Mama network solenoid valve fault diagnosis method based on frequency domain characteristics

Abstract

According to the Mama network solenoid valve fault diagnosis method based on the frequency domain characteristics, the problem of solenoid valve system fault diagnosis can be solved. According to the method, a pneumatic solenoid valve fault data set is constructed by collecting operation signals such as voltage and current, and multi-scale frequency domain features of the signals are extracted by adopting Wavelet Packet Transform (WPT) and Discrete Fourier Transform (DFT). WPT-DFT preprocessing features are embedded into an improved channel-space joint attention mechanism module, and the improved channel-space joint attention mechanism module is combined with a Mamba network with extremely high sequence modeling capability to construct an end-to-end fault diagnosis model. Compared with a traditional convolutional neural network, the method can more effectively capture deep dynamic features in a time sequence and highlight a key frequency region, and experimental results show that the method has higher diagnosis precision and generalization ability and is suitable for intelligent detection of the electromagnetic valve under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of pneumatic solenoid valve fault diagnosis technology, specifically to a solenoid valve fault diagnosis method based on a Mamba network fusion of wavelet packet transform and discrete Fourier transform, for diagnosing solenoid valve faults. Background Technology

[0002] As a core actuator, the stability of the pneumatic solenoid valve's operating state directly affects the system's regulation response speed and efficiency. Therefore, timely and accurate identification of abnormal states during solenoid valve operation is of great significance for ensuring the performance of the thermal management system and improving system reliability.

[0003] However, solenoid valves typically operate in environments with high-frequency switching, complex thermal-fluid coupling, and strong interference. Their fault signals often exhibit non-stationarity, time-frequency coupling, and fuzzy characteristics, making it difficult for traditional fault diagnosis methods to guarantee accuracy and real-time performance under complex operating conditions. Early solenoid valve fault identification relied mainly on disassembly inspection and manual experience, which had significant drawbacks such as low diagnostic efficiency, high subjectivity, and inability to be applied online. To address these issues, researchers both domestically and internationally have proposed various intelligent fault diagnosis schemes based on physical modeling, signal processing, and expert systems. For example... Figure 3 As shown, the prior art solutions most relevant to this application will be briefly described below: 1) Model-based fault diagnosis methods: Model-based fault diagnosis methods utilize physical principles such as electromagnetism and fluid dynamics to analyze the response time, pressure, and other characteristics of solenoid valves to construct models for fault diagnosis. The drawbacks of this method are the high complexity of model construction, the high level of expertise required of researchers, and the difficulty in creating highly accurate models. Furthermore, model-based methods have limitations, being constrained by the accuracy and completeness of the model, and making it difficult to verify whether the model's results conform to reality.

[0004] 2) Fault diagnosis methods based on signal processing: Signal processing-based fault diagnosis methods primarily involve collecting and analyzing various signals generated during the operation of solenoid valves to extract fault features and achieve fault diagnosis. The specific process includes: collecting various signals during system operation using sensors; preprocessing the collected signals, such as filtering and noise reduction, to improve signal quality; using signal processing methods to extract characteristic quantities reflecting fault features, such as frequency components, amplitude, and phase, from the preprocessed signals; and analyzing and comparing the extracted features to determine the presence and severity of a fault. While signal processing methods can effectively extract fault features from various signals, different methods have varying capabilities and adaptability, requiring the selection of an appropriate method based on the specific fault type and signal characteristics. The model construction process is complex and exhibits poor adaptability to changes in operating conditions.

[0005] 3) Fault diagnosis methods based on expert systems and rule-based reasoning: In the field of solenoid valves, fault diagnosis methods based on expert systems and rule-based reasoning rely on manually constructed rule bases and logical reasoning mechanisms. Their core principle is to match fault phenomena with their causes using pre-defined "IF-THEN" rules. However, this method's rule base struggles to effectively integrate multi-dimensional sensor data; for example, the collaborative analysis of vibration, temperature, and pressure signals requires the manual design of complex logic chains.

[0006] In summary, current fault diagnosis methods for solenoid valves mainly suffer from the following problems: 1) Model-based methods rely on complex physical modeling processes, requiring a deep understanding of principles such as electromagnetism and fluid mechanics. Furthermore, model building demands a high level of expertise from researchers and is difficult to adapt to changes in working conditions (such as temperature and pressure fluctuations), which limits the accuracy of the models. 2) Although signal processing-based methods can extract fault features (such as frequency and amplitude), they are easily affected by environmental noise (such as mechanical vibration and electromagnetic interference), sensor errors and non-stationary signals (such as transient impacts). Manual feature extraction relies on experience judgment, has weak adaptability to complex fault modes, and is difficult to automate. 3) Fault diagnosis methods based on expert systems and rule-based reasoning are difficult to meet the dual requirements of high precision and real-time performance in industrial scenarios when dealing with complex signal characteristics and dynamic operating conditions due to problems such as rigid rules, difficulty in integrating multi-source data and insufficient generalization ability. Summary of the Invention

[0007] This invention proposes a fault diagnosis method for solenoid valves based on a Mamba network that integrates wavelet packet transform and discrete Fourier transform. This method classifies and diagnoses fault signals of solenoid valves, and can at least solve one of the technical problems in the background art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for diagnosing solenoid valve faults based on a Mamba network fusion of wavelet packet transform and discrete Fourier transform includes the following steps: S1. Collect the voltage and current signals of the faulty solenoid valve and construct a dataset in a unified format; S2. Perform wavelet packet transform (WPT) and discrete Fourier transform (DFT) on each one-dimensional time series signal to extract multi-resolution frequency domain features; S3. Introduce the frequency domain features of WPT and DFT as inputs into the frequency domain attention mechanism module to obtain the enhanced frequency domain representation; S4. Input the enhanced frequency domain representation into the Mamba network to complete the classification and prediction of solenoid valve fault types; S5. Verify the diagnostic performance and robustness of this method based on a standard pneumatic solenoid valve fault dataset.

