Fault diagnosis method and device
By introducing a pulse neural network with self-attention mechanism into the fault diagnosis model of autonomous underwater vehicles, the problems of poor fault diagnosis and explanation capabilities and high energy consumption in the existing technology are solved, and more efficient and accurate fault detection is achieved .
Patent Information
- Application Number
- CN202410978815.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-19
AI Technical Summary
In complex and unpredictable deep-sea environments, autonomous underwater vehicles fail tasks due to equipment or components failures, which poses a high risk. The existing fault diagnosis technology has poor interpretation capabilities, high resource consumption and is difficult to meet the ability to operate autonomously for a long time. Consume demand.
The fault diagnosis model is constructed using pulse neural network based on self-attention mechanism. Through the convolutional coding layer, feature extraction layer and output layer, the self-attention layer calculates the attention weight of sequence features, and ignores unimportant features to improve diagnostic accuracy. and reduce energy consumption.
It improves the accuracy and energy efficiency of fault diagnosis, reduces the impact of unimportant features on diagnostic results, reduces the power consumption of fault detection, and achieves more efficient fault detection.
Smart Images

Figure CN118940806A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of this specification relate to the field of computers, and more particularly to a fault diagnosis method and device. Background Art
[0002] With the continuous development of artificial intelligence technology, the scope of application of autonomous unmanned systems developed based on artificial intelligence technology is constantly expanding due to their remarkable autonomy and adaptability, such as drones, self-driving vehicles, deep space probes, autonomous underwater vehicles, etc.
[0003] Take autonomous underwater vehicles as an example. Due to their autonomous operation capabilities in deep-sea environments and long-distance and high-precision operating performance, they have been widely used in marine applications, such as offshore oil and gas exploration, maritime rescue operations, and marine environmental surveys. However, due to the complex and unpredictable environment, the use of autonomous underwater vehicles also has certain risks, mainly concentrated on the failure of the mission caused by the failure of equipment or components, and may even lead to the loss of the autonomous underwater vehicle.
[0004] Therefore, it is necessary to provide a solution that can autonomously perform fault diagnosis to ensure the safety and reliability of such autonomous unmanned systems. Summary of the invention
[0005] The embodiments of this specification provide a fault diagnosis method and device.
[0006] According to a first aspect of an embodiment of the present specification, a fault diagnosis method is provided, which is applied to an autonomous unmanned system, wherein the autonomous unmanned system is equipped with multiple sensors and a fault diagnosis model constructed by a pulse neural network based on a self-attention mechanism; the pulse neural network includes a convolutional coding layer, a feature extraction layer and an output layer, the feature extraction layer includes a first type of convolutional layer for extracting features, and a self-attention layer connected to the first type of convolutional layer, and the output layer includes a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer; the method includes:
[0007] Acquire multi-sensor signals representing the operating state of the autonomous unmanned system collected by the multiple sensors; wherein the signals carry the collection time;
[0008] The multivariate sensor signal is input into the convolutional coding layer in the fault diagnosis model, and the convolutional coding layer encodes the multivariate sensor signal into a pulse space-time sequence according to the acquisition time of the sensor signal, and outputs the pulse space-time sequence to the feature extraction layer; the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse space-time sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence feature, and further outputs the sequence feature and the attention weight corresponding to the sequence feature to the output layer; the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features, and outputs the prediction features to the fully connected layer, and the fully connected layer performs fault detection on the autonomous unmanned system based on the prediction features.
[0009] Optionally, the feature extraction layer includes at least two groups of first-type convolutional layers for extracting features stacked in a serial order, and a self-attention layer connected to the first-type convolutional layers.
[0010] Optionally, both the first type of convolutional layer and the second type of convolutional layer include a neuron model for simulating neurons of a human brain, and the neuron model includes:
[0011]
[0012] Among them, τ m represents the membrane time constant, V(t) represents the membrane potential of the neuron, V rest represents the resting potential of the neuron, R m represents the leakage resistance of the neuron, and I(t) represents the input current of the neuron.
[0013] Optionally, the first type of convolutional layer and the second type of convolutional layer further have an internal cyclic feature learning algorithm, so that the first type of convolutional layer and the second type of convolutional layer become cyclic convolutional layers; wherein the cyclic feature learning algorithm includes:
[0014]
[0015] in, represents the potential change of the neuron simulated by the recurrent convolution layer, S[t] represents the input feature of the recurrent convolution layer at time t, which is used to simulate the input potential of the neuron at time t; V[t-1] represents the module potential of the neuron at time t-1, and H[t] represents the membrane potential of the neuron at time t; 1 / τ m represents the update gate used to retain important features in S[t], 1-1 / τ mrepresents a forget gate for discarding unimportant features in V[t-1]; the important features include features with high relevance to fault detection, and the unimportant features include features with low relevance to fault detection.
[0016] Optionally, when the membrane potential of the neuron exceeds the discharge threshold, the neuron will discharge and reset the membrane potential;
[0017] The discharge includes passing the important features in S[t] retained by the update gate as new input potentials to the neuron that resets the membrane potential for iterative processing.
[0018] Optionally, calculating the attention weight corresponding to the sequence feature by a self-attention layer connected to the convolutional layer includes:
[0019] A self-attention layer connected to the convolutional layer encodes each type of feature in the sequence features into a first position code and a second position code according to a time step;
[0020] Performing matrix multiplication on the first position code and the second position code of each time step to obtain an attention score representing the similarity between the first position code and the second position code;
[0021] The attention scores are normalized to obtain attention weights corresponding to the sequence features.
[0022] Optionally, the fault diagnosis model is trained using a gradient substitution algorithm to optimize model parameters in a gradient back-propagation manner.
[0023] Optionally, the gradient substitution algorithm includes a Gaussian error function;
[0024] The optimization of model parameters by using the gradient substitution algorithm in a gradient back propagation manner includes:
[0025] The gradient generated by the Gaussian error function replaces the gradient required by the transition function in the gradient back propagation, and the replaced gradient is used to optimize the model parameters in a back propagation manner.
