute segment temperature data fusion comparison chart of exhaust temperature multi-sensor.

[0063] Figure 6 10-minute segment temperature data fusion chart of exhaust temperature multi-sensor. DETAILED DESCRIPTION

[0064] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to specific examples and the accompanying drawings.

[0065] A data fusion processing method for an automobile engine exhaust temperature sensor, as shown in FIG. 1, specifically includes the following steps: Figure 1

[0066] Step 1, using M (M>1) exhaust temperature sensors installed at different positions of the automobile engine, collecting the exhaust temperature x m (m,t) of the automobile engine within a time T, where m=1,2,…,M, t=1,2,…,T.

[0067] In this embodiment, the number of exhaust temperature sensors of the automobile engine M=4, and the four exhaust temperature sensors are installed at both ends of the three-way catalyst and the particulate trap. The sampling frequency HZ and the sampling time length T of each exhaust temperature sensor are the same, where the sampling frequency HZ is an integer multiple of 1 Hz and is greater than or equal to 1 Hz, and the sampling time length T is an integer multiple of 10 minutes and is greater than or equal to 10 minutes. In this embodiment, the sampling frequency of the exhaust temperature sensor is 10 Hz, the sampling time length is 60 minutes, and the sampling point number is 36000 times.

[0068] Step 2, calculating the exhaust temperature average value x of all temperature sensors at each sampling point, and screening out the abnormal exhaust temperature x m (m,t) of the temperature sensor at the corresponding sampling point and performing calibration.

[0069] Step 2.1, calculating the exhaust temperature average value x

[0070] ​

[0071] Step 2.2, exhaust temperature x m (t) of each temperature sensor at sampling point t is compared with the average exhaust temperature at sampling point t to screen out abnormal temperature data and calibrate them:

[0072] When the difference between the exhaust temperature x m (t) of the temperature sensor at sampling point t and the average exhaust temperature at sampling point t is within the preset allowable range, such as , it indicates that the exhaust temperature x m (t) of the temperature sensor at sampling point t is normal, and no correction is needed. The exhaust temperature x m (t) is retained, i.e., x m (t)=x m (t).

[0073] When the difference between the exhaust temperature x m (t) of the temperature sensor at sampling point t and the average exhaust temperature at sampling point t is beyond the preset allowable range, such as , it indicates that the exhaust temperature x m (t) of the temperature sensor at sampling point t is abnormal, and needs to be corrected. The average exhaust temperature at the previous sampling point t-1 is used to replace the exhaust temperature x m (t), i.e.

[0074] Step 3, after the abnormal exhaust temperature data is processed, the variational mode decomposition algorithm is used to decompose the exhaust temperature collected by each exhaust temperature sensor into K IMF components u m,k and corresponding center frequencies ω m,k , where k=1,2,…,K.

[0075] Variational Mode Decomposition (VMD) is a signal processing method based on variational framework, aiming to decompose complex signals into multiple modal components with different center frequencies. After the abnormal data calibration at all sampling times is completed, the variational mode decomposition is performed on the complete time series of each sensor to extract the modal components and center frequencies. In this embodiment, for each temperature sensor m, the calibrated time series {x m (1), x m (2),…,x m(T)} as the input signal, the VMD is applied to obtain K (K > 1) IMF components of the intrinsic mode function, each of which is a time sequence {u m,k (1), u m,k (2), …, u m,k (T)}. In this embodiment, K = 4. Then, each IMF component u m,k of each exhaust temperature sensor is automatically solved by iteration optimization of the VMD algorithm, and the center frequency ω m,k of each IMF component is obtained.

[0076] Step 4, the K IMF components u m,k of each exhaust temperature sensor are distributed to the corresponding time scales according to the corresponding center frequencies ω m,k , and features are extracted in the time window corresponding to each time scale to obtain a multi-scale feature set F m of each exhaust temperature sensor.

[0077] Multi-scale feature extraction is based on the frequency characteristics of the signal, and the signal is distributed to different time scales, second, minute and 10-minute scales, and features that can represent the characteristics of the signal are extracted under each scale, and then the feature set is formed by fusion. The core is to use the sensitive difference of different time scales to the signal to comprehensively capture the transient, medium-term and long-term characteristics.

