A method for overload and eccentric load detection based on lightweight network
By installing sensors on rails and designing lightweight neural networks to extract the axle and bogie features, the problem of low detection accuracy in the prior art is solved, higher detection accuracy and lower risk of false detection, and reduced Risks caused by excessive loading.
Patent Information
- Application Number
- CN202410767745.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-14
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-06-14
AI Technical Summary
The existing track hyper-biased detection methods have problems such as false detection, missed detection and poor detection results when dealing with different vehicle models, speeds, directions and loads, resulting in low detection accuracy.
Using the ultra-biased load detection method based on lightweight network, by installing shear force and pressure sensors on the rails, collecting data for fusion, designing a lightweight neural network structure, extracting the axle and bogie features, and calculating the ultra-biased load results of the carriage , reduce the risk of false detection and missed detection.
It improves the accuracy of axle feature detection, improves the accuracy of cabin detection, and reduces the risks caused by excessive loading.
Smart Images

Figure CN118568596B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of track overload and eccentric load detection and early warning, and specifically is an overload and eccentric load detection method based on a lightweight network. Background Art
[0002] In recent years, with the continuous increase in train speed, safety accidents have occurred from time to time. Train overloading and unbalanced loading are important factors that cause vehicle derailment and affect railway driving safety. The multi-element detection and measurement device combining pressure and shear force plays a vital role in the intelligent and efficient detection of overloading and unbalanced loading of freight trains due to its advantages of high precision and simple installation.
[0003] In the existing detection methods, when detecting the axle characteristics on the shear force and pressure sensors, due to the different loading rack devices of different models, the same model will present different characteristics at different speeds, directions, and loads; and when calculating overload and unbalanced loads, the existing methods cannot accurately adapt to the actual diverse situations when processing the sensor signal characteristics, resulting in many false detections and missed detections in the axle characteristics, and causing large errors in measuring overload and unbalanced loads, which leads to poor detection results. Summary of the invention
[0004] In view of the above-mentioned problems and shortcomings, in order to solve the problem of low detection accuracy of existing track overload and eccentric load detection methods, the present invention provides an overload and eccentric load detection method based on a lightweight network.
[0005] A method for detecting overload and eccentric load based on a lightweight network, the specific steps are as follows:
[0006] Step 1: Install 4 shear force sensors and 4 pressure sensors in each measurement area set on the rail, and collect data from multiple vehicle models as a data set. The specific installation method is as follows: Figure 1 As shown; wherein each measurement area divides the two tracks into an upper track and a lower track, the shear force sensors installed on the upper track are named S1 and S2, and the pressure sensors are named P1 and P2; the shear force sensors installed on the lower track are named S3 and S4, and the pressure sensors are named P3 and P4.
[0007] In the same measurement area, each sensor collects one-dimensional signal data in time series, so the data on the eight sensors corresponds to a group of data in the measurement area. The data set is expressed as:
[0008] Datas={D1,D2,…,D n}
[0009]
[0010]
[0011] Among them, Datas is the data set collected by the sensors in the measurement area, D1, D2, ..., D n It means there are n groups of data in the data set, i means the i-th sequence in n groups of data, D i represents the i-th group of data structures, represents the i-th data sequence collected by shear sensor S1, Respectively represent the m amplitude data in the sequence; represents the i-th data sequence collected by shear sensor S2, represents the m amplitudes in the sequence; represents the i-th data sequence collected by shear sensor S3, represents the m amplitudes in the sequence; represents the i-th data sequence collected by shear sensor S4, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P1, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P2, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P3, represents the m amplitudes of the sequence; Represents the i-th data sequence collected by pressure sensor P4.
[0012] Step 2: Perform data fusion based on the shear force and pressure sensor data of the upper and lower tracks collected in step 1.