[0026] Optionally, the loss function of the fault diagnosis model includes:
[0027]
[0028] Where N represents the number of training samples. A coded value indicating the type of fault, Represents the predicted features of the nth training sample.
[0029] Optionally, the autonomous unmanned system includes an autonomous underwater vehicle.
[0030] According to a second aspect of an embodiment of the present specification, there is provided a fault diagnosis device, which is applied to an autonomous unmanned system, wherein the autonomous unmanned system is equipped with a plurality of sensors and a fault diagnosis model constructed by a pulse neural network based on a self-attention mechanism; the pulse neural network comprises a convolutional coding layer, a feature extraction layer and an output layer, the feature extraction layer comprises a first type of convolutional layer for extracting features, and a self-attention layer connected to the first type of convolutional layer, the output layer comprises a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer; the device comprises:
[0031] An acquisition unit, which acquires multi-sensor signals representing the operating state of the autonomous unmanned system acquired by the multiple sensors; wherein the signals carry an acquisition time;
[0032] The diagnosis unit inputs the multi-sensor signal into the convolutional coding layer in the fault diagnosis model, and the convolutional coding layer encodes the multi-sensor signal into a pulse space-time sequence according to the acquisition time of the sensor signal, and outputs the pulse space-time sequence to the feature extraction layer; the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse space-time sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence feature, and further outputs the sequence feature and the attention weight corresponding to the sequence feature to the output layer; the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features, and outputs the prediction features to the fully connected layer, and the fully connected layer performs fault detection on the autonomous unmanned system based on the prediction features.
[0033] According to a third aspect of an embodiment of the present specification, there is provided an autonomous underwater vehicle, on which an autonomous unmanned system is deployed, and the autonomous unmanned system implements any one of the above methods.
[0034] The embodiment of this specification provides a fault diagnosis scheme. A self-attention coding mechanism is introduced into the fault diagnosis model constructed based on the pulse neural network (a self-attention layer is connected to the first type of convolution layer for extracting features), so as to construct a fault diagnosis model constructed based on the pulse neural network of the self-attention mechanism. In this way, the attention weight of the sequence features extracted by the first type of convolution layer is calculated by the self-attention coding layer, so that the second type of convolution layer in the output layer can extract important features, that is, features related to fault diagnosis, from the sequence features based on the attention weight. On the one hand, since the second type of convolution layer ignores unimportant features, it can avoid reducing the accuracy of the diagnosis results due to unimportant features, thereby improving the accuracy of the diagnosis results; on the other hand, reducing unimportant features can reduce the power consumption generated by the fully connected layer, thereby further reducing the power consumption of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flow chart of a fault diagnosis method provided by an embodiment of this specification;
[0036] Figure 2 A schematic diagram of the architecture of a pulse neural network provided in one embodiment of this specification;
[0037] Figure 3 A schematic diagram of a fault diagnosis model constructed by a spiking neural network based on a self-attention mechanism provided in an embodiment of this specification;
[0038] Figure 4 A schematic diagram of a learnable recursive feature learning mechanism provided in an embodiment of this specification;
[0039] Figure 5 A schematic diagram of a self-attention layer provided in one embodiment of this specification;
[0040] Figure 6 A schematic diagram of components and common faults of an autonomous underwater vehicle provided in an embodiment of this specification;
[0041] Figure 7 is a hardware structure diagram of a fault diagnosis device provided in an embodiment of this specification;
[0042] Figure 8 It is a module of the fault diagnosis device provided in an embodiment of this specification. DETAILED DESCRIPTION
[0043] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this specification. Instead, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.
[0044] The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit this specification. The singular forms "a", "the" and "the" used in this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.
[0045] It should be understood that although the terms first, second, third, etc. may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0046] With the continuous development of artificial intelligence technology, the autonomous unmanned systems (also called intelligent unmanned systems) developed based on artificial intelligence technology have been continuously expanding their application scope due to their remarkable autonomy and adaptability, such as drones, self-driving vehicles, deep space probes, autonomous underwater vehicles (AUV), etc.
[0047] Taking autonomous underwater vehicles as an example, due to their autonomous operation capabilities in deep-sea environments and long-distance and high-precision operating performance, they have been widely used in marine applications. For example, they have played an important role in offshore oil and gas exploration, maritime rescue operations, and marine environmental surveys.
[0048] However, due to the complex and unpredictable environment, the use of autonomous underwater vehicles also has certain risks, mainly concentrated on the mission failure caused by the failure of equipment or components, and may even lead to the loss of the autonomous underwater vehicle.
[0049] Therefore, a solution that can autonomously perform fault diagnosis is needed, which is particularly important for ensuring the safety and reliability of such autonomous unmanned systems.
[0050] Among the related technologies, the development of fault diagnosis technology for autonomous unmanned systems has gone through the following stages:
[0051] The first stage is data-driven fault diagnosis technology. With the advancement of sensor and operation monitoring technology, a large amount of operation data of autonomous unmanned systems has been accumulated. The use of data-driven fault diagnosis algorithms to mine data features of useful data from these accumulated historical data sets and establish a nonlinear mapping from data features to fault types has gradually matured.
[0052] However, data-driven algorithms mainly rely on experienced experts to manually select meaningful features from the collected historical data sets, and then use these features for fault diagnosis based on machine learning classifiers. However, the number of manually selected features is small, and there are human subjective influences, which leads to the limited nonlinear characterization ability of this type of method, making it difficult to accurately diagnose faults in autonomous unmanned systems.