[0009] Furthermore, step S1 specifically includes, Fault data of solenoid valves were collected using a DHDAS signal acquisition device. In the fault data acquisition experiment, the sampling rate of the data acquisition board of the DHDAS signal acquisition device was set to 5KHz to acquire voltage and current signals. The collected raw fault data was processed, and the fault data of each type was unified to construct a fault dataset of pneumatic solenoid valves.

[0010] Furthermore, step S2 specifically includes, This step aims to extract frequency domain feature information from the original one-dimensional time-series signal of the solenoid valve, so that subsequent models can identify the characteristic distribution of fault modes in the frequency domain. Two frequency domain methods, Wavelet Packet Transform (WPT) and Discrete Fourier Transform (DFT), are used to achieve multi-scale and global spectral representation of the time-series signal. This process mainly includes three steps: wavelet transform processing, Fourier transform processing, and WPT and DFT feature fusion.

[0011] Step S21: Set the transform parameters and perform wavelet packet decomposition. Let the input signal be x[n], its length be N, and the sampling frequency be... After L-layer wavelet packet decomposition, the small value can be obtained. There are 10 frequency bands, where the bandwidth of each frequency band is given by formula (1).

[0012] ; Where L represents the number of decomposition layers, L=4 is chosen to balance frequency, resolution, and computational cost. Wavelet packet decomposition is performed on the signal x[n] using a preset wavelet basis. In each decomposition layer, a pair of orthogonal filters, namely a low-pass filter h[n] and a high-pass filter g[n], are used to perform convolution and downsampling operations on the signal from the previous layer. The coefficient calculation formula for the k-th child node of the j-th layer is as follows: ; in, Indicates the corresponding filter, Given the input signal of the j-th layer, after completing the L-level decomposition, we can obtain... Sub-signals corresponding to each frequency band.

[0013] S22. Extract the features of wavelet decomposition for all Calculate the energy of a frequency band signal; frequency band energy. It can be represented as: ; The final WPT energy eigenvector is formed as follows: ; S23. Discrete Fourier Transform (DFT) processing: Perform a DFT on the original signal x[n] to obtain its amplitude spectrum and phase spectrum in the frequency domain. The specific calculation formula is as follows: ; The amplitude spectrum |X[K]| is taken, and features are further extracted, including key frequency domain indicators such as spectral center, spectral entropy, and dominant frequency. The final DFT feature vector is then formed. ; S24. Feature Concatenation and Normalization. The frequency domain feature vectors extracted by WPT and DFT methods are concatenated to form a unified and complete frequency domain feature, which is represented as follows: The formula is shown in (7). In order to eliminate the influence of dimensions and improve the generalization ability of the model, all features are normalized and Z-score standardization is used to obtain a new feature vector F.

[0014] ; Furthermore, step S3 specifically includes, Based on the normalized multidimensional frequency domain features obtained in step S2, to enhance the model's ability to perceive key frequency bands, a frequency domain feature enhancement module based on an attention mechanism is introduced. This module automatically learns the weight distribution of each frequency domain feature dimension to highlight frequency information that is more sensitive to fault detection. This module references the channel and spatial hybrid attention structure in CBAM (Convolutional Block Attention Module) and, combined with time-frequency characteristics, designs a dual-branch frequency domain attention module. To model the relative importance between each feature channel (frequency band), a channel attention mechanism is first adopted. Global average pooling and max pooling are performed on the feature vector F respectively to obtain the global average pooling vector. and global max pooling vector The two are concatenated and fed into a multilayer perceptron (MLP) with shared weights to calculate channel attention weights. The formula is as follows: ; in, Here is the perceptron weight matrix, and r is the channel compression ratio. This represents the Sigmoid activation function, while ReLU is the activation function used to introduce non-linearity. Finally, the channel weights are multiplied element-wise with the original features to achieve channel-direction enhancement. ; To further capture saliency in the frequency domain, a spatial attention module is designed. First, the original frequency domain features are subjected to average pooling and max pooling along the channel direction, resulting in two scalar sequences. These sequences are then concatenated into a two-dimensional description, which consists of the average pooling vector. and max pooling vector ,pass The functions are concatenated, and the concatenated result is input into a 1×1 convolutional layer to generate attention maps at frequency locations. The formula is as follows: ; Then the attention map With feature vectors Multiplication achieves weighted enhancement based on frequency position direction: ; Finally, the output is fused and enhanced. The enhanced frequency domain features of the final output are expressed as follows: ; Furthermore, step S4 specifically includes, In step S3, the fused WPT and DFT frequency domain features are enhanced by channel and frequency position weighting using a dual time-frequency attention mechanism module to obtain a discriminative frequency domain representation. To further mine global dependency features in the sequence and achieve accurate classification of solenoid valve fault types, a classification network based on the Mamba architecture was constructed, and the enhanced features were input into the Mamba network for processing. First, the enhanced frequency domain features were... The input is fed into the Mamba model as a one-dimensional sequence. Considering the Mamba model's requirements for input format, the input needs to be reshaped as follows: ; Where L represents the sequence length, which is equivalent to the length of the frequency dimension expansion. The input dimension is defined for each frequency point. Mamba is a sequence modeling architecture with linear time complexity, built upon the State-Space Model (SSM) concept, possessing excellent long-range dependency modeling capabilities and suitable for efficiently processing long-term sequences. The enhanced frequency domain representation sequence obtained in the previous stage is input into a multi-layered stacked Mamba Block module. The structure of each Mamba module includes normalization, state-space mapping, and residual connections, as shown below: ; in The representation layer normalization operation is used to stabilize model training; The state-space mapping module is used to model continuous dynamic processes in time series. By stacking N layers of Mamba Block modules, layer-by-layer in-depth frequency domain feature extraction and time series modeling can be achieved, effectively enhancing the model's ability to identify complex fault modes. After completing the time series modeling of frequency domain features, the sequence feature representation output by the Mamba network... This will be further compressed into a fixed-length vector representation using a global pooling mechanism. ; This vector contains comprehensive information about the current signal in both the frequency and time domains. It is then fed into a fully connected layer and a Softmax classifier to output the predicted probability distribution of each solenoid valve fault type: ; in, and Here, C represents the classifier parameters, and C represents the number of possible fault types of the solenoid valve. The predicted probability vector output by the model reflects the confidence level of the input signal belonging to each category. To achieve high-precision classification of solenoid valve fault types, this invention uses the cross-entropy loss function as the supervision signal, defined as follows: ; in, One-hot encoding of the real label. The predicted probability value for the corresponding category, where C is the total number of fault categories. To improve the model's generalization ability and convergence speed, the AdamW optimizer is used during training, combined with a cosine annealing scheduler to adjust the learning rate and avoid getting trapped in local optima.

