An intelligent monitoring fault diagnosis method and system based on a storage tank relief device

Abstract

The application discloses an intelligent monitoring fault diagnosis method and system based on a storage tank relief device, relates to the technical field of industrial equipment fault diagnosis, and comprises the following steps: acquiring running state time series data corresponding to a single opening and closing event of a breathing valve; generating an opening and closing state representation based on the running state time series data, wherein the opening and closing state representation represents dynamic state changes of the breathing valve from opening to back seating in the single opening and closing event; distinguishing normal opening and closing states of the breathing valve from fault abnormal states according to the opening and closing state representation, and generating a fault diagnosis result according to dynamic state changes corresponding to the fault abnormal states; and the method realizes process-based differentiation of the normal opening and closing states and the fault abnormal states, enhances the adaptability of variable-length opening and closing events through length adaptation and mask feature aggregation, and improves the accuracy and availability of fault diagnosis of valve clapper sticking, sealing failure, overpressure opening and the like through fault probability and fault level output.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically to an intelligent monitoring fault diagnosis method and system based on a storage tank venting device. Background Technology

[0002] Breather valves are critical safety accessories for pressure balancing and discharge control in storage tanks, and their operating status affects the tank's pressure release and seal maintenance. During long-term use, the valve disc and sealing surface are susceptible to media deposits, corrosion, and mechanical wear, leading to difficulties in opening, unstable reseating, or reduced sealing capacity. Existing diagnostic methods largely rely on manual maintenance and single-point alarms, with some scenarios using high-frequency vibration monitoring. Manual maintenance is inherently periodic, making it difficult to detect early anomalies promptly; high-frequency monitoring requires specialized equipment and specific on-site installation conditions and is easily affected by ambient noise; single-point alarms typically rely on pressure peak values ​​or rates of change, failing to reflect the complete opening and closing process of the breather valve.

[0003] In actual operation, fluctuations in normal loading and unloading, changes in initial tank pressure, and sensor response delays can all cause the pressure curve to deviate from the theoretical opening value for a short period of time, resulting in false alarms. Some faults may be close to the normal curve during the opening phase, only showing abnormalities during the pressure holding phase after reseating. Therefore, a single threshold is difficult to accurately distinguish between normal opening and closing states and faulty states. At the same time, the duration of different opening and closing events is not consistent, and the processing method of fixed length or fixed threshold is difficult to adapt to the temporal differences in the opening and closing process of the breather valve. An online diagnostic method that can identify the status of the entire opening and closing process is still needed. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing methods for diagnosing breathing valve faults rely on manual experience or single-point threshold judgments, which makes it difficult to reflect the complete opening and closing process of the breathing valve, have low accuracy in distinguishing between normal operating condition fluctuations and abnormal fault states, low real-time performance of early fault identification, and the problem of how to achieve online diagnosis of the opening and closing status of the breathing valve under conventional industrial sensor conditions.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent monitoring and fault diagnosis based on a storage tank venting device, comprising: acquiring time-series data of the operating status corresponding to a single opening and closing event of a breather valve; generating an opening and closing status representation based on the time-series data, wherein the opening and closing status representation indicates the dynamic state change of the breather valve from opening to reseating during a single opening and closing event; distinguishing between the normal opening and closing state and the faulty / abnormal state of the breather valve based on the opening and closing status representation, and generating a fault diagnosis result based on the dynamic state change corresponding to the faulty / abnormal state; wherein the time-series data of the operating status includes a pressure sequence inside the storage tank, a temperature sequence inside the storage tank, and at least one other sensor sequence related to the opening and closing of the breather valve; wherein a portion of the sensor sequence is used for determining the time boundary of a single opening and closing event, and another portion of the sensor sequence can simultaneously serve as model input features to participate in the generation of the opening and closing status representation and fault diagnosis.

