An electro-hydraulic servo valve fault diagnosis method suitable for edge device deployment

Abstract

The application discloses a kind of electro-hydraulic servo valve fault diagnosis methods suitable for edge device deployment, comprising: the vibration signal of electro-hydraulic servo valve is collected, and the vibration signal collected is denoised and normalized, to obtain the vibration signal after pre-processing;Using the improved particle swarm algorithm to optimize the modal number and penalty factor of variational mode decomposition;The vibration signal after pre-processing is decomposed using the optimized modal number and penalty factor, to obtain multiple intrinsic mode function components, and screen out effective intrinsic mode function components;Using parallel channel attention subnetwork and timing attention subnetwork, the effective intrinsic mode function components are respectively weighted in channel dimension and time dimension, and the outputs of channel attention subnetwork and timing attention subnetwork are fused through learnable dynamic proportion coefficient, and the fusion features are input into one-dimensional deep separable convolutional neural network for fault pattern recognition, to obtain fault recognition result.

Description

Technical Field

[0001] This invention belongs to the field of electromechanical hydraulic system fault diagnosis technology, specifically relating to a fault diagnosis method for electro-hydraulic servo valves suitable for deployment in edge devices. Background Technology

[0002] As a core actuator in industrial automation control systems, servo valves are responsible for the precise control of fluid pressure and flow. Their operational stability directly determines the operating accuracy and production efficiency of the entire equipment. In high-frequency operating conditions such as injection molding machines and rolling mills, servo valves must withstand long-term high-frequency reciprocating load impacts, while also facing complex environmental influences such as high-pressure fluid corrosion, mechanical wear, and electromagnetic interference. This makes them prone to various failure modes, including valve core jamming, coil failure, seal aging, and oil circuit blockage. Statistics show that equipment downtime caused by industrial servo valve failures accounts for more than 35% of total downtime, and the economic loss from a single failure can reach hundreds of thousands of yuan. Therefore, achieving early warning and accurate diagnosis of servo valve failures has significant engineering value.

[0003] Traditional servo valve fault diagnosis methods mainly rely on physical model-based analysis methods, spectrum analysis-based signal processing methods, and traditional machine learning methods such as support vector machines and shallow neural networks. These methods have the following significant limitations:

[0004] (1) Insufficient ability to identify weak faults: Traditional signal analysis methods (such as empirical mode decomposition) have mode aliasing problems when processing non-stationary and nonlinear signals, making it difficult to effectively extract weak early fault features from strong background noise.

[0005] (2) Feature extraction depends on expert experience: Traditional machine learning methods rely heavily on manual feature engineering (such as time domain statistics, frequency domain features, etc.), and their diagnostic performance is deeply tied to the domain knowledge of experts, with weak generalization and adaptive capabilities.

[0006] (3) Model complexity and deployment difficulty: Although deep learning methods, represented by deep convolutional neural networks (CNN), can automatically extract deep features, their models usually have a large number of parameters and are computationally complex, making it difficult to directly deploy them to resource-constrained edge computing or embedded devices. This fails to meet the industrial field's demand for real-time diagnosis (millisecond-level response) and low-cost deployment.

[0007] (4) Insufficient utilization of multi-source information: Existing methods are mostly focused on the analysis of single signals (such as pressure or vibration), and fail to effectively integrate the complementary fault information contained in the multi-channel dynamic signals (pressure, flow, current, vibration, etc.) of servo valves, which limits the further improvement of diagnostic accuracy.

[0008] Therefore, there is an urgent need for a new intelligent fault diagnosis method for servo valves that can take into account high diagnostic accuracy, strong generalization ability, high computational efficiency and excellent embedded adaptability. Summary of the Invention

[0009] Purpose of the invention: To address the shortcomings of traditional servo valve fault diagnosis methods, this invention proposes a fault diagnosis method for electro-hydraulic servo valves suitable for edge device deployment. By integrating adaptive variational mode decomposition (VMD), dual-path attention mechanism, and one-dimensional deep separable convolutional neural network (1D-DSCNN), it can achieve real-time and accurate diagnosis of various types of faults such as valve core jamming, seal wear, and oil circuit blockage. It is suitable for online monitoring of industrial servo valves, and is especially suitable for high-frequency reciprocating working scenarios such as injection molding machines and rolling mills.