[0015] Furthermore, step S5 specifically includes, To verify the effectiveness and robustness of the Mamba network-based solenoid valve fault diagnosis method proposed in this invention, which integrates wavelet packet transform and discrete Fourier transform, a standard experimental scheme was designed, and the system was evaluated on publicly available or self-built pneumatic solenoid valve fault datasets. Solenoid valve fault signal data collected from a standard pneumatic experimental platform was selected as the validation object. The dataset contains various typical fault types (such as coil short circuit, mechanical jamming, electrical open circuit, etc.), with multiple samples for each type. To ensure the fairness of the evaluation and the generalization ability of the model, the entire dataset was divided into training, validation, and test sets in a 7:1:2 ratio to ensure that samples of each type are evenly distributed across different subsets. During the experiment, the proposed method was compared with the mainstream benchmark model. The results showed that the proposed method significantly outperformed the comparative model in terms of overall performance, especially in long-term dependency modeling and multi-band fault feature extraction. The accuracy was improved by 3% to 8%, and the harmonic mean (F1-score) of precision and recall was significantly improved. This verified the synergistic effect of wavelet-Fourier fusion frequency domain features and Mamba sequence modeling, which significantly improved its fault diagnosis performance.

[0016] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0017] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0018] As can be seen from the above technical solutions, existing methods can be categorized into model-based fault diagnosis, signal processing-based fault diagnosis, and expert system- and rule-based fault diagnosis methods. Solenoid valve fault diagnosis generally relies on sensors, including characteristic quantities such as voltage, current, and flow rate. These methods all have limitations to varying degrees when dealing with solenoid valve fault diagnosis. The technical solution proposed in this invention can specifically address or improve the following technical shortcomings: 1) Model-based fault diagnosis: These methods heavily rely on accurate prior physical models, requiring modeling of the nonlinear dynamic characteristics and non-stationary signal behavior of solenoid valves. However, such modeling often fails to capture the real system behavior under complex operating conditions, leading to discrepancies between the model and actual signal characteristics. Furthermore, model construction demands a high level of expertise from researchers and is difficult to adapt to various complex operating conditions.

[0019] To address the aforementioned problems, this invention proposes a multi-resolution frequency domain analysis mechanism that integrates wavelet packet transform and discrete Fourier transform, effectively handling the non-stationary signal characteristics that occur during the operation of solenoid valves. Wavelet packet transform possesses excellent time-frequency localization capabilities, making it suitable for capturing local transient features, while discrete Fourier transform offers higher analytical accuracy in overall frequency domain energy analysis. The combination of these two technologies achieves a good balance between frequency and time resolution, enabling comprehensive extraction of the signal's time-frequency features and enhancing the discriminative power of feature representation.

[0020] 2) Signal processing-based methods: Although it can extract fault features, it is easily affected by environmental noise, sensor errors and non-stationary signals. Manual feature extraction relies on experience and judgment, has weak adaptability to complex fault modes, and is difficult to automate.

[0021] To address the aforementioned issues, this invention employs a frequency domain attention mechanism. This mechanism adaptively learns the importance of different frequency components in fault identification, suppressing redundant frequency information and highlighting key characteristic frequency bands. Compared to the uniform weighting or empirical weight allocation methods in traditional signal processing, the frequency domain attention mechanism can dynamically adjust based on the frequency response characteristics of different fault modes, thereby improving the effectiveness of feature representation and classification performance.

[0022] 3) Methods based on expert systems and rule-based reasoning: In the diagnosis of solenoid valve faults, due to problems such as rigid rules, difficulty in integrating multi-source data and insufficient generalization ability, it is difficult to meet the dual requirements of high precision and real-time performance in industrial scenarios when dealing with complex signal characteristics and dynamic operating conditions.

[0023] To address the aforementioned issues, this invention constructs a deep temporal modeling network based on the Mamba architecture. This network can fully exploit the temporal dependencies in the solenoid valve's operating signals and, combined with state-space modeling principles, effectively capture feature evolution trends over long time spans. Compared to traditional convolutional neural networks (CNNs) which rely on local receptive fields and recurrent neural networks (RNNs) which suffer from gradient vanishing, the Mamba network achieves more efficient information transmission and global modeling capabilities through a parallelized state update mechanism, significantly improving the accuracy of fault mode identification under complex, nonlinear, and variable operating conditions.

[0024] Compared with existing technologies, the solenoid valve fault diagnosis method based on the fusion of wavelet packet transform and discrete Fourier transform using a Mamba network proposed in this invention application has the following significant advantages and effects: 1) Possesses stronger time-frequency feature representation capabilities, significantly improving the completeness and discriminative power of fault feature extraction. This invention integrates wavelet packet transform (WPT) and discrete Fourier transform (DFT) for frequency domain analysis. WPT provides multi-level local time-frequency information, suitable for capturing short-term abrupt changes or local fault characteristics; DFT provides global spectral information, helping to identify stable periodic components. The two methods complement each other, and the constructed multi-resolution frequency domain feature matrix can more comprehensively characterize the time-frequency evolution of solenoid valve signals, greatly improving diagnostic accuracy and discrimination capability.