[0078] Step 4.1, time scale distribution and time window division: according to the center frequency ω m,k of the IMF component u m,k , the time scale and the window are divided as follows:

[0079] (1) When ω m,k ≥ 1 Hz, the IMF component u m,k is distributed to the second scale. In the second scale, the IMF component u m,k is divided by using a time window with a window size of 1 second, the number of time windows is T s , and the number of sampling points of exhaust temperature contained in each time window is N s . The second scale time window keeps one time step per second to ensure that the time resolution is consistent with the output frequency.

[0080] In this embodiment, the sampling frequency is 10 Hz, and the sampling time is 60 minutes, so the IMF component u m,k distributed to the second scale is divided into T s = 3600 time windows, and each time window contains N s = 10 sampling points.

[0081] (2) When 1 / 600Hz < ω m,k<1 Hz, the IMF component u m,k is assigned to the minute scale. In the minute scale, the IMF component u m,k is divided using time windows with a window size of 1 minute, and the number of time windows is T min , and each time window contains N min sample points of exhaust gas temperature. The minute scale time window is aligned with the minute scale, and the time granularity of the medium-term trend feature is enhanced.

[0082] In this embodiment, the sampling frequency is 10 Hz, and the sampling duration is 60 minutes. The IMF component u m,k assigned to the minute scale is divided into T min = 60 time windows, and each time window contains N min = 600 sample points.

[0083] (3) When ω m,k ≤ 1 / 600 Hz, the IMF component u m,k is assigned to the 10-minute scale. In the 10-minute scale, the IMF component u m,k is divided using time windows with a window size of 10 minutes, and the number of time windows is T 10min , and each time window contains N 10min sample points of exhaust gas temperature. The 10-minute scale time window is adapted to the 10-minute time scale, and covers the long-term features of the entire monitoring period.

[0084] In this embodiment, the sampling frequency is 10 Hz, and the sampling duration is 60 minutes. The IMF component u m,k assigned to the 10-minute scale is divided into T 10min = 6 time windows, and each time window contains N 10min = 6000 sample points.

[0085] Step 4.2, window-based feature extraction: for each IMF component u m,k of each exhaust gas temperature sensor, according to the time scale to which it is assigned, the features within the corresponding time window are extracted;

[0086] (1) For the IMF component u m,k assigned to the second scale, 4 time-domain features of each time window are calculated using the exhaust gas temperature of N s sample points in each time window i s , and the second feature of the IMF component u m,k is obtained. where D s= 4 is the number of second features, which are used to enhance the high-frequency feature extraction ability and support the accurate modeling of transient temperature changes.

[0087] ① The 4 time-domain features include:

[0088] the mean of the time window i s

[0089]

[0090] the variance of the time window i s

[0091]

[0092] the skewness of the time window i s

[0093]

[0094] the peak amplitude of the time window i s

[0095]

[0096] where i s = 1, 2, …, T s , T s is the number of time windows in the second scale, N s is the number of sampling points in each time window in the second scale, (i s - 1) × N s is the starting position of the time window i s , and n is the offset.

[0097] (2) For the IMF component u m,k assigned to the minute scale, 5 time-domain features and 2 derived features are calculated using the exhaust gas temperature of N min sampling points in each time window i min , and the minute features of the IMF component u m,k are obtained. where D min = 7 is the number of minute features, which are used to balance the calculation amount of medium-term features and the representation accuracy.

[0098] ① The 5 time-domain features include:

[0099] the mean of the time window i min ​​​​​

[0100]

[0101] Time window i min Variance of

[0102]

[0103] Time window i min Energy of

[0104]

[0105] Time window i min Skewness of

[0106]

[0107] Time window i min Kurtosis of

[0108]

[0109] ②2 derived features include:

[0110] Time window i min Adjacent time point mean correlation coefficient of

[0111]

[0112] Time window i min Mean sequence trend factor of

[0113]

[0114] In the formula, i min = 1, 2, …, T min , T min is the number of time windows on the minute scale, N min is the number of sampling points of each time window on the minute scale, (i min - 1) × N min is the starting position of time window i min , n is the offset; μ f is the mean value of all time windows on the minute scale is the mean value of time window i min - 1.

