Gearbox intelligent fault diagnosis method based on Li-MDCCAN

Abstract

The present application relates to a gearbox intelligent fault diagnosis method based on Li-MDCCAN, and a new gearbox intelligent fault diagnosis framework-lightweight multi-scale hollow causal convolution attention network (Li-MDCCAN) is designed for a noise environment, which realizes the balance between robustness and complexity. Firstly, the vibration signals along the horizontal, vertical and axial directions are fused as the network input to obtain complete fault information. Then, the hollow causal convolution is introduced to obtain multi-scale features through three parallel convolution channels, and then the lightweight convolution module and the attention module are stacked in turn to gradually excavate multi-dimensional features. Finally, the three multi-scale multi-dimensional features extracted are fused again as the output, and the global average pooling layer, the full connection layer and the multi-classification function are used to realize the gearbox intelligent fault diagnosis. The Li-MDCCAN of the present application has better diagnosis performance, stronger generalization ability and lower calculation complexity in gearbox fault diagnosis, and especially can highlight the advantages of the method in a noisy environment.

Description

Technical Field

[0001] This invention relates to the field of gearbox fault diagnosis technology, specifically to a gearbox intelligent fault diagnosis method based on Li-MDCCAN, which is particularly suitable for fault diagnosis in noisy environments. Background Technology

[0002] Gearboxes are widely used and indispensable key components in modern industry. However, due to their long operating time and complex working environment, gears and / or bearings in gearboxes are prone to failure. According to statistics, gears and bearings account for the largest proportion of gearbox component failures, at 60% and 19% respectively. Therefore, gears and bearings are critical components with high failure rates in gearboxes. Once gears and bearings fail, they can easily trigger a series of adverse chain reactions, ultimately leading to the failure of the entire mechanical equipment and even causing production safety accidents. Therefore, conducting research on gearbox fault diagnosis is of extremely important practical significance for ensuring the safe operation of mechanical equipment and improving industrial production efficiency.

[0003] Currently, gearbox vibration signals are widely recognized as the best feature extraction source, capable of rapidly, accurately, and comprehensively reflecting the equipment's operating status and the nature of most faults. Therefore, vibration diagnosis plays a crucial role in gearbox fault diagnosis. Vibration signal-based fault diagnosis methods mainly include three steps: data acquisition, feature extraction, and fault identification. Depending on the feature extraction method, they can be further divided into traditional machine learning diagnostic methods and the currently popular deep learning diagnostic methods.

[0004] Traditional machine learning diagnostic methods rely heavily on manual feature extraction, depending on the prior knowledge of diagnostic experts. This is particularly problematic when dealing with complex systems, where extracting effective fault features from noisy background signals is difficult. Furthermore, with the development of IoT technology, fault diagnosis has entered the era of big data. Monitoring systems for mechanical equipment generate vast amounts of data, with data acquisition speeds far exceeding expert analysis speeds, resulting in relatively low efficiency in fault diagnosis. Therefore, traditional machine learning diagnostic methods require extensive signal processing skills and a wealth of diagnostic experience, making it a laborious and time-consuming process.

[0005] Data-driven deep learning diagnostic methods have the advantage of automatically extracting features from a large number of samples, avoiding to some extent the tedious manual feature design process. However, on the one hand, gearboxes often operate in harsh environments with high speed and heavy loads, and coupled with noise interference in industrial settings, the collected vibration signals inevitably contain significant noise, making it very difficult to extract effective features and placing high demands on the robustness of the model. On the other hand, in order to ensure the diagnostic accuracy of the network model under noise interference, the model complexity often increases, resulting in a large number of parameters and high computational resource consumption, making it difficult to meet the deployment requirements of practical systems.

[0006] It should be noted that the above content falls within the inventor's technical knowledge and does not necessarily constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to solve the problems existing in the prior art and provide a gearbox intelligent fault diagnosis method based on Li-MDCCAN. A new lightweight multi-scale dilated causal convolutional attention network (Li-MDCCAN) is designed for fault diagnosis. In noisy environments, Li-MDCCAN has higher recognition accuracy, stronger robustness and lower computational complexity in gearbox fault diagnosis.

[0008] The present invention achieves the above objectives by adopting the following technical solutions:

[0009] The intelligent fault diagnosis method for gearboxes based on Li-MDCCAN includes the following steps:

[0010] 1. A gearbox intelligent fault diagnosis method based on Li-MDCCAN, characterized by comprising the following steps:

[0011] S1. Acquire raw vibration data

[0012] Using an accelerometer Accelerometer to measure the horizontal vibration signal of the gearbox. An accelerometer measures the vertical vibration signal of the gearbox. The vibration signal in the axial direction of the gearbox is measured to obtain the raw vibration data of the gearbox in three directions under different health conditions.

[0013] S2. Construct a lightweight multi-scale dilated causal convolutional attention network

[0014] The lightweight multi-scale dilated causal convolutional attention network includes a multi-sensor data fusion sub-network, a multi-scale multi-dimensional feature extraction sub-network, and a fault recognition sub-network connected in sequence.

[0015] S21, Multi-sensor data fusion subnetwork

[0016] S211, Install an acceleration sensor , and The first collection The data samples are respectively , and ,

[0017] By using a sliding window to sequentially extract equal-length data segments from the above data samples and stacking them together, we obtain the multi-sensor fusion samples. Its expression is as follows:

[0018]

[0019] in, This represents a function that fuses sensor data. This function sequentially splices together data fragments from single sensors to create a multi-sensor fusion dataset, and then processes the multi-sensor fusion dataset into training and testing sets.

[0020] S212. During the network training phase, the Dropout strategy is introduced to randomly delete data points of multi-sensor fusion samples, thereby increasing the network's tolerance to input data samples.

[0021] S22, Multi-scale and Multi-dimensional Feature Extraction Subnetwork

[0022] The multi-scale, multi-dimensional feature extraction sub-network includes three parallel convolution channels. Each convolution channel is composed of a multi-scale dilated causal convolution and several lightweight convolution modules stacked together. The dilated causal convolution is the first layer of the network, and a corresponding attention mechanism is embedded after each convolution operation.