[0053] In the second stage, with the progress of artificial neural networks (ANNs) in various fields, it became possible to apply ANNs to fault diagnosis of autonomous unmanned systems. Among the various variant networks of ANNs, convolutional neural networks (CNNs) have gradually become the mainstream solution for fault diagnosis of autonomous unmanned systems due to their excellent performance in image analysis and signal processing. However, although ANNs are inspired by the human brain, they lack the dynamic mechanism of neurons in the human brain and cannot accurately simulate the complex neural operations of the human brain. This limitation leads to poor explanatory power, limited generalization performance, and high resource consumption of existing ANNs models.
[0054] In the third stage, due to the limitations of artificial neural networks, researchers have built a neural network model with stronger cognitive ability by deeply studying the characteristics of the human brain, namely the spiking neural network (SNNs). SNNs imitate the information transmission mechanism and dynamic characteristics of neurons in the brain, have unique spatiotemporal dynamics, diverse encoding machines, and advantages in time series processing, and are known as the third generation of neural network models.
[0055] However, there are still some technical problems to be overcome in the research and application of SNNs, especially in the fault identification of autonomous unmanned systems. For example, autonomous unmanned systems are highly sensitive to energy consumption because they need to operate autonomously for a long time. Although SNNs themselves have the characteristics of low energy consumption, they still cannot meet the use requirements of such autonomous unmanned systems that are sensitive to energy consumption in fault diagnosis, that is, how to further reduce energy consumption without reducing fault diagnosis performance.
[0056] In view of this, the present invention proposes a new fault diagnosis scheme, which introduces a self-attention coding mechanism into the fault diagnosis model constructed based on the pulse neural network, thereby connecting a self-attention layer to the first type of convolution layer used to extract features in the pulse neural network, so that a fault diagnosis model constructed based on the pulse neural network with a self-attention mechanism can be constructed. In this way, the attention weight of the sequence features extracted by the first type of convolution layer is calculated through the self-attention coding layer, so that the second type of convolution layer in the output layer can extract important features (i.e., prediction features related to fault detection) from the sequence features based on the attention weight and ignore unimportant features. On the one hand, it can avoid reducing the accuracy of the diagnosis result due to unimportant features (i.e., the embodiment can improve the accuracy of fault diagnosis); on the other hand, reducing unimportant features can reduce the power consumption generated by the fully connected layer, thereby achieving the purpose of further reducing the power consumption of fault detection.
[0057] Please refer to the following Figure 1 The schematic diagram of the fault diagnosis method flow chart shown in the figure can be applied to an autonomous unmanned system, wherein the autonomous unmanned system is equipped with multiple sensors and a pre-trained fault diagnosis model constructed by a spiking neural network based on a self-attention mechanism (SRA-SNN); wherein the spiking neural network based on the self-attention mechanism includes a convolutional coding layer, a feature extraction layer and an output layer, wherein the feature extraction layer includes a first type of convolutional layer for extracting features, and a self-attention layer connected to the first type of convolutional layer, and the output layer includes a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer; the method may include the following steps:
[0058] Step 110, obtaining multi-sensor signals representing the operating status of the autonomous unmanned system collected by the multiple sensors; wherein the signals carry the collection time.
[0059] A plurality of sensors are installed on the autonomous unmanned system, and the plurality of sensors can collect multi-sensor signals when the autonomous unmanned system is running, and then the collected multi-sensor signals are used to perform fault detection on the autonomous unmanned system.
[0060] The multi-sensor signals collected in this embodiment may include but are not limited to: acceleration of the XYZ axis, angular velocity of the XYZ axis, longitude and latitude of the autonomous unmanned system, voltage and current of the power equipment, etc. How to use these sensor signals will be introduced in subsequent embodiments.
[0061] Since sensors collect signals continuously or periodically, the sensor signal collected by a sensor can be regarded as a data stream; and the data stream can be represented as a time series vector in, It can be the number of data in the data stream represented by the time series vector Z. Assuming there are y sensors, the data stream consisting of y-element sensor signals can be expressed as
[0062] In addition, since the sensor signals collected by different sensors may have different dimensions, in order to eliminate the dimension effect, the multi-sensor signals may be pre-processed to obtain standardized sensor signals. For example, the Z-score standardization algorithm may be used to standardize the multi-sensor signals.
[0063] Step 120, input the multi-sensor signal to the convolutional coding layer in the fault diagnosis model, the convolutional coding layer encodes the multi-sensor signal into a pulse space-time sequence according to the acquisition time of the sensor signal, and outputs the pulse space-time sequence to the feature extraction layer; the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse space-time sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence feature, and further outputs the sequence feature and the attention weight corresponding to the sequence feature to the output layer; the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features, and outputs the prediction features to the fully connected layer, and the fully connected layer performs fault detection on the autonomous unmanned system based on the prediction features.
[0064] In this specification, the fault diagnosis model constructed based on the pulse neural network of the self-attention mechanism may refer to a fault diagnosis model that introduces the self-attention mechanism into the pulse neural network model (SNNs).
[0065] In order to better understand the fault diagnosis model constructed by the pulse neural network based on the self-attention mechanism provided by the present invention, the pulse neural network is first introduced below.
[0066] Please refer to Figure 2 The schematic diagram of the pulse neural network shown in Figure 1 is as follows. Figure 2 As shown in the upper part of the figure, the spiking neural network model can simulate the working mechanism of neurons in the human brain. A typical neuron can be composed of dendrites, axons, and cell bodies. Dendrites are used to receive pulse signals from pre-neurons and transmit the received pulse signals to the cell body. When the membrane potential of a neuron exceeds the discharge threshold, the neuron generates an output pulse signal, which is transmitted to the post-neurons through the axon. The connection between two neurons can be called a synapse. The goal of the spiking neural network model is to use neuroscience and machine learning techniques to achieve efficient intelligent computing with biological credibility mechanisms.
[0067] This specification is based on the idea of the pulse neural network model, introduces the concept of convolution in the pulse neural network model, and uses the convolution's ability to process high-dimensional data to simulate the neurons of the human brain. It is worth mentioning that the first type of convolution layer and the second type of convolution layer are named the first type of convolution layer and the second type of convolution layer for the convenience of distinction due to their different locations, and their convolution layers are the same.