[0010] Technical solution: A fault diagnosis method for electro-hydraulic servo valves suitable for edge device deployment, comprising the following steps:

[0011] Step 1: Collect the vibration signal of the electro-hydraulic servo valve, and perform noise reduction and normalization on the collected vibration signal to obtain the pre-processed vibration signal;

[0012] Step 2: Optimize the number of modes and penalty factor in variational mode decomposition using an improved particle swarm optimization algorithm; decompose the preprocessed vibration signal using the optimized number of modes and penalty factor to obtain multiple intrinsic mode function components, and select the effective intrinsic mode function components.

[0013] Step 3: Using parallel channel attention subnetworks and temporal attention subnetworks, the effective intrinsic mode function components are weighted by channel and temporal dimensions respectively. The outputs of the channel attention subnetworks and temporal attention subnetworks are then fused using learnable dynamic scaling coefficients to obtain fused features. The channel attention subnetwork generates channel weights based on one-dimensional convolution, and the temporal attention subnetwork generates temporal weights based on a bidirectional long short-term memory network.

[0014] Step 4: Input the fused features into a one-dimensional deep separable convolutional neural network for fault mode recognition to obtain the fault recognition result; the one-dimensional deep separable convolutional neural network is obtained by quantizing the trained floating-point model into INT8 format.

[0015] Furthermore, the one-dimensional depthwise separable convolutional neural network includes an input layer, a first depthwise separable convolutional block, a second depthwise separable convolutional block, a global average pooling layer, a Dropout layer, and a Softmax output layer; each depthwise separable convolutional block sequentially includes a depthwise convolutional layer, a pointwise convolutional layer, a batch normalization layer, and an activation function layer.

[0016] Furthermore, the channel attention subnetwork includes a one-dimensional convolutional layer, a LeakyReLU activation function layer, and a Sigmoid activation function layer; the set of effective intrinsic mode function components obtained through screening is used as input;

[0017] Generate channel weight vectors:

[0018]

[0019] Among them, among them, The number of effective intrinsic mode function components, The channel weights are the values ​​corresponding to the i-th valid intrinsic mode function components.

[0020] The number of output channels of the one-dimensional convolutional layer is equal to the number of effective intrinsic mode function components. .

[0021] Furthermore, the temporal attention subnetwork includes a bidirectional long short-term memory network and a sigmoid activation function layer, wherein the hidden layer dimension of the bidirectional long short-term memory network is set to the number of effective intrinsic mode function components. 2 times;

[0022] Each effective intrinsic mode function (EMF) component in the set of effective EEM components is used as a time series input to a bidirectional long short-term memory (LSTM) network to extract dynamic features in the time dimension and output a time-series weight vector.

[0023]

[0024] in, The length of the time series. Let be the temporal weight for the j-th time step.

[0025] Furthermore, the fusion of the outputs of the channel attention subnet and the temporal attention subnet using learnable dynamic scaling coefficients specifically includes:

[0026] The characteristic matrix composed of effective intrinsic mode function components As input, based on the channel weight vector Generate channel-weighted features Based on time-series weight vectors Generate time-weighted features According to the channel weighting characteristics Time-weighted features The difference information is used to generate dynamic scaling coefficients through global average pooling and a multilayer perceptron. Then weight the channel features. Time-weighted features According to the dynamic proportional coefficient Perform weighted fusion and combine it with learnable coupling coefficients. Obtain fusion features .

[0027] Furthermore, the first depthwise separable convolutional block and the second depthwise separable convolutional block have the same structure, and both perform depthwise convolution, pointwise convolution, batch normalization and activation operations in sequence.

[0028] The first depth-separable convolutional block includes:

[0029] Perform channel-wise convolution as follows:

[0030]

[0031] in, For time step, Indicates the first Each input channel corresponds to a depthwise convolutional kernel;

[0032] Performing pointwise convolution on the output of channel-wise convolution is represented as:

[0033]

[0034] in, This represents the pointwise convolution kernel used for channel information fusion;

[0035] Batch normalization:

[0036]

[0037] Introducing MLP+Sigmoid to dynamically generate the scaling factor, expressed as:

[0038]

[0039] In the formula, This represents a dynamic scaling factor generated by the multilayer perceptron and the sigmoid function, used for adaptive scaling of the normalized features;

[0040] in,

[0041]

[0042] Among them, weight , , For batch statistics, These are learnable parameters;

[0043] ReLU activation: .

