Sensor fault detection method based on wavelet packet decomposition and long-short term memory network

Abstract

The invention discloses a sensor fault detection method based on wavelet packet decomposition and a long-short term memory network. The method comprises the following steps: acquiring a vibration signal of a mechanical transmission system by using a sensor; setting a sample and a corresponding label for the vibration signal by using a sliding window; performing wavelet packet decomposition on the sample to form multi-scale input and construct a data set; constructing a prediction model, and training by using the data set and the label; and inputting a test sample into the trained model to predict a future signal, and comparing the future signal with a real measurement value so as to judge the working state of the sensor. The method does not need to depend on fault data to carry out model training, can avoid diagnosis system misjudgment caused by sensor faults, and improves equipment operation efficiency and fault detection reliability.

Description

Technical Field

[0001] The present invention belongs to the technical field of sensor fault detection, and particularly relates to a sensor fault detection method based on wavelet packet decomposition and long short-term memory network. Background Art

[0002] In the field of mechanical system fault monitoring, vibration signal analysis is one of the most commonly used means. By capturing and analyzing the vibration characteristics generated during the operation of machinery, the abnormal states inside the equipment can be effectively detected. For some complex equipment, its mechanical transmission system usually serves in extremely harsh environments such as strong vibration, high impact, and multi-dust for a long time. Under such harsh conditions, vibration sensors are prone to failures due to external factors, directly affecting the accuracy of signal acquisition, and even leading to misjudgments of equipment failures, thus threatening the normal operation and safety of the equipment. Therefore, it is very important to develop a sensor fault detection method to accurately judge signals.

[0003] Sensor fault detection methods usually take data-driven as the core. Traditional methods mainly rely on statistical theories and signal processing techniques. Common representatives include the Autoregressive Integrated Moving Average (ARIMA) model, Kalman filter, etc. The ARIMA model can predict the normal change trend of signals and detect abnormalities by modeling time series data. However, its ability to process non-stationary data is relatively limited, and it is difficult to handle dynamic signals under complex working conditions. The Kalman filter is widely used in the state estimation of dynamic systems and is good at dealing with environments with less noise interference. However, its performance highly depends on noise characteristics and model assumptions. Once these assumptions are violated, its accuracy may drop significantly. Generally speaking, these traditional methods still have application value in specific scenarios. However, with the increase in the complexity and data volume of modern equipment signals, they gradually show limitations in complex environments with high dimensions, non-linearity, and dynamic changes.

[0004] In recent years, the rapid development of deep learning technology has provided a new research direction for sensor fault detection. Due to its excellent time series modeling ability, the classical Recurrent Neural Network (RNN) is widely used in vibration sensor fault detection. However, RNN often faces the problems of gradient explosion and gradient disappearance when dealing with long time series. To address these limitations, the Long Short-Term Memory Network (LSTM), as an improved model, can effectively solve the above problems and has the ability to preserve long-term historical information, making it very suitable for modeling complex time series. The present invention constructs a sequence-to-sequence network framework based on LSTM, and realizes sensor fault detection by predicting the change trend of future vibration signals and comparing it with the actual value. To further improve the prediction performance, wavelet packet decomposition is used to preprocess the input signal to construct multi-scale inputs. With its excellent time-frequency resolution ability, wavelet packet decomposition can refine and decompose the signal into different frequency bands and time scales, extract the key features of each frequency band, and combine local characteristics with the overall trend to provide high-quality input signals for the network. In addition, the present invention introduces a temporal attention mechanism into the network to enhance the model's ability to focus on key time steps, thereby more accurately capturing the long-term dependencies in the signal. By training the network with sensor data in the normal working state and saving the model, sensor faults can be detected efficiently and accurately, thus improving the reliability and safety of equipment operation. Summary of the Invention

[0005] The present invention aims to solve the deficiencies of the prior art. By integrating wavelet packet decomposition preprocessing, LSTM network, and temporal attention mechanism, a sequence-to-sequence vibration signal prediction model is constructed, and sensor faults are detected by comparing the predicted value with the measured value.