[0013] The data collected by the shear and pressure sensors on the upper and lower tracks are fused separately to obtain the peak shape of the weight borne by each axle, and the horizontal and vertical coordinate positions of the peaks are extracted and recorded as axle features; the horizontal axis represents the time series of the sensor data collected, and the vertical axis represents the amplitude corresponding to each time series. The fusion method for the pressure and shear data of the same group is as follows:
[0014]
[0015] Among them, t is the number of amplitude sequences in the same set of data, t∈[1,m], Fd top (t) represents the t-th amplitude of the upper orbit fusion, represents the t-th data amplitude in the i-th data set of shear sensor S1, represents the t-th data amplitude in the i-th group of data of pressure sensor P1, represents the t-th data amplitude in the i-th group of data of pressure sensor P2, Fd represents the tth data amplitude in the i-th data set of shear sensor S2; down (t) represents the tth amplitude after the fusion of the lower track, represents the t-th data amplitude in the i-th data set of shear sensor S3, represents the t-th data amplitude in the i-th group of data of the pressure sensor P3, represents the t-th data amplitude in the i-th group of data of pressure sensor P4, Represents the amplitude of the tth data in the i-th group of data of the shear sensor S4.
[0016] According to common sense of train structure, the same carriage consists of two bogies on the upper track or the lower track respectively. A bogie usually has 2-4 axles, and according to the different loads of each carriage on the train, the carriages are divided into light carriages and heavy carriages. The axle features of the fused data of each axle on the light carriage have a large amplitude of the vertical coordinate, and the axle features of the fused data of each axle on the heavy carriage have a small amplitude of the vertical coordinate. In traditional methods, the extraction of axle features of different weights is often affected, and there are many cases of false detection and missed detection.
[0017] Step 3: Label the one-dimensional shear force and pressure sensor signals, and divide the labeled data set into training set, validation set and test set in proportion.
[0018] Different from the annotation of image data, each set of data in the data set is one-dimensional data. The two key features of axle feature and bogie feature are annotated on the fused data obtained in step 2, such as Figure 6 The coordinates of the red points in the middle represent the axle features. The axle features in the same bogie constitute the bogie features.
[0019] The specific marking method is as follows:
[0020] Feature(peak)={P1,P2,…,P l}
[0021] Feature(bogie)={B1,B2,…,B k}
[0022] Among them, Feature(peak) represents the axle feature set, each element is a coordinate, P1 represents the feature of the first axle, P2 represents the feature of the second axle, P l Indicates the feature of the lth axle, there are l in total; Feature(bogie) represents the bogie feature set, each bogie feature contains 2-4 axle features, B1 represents the first bogie feature, B2 represents the second bogie feature, B k Represents the k-th bogie feature, there are k of them; 2k≤l.
[0023] Step 4: Design a lightweight neural network structure and train the model with the dataset annotated in step 3 to extract axle features (i.e., the peaks on the axle data). Figure 7 shown) is:
[0024] Input layer: The horizontal axis x_axiel set and the vertical axis y_axiel. The horizontal axis x_axiel set is the time series of data collected by 8 sensors in the measurement area. The vertical axis y_axiel is the amplitude sequence fused according to the method in step 1 in the upper or lower track corresponding to each sampling time point in the measurement area. x_axiel and y_axiel form a matrix of size 2×T, where T is the length of the time series.
[0025] Hidden layer: Each time series collected is only associated with L consecutive data, where L is the current collection time point and the length of time with continuity. To ensure the accuracy of the extracted features, the hidden layer extracts features through 10 consecutive convolutional layers and activation layers (ReLU). The calculation method of L is as follows:
[0026]
[0027] Among them, Len is the maximum length of the train passing through the measurement area, V is the maximum speed of the train currently passing through the measurement area, and Fre is the sensor acquisition frequency in the current measurement area. The acquisition frequencies of the eight sensors are the same.
[0028] The expression formula of ReLU is:
[0029] f(x)=max(0,w T x+b)
[0030] Where x is the vector of the input layer, f(x) is the result after the x vector is activated, max represents the maximum value function, and w T represents the transpose of the matrix w, w represents the weight matrix, and b represents the offset matrix.
[0031] Output layer: Through the feature vector extracted by the hidden layer, the probability of each element for the 0-class label label_0 and the 1-class label label_1 is calculated, and the final classification result is calculated through the Softmax function. The expression of the Softmax function is as follows:
[0032]
[0033] Among them, f(0) represents the probability of the category being label 0, f(1) represents the probability of the category being label 1, e represents the exponential function, y0 represents the probability distribution of the category being label 0, and y1 represents the probability distribution of the category being label 1. This lightweight neural network is trained with the dataset annotated in step 3 to obtain a model for detecting axle features, thereby realizing axle feature extraction.