[0068] In an exemplary embodiment, the first type of convolutional layer and the second type of convolutional layer both include a neuron model for simulating neurons of a human brain, and the neuron model includes:
[0069]
[0070] Among them, τ m represents the membrane time constant, V(t) represents the membrane potential of the neuron, V rest represents the resting potential of the neuron, R m represents the leakage resistance of the neuron, and I(t) represents the input current of the neuron.
[0071] In addition, the neuron also has a pulse firing mechanism, which specifically means that when the membrane potential exceeds the discharge threshold, the neuron will also discharge and the membrane potential will also be reset:
[0072]
[0073] Where θ represents the discharge threshold, V r Indicates the reset potential.
[0074] like Figure 2 The lower part of the figure shows a schematic diagram of the neuron model. The rear neuron receives the pulse signals input by the n front neurons, as well as the weights w1, w2, …, wn of each of these n pulse signals; accordingly, the neural dynamics mechanism of the neuron model in the convolutional layer can be expressed as:
[0075]
[0076] O(t f )=1,V(t)=V r , if V(t)≥θ
[0077] O(t)=0,if V(t)≤θ
[0078] Among them, S i (t) is the action potential output by the i-th front neuron, and O(t) represents the action potential output by the rear neuron.
[0079] Unlike artificial neural network models, pulse neural network models use pulse signals as the medium for information transmission, and each neuron can exhibit different dynamic properties. Since the pulse neural network model has the aforementioned pulse emission mechanism and the characteristics of dynamic transformation of membrane potential based on discharge threshold, the pulse neural network model is endowed with good spatiotemporal performance while maintaining biological authenticity.
[0080] The following further introduces the pulse neural network model based on the self-attention mechanism obtained by further improving the pulse neural network model provided in this specification.
[0081] like Figure 3 The schematic diagram of the pulse neural network model based on the self-attention mechanism shown in the figure can be composed of a convolutional coding layer, a feature extraction layer and an output layer, wherein the feature extraction layer includes a first type of convolutional layer for extracting features and a self-attention layer connected to the first type of convolutional layer, and the output layer includes a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer.
[0082] Among them, a first-class convolutional layer and a self-attention layer connected thereto can be regarded as a group. In order to obtain more accurate prediction features, the feature extraction layer can stack at least two groups of first-class convolutional layers for extracting features and self-attention layers connected to the first-class convolutional layers in a serial order. More accurate attention weights can be calculated through the features of the at least two groups, thereby improving the accuracy of the prediction features extracted by the second-class convolutional layer in the output layer.
[0083] The present invention introduces the concept of convolution into the spiking neural network model to realize synaptic weight sharing and hierarchical spiking representation learning. Assume that at time t, the input feature map is Its length is l, then U t (v) represents the output of the front neuron at the v-th position on the feature map.
[0084] The size of the neuron synaptic weight is set to 2k+1 to construct the convolution kernel W(i), -k≤i≤k. Assume that the feature map U t The convolution kernel W(i) is convolved with the position v as the center, so the output O of the convolution layer at position v can be obtained. t (v):
[0085]
[0086] Taking the above neuron model as an example, it can be expressed as:
[0087]
[0088] O t (l, h) = 1, V(t) = Vr , if V(t)≥θ
[0089] O t (l, h) = 0, if V(t) ≤ θ
[0090] When the rear neuron at position v receives a pulse signal output from the front neuron at position 2k+1, if the membrane potential of the rear neuron reaches the preset discharge threshold, it will output a pulse signal; otherwise, it will remain silent.
[0091] Based on the above-mentioned neuron model, this specification introduces a learnable recursive feature learning mechanism; by setting a cyclic feature learning algorithm inside the first type of convolutional layer and the second type of convolutional layer, the first type of convolutional layer and the second type of convolutional layer become cyclic convolutional layers.
[0092] Such a recurrent convolutional layer can not only achieve spatial feature learning similar to convolutional neural networks through convolutional topology, but also has the following advantages: Figure 4 The learnable recursive feature learning mechanism shown in Figure 1 allows the neural dynamics mechanism of the recurrent convolutional layer to be discretized into the following recursive formula:
[0093] The first formula:
[0094] The second formula:
[0095] The third formula:
[0096] The fourth formula: V[t]=H[t](1-O[t])+V rest O[t].
[0097] Among them, the second formula describes the changing state of neuron potential, represents the potential change of the neuron simulated by the recurrent convolution layer, S[t] represents the input feature of the recurrent convolution layer at time t, which is used to simulate the input potential of the neuron at time t; V[t-1] represents the module potential of the neuron at time t-1, and H[t] represents the membrane potential of the neuron at time t. Θ(·) is the Heaviside step function. 9 represents the synaptic threshold. O[t] represents the action potential output at time t.
[0098] The third formula describes the firing process of a neuron, and the fourth formula shows the resetting process of a neuron. rest Typically set to 0.
[0099] The second formula above can be transformed into a cyclic feature learning algorithm:
[0100]
[0101] 1 / τ m represents the update gate used to retain important features in S[t], 1-1 / τ m represents a forget gate for discarding unimportant features in V[t-1]; the important features include features with high relevance to fault detection, and the unimportant features include features with low relevance to fault detection.
[0102] As mentioned above, when the membrane potential of a neuron exceeds the discharge threshold, it will discharge and reset the membrane potential; combined with the cyclic feature learning algorithm, it can be seen that the discharge can include passing the important features in S[t] retained by the update gate as new input potentials to the neuron that resets the membrane potential for iterative processing.
[0103] After introducing the first type of convolutional layer and the second type of convolutional layer (also known as the recurrent convolutional layer), the self-attention layer (also known as the self-attention module) that cooperates with it will be introduced below.