[0115] (3) For the IMF component u of the exhaust temperature sensor assigned to the 10-minute scale​m,k , the exhaust temperature of each of the N 10min sampling points within each time window i 10min is used to calculate 4 time-domain features and 2 derived long-term trend features for each time window, and 10-minute features of the IMF component u m,k are obtained. where D 10min = 6 is the number of 10-minute features used to preserve key statistics of long-term steady features and improve the stability of the fusion results.

[0116] ① The 4 time-domain features include:

[0117] the mean of the time window i 10min

[0118]

[0119] the variance of the time window i 10min

[0120]

[0121] the energy of the time window i 10min

[0122]

[0123] the kurtosis of the time window i 10min

[0124]

[0125] ② The 2 derived long-term trend features include:

[0126] the long-term standard deviation of the time window i 10min

[0127]

[0128] the sequence cumulative sum of the time window i 10min

[0129]

[0130] where i 10min = 1, 2, …, T 10min , T 10min is the number of 10-minute scale time windows, and N 10min is the number of sampling points of each 10-minute scale time window.​​​​​​10min -1)×N 10min For time window i 10min The starting position, where n is the offset; μ m,k,n This represents the average of the nth 10-minute time window; From the first time window to the i-th time window 10min Mean of a 10-minute window sequence of means

[0131]

[0132] It is important to note that if an exhaust temperature sensor has two IMF components assigned to the same time scale, the average of the characteristic values ​​for the same feature should be used as the characteristic value for that feature. For example, if an exhaust temperature sensor has two IMF components assigned to the second scale, the mean, variance, skewness, and peak amplitude of each IMF component should be calculated separately. Then, the average of the means of the two IMF components should be used as the mean of the second feature, the average of the variances of the two IMF components should be used as the variance of the second feature, the average of the skewnesses of the two IMF components should be used as the skewness of the second feature, and the average of the peak amplitudes of the two IMF components should be used as the peak amplitude of the second feature. If an exhaust temperature sensor has no IMF components assigned to a certain time scale, all features corresponding to that time window should be set to 0. For example, if an exhaust temperature sensor has 0 IMF components assigned to the minute scale, then all 5 time-domain features and 2 derived features should be set to 0.

[0133] Step 4.3, Construction of Multi-Scale Feature Set: For each exhaust temperature sensor, construct the second feature F of all IMF components assigned to the second scale. m,k,s Substitute into the second feature set F m,s In the middle, the minute features F of all IMF components assigned to the minute scale m,k,min Included in the minute feature set F m,min In the middle, the 10-minute features F of all IMF components assigned to the 10-minute scale m,k,10min Included in the 10-minute feature set F m,10min Thus, the multi-scale feature set F of each exhaust temperature sensor is obtained. m ={F m,s ,F m,min ,F m,10min}

[0134] Step 5: Assemble the multi-scale feature set F for each exhaust temperature sensor. m The feature sets F at various time scales m,s ,F m,min ,F m,10minThe data are fed into a temporal convolutional network to obtain the temperature features X3 of each exhaust temperature sensor at different time scales.

[0135] The multi-scale feature set F of the exhaust temperature sensor m m ={F m,s ,F m,min ,F m,10mim The input is used as the input to the temporal convolutional network. A temporal convolutional network (TCN) is a convolutional neural network specifically designed for temporal data. It consists of dilated, causal 1D convolutional layers with the same input and output lengths, designed to capture temporal dependencies through convolutional operations. In this embodiment, the temporal convolutional network consists of two dilated causal convolutions and a residual layer.

[0136] Step 5.1, the first dilated causal convolution output X1 is:

[0137]

[0138] In the formula, b1 is the weight matrix when the expansion factor d1 = 1, b1 is the bias vector, BatchNorm is the batch normalization operation, tanh is the hyperbolic tangent activation function, Dropout is the random deactivation operation to prevent overfitting, C1 = 32 output channels, and p1 = 0.1 dropout rate.

[0139] For the second-scale feature set F m,s After applying the first dilated causal convolution, we obtain Its channel count has expanded from 4 dimensions to 32 dimensions.