[0025] 2) Possesses higher fault identification accuracy and stronger environmental adaptability. This invention introduces a feature enhancement module based on a frequency domain attention mechanism. Through an adaptive weighting strategy, it highlights the response to key frequency bands and important moments, guiding the model to focus on information regions closely related to the fault. This enables the effective extraction of discriminative features even in the context of noise interference and signal ambiguity, significantly enhancing the robustness of fault identification.

[0026] 3) The diagnostic model is lightweight and efficient, supporting low-latency real-time fault detection. This invention employs the Mamba network as the backbone structure for feature extraction, possessing excellent long sequence modeling capabilities and high parallelism, effectively modeling long-term dependencies in solenoid valve operation data. Compared to traditional RNNs and LSTMs, the Mamba network significantly reduces computational resource consumption while maintaining accuracy, making it more suitable for deployment in edge computing devices or industrial control terminals for online inference.

[0027] 4) It has a stronger ability to adapt to non-stationary signals and diverse fault modes, and its generalization ability is better. Solenoid valves often face signal non-stationarity, dynamic changes in operating conditions, and interference from multiple types of complex faults in actual operation. This invention effectively captures the characteristic differences of different types of faults by combining time-frequency domain analysis and deep modeling, and has strong anti-interference and generalization capabilities. It is applicable to fault diagnosis tasks of pneumatic solenoid valves under various operating modes, structural types, and load environments. Attached Figure Description

[0028] Figure 1 This is the overall process flow of the embodiments of the present invention; Figure 2 This is a schematic diagram of a solenoid valve fault data acquisition system. Figure 3 Classify existing solenoid valve fault diagnosis methods. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0030] like Figure 1 As shown in this embodiment, the method for diagnosing fault signals of solenoid valves based on a Mamba network fusion of wavelet packet transform and discrete Fourier transform aims to solve the problems of low accuracy in identifying the operating status of solenoid valves and weak feature extraction capabilities under complex working conditions. This method uses the acquired one-dimensional time-series voltage and current signals as a foundation, employing a frequency domain feature extraction mechanism that fuses wavelet packet transform (WPT) and discrete Fourier transform (DFT), combined with a frequency domain attention enhancement module and a Mamba deep time-series modeling network, to achieve efficient identification of solenoid valve fault states. The overall process begins with acquiring voltage and current signals, performing WPT decomposition and DFT spectral analysis on each time-series signal to extract multi-scale, multi-frequency band frequency domain feature information, and constructing a fused frequency domain feature matrix. Subsequently, the fused frequency domain features are input into the frequency domain attention mechanism module to adaptively enhance key frequency channels and high-weight time segments. Then, the enhanced frequency domain representation is fed into the Mamba network model for time-series modeling and fault classification, utilizing its state-space modeling structure to achieve efficient expression of long sequence dependencies. Finally, the WPT-DFT-Mamba diagnostic network was tested and validated on a standard pneumatic solenoid valve fault dataset to evaluate its fault diagnosis accuracy, robustness, and real-time performance, and to verify the applicability and promotion value of this method in multi-operating-condition scenarios.

[0031] The specific steps of this plan are as follows: Step 1: Collect the current and voltage signals of the faulty solenoid valve and construct a dataset. The solenoid valve fault data acquisition system mainly consists of a solenoid valve fault test platform, solenoid valves, air source, power supply, sensors, and a data acquisition system. A 2-position 5-way pilot-operated solenoid valve with a DC 24V power supply is selected. Voltage and current signals are acquired using an electrical DHDAS signal acquisition device and its matching DHDAS dynamic signal acquisition and analysis system software, with a sampling frequency set at 5kHz. A current sensor is connected in series and a voltage sensor is connected in parallel in the solenoid valve's operating circuit. The sensors convert the physical signals driving the solenoid valve on the test platform into standard signals and transmit them to the data acquisition system. The schematic diagram is shown below. Figure 2As shown. Because it's impossible to control the start and end times of each solenoid valve operation precisely during raw data collection, the collected raw data includes useless data from before and after the solenoid valve's operation, and the number of data points in each raw data set is inconsistent. Therefore, to construct a solenoid valve fault dataset, the raw fault feature data needs to be preprocessed. This involves standardizing the number of fault feature data points for each set, removing useless non-feature data collected before and after solenoid valve operation, and then constructing the dataset. All data files in the dataset are saved in MATLAB data file format (.mat files).

[0032] It should be explained that the solenoid valve performance testing platform is used to test solenoid valves, requiring the setting of some basic parameters of the solenoid valve under different fault conditions. The solenoid valve is installed on the testing platform, and a stable supply of compressed air is provided by an air source (air compressor). The air pressure at the inlet of the solenoid valve is regulated by the inlet shut-off valve on the platform. The air compressor used can provide compressed air up to 1 MPa; to ensure a relatively stable supply of compressed air during the experiment, the maximum pressure at the air compressor inlet is limited to 0.7 MPa. This dataset covers four solenoid valve states: normal state, internal leakage, external leakage, and blockage fault. Each state contains four sets of data, with 10 data points collected from different samples in each set, totaling 160 feature data points.

[0033] Step 2: Perform wavelet packet transform (WPT) and discrete Fourier transform (DFT) on each one-dimensional time series signal to extract multi-resolution frequency domain features; This step aims to extract frequency domain feature information from the original one-dimensional time-series signal of the solenoid valve, so that subsequent models can identify the characteristic distribution of fault modes in the frequency domain. Two frequency domain methods, Wavelet Packet Transform (WPT) and Discrete Fourier Transform (DFT), are used to achieve multi-scale and global spectral representation of the time-series signal. This process mainly includes three steps: wavelet transform processing, Fourier transform processing, and WPT and DFT feature fusion.

[0034] Step 21: Set the transformation parameters and perform wavelet packet decomposition.