[0104] like Figure 5 Schematic diagram of the self-attention layer shown in Figure 2. If the input of the self-attention layer is s l [i] represents the features of the first layer at the i-th time step, C and L are the number of channels and feature length respectively;
[0105] For each type of feature in the feature sequence, it is encoded as the first position code Q according to the time step i and the second position code K i , and the third position code V i ;
[0106] Q i =W Q *S l [i]+b Q , 1≤i≤N
[0107] K i =W K *S i [i]+b K , 1≤i≤N
[0108] V i =W V *S l [i]+b V , 1≤i≤N
[0109] Then, matrix multiplication is performed on the first position code and the second position code of each time step to obtain an attention score representing the similarity between the first position code and the second position code.
[0110] For example, the obtained attention scores can be normalized by a classifier (such as Softmax) to obtain the attention weight corresponding to the sequence feature.
[0111]
[0112] We can further add the attention weight and the third position code V i Matrix multiplication is performed to obtain a feature sequence optimized by attention weights, thereby helping the second type of convolutional layer in the output layer to further extract predictive features related to fault detection from this optimized feature sequence.
[0113] After introducing the architecture of the above fault diagnosis model, the following describes the training of the fault diagnosis model:
[0114] Collecting training samples for model training; the training samples include multivariate sensor signals annotated with fault detection results;
[0115] The training samples are used to perform supervised training on the fault diagnosis model, and a gradient substitution algorithm can be used to optimize the model parameters in a gradient back propagation manner.
[0116] Since the transition function Θ(x) used in traditional back-propagation is non-differentiable, it is difficult to perform gradient propagation and model parameter optimization in the spiking neural network model.
[0117] To this end, this patent uses a gate function σ(x) that is similar to the transition function Θ(x) but differentiable for back propagation. Specifically, in the forward propagation process, Θ(x) is used, and neurons transmit information in the form of pulse spatiotemporal sequences and membrane dynamics; in the back propagation process, the gradient of the gradient substitution algorithm is used to replace the gradient Θ′(x)≈σ′(x) of the transition function Θ(x), thereby realizing gradient propagation and parameter optimization adjustment.
[0118] In an exemplary embodiment, the gradient substitution algorithm may include a Gaussian error function;
[0119] The optimization of model parameters by using the gradient substitution algorithm in a gradient back propagation manner includes:
[0120] The gradient generated by the Gaussian error function replaces the gradient required by the transition function in the gradient back propagation, and the replaced gradient is used to optimize the model parameters in a back propagation manner.
[0121] Exemplarily, the above replacement algorithm can be expressed as:
[0122]
[0123] Through the above embodiments, the gradient substitution algorithm can be used to solve the problem that it is difficult to perform gradient propagation and model parameter optimization in a pulse neural network model due to the non-differentiability of the transition function.
[0124] Assume that the number of failure types of autonomous unmanned systems is Then the number of neurons in the output layer is also Output layer that accumulates T time steps The output pulses of neurons are counted to obtain the total number of impulse releases for each neuron. Then, the total number of impulse releases is divided by the time step to obtain the impulse release frequency:
[0125]
[0126] Ideally, when the category of the input data is i, the i-th neuron in the output layer of the spiking neural network model based on the self-attention mechanism should have the maximum activation degree, while other neurons remain inactive.
[0127] Therefore, the loss function can use the cross entropy function to measure the distance between the fault type output by the fully connected layer in the output layer and the actual labeled fault type, which can be expressed as:
[0128]
[0129] Where N represents the number of training data. A coded value indicating the type of fault, Represents the predicted features of the nth training sample. This encoding value can be obtained by ne-hot encoding. It is worth mentioning that since the pulse neural network model is a memory model, when new data is input, the state of the model needs to be reset.
[0130] After introducing the training of the fault diagnosis model, let’s look back at Figure 1 The fault diagnosis method shown in the embodiment is combined with Figure 3 The lower part of the figure shows the model architecture of the spiking neural network model based on the self-attention mechanism (SRA-SNN). Figure 3 As shown, the SRA-SNN model can be composed of a convolutional coding layer, a feature extraction layer and an output layer, wherein the feature extraction layer includes a first type of convolutional layer for feature extraction and a self-attention layer connected thereto, and the output layer includes a second type of convolutional layer and a fully connected layer connected thereto.
[0131] When the autonomous unmanned system acquires a multivariate sensor signal, the multivariate sensor signal can be input into the SRA-SNN model. First, the convolutional coding layer encodes the input multivariate sensor signal into a pulse spatiotemporal sequence. Then, the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse spatiotemporal sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence features. Then, the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features. Finally, the fully connected layer is used as a classifier to accurately map the prediction features to the corresponding fault types.
[0132] Below Figure 6 Take the schematic diagram of the autonomous underwater vehicle shown in Figure 2 for example. Figure 6 As shown, the autonomous underwater vehicle is a small autonomous unmanned system that is sensitive to energy consumption. The main body of the autonomous underwater vehicle is a control cabin and four engines deployed around the control cabin (each engine can be composed of a motor and a propeller). In addition, the autonomous underwater vehicle is also equipped with multiple sensors (such as the depth sensor shown in the figure, and satellite positioning equipment, inertial measurement unit IMU, etc. are not shown in the figure).
[0133] Multi-sensor signals can be collected through multiple sensors. Please refer to Table 1 for the exemplary 16 different sensor data provided below, such as motor control signals, battery voltage, depth, angular velocity, acceleration, and posture information.
[0134] Table 1
[0135]
[0136] like Figure 6 As shown, there are five types of faults for the autonomous underwater vehicle, namely normal state (F1), load abnormality (F2), depth sensor failure (F3), serious propeller failure (F4) and minor propeller failure (F5).
[0137] When collecting training samples, the AUV was placed in each fault type and run multiple times to collect multi-sensor signals under each fault type. Each run lasted about 10 to 20 seconds to obtain enough diverse sensor data. Finally, for each fault type, about 250 training samples were obtained, and a total of 1220 training samples were obtained.
