A bellows seal valve leakage fault intelligent diagnosis method

Abstract

The present application relates to the field of electric digital data processing, and in particular to a bellows seal valve leakage fault intelligent diagnosis method, comprising collecting operation data of the bellows seal valve under various states, constructing an original sensor time sequence matrix corresponding to each sample for each sample; labeling the working condition category of each sample as the true label of each sample; calculating the physical information enhanced input matrix of each sample; calculating the global feature vector; training the diagnosis model to obtain the trained diagnosis model; real-time collecting new operation data, inputting the trained diagnosis model after processing, obtaining the prediction probability distribution corresponding to each working condition category, taking the working condition category with the maximum probability as the diagnosis result output, enabling the dynamic state regulation and control deep time sequence network to adaptively adjust the information retention and update strategy according to the current leakage evolution degree, enabling the diagnosis model to focus on the key time segments in the leakage evolution process, and being used for online monitoring and early warning of bellows seal valve leakage in the field of nuclear power and petrochemical industry.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and specifically to an intelligent diagnostic method for leakage faults in bellows-sealed valves. Background Technology

[0002] In the semiconductor industry, bellows are mainly used for thermal expansion compensation, vibration absorption, and pressure regulation in piping systems. Bellows-sealed valves are typically used to control the flow of fluid within the bellows.

[0003] Bellows-sealed valves are prone to malfunction during operation due to factors such as alternating stress, corrosive media, temperature cycling, and mechanical fatigue, leading to fluid leakage within the bellows. To reduce safety incidents, multiple sensors need to be deployed to monitor the bellows-sealed valve's status in real time. By analyzing the signals collected by each sensor through independent channels, it can be determined whether the valve is operating normally, experiencing external or internal leakage, or exhibiting early, minor leakage. Existing diagnostic methods have the following main shortcomings:

[0004] (1) Existing diagnostic methods often treat the data of each channel as independent inputs, ignoring the propagation path and causal relationship between physical quantities. This makes it easy to misjudge the vibration and acoustic emission response caused by normal pressure fluctuations as leakage. The weak abnormal signals in the early micro-leakage stage are easily submerged in background noise and normal coupling response, making it difficult to extract the weak correlation features corresponding to the early micro-leakage stably, thus reducing the diagnostic accuracy.

[0005] (2) The formation and aggravation of leakage faults in bellows-sealed valves have obvious stages. Early minor leaks often manifest as weak features that accumulate slowly over a long period of time, while severe leaks manifest as strong features that increase rapidly in a short period of time. However, the temporal neural network of the existing method adopts a fixed state update mechanism. Regardless of the stage of the input sequence, its information retention and forgetting methods are exactly the same. This leads to the easy forgetting of the slowly accumulated weak features in the early stage and insufficient response to the rapidly increased strong features in the later stage. It is impossible to effectively model the complete evolution process of leakage from its inception to its aggravation, which reduces the diagnostic accuracy.

[0006] (3) Existing methods rely solely on cross-entropy loss for probability classification. For cases where early minor leakage conditions and normal conditions have overlapping features, or where internal leakage conditions and external leakage conditions have different physical mechanisms but similar features, it is difficult to learn discriminative feature distributions. This leads to blurred classification boundaries between confused categories and makes it difficult to improve diagnostic accuracy.

[0007] (4) The existing methods use fixed-step recurrent networks or single-scale convolutions that cannot take into account the time scale differences of different leakage types. They are difficult to specifically enhance short but critical leakage anomalies while suppressing long periods of no leakage, resulting in a decrease in diagnostic accuracy.

[0008] Therefore, existing intelligent diagnostic methods for bellows-sealed valve leakage faults have the problem of low diagnostic accuracy. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an intelligent diagnostic method for leakage faults in bellows-sealed valves with high diagnostic accuracy.

[0010] To solve the above-mentioned technical problems, the present invention provides an intelligent diagnostic method for leakage faults in bellows-sealed valves, comprising:

[0011] S1. Continuously collect the operating data of the bellows sealing valve under various conditions through sensor devices. The operating data within the same collection cycle is a sample. Construct the original sensor time series matrix corresponding to each sample. Label the operating condition category of each sample as the true label of each sample.

[0012] S2. Calculate the physical information enhancement input matrix for each sample based on the original sensor time-series matrix of each sample;

[0013] S3. Calculate the global feature vector based on the physical information of all samples to enhance the input matrix;

[0014] S4. The trained diagnostic model is obtained by training the diagnostic model based on the global feature vector;

[0015] S401. Calculate the predicted probability distribution of each working condition category based on the global feature vector;

[0016] S402. Construct a total loss function based on the predicted probability distribution and train the diagnostic model to obtain the trained diagnostic model;

[0017] S403. The diagnostic model is trained using the mini-batch stochastic gradient descent method to obtain the trained diagnostic model;

[0018] S5. Collect new operational data in real time, input it into the trained diagnostic model after executing S2, obtain the predicted probability distribution corresponding to each working condition category, and output the working condition category with the highest probability as the diagnostic result.

[0019] As a further improvement of the present invention: step S2 includes:

[0020] S201. Extract pressure field data and non-pressure field data from the original sensor time series matrix. The non-pressure field data includes mechanical motion field data, acoustic vibration field data and temperature field data. Calculate the complex transfer function of each non-pressure field relative to the pressure field data.

[0021] S202. Calculate the physical information enhancement input matrix based on the complex transfer function of each non-pressure field relative to the pressure field data.

[0022] Preferably, step S201 includes:

[0023] S2011. Extract mechanical motion field data, pressure field data, acoustic vibration field data, and temperature field data from the original sensor time series matrix;

[0024] S2012. Perform time-frequency transformation on the downstream pressure signal, mechanical motion field data, acoustic vibration field data, and temperature field data respectively to obtain the spectrum of the downstream pressure signal, the spectrum of the mechanical motion field data, the spectrum of the acoustic vibration field data, and the spectrum of the temperature field data.

[0025] S2013. Using the spectrum of the downstream pressure signal as a reference, calculate the complex transfer function from the downstream pressure to the mechanical motion field, the complex transfer function from the downstream pressure to the acoustic vibration field, and the complex transfer function from the downstream pressure to the temperature field based on the spectrum of the mechanical motion field data, the spectrum of the acoustic vibration field data, and the spectrum of the temperature field data.

[0026] Preferably, step S202 includes:

[0027] S2021. Multiply the spectrum of the downstream pressure signal by the complex transfer function of each non-pressure field relative to the pressure field data to obtain the expected response spectrum of each non-pressure field.

[0028] S2022. Perform inverse time-frequency transformation on the expected response spectrum of each non-pressure field to restore it to the time domain and obtain the estimated value of each non-pressure field;

[0029] S2023. Subtract the estimated value of each non-pressure field from the data of each non-pressure field to calculate the leakage-related residual components of each non-pressure field. The leakage-related residual components of each non-pressure field include the leakage-related residual components of the mechanical motion field, the leakage-related residual components of the acoustic vibration field, and the leakage-related residual components of the temperature field.

[0030] S2024. The pressure field data, the residual components related to leakage in the mechanical motion field, the residual components related to leakage in the acoustic vibration field, and the residual components related to leakage in the temperature field are concatenated column by column to obtain the physical information enhancement input matrix.