[0034] Step 5: Use the axle features extracted in step 4 as a new input layer, and then train the model for extracting bogie features through the lightweight neural network described in step 4 to achieve bogie feature extraction.
[0035] The bogie feature extraction process is the same as the peak calculation process. The axle feature (peak feature) result is used as the new input layer. The bogie feature is calculated through 3 hidden layers and ReLU layers. Then the output layer calculates the feature probability of the same bogie. Label_2 indicates that it belongs to the same bogie, and label_3 indicates that it is not the same bogie. Finally, after passing through the Softmax layer, the final bogie feature result is obtained.
[0036] Feature(bogie) new = {B 1_new ,B 2_new ,…}
[0037] Among them, B 1_new and B 2_new It is the bogie feature, and each bogie feature is a collection of 2-4 axle features.
[0038] Step 6: Extract the features of each carriage based on the axle features obtained in step 4 and the bogie features obtained in step 5.
[0039] A carriage consists of two bogies adjacent to the upper and lower tracks. Therefore, the characteristics of the carriage composed of four bogies adjacent to the upper and lower tracks are expressed as follows:
[0040] Feature(car)={{B 1_top ,B 2_top ,B 1_blow ,B 2_blow},{B 3_top ,B 4_top ,B 3_blow ,B 4_blow}…}
[0041] Among them, Feature(car) is the feature of the car, B 1_top and B 2_too The first and second bogie features detected by the upper track fusion data, B 1_blow ,B 2_blowrepresents the first and second bogie features detected by the lower track fusion data, B 3_top and B 4_top The third and fourth bogie features detected by the upper track fusion data; B 1_blow ,B 2_blow represents the first and second bogie features detected by the lower track fusion data, B 3_blow ,B 4_blow It represents the 3rd and 4th bogie features detected by the lower track fusion data. Therefore, the number of upper track or lower track bogie features detected for a train is twice the number of carriage features.
[0042] Step 7: Based on the model trained in steps 4-5, the axle and bogie features are quickly and accurately extracted from the data collected in the measurement area, and the car features are composed according to the bogie features, so as to further calculate the overload and eccentricity results of each car.
[0043] Furthermore, the specific elements for calculating the overload and eccentricity results of each carriage are:
[0044] 1) Bogie weight: The bogie weight of each carriage is the sum of the amplitudes of all axle characteristics in the bogie. Each carriage has four bogies, represented by front, rear, left and right, respectively. The expression is W p (c,r),W b (c,r),W left (c,r),W rig (c,r). c represents the cth group of data in the current data, and r represents the rth carriage in the data.
[0045] 2) The total weight of the carriage is W(c,r) = W p (c,r)+W b (c,r)+W left (c,r)+W rig (c,r).
[0046] 3) When the total weight of the carriage exceeds the set threshold, it is recorded as overloaded.
[0047] 4) Carriage front and rear weight difference ΔW pb (r,c) has ΔW pb (r,c)=W p (r,c)-W b (r,c).
[0048] 5) Carriage left and right weight difference ΔW lr (r,c) has ΔW lr (r,c)=W left (r,c)-W rig (r,c).
[0049] Furthermore, the axle features extracted in step 4 are subjected to peak deduplication and then used in step 5. The specific method is as follows:
[0050] In actual data, there are often abnormal feature interferences in the horizontal coordinate neighborhood of the same axle feature, and two or even more axle features are extracted from one axle. Therefore, the extracted axle features need to be screened and deduplicated.