[0140] For the minute-scale feature set F m,min After applying the first dilated causal convolution, we obtain Its channel count has expanded from 7 dimensions to 32 dimensions.

[0141] For the 10-minute scale feature set F m,10min After applying the first dilated causal convolution, we obtain Its channel count has expanded from 6 dimensions to 32 dimensions.

[0142] Step 5.2, the output X2 of the second dilated causal convolution is:

[0143]

[0144] In the formula, X1 is the output tensor of the first dilated causal convolution. W2is the weight matrix, b2is the bias vector, BatchNorm is the batch normalization operation, tanh is the hyperbolic tangent activation function, Dropout is the random inactivation operation to prevent overfitting, the output channel C2= 64, and the dropout rate p2= 0.2.

[0145] For the second scale feature set F m,s After applying the second dilated causal convolution, we get The number of channels is expanded from 32 dimensions to 64 dimensions.

[0146] For the second scale feature set F m,min After applying the second dilated causal convolution, we get The number of channels is expanded from 32 dimensions to 64 dimensions.

[0147] For the second scale feature set F m,10min After applying the second dilated causal convolution, we get The number of channels is expanded from 32 dimensions to 64 dimensions.

[0148] Step 5.3, the residual operation output, i.e., the temperature feature X3, is:

[0149] X3= tanh(X2+ X res )

[0150] In the formula, T * is the number of window for the corresponding time scale, T s is the number of time window for the second scale, T min is the number of time window for the minute scale, T 10min is the number of time window for the 10-minute scale.

[0151] In the second scale:

[0152] X res,s = W s1×1 * F m,s

[0153] In the formula, is a 1x1 convolution kernel for adjusting the feature dimension, * is a convolution operation, F m,s is the input tensor corresponding to the second scale feature set.

[0154] In the minute scale:

[0155] X res,min = W min×1 * F m,min

[0156] In the formula, is a 1x1 convolution kernel for adjusting the feature dimension, * is a convolution operation, F m,minis the input tensor corresponding to the 10-minute scale feature set.

[0157] In the 10-minute scale:

[0158] X res,10min = W 10min1×1 * F m,10min

[0159] wherein, is a 1x1 convolution kernel for adjusting the feature dimension, * is a convolution operation, and F m,10min is the input tensor corresponding to the 10-minute scale feature set.

[0160] Step 6, the temperature features X3 of each exhaust temperature sensor at different time scales are input into the spatio-temporal attention mechanism network to obtain temperature enhanced features of each exhaust temperature sensor at different time scales

[0161] The basic principle of the attention mechanism is to calculate the similarity or correlation between the query and the key, obtain the weight coefficient of the value corresponding to each key, and then perform weighted summation on the value to obtain the final attention value. The spatio-temporal attention mechanism network combines time delay and spatial distance, and by simultaneously introducing bidirectional time context information in the spatial attention mechanism, it combines the relative distance of sensors to more comprehensively explore the spatio-temporal dependence relationship between sensors to enhance the attention to key spatio-temporal information. The principle is as shown in Figure 2 , and the specific process is as follows:

[0162] Step 6.1, the related features of the current exhaust temperature sensor m and other exhaust temperature sensors m' are constructed

[0163]

[0164] wherein, concat(*) represents concatenation. is the temperature feature of the current exhaust temperature sensor m, is the temperature feature of the other exhaust temperature sensor m', m and m' are the index values of the exhaust temperature sensors, m, m' = 1, 2, …, M, and M is the number of exhaust temperature sensors.

[0165] Step 6.2, the related features of the current exhaust temperature sensor m and other exhaust temperature sensors m' are dimensionally compressed using a 1x1 convolution to obtain dimensionally compressed features of the current exhaust temperature sensor m and other exhaust temperature sensors m'

[0166] ​

[0167] where conv 1×1 (*) denotes 1 x 1 convolution, out channels = C2denotes the output channel number is C2.