[0138] In order to test the performance of the model, the 1220 training samples were randomly divided into a training set and a test set in a ratio of 4:1, that is, the training set contained 975 training samples and the test set contained 245 training samples. Subsequently, each training sample was Z-score standardized in the time direction to remove the influence of the dimension. At this point, the training sample collection and preprocessing process was completed.
[0139] Next, the pulse neural network model SRA-SNN based on the self-attention mechanism is compared with the artificial neural network model ANN to prove that the SRA-SNN provided by the present invention has significant fault detection performance. The evaluation indicators are: accuracy (ACC), F1-Score and AUC.
[0140] In the comparative experiment, ten ANN-based fault diagnosis models (hereinafter referred to as experimental models) were implemented on the AUV dataset, namely Multi-ANN, Wen-CNN, Sheng-CNN, AD-CNN, Multi-GRU, Multi-LSTM, Multi-RNN, MFCNN, Gu-1DCNN and Conv-Trans.
[0141] Among them, Multi-ANN is composed of multiple fully connected neural architectures. Wen-CNN, Sheng-CNN and AD-CNN use 2D convolutional layers to extract valuable features and use FC-Layers to implement fault diagnosis. Multi-GRU, Mult i-LSTM and Multi-RNN use different recurrent neural units to implement temporal feature learning and use FC-Layers to implement fault diagnosis. MFCNN is a convolutional network model for rolling bearing multi-fault classification, using average aggregation instead of maximum aggregation. Gu-1DCNN is a convolutional model for multi-sensor information fusion, which achieves excellent motor fault diagnosis performance. Finally, according to the idea of ViT architecture, CNN and Transformer are combined to construct Conv-Trans.
[0142] The comparative test is divided into three parts: (1) no noise is added; (2) Gaussian white noise is added; (3) impulse noise is added.
[0143] (1) No noise is added. The experimental results of the comparison of various models without noise are shown in Table 2:
[0144] Table 2
[0145]
[0146] (2) Add Gaussian white noise respectively:
[0147] In order to further verify the noise robustness of SRA-SNN, the above experimental model was experimented in a noisy environment. To simulate the noise interference in the AUV data set, Gaussian white noise with different signal-to-noise ratios (SNR) was added. Specifically, noise with SNRs of 8dB and 4dB was added to the data set. Then, the above experimental model was trained and tested on the noisy data set to demonstrate the noise resistance of the experimental model. The experimental results of each model with the addition of two types of Gaussian noise (8db and 4db) are shown in Table 3 below. The bold words in the table indicate the best diagnostic performance or the smallest standard deviation.
[0148] The experimental results show that the performance of all experimental models is greatly reduced under different Gaussian noise conditions. For example, when the SNR is 8dB, the F1-Score of SRA-SNN drops from 97.32% to 89.72%; the F1-Score of Wen-CNN drops from 95.27% to 84.73%; the F1-Score of Multi-LSTM drops from 91.85% to 79.04%; the F1-Score of Gu-1DCNN drops from 95.62% to 83.80%. However, the SRA-SNN provided by the present invention still achieves the best fault diagnosis performance and is at least 3% higher than Wen-CNN. This shows that SRA-SNN has good anti-noise performance, mainly due to its pulse-based information transmission mechanism. The pulse-based information transmission mechanism transmits information only through the pulse spatiotemporal sequence, thereby suppressing the influence of noise signals on the model. However, ANNs use floating-point numbers to transmit information, and noise signals will affect the entire model along with the information transmission of the model, so its anti-noise performance is poor. When the SNR is 4 dB, the performance of these models deteriorates more seriously. However, the SRA-SNN proposed in the present invention still has better diagnostic performance than Wen-CNN, MFCNN and Conv-Trans, which once again shows that SRA-SNN has good noise resistance in a strong noise environment.
[0149] Table 3
[0150]
[0151] (3) Add impulse noise separately:
[0152] This section further evaluates the performance of SRA-SNN and ten experimental models in the presence of impulse noise. Impulse noise is characterized by intermittency, consisting of irregular pulses or noise spikes, short duration and large amplitude. Considering the intermittent nature of impulse noise, its impact on the signal is quantified by calculating the percentage of signal values affected by impulses and the total number of signal values. For example, assuming a 10% impulse noise level means that 10% of the signal values behave as impulses. The experimental results of each model under two different impulse noises (10% and 20%) are shown in Table 4 below.
[0153] The experimental results reveal the significant impact of impulse noise on SRA-SNN and other ANN-based experimental models. Specifically, when 10% impulse noise is present in the signal, the diagnostic accuracy of SRA-SNN drops from 97.55% to 93.57%. Similarly, the diagnostic accuracy of Wen-CNN drops from 95.76% to 88.98%, the diagnostic accuracy of Multi-GRU drops from 93.61% to 86.73%, and the diagnostic accuracy of Conv-Trans drops from 96.73% to 91.33%. Obviously, SRA-SNN shows a relatively small drop and achieves the best fault diagnosis performance. This observation emphasizes that SNN models show high noise robustness in the presence of impulse noise. Even when the noise level increases to 20%, SRA-SNN achieves 88.31% accuracy and 87.45% F1-Score, surpassing the performance of other experimental models. The experimental results provide strong evidence that the proposed SRA-SNN achieves noise-resistant performance in various noise types and levels.
[0154] Table 4
[0155]
[0156] Since SRA-SNN can achieve 97.55% accuracy on a real multi-sensor AUV dataset, it shows that SRA-SNN can be used as an efficient and low-power fault diagnosis solution.
[0157] Corresponding to the above-mentioned fault diagnosis method embodiment, this specification also provides an embodiment of a fault diagnosis device. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of the device in which it is located reading the corresponding computer program in the non-volatile memory into the memory and running it. From the hardware level, such as Figure 7 The figure is a hardware structure diagram of the device where the fault diagnosis device in this manual is located, except Figure 7In addition to the processor, network interface, memory and non-volatile memory shown, the device in which the apparatus in the embodiment is located may also include other hardware according to the actual communication function, which will not be described in detail.