[0044] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0045] This invention meets the requirements for accurate feature extraction: it introduces adaptive VMD with optimized parameters, and accurately matches the intrinsic characteristics of the signal by improving the particle swarm algorithm, suppressing mode aliasing and providing pure intrinsic mode components (IMF) for subsequent analysis.

[0046] This invention meets the need for focusing on key information: it introduces a parallel dual-path attention mechanism, which adaptively weights the components after VMD decomposition from the channel dimension and the time dimension, respectively, so that the model can autonomously focus on the feature segments most relevant to the fault and enhance the sensitivity to minor faults.

[0047] This invention meets the requirements of lightweight and real-time models: it uses a one-dimensional depthwise separable convolutional network to replace the standard convolution, which greatly reduces the number of model parameters and computational load while ensuring feature extraction capabilities, laying the foundation for embedding models into resource-constrained microcontrollers (such as STM32).

[0048] By organically integrating the above technologies, this invention breaks through the technical bottlenecks of traditional methods in weak fault identification and model lightweighting, and realizes high-precision, high-efficiency, and low-cost online intelligent fault diagnosis of servo valves. Attached Figure Description

[0049] Figure 1 A flowchart of a fault diagnosis method for an electro-hydraulic servo valve suitable for edge device deployment;

[0050] Figure 2 Flowchart for VMD decomposition and modality selection;

[0051] Figure 3 Flowchart for weighted and fused dual-path Attention features;

[0052] Figure 4 This is a schematic diagram of the 1D-DSCNN network structure. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Example 1:

[0055] like Figure 1 As shown in the figure, this invention proposes a fault diagnosis method for electro-hydraulic servo valves suitable for edge device deployment. The method mainly includes the following steps, which are closely linked to ensure synergistic optimization of diagnostic accuracy, real-time performance, and embedded adaptability:

[0056] Step 1: Use a composite sensor array to synchronously monitor the operating status parameters of the electro-hydraulic servo valve. The operating status parameters include pressure, flow rate, temperature, current, displacement, and vibration signals. Among them, the vibration signal is the main analysis object for subsequent fault diagnosis, while the other operating status parameters are used for acquisition triggering, operating condition monitoring, or auxiliary status verification.

[0057] Data acquisition will begin when any of the following trigger conditions are met:

[0058] (1) The peak value of the pressure signal exceeds 80% of the rated value and the duration is greater than 1ms;

[0059] (2) Configure a hardware comparator to monitor the pressure signal in real time. When the instantaneous pressure value at port A or port B exceeds the threshold... An interrupt signal is generated at this time;

[0060] (3) The THD value of the synchronously detected current signal is calculated using the following formula:

[0061]

[0062] in, The effective value of the nth harmonic. This is the effective value of the fundamental frequency.

[0063] When the trigger condition is met, the data acquisition card synchronously acquires data from all channels via DMA, with a fixed sampling window length of 5 seconds.

[0064] Step 2: Perform noise reduction, normalization, and resampling on the collected vibration signal in sequence to obtain the preprocessed vibration signal.

[0065] In the preferred embodiment:

[0066] For vibration signals Wavelet thresholding is used for noise reduction:

[0067] The threshold selected is a general threshold:

[0068] ,

[0069] in, For wavelet threshold, This is an estimate of the noise standard deviation. This represents the number of sampling points.

[0070] Perform maximum-minimum normalization on the denoised vibration signal:

[0071]

[0072] Where T=0.5 seconds is the length of the acquisition window.

[0073] The normalized vibration signal was resampled to 25.6 kHz to obtain a length of

[0074]

[0075] vibration time series vector .

[0076] Step 3: Optimize the number of modes in variational mode decomposition using an improved particle swarm optimization algorithm. and penalty factor The optimized parameters are then used to perform variational mode decomposition on the preprocessed vibration signal to obtain multiple intrinsic mode function components. These components are then filtered based on signal correlation and envelope entropy to obtain a set of effective components. The specific operations include:

[0077] In the preferred embodiment:

[0078] The population size is set to 30, and the maximum number of iterations is [number missing].