[0031] As a further improvement of the present invention: step S3 includes:

[0032] S301. Using the physical information enhancement input matrix of each sample as input, extract local change patterns at different time scales along the time axis, and obtain multi-scale local features at each time step after fusion;

[0033] S302. Calculate the leakage evolution stage factor based on the normalized physical information enhanced input matrix;

[0034] S303. Calculate the output sequence of the second-layer evolution-sensing gated recurrent unit based on the leakage evolution stage factor;

[0035] S304. Calculate the global feature vector based on the output sequence of the second-layer evolution-aware gated recurrent unit.

[0036] Preferably, step S301 includes:

[0037] S3011. The physical information of each sample is used to enhance the input matrix to form an enhanced training set. The mean of each channel of all samples in the enhanced training set at all time steps and the standard deviation of each channel of all samples in the enhanced training set at all time steps are calculated.

[0038] S3012. Based on the mean of each channel of all samples in the augmented training set at all time steps and the standard deviation of each channel of all samples in the augmented training set at all time steps, perform channel normalization processing on each physical information augmentation input matrix to obtain the normalized physical information augmentation input matrix.

[0039] S3013. Extract short-time local time segments, medium-time local time segments, and long-time local time segments from the normalized physical information enhancement input matrix, respectively;

[0040] S3014. Perform 1D convolution operations on the short-time local time segment, the medium-time local time segment, and the long-time local time segment respectively to obtain the short-time local convolution feature, the medium-time local convolution feature, and the long-time local convolution feature;

[0041] S3015. Calculate the multi-scale local features at each time step based on the short-time local convolution features, medium-time local convolution features, and long-time local convolution features.

[0042] Preferably, step S303 includes:

[0043] S3031. Input the second-layer hidden state of each time step, the multi-scale local features of each time step, and the leakage evolution stage factor into the update gate calculation module to calculate the update gate;

[0044] S3032. Input the second-layer hidden state of each time step, the multi-scale local features of each time step, and the leakage evolution stage factor into the reset gate calculation module to calculate the reset gate;

[0045] S3033. Calculate the historical state after adjustment by the reset door based on the reset door;

[0046] S3034. The hyperbolic tangent activation function is used to calculate the candidate hidden state of the new state candidate value based on the multi-scale local features of each time step, the historical state after reset gate adjustment, and the leakage evolution stage factor.

[0047] S3035. Use the update gate to perform a weighted fusion of the second-level hidden state and the candidate hidden state of the previous time step to obtain the hidden state of the current time step;

[0048] S3036. Calculate the hidden state of the current time step for all time steps along the time axis in sequence to obtain the output sequence of the first layer evolutionary perception gated loop unit; input the output sequence of the first layer evolutionary perception gated loop unit into the second layer evolutionary perception gated loop unit to obtain the output sequence of the second layer evolutionary perception gated loop unit.

[0049] Preferably, step S304 includes:

[0050] S3041. Extract the second-layer hidden state at each time step from the output sequence of the second-layer evolution-aware gated recurrent unit;

[0051] S3042. Input the second-layer hidden state of each time step into the attention scoring module to calculate the unnormalized importance score of each time step;

[0052] S3043. Calculate the attention weight for each time step based on the unnormalized importance score of each time step;

[0053] S3044. Calculate the global feature vector based on the attention weights at all time steps and the second-layer hidden states at all time steps.

[0054] As a further improvement of the present invention: step S401 includes:

[0055] S4011. Input the global feature vector into the fully connected classification layer and then map it to all working condition categories to obtain the classification score vector;

[0056] S4012. Perform a Softmax transformation on the classification score vector to obtain the predicted probability distribution.

[0057] As a further improvement of the present invention: step S402 includes:

[0058] S4021. Construct cross-entropy loss based on the predicted probability distribution and the true labels;

[0059] S4022. Set a category center for each working condition category in the feature space;

[0060] S4023. Calculate the class center distance loss based on the distance between the global feature vector of each sample and the class centers;

[0061] S4024. The total loss is obtained by weighted summation of the cross-entropy loss and the class center distance loss.

[0062] The beneficial effects of this invention are as follows: The intelligent diagnostic method for leakage faults in bellows-sealed valves provided by this invention has a high diagnostic accuracy. To separate the changes caused by normal pressure coupling from the additional anomalies induced by leakage, S2, after obtaining the complex transfer function, reconstructs the expected responses of each physical field under normal pressure disturbances using the downstream pressure signal. This expected response is then removed from the original signal, and targeted separation is performed based on the physical propagation path of the leakage fault. This highlights the weak anomalies corresponding to early, minor leaks, improves the separability between different leakage types, and extracts residual components that better reflect leakage anomalies. Simultaneously, the original pressure field data is retained to directly reflect the pressure drop and pressure fluctuations caused by leakage. S3 enables the dynamic state control deep temporal network to adaptively adjust its information retention and update strategies according to the current degree of leakage evolution, and allows the diagnostic model to focus on key time segments in the leakage evolution process. S301's multi-scale local feature extraction can simultaneously cover fault modes at different time scales, enabling the dynamic state control deep temporal network to simultaneously focus on short-term, medium-term, and long-term local temporal features at the same time location. Conventional methods typically only append the degree of leakage evolution as an additional input feature to the beginning of a dynamic state-controlled deep temporal network. However, S303 directly applies the degree of leakage evolution to the update gate, reset gate, and candidate state generation process of the gated recurrent unit. This allows the internal state update mechanism to truly change with the leakage stage, enabling early, minor leakage stages to retain historical weak features more fully, and severe leakage stages to absorb current strong anomalies more quickly, thereby simultaneously improving early detection capability and severe fault response capability. Compared to direct average pooling or directly taking the features from the last time step, S304's adaptive attention temporal pooling is more suitable for retaining short but crucial leakage anomaly fragments, while also accommodating slow trend fragments that evolve over long periods. S4 introduces a class center distance constraint during classifier training, making similar samples more concentrated in the feature space and dissimilar samples more separated, forming a more compact intra-class distribution and clearer inter-class boundaries in the feature space, thereby improving the ability to distinguish confused classes. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the principle of the present invention; Figure 2 The image shows a comparison of the time-frequency spectra of acoustic emissions under different operating conditions. Figure 2 (a) is the spectrum of the downstream pressure when there is internal leakage in the valve; Figure 2 (b) is the acoustic emission spectrum during valve leakage; Figure 2 (c) is the spectrum of the downstream pressure during early minor leakage; Figure 2 (d) is the acoustic emission spectrum during early, minute leakage; Figure 3 Figure 1 shows the experimental results related to residual components of acoustic emission channel leakage. Figure 4 The attention weight distribution is adapted for attention-sequential pooling. Detailed Implementation

[0064] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0065] Definitions of relevant terms in this article:

[0066] (1) Normal: refers to the working condition where the bellows sealing valve has no leakage and operates smoothly;

[0067] (2) External leakage of valve: refers to the condition in which the medium leaks outside the bellows sealing structure, usually accompanied by acoustic emission impact and local temperature change;

[0068] (3) Internal leakage of valve: refers to the working condition in which the medium leaks from the high-pressure side to the low-pressure side due to damage to the internal sealing surface of the valve. It is mainly manifested as pressure fluctuation after the valve and abnormal valve stem chatter.

[0069] (4) Early micro-leakage: refers to the initial leakage stage where the leakage amount is extremely small and has not yet caused a significant pressure drop and acoustic emission energy. Its characteristics are weak and easily submerged in background noise.