[0051] Assume that the same axle detects two axle features, respectively (x i ,y i ),(x i+1 ,y i+1 ) represents the coordinates. When traversing the axle feature set to (x i ,y i ) element, by setting the y value of the ordinate of the axle feature i The proportional coefficient α (α is set to 0.1-0.5 according to different vehicle models and wheel weights) is used to find the left and right threshold points of the waveform data formed by the axle fusion data, and to remove repeated axle features. The coordinates of the left and right threshold points of the waveform are expressed as (sx, sy) and (ex, ey) respectively. The coordinates of the left and right threshold points satisfy the following relationship:
[0052] sy=y i ×α
[0053]
[0054] ey=y i ×α
[0055]
[0056] Where d is the fusion data of the upper or lower track. If sx≤x i+1 ≤ex, then remove (x i+1 ,y i+1 ) peak point, and continue to traverse the subsequent peak features. Through deduplication calculation, the following new peak feature set Feature(peak) is obtained: new :
[0057] Feature(peak) new = {P 1_new ,P 2_new ,…,P r_new}
[0058] P 1_new , P 2_new ,…,P r_new It is the unfiltered peak in the Feature(peak) set, that is, the axle feature after deduplication.
[0059] In summary, the present invention fuses the data collected by 8 sensors in each measurement area, fuses the data separately according to the upper and lower tracks, and transmits the fused data to the designed lightweight neural network structure to train the hyperparameter model, thereby extracting the axle characteristics and bogie characteristics of the train car, and forming the car characteristics to calculate the overload and unbalanced load results of each car. The use of lightweight neural networks in the present invention can effectively reduce the risks of false detection and missed detection, and improve the accuracy of axle feature detection; based on this, the accuracy of car detection will also be greatly improved, ultimately achieving the goal of improving the accuracy of train overload and unbalanced load detection and reducing the risks of freight trains due to overload and unbalanced load. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a schematic diagram of the layout of shear force and pressure sensors on the upper and lower tracks of the present invention;
[0061] Figure 2 A schematic diagram of data fusion of a two-axle bogie of an embodiment;
[0062] Figure 3 A schematic diagram of data fusion of a three-axle bogie of an embodiment;
[0063] Figure 4 A schematic diagram of data fusion of two adjacent carriages in an embodiment;
[0064] Figure 5 A schematic diagram of data fusion of adjacent light carriages and heavy carriages in an embodiment;
[0065] Figure 6 It is a schematic diagram of the characteristic marking of the embodiment;
[0066] Figure 7 A lightweight neural network and model training graph;
[0067] Figure 8 This is a schematic diagram of the peaks of the embodiment before deduplication. DETAILED DESCRIPTION
[0068] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.
[0069] A method for detecting overload and eccentric load based on a lightweight network, the specific steps are as follows:
[0070] Step 1: Install shear force sensors and pressure sensors on the upper and lower rails in a single measurement area. The overall installation structure is as follows: Figure 1 As shown in the figure: By collecting one-dimensional signals from each sensor for multiple vehicle models in a single measurement area, each set of data includes data from 8 sensors, expressed as:
[0071] Datas={D1,D2,…,D n}
[0072]
[0073] Step 2: Perform data fusion based on the shear force and pressure sensor data of the upper and lower tracks collected in step 1.
[0074] The data collected by the upper and lower rail shear and pressure sensors are integrated to obtain the wave crest shape of the wheel weight. The fusion method for the same group of pressure and shear data is as follows:
[0075]
[0076]
[0077] The schematic diagram of synthetic data is as follows Figure 2 , Figure 3 As shown, the same carriage has two bogies on the upper track or the lower track, and one bogie has two, three, four axles, etc. The bogies of adjacent carriages are closely spaced.
[0078] like Figure 4 In the train, two carriages are adjacent, and the axle of the front carriage is close to the axle of the rear carriage.
[0079] like Figure 5 In the figure, the light car is adjacent to the heavy car, the amplitude of the axle fusion data of the front car is large, and the amplitude of the axle fusion data of the rear car is small.
[0080] Step 3: Label the one-dimensional shear force and pressure sensor signals and divide them into training set, validation set and test set in proportion.
[0081] Different from the annotation of image data, the fused data is one-dimensional data. The two key features of axle feature and bogie feature are annotated on the synthetic data. The annotation method is as follows:
[0082] Figure 6 The coordinates of the red points in the middle represent the peak features, and the coordinates in the same bogie constitute the bogie features.
