Damage signal detection method based on adaptive weighted multi-sensor data fusion

Abstract

This invention discloses a damage signal detection method based on adaptive weighted multi-sensor data fusion. The method comprises the following steps: 1. Loading acoustic emission signals containing damage information from three sensor groups, calculating the cosine similarity and real-time spatial information between the sensor signals of each group to form a support matrix, obtaining weight coefficients, and linearly weighting the fused signal; 2. Applying a sliding time window operation to the fused signal, performing wavelet decomposition on the fused signal within each time window, calculating the wavelet coefficients, and then calculating the Tsallis entropy based on the energy proportions of these coefficients to obtain the time-wavelet Tsallis entropy; 3. Frame-by-frame (TWTE) data, calculating the standard deviation of the entropy within each group to form a detection threshold, and combining the thresholds within each frame to form an adaptive rail damage detection threshold. Signals corresponding to TWTE values below the threshold are detected as rail damage signals. This invention improves the detection accuracy and real-time performance of rail damage signals.

Description

Technical Field

[0001] The present invention relates to a rail damage signal processing and detection method, and in particular to a damage signal detection method based on adaptive weighted multi-sensor data fusion. Background Art

[0002] High-speed rail is a crucial cornerstone and backbone of China's modern transportation network. With its speed, high capacity, and high punctuality, it has become the preferred mode of transportation for many passengers and freight. The rapid expansion of China's high-speed rail network and the increasing frequency of its operations have posed greater challenges to the maintenance of rail infrastructure, particularly rails. As the core load-bearing structure of the railway system, rails are constantly subjected to enormous loads and continuous vibration. This high-speed operating environment inevitably leads to surface wear and damage. Fatigue damage is a typical type of rail damage. If not detected and addressed promptly, these damages may expand during continued operation, ultimately leading to serious safety accidents such as rail breakage or derailment. Therefore, implementing effective rail damage detection is crucial to ensuring the safe and stable operation of high-speed railways.

[0003] Currently, rail damage signal detection methods can be categorized into two main categories: those based on single-sensor data and those based on multi-sensor data fusion. Single-sensor data methods are more common in rail damage signal detection. They rely primarily on signals from a single source and are relatively simple to process. However, since they only acquire data from a single perspective, this can lead to incomplete information, compromising the accuracy and reliability of damage detection. In contrast, multi-sensor data fusion methods integrate data from different sensors to analyze signals from multiple dimensions, thereby improving detection accuracy and robustness. Multi-sensor data fusion typically employs supervised learning algorithms, which require a large amount of labeled data with detailed damage information to train the model. During this process, each set of collected signals must be manually labeled to determine whether it is a damage signal. This process is both labor-intensive and time-consuming. Furthermore, training and optimizing supervised learning models requires high-performance computing resources, which limits their effectiveness in practical applications. Given the rapid development of China's modern railway network, the limitations of existing rail damage signal detection methods highlight the need for more efficient and intelligent detection methods. Therefore, it is of extraordinary significance to propose a multi-sensor data fusion rail damage signal detection method that does not require any prior information and has high detection accuracy and strong real-time performance. Summary of the Invention

[0004] To address the slow speed and low accuracy of traditional damage signal detection methods, the present invention provides a damage signal detection method based on adaptive weighted multi-sensor data fusion. This method, based on adaptive weighted multi-sensor data fusion, constructs a support matrix based on the cosine similarity between each set of sensor data. This matrix is updated using real-time spatial information from each sensor. Weight coefficients for each set of sensor data are calculated, and a fused signal is obtained through weighting. The resulting fused signal is then subjected to a sliding time window operation. Wavelet decomposition is performed on the fused signal within each time window, and wavelet coefficients are calculated. The Tsallis entropy is calculated based on the distribution of these coefficients, and the time-wavelet Tsallis entropy is obtained as a detection index. The standard deviation of this detection index is used to construct an adaptive rail damage detection threshold. Signals corresponding to a time-wavelet Tsallis entropy below the threshold are identified as rail damage signals, achieving highly accurate damage signal detection. This method boasts high computational speed and high detection accuracy, improving the accuracy and real-time performance of rail damage signal detection. It holds great social and economic value in the field of high-speed rail damage detection.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A damage signal detection method based on adaptive weighted multi-sensor data fusion includes the following steps:

[0007] Step 1: Load the acoustic emission signals containing damage information from the three sensor groups, calculate the cosine similarity between the sensor signals of each group, construct a support matrix, and then calculate the real-time spatial information of each sensor to update the matrix. Obtain the weight coefficient of each sensor data group, and linearly weight it to obtain the fusion signal FS.