[0168] Step 6.3, dimension compression feature of current exhaust temperature sensor m and other exhaust temperature sensors m' into the tanh function activation layer, to get the time attention score function of current exhaust temperature sensor m and other exhaust temperature sensors m'

[0169]

[0170] where τ m,m′ is the signal transmission time delay of current exhaust temperature sensor m and other exhaust temperature sensors m', τ m,m′ = d m,m′ / v, d m,m′ is the relative distance of current exhaust temperature sensor m and other exhaust temperature sensors m', v is the exhaust flow rate; tanh(*) denotes the tanh function, W time1 is the score output weight matrix, U time1 is the feature transformation weight matrix, V time1 is the time delay weight matrix,

[0171] Step 6.4, the time attention score function of current exhaust temperature sensor m and other exhaust temperature sensors m' into the softmax function activation layer, to get the time attention weight of current exhaust temperature sensor m and other exhaust temperature sensors m'

[0172]

[0173] Step 6.5, the time attention weight of current exhaust temperature sensor m and other exhaust temperature sensors m' and the time delay feature of current exhaust temperature sensor m and other exhaust temperature sensors m' into the weighted fusion layer, to get the time attention feature of current exhaust temperature sensor m

[0174]

[0175] Step 6.6, construct the relative distance feature of current exhaust temperature sensor m and other exhaust temperature sensors m v ​

[0176]

[0177] In the formula, concat(*) represents concatenation, and d m,m′ The distance between the current exhaust temperature sensor m and other exhaust temperature sensors m′ is given.

[0178] Step 6.7: Determine the relative distance characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′. The input is fed into the tanh function activation layer to obtain the spatial attention score function between the current exhaust temperature sensor m and other exhaust temperature sensors m′.

[0179]

[0180] In the formula, W sp1 Here is the spatial attention weight matrix. U sp1 Input the weight matrix for spatial attention. V sp1 This is the spatial attention distance weight matrix.

[0181] Step 6.8: Combine the spatial attention score function of the current exhaust temperature sensor m with that of other exhaust temperature sensors m′. The data is fed into the softmax function activation layer to obtain the spatial attention weights between the current exhaust temperature sensor m and other exhaust temperature sensors m′.

[0182]

[0183] Step 6.9: Assign spatial attention weights to the current exhaust temperature sensor m and other exhaust temperature sensors m′. The relative distance characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′ The data are fed into the weighted fusion layer to obtain the spatial attention features of the current exhaust temperature sensor m.

[0184]

[0185] Step 6.10: Utilize 1×1 convolution to apply the spatial attention features of the current exhaust temperature sensor m, i.e., the exhaust temperature enhancement features. Dimensional compression is performed to obtain the temperature enhancement characteristics of the current exhaust temperature sensor m and other exhaust temperature sensors m′.

[0186]

[0187] In the second scale, the temperature enhancement feature In the minute scale, the temperature enhancement feature In the 10-minute scale, the temperature enhancement feature

[0188] Step 7, using a linear interpolation function to obtain the temperature enhancement feature of each exhaust temperature sensor in different time scales Perform dimension alignment to obtain the temperature alignment feature H of each exhaust temperature sensor in different time scales s,unsampled , H min,unsampled , H 10min,unsampled .

[0189] Linear interpolation refers to an interpolation method in which the interpolation function is a first-order polynomial, and the interpolation error at the interpolation nodes is zero. In the present application, the features in the minute scale and the 10-minute scale are unified to the time resolution of the second scale, ensuring that the features in different scales can be directly fused in the time dimension, i.e.

[0190]

[0191] In the formula, LinearInterp(*) represents a linear interpolation function.

[0192] Step 8, dynamically assign time scale weights according to the importance of features in different time scales, and perform weighted summation of the temperature alignment features of each exhaust temperature sensor in different time scales using the time scale weights to obtain the temperature fusion feature of each exhaust temperature sensor

[0193] Step 8.1, dynamically assign weights for fusion according to the importance of features in different time scales to improve the representation ability of the model to temperature changes:

[0194] α = Softmax(w)

[0195] In the formula, Softmax(*) represents a Softmax function. w = [w s , w min , w 10min ] is a learnable parameter vector, and α = [α s , α min , α 10min ] is a learnable scale weight.

[0196] Step 8.2, weighting based on assigned weights to obtain the temperature fusion feature

[0197] where α s +α min +α 10min =1,

[0198] Step 9, the temperature fusion features of all exhaust temperature sensors are integrated into an average pooling layer to obtain the multi-sensor temperature integration features H pool , and the multi-sensor temperature integration features H pool are mapped into the final temperature value Y out .