[0158] See also Figure 8 , is a module diagram of a fault diagnosis device provided in an embodiment of this specification, and the device corresponds to Figure 1 The embodiment shown is applied to an autonomous unmanned system, wherein the autonomous unmanned system is equipped with multiple sensors and a fault diagnosis model constructed by a pulse neural network based on a self-attention mechanism; the pulse neural network includes a convolutional coding layer, a feature extraction layer and an output layer, the feature extraction layer includes a first type of convolutional layer for extracting features, and a self-attention layer connected to the first type of convolutional layer, the output layer includes a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer; the device includes:
[0159] An acquisition unit 910 acquires a multi-sensor signal representing the operating state of the autonomous unmanned system acquired by the multiple sensors; wherein the signal carries an acquisition time;
[0160] The diagnostic unit 920 inputs the multivariate sensor signal into the convolutional coding layer in the fault diagnosis model, and the convolutional coding layer encodes the multivariate sensor signal into a pulse space-time sequence according to the acquisition time of the sensor signal, and outputs the pulse space-time sequence to the feature extraction layer; the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse space-time sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence feature, and further outputs the sequence feature and the attention weight corresponding to the sequence feature to the output layer; the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features, and outputs the prediction features to the fully connected layer, and the fully connected layer performs fault detection on the autonomous unmanned system based on the prediction features.
[0161] Optionally, the feature extraction layer includes at least two groups of first-type convolutional layers for extracting features stacked in a serial order, and a self-attention layer connected to the first-type convolutional layers.
[0162] Optionally, both the first type of convolutional layer and the second type of convolutional layer include a neuron model for simulating neurons of a human brain, and the neuron model includes:
[0163]
[0164] Among them, τ mrepresents the membrane time constant, V(t) represents the membrane potential of the neuron, V rest represents the resting potential of the neuron, R m represents the leakage resistance of the neuron, and I(t) represents the input current of the neuron.
[0165] Optionally, the first type of convolutional layer and the second type of convolutional layer further have an internal cyclic feature learning algorithm, so that the first type of convolutional layer and the second type of convolutional layer become cyclic convolutional layers; wherein the cyclic feature learning algorithm includes:
[0166]
[0167] in, represents the potential change of the neuron simulated by the recurrent convolution layer, S[t] represents the input feature of the recurrent convolution layer at time t, which is used to simulate the input potential of the neuron at time t; V[t-1] represents the module potential of the neuron at time t-1, and H[t] represents the membrane potential of the neuron at time t; 1 / τ m represents the update gate used to retain important features in S[t], 1-1 / τ m represents a forget gate for discarding unimportant features in V[t-1]; the important features include features with high relevance to fault detection, and the unimportant features include features with low relevance to fault detection.
[0168] Optionally, when the membrane potential of the neuron exceeds a threshold, the neuron will discharge and reset the membrane potential;
[0169] The discharge includes passing the important features in S[t] retained by the update gate as new input potentials to the neuron that resets the membrane potential for iterative processing.
[0170] Optionally, the step of calculating the attention weight corresponding to the sequence feature by the self-attention layer connected to the convolutional layer in the diagnosis unit 920 includes:
[0171] A self-attention layer connected to the convolutional layer encodes each type of feature in the sequence features into a first position code and a second position code according to the time step; matrix multiplication is performed on the first position code and the second position code of each time step to obtain an attention score representing the similarity between the first position code and the second position code; the attention score is normalized to obtain an attention weight corresponding to the sequence feature.
[0172] Optionally, the fault diagnosis model is trained using a gradient substitution algorithm to optimize model parameters in a gradient back-propagation manner.
[0173] Optionally, the gradient substitution algorithm includes a Gaussian error function;
[0174] The optimization of model parameters by using the gradient substitution algorithm in a gradient back propagation manner includes:
[0175] The gradient generated by the Gaussian error function replaces the gradient required by the transition function in the gradient back propagation, and the replaced gradient is used to optimize the model parameters in a back propagation manner.
[0176] Optionally, the loss function of the fault diagnosis model includes:
[0177]
[0178] Where N represents the number of training samples. A coded value indicating the type of fault, Represents the predicted features of the nth training sample.
[0179] Optionally, the autonomous unmanned system includes an autonomous underwater vehicle.
[0180] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which may be in the form of a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email transceiver, a game console, a tablet computer, a wearable device or a combination of any of these devices.
[0181] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.
[0182] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial description of the method embodiments. The device embodiments described above are only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this specification. Ordinary technicians in this field can understand and implement it without paying creative work.
[0183] Corresponding to the aforementioned fault diagnosis method embodiment, this specification also provides an autonomous underwater vehicle, on which an autonomous unmanned system is deployed, and the autonomous unmanned system implements the aforementioned fault diagnosis method.
[0184] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.
[0185] Those skilled in the art will readily appreciate other embodiments of the specification after considering the specification and practicing the invention disclosed herein. The specification is intended to cover any variations, uses, or adaptations of the specification that follow the general principles of the specification and include common knowledge or customary techniques in the art that are not disclosed in the specification. The specification and examples are to be considered exemplary only, and the true scope and spirit of the specification are indicated by the following claims.
[0186] It should be understood that the present description is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present description is limited only by the appended claims.