[0070] like Figure 1 As shown, the intelligent diagnostic method for leakage faults in bellows-sealed valves provided by this invention includes:

[0071] S1. Continuously collect the operating data of the bellows sealing valve under various conditions through sensor devices. The operating data within the same collection cycle is a sample. Construct the original sensor time series matrix corresponding to each sample. Label the operating condition category of each sample as the true label of each sample, and provide a true category basis for subsequent supervised learning.

[0072] The sensing equipment includes a displacement sensor mounted on the stem of the bellows-sealed valve, a first pressure sensor mounted at the inlet of the bellows-sealed valve to measure the upstream pressure, a second pressure sensor mounted at the outlet of the bellows-sealed valve to measure the downstream pressure, an acoustic emission sensor mounted on the surface of the bellows-sealed valve body near the connection between the bellows and the valve cover, a vibration sensor mounted on the valve body support or pipeline support structure of the bellows-sealed valve, and a temperature sensor mounted on the outer wall of the valve body or the outer wall of the outlet pipeline of the bellows-sealed valve. All sensors in the sensing equipment adopt a synchronous sampling method, and the sampling rate is uniformly set according to the rate of change of each physical quantity to a value sufficient to cover the frequency range of pressure fluctuations, acoustic emission impacts, and mechanical vibrations caused by leakage, so as to ensure sufficient time resolution for transient anomalies corresponding to early minor leaks.

[0073] Operational data can include actual data and simulated data. Actual data is the state data generated by the bellows-sealed valve operating under actual conditions, which is continuously collected by sensor devices. Simulated data is the state data generated by the bellows-sealed valve operating under different artificially simulated conditions, which is continuously collected by sensor devices. Artificially simulated conditions include normal, external leakage (different severity), internal leakage (different severity), and early minor leakage (different severity).

[0074] Both actual and simulated data include valve stem displacement signals measured by displacement sensors, inlet valve pressure signals measured by first pressure sensors, outlet valve pressure signals measured by second pressure sensors, emission signals measured by acoustic emission sensors, vibration signals measured by vibration sensors, and temperature signals measured by temperature sensors. All signals are time-series data.

[0075] The size is constructed based on each operational data. The original sensor time series matrices, The total number of channels. , The total number of time steps; the original sensor timing matrix. line, number The elements of the column are , It is a row index, and also a time step index. ; It is a column index, and also a channel index. ; Also represents the first The time step, the first Sensor data from the first channel, the first The time step represents the current time step, the _th time step The time step represents the previous time step, the _th time step Each time step represents the next time step;

[0076] Original sensor timing matrix Represented as: ,

[0077] In the formula, for The Each channel corresponds to a specific sensor type. This is the valve stem displacement signal, which is also the sensor data for the first channel; This is the pressure signal before the valve, and also the sensor data for the second channel; This is the pressure signal after the valve, and also the sensor data from the third channel; This is the acoustic emission signal, which is also the sensor data from the fourth channel; This is a vibration signal, and also the sensor data from the 5th channel; This is the temperature signal, and also the sensor data from the 6th channel;

[0078] Operating conditions include normal, external leakage, internal leakage, and early minor leakage. In practice, external leakage, internal leakage, and early minor leakage can be further classified according to their severity, which can be based on the diameter of the leakage orifice.

[0079] The operating condition category of the simulated data is determined directly according to the category of the artificially simulated operating condition; the operating condition category of the actual data needs to be determined based on the detection results of offline detection methods; offline detection methods include bubble leak detection, pressure decay test, helium mass spectrometry leak detection, etc. First, the existence of a leak is confirmed by offline detection methods, and if it is confirmed, the type of leak is confirmed.

[0080] All samples are divided into a training set for subsequent diagnostic model parameter learning, a validation set for hyperparameter tuning and early stop control, and a test set for final diagnostic performance evaluation, either chronologically or randomly. The number of samples in each class in the training set should be kept basically balanced. For classes with fewer samples (such as early micro-leakage), data augmentation or resampling strategies can be used to appropriately expand the sample size to avoid bias in the diagnostic model towards the majority class.

[0081] S2. Calculate the physical information enhancement input matrix for each sample based on the original sensor time-series matrix of each sample, so as to realize the physical information decoupling and reconstruction of multi-sensor data;

[0082] S201. Extract pressure field data and non-pressure field data from the original sensor time series matrix. The non-pressure field data includes mechanical motion field data, acoustic vibration field data and temperature field data. Calculate the complex transfer function of each non-pressure field relative to the pressure signal after the valve in the pressure field data.

[0083] S2011. Extract mechanical motion field data, pressure field data, acoustic vibration field data, and temperature field data from the original sensor time series matrix column by column. Each physical field data is a matrix.

[0084] Mechanical sports field data The size is Corresponding to the valve stem displacement signal Pressure field data The size is Including the valve inlet pressure signal With valve downstream pressure signal Therefore Acoustic vibration field data The size is Including acoustic emission signals With vibration signal Therefore Temperature field data The size is Corresponding temperature signal Therefore It can also be expressed as: ;

[0085] S2012. Perform time-frequency transformation on the downstream pressure signal, mechanical motion field data, acoustic vibration field data, and temperature field data respectively to obtain the spectrum of the downstream pressure signal, the spectrum of the mechanical motion field data, the spectrum of the acoustic vibration field data, and the spectrum of the temperature field data, so as to realize the conversion of the original time domain signal into a frequency domain representation that varies with time, which is convenient for analyzing the frequency response characteristics of leakage disturbances propagating from the pressure field to other physical fields;

[0086] Time-frequency transformation refers to using short-time Fourier transform to perform frame-by-frame analysis on each channel to obtain spectral representations at different time frames and frequency positions. The window length can be 64, 128, or 256 sampling points, the frame shift can be 1 / 4 to 1 / 2 of the window length, and the window function can be a Hamming window.

[0087] No. The frequency, the first Spectrum of the pressure signal after the frame valve The calculation formula is: ,

[0088] In the formula, Frequency index; For time frame indexing; For the time index of the signal; It is a discrete downstream pressure signal; For The window function centered on; The imaginary unit;

[0089] No. The frequency, the first Spectrum of frame mechanical motion field data The calculation formula is: ,

[0090] In the formula, It represents a discrete valve stem displacement signal;

[0091] No. The frequency, the first Spectrum of frame acoustic field data The calculation formula is: ,

[0092] In the formula, This is a general term for discrete acoustic emission signals or vibration signals; in actual calculations, it refers to... , Perform short-time Fourier transforms on each;

[0093] No. The frequency, the first Spectrum of frame temperature field data The calculation formula is: ,

[0094] In the formula, It is a discrete temperature signal;

[0095] S2013. Using the spectrum of the downstream pressure signal as a reference, calculate the complex transfer function from the downstream pressure to the mechanical motion field, the complex transfer function from the downstream pressure to the acoustic vibration field, and the complex transfer function from the downstream pressure to the temperature field based on the spectrum of the mechanical motion field data, the spectrum of the acoustic vibration field data, and the spectrum of the temperature field data.

[0096] The complex transfer functions quantitatively describe the amplitude gain and phase shift of the pressure downstream of the valve at a certain frequency after the disturbance propagates through the physical propagation path in the mechanical motion field, acoustic vibration field, and temperature field. In other words, the complex transfer functions represent the dynamic transmission relationship. The downstream pressure signal is selected as the reference signal because the downstream pressure is most sensitive to leakage and can reflect the local pressure drop and flow state changes caused by leakage earlier. By establishing the propagation relationship between various physical fields through a unified reference signal, the residual components obtained by subsequent reconstruction have stronger physical interpretability and are conducive to distinguishing between internal leakage, external leakage, and early micro-leakage.