[0007] As a preferred embodiment of the intelligent monitoring and fault diagnosis method based on the tank venting device described in this invention, the following steps are taken: the operating status time-series data is extracted according to the start and end times of a single opening and closing event; the start and end times are determined by at least one of the changes in airflow and valve position height, so that the extracted tank internal pressure sequence and tank internal temperature sequence correspond to the actual opening and reseating process of the breather valve; the tank internal pressure sequence and tank internal temperature sequence in the operating status time-series data are standardized to form a normalized time-series input; the normalized time-series input retains the time correspondence between pressure change and temperature change in the same opening and closing event, forming an opening and closing status representation that distinguishes between normal opening and closing states and faulty abnormal states.

[0008] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the normalized timing input is adapted to a preset unified timing length; normalized timing inputs shorter than the preset unified timing length are padded with backward padding; normalized timing inputs longer than the preset unified timing length are truncated from the starting position, and a binary mask matrix is ​​generated based on the effective timing length before adaptation.

[0009] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the binary mask matrix maintains a time-position correspondence with the normalized timing input after length adaptation; when forming the opening and closing state representation, the valid marker positions in the binary mask matrix participate in feature aggregation, while the filling marker positions are excluded from feature aggregation.

[0010] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the following steps are performed: The pressure sequence and valve position height signal of a single breather valve opening / closing event are pre-screened and classified using physical rules; the physical rules include: when the pressure sequence exceeds a preset threshold, but the valve position height signal is still in the closed state, it is determined that the valve disc is faulty, used to determine whether it is a valve disc jamming fault or a valve disc overload fault; when the pressure sequence is below a preset threshold, but the valve position signal is close to the open state, it is determined that the valve guide rod is misaligned and the valve disc cannot fall back normally; the event samples pre-screened and classified by physical rules are input into the intelligent breather valve fault diagnosis model for fine classification.

[0011] As a preferred embodiment of the intelligent monitoring and fault diagnosis method based on the tank venting device described in this invention, the opening and closing state representation includes local waveform state representation and long-range temporal state representation; the local waveform state representation is formed by a one-dimensional convolutional network based on the normalized temporal input; the long-range temporal state representation is formed by a long short-term memory network based on the local waveform state representation.

[0012] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the long-range time-series state representation is transformed into an event-level state representation through mask-aware weighted global pooling; the mask-aware weighted global pooling excludes filling time steps according to the binary mask matrix and assigns weights that increase with time to the effective time steps, so that the pressure convergence state or pressure decay state in the later stage of a single opening and closing event participates in the event-level state representation.

[0013] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the fault diagnosis result includes a fault probability value; the fault probability value is obtained by outputting the event-level state representation through a fully connected network and transforming it using a Sigmoid function; when the fault probability value is greater than or equal to a preset probability threshold, the corresponding single opening / closing event is determined to be a fault abnormal state; when the fault probability value is less than the preset probability threshold, the corresponding single opening / closing event is determined to be a normal opening / closing state.

[0014] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the abnormal fault state includes valve disc sticking fault; the opening and closing process state characterization corresponding to the valve disc sticking fault includes pressure abnormal overshoot state; the pressure abnormal overshoot state is manifested as the pressure peak value during the breather valve opening stage being significantly higher than that during the normal opening and closing state.

[0015] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the storage tank venting device described in this invention, the abnormal fault state includes a sealing failure fault; the opening and closing process state characterization corresponding to the sealing failure fault includes a pressure holding abnormal state; the pressure holding abnormal state is manifested as the breather valve opening stage being close to the normal opening and closing state, while the internal pressure of the storage tank continues to drop after the feeding is completed.

[0016] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the fault abnormal state also includes an overpressure opening fault; the opening and closing process state characterization corresponding to the overpressure opening fault includes an opening pressure deviation state; the opening pressure deviation state is manifested as the deviation of the breather valve opening pressure peak and the opening and closing oscillation state from the normal opening and closing state.