Claims
1. A fault diagnosis method, characterized in that: The invention is applied to an autonomous unmanned system, wherein the autonomous unmanned system is equipped with multiple sensors and a fault diagnosis model constructed by a pulse neural network based on a self-attention mechanism; the pulse neural network comprises a convolutional coding layer, a feature extraction layer and an output layer, the feature extraction layer comprises a first type of convolutional layer for extracting features, and a self-attention layer connected to the first type of convolutional layer, and the output layer comprises a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer; the method comprises: Acquire multi-sensor signals representing the operating state of the autonomous unmanned system collected by the multiple sensors; wherein the signals carry the collection time; The multivariate sensor signal is input into the convolutional coding layer in the fault diagnosis model, and the convolutional coding layer encodes the multivariate sensor signal into a pulse space-time sequence according to the acquisition time of the sensor signal, and outputs the pulse space-time sequence to the feature extraction layer; the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse space-time sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence feature, and further outputs the sequence feature and the attention weight corresponding to the sequence feature to the output layer; the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features, and outputs the prediction features to the fully connected layer, and the fully connected layer performs fault detection on the autonomous unmanned system based on the prediction features.
2. The method according to claim 1, characterized in that The feature extraction layer includes at least two groups of first-type convolutional layers for extracting features stacked in a serial order, and a self-attention layer connected to the first-type convolutional layer.
3. The method according to claim 1, characterized in that The first type of convolutional layer and the second type of convolutional layer both include a neuron model for simulating neurons of a human brain, and the neuron model includes: Among them, τ m represents the membrane time constant, V(t) represents the membrane potential of the neuron, V rest represents the resting potential of the neuron, R m represents the leakage resistance of the neuron, and I(t) represents the input current of the neuron.
4. The method according to claim 3, characterized in that The first type of convolutional layer and the second type of convolutional layer also have an internal cyclic feature learning algorithm, so that the first type of convolutional layer and the second type of convolutional layer become cyclic convolutional layers; wherein the cyclic feature learning algorithm includes: in, represents the potential change of the neuron simulated by the recurrent convolution layer, S[t] represents the input feature of the recurrent convolution layer at time t, which is used to simulate the input potential of the neuron at time t; V[t-1] represents the module potential of the neuron at time t-1, and H[t] represents the membrane potential of the neuron at time t; 1 / τ m represents the update gate used to retain important features in S[t], 1-1 / τ m represents a forget gate for discarding unimportant features in V[t-1]; the important features include features with high relevance to fault detection, and the unimportant features include features with low relevance to fault detection.
5. The method according to claim 4, characterized in that When the membrane potential of the neuron exceeds the discharge threshold, the neuron will discharge and reset the membrane potential; The discharge includes passing the important features in S[t] retained by the update gate as new input potentials to the neuron that resets the membrane potential for iterative processing.
6. The method according to claim 1, characterized in that The calculating, by the self-attention layer connected to the convolutional layer, the attention weight corresponding to the sequence feature comprises: A self-attention layer connected to the convolutional layer encodes each type of feature in the sequence features into a first position code and a second position code according to a time step; Performing matrix multiplication on the first position code and the second position code of each time step to obtain an attention score representing the similarity between the first position code and the second position code; The attention scores are normalized to obtain attention weights corresponding to the sequence features.
7. The method according to claim 1, characterized in that The fault diagnosis model is trained by using a gradient substitution algorithm to optimize model parameters in a gradient back-propagation manner.
8. The method according to claim 7, characterized in that The gradient substitution algorithm includes a Gaussian error function; The optimization of model parameters by using the gradient substitution algorithm in a gradient back propagation manner includes: The gradient generated by the Gaussian error function replaces the gradient required by the transition function in the gradient back propagation, and the replaced gradient is used to optimize the model parameters in a back propagation manner.
9. The method according to claim 1, characterized in that: The loss function of the fault diagnosis model includes: Where N represents the number of training samples. A coded value indicating the type of fault, Represents the predicted features of the nth training sample.
10. A fault diagnosis device, characterized in that: Applied to an autonomous unmanned system, the autonomous unmanned system is equipped with multiple sensors and a fault diagnosis model constructed by a pulse neural network based on a self-attention mechanism; the pulse neural network includes a convolutional coding layer, a feature extraction layer and an output layer, the feature extraction layer includes a first type of convolutional layer for extracting features, and a self-attention layer connected to the first type of convolutional layer, the output layer includes a second type of convolutional layer and a fully connected layer connected to the second type of convolutional layer; the device includes: An acquisition unit, which acquires multi-sensor signals representing the operating state of the autonomous unmanned system acquired by the multiple sensors; wherein the signals carry an acquisition time; The diagnosis unit inputs the multi-sensor signal into the convolutional coding layer in the fault diagnosis model, and the convolutional coding layer encodes the multi-sensor signal into a pulse space-time sequence according to the acquisition time of the sensor signal, and outputs the pulse space-time sequence to the feature extraction layer; the first type of convolutional layer in the feature extraction layer extracts sequence features from the pulse space-time sequence, and the self-attention layer connected to the first type of convolutional layer calculates the attention weight corresponding to the sequence feature, and further outputs the sequence feature and the attention weight corresponding to the sequence feature to the output layer; the second type of convolutional layer in the output layer further extracts prediction features related to fault detection from the sequence features based on the attention weight corresponding to the sequence features, and outputs the prediction features to the fully connected layer, and the fully connected layer performs fault detection on the autonomous unmanned system based on the prediction features.
11. An autonomous underwater vehicle, characterized in that: An autonomous unmanned system is deployed on the autonomous underwater vehicle, and the autonomous unmanned system implements the method described in any one of claims 1-10.
Citation Information
Patent Citations
Musical emotion recognition method based on bidirectional convolution-recurrent sparse network
CN110223712A
Image character removing method, system and device based on gate circulation unit
CN112419174A
Unmanned equipment fault diagnosis method and device, equipment, storage medium and product
CN116300796A
Signal processing method for neuron in spiking neural network and method for training said network
WO2023284142A1
Cited By
Mechanical fault diagnosis method and system based on spiking neural network
CN119443153A
Fault diagnosis method, device, medium and computer program product
CN120121275A
Multi-modal sensing electromechanical actuator fault prediction system and control method thereof
CN120143610A
Numerical control program control fault diagnosis method and system
CN120848463A