[0199] Step 9.1, the temperature fusion features of all exhaust temperature sensors are integrated :

[0200]

[0201] Step 9.2, the multi-sensor temperature integration features H pool are mapped and output:

[0202] Y out = H pool · W out + b out

[0203] wherein, W is a weight, b out ∈R 1 is a bias term, and output

[0204] The multi-sensor information is integrated, the fused temperature value is finally output, the multi-scale information is complementary, and the result accuracy is improved.

[0205] The application firstly collects engine exhaust temperature signals through multiple exhaust temperature sensors in different positions, and uses the average value method to correct abnormal value data; then the corrected data is decomposed into multiple intrinsic mode functions through variational mode decomposition, divided into components of different time scales (second, minute and 10 minute scale) based on the center frequency of the modal component, and a customized feature set (time domain feature and derived feature) is extracted. Then an improved attention mechanism time convolution network with physical constraints is adopted to fuse the sensor distance and time delay parameters, realize multi-scale spatio-temporal feature weighted fusion. Finally, the final fused temperature value is output through weighted pooling and full connection. The method of multi-scale spatio-temporal feature fusion based on time convolution network is adopted to assign weights to the data collected and processed by multiple temperature sensors, which can adapt to the random changes of sensor measurement data, capture the long-term dependence in time series data, and better understand the importance of sensor data at different time points. The method significantly improves the accuracy and robustness of exhaust temperature monitoring, solves the fusion deviation problem caused by the spatio-temporal asynchrony of multi-sensor data, and is particularly suitable for engine exhaust temperature state real-time monitoring system.

[0206] The four temperature sensors collected by the method of the application are processed, and the processing results are shown in Figure 3 and Figure 4 It can be seen that the measurement results of each temperature sensor have relatively large fluctuations, and the engine exhaust temperature processed by the time convolution network multi-scale spatio-temporal feature fusion is more stable and accurate, eliminating random noise interference and random fluctuations of measurement results of different position sensors, and more accurately reflecting the change of engine exhaust temperature. In order to show the results of different time scales, the four temperature sensors and the fusion processing results collected for 10 minutes are plotted as shown in Figure 5 and Figure 6 It can be seen that at the time scale of 10 minutes, the change of temperature is more stable, and the change trend of engine exhaust temperature can be more accurately reflected.

[0207] It should be noted that although the above embodiments of the application are illustrative, this is not a limitation of the application, therefore the application is not limited to the above specific embodiments. Any other embodiments obtained by those skilled in the art under the inspiration of the application without departing from the principles of the application are considered to be within the protection scope of the application.

Claims

1. A method for data fusion processing of automotive engine exhaust temperature sensor data, characterized in that, The steps include the following: Step 1: Use two or more exhaust temperature sensors installed at different locations on the car engine to collect the exhaust temperature of the car engine over a period of time. Step 2: Calculate the average exhaust temperature of all temperature sensors at each sampling point, and based on this, screen out the abnormal exhaust temperatures of the temperature sensors at the corresponding sampling points and calibrate them. Step 3: Use variational mode decomposition algorithm to perform variational mode decomposition on the exhaust temperature collected by each exhaust temperature sensor to obtain more than two IMF components and their corresponding center frequencies for each exhaust temperature sensor. Step 4: Assign each IMF component of each exhaust temperature sensor to the corresponding time scale according to its corresponding center frequency, and extract features within the time window corresponding to each time scale to obtain the multi-scale feature set of each exhaust temperature sensor. Step 5: Input the feature sets of each time scale of the multi-scale feature set of each exhaust temperature sensor into the temporal convolutional network to obtain the temperature features of each exhaust temperature sensor at different time scales. Step 6: Input the temperature features of each exhaust temperature sensor at different time scales into the spatiotemporal attention mechanism network to obtain the temperature enhancement features of each exhaust temperature sensor at different time scales; Step 7: Use a linear interpolation function to dimensionally align the temperature enhancement features of the exhaust temperature sensor at different time scales to obtain the temperature alignment features of each exhaust temperature sensor at different time scales. Step 8: Dynamically assign time scale weights according to the importance of features at different time scales, and use the time scale weights to perform weighted summation of the temperature alignment features of each exhaust temperature sensor at different time scales to obtain the temperature fusion features of each exhaust temperature sensor. Step 9: Perform average pooling to integrate the temperature fusion features of all exhaust temperature sensors to obtain multi-sensor temperature integration features, and map the multi-sensor temperature integration features to the final exhaust temperature value.