[0097] When calculating each complex transfer function, the target physical field spectrum is divided by the downstream pressure spectrum, and the transfer relationship in complex form is obtained point by point of frequency. When the downstream pressure amplitude at certain frequency positions is too small and direct division is prone to numerical instability, a very small positive stable term can be added to the denominator to avoid abnormal amplification. Complex division is performed point by point of frequency and frame by frame to obtain the transfer relationship in complex form.

[0098] Complex transfer function from downstream pressure to non-pressure field The calculation formula is: ,

[0099] In the formula, This is a type identifier for non-pressure fields. , For the pressure after the valve to the mechanical motion field at frequency ,time The complex transfer function at the point, The gain and phase changes when the pressure disturbance propagates to the mechanical motion field are given by a dimension of . The complex matrix represents the transmission relationship between the downstream pressure and the valve stem displacement, where, The number of frequency points in the short-time Fourier transform. For time frames; The frequency of the pressure after the valve to the acoustic field ,time The complex transfer function at the point, The gain and phase changes of the pressure disturbance propagating to the acoustic field are of size . The complex tensor has its first dimension corresponding to two channels: acoustic emission and vibration. Each channel has its own independent transmission relationship. For the valve downstream pressure to temperature field at frequency ,time Complex transfer function at point; The gain and phase changes when the pressure disturbance propagates to the temperature field are of size . The complex matrix represents the transmission relationship between pressure and temperature downstream of the valve; This is a positive stable term used to prevent division by zero; its value can be set to [value missing]. , or ;

[0100] S2014. Smooth the obtained complex transfer functions to improve the stability of the transfer relationship;

[0101] Smoothing refers to local mean smoothing or median smoothing along the time axis and frequency axis, respectively, to suppress random noise and occasional spikes, so that the transmission relationship can better reflect the stable physical laws of leakage propagation, rather than the random fluctuations at a single moment.

[0102] By comparing the time-spectrum data of the valve downstream pressure and acoustic emission under conditions of internal leakage and early minor leakage, the following results were obtained. Figure 2 , Figure 2 (a) is the spectrum of the downstream pressure when there is internal leakage in the valve; Figure 2 (b) is the acoustic emission spectrum during valve leakage; Figure 2 (c) is the spectrum of the downstream pressure during early minor leakage; Figure 2 (d) shows the acoustic emission spectrum during early, minute leaks. In each graph, the horizontal axis represents time (seconds), the vertical axis represents frequency (Hz), and the color represents the power spectral density (dB). Figure 2 (a) ~ Figure 2 (d) reveals the non-stationary frequency components caused by leakage, proving that the short-time Fourier transform can capture the transient frequency characteristics caused by leakage, providing a time-frequency domain basis for subsequent complex transfer function calculations.

[0103] S202. Calculate the physical information enhancement input matrix based on the complex transfer function of each non-pressure field relative to the pressure field data, so as to realize the reconstruction of the leakage-related components according to the dynamic transfer relationship;

[0104] S2021. Multiply the spectrum of the downstream pressure signal by the complex transfer function of each non-pressure field relative to the pressure field data to obtain the expected response spectrum of each non-pressure field. The expected response spectrum of each non-pressure field includes the expected response spectrum of the mechanical motion field. Expected response spectrum of acoustic vibration field Expected response spectrum of temperature field If S2014 has smoothed out the transitive relationships, then the complex transfer function here refers to the smoothed complex transfer function.

[0105] Non-stress field expected response spectrum The calculation formula is: ;

[0106] The mechanical motion field corresponds to only one channel, thus yielding one expected response spectrum; the acoustic vibration field corresponds to two channels, therefore the acoustic emission channel and the vibration channel are reconstructed independently; the temperature field corresponds to one channel, thus yielding one expected temperature response spectrum. The dimensions of the expected response spectra of each non-pressure field are consistent with the dimensions of the spectrum of each non-pressure field. For example... The size is ,and Consistent, the expected response spectrum only includes the component contributed by the downstream pressure disturbance through physical transmission, and does not include additional components generated by the leakage itself (such as shock, friction, local overheating, etc.), representing the response that should occur under normal pressure fluctuations.

[0107] S2022. Perform inverse time-frequency transformation on the expected response spectrum of each non-pressure field to restore it to the time domain, obtaining the estimated values ​​of each non-pressure field as a benchmark for the physical field response that should be generated by normal pressure fluctuations under leak-free conditions. The estimated values ​​of each non-pressure field include the estimated values ​​of the mechanical motion field. Acoustic field estimation value and temperature field estimates The inverse time-frequency transform is implemented using the inverse short-time Fourier transform.

[0108] The signal's first The formula for calculating the estimated value of the non-pressure field at each time point is: ;

[0109] The size is This indicates the expected valve stem displacement response derived solely from normal pressure fluctuations; the signal's first... The estimated mechanical motion field at time point is: ; The size is This represents the expected acoustic emission and vibration response derived solely from normal pressure fluctuations; the signal's first... The estimated acoustic field at time point is: ; The size is This indicates the expected temperature response derived solely from normal pressure fluctuations; the signal's first... The estimated temperature field at time point is: ;

[0110] S023. Subtract the estimated value of each non-pressure field from each non-pressure field to calculate the leakage-related residual components of each non-pressure field. The leakage-related residual components of each non-pressure field include the leakage-related residual components of the mechanical motion field. Residual components related to acoustic vibration field leakage Residual components related to temperature field leakage To remove the influence of normal pressure disturbances, making it easier for dynamic state control deep temporal networks to capture the weak features of early small leaks. The size is , used to characterize abnormal valve stem chatter caused by leakage; The size is The two columns correspond to the acoustic emission leakage component and the vibration leakage component, respectively, and are used to characterize abnormal phenomena such as turbulent noise and impact vibration. The size is , used to characterize temperature anomalies caused by leaks;

[0111] Non-pressure field leakage related residual components The calculation formula is: ;

[0112] S024. The pressure field data, the residual components related to leakage in the mechanical motion field, the residual components related to leakage in the acoustic vibration field, and the residual components related to leakage in the temperature field are concatenated column-wise to obtain the physical information enhancement input matrix; Physical Information Enhancement Input Matrix The size is The six columns correspond sequentially to the inlet valve pressure signal, outlet valve pressure signal, residual components related to mechanical motion field leakage, acoustic emission leakage component, vibration leakage component, and residual components related to temperature field leakage. ;

[0113] Using pressure field data without residual processing allows subsequent diagnostic models to utilize both direct changes in inlet and outlet pressures, as well as leak-related anomalous components in other physical fields. Specifically, the inlet pressure reflects changes in the inlet pressure, while the outlet pressure reflects the sensitive response to leakage disturbances; both together constitute crucial foundational information for leak detection. This also ensures that the physical information enhancement input matrix retains the original state information of the pressure field while amplifying anomalous information from other physical fields that deviates from normal propagation patterns, providing clearer and more physically meaningful input.

[0114] The residual components related to the leakage of the acoustic emission channel were obtained. Figure 3 , Figure 3 The horizontal axis represents time in seconds (s), and the vertical axis represents the acoustic emission residual component in volts (V). It can be seen that the waveform in the 0s~0.2s segment has almost no spikes, which indicates that the background noise and normal flow noise in the original acoustic emission signal are significantly suppressed. The significant negative spikes near 0.1~0.5s indicate that the impact spikes caused by leakage in the residual component are significantly enhanced. This shows that decoupling the physical information of the pressure reference can extract weak acoustic emission leakage characteristics, which is especially beneficial for the detection of early micro-leakage.