[0083] Fd top Taking the data in (t) as an example, the annotated results are expressed as follows:
[0084] Feature(peak)={(x1,y1),(x2,y2),…,(xl,yl)}
[0085] Feature(bogie)={{(x1,y1),(x2,y2)},{(x3,y3),(x4,y4)}…}
[0086] Among them, Feature(peak) represents the axle feature set, Feature(bogie) represents the bogie feature set, (x1,y1),(x2,y2),…,(xl,yl) represent l axle features respectively, {(x1,y1),(x2,y2)},{(x3,y3),(x4,y4)}… represent that two axle features constitute a bogie feature.
[0087] The labeled data set is divided into training set, validation set and test set in a ratio of 7:2:1.
[0088] Step 4: Design a lightweight neural network structure, and train the model with the data set annotated in step 3 to obtain a model for detecting axle features and realize axle feature extraction; the lightweight neural network structure is:
[0089] like Figure 7 As shown in:
[0090] Input layer: time series x_axiel of collected data, amplitude sequence y_axiel corresponding to each time point, 2×T matrix, T is the length of the collected time series.
[0091] Hidden layer: Each time series collected is only associated with L consecutive data, where L is the current collection time point and the length of continuous time. To ensure the accuracy of the extracted features, the hidden layer extracts features through 10 consecutive convolutional layers and activation layers (ReLU).
[0092] Calculation method of L:
[0093]
[0094] The expression formula of ReLU is:
[0095] f(x)=max(0,w T x+b)
[0096] Output layer: The feature vector extracted by the hidden layer is used to calculate the probability of each element for the 0-class label (label_0) and the 1-class label (label_1), and the final classification result is calculated through the Softmax function. The expression of the Softmax function is as follows:
[0097]
[0098] The axle feature detection model is obtained by training this lightweight neural network and labeled dataset.
[0099] Step 5: Remove duplicate axle (peak) features.
[0100] In actual data, the same axle often has abnormal axle feature interference, such as Figure 8 As shown, two or even more axle features are extracted from one axle, so the extracted axle features need to be screened and deduplicated.
[0101] set up Figure 8 The two axle features detected in the i ,y i ),(x i+1 ,y i+1 ) indicates that the two axle features belong to the same axle, and traversing to (x i ,y i ) point, by setting the proportional coefficient α (0.1-0.5) of the ordinate, find the left and right threshold points of the axle synthetic data waveform, remove the repeated axle features, and the coordinates of the left and right threshold points of the waveform are expressed as (sx, sy) and (ex, ey), respectively. The coordinates of the left and right threshold points satisfy the following relationship:
[0102] sy=y i ×α
[0103]
[0104] ey=y i ×α
[0105]
[0106] d is the upper orbit fusion data or the lower orbit fusion data; if sx≤x i+1 ≤ex, then remove (x i+1 ,y i+1 ) peak point. After the deduplication step, the final axle feature is obtained.
[0107] Step 6: The bogie feature extraction process is the same as the axle feature extraction process. The deduplicated axle feature results are used as the new input layer. After the bogie features are calculated through 3 hidden layers and the ReLU layer, the feature probability of whether it is the same bogie is calculated through the output layer. label_2 indicates that it belongs to the same bogie, and label_3 indicates that it is not the same bogie. After the Softmax layer, the final bogie feature result is obtained. The set of axle feature results and the set of bogie feature results are expressed as follows:
[0108] Feature(bogie) top ={{(x1 top ,y1 top ),(x2 top ,y2 top )},{(x3 top ,y3 top ),(x4top ,y4 top )}…}Feature(bogie) blow
[0109] ={{(x1 vlow ,y1 blow ),(x2 blow ,y2 blow )},{(x3 blow ,y3 blow ),(x4 blow ,y4 blow )}…}
[0110] Among them, top represents the upper track fusion data in the current data, Feature (bogie) top Indicates the bogie features of the upper track fusion data in the current data, (x1 top ,y1 top ),(x2 top ,y2 top ),(x3 top ,y3 top ),(x4 top ,y4 top ) is the feature of each axle on the upper track; Feature (bogie) blow Indicates the bogie features of the lower track fusion data in the current data, (x1 blow ,y1 blow ),(x2 blow ,y2 blow ),(x3 blow ,y3 blow ),(x4 blow ,y4 blow ) represents the characteristics of each axle of the lower track.