[0017] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the normal opening and closing state includes the normal opening and closing state under the condition of slight valve disc mass drift; when the valve disc attached mass does not exceed the preset slight drift range, and the corresponding opening and closing process state characterization is consistent with the normal opening and closing state, the corresponding single opening and closing event is classified as the normal opening and closing state.

[0018] As a preferred embodiment of the intelligent monitoring fault diagnosis method based on the tank venting device described in this invention, the fault diagnosis result further includes a fault level; the fault level is determined by a preset probability interval in which the fault probability value is located, so that the same fault abnormal state forms a graded diagnosis result according to the probability magnitude; the low probability interval in the preset probability interval is defined as the first level, the medium probability interval is defined as the second level, and the high probability interval is defined as the third level. The fault probability value of each opening and closing event is obtained by the event-level state representation through a fully connected network and the output of the Sigmoid function, representing the probability that the current opening and closing event belongs to a fault abnormal state.

[0019] As a preferred embodiment of the intelligent monitoring and fault diagnosis method based on a storage tank venting device described in this invention, the fault probability value is expressed as: , in, This represents the binary cross-entropy loss value, used to measure the difference between the fault probability output by the intelligent breathing valve fault diagnosis model and the true state label. This represents the total number of samples participating in a single loss calculation. [0,1], Indicates the first The true label of each sample This indicates that the fault diagnosis model for the intelligent breathing valve is applicable to the first... The logit value output by each sample. This represents the Sigmoid function.

[0020] Another objective of this invention is to provide an intelligent monitoring and fault diagnosis system based on a storage tank venting device. This system can correlate and process the pressure and temperature time-series data in a single opening and closing event of a breather valve through a data acquisition module, a status characterization module, and a fault diagnosis module, and generate corresponding opening and closing process status characterization and fault diagnosis results. This solves the problems of current breather valve fault diagnosis technology, which relies on manual experience or single-point threshold alarms, has difficulty in characterizing the complete opening and closing process, and has low accuracy in distinguishing between normal opening and closing fluctuations and abnormal fault states.

[0021] As a preferred embodiment of the intelligent monitoring and fault diagnosis system based on a storage tank venting device according to the present invention, the system includes: a data acquisition module, a rule determination module, a state representation module, and a fault diagnosis module; the data acquisition module is used to determine the start and end times of a single opening and closing event of the breather valve using at least one of the inlet flow signal and the valve position height signal, and to extract the internal pressure sequence and internal temperature sequence of the storage tank within the start and end times to form the operating state time series data corresponding to the single opening and closing event; the rule determination module is used to perform preliminary classification of the single opening and closing event according to preset physical rules, directly identify obvious abnormal events, and determine the event handling method; the state representation module is used to standardize and unify the time series length of the operating state time series data, generate a binary mask matrix, and form an opening and closing process state representation through a one-dimensional convolutional network, a long short-term memory network, and mask-aware weighted global pooling; the fault diagnosis module is used to generate a fault probability value based on the opening and closing process state representation, and to determine the normal opening and closing state, the fault abnormal state, or the fault level according to the fault probability value and its corresponding opening and closing process state representation.

[0022] Another object of the present invention is to provide an intelligent monitoring and fault diagnosis device based on a storage tank venting device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the intelligent monitoring and fault diagnosis method based on the storage tank venting device.

[0023] Another object of the present invention is to provide an intelligent monitoring fault diagnosis storage medium based on a tank venting device, wherein a computer program is stored thereon, and when the computer program is executed by a processor, the steps of an intelligent monitoring fault diagnosis method based on a tank venting device are implemented.