[0023] The dilated causal convolution introduces a concept called the dilation rate. hyperparameters, void ratio The spacing between values ​​is defined when the convolution kernel processes data;

[0024] Suppose the input is one-dimensional time series data Using a convolution kernel size of void ratio Filter exist Swipe up to navigate to sequence elements dilated causal convolution operation The mathematical expression is:

[0025]

[0026] in, Refers to historical moment data;

[0027] The three convolutional channels are identical in network structure and hyperparameters except for their dilation rates. The dilation rates of the three convolutional channels are sequentially adjusted. , , Dilated causal convolutions extract multi-scale features from input sensor data;

[0028] The lightweight convolution module consists of two steps: temporal feature learning and spatial feature learning. First, along the direction of the time step, max pooling is performed on the input spatial features. Then, along the direction of the feature channels, the output spatial features are mixed by pointwise convolution.

[0029] The attention mechanism starts from two dimensions: space and time. It constructs corresponding spatial attention and temporal attention and embeds them after each convolution operation. Spatial attention is used to focus on which sensor the important fault information comes from, while temporal attention is used to focus on which moment in the data of the corresponding sensor the important fault information corresponds to, thereby locking in the key features that play a decisive role in the fault identification task.

[0030] S23, Fault Classification Subnetwork

[0031] S231. Set up a feature fusion layer to fuse the three multi-scale and multi-dimensional features output by the previous sub-network;

[0032] S232. Set up a global average pooling layer. By using global average pooling, the input features are compressed into a single-channel feature vector, reducing the number of parameters in the model.

[0033] S233. Set up a fully connected layer to map distributed features to the sample label space;

[0034] S234. Use a softmax feature classifier to classify the extracted features and map the output to (0,1) values. The sum of these values ​​is 1.

[0035] S3. Input the training set into a lightweight multi-scale dilated causal convolutional attention network, perform network training, and optimize network parameters.

[0036] S4. Input the test set into the trained lightweight multi-scale dilated causal convolutional attention network for testing to achieve gearbox fault diagnosis.

[0037] In step S211, the multi-sensor fusion dataset is randomly divided into a training set and a test set at a ratio of 8:2.

[0038] The convolutional channel in step S2 is composed of a multi-scale dilated causal convolution and four lightweight convolutional modules stacked together.

[0039] The present invention, employing the above-described structure, can bring the following beneficial effects:

[0040] (1) Construct an end-to-end intelligent fault diagnosis network that directly processes the original vibration signal. By reducing the reliance on prior knowledge, the proposed method has stronger adaptability and is easier to apply to industrial scenarios.

[0041] (2) Using multi-sensor vibration data as input signal overcomes the drawback that a single sensor signal cannot fully reflect the fault characteristics and can effectively solve the problem of gearbox fault diagnosis under noise interference.

[0042] (3) Design a multi-scale learning strategy, focusing on global fault characteristics. The first layer of the network adopts dilated causal convolution with different dilation rates. With the help of three parallel convolution channels, the network can automatically obtain rich temporal features from multi-sensor data to ensure the network's overall learning ability for fault signals.

[0043] (4) Design a lightweight convolution module that focuses on local fault characteristics. This module can reduce the number of model parameters and computational load while ensuring the model's feature extraction capability, thus laying the foundation for the porting of the fault diagnosis system.

[0044] (5) Design an attention mechanism to achieve adaptive identification of fault features. The importance of each feature information is obtained through similarity calculation. The features are dynamically rewarded or punished by allocating weights, which enhances the model's attention to non-noise components and weakens information unrelated to classification features, thereby improving the network's feature extraction capability.

[0045] (6) Discuss the influence of sensor installation direction on gearbox fault diagnosis results, obtain the installation priority level of sensors in different directions, and provide certain guidance for gearbox data acquisition system. Attached Figure Description

[0046] Figure 1 This is a diagram of the Li-MDCCAN network structure of the present invention;

[0047] Figure 2 This is a schematic diagram of the dilated convolution operation of the present invention;

[0048] Figure 3 This is a schematic diagram of the operation of the lightweight convolution module of the present invention;

[0049] Figure 4 This is a schematic diagram of the operation of the attention module of the present invention;

[0050] Figure 5 The experimental system used in the embodiments of the present invention;

[0051] Figure 6 The image shown is of the actual bearing used in the embodiments of the present invention.

[0052] Figure 7 This is a schematic diagram of three typical vibration signals of the present invention;

[0053] Figure 8 The diagnostic results of the four bearing datasets of this invention under different noise environments;

[0054] Figure 9 This invention relates to the impact of the multi-scale learning strategy and attention module on network diagnostic performance.

[0055] Figure 10 These are the diagnostic results of the present invention under different numbers of sensor fusion data;

[0056] Figure 11 To illustrate this invention, we use two bearing datasets to visualize the features of eight health states at different layers of the model when SNR=0dB.

[0057] Figure 12 This invention uses diagnostic results from two gear datasets under different noise environments.

[0058] Figure 13 The present invention uses the confusion matrix of two gear datasets at SNR=0dB;

[0059] Figure 14 This is a comparison of the diagnostic results of the present invention and four advanced methods under different noise environments; Detailed Implementation

[0060] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0061] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0062] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0063] The intelligent fault diagnosis method for gearboxes based on Li-MDCCAN includes the following steps:

[0064] S1. Acquire raw vibration data

[0065] Measuring points serve as the window for acquiring fault diagnosis data. In gearbox fault diagnosis based on vibration signal analysis, the ability to resist vibration deformation varies in different directions due to the influence of the transmission path. This results in different vibration manifestations of the same fault in different directions, and the directions in which different faults exhibit significant vibration manifestations also differ. Furthermore, due to the complexity of real-world operating conditions, fault characteristics are often interfered with or even masked by noise. These factors mean that single-sensor data cannot fully reflect fault characteristics, easily leading to problems such as information loss, reduced fault diagnosis accuracy, and decreased diagnostic reliability.