[0115] S3. Calculate the global feature vector based on the physical information of all samples to construct a dynamic state-controlled deep temporal network;

[0116] S301. Using the physical information enhancement input matrix of each sample as input, extract local change patterns at different time scales along the time axis, and obtain multi-scale local features at each time step after fusion;

[0117] S3011. Use the physical information of each sample to enhance the input matrix to form an augmentation training set, and calculate the mean of each channel for all samples in the augmentation training set at all time steps. Compared with the standard deviation of each channel of all samples in the augmented training set over all time steps , and All are vectors with a dimension of 6;

[0118] S3012. Based on the mean of each channel of all samples in the augmented training set at all time steps and the standard deviation of each channel of all samples in the augmented training set at all time steps, the physical information augmentation input matrix is ​​normalized to obtain a normalized physical information augmentation input matrix, so that the data of each channel are in a similar numerical range, and avoid one channel from dominating the training of the deep temporal network for dynamic state regulation due to its large dimension.

[0119] Normalized physical information enhancement input matrix The calculation formula is: ;

[0120] S3013. Extract short-time local time segments from the normalized physical information enhancement input matrix to capture sudden acoustic and vibration shocks and transient pressure disturbances, respectively. Medium-duration local time segments used to capture pressure fluctuations and mechanical responses of moderate duration. And long-term local time segments used to capture temperature drift and slowly evolving leakage trends. ;

[0121] During extraction, fixed-length neighborhood segments are symmetrically extracted bidirectionally along the time axis in the normalized physical information enhancement input matrix, with each time step as the center. For parts with insufficient length at the sequence boundary, zero padding, mirror padding, or repeated boundary value padding can be used to complete the length.

[0122] No. Each time step, across Local time segments of each time step The calculation formula is: ,

[0123] In the formula, This is the window length of the local time segment, i.e., the number of time steps the local time segment spans. Therefore, a short local time segment is a local time segment spanning 3 time steps, a medium local time segment is a local time segment spanning 7 time steps, and a long local time segment is a local time segment spanning 15 time steps. To enhance the normalized physical information input matrix in the th... A single-line slice at each time step. , To obtain the matrix The The rows and all columns are combined to form a row vector with a dimension of 6.

[0124] Substitution It can be calculated that: ;

[0125] S3014. Perform one-dimensional convolution operations on the short-time local time segment, the medium-time local time segment, and the long-time local time segment respectively to obtain the short-time local convolution feature. Local convolution features in mid-time Long-term local convolution features The convolution direction is the time axis direction, the number of convolution input channels is 6, and the number of convolution kernels is uniformly set to 64. , , The dimension of each element is the number of convolutional kernels. After boundary padding, the length of the output sequence is the same as the number of input time steps.

[0126] No. Each time step, across The formula for calculating the local convolutional features at each time step is: ,

[0127] In the formula, It is a linear rectification activation function; For a window length of A 1D convolution kernel for a local time segment; This is a convolution operation; For a window length of The bias vector of a local time segment, p=3 corresponds to short-time scale bias, p=7 corresponds to medium-time scale bias, and p=15 corresponds to long-time scale bias;

[0128] Short-time local convolution features, i.e., the first Local convolutional features spanning three time steps; local convolutional features in the middle time step, i.e., the first time step... Local convolutional features spanning 7 time steps; long-term local convolutional features are the first... Local convolutional features spanning 15 time steps; 1D convolutional kernels for local time segments with a window length of 3. The size is 1D convolution kernel for local time segments with a window length of 7 The size is A 1D convolution kernel for a local time segment with a window length of 15. The size is ; Let be the bias vector for a local time segment with a window length of 3. The bias vector for a local time segment with a window length of 7; The bias vector for a local time segment with a window length of 15;

[0129] Used to capture sudden anomalies such as short-term (approximately several sampling points) impacts and transient acoustic emission spikes; Used to capture pressure fluctuations and mechanical response patterns with medium cycles (approximately tens of milliseconds); It is used to capture long-term (hundreds of milliseconds or seconds) drift, slow temperature changes, and leakage trends.

[0130] S3015. Calculate multi-scale local features for each time step to express short-term shocks, medium-term fluctuations and long-term drift information around the current time position based on short-term local convolution features, medium-term local convolution features and long-term local convolution features;

[0131] No. Multiscale local features at each time step The dimension is , can be set , The calculation formula is: ,

[0132] In the formula, To be , , Concatenate the features into a 192-dimensional vector; Here is the weight matrix of the linear combination layer, with size . The linear combination layer is a fully connected layer with no bias, and its input is the concatenated layer. 3D vector, output is The dimensional vector is used to weight and fuse local features at different scales, enabling dynamic state-controlled deep temporal networks to automatically select the most effective scale combination.

[0133] S302. Calculate the leakage evolution stage factor based on the normalized physical information enhanced input matrix to estimate the leakage evolution stage corresponding to the current time position, so that the dynamic state control deep temporal network can perceive whether the current state is closer to normal, early small leakage state or serious leakage state.

[0134] S3021. Let the length of the historical time window used to estimate the leakage evolution stage be... , can be set ; in the Each time step, for arrive Within the range, the residual components related to the leakage of the acoustic emission field in the normalized physical information enhancement input matrix are statistically analyzed to obtain the residual components related to the leakage of the acoustic emission field in the historical time window of each time step.

[0135] when Less than In such cases, the length of an existing historical time window can be used for calculation, or the missing portion can be filled in by boundary copying. For example, when At that time, the length of the historical time window ,but Since the data exceeds the sequence range, the data from the 1st to the 10th time step is used as the available historical time window length (a total of 10 points). Statistical measures (such as mean, difference, etc.) are calculated for this insufficient window. No further padding or repeated padding with the value from the 1st time step is performed to make the window length reach 32. The former approach is preferred, directly using existing actual data without forced padding to avoid introducing spurious trends.

[0136] S3022. Calculate the downstream pressure trend index for each time step based on the downstream pressure signal within the historical time window of each time step to describe the downstream pressure trend. The downstream pressure trend index for each time step can reflect whether the downstream pressure continues to decrease. The continuous decrease of the downstream pressure usually means that leakage is developing.

[0137] The post-valve pressure trend indicators at each time step include one or more of the following: the average post-valve pressure within the historical time window, the rate of decrease relative to the initial value, and the first-order difference mean within the historical time window.

[0138] The decrease relative to the initial value is the difference between the start and end points of the historical time window;

[0139] During implementation, the downstream pressure trend indicators at each time step may only include the average downstream pressure within the historical time window and the decrease relative to the initial value.

[0140] No. Post-valve pressure trend indicators at each time step The calculation formula is: = ,

[0141] In the formula, This represents the average post-valve pressure within a historical time window. This represents the decrease relative to the initial value; This is an operation to retrieve the minimum value; This is for the operation of retrieving the maximum value; for The second column (i.e., the pressure after the valve) is in the... The value of each time step; To distinguish from Time step index;

[0142] S3023. Calculate the acoustic emission energy index of each time step based on the residual components related to the acoustic emission field leakage within the historical time window of each time step. The acoustic emission energy index of each time step reflects whether the acoustic emission energy continues to increase. A continuous increase in acoustic emission energy usually means that the fluid disturbance and leakage noise are enhanced.