[0035] The acquired original timing signal x[n] of the solenoid valve is subjected to wavelet packet decomposition. Let the signal length be N and the sampling frequency be... =5kHz. At this sampling frequency, according to the Nyquist sampling theorem, the highest analyzable frequency of the signal is... The frequency range is 2500Hz. Therefore, frequency domain analysis will primarily focus on the spectral distribution within the range of 0–2500Hz. To achieve multi-scale frequency domain analysis of the signal, a wavelet packet decomposition layer number of L=4 is selected, striking a balance between frequency resolution and computational complexity. After L layers of decomposition, the signal is divided into… Each of the four equal-width sub-bands has a bandwidth of 312.5 Hz, which can be calculated using formula (1).

[0036] ; The frequency range of 0~2500Hz is subdivided into 16 sub-bands, each covering a spectrum of 312.5Hz, which can more finely characterize the frequency distribution features of the high-frequency part. In each layer of decomposition, a pair of orthogonal filters, namely a low-pass filter h[n] and a high-pass filter g[n], are used to perform convolution and downsampling operations on the signal of the previous layer. The coefficient calculation formula corresponding to the k-th sub-node of the j-th layer is as follows: ; in, This indicates the filter used at the corresponding node. This is the input signal for the j-th layer. After completing the L-layer decomposition, 16 sub-signals are obtained, corresponding to the component coefficients of 16 frequency bands.

[0037] S22. Extracting features from wavelet decomposition. To extract the energy information contained in each frequency band signal, the energy of the k-th frequency band is defined as: ; Energy in this frequency band This reflects the local energy intensity of the signal in that frequency sub-band, and can characterize the energy concentration characteristics of a specific fault in certain frequency bands. Combining all 16 sub-band energy features forms a wavelet packet energy feature vector: ; S23. Discrete Fourier Transform Processing. To obtain the global spectral structure characteristics, a Discrete Fourier Transform is performed on the original signal x[n] to obtain its amplitude spectrum and phase spectrum in the frequency domain. The specific calculation formula is as follows: ; The transformation result is a complex sequence. Extract its amplitude spectrum from it. Based on this, frequency domain statistical features are constructed, including spectral centroid, spectral entropy, and dominant frequency. The spectral centroid represents the location of the spectral "center of gravity," and its calculation formula is: ; Spectral entropy represents the degree of dispersion of the spectral energy distribution, and is calculated as follows: ; Dominant frequency, which is the frequency component with the largest amplitude. ; By extracting the above features, a discrete Fourier transform feature vector is formed: ; S24. Feature Concatenation and Normalization. To combine the local multi-scale frequency domain energy information provided by WPT with the global spectral statistical feature information provided by DFT, the two are concatenated to form a complete frequency domain feature vector: ; Considering that the inconsistent numerical units among different features may affect the convergence efficiency and classification accuracy of subsequent models, the Z-score normalization method is used to normalize the feature vectors. To improve the model's generalization ability, normalization is performed using the following standardized formula: ; in, and Here, represents the mean and standard deviation of the i-th feature in the training set, respectively. The normalized feature vector serves as the input to the subsequent attention mechanism and temporal modeling network, laying a solid foundation for the final realization of solenoid valve fault mode recognition.

[0038] Step 3: Input the WPT and DFT frequency domain features into the frequency domain attention mechanism module to obtain the enhanced frequency domain representation.

[0039] In the method described in this invention, to further enhance the model's ability to identify solenoid valve fault features in the frequency domain, after completing the frequency domain feature extraction and normalization process in step 2, a frequency domain feature enhancement module based on an attention mechanism is proposed. This module is used to weight the frequency domain feature vectors, thereby strengthening the focus on key frequency bands and suppressing redundant or non-discriminative frequency band information. Structurally, this module draws on the design concept of the Convolutional Block Attention Module (CBAM), while improving and adjusting it to address the distribution pattern and one-dimensional nature of the frequency domain features in this invention. A dual-branch attention enhancement structure adapted to the frequency domain feature representation is constructed, including two branches: a channel attention mechanism and a frequency position attention mechanism. This aims to achieve weighted selection and spatial saliency enhancement of frequency information, thereby extracting more discriminative fault features. To model the relative importance between each feature channel (frequency band), a channel attention mechanism is first adopted. In this branch, to measure the relative importance of each frequency sub-band feature in the fault identification task, global average pooling (GAP) and global max pooling (GMP) operations are applied to the input frequency domain feature vector F along the frequency channel dimension, respectively, to obtain two global statistical vectors: ; The two vectors are concatenated and fed into a multilayer perceptron (MLP) with shared weights for nonlinear modeling to calculate the channel attention weights. The formula is as follows: ; in, Here, r is the perceptron weight matrix, and r is the channel compression ratio, with the sign... This represents the Sigmoid activation function, while ReLU is the activation function used to introduce nonlinearity. Ultimately, the trace weights are... Channel-wise weighting is performed on the original feature F to obtain the channel-enhanced feature representation: ; in, This represents a broadcast multiplication operation, where each channel is multiplied by its corresponding attention weight. This mechanism effectively highlights the frequency sub-band responses that are more critical for fault detection in the frequency domain, enhancing the model's ability to focus on fault characteristic frequency bands. This mechanism can automatically enhance the model's ability to perceive fault frequency bands that are more active or sensitive in the frequency domain. To further capture local saliency in the frequency domain, a spatial attention module is introduced. First, the original frequency domain features are subjected to average pooling and max pooling along the channel dimension, resulting in two scalar sequences, respectively. and Then, the two are spliced ​​together along the frequency position dimension, through... The functions are concatenated, and the concatenated data is then input into a 1×1 convolutional layer to generate attention maps at frequency locations. The announcement is as follows: ; The frequency position weighting map Features enhanced by the aforementioned channels Multiplication achieves weighted enhancement of the frequency position dimension, resulting in a doubly enhanced frequency domain feature representation: ; Considering the synergistic weighting effect of the dual attention mechanism on channel and frequency positions, the two-level enhancement operations are fused into the following output feature: ; The aforementioned enhancement module, by introducing a feature weighting strategy based on a frequency domain attention mechanism, effectively improves the robustness and discriminative ability of the neural network in high-noise environments. This mechanism enables the model to adaptively focus on frequency sub-bands closely related to the fault and their salient positions, thereby extracting highly discriminative key features even when spectral information is ambiguous or heavily interfered with, thus improving the overall accuracy and stability of fault diagnosis.