[0066] To avoid the aforementioned problems, three accelerometers are used to measure vibration signals in the horizontal, vertical, and axial directions, respectively, reflecting fault information from different angles and complementing each other. Compared to single-sensor data, multi-sensor data has less information loss and can provide a more accurate and complete description of fault characteristics. For each fault type, the vibration signals collected by the accelerometers in the horizontal, vertical, and axial directions are used to determine its characteristics.

[0067] Specifically, using an acceleration sensor Accelerometer to measure the horizontal vibration signal of the gearbox. An accelerometer measures the vertical vibration signal of the gearbox. The vibration signal in the axial direction of the gearbox is measured using an accelerometer. Accelerometer and acceleration sensor Vibration signals of the gearbox in three directions under different health conditions were acquired;

[0068] S2. Construct a lightweight multi-scale dilated causal convolutional attention network

[0069] like Figure 1 As shown, the lightweight multi-scale dilated causal convolutional attention network includes a multi-sensor data fusion sub-network, a multi-scale multi-dimensional feature extraction sub-network, and a fault recognition sub-network connected in sequence.

[0070] S21, Multi-sensor data fusion subnetwork

[0071] S211, Install an acceleration sensor , and The first collection The data samples are respectively , and ,

[0072] By using a sliding window to sequentially extract equal-length data segments from the above data samples and stacking them together, we obtain the multi-sensor fusion samples. Its expression is as follows:

[0073]

[0074] in, This function represents the fusion of sensor data. It sequentially splices together single-sensor data fragments to create a multi-sensor fusion dataset, which is then randomly divided into a training set and a test set in an 8:2 ratio.

[0075] The multi-sensor fusion method of the present invention uses sensor data collected from different directions to construct a fault diagnosis dataset, which comprehensively reflects the fault characteristics and achieves reliable diagnosis under real working conditions and noise interference.

[0076] S212. During the network training phase, a Dropout strategy is introduced to randomly delete data points from multi-sensor fusion samples, increasing the network's tolerance to input data samples.

[0077] S22, Multi-scale and Multi-dimensional Feature Extraction Subnetwork

[0078] The multi-scale, multi-dimensional feature extraction sub-network includes three parallel convolutional channels. Each convolutional channel consists of a multi-scale dilated causal convolution and several lightweight convolutional modules stacked together. The dilated causal convolution is the first layer of this network, strictly adhering to the causality of temporal signals, enabling coarse-grained feature extraction from long sequence samples and improving the network's tolerance to noisy scenes. Lightweight convolutional modules are used to perform convolutions from both spatial and temporal dimensions, improving the network's feature learning ability and efficiency. To adaptively identify key features and suppress irrelevant features, a corresponding attention mechanism is embedded after each convolutional operation.

[0079] One-dimensional convolutional neural networks (CNNs) slide convolutional kernels along the time axis, traversing the input data and mining its features—a process of continuously fitting the mapping relationship between data and features. However, vibration signals acquired for fault diagnosis are time-series data. This process is limited by the size of the convolutional kernel and does not consider the correlation between data points in the time-series vibration signals. The dilated causal convolution combines the characteristics of dilated and causal convolutions, broadening the receptive field of the convolutional kernel and considering the dependencies between data points. Simultaneously, it introduces a concept called the dilation rate. hyperparameters, void ratio The spacing between values ​​is defined when the convolution kernel processes data;

[0080] Suppose the input is one-dimensional time series data Using a convolution kernel size of void ratio Filter exist Swipe up to navigate to sequence elements dilated causal convolution operation The mathematical expression is:

[0081]

[0082] in, Refers to historical moment data;

[0083] Expanded convolution window size It can be calculated using the following formula:

[0084]

[0085] like Figure 2 As shown, taking a dilated causal convolution with a kernel size of 2 and a stride of 1 as an example, the effect of the dilation rate on the convolution operation is illustrated. Figure 2 (a) indicates the use of dilated causal convolution with a dilation rate of 1. In this case, the receptive field is equal to 2, which corresponds to the information of the current time step and the previous time step, which is equivalent to the standard convolution operation. Figure 2 (b) Using dilated causal convolution with a dilation rate of 4, the receptive field is equal to 5. Obviously, the receptive field of the convolution is larger, and 3 data points are skipped on the basis of the standard convolution operation. Figure 2 (c) Using dilated convolution with a dilation rate of 8, the receptive field becomes 9, skipping 4 more data points compared to dilated convolution with a dilation rate of 4, resulting in a larger receptive field. Simultaneously, to ensure the output sequence length matches the input sequence length, the convolution operation uses zero-padded vectors on the left. It can be seen that the temporal range mapped by dilated convolution is significantly increased, thus capturing more global and richer features. Furthermore, it maintains sequence sensitivity, does not incur additional computational overhead, and reduces the occurrence of model overfitting.

[0086] Furthermore, the vibration signals from different parts of the gearbox have different frequency characteristics, making the measured signal the result of superposition and coupling of multiple frequency signals. In addition, fault signals inevitably introduce low-frequency pulses, which are easily masked by strong background noise. If a single-scale convolution is used, some feature information will inevitably be missed during feature extraction, leading to reduced model accuracy and poor generalization performance. Therefore, a multi-scale strategy is introduced to sample signals of different granularities from multiple angles, providing the network with rich features and enabling it to have better fault identification capabilities.

[0087] like Figure 1 As shown, three convolutional channels with identical network structures and hyperparameter settings are constructed, except for the dilation rate. They are parallel to each other and sequentially use dilation rates of different values. , , The dilated causal convolution extracts multi-scale features from the input sensor signals, fuses data from different historical moments, ensures the integrity of the features, and suppresses noise interference to a certain extent, thereby enhancing the generalization ability of the network.

[0088] Accuracy and complexity are two metrics for judging the quality of a model, and they are often contradictory. In other words, increasing the number of variables in a model can improve its accuracy, but it can also increase its complexity, leading to the risk of overfitting. The lightweight convolutional module aims to find a balance between model accuracy and complexity. It consists of two steps: temporal feature learning and spatial feature learning. First, along the direction of the time step, max pooling is performed on the input spatial features, i.e., the time dimension operation; then, along the direction of the feature channels, pointwise convolution is used to mix the output spatial features, i.e., the spatial dimension operation. Thus, this convolutional module can capture local features, improving model performance, while reducing the model's computational complexity, making it easier to port to real-world systems.