[0006] To achieve the above object, the present invention proposes the following technical solutions: A sensor fault detection method based on wavelet packet decomposition and long short-term memory network, comprising the following steps:

[0007] S1: Collect the vibration signals of the mechanical transmission system, and simulate sensor deviation faults and sensor impact faults on the basis of partial normal signals;

[0008] S2: Set samples and corresponding labels for the vibration signals using a sliding window;

[0009] S3: Perform wavelet packet decomposition on the samples to form multi-scale inputs and construct a data set;

[0010] Use the wavelet packet decomposition algorithm to decompose one-dimensional input samples into detail signals and approximation signals to form multi-scale inputs of different frequency bands for network training;

[0011] S4: Construct a prediction model and train it;

[0012] S5: Input the test samples into the trained model to predict future signals, and compare the output results with the true measured values to judge the working state of the sensor.

[0013] Preferably, in step S1, the vibration signal is collected by a vibration sensor.

[0014] Preferably, the vibration sensor includes: a normal vibration sensor and a fault vibration sensor.

[0015] Preferably, the fault vibration sensor includes: a soft drift fault sensor.

[0016] Preferably, the deviation fault is simulated by adding a random small signal to the normal signal.

[0017] Preferably, the impact fault is simulated by adding multiple pulse signals to the normal signal.

[0018] Preferably, in step S2, a sliding window with a step size of S and a window length of W is used to slide on the collected vibration signal to segment the samples, and the signal with a length of L after the samples is used as the corresponding label.

[0019] Preferably, in step S4, the prediction model is a sequence-to-sequence prediction model that fuses temporal attention and long short-term memory networks.

[0020] The present invention also discloses a sensor fault detection system based on wavelet packet decomposition and long short-term memory networks, including:

[0021] An acquisition module for acquiring vibration signals of a mechanical transmission system;

[0022] A construction module for constructing a prediction model;

[0023] A training module for training the prediction model;

[0024] A judgment module for comparing the predicted future signal changes of the model with the true measured values to judge the working state of the sensor.

[0025] The present invention also discloses a computer storage medium, wherein the storage medium includes computer instructions, and when it runs on a computer, it causes the computer to execute the method described in any one of the above.

[0026] The present invention also discloses an electronic device, wherein the electronic device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in any one of the above.

[0027] Compared with the prior art, the present invention has the following advantages: The multi-scale input after wavelet packet decomposition can capture the noise and complex patterns mixed in the original signal, and refine the extraction of the characteristics of each frequency band in the signal, providing high-quality input for the LSTM; The sequence-to-sequence prediction framework can effectively model the non-linear dependencies in long time series, thereby improving the prediction ability for signal trends and anomalies; The introduction of the time series attention mechanism endows the model with the ability to dynamically adjust the focus of attention on signal characteristics, so as to more accurately capture complex fault characteristics and long-term dependencies; The fault detection method based on prediction residuals not only avoids the prior assumptions about sensor fault modes and does not require fault data for training, but also has the advantages of strong real-time performance and high robustness. The present invention can also avoid misjudgment of the diagnosis system caused by sensor faults, improve the operation efficiency of the equipment and the reliability of fault detection. Through the organic combination of multiple technologies, the present invention not only overcomes the limitations of traditional methods, but also provides an efficient and accurate solution for sensor fault detection. Description of the Drawings

[0028] Figure 1 is a flowchart of a sensor fault detection method based on wavelet packet decomposition and long short-term memory network provided by an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of the vibration signal collected by the experimental bench sensor provided by an embodiment of the present invention;

[0030] Figure 3 is a schematic diagram of vibration signals in different sensor states provided by an embodiment of the present invention;

[0031] Figure 4 is an example diagram of three-layer wavelet packet decomposition of a one-dimensional vibration signal provided by an embodiment of the present invention;

[0032] Figure 5 is a framework diagram of a vibration signal prediction model provided by an embodiment of the present invention;

[0033] Figure 6 is a schematic diagram of the signal prediction result when the sensor is normal provided by an embodiment of the present invention;

[0034] Figure 7 is a schematic diagram of the signal prediction result when the sensor fails provided by an embodiment of the present invention.