[0079]

[0080] The learning factors are respectively

[0081] ,

[0082] The inertia weight adopts a linear decreasing strategy:

[0083]

[0084] in,

[0085] =0.9, =0.4

[0086] The parameter search space is set as follows:

[0087] ∈[3,8], ∈[1000,3000]

[0088] Fitness function: The objective is to minimize the mean envelope entropy of each mode after decomposition. The fitness function expression is:

[0089]

[0090] Wherein, the envelope entropy of the k-th IMF The calculation formula is:

[0091] ,

[0092] Among them, is the envelope signal of the IMF component;

[0093] The specific process of the improved particle swarm optimization (IPSO) includes:

[0094] First, input: vibration signal , parameter search range ∈[3, 8], ∈[1000, 3000]

[0095] Specific steps:

[0096] Initialization: Randomly generate 30 particles, and the position of each particle is ( , ), the velocity = 0; Set the adaptive inertia weight parameter = 0.9, = 0.4, early stop threshold |Δf| < 10 −4 . Randomly generate 30 particles, and the position of each particle is ( , ), the velocity v i = 0

[0097] Calculate from i = 1 to 30; Calculate the fitness function for all particles, record the historical optimal solution of each individual, and determine the current global optimal solution.

[0098] Perform VMD decomposition on x(t) to obtain IMF components.

[0099] Initial fitness calculation: For all particles, calculate the multi-objective fitness function:

[0100]

[0101] Among them , , T max is the maximum number of iterations. In the embodiment of the present invention, the dynamic balance weight λ(t) is introduced. In the early stage (t < Tmid), more attention is paid to the envelope entropy (λ(t) → 1), and the clear fault features with low entropy are preferentially screened. In the later stage (t > Tmid), more attention is paid to the correlation coefficient (λ(t) → 0), ensuring that the retained components are strongly correlated with the original signal, and the effective IMF components of "low entropy + high correlation" can be more accurately screened, improving the sensitivity of VMD decomposition to weak faults of servo valves. j is the envelope entropy of the jth IMF component, r jUpdate the individual optimal value using the Pearson correlation coefficient between the j-th IMF component and the original signal. and global optimal .

[0102] Calculate the adaptive inertia weights:

[0103]

[0104] Update speed: ,in ;

[0105] Update location: ;

[0106] Boundary checks:

[0107] like

[0108] like

[0109] like

[0110] like

[0111] Adaptive mutation operation: Gaussian perturbation is added to the particle position with a 15% probability. The globally optimal particle triggers mutation additionally, and the perturbation amplitude decreases with the number of iterations.

[0112]

[0113] Finish

[0114] return As the optimal parameter ( , )

[0115] Output: Optimal parameters ( , ).

[0116] like Figure 2 As shown, according to the optimized core parameters ( , The valve body vibration signal, one of the core operating parameters of the preprocessed servo valve, is decomposed using VMD to establish a constrained variational problem:

[0117]

[0118] Constraints:

[0119] The solution was obtained iteratively using the alternating direction multiplier method (ADMM), with the number of iterations set to 500 and a convergence tolerance of 1e-7. IMF components and the corresponding center frequency.

[0120] According to the calculation of each IMF component Correlation coefficient with the original signal x(t):

[0121]

[0122] Calculate the envelope entropy of each IMF component. (Formula as before)

[0123] The intrinsic mode function components that meet the screening criteria are sorted according to their correlation coefficients from largest to smallest and their envelope entropies from smallest to largest. The top [components] are selected. The effective component set consists of several components.

[0124]

[0125] in, The number of valid intrinsic mode function components to be reserved by default.

[0126] Step 4: Set the effective components The input is fed into parallel channel attention subnets and temporal attention subnets, where weighted features in the channel and time dimensions are extracted respectively. These features are then fused using dynamic scaling factors to obtain a fused feature matrix. .

[0127] In an embodiment of the present invention:

[0128] Let the characteristic matrix corresponding to the set of effective components be...

[0129]

[0130] in, The number of effective intrinsic mode function components, The time series length for each component.

[0131] like Figure 3 As shown, this channel attention subnetwork contains a one-dimensional convolutional layer (kernel size 3, output channel number equal to the number of effective intrinsic mode function components M), a LeakyReLU activation function, and a Sigmoid activation function; it extracts inter-channel correlation features through one-dimensional convolution, and generates channel weight vectors by mapping through activation functions:

[0132]

[0133] in, Let be the channel weight corresponding to the i-th valid intrinsic mode function component.