[0024] The beneficial effects of this invention are as follows: The intelligent monitoring fault diagnosis method based on the tank venting device provided by this invention transforms the pressure and temperature time-series data in a single opening and closing event of the breather valve into a representation of the opening and closing process state, changing the diagnostic object from a single point of pressure exceeding the limit to a judgment of the complete opening and closing process, thereby reducing the risk of normal operating condition fluctuations being misjudged as faults. By adapting the length of opening and closing events of different durations and combining them with mask perception feature aggregation, the effective opening and closing time period is included in the diagnostic calculation, which can improve the stability of the diagnostic results under variable-length opening and closing processes. By using local waveform states and long-range time-series states to generate fault probabilities, abnormal states such as valve disc jamming, sealing failure, and overpressure opening can be presented with quantitative results, which can improve the precision of fault identification and maintenance orientation. This invention achieves better results in distinguishing between normal opening and closing states and abnormal fault states, adapting to variable-length opening and closing events, and probabilistic fault diagnosis. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is an overall flowchart of an intelligent monitoring and fault diagnosis method based on a storage tank venting device provided in Embodiment 1 of the present invention.

[0027] Figure 2 This is a flowchart illustrating the diagnostic process of an intelligent breathing valve fault diagnosis model based on an intelligent monitoring fault diagnosis method for a storage tank venting device, as provided in Embodiment 1 of the present invention.

[0028] Figure 3 This is a simulation configuration diagram of a sealing failure fault for an intelligent monitoring and fault diagnosis method based on a storage tank venting device, provided in Embodiment 2 of the present invention.

[0029] Figure 4 This is a simulated configuration diagram of valve jamming fault in an intelligent monitoring fault diagnosis method based on a storage tank venting device provided in Embodiment 2 of the present invention.

[0030] Figure 5 This is a simulation configuration diagram of an overpressure start-up fault for an intelligent monitoring fault diagnosis method based on a storage tank venting device, provided in Embodiment 2 of the present invention.

[0031] Figure 6 Pressure curves of normal operating conditions and fault samples for an intelligent monitoring and fault diagnosis method based on a storage tank venting device provided in Embodiment 2 of the present invention.

[0032] Figure 7 The loss curve and accuracy curve are shown in Embodiment 2 of the present invention for a smart monitoring fault diagnosis method based on a storage tank venting device. Detailed Implementation

[0033] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0034] Example 1, referring to Figures 1-2 This invention provides an intelligent monitoring and fault diagnosis method based on a tank venting device, applicable to a tank system equipped with a breather valve. The tank system includes a tank, a breather valve, a pressure sensor, a temperature sensor, an air inlet flow detection component, a valve position height detection component, and a processing device.

[0035] Pressure sensors are used to collect the internal pressure of the storage tank, temperature sensors are used to collect the internal temperature of the storage tank, air intake flow detection components are used to reflect the air intake or material feeding process of the storage tank, and valve position height detection components are used to reflect whether the breather valve disc has opened and reseated.

[0036] The processing device is used to receive the above data and perform single opening / closing event identification, opening / closing status characterization generation, and fault diagnosis result output.

[0037] In this embodiment, the intelligent breather valve refers to a tank breather valve that can open and reseat in response to changes in the internal pressure of the tank; a single opening and closing event refers to a complete action process in which the breather valve disc rises from the reseat state, opens, releases pressure, and then returns to the reseat state; the operating status time sequence data refers to the tank internal pressure data and tank internal temperature data arranged in the order of sampling time during this complete action process; the opening and closing status representation refers to the data expression formed based on the pressure data and temperature data that can represent the dynamic state change of the breather valve from opening to reseat; the fault diagnosis result refers to the normal opening and closing status, fault abnormal status, fault probability value, fault type, or fault level output based on the opening and closing status representation.

[0038] S1: Obtain the operating status timing data 100 corresponding to a single opening and closing event of the breathing valve.

[0039] The purpose of this step is to determine the complete opening and closing process of a breather valve from the continuous operation data of the storage tank, and to extract the pressure and temperature data corresponding to the opening and closing process. This is to ensure that the subsequent diagnostic object is not an arbitrary segment in the continuous operation curve, nor a certain instantaneous pressure point, but the complete event segment from the opening to the reseating of the breather valve.