[0040] Step 4: Input the enhanced frequency domain representation into the Mamba network to complete the classification and prediction of solenoid valve fault types.

[0041] In step S3, the fused WPT and DFT frequency domain features are enhanced by channel and frequency position weighting using a dual time-frequency attention mechanism module to obtain a discriminative frequency domain representation. To further mine global dependency features in the sequence and achieve accurate classification of solenoid valve fault types, a classification network based on the Mamba architecture was constructed, and the enhanced features were input into the Mamba network for processing. First, the enhanced frequency domain features were... The input is fed into the Mamba model as a one-dimensional sequence. Considering the Mamba model's requirements for input format, the input needs to be reshaped as follows: ; Where L represents the sequence length, which is equivalent to the length of the frequency dimension expansion. The input dimension is defined for each frequency point. Mamba is a sequence modeling architecture with linear time complexity, built upon the State-Space Model (SSM) concept, possessing excellent long-range dependency modeling capabilities and suitable for efficiently processing long-term sequences. The enhanced frequency domain representation sequence obtained in the previous stage is input into a multi-layered stacked Mamba Block module. The structure of each Mamba module includes normalization, state-space mapping, and residual connections, as shown below: ; in The representation layer normalization operation is used to stabilize model training; The state-space mapping module is used to model continuous dynamic processes in time series. By stacking N layers of Mamba Block modules, layer-by-layer in-depth frequency domain feature extraction and time series modeling can be achieved, effectively enhancing the model's ability to identify complex fault modes. After completing the time series modeling of frequency domain features, the sequence feature representation output by the Mamba network... This will be further compressed into a fixed-length vector representation using a global pooling mechanism. ; This vector contains comprehensive information about the current signal in both the frequency and time domains. It is then fed into a fully connected layer and a Softmax classifier to output the predicted probability distribution of each solenoid valve fault type: ; in, and Here, C represents the classifier parameters, and C represents the number of possible fault types of the solenoid valve. The predicted probability vector output by the model reflects the confidence level of the input signal belonging to each category. To achieve high-precision classification of solenoid valve fault types, this invention uses the cross-entropy loss function as the supervision signal, defined as follows: ; in, One-hot encoding of the real label. The predicted probability value for the corresponding category, where C is the total number of fault categories. A smaller loss function value indicates that the probability distribution output by the model is closer to the true distribution, resulting in better classification performance. During training, to optimize model parameters and minimize the aforementioned cross-entropy loss function, this invention employs the AdamW optimizer for gradient updates. Compared to the traditional Adam optimizer, the AdamW optimizer introduces a weight decay mechanism, which can more effectively prevent overfitting and improve the model's generalization performance. The parameter update rules of the AdamW optimizer are as follows: ; in, This represents the model parameters at step t. For learning rate, and These are the first and second moment estimates of the gradient after bias correction. It is a tiny constant to prevent the denominator from being zero. This refers to the weight decay coefficient. By directly incorporating a decay term into the gradient update, AdamW not only suppresses the unlimited growth of weight values ​​but also promotes the rational exploration of the model's parameter space, avoiding the bias issues in weight decay inherent in traditional Adam. To further improve training effectiveness and convergence speed, this invention also introduces a Cosine Annealing Learning Rate Scheduler. This scheduler dynamically adjusts the learning rate, causing it to periodically change according to the law of a cosine function throughout the training process, slowly decreasing from a large initial value to a set minimum value, and then repeating this process. This strategy helps the model escape local optima, promotes the search for the global optimum, reduces oscillations, and improves convergence speed. Its core idea is to gradually reduce the initial learning rate to the minimum learning rate according to a cosine curve during training. Its mathematical expression is: ; in, This represents the learning rate at step t. and These are the preset maximum and minimum learning rates, respectively. It is the number of steps in the current training cycle. This represents the total number of steps throughout the entire learning rate cycle. Using this method, the model rapidly explores the parameter space with a large learning rate in the early stages of training, then gradually reduces the learning rate to achieve stable convergence, balancing training speed and final performance.

[0042] Step 5: Verify the diagnostic performance and robustness of this method based on a standard pneumatic solenoid valve fault dataset.

[0043] The effectiveness and performance of the designed solenoid valve fault diagnosis method based on a Mamba network fusion of Wavelet Packet Transform (WPT) and Discrete Fourier Transform (DFT) were verified using the pneumatic solenoid valve fault dataset collected in step 1. To fully evaluate the diagnostic performance and generalization ability of the method, the dataset was divided into training, validation, and test sets in a ratio of 7:1:2 to ensure a balanced distribution of samples across different subsets and avoid sample bias affecting the fairness of the experimental results. Furthermore, to comprehensively compare the performance stability of the method under different training rounds, the training process was set to 10, 20, 50, 100, and 200 rounds, and the performance of multiple models under different round numbers was compared and analyzed. The invention was compared with a one-dimensional convolutional neural network model, a dual time-frequency attention network model based on STFT time-frequency analysis (DTFAN_STFT), and a dual time-frequency attention network model based on Morlet wavelet analysis. All models were compared using the same dataset, partitioning strategy, and training configuration to uniformly evaluate their diagnostic performance under different training epochs. Experiments show that, compared to the three methods mentioned above, this invention employs a fusion strategy of wavelet packet transform and discrete Fourier transform, achieving high-resolution spectrum extraction and multi-scale frequency band analysis. WPT has full-band decomposition capabilities, accurately locating the energy distribution of fault signals in each frequency band, thus compensating for the shortcomings of STFT and Morlet analysis in terms of spectral resolution and adaptability; while DFT possesses global frequency domain representation capabilities, providing accurate characterization of stable periodic features. Under different training epochs, the method of this invention outperforms the comparative models, exhibiting higher accuracy and stability. Under 200 training rounds, the average accuracy of the proposed method is improved by approximately 5.7%, 4.3%, and 3.1% compared to Backbone_CNN, DTFAN_STFT, and DTFAN_Morlet, respectively. Significant improvements are achieved in comprehensive performance indicators such as Precision, Recall, and F1-score. The F1-score remains above 0.94 across multiple fault categories, demonstrating good generalization ability and diagnostic robustness.