[0134] In this embodiment of the invention, the temporal attention subnetwork includes a one-layer bidirectional long short-term memory network (Bi-LSTM) and a sigmoid activation function. The hidden layer dimension of the Bi-LSTM is set to twice the number of effective modes.

[0135] Each effective intrinsic mode function (EMF) component in the effective EEM component set is used as a time-series input to the Bi-LSTM to capture temporal dynamic features and output a temporal weight vector:

[0136] ,

[0137] in, Let be the temporal weight for the j-th time step.

[0138] Based on the channel weight vector and the time-series weight vector, the channel-weighted features and the time-series weighted features are obtained respectively:

[0139]

[0140] in, This represents element-wise multiplication, the... Broadcast extension along the time dimension, the Broadcast extension along the channel dimension.

[0141] Calculate the feature differences between the two paths:

[0142]

[0143] right Perform global average pooling to obtain the global descriptor:

[0144]

[0145] Then, dynamic scaling factors are generated using a multilayer perceptron and a sigmoid function:

[0146]

[0147] The final fusion feature is represented as:

[0148]

[0149] in, This is the fused feature matrix. The Fc and Ft terms capture the coupling information between channel and temporal features. It is a learnable dynamic scaling factor. These are learnable coupling coefficients that can further enhance the representation of weak fault characteristics. This indicates a global average pooling operation.

[0150] Step 5: Input the fused feature matrix into a lightweight one-dimensional deep separable convolutional neural network to obtain the servo valve fault diagnosis classification result.

[0151]

[0152] in, The number of valid intrinsic mode function components to be reserved is preset. The time series length corresponding to each component; in this embodiment, .

[0153] like Figure 4 As shown, the total number of parameters of the lightweight one-dimensional deep separable convolutional neural network does not exceed 500KB. The network structure includes, in sequence: an input layer, a first deep separable convolutional block, a second deep separable convolutional block, a one-dimensional convolutional layer, a global average pooling layer, a Dropout layer, a fully connected output layer, and a Softmax classification layer.

[0154] in:

[0155] The first depthwise separable convolutional block sequentially includes depthwise convolution, pointwise convolution, batch normalization, and activation function operations, specifically represented as follows:

[0156] For the input feature matrix Perform channel-wise convolution:

[0157]

[0158] in, Indicates the first The first depthwise convolutional kernel corresponds to each input channel; in this embodiment, the kernel size is 64 and the stride is 4.

[0159] Perform pointwise convolution on the channel-wise convolution output:

[0160]

[0161] in, This represents the first pointwise convolution kernel, used to achieve feature fusion between channels; in this embodiment, the number of output channels is 32.

[0162] The pointwise convolution outputs are batch normalized, and a dynamic scaling factor is introduced. The normalization result is adaptively scaled, and is represented as follows:

[0163]

[0164] By introducing MLP+Sigmoid to dynamically generate the scaling factor, this design allows the scaling parameter to change dynamically with the input features. For example, in the scenario of a servo valve with a minor fault, it automatically enhances the scaling weights corresponding to the fault features.

[0165]

[0166] Adding locally weighted statistics before standardization reduces the impact of outliers on the mean and variance.

[0167]

[0168] Among them, weight Outliers are penalized using a Gaussian kernel function. , For batch statistics, For learnable parameters, .

[0169] The normalized features are then activated:

[0170] .

[0171] The structure of the second depthwise separable convolutional block is the same as that of the first depthwise separable convolutional block, and its output is represented as follows: In this embodiment, the depth convolution kernel size of the second depth convolution block is 32, the stride is 2, and the number of pointwise convolution output channels is 64.

[0172] A one-dimensional convolutional layer is set after the second depthwise separable convolutional block to further extract features, represented as:

[0173]

[0174] in, This represents a one-dimensional convolution kernel; in this embodiment, the kernel size is 16, the stride is 2, and the number of output channels is 128.

[0175] Convolution output Perform global average pooling to obtain the global feature vector:

[0176]

[0177] in, This indicates the length of the time dimension of the convolution output. .

[0178] For global feature vectors After performing Dropout processing, we get:

[0179]

[0180] in, This is a random deactivation mask; in this embodiment, the Dropout probability is... .

[0181] The features processed by Dropout are input into the fully connected output layer to obtain classification scores for each category:

[0182]

[0183] in, This represents the output layer weight matrix. This indicates the bias term.