[0040] When the processing device continuously receives tank internal pressure data from the pressure sensor, tank internal temperature data from the temperature sensor, air flow change data from the air flow detection component, and valve height change data from the valve height detection component, the processing device first determines whether the breather valve has started to open based on the valve height change. When the valve height rises from the reseating position to the open position and remains stable, the corresponding sampling time is determined as the start time of a single opening and closing event; when the valve height falls back to the reseating position and remains stable, the corresponding sampling time is determined as the end time of a single opening and closing event.

[0041] If there are short-term fluctuations in the valve position height signal, the processing device further corrects it by combining the changes in the intake flow rate. For example, it confirms that the start time is valid when the intake flow rate is in the intake process, and confirms that the end time is valid when the intake flow rate ends and the pressure enters the reseating stage.

[0042] After determining the start and end times, the processing device extracts the internal pressure data of the tank within the specified time period from the continuous pressure data, extracts the internal temperature data of the tank within the same time period from the continuous temperature data, and aligns them according to the sampling time sequence.

[0043] In practical systems, in addition to pressure and temperature sequences, intake flow rate, valve position height, or other relevant signals can be used as model input features to improve the accuracy of fault type identification. However, these additional signals can also be used only to determine the start and end times of a single opening and closing event. Regardless of the method, the opening and closing state representation is generated in the model.

[0044] In this step, the intake flow rate signal and valve position height signal are only used as the basis for determining the time period boundary of a single opening and closing event, and are not used as input data for the subsequent opening and closing state characterization 101. The intake flow rate signal and valve position height signal are only used to determine when the pressure sequence Y and temperature sequence W start to be intercepted and when the interception ends. The data used to form the opening and closing state characterization 101 are only the internal pressure sequence Y and internal temperature sequence W of the storage tank.

[0045] The start and end times of a single opening and closing event can be determined solely based on changes in valve position height, solely based on changes in air intake flow, or jointly based on changes in both valve position height and air intake flow. When using changes in valve position height to determine the boundary, the start time can be the moment when the valve disc rises above the preset opening height and stabilizes, and the end time can be the moment when the valve disc falls below the preset reseating height and stabilizes. When using changes in air intake flow to determine the boundary, the start time can be the moment when the air intake flow enters the stable air intake process, and the end time can be the moment when the air intake flow ends and the internal pressure of the storage tank enters the reseating change phase.

[0046] In a specific implementation process, the processing device first reads the valve position height detection result. When the valve disc rises from the reseating position and remains continuously above the opening height, the processing device records this moment as the start of the current opening and closing event. When the valve disc falls back to below the reseating height and remains continuously stable, the processing device records this moment as the end of the current opening and closing event. The processing device only extracts the pressure and temperature data within this time period and uses the extracted results as the operating status timing data 100 for this opening and closing event.

[0047] S2: Generate an opening / closing state representation 101 based on the operating state timing data 100. The opening / closing state representation 101 represents the dynamic state change of the breathing valve from opening to reseating in a single opening / closing event.

[0048] The purpose of this step is to convert the pressure and temperature time series data obtained in S1 into an opening and closing state characterization 101 that can be used for diagnosis. This characterization describes the dynamic state changes of the breathing valve from opening to reseating during a complete opening and closing event, including both the local pressure changes during the opening phase and the pressure maintenance or pressure decay changes after reseating.

[0049] The processing device first standardizes the pressure sequence Y and temperature sequence W obtained in step one, so that the pressure data and temperature data are on a unified numerical scale, forming a normalized time series input. Here, the normalized time series input specifically refers to the pressure sequence Y and temperature sequence W after standardization, which still retain the original sampling time order, and the time correspondence between pressure change and temperature change in the same opening and closing event does not change.