[0035] The present invention will be further explained below with reference to the drawings and embodiments. Detailed Embodiments

[0036] The following will be combined with the attached Figures 1 to 7A detailed description is given of specific embodiments of the present invention. It should be noted that the embodiments shown in the drawings are only specific application forms of the present invention and should not be regarded as limitations on the present invention. The terms mentioned in this specification and claims are only for descriptive purposes, and those skilled in the art may use other terms to refer to the same components, and the key to differentiation lies in their functional differences. The expressions "comprising" or "including" mentioned herein are open-ended expressions and should be interpreted as "including but not limited to". The following description aims to clarify the core idea and implementation method of the present invention, and the specific scope of protection shall be subject to the appended claims.

[0037] For the convenience of understanding the embodiments of the present invention, the following will further explain with specific embodiments as examples in conjunction with the drawings, and each drawing does not constitute a limitation on the embodiments of the present invention.

[0038] The present invention provides a sensor fault detection method based on wavelet packet decomposition and long short-term memory network, as Figure 1 shown, the method includes the following steps:

[0039] S1: Collect the vibration signals of the mechanical transmission system, and simulate sensor deviation faults and sensor impact faults on the basis of part of the normal signals;

[0040] S2: Set samples and corresponding labels for the vibration signals by using a sliding window;

[0041] S3: Perform wavelet packet decomposition on the samples to form multi-scale inputs and construct a data set;

[0042] S4: Build a prediction model and train it using the data set and the labels;

[0043] S5: Input the test samples into the trained model to predict future signals, and compare the output results with the true measured values to further judge the working state of the sensor.

[0044] In one embodiment, a sensor fault detection method based on wavelet packet decomposition and long short-term memory network includes the following steps:

[0045] Step S1: Use a soft drift fault sensor and a normal sensor to collect the vibration signals of the mechanical transmission system respectively, randomly select part of the normal signals, and simulate sensor deviation faults and sensor impact faults on the basis of the part of the normal signals; as Figure 2 shown, the vibration sensor is installed on the outer surface of the transmission mechanism for monitoring the operating state of the mechanical system.

[0046] Step S2: Use a sliding window to segment the input samples in the collected one-dimensional vibration signals, and use the latter segment of the input samples as the label;

[0047] Step S3: Use the wavelet packet decomposition algorithm to decompose the one-dimensional input samples into detail signals and approximation signals, so as to form multi-scale inputs in different frequency bands for network training;

[0048] Step S4: Construct a sequence-to-sequence prediction model based on the Long Short-Term Memory (LSTM) network, fuse the temporal attention mechanism to adaptively assign different weights to the input time steps, input the multi-scale into the model to output the predicted values of a future segment of signals, and calculate the mean square error with the labels as the loss. Optimize the model weights through the backpropagation algorithm to continuously improve the model's ability to predict future signal changes.

[0049] Step S5: Input the normal signals and various fault signals into the trained model to predict the signal changes in the future for a period of time, and compare them with the measured values in the future for the same period of time, so as to judge the working state of the sensor. Exemplarily, for a test vibration signal with a sample length of 1000, the present invention uses the first 800-length signal to predict the last 200-length signal, and then compares the predicted values with the last 200-length signal of the original signal.

[0050] In a preferred embodiment of the described method, in step S1, the vibration sensor collects vibration signals from the mechanical transmission system, and injects two fault modes of impact fault and deviation fault into the collected vibration signals. For the deviation fault, a random small signal can be added to the normal signal for simulation. For the impact fault, multiple pulse signals can be added to the normal signal. Specifically, the simulation method of the deviation fault can be written as:

[0051]

[0052] where is the normal signal, and is the simulated deviation fault signal. represents a random small deviation signal, uniformly sampled from the interval, where is the maximum amplitude of the normal vibration signal.