[0008] Step 2: Perform a sliding time window operation on the fused signal FS, perform wavelet decomposition on the fused signal within each time window, calculate the wavelet coefficients, and then calculate the Tsallis entropy based on the energy proportion of these coefficients to obtain the Time-Wavelet Tsallis Entropy (TWTE);

[0009] Step 3: Divide the TWTE into frames and calculate the standard deviation of the entropy value within each group to form a detection threshold. The thresholds within each frame are combined to form an adaptive rail damage detection threshold. Signals corresponding to TWTEs below the threshold are detected as rail damage signals, achieving high-precision detection.

[0010] Compared with the prior art, the present invention has the following advantages:

[0011] 1. This paper proposes a damage signal detection method based on multi-sensor data fusion. Sensor data from different locations are adaptively weighted and fused based on cosine similarity and real-time spatial position information. The obtained fused data has a higher confidence level than other single-sensor data, and contains more angles and richer damage information, providing more stable data support for rail damage detection.

[0012] 2. The present invention proposes a time-wavelet Tsallis entropy as a rail damage detection index. A sliding time window method is used to divide the obtained fusion signal into multiple time windows. The Tsallis entropy characteristics of the signal within each time window are calculated to reveal local damage information that may be obscured by the global perspective of the signal. Then, combined with the time domain characteristics of the time-wavelet Tsallis entropy, an adaptive rail damage detection threshold is constructed. Outliers within each detection index are searched for, damage information is identified, and accurate detection of rail damage signals is achieved.

[0013] 3. Most existing rail damage signal detection methods based on multi-sensor data fusion are based on supervised learning algorithms. These methods require labeled data with detailed damage information to train deep learning models. Each signal must be labeled as a damage signal, resulting in low algorithm efficiency, significant time consumption, and difficulty in practical detection. However, the damage signal detection method based on adaptive weighted multi-sensor data fusion proposed in this paper requires no prior information, offers rapid detection speed, and significantly improves detection accuracy compared to existing damage signal detection methods. This method can simultaneously meet the high-speed and high-precision damage detection requirements of actual railway operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The flowchart of the damage signal detection method based on adaptive weighted multi-sensor data fusion of the present invention.

[0015] Figure 2 is the time-amplitude diagram of the fused signal.

[0016] Figure 3 It is the time-amplitude diagram of the time-wavelet Tsallis entropy.

[0017] Figure 4 This is a signal detection example diagram

[0018] Figure 5 A comparison chart of the test results. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0020] The present invention provides a damage signal detection method based on adaptive weighted multi-sensor data fusion, such as Figure 1 The specific steps are as follows:

[0021] Step 1: Load the acoustic emission signals containing damage information from the three sensor groups, calculate the cosine similarity between the sensor signals, construct a support matrix, and then calculate the real-time spatial information of each sensor to update the matrix, obtain the weight coefficient of each sensor data group, and linearly weight the fusion signal FS. The specific steps are as follows:

[0022] Step 1: Load the acoustic emission signals containing damage information of three sets of sensors, marked as and Where L represents the number of electromagnetic acoustic emission signal sequences collected by each sensor, represents the signal sequence of sensor 1 at time t, represents the signal sequence of sensor 2 at time t, represents the signal sequence of sensor 3 at time t, t∈[1,2,…,L], containing N sampling points;

[0023] Step 1 and 2: Calculate the cosine similarity between the sensor signals at time t to form the support matrix SM:

[0024]