[0111] Step 7: Extract the features of each carriage based on the axle features obtained in step 5 and the bogie features obtained in step 6. There are two bogies on the upper and lower tracks of a carriage, respectively, which can be expressed as follows:
[0112] Feature(car)={{{(x1 top ,y1 top ),(x2 top ,y2 top ),(x1 blow ,y1 blow ),(x2 blow ,y2 blow )}},{{(x3 top ,y3 top ),(x4 top ,y4top ),(x3 blow ,y3 blow ),(x3 blow ,y3 blow )}}…}
[0113] Among them, Feature(car) is the car feature.
[0114] Step 8: Check the overload and unbalanced load results of each carriage.
[0115] 1) Bogie weight: The bogie weight of each carriage is the sum of the ordinate values of all axle features (i.e., crest coordinates) in the bogie. Each carriage has four bogies, represented by front, rear, left, and right, respectively. The expression is W p (c,r),W b (c,r),W left (c,r),W rig (c,r).
[0116] 2) Total weight of the carriage W(c,r): W(c,r)=W p (c,r)+W b (c,r)+W left (c,r)+W rig (c,r).
[0117] 3) When the total weight of the carriage exceeds the set threshold, it is recorded as overloaded.
[0118] 4) Carriage front and rear weight difference ΔW pb (r,c):ΔW pb (r,c)=W p (r,c)-W b (r,c).
[0119] 5) Carriage left and right weight difference ΔW lr (r,c):ΔW lr (r,c)=W left (r,c)-W rig (r,c).
[0120] It can be seen from the above embodiments that the present invention installs 4 shear sensors and 4 pressure sensors in each measuring area of the railroad track, fuses the data collected by the 8 sensors in each measuring area, fuses the data according to the upper and lower tracks respectively, and transmits the fused data to the designed lightweight neural network structure to train the hyperparameter model, thereby extracting the axle characteristics and bogie characteristics of the train car, and composing the car characteristics to calculate the overload and eccentricity results of each car. The use of lightweight neural networks in the present invention can effectively reduce the risks of false detection and missed detection, and improve the accuracy of the detection of axle characteristics; and provide a deduplication step for axle characteristics to further improve the accuracy; based on this, the detection accuracy of the car will also be greatly improved, and ultimately achieve the goal of improving the accuracy of train overload and eccentricity detection and reducing the risks of freight trains due to overload and eccentricity.
Claims
1. A method for detecting overload and unbalanced load based on a lightweight network, characterized in that: The specific steps are as follows: Step 1: Install 4 shear sensors and 4 pressure sensors in each measurement area set on the rails, and collect data from multiple vehicle models as a data set; wherein each measurement area divides the two rails into upper and lower rails, the shear sensors installed on the upper rails are named S1 and S2, and the pressure sensors are named P1 and P2; the shear sensors installed on the lower rails are named S3 and S4, and the pressure sensors are named P3 and P4; In the same measurement area, each sensor collects one-dimensional signal data in time series. The data on the eight sensors correspond to a group of data in the measurement area. The data set is expressed as: Datas={D1,D2,…,D n } Among them, Datas is the data set collected by the sensors in the measurement area, D1, D2, ..., D n It means there are n groups of data in the data set, i means the i-th sequence in n groups of data, D i represents the i-th group of data structures, represents the i-th data sequence collected by shear sensor S1, Respectively represent the m amplitude data in the sequence; represents the i-th data sequence collected by shear sensor S2, represents the m amplitudes in the sequence; represents the i-th data sequence collected by shear sensor S3, represents the m amplitudes in the sequence; represents the i-th data sequence collected by shear sensor S4, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P1, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P2, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P3, represents the m amplitudes of the sequence; represents the i-th data sequence collected by pressure sensor P4; Step 2: Perform data fusion based on the shear force and pressure sensor data of the upper and lower tracks collected in step 1; The data collected by the shear and pressure sensors on the upper and lower tracks are fused separately to obtain the peak shape of the weight borne by each axle, and the horizontal and vertical coordinate positions of the peaks are extracted and recorded as axle features; the horizontal axis represents the time series of the sensor data collected, and the vertical axis