[0053] The simulation method of the impact fault can be written as:

[0054]

[0055] where is the simulated soft drift fault signal, is the unit impulse function, which is 1 at and 0 at other times, is the pulse amplitude, randomly taking values in the range of and . is the time point of pulse generation, sampled from a random distribution, and N is the number of impact pulses, with a value range of [1000, 5000].

[0056] Figure 3 Shows the normal signal, soft drift fault signal, and two simulated sensor fault signals (sensor impact fault and sensor deviation fault). It can be seen from the figure that the fault signal has more drastic change characteristics compared to the normal signal, and there are significant differences in their distributions.

[0057] In the preferred embodiment of the described method, in step S2, a sliding window with a step size of S and a window length of W is used to slide in the collected one-dimensional vibration signal to segment the samples, and the signal with a final length of L of the samples is used as the corresponding label. Exemplarily, the step size is set to 100, the window length is set to 4096, and the label length is set to 128.

[0058] In the preferred embodiment of the described method, in step S3, the node coefficients of wavelet packet decomposition can be expressed as:

[0059]

[0060] Among them, and represent the low-pass filter bank and the high-pass filter bank respectively, k is the time step index, and k - 2l represents the index offset during the decimation operation on the signal; is the wavelet coefficient of node 2n + 1 in the j-th layer of wavelet packet decomposition, represents the wavelet coefficient of the n-th node in the (j - 1)-th layer, which is the input signal obtained from the decomposition of the previous layer. Here, j and n are wavelet packet node indices, and their value ranges are (1, i) and (0, 2j - 1) respectively. The 4th-order Daubechies wavelet is used as the mother wavelet to perform multi-scale decomposition on the segmented samples to obtain the multi-scale sequence . Exemplarily, the level j of wavelet packet decomposition is set to 3, then a total of 8 sub-sequences are decomposed from the original signal as the network input. Specifically, in each layer, the signal passes through a low-pass filter to extract the approximation (Approximation) signal, corresponding to the low-frequency part. The detail signal is extracted through a high-pass filter, corresponding to the high-frequency part. Subsequently, the low-frequency and high-frequency parts are further decomposed to form a complete wavelet packet tree, as Figure 4 shown. Figure 4 Shows the process of decomposing a signal with a length of 4096 into 8 sub-sequences, where at each node, the signal is decomposed into a detail signal and an approximation signal, which are represented by the letters D and A respectively.

[0061] In a preferred embodiment of the method described above, in step S4, the sequence-to-sequence vibration signal prediction model includes an encoder and a decoder, both composed of two layers of LSTM units, as Figure 5 shown. The basic LSTM unit includes an input gate , a forget gate , an output gate and a storage unit , which can be expressed as:

[0062]

[0063] where is the hidden state extracted from the input, and represent the hidden state and cell state at the previous moment, U represents the input state the corresponding weight matrix (for example represents the weight of the forget gate, represents the weight of the input gate, represents the weight of the storage unit, represents the weight of the output gate), W represents the hidden state the corresponding weight matrix (for example represents the weight of the forget gate, represents the weight of the input gate, represents the weight of the storage unit, represents the weight of the output gate), b represents the threshold ( is the threshold of the forget gate, is the threshold of the input gate, is the threshold of the output gate, is the threshold of the storage unit), σ and tanh represent the sigmoid activation function and the tanh activation function respectively, then represents the dot product operation. The LSTM network can selectively remember or forget past information, thus effectively handling tasks with long-term dependencies.