[0143] The acoustic emission energy indicators at each time step include the sum of the squares of the amplitudes of each sampling point within the historical time window, the average of the squares of the amplitudes of each sampling point within the historical time window, and the average value of the signal after taking the absolute value within the historical time window.

[0144] In practice, the sum of the squares of the amplitudes of each sampling point within the historical time window and the average of the squares of the amplitudes of each sampling point within the historical time window can be used as the acoustic emission energy index for each time step. Alternatively, the average value of the signal after taking the absolute value within the historical time window can be used as the acoustic emission energy index for each time step.

[0145] The formula for calculating the acoustic emission energy index at each time step is as follows: ,

[0146] In the formula, It is the sum of the squares of the amplitudes of each sampling point within the historical time window; The average value of the signal after taking the absolute value within the historical time window; for The fourth column (i.e., the acoustic emission leakage component) is in the... The value of each time step;

[0147] S3024. Calculate the leakage evolution stage factor for each time step based on the downstream pressure trend index and the acoustic emission energy index for each time step.

[0148] During calculation, the post-valve pressure trend index and the acoustic emission energy index at each time step are normalized to a unified numerical range, resulting in normalized post-valve pressure trend index and normalized acoustic emission energy index at each time step. The normalized post-valve pressure trend index and normalized acoustic emission energy index at each time step are then concatenated and input into a small fully connected mapping layer, which is activated by the Sigmoid activation function. Normalization can avoid unreasonable bias in the evolution stage estimation results caused by different dimensions, so that the obtained leakage evolution stage factor is not a fixed threshold judgment result set by humans, but a continuous state description quantity that can be automatically optimized with training.

[0149] The dimension of the leakage evolution stage factor at each time step is: , can be set The leakage evolution stage factor at each time step ranges from 0 to 1. The closer the value is to 0, the closer the current state is to normal or early minor leakage; the closer the value is to 1, the closer the current state is to obvious leakage or serious leakage.

[0150] No. Leakage evolution stage factor at each time step The calculation formula is: ,

[0151] In the formula, Use the Sigmoid activation function; Here is the weight matrix of the fully connected mapping layer, with size . ; This is a pressure trend indicator normalized by Min-Max or Z-score. The acoustic emission energy index after Min-Max or Z-score normalization; The bias vector of the fully connected mapping layer;

[0152] S303. Calculate the output sequence of the second-layer evolution-aware gated recurrent unit based on the second-layer hidden state at each time step, the multi-scale local features at each time step, and the leakage evolution stage factor, so as to complete the state update of the evolution-aware gated recurrent unit and enable the dynamic state-controlled deep temporal network to adopt different information accumulation and forgetting strategies at different leakage evolution stages.

[0153] S3031. The second-layer hidden state at each time step, the multi-scale local features at each time step, and the leakage evolution stage factor are jointly input into the update gate calculation module to calculate the update gate used to represent the amount of new information received and the amount of old information retained at the current time step. Introducing the leakage evolution stage factor enables the dynamic state-controlled deep temporal network to dynamically adjust the opening and closing degree of the update gate according to the severity of the current leakage. The values ​​of each element of the update gate are between 0 and 1. The calculation formula is: ,

[0154] In the formula, For the first The second hidden state at each time step represents the temporal memory information retained from the previous time step, denoted by dimension . ; To update the weight matrix of the gate; To be , , The joint vector obtained by concatenation; To update the bias vector of the gate;

[0155] S3032. The second-layer hidden states at each time step, the multi-scale local features at each time step, and the leakage evolution stage factor are all input into the reset gate calculation module to calculate the reset gate used to determine the amount of historical information ignored and the amount of current local features introduced when generating the current candidate state. Introducing the leakage evolution stage factor allows the dynamic state control deep temporal network to ignore more old historical information that is no longer suitable for the current state when the leakage is severe. This allows the dynamic state control deep temporal network to not only rely on the current local features, but also to combine the current leakage evolution stage to determine whether to weaken the influence of the old state. Reset gate The calculation formula is: ,

[0156] In the formula, To reset the weight matrix of the gate; To reset the gate's bias vector;

[0157] S3033. Calculate the reset-gated historical state based on the reset gate to represent the portion of historical information that was allowed to be retained when generating candidate states. , The calculation formula is: ,

[0158] In the formula, Represents the Hadamard product;

[0159] S3034. A hyperbolic tangent activation function is used to calculate the candidate hidden state representing the new state candidate value derived from the new input and some historical information at the current time step, based on the multi-scale local features of each time step, the historical state after reset gate adjustment, and the leakage evolution stage factor. , The dimension is , The calculation formula is: ,

[0160] In the formula, It is the hyperbolic tangent activation function; Let be the weight matrix of the candidate hidden states; is the bias vector of the candidate hidden state;

[0161] S3035. Use the update gate to perform a weighted fusion of the second-level hidden state and the candidate hidden state from the previous time step to obtain the hidden state of the current time step. , The calculation formula is: ;

[0162] In the above formula, This term is used to determine the number of new candidate states to accept, because The value of the old state retention factor determines the tendency to retain historical states; a larger value indicates a greater preference for preserving historical states, while a smaller value indicates a greater preference for accepting new states. Since both the update gate and reset gate are affected by the leakage evolution stage factor, the hidden state update strategy automatically changes with the degree of leakage evolution. When leakage is severe, the leakage evolution stage factor causes the update gate to decrease, allowing the dynamically regulated deep temporal network to absorb more new features. In the early stages with minor leakage, the update gate is close to 1, preserving weak historical features.

[0163] S3036. Calculate the hidden state of the current time step along the time axis in sequence to obtain the output sequence of the first layer evolution-aware gated recurrent unit. The output sequence of the first-layer evolution-aware gated recurrent unit is input into the second-layer evolution-aware gated recurrent unit to obtain the output sequence of the second-layer evolution-aware gated recurrent unit. ;

[0164] The set of hidden states of the first layer at all time steps, with size . , can be set , The calculation formula is: ,

[0165] In the formula, For the first layer in the first Hidden states at each time step;

[0166] It can represent more abstract and stable leakage evolution time-series characteristics, with a size of , can be set ; The calculation formula is: ,

[0167] In the formula, For the second layer in the first Hidden states at each time step;

[0168] The first-layer evolution-aware gated recurrent unit and the second-layer evolution-aware gated recurrent unit are stacked with the same structure. Except for the leakage evolution stage factor, which shares the same estimate, all other parameters are independent. The first-layer evolution-aware gated recurrent unit has an input dimension of 64, a hidden dimension of 64, and an output of... The second-layer evolutionary perception gated recurrent unit is... As input, the hidden dimension is 32, and the output is... .