represents the amplitude corresponding to each time series; the fusion method for the pressure and shear data of the same group is as follows: Among them, t is the number of amplitude sequences in the same set of data, t∈[1,m], Fd top (t) represents the t-th amplitude of the upper orbit fusion, represents the t-th data amplitude in the i-th data set of shear sensor S1, represents the t-th data amplitude in the i-th group of data of pressure sensor P1, represents the t-th data amplitude in the i-th group of data of pressure sensor P2, Fd represents the t-th data amplitude in the i-th data set of shear sensor S2; down (t) represents the tth amplitude after the fusion of the lower track, represents the t-th data amplitude in the i-th data set of shear sensor S3, represents the t-th data amplitude in the i-th group of data of the pressure sensor P3, represents the t-th data amplitude in the i-th group of data of pressure sensor P4, represents the t-th data amplitude in the i-th group of data of the shear sensor S4; The same carriage consists of two bogies on the upper track or the lower track, and one bogie has 2-4 axles; Step 3: annotate the one-dimensional shear force and pressure sensor signals, and divide them into training set, validation set and test set according to the proportion; The two key features of axle feature and bogie feature are marked on the fused data obtained in step 2. The axle feature in the same bogie constitutes the bogie feature; the specific marking method is: Feature(peak)={P1,P2,...,P l } Feature(bogie)={B1,B2,...,B k } Among them, Feature(peak) represents the axle feature set, each element is a coordinate, P1 represents the feature of the first axle, P2 represents the feature of the second axle, P l Indicates the feature of the lth axle, there are l in total; Feature (bogie) represents the bogie feature set, each bogie feature contains 2-4 axle features, B1 represents the first bogie feature, B2 represents the second bogie feature, Bk represents the kth bogie feature, there are k in total; 2k≤l; Step 4: Design a lightweight neural network structure, and train the model with the data set annotated in step 3 to obtain a model for detecting axle features and realize axle feature extraction; the lightweight neural network structure is: Input layer: The horizontal axis x_axiel set and the vertical axis y_axiel. The horizontal axis x_axiel set is the time series of data collected by 8 sensors in the measurement area. The vertical axis y_axiel is the amplitude sequence fused according to the method in step 1 in the upper or lower track corresponding to each sampling time point in the measurement area. x_axiel and y_axiel form a matrix of size 2×T, where T is the length of the time series. Hidden layer: Each time series collected is only associated with L consecutive data, where L is the current collection time point and the length of time with continuity. To ensure the accuracy of the extracted features, the hidden layer extracts features through 10 consecutive convolutional layers and activation layers; L is calculated as follows: Among them, Len is the maximum length of the train passing through the measurement area, V is the maximum speed of the train currently passing through the measurement area, and Fre is the sensor acquisition frequency in the current measurement area. The acquisition frequency of the 8 sensors is the same; The expression formula of the activation layer is: f(x)=max(0,w T x+b) Where x is the vector of the input layer, f(x) is the result after the x vector is activated, max represents the maximum value function, and w T represents the transpose of the matrix w, w represents the weight matrix, and b represents the offset matrix; Output layer: The feature vector extracted by the hidden layer is used to calculate the probability of each element for the class 0 label label_0 and the class 1 label label_1. The Softmax function is used to calculate the final classification result. The expression of the Softmax function is as follows: Among them, f(0) represents the probability of the category being label 0, f(1) represents the probability of the category being label 1, e represents the exponential function, y0 represents the probability distribution of the category being label 0, and y1 represents the probability distribution of the category being label 1; Step 5, using the axle features extracted in step 4 as a new input layer, and then training a model for extracting bogie features through the lightweight neural network described in step 4 to achieve bogie feature extraction; The bogie feature extraction process is the same as the peak calculation process. The axle feature result is used as the new input layer. The bogie feature is calculated through 3 hidden layers and ReLU layers. Then the feature probability of the same bogie is calculated through the output layer. Label_2 indicates that it belongs to the same bogie, and label_3 indicates that it is not the same bogie. Finally, after the Softmax layer, the final bogie feature result is obtained. Feature(bogie) new ={B 1_new ,B 2_new ,...