2. The method for data fusion processing of automotive engine exhaust temperature sensor data according to claim 1, characterized in that, In step 1, the sampling frequency of the exhaust temperature sensor is an integer multiple of 1 Hz and greater than or equal to 1 Hz, and the sampling duration is an integer multiple of 10 minutes and greater than or equal to 10 minutes.

3. The method for data fusion processing of automotive engine exhaust temperature sensor data according to claim 1, characterized in that, In step 2, the specific process of screening out the abnormal exhaust temperatures of each temperature sensor at the corresponding sampling point and calibrating them is as follows: When the temperature sensor detects the exhaust temperature x at sampling point t m (t) and the average exhaust temperature at sampling point t If the difference is within the preset allowable range, then the exhaust temperature x is retained. m (t); When the temperature sensor detects the exhaust temperature x at sampling point t m (t) and the average exhaust temperature at sampling point t If the temperature exceeds the preset allowable range, the average exhaust temperature at the previous sampling point t-1 will be used. Replace this exhaust temperature x m (t); In the above, m = 1, 2, ..., M, where M is the number of exhaust temperature sensors, and t = 1, 2, ..., T, where T is the sampling duration.

4. The method for data fusion processing of automotive engine exhaust temperature sensor data according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: Convert the individual IMF components u of each exhaust temperature sensor m,k According to its corresponding center frequency ω m,k Assigned to the corresponding time scale, that is: When ω m,k When the frequency is ≥1Hz, the IMF component of the exhaust temperature sensor will be u m,k Allocation to a second scale; In the second scale, a time window of size 1 second is used to evaluate the IMF components u. m,k The time windows are divided into segments, with the number of segments being T. s ; When 1 / 600Hz < ω m,k When <1Hz, the IMF component u of the exhaust temperature sensor will be... m,k Assign to a minute scale; within the minute scale, use a time window with a window size of 1 minute for the IMF component u. m,k The time windows are divided into segments, with the number of segments being T. min ; When ω m,k When the frequency is ≤1 / 600Hz, the IMF component u of the exhaust temperature sensor will be... m,k Assigned to a 10-minute scale; within the 10-minute scale, IMF components are assigned using a time window with a window size of 10 minutes. m,k The time windows are divided into segments, with the number of segments being T. 10min ; Step 4.2: For each IMF component u of each exhaust temperature sensor m,k Based on the allocated time scale, features within the corresponding time window are extracted; For IMF components allocated to the second scale u m,k Utilizing each time window i s The exhaust temperature of all sampling points within the range is calculated, and the mean value of the four time-domain features of each time window is obtained. variance Skewness and peak amplitude And obtain the IMF component u m,k The second feature F m,k,s ; For IMF components allocated to the minute scale u m,k Utilizing each time window i min The exhaust temperature of all sampling points within the range is calculated, and the mean value of five time-domain features for each time window is obtained. variance energy Skewness and kurtosis And two derived features, namely the correlation coefficient of the mean values ​​at adjacent time points. and mean series trend factor And obtain the IMF component u m,k minute feature F m,k,min ; For IMF components assigned to a 10-minute scale, u m,k Utilizing each time window i 10min The exhaust temperature of all sampling points within the range is calculated, and the mean value of the four time-domain features of each time window is obtained. variance energy and kurtosis And two derived long-term trend characteristics, namely long-term standard deviation. and sequence cumulative sum And obtain the IMF component u m,k 10-minute characteristic F m,k,10min ; Step 4.3: For each exhaust temperature sensor, calculate the second characteristic F of all IMF components assigned to the second scale. m,k,s Substitute into the second feature set F m,s In the middle, the minute features W of all IMF components assigned to the minute scale m,k,min Included in the minute feature set F m,min In the middle, the 10-minute features F of all IMF components assigned to the 10-minute scale m,k,10min Included in the 10-minute feature set F m,10min Thus, the multi-scale feature set F of each exhaust temperature sensor is obtained. m ={F m,s ,F m,min ,F m,10min }; In the above, m = 1, 2, ..., M, where M is the number of exhaust temperature sensors; k = 1, 2, ..., K, where K is the number of IMF components for each exhaust temperature sensor; i s =1,2,…,T s T s i is the number of time windows on a second scale. min =1,2,…,T min T min i represents the number of time windows on a minute scale. 10min =1,2,…,T 10min T 10min This represents the number of time windows on a 10-minute scale.