[0089] Next, we will explain in detail how the lightweight convolution module reduces model complexity from two aspects: parameter count and computational cost. Parameter count refers to the total number of variables in the model, generally referring to trainable parameters. Computational cost refers to the amount of floating-point computation in the model. For example... Figure 3 As shown, assuming the feature map output by the previous convolution operation... .in, Indicates the feature channel dimension; This indicates the time step dimension. If using... The size of each convolutional kernel is A standard convolution with a stride of 1 and the same padding mode selected will output the following features: The matrix plane. The number of parameters required for this traditional convolution operation. The computational complexity is In comparison, to output a feature plane of the same size, the lightweight convolutional module requires fewer parameters. The computational complexity is All are reduced to the standard convolution operation. This effectively improves the speed of model fault diagnosis.

[0090] Attention mechanisms essentially mimic how the human brain's visual system processes information when observing objects. They eliminate the negative impact of external interference, focusing limited attention on important areas among a wealth of information, selecting the most relevant and critical features for the current task. This invention proposes a method that uses sensor data from three directions as input. Sensor data from different directions contains varying degrees of fault information; some sensor data has obvious fault characteristics, while others have only a few. Furthermore, even data acquired by sensors in the same direction at different times contributes differently to fault information. Therefore, spatial and temporal attention mechanisms are constructed and embedded after each convolution operation. Spatial attention focuses on which sensor the important fault information originates from, while temporal attention focuses on which moment in the data from that sensor corresponds to the key information, thus identifying the crucial features that determine the fault identification task.

[0091] like Figure 4 As shown, assuming the feature map obtained after the convolution operation is still... ; Corresponding to different sensor characteristics; This corresponds to the data features of the same sensor at different times. The purpose of spatial attention is to explore the relationships between sensor data features from different directions; it tells the model which direction of sensor data to focus on. First, along the direction of the time step, the features of each channel are aggregated through a global average pooling operation to obtain the aggregated features. The calculation formula is as follows:

[0092]

[0093] Then, the above-mentioned aggregated features are learned through a one-dimensional convolution with a kernel size of k. The kernel size... The kernel size directly affects the number of channels involved in the calculation of each weight in the attention mechanism, thus determining the coverage of cross-channel interactions. Generally, it is believed that larger channel sizes result in stronger long-term interactions, while smaller channel sizes result in stronger short-term interactions; that is, there is a certain mapping relationship between kernel size and the number of interaction channels. It can be determined by the number of channels. The function is derived from adaptive selection, and the calculation formula is as follows:

[0094]

[0095] here, Represents the closest Odd numbers; and These are constants, set to 2 and 1 respectively. Therefore, the formula for calculating spatial attention weights is as follows:

[0096]

[0097] in, Indicates the kernel size as One-dimensional convolution, This is an aggregation feature. The hardsigmoid activation function is defined as follows:

[0098]

[0099] The Hard sigmoid function is a piecewise linear approximation of the sigmoid function. This function simplifies computation while retaining the main properties of the sigmoid function. Finally, by multiplying the spatial attention weights and the input feature map along the feature channel direction, and assigning weights to the features of each channel, an intermediate feature vector is obtained. .

[0100] The purpose of temporal attention is to explore the relationships between data features from a sensor at different times; temporal attention can be seen as a complement to spatial attention, telling the model which time of day's data from a particular sensor to focus on. This is relevant for intermediate feature vectors. Along the direction of the channel features, the maximum and average values ​​of the features at different times are calculated, resulting in two features compressed into a single channel. and The calculation formula is as follows:

[0101]

[0102]

[0103] Then, these two features are concatenated along the feature channel direction, and the feature information is integrated using a one-dimensional convolution with a kernel size of 1. Finally, the hard sigmoid function is used to obtain the temporal attention weights. The calculation formula is as follows:

[0104]

[0105] Finally, temporal attention weights and intermediate feature vectors Multiply the features along the time step direction, assign weights to the features at each time step, and finally obtain the output feature vector. Then input the next convolutional module.

[0106] S23, Fault Classification Subnetwork

[0107] S231. Set up a feature fusion layer to fuse the three multi-scale and multi-dimensional features output by the previous sub-network;

[0108] S232. Set up a global average pooling layer. By using global average pooling, the input features are compressed into a single-channel feature vector, reducing the number of parameters in the model.

[0109] S233. Set up a fully connected layer to map distributed features to the sample label space;

[0110] S234. Use a softmax feature classifier to classify the extracted features and map the output to (0,1) values. The sum of these values ​​is 1.

[0111] Specifically, the three multi-scale, multi-dimensional features output from the previous sub-network are first fused and then fed into a fully connected layer for feature classification. Traditional convolutional neural networks typically use two fully connected layers. The first fully connected layer flattens and quantizes the input features, while the second fully connected layer and softmax function perform classification. However, this classification structure has a large number of parameters, leading to increased computation and significantly limiting the convergence speed of the neural network. It is also prone to overfitting, resulting in a high convergence value for the model's loss function and low accuracy. Therefore, this model replaces the first fully connected layer with a global average pooling layer. Global average pooling compresses the input features into a single-channel feature vector, reducing the number of model parameters. Then, the softmax function maps the output to (0,1) values, the sum of which is 1. This value represents the probability that the input sample belongs to each label, and the node with the largest value is selected as the prediction target. The formula is as follows:

[0112]

[0113] In the formula, Is the predicted object the first The probability of a class It is the output layer. The logits value of each neuron. This represents the number of neurons in the output layer.

[0114] S3. Input the training set into a lightweight multi-scale dilated causal convolutional attention network, perform network training, and optimize network parameters.