[0050] Since the durations of different opening and closing events are not consistent, the processing device further adapts the normalized timing input to a certain length. The preset unified timing length refers to the fixed sequence length received by the intelligent breathing valve fault diagnosis model Q, which can be determined based on the effective length distribution of opening and closing events in historical samples. If the normalized timing input of a certain opening and closing event is shorter than the preset unified timing length, the processing device fills the end of the sequence with backfill. If the normalized timing input of a certain opening and closing event is longer than the preset unified timing length, the processing device truncates the sequence from the beginning position to ensure that the length of the truncated sequence meets the preset unified timing length.

[0051] While performing length adaptation, the processing device generates a binary mask matrix based on the effective timing length before adaptation. Specifically, the binary mask matrix refers to the marked data that maintains a time position correspondence with the normalized timing input after length adaptation. The effective sampling position is marked as the effective position, and the position formed by backward padding is marked as the padding position. In the subsequent feature aggregation process, the effective marked position participates in feature aggregation, while the padding marked position is excluded from feature aggregation. The padding data will not be mistaken for part of the actual opening and closing process of the breathing valve.

[0052] Building a CNN-LSTM network: Input layer: Accepts tensors of shape (batch, seq_len, 2) (2 channels: pressure, temperature) and passes a mask matrix (batch, seq_len).

[0053] CNN Feature Extraction Module: First layer: 32 filters, convolution kernel size 5, followed by BatchNorm, ReLU, Dropout, and a max pooling layer with a stride of 2 (kernel size 2).

[0054] The second layer consists of 64 filters, a kernel size of 3, followed by BatchNorm, ReLU, and Dropout, with no pooling.

[0055] LSTM module: Receives temporal features from the CNN output and captures long-range dependencies.

[0056] Mask-aware global pooling: After the LSTM output, a new mask is generated based on the original effective length (the length will be shortened after CNN pooling), and the LSTM output is time-weighted averaged (the weights increase linearly with time, emphasizing later information) to ensure that the filled region does not participate in the calculation.

[0057] Fully connected layer and output layer: The first layer is 64-dimensional (ReLU+Dropout), the second layer is 16-dimensional (ReLU+Dropout), and finally a single output neuron is connected to output logit, which is converted into fault probability by Sigmoid.

[0058] Before generating an opening and closing state representation for a single breathing valve opening and closing event, the pressure sequence and valve position signal are first pre-screened and classified using physical rules.

[0059] Specifically, when the pressure sequence exceeds the preset threshold, but the valve position signal is still in the closed state, the system determines that there is an abnormality in the valve disc, which is used to distinguish whether the valve disc is stuck or the valve disc is overloaded. When the pressure sequence is below the preset threshold, but the valve position signal is close to the open state, the system determines that the valve guide rod may be tilted, making the valve disc unable to fall back normally.

[0060] After the samples are pre-screened and classified according to physical rules, they are then input into the intelligent breathing valve fault diagnosis model Q. The model uses event-level opening and closing state representation to extract local waveform features and long-term time-series dependent features, and generates corresponding fault types and fault probability values.

[0061] For the opening and closing state characterization of a single opening and closing event 101, it is determined that the internal pressure of the storage tank is lower than 7 mbar and the valve position height signal shows that the valve disc is still in the open state. Based on the above judgment conditions, this event is directly identified as a guide rod misalignment fault and is not entered into the CNN+LSTM model for subsequent classification. This event judgment indicates that the valve disc should return to its seat and close under the action of gravity, but it cannot fall back normally due to the misalignment or jamming of the guide rod, thus triggering the direct physical rule judgment.

[0062] The opening and closing status characterization of a single opening and closing event is 101. It is determined that the internal pressure of the storage tank is greater than or equal to 10.5 mbar and the valve position height signal shows that the valve disc is still in the closed state. Based on the above judgment conditions, this event is fed into the CNN+LSTM model for fine classification to distinguish between valve disc sticking fault and valve disc overload fault. This event judgment indicates that the storage tank pressure has exceeded the design start pressure, but the valve disc has not opened. It may be due to the valve disc sticking to the valve seat or the valve disc having excessive additional mass, which leads to an increase in opening pressure. A deep learning model is needed for further fine judgment.