5. The method for data fusion processing of automotive engine exhaust temperature sensor data according to claim 1, characterized in that, The specific process of step 6 is as follows: Step 6.1: Construct the correlation features between the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.2: Use 1×1 convolution to analyze the correlation features between the current exhaust temperature sensor m and other exhaust temperature sensors m′. Dimensional compression is performed to obtain the dimensional compression features of the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.3: Compress the dimensionality of the current exhaust temperature sensor m with that of other exhaust temperature sensors m′. The data is fed into the tanh function activation layer to obtain the time attention score function between the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.4: Combine the time attention score function of the current exhaust temperature sensor m with that of other exhaust temperature sensors m′. The weights are fed into the softmax function activation layer to obtain the time attention weights of the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.5: Weight the time attention of the current exhaust temperature sensor m with that of other exhaust temperature sensors m′. The time delay characteristics of the current exhaust temperature sensor m and other exhaust temperature sensors m′ The data is fed into the weighted fusion layer to obtain the time attention features of the current exhaust temperature sensor m. Step 6.6: Construct the relative distance characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.7: Determine the relative distance characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′. The input is fed into the tanh function activation layer to obtain the spatial attention score function between the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.8: Combine the spatial attention score function of the current exhaust temperature sensor m with that of other exhaust temperature sensors m′. The data is fed into the softmax function activation layer to obtain the spatial attention weights between the current exhaust temperature sensor m and other exhaust temperature sensors m′. Step 6.9: Assign spatial attention weights to the current exhaust temperature sensor m and other exhaust temperature sensors m′. The relative distance characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′ The data are fed into the weighted fusion layer to obtain the spatial attention features of the current exhaust temperature sensor m. Step 6.10: Utilize 1×1 convolution to apply spatial attention features of the current exhaust temperature sensor m. Dimensional compression is performed to obtain the temperature enhancement characteristics of the current exhaust temperature sensor m and other exhaust temperature sensors m′. In the above formula, This section describes the correlation characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′. `concat(*)` indicates concatenation. The temperature characteristics of the current exhaust temperature sensor m. Temperature characteristics of other exhaust temperature sensors m′; m,m′=1,2,…,M, where M is the number of exhaust temperature sensors; Let tanh(*) be the time attention score function between the current exhaust temperature sensor m and other exhaust temperature sensors m′, and let τ be the tanh function. m,m′ τ is the signal transmission time delay between the current exhaust temperature sensor m and other exhaust temperature sensors m′. m,m′ =d m,m′ / v,d m,m′ W represents the relative distance between the current exhaust temperature sensor m and other exhaust temperature sensors m′, where v is the exhaust flow velocity; time1 Output the weight matrix U for the score. time1 V is the feature transformation weight matrix. time1 This is the time delay weight matrix; The time attention weights of the current exhaust temperature sensor m and other exhaust temperature sensors m′; The time attention characteristics of the current exhaust temperature sensor m, The time attention weights of the current exhaust temperature sensor m and other exhaust temperature sensors m′; The relative distance characteristics between the current exhaust temperature sensor m and other exhaust temperature sensors m′ are given. For the time attention characteristics of other exhaust temperature sensors m′; W is the spatial attention score function between the current exhaust temperature sensor m and other exhaust temperature sensors m′. sp U is the spatial attention weight matrix. sp Input the weight matrix V for spatial attention. sp This is the spatial attention distance weight matrix; The spatial attention weights of the current exhaust temperature sensor m and other exhaust temperature sensors m′ are used. The spatial attention characteristics of the current exhaust temperature sensor m.

Images