[0115] The training of this model is supervised training, which can be viewed as an optimization problem. Its goal is to minimize the error between the predicted value and the true label for each sample. Let... This represents the predicted value, which is the output of the Softmax function in the last layer of the network; To represent the true value, i.e., the fault type corresponding to the sample, the classification cross-entropy loss function is defined as follows:

[0116]

[0117] The Adam optimization algorithm, known for its fast and stable convergence, is selected to optimize the parameters until the aforementioned loss function is minimized. Let t be the iteration number; then the parameter vector update steps at iteration t are shown in the following equation:

[0118]

[0119]

[0120] In the formula, It is a first-order gradient matrix. It is a second-order gradient matrix. For the number of iterations, This is an estimate of the current rate of index decay, where, , .

[0121] To avoid and When the initial value is a zero vector, the resulting overall gradient weights are too small, affecting training accuracy. Therefore, a bias correction is needed for the initial value, calculated using the following formula:

[0122]

[0123]

[0124] Based on the above calculation formula, the parameters can be updated in real time. The calculation formula is as follows:

[0125]

[0126] In the formula, The learning rate is the default. ; To avoid division by zero, the step size for updating parameters depends on the initial learning rate multiplied by the ratio of the square root of the gradient mean to the gradient variance. Therefore, it can be adaptively adjusted from two perspectives: the gradient mean and the gradient variance, rather than being directly determined by the current gradient.

[0127] S4. Input the test set into the trained lightweight multi-scale dilated causal convolutional attention network for testing to achieve gearbox fault diagnosis.

[0128] Experimental verification and analysis:

[0129] This invention uses two datasets to evaluate the performance of the proposed Li-MDCCAN method and introduces ablation experiments to verify the effectiveness of each module in the method. Finally, the fault diagnosis results of Li-MDCCAN are compared with four advanced diagnostic methods.

[0130] Example 1: Gearbox Bearing Fault Diagnosis

[0131] (1) Bearing dataset

[0132] The bearing experiments were conducted on a comprehensive fault simulation test bench. For example... Figure 5 As shown, the experimental system consists of a three-phase variable frequency motor, a torque and speed sensor, a single-span rotor shaft system, rolling bearing housings, a shaft load plate, and a radial loading device. The motor's rated voltage is 380V, and its rated power is 3KW. The loading experiment applies a radial loading force to the rotating shaft via a rotating handwheel above the radial loading device, with a loading range of -1000N to 1000N. The widely used single-row cylindrical roller bearing SKF N205 is used as the research object. To ensure a relatively complete vibration signal can be collected after one bearing cycle, three accelerometers (model: RH103) are installed on the bearing housing to measure vibration signals in the horizontal (CH1), vertical (CH2), and axial (CH3) directions, respectively. Data acquisition is performed by an acquisition station (model: RH1000V4).

[0133] This dataset collected eight health states of bearings, including five single faults, two compound faults, and one normal state, as shown in Table 1 and... Figure 6 As shown in the figure, signals were collected under four different operating conditions at a sampling frequency of 51.2 kHz, denoted as dataset A, dataset B, dataset C, and dataset D. In dataset A, the drive motor speed and radial force were set to 1000 rpm and 0 N, respectively. In dataset B, the drive motor speed was maintained at 1000 rpm, and the radial force was increased to 500 N. In dataset C, the drive motor speed was adjusted to 1500 rpm, and the radial force was reduced to 0 N. In dataset D, the drive motor speed was maintained at 1500 rpm, and the radial force was increased to 500 N. Each of the four datasets contained 1000 samples, therefore, the eight datasets for bearing health status contained a total of 8000 samples, which were randomly divided into training and test sets in an 8:2 ratio. To avoid overlapping regions between samples, the original signals were segmented using non-overlapping sampling, with each sample containing 3072 sampling points.

[0134] Table 1. Detailed description of bearing health status

[0135]

[0136] (2) Specific implementation

[0137] The experiment was set up on a Dell Precision 5820 Tower workstation, with an Intel(R) Xeon(R) w-2255 CPU @ 3.70GHz and an NVIDIA TRX A4000 GPU. The software used was the Tensorflow open-source library, and the experiment was implemented in the PyCharm development environment using Python.

[0138] During the fault diagnosis process, the hyperparameters of the proposed Li-MDCCAN were adjusted and selected using a 10-fold cross-validation method on the training set. The selection results are shown in Table 2. Furthermore, in the subsequent experiments, to avoid the influence of randomness, each method was run ten times, and the average of the ten results was taken as the final test result.

[0139] Table 2 Hyperparameters of Li-MDCCAN

[0140]

[0141] (3) Performance testing in noisy scenarios

[0142] In industrial settings, vibration signals often contain significant noise interference. To investigate the fault diagnosis capabilities of Li-MDCCAN in noisy environments, different levels of noise were added to the original signal to generate noisy signals. Gaussian white noise, as one of the most typical additive noises in simulating industrial production, has an instantaneous value and power spectral density that follow a Gaussian and uniform distribution, respectively, and is widely used in noise analysis. Therefore, in the experiment, test signals were constructed by adding Gaussian white noise with different signal-to-noise ratios to the original signal. The definition is as follows:

[0143]

[0144] in, The power spectral density of the original signal. Let be the power spectral density of the noise signal. Clearly, the higher the signal-to-noise ratio (SNR), the weaker the noise interference on the vibration signal; conversely, the lower the SNR, the stronger the noise interference. And when... A value of 0dB means that the noise energy is equal to the original signal energy.

[0145] To clearly demonstrate the interference of noise on the original vibration signal, three typical health states, BN, BSF3, and BCF5, are used as examples. Figure 7 The original signal, additive white Gaussian noise, and... are given respectively. The time-domain waveform of the composite noise signal equal to 0dB. It can be seen that no pulses were observed in BN, while obvious pulses were observed in BSF3 and BCF5. Although the vibration signals of these three health states exhibit different characteristics, it is still difficult to directly identify the matching health state. In particular, due to the presence of additive noise, the periodic impact component of the noisy signal is significantly weakened compared to the original vibration signal, making the vibration characteristics of the original signal even more difficult to distinguish by the human eye, greatly increasing the difficulty of fault diagnosis.