[0169] S304. Calculate the global feature vector for classification based on the output sequence of the second-layer evolution-aware gated recurrent unit to achieve adaptive attention temporal pooling, enabling the diagnostic model to automatically focus on key time segments in the leakage evolution process;

[0170] S3041. Extract the second-layer hidden state at each time step from the output sequence of the second-layer evolution-aware gated recurrent unit;

[0171] No. The second hidden state at each time step This represents the output feature of the second-layer evolutionary perception gated recurrent unit at time step t, with dimension . ; The calculation formula is: ,

[0172] In the formula, for The Rows, all columns;

[0173] S3042. Input the second-layer hidden state of each time step into the attention scoring module to calculate the unnormalized importance score of each time step. Unnormalized importance score at each time step The score is a scalar representing the importance of the current time step. The higher the score, the more likely the corresponding time step contains key information that helps distinguish the type of leak. The calculation formula is: ,

[0174] In the formula, The attention weight vector has a dimension of 32 and is a learnable and trainable parameter that is automatically updated during the training of the deep temporal network under dynamic state control. Here is the weight matrix of the attention fully connected layer, with size . ; Here is the bias vector of the attention fully connected layer;

[0175] S3043. Calculate the attention weight for each time step based on the unnormalized importance score of each time step. Attention weights at each time step Indicates the first The contribution ratio of each time step to the global representation, where the sum of the attention weights for all time steps is 1. The calculation formula is: ,

[0176] In the formula, To find the sum index, iterate through all time steps;

[0177] S3044. Calculate a global feature vector that incorporates key dynamic information from the entire leakage evolution process, based on the attention weights at all time steps and the second-layer hidden states at all time steps. , dimension , The calculation formula is: .

[0178] Analysis of the attention weight distribution obtained by adaptive attention temporal pooling yields... Figure 4 , Figure 4 The horizontal axis represents time, measured in seconds (s), and the vertical axis represents attention weights. Attention weights are dimensionless; a larger value indicates a higher level of attention the diagnostic model pays to the data at that moment. For example... Figure 4 As shown, the attention weights for the time segments of approximately 0.3s to 0.5s and approximately 0.6s to 0.75s are significantly higher than those for other regions, indicating that the diagnostic model can automatically focus attention on the time segments containing key leakage information.

[0179] The dynamic state-regulating deep temporal network is the core of the diagnostic model, consisting of three interconnected functional modules:

[0180] Multi-scale local feature extraction module: Taking the normalized physical information enhancement input matrix as input, it extracts local convolutional features at three scales (short-time, medium-time, and long-time) in parallel along the time axis, and merges the three into multi-scale local features at each time step;

[0181] Evolutionary sensing temporal modeling module: First, the leakage evolution stage factor is calculated based on the normalized physical information enhanced input matrix. The multi-scale local features and the leakage evolution stage factor are input into two stacked evolutionary sensing gated recurrent units. The output sequence of the second layer evolutionary sensing gated recurrent unit is obtained through the state update mechanism.

[0182] The adaptive attention temporal pooling module performs attention scoring and weighted summation on the second-layer hidden states at each time step in the output sequence of the second-layer evolutionary perception gated recurrent unit to obtain a global feature vector for final classification.

[0183] S4. The trained diagnostic model is obtained by training the diagnostic model based on the global feature vector;

[0184] S401. Calculate the predicted probability distribution of each working condition category based on the global feature vector; during the training phase, compare the predicted probability distribution with the true label for subsequent calculation of cross-entropy loss; during the inference phase, directly use the category with the highest predicted probability as the classification result of the sample or output the probabilities of all categories simultaneously for the diagnostic model to evaluate confidence.

[0185] S4011. After inputting the global feature vector into the fully connected classification layer, it is mapped to all working condition categories. There are 4 working condition categories in total, resulting in a classification score vector of length 4. , This represents the unnormalized classification score corresponding to the four operating condition categories. The calculation formula is: ,

[0186] In the formula, Here is the weight matrix of the fully connected classification layer, with size . ; This is the bias vector of the fully connected classification layer, with a dimension of 4;

[0187] S4012. Perform a Softmax transformation on the classification score vector to obtain the predicted probability distribution. , This represents the predicted probability distribution of the input sample belonging to one of the four working condition categories. It has a length of 4 and the sum of all elements is 1. The calculation formula is: ,

[0188] In the formula, For the Softmax function; Indexed by category; The first of the classification score vectors One element;

[0189] S402. Construct a total loss function based on the predicted probability distribution and train the diagnostic model to obtain the trained diagnostic model;

[0190] S4021. Construct a cross-entropy loss function based on the predicted probability distribution and the true labels to constrain the output of the diagnostic model so that the predicted probability distribution is consistent with the true labels. , The smaller the value, the closer the classification result of the diagnostic model is to the true label; The calculation formula is: ,

[0191] In the formula, For the first The true label of the nth sample represents the nth sample. The true category number of each sample; For sample index; The total number of samples; For the first The predicted probability distribution of each sample;

[0192] S4022. Set a category center for each working condition category in the feature space, and the first... The learnable feature centers corresponding to each category are: The dimension is the same as the global feature vector; when hour, For normally corresponding learnable feature centers; when hour, The learnable feature center corresponding to valve leakage; when hour, The learnable feature center corresponding to the valve internal leakage; when hour, Learnable feature centers corresponding to early, minor leaks; There are two update methods: one is to calculate the mean of global feature vectors belonging to the same category within each training batch and update the corresponding category center accordingly; the other is to use the category center as a learnable parameter and define four dimensions in the dynamically state-controlled deep temporal network. The trainable vectors are updated along with the diagnostic model parameters. The initial values ​​can be set to random small values ​​or initialized from the mean of the sample features in the early training phase.

[0193] S4023. Calculate the class center distance loss based on the distance between the global feature vector of each sample and the class centers. To ensure that the global feature vector of each sample is as close as possible to the class center corresponding to its true class (i.e., the center of its own class) and as far away as possible from the class centers of other classes (i.e., the centers of other classes), the two requirements of being close to the center of its own class and far away from the centers of other classes are combined into an auxiliary loss to reduce the degree of dispersion within the same class and increase the margin between different classes. The calculation formula is: ,

[0194] In the formula, It is the L2 norm, also known as the Euclidean norm; To distinguish from The category index is a positive integer; It is a positive interval, and can take the value 1.0; For the first Learnable feature centers corresponding to each category This means representing categories other than the true category;

[0195] S4024. Weight the cross-entropy loss and the class center distance loss to obtain the total loss. This enables the diagnostic model to have both high classification accuracy and a clearer feature space structure.

[0196] The calculation formula is: ,

[0197] In the formula, To balance the weighting coefficients of the two parts of the loss, When the size is small, the focus should be on optimizing classification accuracy. When the value is large, more emphasis is placed on the structural constraints of the feature space distribution. This can be achieved by setting appropriate... It can strike a balance between classification accuracy and feature separability, and can take a value of 0.1;

[0198] S403. The diagnostic model is trained using the mini-batch stochastic gradient descent method to obtain the trained diagnostic model;

[0199] Mini-batch stochastic gradient descent divides all samples into multiple training batches. For each training batch, the original sensor time-series matrix corresponding to each sample is input sequentially. S2 is executed to obtain the physical information enhancement input matrix, S3 is executed to obtain the global feature vector, and S401 and S402 are executed to obtain the total loss of each sample. The average of the total loss of all samples is calculated to obtain the total loss of the current batch. Backpropagation is performed on the total loss of the current batch to calculate the gradient of each parameter of the dynamic state-controlled deep temporal network and update the network parameters using the Adam optimizer. When the validation set index no longer improves for several consecutive rounds, training is stopped and the optimal diagnostic model parameters are saved to obtain the trained diagnostic model.

[0200] The validation set metrics are validation set classification accuracy, validation set loss, or F1 score.