} Among them, B 1_new and B 2_new Bogie features, each bogie feature is a collection of 2-4 axle features; Step 6, extracting the features of each carriage according to the axle features obtained in step 4 and the bogie features obtained in step 5; A carriage consists of two bogies adjacent to the upper and lower tracks. Therefore, the characteristics of the carriage composed of four bogies adjacent to the upper and lower tracks are expressed as follows: Feature(car)={{B 1_top ,B 2_top ,B 1_blow ,B 2_blow },{B 3_top ,B 4_top ,B 3_blow ,B 4_blow }...} Among them, Feature(car) is the feature of the car, B 1_top and B 2_top The first and second bogie features detected by the upper track fusion data, B 1_blow , B 2_blow represents the first and second bogie features detected by the lower track fusion data, B 3_top and B 4_top The third and fourth bogie features detected by the upper track fusion data; B 1_blow , B 2_blow represents the first and second bogie features detected by the lower track fusion data, B 3_blow , B 4_blow Indicates the 3rd and 4th bogie features detected by the lower track fusion data; the number of upper track or lower track bogie features detected for a train is twice the number of carriage features; Step 7: Based on the model trained in steps 4-5, extract the axle and bogie features from the data collected in the measurement area, and compose the car features based on the bogie features, so as to calculate the overload result of each car.
2. The overload and unbalanced load detection method based on a lightweight network as claimed in claim 1, characterized in that: The specific elements for calculating the overload and eccentric load results of each carriage in step 7 are: 1) Bogie weight: The bogie weight of each carriage is the sum of the amplitudes of all axle characteristics in the bogie. Each carriage has four bogies, represented by front, rear, left and right, respectively. The expression is W p (c, r), W b (c,r),W left (c, r), W rig (c, r); where c represents the cth group of data in the current data, and r represents the rth carriage in the data; 2) The total weight of the carriage is W(c, r) = W p (c,r)+W b (c,r)+W left (c,r)+W rig (c, r); 3) When the total weight of the carriage exceeds the set threshold, it is recorded as overloaded; 4) Carriage front and rear weight difference ΔW pb (r, c) has ΔW pb (r, c) = W p (r,c)-W b (r, c); 5) Carriage left and right weight difference ΔW lr (r, c) has ΔW lr (r, c) = W left (r,c)-W rig (r, c).
3. The overload and unbalanced load detection method based on a lightweight network as claimed in claim 1, characterized in that: In step 3, the ratio of the training set, validation set and test set is 7:2:
1.
4. The overload and unbalanced load detection method based on a lightweight network as claimed in claim 1, characterized in that: The axle features extracted in step 4 are subjected to peak deduplication and then used in step 5. The specific method is as follows: Assume that the same axle detects two axle features, respectively (x i ,y i ), (x i+1 ,y i+1 ) represents the coordinates. When traversing the axle feature set to (x i ,y i ) element, by setting the y value of the ordinate of the axle feature i The proportional coefficient α is used to find the left and right threshold points of the waveform data formed by the axle fusion data and remove duplicate axle features. α is set to 0.1-0.5 according to different vehicle models and wheel weights. The coordinates of the left and right threshold points of the waveform are expressed as (sx, sy) and (ex, ey) respectively. The coordinates of the left and right threshold points satisfy the following relationship: and = y i ×α hey=y i ×α Where d is the fusion data of the upper or lower track; if sx≤x i+1 ≤ex, then remove (x i+1 ,y i+1 ) peak point, continue to traverse the subsequent peak features; through deduplication calculation, the following new peak feature set Feature(peak) is obtained: new : Feature(peak) new ={P 1_new ,P 2_new ,...,P r_new } P 1_new , P 2_new , ..., P r_new It is the unfiltered peak in the Feature(peak) set, that is, the axle feature after deduplication.
Citation Information
Patent Citations
Railway vehicle overload and imbalanced load detection system
CN106644010A
Freight train intelligent online monitoring device
CN206255021U