[0146] Therefore, to ensure the effectiveness of the proposed model, it was tested under noise levels ranging from -3dB to 9dB. Notably, during training, the input signal was randomly corrupted by dropout, rather than by adding white noise. During testing, the test signal was corrupted by adding white noise, while dropout did not interfere with the test signal. This training and testing method more closely resembles real-world industrial processes.

[0147] The proposed Li-MDCCAN model is tested using four datasets, each containing varying levels of noise. For example... Figure 8 As shown, when SNR = -3dB, 0dB, 3dB, 6dB, and 9dB, the average test accuracies are 82.41%, 90.31%, 96.38%, 98.53%, and 100%, respectively. Here, SNR = -3dB means the noise is almost twice the energy of the original signal. In other words, the original signal is completely submerged in noise. Despite this, the test accuracy is still above 80%, indicating that the model has strong noise resistance. Furthermore, tests using four different datasets all yielded similar performance, fully demonstrating that the model performs well and stably in noisy scenarios.

[0148] (4) Multi-scale learning strategies and attention modules

[0149] To explore the role of multi-scale learning strategies and attention mechanisms in Li-MDCCAN, three networks were constructed to compare their noise resistance performance. The networks were labeled Network 1, Network 2, and Network 3, depending on whether they incorporated multi-scale learning strategies and attention mechanisms. These three networks are similar to the Li-MDCCAN architecture, with subtle differences: 1) Network 1 employs neither multi-scale learning strategies nor attention mechanisms; 2) Network 2 employs only multi-scale learning strategies; and 3) Network 3 employs only attention mechanisms. Furthermore, the hyperparameter settings of the three networks are identical to those of Li-MDCCAN.

[0150] Figure 9The average diagnostic results of the four networks described above are shown under different noise levels and operating conditions. Network 1 is equivalent to a standard convolutional network; obviously, since it does not incorporate a multi-scale learning strategy or an attention mechanism, its diagnostic accuracy is the lowest. Network 2, due to the application of a multi-scale learning strategy, is able to learn richer and more complete representations, thereby improving its diagnostic accuracy. Figure 9 As shown in (d), when SNR=9dB, the test accuracy only slightly improves from 99.99% to 100%. When SNR=-3dB, the test accuracy shows a significant improvement, increasing from 72.37% to 77.13%. This demonstrates that as noise intensity increases, the improvement in test accuracy becomes more pronounced, and the multi-scale learning strategy becomes increasingly effective. Network 3, by employing an attention mechanism, adaptively identifies fault features from both spatial and temporal dimensions, resulting in a certain degree of performance improvement. Figure 9 In (b), the diagnostic results improved to some extent when SNR >= 0 dB. However, when SNR = -3 dB, the test accuracy decreased slightly, from 71.84% to 70.41%. Similar test results were observed in other datasets, indicating that the attention mechanism works well when noise does not completely cover the original signal. In other words, without the assistance of other strategies, when the signal-to-noise ratio is negative, i.e., the additive noise is several times that of the original signal, fault features are difficult to extract accurately. Therefore, by systematically combining multi-scale learning strategies and attention mechanisms to extract multi-scale and multi-dimensional features, the proposed Li-MDCCAN exhibits excellent diagnostic performance in noisy environments.

[0151] (5) The necessity of lightweight convolution modules

[0152] To test the lightweight convolutional blocks used in the proposed Li-MDCCAN, four diagnostic networks with almost identical structures to Li-MDCCAN were designed. The difference lies in whether the four feature extraction layers use standard convolutional modules or lightweight convolutional modules. If all feature extraction layers use standard convolutional modules, the network is named Network 4. Then, by replacing each standard module with a lightweight convolutional module from the shallow to the deep feature extraction layers, a new network is generated, which are subsequently named Network 5, Network 6, and Network 7.

[0153] Table 3 shows the comparison results of the five networks in terms of diagnostic accuracy and model complexity. Overall, under different noise environments, the proposed Li-MDCCAN exhibits better diagnostic accuracy and lower computational complexity than the other four networks. Regarding diagnostic accuracy, the proposed network demonstrates strong robustness in noisy environments. For example, when the signal-to-noise ratio is 0dB, replacing each standard convolutional module with a lightweight convolutional module improves the average test accuracy by approximately 2%. Furthermore, due to the use of lightweight convolutional modules, Li-MDCCAN achieves the highest test accuracy. In terms of model complexity, the proposed network has the lowest computational complexity, requiring only 68,500 Params and 541,310 Flops, less than half that of Network 4. Moreover, both complexity metrics show significant improvement after replacing the standard convolutional modules with lightweight convolutional modules. Therefore, the proposed Li-MDCCAN, due to its use of lightweight convolutional modules, has significant advantages in both diagnostic accuracy and model complexity, and can be well ported to fault diagnosis systems with limited hardware resources.

[0154] Table 3. Performance comparison of networks composed of different numbers of lightweight convolutional modules.

[0155]

[0156] (6) Multi-sensor data fusion analysis

[0157] Considering that different health states exhibit different vibration characteristics, three accelerometers are used to acquire vibration signals in the horizontal, vertical, and axial directions, respectively. These three signals are then fused in the proposed Li-MDCCAN for fault diagnosis. To illustrate the necessity of multi-sensor data fusion, Figure 10 The diagnostic results are compared using seven types of sensor data. For clarity, H, V, and A represent single-sensor data acquired from the horizontal, vertical, and axial directions, respectively. Similarly, HV, HA, and VA represent the fusion of any two of the above single-sensor data. HVA represents the fusion of all three sensor data, i.e., the multi-sensor data used in Li-MDCCAN.

[0158] first, Figure 10 (a) The test accuracy was compared under different types of sensor data. Regardless of the signal-to-noise ratio, the test accuracy of single-sensor data was consistently the lowest, the test accuracy of dual-sensor data was intermediate, and the test accuracy of multi-sensor data was the highest. The same experiment was conducted on three other datasets. Figure 10 (b) Figure 10 (c) and Figure 10 The same diagnostic result was obtained in (d), which once again proves the necessity of multi-sensor data fusion.