[0184] Finally, the failure category probability is output through the Softmax classification layer:

[0185] in, This indicates the total number of fault categories; in this embodiment, .

[0186] In a preferred embodiment, the lightweight one-dimensional deep separable convolutional neural network is trained offline before deployment. The training process includes: dividing the sample dataset into training, validation, and test sets; using the cross-entropy loss function as the optimization objective; and using the Adam optimizer to iteratively update the network parameters until the model converges.

[0187] In this embodiment, the sample dataset includes 12,000 samples, divided into training, validation, and test sets in a 7:2:1 ratio; the loss function used is cross-entropy loss.

[0188]

[0189] in, For the batch sample size, For the true labels of the samples, Predict probabilities for the model.

[0190] The Adam optimizer parameters are set as follows in this embodiment:

[0191] .

[0192] The initial learning rate is 0.001, and it decays to 0.5 of the original value every 20 rounds. The weight decay coefficient is... .

[0193] Step 6: Perform INT8 quantization on the trained floating-point model to convert it into a low-bit integer model, thereby reducing model storage overhead, improving inference speed, and enabling real-time deployment of the fault diagnosis algorithm on embedded devices.

[0194] In a preferred embodiment of the present invention, the quantization method is post-training quantization (PTQ), which uses a calibration dataset to calibrate the dynamic range of the weights and activation values ​​of each layer of the network; in this embodiment, the calibration dataset is selected from a portion of the validation set.

[0195] For model weights The symmetric INT8 quantization method is used, and the quantization scaling factor is expressed as:

[0196]

[0197] The quantized weights are represented as follows:

[0198] in, This represents a truncation function used to restrict the result to the range of INT8 representation.

[0199] For activation value The INT8 quantization method based on the statistical dynamic range of the calibration dataset is adopted, and the minimum activation value is set to be INT8. The maximum value is Then the activation quantization scaling factor and zero point are expressed as follows:

[0200]

[0201] The quantized activation value is represented as:

[0202] During the inference phase, integer field convolution or matrix multiplication can be represented as:

[0203]

[0204] in, For the quantized input, and These are the input quantization scaling factor and zero point, respectively.

[0205] In this embodiment, the TensorFlow Lite Converter tool is used to convert the trained FP32 model into an INT8 format model and deploy it to an embedded microcontroller for inference.

[0206] In a preferred embodiment, the embedded microcontroller is an STM32F407VET6 with a main frequency of 168MHz, 192KB of RAM, and 512KB of Flash. The embedded terminal includes a data acquisition module, a preprocessing module, a model inference module, an alarm module, and a communication module.

[0207] The data acquisition module is used to collect vibration signals and, as needed, simultaneously collect auxiliary operating status parameters such as pressure, flow rate, temperature, current, and displacement. To ensure consistency with the input data length during the training phase, the embedded terminal processes the vibration signals using the same sampling window and preprocessing method as during the training phase. In this embodiment, the vibration signal sampling window length is 0.5 s, and after preprocessing, a data sample of length of [missing data] is generated. The input sequence.

[0208] The software architecture can implement task scheduling based on a real-time operating system. The tasks include at least: data acquisition tasks, preprocessing tasks, model inference tasks, alarm tasks, and data communication tasks. Among them, the model inference task calls a lightweight INT8 model to identify faults in the vibration signals within the current sampling window and outputs the probability of each type of fault.

[0209] In this embodiment, the delay for a single fault diagnosis is less than 5 ms.

[0210] Fault determination is performed based on the fault category probabilities output by the model. In a preferred embodiment, when the probability of a certain fault category is greater than a preset threshold, it is determined to be a corresponding fault state; when the probabilities of all categories do not reach the preset threshold, it is determined to be a suspicious state and a secondary data acquisition and diagnosis is triggered. The preset threshold is 0.8 in this embodiment.

[0211] When a fault condition is detected, the embedded terminal triggers an audible and visual alarm module and uploads the fault type, occurrence time, and corresponding characteristic parameters to the host computer monitoring platform. The host computer monitoring platform is used to display the servo valve's operating status and diagnostic results in real time, and supports historical fault data querying and trend analysis.