[0064] The vibration signal prediction model sequentially includes: an input layer, an encoder layer, an attention layer, a decoder layer, and an output layer. The vibration signal is processed by the input layer to obtain a multi-scale sequence , and is input into the encoder layer, and the hidden state is generated through recursive processing, where the subscript T is the length of the input sequence. Among them, in the encoder layer, for each X k (k = 1:T in the normalization step is used to traverse all input moments to ensure that the sum of the attention weights is 1), it is sequentially input into two cascaded LSTM basic units. Among them, for any two adjacent X k and X k+1 , for example, taking X1 and X 2 For example, in the encoder layer, the output of the first LSTM basic unit of X 1 is further connected to the input of the first LSTM unit of X 2 ; for the second LSTM unit of X 1 , its output is further connected to the input of the second LSTM unit of X 2 , and the output of its second LSTM basic unit serves as the input to the attention layer. The attention layer further processes the from the encoder layer into a temporal attention context vector and outputs it to the decoder layer. Among them, for any processing object in the decoder layer, that is, the kth processed object, the output of the second LSTM unit in the corresponding decoder layer, in addition to being connected to the output layer as the input to the output layer, is also fed back to the attention layer. For each h t， in the decoder layer, it is sequentially input into two cascaded LSTM basic units. Among them, for any two connected h t and h t+1 , for example, taking h 1 and h 2 as an example, the output of the first LSTM basic unit of h 1 in the decoder layer is further connected to the input of the first LSTM unit of h 2 ; for the second LSTM unit of h 1 , its output is further connected to the input of the second LSTM unit of h 2 , and the output of its second LSTM basic unit serves as the input to the output layer. The output layer further processes the output of the temporal attention layer through a Linear unit into X T+1 , X T+2 , …… X T+H and outputs them, as shown in Figure 5 , where H is the length of the prediction sequence.

[0065] In the preferred embodiment of the described method, the present invention designs a temporal attention layer to weight the hidden states, adaptively assigns different weights to the input time steps to capture long-term temporal dependencies, which can be described as:

[0066]

[0067] where is the hidden state of the decoder, and the superscript T represents the transpose operation, is the hidden state of the encoder, and their product represents the unnormalized attention weight of the output corresponding to the encoder time step t input feature at the ith prediction time step, denotes the unnormalized attention weight of the output at the \(i\)-th predicted time step corresponding to the input feature at the encoder time step \(k\). denotes the normalized attention weight, and is the weighted hidden state. The normalization method of attention is Softmax normalization, where is the exponential operation. The weighted hidden state is input into the decoder to recursively predict future signal changes. It should be noted that the role of \(k = 1:T\) in the normalization step is to traverse all input time moments to ensure that the sum of attention weights is 1, and \(T\) is the length of the input sequence. \(t = 1:T\) in the weighted summation means traversing the hidden states at the input time moments, weighting each moment of the input according to the attention weights to generate a context vector. Different indices (\(k\) and \(t\)) are used in the formula to ensure the logical distinction between the normalization operation and the weighted summation operation and avoid computational confusion.

[0068] In the preferred embodiment of the described method, in step S4, 100 normal samples are used to train the algorithm. The mean squared error loss between the output of the sequence-to-sequence prediction model and the label is used as the loss function of the model. The Adam algorithm is used to optimize the model weights. When ensuring the convergence of the model, the training loop stops at 100 and the model at this time is saved as the best model. It should be emphasized that this training process adopts an unsupervised learning method, that is, it does not need to rely on fault data and their corresponding labels for training. By learning the characteristics of normal data, the model can identify abnormal samples that deviate from the normal signal distribution, so as to realize the detection of sensor faults. The present invention better meets the actual industrial needs in terms of data dependence.

[0069] In the preferred embodiment of the described method, in step S5, multiple sensor faults are injected to form a total of 400 on normal signals and three types of fault type test samples. The test samples are input into the trained model to predict future signal changes and compared with the true measured values. If the prediction error is within a certain threshold, the sensor is considered normal, otherwise the sensor is considered faulty. The experimental results are shown in Table 1. The overall accuracy of the proposed method on the test set is 91.50%. The detection accuracy for bias faults is relatively low, but still as high as 76%, far exceeding the accuracy of the existing technologies in this field, and the detection accuracies for the other types of signals are all higher than 90%. From the experimental results, the proposed method can effectively judge sensor faults.