[0025] Among them, i1,j1∈[1,2,3], represents the cosine similarity between the signals of sensor i1 and sensor j1 at time t;

[0026] Step 13: Calculate the real-time spatial position of sensor 1 at time t

[0027]

[0028] Where r represents the distance from the sensor to the detection wheel, R represents the radius of the detection wheel, V represents the rolling speed of the detection wheel, θ1 represents the initial position of sensor 1, and similarly, based on the initial positions θ2 and θ3 of sensor 2 and sensor 3, the real-time spatial positions of sensor 2 and sensor 3 at time t can be calculated. and

[0029] Step 14: Update the support matrix based on the real-time position information of the three sets of sensors to obtain the updated matrix

[0030]

[0031] in,[·] T represents the transpose operation;

[0032] Step 15: Calculate the weight coefficient F of each sensor at time t based on the updated support matrix a :

[0033]

[0034]

[0035] Step 16: Multiply these weight coefficients by the corresponding sensor signals to obtain the fusion signal FS at time t t :

[0036]

[0037] Step 17: Repeat the above steps 11 to 16 L times to obtain the fusion signal of multi-sensor data at all times, marked as: FS = {FS 1 ,FS 2 ,…,FS L}.

[0038] Step 2: Perform a sliding time window operation on the fused signal FS, perform wavelet decomposition on the fused signal within each time window, calculate the wavelet coefficients, and then calculate the Tsallis entropy based on the energy proportion of these coefficients to obtain the time-wavelet Tsallis entropy. The specific steps are as follows:

[0039] Step 21: Arrange the fusion signal FS obtained in step 1 in time sequence to form a continuous signal

[0040]

[0041] Step 22: Perform a sliding time window operation on the fusion signal:

[0042]

[0043] Among them, L T represents the length of the time window in milliseconds, fs represents the signal sampling rate, represents the floor function, ε T Represents the sliding distance, usually L T / 2,b1∈[1,2,…,N T ],N T Represents the total number of time windows;

[0044] Step 2 and 3: Signal within the time window Perform discrete wavelet transform to obtain approximate wavelet coefficients A n (b1) and detailed wavelet coefficients D n (b1):

[0045]

[0046] Where n represents the number of wavelet decomposition layers;

[0047] Step 24: Calculate the n-layer detailed wavelet coefficient energy E n and approximate wavelet coefficient energy E approx :

[0048]

[0049] Step 25: Calculate the total energy E of the signal total And the energy ratio P of each level n :

[0050]

[0051] P n =E n / E total

[0052] Step 26: Based on the energy ratio P obtained n Calculate the Tsallis entropy of the fused signal within the time window:

[0053]

[0054] Where q represents the non-extensive parameter, and its value is set to n / 2; repeat the calculation N T The TWTE of the full cycle of the fusion signal can be obtained by calculating the Tsallis entropy in a time window.

[0055] Step 3: Divide the TWTE into frames, calculate the standard deviation of the entropy value in each group, form the detection threshold, and combine the thresholds in each frame to form the adaptive rail damage detection threshold. The signal corresponding to the TWTE below the threshold is detected as a rail damage signal, achieving high-precision detection. The specific steps are as follows:

[0056] Step 31: Divide the TWTE obtained in step 2 into frames, that is, divide all TWTE into N H Frames:

[0057] H(j H )={H(i H ),i H =1+(j H -1)ε H ,…,j H *ε H}

[0058] Where H represents the total TWTE, ε H Represents the number of TWTEs contained in each frame, j H ∈[1,2,…,N H ],

[0059] Step 32: Calculate the jth H Intra-frame TWTE, namely H(j H )

[0060]

[0061] Among them, mean(·) represents the mean operation;

[0062] Step 33: Construct the jth H Detection threshold within the frame

[0063]

[0064] Among them, min(·) represents the minimum value operation;

[0065] Step 34: Repeat steps 31 to 33 to calculate N H The detection thresholds of TWTE in each frame are connected in parallel to obtain the overall adaptive impairment detection threshold ADH, and the impairment signal detection standard is established:

[0066]

[0067] Among them, Y represents the final damage signal detection result. When Y = 1, it means that the signal corresponding to the TWTE is a damage signal, that is, the detection wheel rolls over the damaged part of the rail surface. Conversely, when Y = 0, it means that the signal corresponding to the TWTE is a noise signal. Following this detection process, high-precision and real-time rail damage signal detection can be achieved.