[0063] The opening and closing state characterization of a single opening and closing event is 101. It is determined that the internal pressure of the storage tank reaches the preset start-up pressure threshold and the valve position height signal shows that the valve disc has been opened normally. Based on the above judgment conditions, the event is directly determined as a normal opening and closing state and is not entered into the CNN+LSTM model for subsequent classification. This event judgment indicates that the breathing valve is raised and depressurized normally according to the design requirements, and the opening and closing behavior is in line with the expected operation, without the need for deep learning model intervention.

[0064] A fault diagnosis model Q for an intelligent breathing valve, based on a pressure sequence Y and a temperature sequence W, is trained using a binary cross-entropy loss function. The intelligent breathing valve fault diagnosis model Q outputs a fault probability value 201 based on the opening / closing state representation 101, expressed as: , in, This represents the binary cross-entropy loss value, used to measure the difference between the fault probability output by the Q-classification fault diagnosis model for intelligent breathing valves and the true state label. This represents the total number of samples participating in a single loss calculation. [0,1], Indicates the first The true label of each sample The fault diagnosis model Q of the intelligent breathing valve represents the first... The logit value output by each sample. Representing the Sigmoid function: , in, This represents the probability value of a fault, that is, the probability that a single start-up or shutdown event belongs to a faulty or abnormal state. This represents a certain logit value output by the fault diagnosis model Q of the intelligent breathing valve.

[0065] The intelligent breathing valve fault diagnosis model Q outputs a fault probability value 201 for each sample. The judgment conditions are as follows: if the fault probability value is ≥0.5, it is judged as "fault" (label 1); if the fault probability value is <0.5, it is judged as "normal" (label 0).

[0066] Subsequently, the processing device feeds the normalized time-series input with the adapted length into a one-dimensional convolutional network. The one-dimensional convolutional network extracts local features of the pressure channel and temperature channel along the time direction to form a local waveform state representation 101a. The local waveform state representation 101a specifically refers to features that can reflect changes in pressure peak value, local pressure overshoot, local oscillation changes, and temperature response changes. For faults such as valve disc jamming, an abnormal increase in pressure may occur during the opening phase, and this abnormal change can be preserved through the local waveform state representation 101a.

[0067] The processing device further inputs the local waveform state representation 101a into the long short-term memory network. The long short-term memory network extracts the preceding and following state relationships between the opening, depressurization, reseating, and pressure changes after reseating of the breathing valve in chronological order, forming a long-term time-series state representation 101b. The long-term time-series state representation 101b specifically refers to the characteristics that can reflect the succession relationship between preceding and following states in a single opening and closing event of the breathing valve, especially used to express the situation where the opening phase is close to normal but the pressure continues to drop after reseating.

[0068] After the long-term temporal state representation 101b is formed, the processing device performs mask-aware weighted global pooling. Mask-aware weighted global pooling specifically refers to reading the binary mask matrix during the pooling process, excluding the padding time steps, and only weighting and pooling the real and valid time steps. Furthermore, it assigns higher weights to valid time steps that are closer to the later part of the opening and closing event. The pressure convergence state or pressure decay state after the breathing valve reseats can enter the final event-level state representation. The event-level state representation specifically refers to the overall state expression formed by pooling the complete temporal information of a single opening and closing event, which is part of the opening and closing state representation 101.

[0069] The local waveform state representation 101a can be formed by a one-dimensional convolutional network or by other networks capable of extracting features from local temporal segments; the long-range temporal state representation 101b can be formed by a long short-term memory network or by other temporal networks capable of maintaining temporal dependencies; the mask marking can be achieved by marking valid and invalid positions or by using valid and invalid weights to achieve the same function.