[0201] The initial learning rate of the Adam optimizer can be taken as... or The batch size can be 32 or 64. During training;

[0202] S5. Real-time acquisition of new operational data, inputted into the trained diagnostic model after S2, yields the predicted probability distribution for each operating condition category. The operating condition category with the highest probability is output as the diagnostic result. The entire diagnostic process can be executed sequentially across time windows, enabling continuous online monitoring and real-time alarm of the bellows sealing valve's leakage status. Simultaneously, it records the changing trends of the four probability values ​​at each diagnostic moment, providing maintenance personnel with auxiliary information for leak evolution and fault type determination.

Claims

1. A method for intelligent diagnosis of leakage faults in bellows-sealed valves, characterized in that, include: S1. Continuously collect the operating data of the bellows sealing valve under various conditions through sensor devices. The operating data within the same collection cycle is a sample. Construct the original sensor time series matrix corresponding to each sample. Label the operating condition category of each sample as the true label of each sample. S2. Calculate the physical information enhancement input matrix for each sample based on the original sensor time-series matrix of each sample; S3. Calculate the global feature vector based on the physical information of all samples to enhance the input matrix; S4. The trained diagnostic model is obtained by training the diagnostic model based on the global feature vector; S401. Calculate the predicted probability distribution of each working condition category based on the global feature vector; S402. Construct a total loss function based on the predicted probability distribution and train the diagnostic model to obtain the trained diagnostic model; S403. The diagnostic model is trained using the mini-batch stochastic gradient descent method to obtain the trained diagnostic model; S5. Collect new operational data in real time, input it into the trained diagnostic model after executing S2, obtain the predicted probability distribution corresponding to each working condition category, and output the working condition category with the highest probability as the diagnostic result.

2. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 1, characterized in that, The steps in S2 include: S201. Extract pressure field data and non-pressure field data from the original sensor time series matrix. The non-pressure field data includes mechanical motion field data, acoustic vibration field data and temperature field data. Calculate the complex transfer function of each non-pressure field relative to the pressure field data. S202. Calculate the physical information enhancement input matrix based on the complex transfer function of each non-pressure field relative to the pressure field data.

3. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 2, characterized in that, The steps in S201 include: S2011. Extract mechanical motion field data, pressure field data, acoustic vibration field data, and temperature field data from the original sensor time series matrix; S2012. Perform time-frequency transformation on the downstream pressure signal, mechanical motion field data, acoustic vibration field data, and temperature field data respectively to obtain the spectrum of the downstream pressure signal, the spectrum of the mechanical motion field data, the spectrum of the acoustic vibration field data, and the spectrum of the temperature field data. S2013. Using the spectrum of the downstream pressure signal as a reference, calculate the complex transfer function from the downstream pressure to the mechanical motion field, the complex transfer function from the downstream pressure to the acoustic vibration field, and the complex transfer function from the downstream pressure to the temperature field based on the spectrum of the mechanical motion field data, the spectrum of the acoustic vibration field data, and the spectrum of the temperature field data.

4. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 2, characterized in that, The steps in S202 include: S2021. Multiply the spectrum of the downstream pressure signal by the complex transfer function of each non-pressure field relative to the pressure field data to obtain the expected response spectrum of each non-pressure field. S2022. Perform inverse time-frequency transformation on the expected response spectrum of each non-pressure field to restore it to the time domain and obtain the estimated value of each non-pressure field; S2023. Subtract the estimated value of each non-pressure field from the data of each non-pressure field to calculate the leakage-related residual components of each non-pressure field. The leakage-related residual components of each non-pressure field include the leakage-related residual components of the mechanical motion field, the leakage-related residual components of the acoustic vibration field, and the leakage-related residual components of the temperature field. S2024. The pressure field data, the residual components related to leakage in the mechanical motion field, the residual components related to leakage in the acoustic vibration field, and the residual components related to leakage in the temperature field are concatenated column by column to obtain the physical information enhancement input matrix.

5. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 1, characterized in that, The steps in S3 include: S301. Using the physical information enhancement input matrix of each sample as input, extract local change patterns at different time scales along the time axis, and obtain multi-scale local features at each time step after fusion; S302. Calculate the leakage evolution stage factor based on the normalized physical information enhanced input matrix; S303. Calculate the output sequence of the second-layer evolution-sensing gated recurrent unit based on the leakage evolution stage factor; S304. Calculate the global feature vector based on the output sequence of the second-layer evolution-aware gated recurrent unit.

6. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 5, characterized in that, The steps in S301 include: S3011. The physical information of each sample is used to enhance the input matrix to form an enhanced training set. The mean of each channel of all samples in the enhanced training set at all time steps and the standard deviation of each channel of all samples in the enhanced training set at all time steps are calculated. S3012. Based on the mean of each channel of all samples in the augmented training set at all time steps and the standard deviation of each channel of all samples in the augmented training set at all time steps, perform channel normalization processing on each physical information augmentation input matrix to obtain the normalized physical information augmentation input matrix. S3013. Extract short-time local time segments, medium-time local time segments, and long-time local time segments from the normalized physical information enhancement input matrix, respectively; S3014. Perform 1D convolution operations on the short-time local time segment, the medium-time local time segment, and the long-time local time segment respectively to obtain the short-time local convolution feature, the medium-time local convolution feature, and the long-time local convolution feature; S3015. Calculate the multi-scale local features at each time step based on the short-time local convolution features, medium-time local convolution features, and long-time local convolution features.

7. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 5, characterized in that, The steps in S303 include: S3031. Input the second-layer hidden state of each time step, the multi-scale local features of each time step, and the leakage evolution stage factor into the update gate calculation module to calculate the update gate; S3032. Input the second-layer hidden state of each time step, the multi-scale local features of each time step, and the leakage evolution stage factor into the reset gate calculation module to calculate the reset gate; S3033. Calculate the historical state after adjustment by the reset door based on the reset door; S3034. The hyperbolic tangent activation function is used to calculate the candidate hidden state of the new state candidate value based on the multi-scale local features of each time step, the historical state after reset gate adjustment, and the leakage evolution stage factor. S3035. Use the update gate to perform a weighted fusion of the second-level hidden state and the candidate hidden state of the previous time step to obtain the hidden state of the current time step; S3036. Calculate the hidden state of the current time step for all time steps along the time axis in sequence to obtain the output sequence of the first layer evolutionary perception gated loop unit; input the output sequence of the first layer evolutionary perception gated loop unit into the second layer evolutionary perception gated loop unit to obtain the output sequence of the second layer evolutionary perception gated loop unit.

8. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 5, characterized in that, The steps in S304 include: S3041. Extract the second-layer hidden state at each time step from the output sequence of the second-layer evolution-aware gated recurrent unit; S3042. Input the second-layer hidden state of each time step into the attention scoring module to calculate the unnormalized importance score of each time step; S3043. Calculate the attention weight for each time step based on the unnormalized importance score of each time step; S3044. Calculate the global feature vector based on the attention weights at all time steps and the second-layer hidden states at all time steps.

9. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 1, characterized in that, The steps in S401 include: S4011. Input the global feature vector into the fully connected classification layer and then map it to all working condition categories to obtain the classification score vector; S4012. Perform a Softmax transformation on the classification score vector to obtain the predicted probability distribution.

10. The intelligent diagnostic method for leakage faults in a bellows-sealed valve according to claim 1, characterized in that, The steps in S402 include: S4021. Construct cross-entropy loss based on the predicted probability distribution and the true labels; S4022. Set a category center for each working condition category in the feature space; S4023. Calculate the class center distance loss based on the distance between the global feature vector of each sample and the class centers; S4024. The total loss is obtained by weighted summation of the cross-entropy loss and the class center distance loss.

Images