[0068] Example:

[0069] This embodiment illustrates the specific implementation of the present invention by combining the actual measurement of rail damage and the simulation signal data:

[0070] Execute step 1: Load 7936 sets of damage signal sequences collected by three sets of sensors during the full cycle of wheel rolling on the rail. and That is, L=7936, each signal sequence contains 4096 sampling points, that is, M=4096, and the entire signal contains two impairments.

[0071] Take the first moment signal sequence of the three groups of sensors, and The cosine similarities between them are calculated to be 0.40, 0.20 and 0.23 respectively, so the support matrix at this time is:

[0072]

[0073] Then calculate the spatial positions of the three groups of sensors at the first moment as 0.1, 0.45 and 0.45, and use these parameters to update the support matrix The weight coefficients of the three groups of sensors are 0.18, 0.21 and 0.61 respectively, so that the fusion signal at the first moment can be obtained by weighting.

[0074] Repeat the above steps to calculate all 7936 groups of fusion signals, and connect these signal sequences together in chronological order to obtain the fusion signal time-amplitude diagram in all cycles, as shown in Figure 2 shown.

[0075] Execute step 2: perform sliding time window operation on the fusion signal, with a total of 1856 time windows, namely N T =1856, and then perform 4-layer wavelet decomposition on the signal in each time window, that is, n=4, the wavelet basis is db3, and the wavelet coefficients are obtained. Based on these coefficients, the Tsallis entropy of the signal in the time window is extracted in each layer of energy distribution.

[0076] Repeat the above steps to calculate the Tsallis entropy of the fused signal in all 1856 time windows, and connect these entropy values in sequence to obtain the time-wavelet Tsallis entropy proposed in the present invention. Its time-amplitude value is as follows: Figure 3 shown.

[0077] Execute step 3: divide the obtained time-wavelet Tsallis entropy into frames, with each frame consisting of 40 entropy values. A total of 46 frames are obtained, i.e., ε H =40, N H =46, calculate the standard deviation of the 40 entropy values in each frame, and then construct the detection threshold in the frame. The adaptive damage detection threshold proposed in this invention can be obtained by sequentially connecting 46 groups of detection thresholds. Damage detection is performed according to the damage discrimination standard. The final detection result is as follows: Figure 4 As shown, it can be found that both damages are fully detected, and the entire detection method takes 3.91 seconds, which is much less than the 6.5 seconds that the entire signal lasts. This means that the method proposed in the present invention can not only ensure high-precision detection, but also achieve real-time detection.

[0078] In order to verify the superiority of the damage signal detection method proposed in this invention, signals were collected under 6 different experimental conditions and compared with the detection based on single sensor data. The results are as follows: Figure 5 As shown in the figure, it can be found that the damage signal detection method based on adaptive weighted multi-sensor data fusion proposed in the present invention has a damage detection accuracy of 100% under 6 different conditions, which is much higher than the detection results of the three sets of sensor data alone.