Claims

1. A fault diagnosis method for electro-hydraulic servo valves suitable for edge device deployment, characterized in that: Includes the following steps: Step 1: Collect the vibration signal of the electro-hydraulic servo valve, and perform noise reduction and normalization on the collected vibration signal to obtain the pre-processed vibration signal; Step 2: Optimize the number of modes and penalty factor in variational mode decomposition using an improved particle swarm optimization algorithm; decompose the preprocessed vibration signal using the optimized number of modes and penalty factor to obtain multiple intrinsic mode function components, and select the effective intrinsic mode function components. Step 3: Using parallel channel attention subnetworks and temporal attention subnetworks, the effective intrinsic mode function components are weighted by channel and temporal dimensions respectively. The outputs of the channel attention subnetworks and temporal attention subnetworks are then fused using learnable dynamic scaling coefficients to obtain fused features. The channel attention subnetwork generates channel weights based on one-dimensional convolution, and the temporal attention subnetwork generates temporal weights based on a bidirectional long short-term memory network. Step 4: Input the fused features into a one-dimensional deep separable convolutional neural network for fault mode recognition to obtain the fault recognition result; the one-dimensional deep separable convolutional neural network is obtained by quantizing the trained floating-point model into INT8 format.

2. The method for fault diagnosis of an electro-hydraulic servo valve suitable for edge device deployment according to claim 1, characterized in that: The one-dimensional depthwise separable convolutional neural network includes an input layer, a first depthwise separable convolutional block, a second depthwise separable convolutional block, a global average pooling layer, a Dropout layer, and a Softmax output layer; each depthwise separable convolutional block sequentially includes a depthwise convolutional layer, a pointwise convolutional layer, a batch normalization layer, and an activation function layer.

3. The method for fault diagnosis of an electro-hydraulic servo valve suitable for edge device deployment according to claim 1, characterized in that: The channel attention subnetwork includes a one-dimensional convolutional layer, a LeakyReLU activation function layer, and a Sigmoid activation function layer; the set of effective intrinsic mode function components obtained through screening is used as input; Generate channel weight vectors: ； Among them, among them, The number of effective intrinsic mode function components, The channel weights are the values ​​corresponding to the i-th valid intrinsic mode function components. The number of output channels of the one-dimensional convolutional layer is equal to the number of effective intrinsic mode function components. .

4. The method for fault diagnosis of an electro-hydraulic servo valve suitable for edge device deployment according to claim 1, characterized in that: The temporal attention subnetwork includes a bidirectional long short-term memory network and a sigmoid activation function layer. The hidden layer dimension of the bidirectional long short-term memory network is set to the number of effective intrinsic mode function components. 2 times; Each effective intrinsic mode function (EMF) component in the set of effective EEM components is used as a time series input to a bidirectional long short-term memory (LSTM) network to extract dynamic features in the time dimension and output a time-series weight vector. ； in, The length of the time series. Let be the temporal weight for the j-th time step.

5. The method for fault diagnosis of an electro-hydraulic servo valve suitable for edge device deployment according to claim 1, characterized in that: The process of fusing the outputs of the channel attention subnet and the temporal attention subnet using learnable dynamic scaling coefficients specifically includes: The characteristic matrix composed of effective intrinsic mode function components As input, based on the channel weight vector Generate channel weighted features Based on time-series weight vectors Generate time-weighted features According to the channel weighting characteristics Time-weighted features The difference information is used to generate dynamic scaling coefficients through global average pooling and a multilayer perceptron. Then weight the channel features. Time-weighted features According to the dynamic proportional coefficient Perform weighted fusion and combine it with learnable coupling coefficients. Obtain fusion features .

6. The method for fault diagnosis of an electro-hydraulic servo valve suitable for edge device deployment according to claim 2, characterized in that: The first and second depthwise separable convolutional blocks have the same structure, and both perform depthwise convolution, pointwise convolution, batch normalization and activation operations in sequence. The first depth-separable convolutional block includes: Perform channel-wise convolution as follows: ; in, For time step, Indicates the first Each input channel corresponds to a depthwise convolutional kernel; Performing pointwise convolution on the output of channel-wise convolution is represented as: ； in, This represents the pointwise convolution kernel used for channel information fusion; Batch normalization: ； Introducing MLP+Sigmoid to dynamically generate the scaling factor, expressed as: ； In the formula, This represents a dynamic scaling factor generated by the multilayer perceptron and the sigmoid function, used for adaptive scaling of the normalized features; in, ； Among them, weight , , For batch statistics, These are learnable parameters; ReLU activation: .

Images