[0159] However, due to limitations in data acquisition systems and on-site installation conditions in actual engineering projects, it is sometimes impossible to install accelerometers in all three directions simultaneously. Therefore, the diagnostic results described above are re-analyzed from the perspectives of noise signals and datasets to better guide engineering practice. Taking dataset B as an example, from... Figure 10 (f) It can be seen that the multi-sensor data achieved the highest accuracy of 94.45%. Next, ranking the two sensor data sets HA, HV, and VA, the test accuracies were 89.34%, 88.35%, and 87.43%, respectively. Following that were the single-sensor data sets H and A, with accuracies of 85.32% and 83.94%, respectively. Finally, the single-sensor data set V achieved the lowest accuracy at only 81.99%. It is important to note that, as two completely different perspectives, from... Figure 10 (f) and Figure 10 The conclusion reached in (e) is consistent. Furthermore, in Figure 10 In (e), as the noise intensity decreases, the difference in diagnostic results from different sensor data becomes smaller. This means that in a low-noise environment, each sensor data will achieve equivalent performance. However, under relatively strong noise conditions, multi-sensor data fusion is undoubtedly the optimal choice.

[0160] (7) Visualization of network models

[0161] To better understand the feature learning process of Li-MDCCAN, the extracted features are dimensionality-reduced and visualized using the t-SNE algorithm, where different bearing health conditions are represented by different colors. Datasets C and D with SNR=0dB are selected for study. The features extracted from the second convolutional layer C2, the fourth convolutional layer C4, and the last fully connected layer are mapped to two-dimensional features, and the visualization results are shown below. Figure 11 As shown.

[0162] As can be seen, the extracted features have already begun to cluster in convolutional layer C2. With the increase in the number of network layers, Li-MDCCAN can extract more abstract and higher-level features. The feature distributions corresponding to the eight health states of the bearing gradually separate, and the classifiability improves. After the operation in convolutional layer C4, the clustering of the extracted features can clearly identify BSF4, BCF5, and BCF6. In the last fully connected layer, except for slight overlap between BSF0, BSF3, and BN, most conditions show obvious clustering. Experimental results show that Li-MDCCAN can extract more effective features from the original signal with SNR=0dB, while also possessing strong generalization ability.

[0163] Example 2: Gear Fault Diagnosis in Gearbox

[0164] (1) Gear dataset

[0165] To further test the performance of Li-MDCCAN, experiments were conducted using the publicly available gearbox dataset from Southeast University. The experimental setup mainly consists of a planetary gearbox, a motor, and a braking system. Accelerometers were placed in three directions of the planetary gearbox to collect vibration signals. The dataset collected five health states of the gears, as shown in Table 4. Therefore, fault diagnosis of the gear dataset can be considered a five-class classification task. Furthermore, the experimental setup's speed-load system was set to two operating conditions: 20Hz-0V and 30Hz-2V, designated as datasets E and F, respectively. Each dataset contains 5000 samples, with a length of 3072, and the samples were divided according to the same proportion as the bearing samples.

[0166] Table 4. Detailed description of gear health status

[0167]

[0168] (2) Noise resistance experiment results and visualization

[0169] To fully verify the diagnostic performance of Li-MDCCAN in noisy environments, noise immunity experiments were conducted using the gear dataset. Figure 12 As can be seen, the model demonstrates high accuracy regardless of the dataset used. For example, when SNR=0dB, the model's average test accuracy is 93.3%. When SNR=9dB, the accuracy even reaches 100%.

[0170] To further analyze the above results, a confusion matrix was used to specifically reveal the classification accuracy of each health state. For example... Figure 13 As shown in (a), when SNR=0dB, the testing accuracy for most health conditions is above 95%. Furthermore, the model's identification discrepancies are mainly concentrated on GSF1, with 1% of GSF1 cases misclassified as GSF2 and 13% of GSF1 cases misclassified as GSF3. In addition, Figure 13 (b) also shows a similar classification result. Clearly, except for GSF1, all health conditions achieved accuracy rates above 90%, indicating good noise resistance of the model. However, it also reminds us that under noise interference, conditions GSF1 and GSF3 are easily confused.

[0171] Example 3, Method Comparison

[0172] Unlike most intelligent fault diagnosis networks, the proposed Li-MDCCAN studies the model's diagnostic performance in terms of accuracy, robustness, and complexity. To demonstrate the effectiveness and superiority of the proposed method, four state-of-the-art methods are compared: one-dimensional convolutional long short-term memory network (1DCLN), lightweight neural network (LiNet), improved multi-scale convolutional neural network with feature attention mechanism (IMS-FACNN), and multi-scale convolutional model based on multiple dilation rate and multiple attention mechanism (MDRMA-MSCN). Among them, 1DCLN and MDRMA-MSCN mainly focus on improving the robustness of the model, achieving approximately 95% and 90% accuracy, respectively, at SNR=0dB. The research on LiNet focuses on improving the efficiency of fault diagnosis, requiring only 14816 trainable parameters. IMS-FACNN has the same research objective as the proposed network and has been proven to outperform all existing methods. Therefore, the comparison of the above five methods is sufficient to demonstrate the effectiveness and superiority of the proposed Li-MDCCAN method (comparison results are shown in Table 5).

[0173] First, regarding robustness, the five methods were compared using six different datasets with signal-to-noise ratios ranging from 9dB to -3dB. To ensure fairness, the same original vibration signals were directly input for model training, and corresponding noise signals were used for testing. Furthermore, each experiment was performed 10 times, and the average of the results was taken as the model's test accuracy. Figure 14 As shown, when SNR=9dB, i.e. in a weak noise environment, the accuracy of each model approaches or even reaches 100%, demonstrating good feature extraction capabilities. However, with the increase of noise level, the accuracy of all five models decreases to varying degrees. Table IV presents the average accuracy of the five models tested on six datasets. When the SNR decreases from 9dB to -3dB, the accuracy of LiNet and 1DCLN decreases by 28.13% and 22.74%, respectively. The other three models, IMS-FACNN, MDRMA-MSCN, and Li-MDCCAN, decrease by 19.3%, 18.19%, and 17.09%, respectively, significantly outperforming LiNet and 1DCLN. Although these three models all experienced similar declines, the proposed Li-MDCCAN model exhibits more stable accuracy in varying noise compared to other models. In other words, Li-MDCCAN's advantage is more pronounced in noisy environments.