Claims

1. A damage signal detection method based on adaptive weighted multi-sensor data fusion, characterized by The method comprises the following steps: Step 1: Load the acoustic emission signals containing damage information of the three groups of sensors, calculate the cosine similarity between the sensor signals of each group, form a support matrix, and then calculate the real-time spatial information of each sensor to update the matrix, obtain the weight coefficient of each group of sensor data, and obtain the fusion signal by linear weighting. , the specific steps are as follows: Step 1: Load the acoustic emission signals containing damage information of three sets of sensors, marked as , and ,in Represents the number of electromagnetic acoustic emission signal sequences collected by each sensor, Represents sensor 1 in The signal sequence at each moment, Represents sensor 2 in The signal sequence at each moment, Represents sensor 3 in The signal sequence at each moment, ,contain sampling points; Steps 1 and 2: Calculation The cosine similarity between the sensor signals at each moment constitutes the support matrix : in, , Representatives in Time sensor With sensor Cosine similarity between signals; Step 13: Calculate sensor 1 Real-time spatial location at the moment : in, Represents the distance from the sensor to the detection wheel, Represents the radius of the detection wheel, Represents the detection of wheel rolling speed, Represents the initial position of sensor 1, and similarly based on the initial positions of sensor 2 and sensor 3 and , we can calculate the sensor 2 and sensor 3 Real-time spatial location at the moment and ; Step 14: Update the support matrix based on the real-time position information of the three sets of sensors to obtain the updated matrix : in, represents the transpose operation; Step 15: Calculate the support matrix of each sensor based on the updated support matrix. Weight coefficient of time : Step 16: Multiply these weight coefficients by the corresponding sensor signals to obtain Fusion of signals at all times : Step 17: Repeat the above steps 11 to 16 times, that is, obtaining the fusion signal of multi-sensor data at all times, marked as: ; Step 2: Fusion signal Perform a sliding time window operation, perform wavelet decomposition on the fusion signal in each time window, calculate the wavelet coefficients, and then calculate the Tsallis entropy based on the energy proportion of these coefficients to obtain the time-wavelet Tsallis entropy; Step 3: Divide the time-wavelet Tsallis entropy into frames, calculate the standard deviation of the entropy value within each group, and form a detection threshold. The thresholds within each frame are combined to form an adaptive rail damage detection threshold. Signals corresponding to the time-wavelet Tsallis entropy below the threshold are detected as rail damage signals, achieving high-precision detection.

2. The damage signal detection method based on adaptive weighted multi-sensor data fusion according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 21: The fusion signal obtained in step 1 Arranged in time sequence to form a continuous signal : Step 22: Perform a sliding time window operation on the fusion signal: in, Represents the length of the time window in milliseconds. represents the signal sampling rate, represents the floor function, Represents the sliding distance, , Represents the total number of time windows; Step 2 and 3: Signal within the time window Perform discrete wavelet transform to obtain approximate wavelet coefficients and detailed wavelet coefficients : in, represents the number of wavelet decomposition layers; Step 24: Calculation Layer-detailed wavelet coefficient energy and approximate wavelet coefficient energy : Step 25: Calculate the total energy of the signal And the energy ratio of each level : Step 26: Based on the energy ratio obtained Calculate the Tsallis entropy of the fused signal within the time window: in, Represents a non-extensive parameter; repeated counting The Tsallis entropy in a time window can be used to obtain the time-wavelet Tsallis entropy in the full cycle of the fusion signal.

3. The damage signal detection method based on adaptive weighted multi-sensor data fusion according to claim 2 is characterized in that described The value is .

4. The damage signal detection method based on adaptive weighted multi-sensor data fusion according to claim 2 is characterized in that described The value is .

5. The damage signal detection method based on adaptive weighted multi-sensor data fusion according to claim 2 is characterized in that The specific steps of step three are as follows: Step 31: Divide the time-wavelet Tsallis entropy obtained in step 2 into frames, that is, divide all the time-wavelet Tsallis entropy into Frames: in, represents the total time-wavelet Tsallis entropy, Represents the amount of time-wavelet Tsallis entropy contained in each frame, , ; Step 32: Calculate the Intra-frame time-wavelet Tsallis entropy, i.e. Standard deviation : in, Indicates the mean operation; Step 3: Build the Detection threshold within the frame : in, Indicates the minimum operation; Step 34: Repeat steps 31 to 33 to calculate The detection thresholds of time-wavelet Tsallis entropy in each frame are combined to obtain the overall adaptive damage detection threshold , establish damage signal detection standards: in, represents the time-wavelet Tsallis entropy, Represents the final damage signal detection result. When , the signal corresponding to the time-wavelet Tsallis entropy belongs to the damage signal, that is, the detection wheel rolls over the damaged part of the rail surface. On the contrary, when When , it means that the signal corresponding to the time-wavelet Tsallis entropy belongs to a noise signal.

Images