[0070] Table 1 Sensor Fault Detection Results

[0071]

[0072] It should be noted that the present invention can detect all faults with significant differences from the data collected by normal sensors, such as periodic interference faults and open circuit faults. It is not limited to the three simulated faults above. Only because these three fault types are relatively common, simulations are carried out to prove the effect of the present invention.

[0073] Figure 6 It shows the prediction result when a sample selected by a sliding window in a normal signal is input into the model. It can be found that the difference between the actual measurement value and the predicted value is not large. The lower right shows the change of the root mean square error index after all samples formed by this section of the signal are input. It can be seen that both the amplitude and the fluctuation degree are within an acceptable range.

[0074] Figure 7 It shows the prediction result of the sensor soft drift fault signal. It can be seen that the deviation between the predicted signal and the actual measurement value is large and the error index increases significantly when the soft drift phenomenon occurs, indicating that the sensor has a fault.

[0075] In one embodiment, the present invention also discloses a sensor fault detection system based on wavelet packet decomposition and long short-term memory network, including:

[0076] An acquisition module for acquiring vibration signals of a mechanical transmission system;

[0077] A construction module for constructing a prediction model;

[0078] A training module for training the prediction model;

[0079] A judgment module for comparing the predicted future signal change of the model with the true measurement value, and then judging the working state of the sensor.

[0080] In one embodiment, the present invention also discloses a computer storage medium, wherein the storage medium includes computer instructions, which when running on a computer, cause the computer to execute the method described in any one of the above.

[0081] In one embodiment, the present invention also discloses an electronic device, wherein the electronic device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in any one of the above.

[0082] Finally, it should be noted that the embodiments described above are only some specific implementation manners of the present application, not all embodiments. Based on the content of the present application, other implementation manners obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present application. The above content has illustrated the present invention in the form of examples, but should not be construed as a limitation on the protection scope of the claims. Without departing from the spirit and scope of the present invention, those skilled in the art can make various forms of modifications or adjustments to the embodiments, and these deformations and extensions all fall within the protection scope of the present invention.

Claims

1. A sensor fault detection method based on wavelet packet decomposition and long short-term memory network, characterized in that: It includes the following steps: S1: Collect vibration signals of the mechanical transmission system and simulate sensor deviation faults and sensor impact faults based on some normal signals; S2: using a sliding window to set samples and corresponding labels for the vibration signal; S3: performing wavelet packet decomposition on the sample to form a multi-scale input and construct a data set; S4: construct a prediction model and perform training using the data set and the labels; S5: Input the test sample into the trained model to predict the future signal, and compare the output result with the actual measurement value to determine the working status of the sensor.

2. The method according to claim 1, characterized in that Preferably, in step S1, the vibration signal is collected by a vibration sensor.

3. The method according to claim 2, characterized in that The vibration sensor includes: a normal vibration sensor and a fault vibration sensor.

4. The method according to claim 1, characterized in that: The deviation fault is simulated by adding a random small signal to the normal signal.

5. The method according to claim 1, characterized in that The impact fault is simulated by adding multiple pulse signals to the normal signal.

6. The method according to claim 1, characterized in that In step S2, a sliding window with a step size of S and a window length of W is used to slide on the collected vibration signal to segment the samples, and the signal with a length of L after the sample is used as the corresponding label.

7. The method according to claim 1, characterized in that In step S4, the prediction model is a sequence-to-sequence prediction model that integrates temporal attention and long short-term memory network.

8. A sensor fault detection system based on wavelet packet decomposition and long short-term memory network, comprising: A collection module, used for collecting vibration signals of a mechanical transmission system; Building modules for building prediction models; Training module, used to train the prediction model; The judgment module is used to compare the future signal changes predicted by the model with the actual measured values ​​to determine the working status of the sensor.

9. A computer storage medium, wherein: The storage medium includes computer instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 7.

10. An electronic device, wherein: The electronic device comprises: a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 7 when executing the program.

Images