[0044] There are alternative solutions for some structures, devices, and method steps in the embodiments of the present invention, which will be described in detail below.

[0045] 1) Model Alternatives. In terms of sequence modeling and classification modules, this invention uses the Mamba neural network as the backbone modeling structure, which relies on the state-space model and long-term dependency extraction mechanism to achieve efficient identification of fault modes. However, in resource-constrained edge computing environments, it can also be replaced by lighter-weight temporal neural network structures such as TCN (Temporal Convolutional Network), GRU (Gated Recurrent Unit), or low-rank Transformer.

[0046] 2) Data augmentation and preprocessing method substitution. In the preprocessing of the original signal, in order to improve the robustness of the model, in addition to standardization, denoising and other operations, various data augmentation techniques can be introduced. For example, in the time domain, methods such as adding noise, time stretching and amplitude scaling can be used, and in the frequency domain, methods such as frequency masking and spectrum shifting can be used to simulate various faults and operating conditions and improve the model's generalization ability.

[0047] In summary, the core innovative technologies of this invention are as follows: 1) Construct a frequency domain feature extraction mechanism that integrates wavelet packet transform and discrete Fourier transform. This invention addresses the problem of non-stationary operation and complex frequency domain feature distribution of solenoid valve operating status signals under complex conditions. It proposes a frequency domain preprocessing method that integrates wavelet packet transform (WPT) and discrete Fourier transform (DFT). Wavelet packet transform, through multi-level signal decomposition, can extract local features of different frequency bands with fine granularity, effectively reflecting the changing trends of weak fault signals within specific frequency bands. Simultaneously, discrete Fourier transform provides global spectral information, revealing the overall frequency component distribution characteristics of the signal. By fusing the features extracted by these two frequency domain processing methods, a frequency domain feature matrix with multi-resolution representation capabilities is constructed, enhancing not only the identification ability of key frequency bands but also improving the robustness and completeness of feature representation.

[0048] 2) Design a feature weighting enhancement module based on frequency domain attention mechanism This invention, building upon frequency domain feature extraction, further introduces an attention mechanism to perform weighted modeling of the fused frequency domain feature matrix. Specifically, it is designed as a module fusing channel attention and spatial attention, referencing the CBAM structure and improving it based on the spectral distribution characteristics of solenoid valve signals to form a dual-frequency domain enhancement mechanism. Channel attention is used to capture the importance of features in different frequency bands, adaptively adjusting the weight distribution of different frequency bands; spatial attention focuses on regions with corresponding spectral changes in key time segments, enhancing the response capability to time-varying fault modes. This module effectively suppresses background noise and non-critical components during feature enhancement, significantly improving the model's ability to perceive weak fault signals, making it particularly suitable for complex application scenarios such as high-frequency background interference or the coexistence of multiple fault types.

[0049] 3) Using Mamba networks for time series modeling and fault identification To achieve efficient sequence modeling and real-time fault identification, this invention selects the Mamba neural network as the core feature extraction and classification model. Mamba possesses excellent long sequence modeling capabilities and high computational efficiency. By replacing traditional convolutional or attention mechanisms with state-space modeling, it can learn complex sequence dependencies while maintaining a lightweight model. Compared to traditional CNN, RNN, or Transformer structures, Mamba can handle fault evolution information over longer time spans with fewer parameters, improving its responsiveness to both slowly changing and abruptly changing fault types.

[0050] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0051] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0052] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the method for diagnosing fault signals of a solenoid valve using a CNN neural network with embedded time-frequency analysis as described in the above embodiments. It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0053] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0055] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing solenoid valve faults based on a Mamba network, characterized in that, Includes the following steps, S1. Collect the voltage and current signals of the faulty solenoid valve and construct a dataset in a unified format; S2. Perform wavelet packet transform (WPT) and discrete Fourier transform (DFT) on each one-dimensional time series signal to extract multi-resolution frequency domain features. S3. Introduce the frequency domain features of wavelet packet transform (WPT) and discrete Fourier transform (DFT) as inputs into the frequency domain attention mechanism module to obtain the enhanced frequency domain representation; S4. Input the enhanced frequency domain representation into the Mamba network to complete the classification and prediction of solenoid valve fault types; S5. The diagnostic performance and robustness of this method are verified based on a standard pneumatic solenoid valve fault dataset, and applied to solenoid valve fault diagnosis.

2. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 1, characterized in that: Step S1 specifically includes, Fault data of solenoid valves were collected using a DHDAS signal acquisition device. In the fault data acquisition experiment, the sampling rate of the data acquisition board of the DHDAS signal acquisition device was set to 5KHz to acquire voltage and current signals. The collected raw fault data was processed, and the fault data of each type was unified to construct a fault dataset of pneumatic solenoid valves.

3. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 1, characterized in that: Step S2 specifically includes three steps: wavelet transform processing, Fourier transform processing, and WPT and DFT feature fusion method. Step S21: Set the transform parameters and perform wavelet packet decomposition; let the input signal be x[n], the length be N, and the sampling frequency be... After L-level wavelet packet decomposition, the small value can be obtained. There are 10 frequency bands, where the bandwidth of each frequency band is given by formula (1); ； Where L represents the number of decomposition layers, and L=4 is chosen to balance frequency, resolution, and computational cost; wavelet packet decomposition is performed on signal x[n] using a preset wavelet basis; in each decomposition layer, a pair of orthogonal filters, namely low-pass filter h[n] and high-pass filter g[n], are used to perform convolution and downsampling operations on the signal of the previous layer; the coefficient calculation formula corresponding to the k-th child node of the j-th layer is as follows: ； in, Indicates the corresponding filter, Given the input signal of the j-th layer, after completing the L-level decomposition, we can obtain... Sub-signals corresponding to each frequency band; S22. Extract the features of wavelet decomposition for all Calculate the energy of a frequency band signal; frequency band energy. Represented as: ； The final WPT energy eigenvector is formed as follows: ； S23. Discrete Fourier Transform Processing: Perform a Discrete Fourier Transform on the original signal x[n] to obtain its amplitude spectrum and phase spectrum in the frequency domain; the specific calculation formula is as follows: ； The amplitude spectrum |X[K]| is taken, and features are further extracted, including key frequency domain indicators such as spectral center, spectral entropy, and dominant frequency; finally, a DFT feature vector is formed: ； S24. Feature Concatenation and Normalization: The frequency domain feature vectors extracted by the wavelet packet transform (WPT) and discrete Fourier transform (DFT) methods are concatenated to form a unified and complete frequency domain feature, which is represented as follows: The formula is shown in (7); in order to eliminate the influence of the dimension and improve the generalization ability of the model, all features are normalized and Z-score standardization is used to obtain the new feature vector F; 。 4. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 3, characterized in that: Step S3 specifically includes, Based on the normalized multi-dimensional frequency domain features obtained in step S2, in order to improve the model's ability to perceive key frequency bands, a frequency domain feature enhancement module based on an attention mechanism is introduced to automatically learn the weight distribution of each frequency domain feature dimension in order to highlight frequency information that is more sensitive to fault identification. Referring to the channel and spatial hybrid attention structure in CBAM (Convolutional Block Attention Module), and combining time-frequency characteristics, a dual-branch frequency domain attention module is designed.

5. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 4, characterized in that: Step S3 specifically includes, First, a channel attention mechanism is employed; then, global average pooling and max pooling are performed on the feature vector F respectively to obtain the global average pooling vector. and global max pooling vector The two are concatenated and fed into a multilayer perceptron (MLP) with shared weights to calculate channel attention weights. The formula is as follows: ； in, Here is the perceptron weight matrix, and r is the channel compression ratio. This represents the Sigmoid activation function, while ReLU is an activation function used to introduce non-linearity. Ultimately, the channel weights are multiplied element-wise with the original features to achieve channel-wise enhancement. ； To further capture saliency in the frequency domain, a spatial attention module is designed. First, the original frequency domain features are subjected to average pooling and max pooling along the channel direction, resulting in two scalar sequences, which are then concatenated into a two-dimensional description, namely the average pooling vector. and max pooling vector ,pass The functions are concatenated, and the concatenated result is input into a 1×1 convolutional layer to generate attention maps at frequency locations. The formula is as follows: ； Then the attention map With feature vectors Multiplication achieves weighted enhancement based on frequency position direction: ； Finally, the output is fused and enhanced. The enhanced frequency domain features of the final output are expressed as follows: 。 6. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 5, characterized in that: Step S4 specifically includes, In step S3, the fused wavelet packet transform (WPT) and discrete Fourier transform (DFT) frequency domain features are weighted and enhanced by channel and frequency position using a dual time-frequency attention mechanism module to obtain a discriminative frequency domain representation. To further explore the global dependency features in the sequence and achieve accurate classification of solenoid valve fault types, a classification network based on the Mamba architecture was constructed, and the enhanced features were input into the Mamba network for processing.

7. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 6, characterized in that: Step S4 specifically includes, First, the enhanced frequency domain features The input is fed into the Mamba model as a one-dimensional sequence; considering the Mamba model's requirements for input format, the input needs to be reshaped as follows: ； Where L represents the sequence length, which is equivalent to the length of the frequency dimension expansion. Input dimension for each frequency point; The enhanced frequency domain representation sequence obtained in the previous stage is input into a multi-layered stacked Mamba Block module; the structure of each Mamba module includes normalization, state space mapping, and residual connections, as shown below: ； in The representation layer normalization operation is used to stabilize model training; This represents the state-space mapping module, used for continuous dynamic processes in modulo time series. By stacking N layers of Mamba Block modules, we can achieve in-depth frequency domain feature extraction and time series modeling, which effectively enhances the model's ability to identify complex fault modes. After completing the temporal modeling of frequency domain features, the sequence feature representation output by the Mamba network... This will be further compressed into a fixed-length vector representation using a global pooling mechanism. ； This vector contains comprehensive information about the current signal in both the frequency and time domains. It is then fed into a fully connected layer and a Softmax classifier to output the predicted probability distribution of each solenoid valve fault type: ； in, and Here, C represents the classifier parameters, and C represents the number of possible fault types of the solenoid valve. This is the predicted probability vector output by the model, reflecting the confidence level of the input signal belonging to each category.

8. A method for diagnosing solenoid valve faults based on a Mamba network as described in claim 7, characterized in that: Step S4 also includes, To achieve high-precision classification of solenoid valve fault types, the cross-entropy loss function is used as the monitoring signal, defined as follows: ； in, One-hot encoding of the real label. The predicted probability value for the corresponding category, where C is the total number of fault categories.

9. A method for diagnosing solenoid valve faults based on a Mamba network, as described in claim 8, is characterized in that: Step S4 also includes, To improve the model's generalization ability and convergence speed, the AdamW optimizer is used during training, combined with a cosine annealing scheduler to adjust the learning rate and avoid getting trapped in local optima.

10. The method for diagnosing solenoid valve faults based on a Mamba network according to claim 1, characterized in that: Step S5 specifically includes, Solenoid valve fault signal data collected from a standard pneumatic test platform were selected as the verification object. The dataset contains a variety of typical fault types, and each type contains multiple samples. The entire dataset is divided into training, validation, and test sets in a 7:1:2 ratio to ensure that samples of each type are evenly distributed across different subsets.

Images