[0174] Secondly, in terms of complexity, these five methods still exhibit varying performance. LiNet has the lowest complexity, with 15224 Params and 119998 Flops. The Li-MDCCAN algorithm follows closely behind, ranking second with 68500 Params and 541310 Flops. Next is IMS-FACNN, containing 78113 Params and 613163 Flops. In stark contrast, 1DCLN requires 1081864 Params and 3935161 Flops, while MDRMA-MSCN requires 5905631 Params and 47226471 Flops, making it the most complex of the five models.

[0175] Third, while MDRMA-MSCN exhibits strong robustness, it is computationally intensive; while LiNet has low complexity but performs poorly in noisy environments. Clearly, balancing robustness and complexity is not an easy task. Therefore, the proposed Li-MDCCAN demonstrates both high robustness and low complexity, outperforming the other four diagnostic methods overall. This result further confirms the advantages of multi-scale learning strategies, lightweight convolutional modules, and attention mechanisms.

[0176] Table 5 Performance Comparison of Five Diagnostic Methods

[0177]

[0178] The above analysis verifies that the novel intelligent fault diagnosis framework Li-MDCCAN designed in this invention achieves a balance between robustness and complexity. Li-MDCCAN's strong robustness is a result of combining a dual-fusion mechanism with the extraction of multi-scale, multi-sensor features. Li-MDCCAN provides accurate diagnostic results, satisfactory noise immunity, and user-friendly portability. Furthermore, the impact of different sensor placement schemes on diagnostic performance is discussed to better guide engineering practice.

[0179] The above specific embodiments should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, any alternative improvements or modifications made to the embodiments of the present invention shall fall within the scope of protection of the present invention.

[0180] Any aspects of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A gearbox intelligent fault diagnosis method based on Li-MDCCAN, characterized in that, Includes the following steps: S1. Acquire raw vibration data Using an accelerometer Accelerometer to measure the horizontal vibration signal of the gearbox. An accelerometer measures the vertical vibration signal of the gearbox. The vibration signal in the axial direction of the gearbox is measured to obtain the raw vibration data of the gearbox in three directions under different health conditions. S2. Construct a lightweight multi-scale dilated causal convolutional attention network The lightweight multi-scale dilated causal convolutional attention network includes a multi-sensor data fusion sub-network, a multi-scale multi-dimensional feature extraction sub-network, and a fault recognition sub-network connected in sequence. S21, Multi-sensor data fusion subnetwork S211, Install an acceleration sensor , and The first collection The data samples are respectively , and , By using a sliding window to sequentially extract equal-length data segments from the above data samples and stacking them together, we obtain the multi-sensor fusion samples. Its expression is as follows: ； in, This represents a function that fuses sensor data. This function sequentially splices together data fragments from single sensors to create a multi-sensor fusion dataset, and then processes the multi-sensor fusion dataset into training and testing sets. S212. During the network training phase, the Dropout strategy is introduced to randomly delete data points of multi-sensor fusion samples, thereby increasing the network's tolerance to input data samples. S22, Multi-scale and Multi-dimensional Feature Extraction Subnetwork The multi-scale, multi-dimensional feature extraction sub-network includes three parallel convolution channels. Each convolution channel is composed of a multi-scale dilated causal convolution and several lightweight convolution modules stacked together. The dilated causal convolution is the first layer of the network, and a corresponding attention mechanism is embedded after each convolution operation. The dilated causal convolution introduces a concept called the dilation rate. hyperparameters, void ratio The spacing between values ​​is defined when the convolution kernel processes data; Suppose the input is one-dimensional time series data Using a convolution kernel size of void ratio Filter exist Swipe up to navigate to sequence elements dilated causal convolution operation The mathematical expression is: ； in, Refers to historical moment data; The three convolutional channels are identical in network structure and hyperparameters except for their dilation rates. The dilation rates of the three convolutional channels are sequentially adjusted. , , Dilated causal convolutions extract multi-scale features from input sensor data; The lightweight convolution module consists of two steps: temporal feature learning and spatial feature learning. First, along the direction of the time step, max pooling is performed on the input spatial features. Then, along the direction of the feature channels, the output spatial features are mixed by pointwise convolution. The attention mechanism starts from two dimensions: space and time. It constructs corresponding spatial attention and temporal attention and embeds them after each convolution operation. Spatial attention is used to focus on which sensor the important fault information comes from, while temporal attention is used to focus on which moment in the data of the corresponding sensor the important fault information corresponds to, thereby locking in the key features that play a decisive role in the fault identification task. S23, Fault Classification Subnetwork S231. Set up a feature fusion layer to fuse the three multi-scale and multi-dimensional features output by the previous sub-network; S232. Set up a global average pooling layer. By using global average pooling, the input features are compressed into a single-channel feature vector, reducing the number of parameters in the model. S233. Set up a fully connected layer to map distributed features to the sample label space; S234. Use a softmax feature classifier to classify the extracted features and map the output to (0,1) values. The sum of these values ​​is 1. S3. Input the training samples into a lightweight multi-scale dilated causal convolutional attention network, perform network training, and optimize network parameters. S4. Input the test samples into the trained lightweight multi-scale dilated causal convolutional attention network for testing to achieve gearbox fault diagnosis.

2. The intelligent fault diagnosis method for gearboxes based on Li-MDCCAN according to claim 1, characterized in that, In step S211, the multi-sensor fusion dataset is randomly divided into a training set and a test set at a ratio of 8:

2.

3. The intelligent fault diagnosis method for gearboxes based on Li-MDCCAN according to claim 2, characterized in that, The convolutional channel in step S2 is composed of a multi-scale dilated causal convolution and four lightweight convolutional modules stacked together.

Images