[0070] Normalized pressure and temperature data first enter a two-layer one-dimensional convolutional network. The first convolutional network extracts local changes in pressure and temperature over a relatively wide time range, while the second convolutional network further extracts more fine-grained waveform changes. Subsequently, the convolutional output enters a long short-term memory network, which outputs the temporal state of each time step. Finally, the processing device excludes padding time steps based on a binary mask matrix and performs time-weighted aggregation on the effective time steps to obtain an event-level state representation.

[0071] S3: Based on the opening and closing state characterization 101, distinguish between the normal opening and closing state and the fault abnormal state of the breathing valve, and generate a fault diagnosis result 200 based on the dynamic state change corresponding to the fault abnormal state.

[0072] The purpose of this step is to convert the opening and closing state representation 101 formed by S2 into an output diagnostic conclusion. The fault diagnosis result 200 specifically includes the fault probability value 201, the judgment result of normal opening and closing state or fault abnormal state, the fault type corresponding to the fault abnormal state, and the fault level 202.

[0073] Fault level 202 is determined by the preset probability interval where the fault probability value 201 is located, so that the same fault abnormal state forms a graded diagnosis result according to the probability.

[0074] The low probability range in the preset probability interval is defined as the first level, the medium probability range as the second level, and the high probability range as the third level. The fault probability value 201 of each start-up and shutdown event is obtained by the event-level state representation through a fully connected network and the output of the Sigmoid function, representing the probability that the current start-up and shutdown event belongs to a fault abnormal state.

[0075] In specific implementation, the processing device inputs the event-level state representation into the fully connected network, the fully connected network outputs diagnostic values, and converts them into a fault probability value 201 through a probability transformation function. The fault probability value 201 specifically refers to the probability that a single opening / closing event belongs to a faulty or abnormal state. The processing device compares the fault probability value 201 with a preset probability threshold. The preset probability threshold specifically refers to the probability judgment benchmark used to distinguish between normal opening / closing states and faulty or abnormal states. In one specific implementation, the preset probability threshold is set to one-half. When the fault probability value 201 is greater than or equal to the threshold, the corresponding single opening / closing event is determined to be a faulty or abnormal state; when the fault probability value 201 is less than the threshold, the corresponding single opening / closing event is determined to be a normal opening / closing state.

[0076] Once a single opening / closing event is identified as a faulty or abnormal state, the processing device further determines the fault type based on the corresponding dynamic state changes in the opening / closing state characterization 101. The dynamic state changes specifically refer to temporal changes that reflect the mechanical state of the breather valve, such as changes in pressure peak value, pressure holding, pressure decay, and opening / closing oscillations.

[0077] If the opening / closing status characterization 101 includes a pressure abnormal overshoot state, the processing device will determine the fault abnormal status as a valve disc jamming fault. The pressure abnormal overshoot state specifically refers to the pressure peak during the opening phase of the breathing valve being significantly higher than the normal opening / closing state, manifested as excessive pressure accumulation before the valve disc opens or a rapid pressure surge at the moment of opening. This state corresponds to the situation where the valve disc and valve seat are obstructed due to adhesion, deposition, or attachment.

[0078] If the opening / closing status characterization 101 includes an abnormal pressure holding status, the processing device will determine the abnormal fault status as a sealing failure fault. The abnormal pressure holding status specifically refers to the fact that the curve of the breather valve during the opening phase is close to the normal opening / closing state, but the internal pressure of the storage tank continues to drop after the feeding is completed, which is manifested as the pressure not being able to be maintained stably after reseating. This status corresponds to the situation where the breather valve sealing surface is damaged or the sealing capacity is reduced.

[0079] If the opening and closing status characterization 101 includes an opening pressure deviation status, the processing device will determine the fault abnormal status as an overpressure opening fault. The opening pressure deviation status specifically refers to the deviation of the peak opening pressure of the breather valve and the opening and closing oscillation status from the normal opening and closing status, which is manifested as an increase in opening pressure, a change in oscillation amplitude, or a change in oscillation frequency. This status corresponds to a situation where the valve disc has a large additional mass or a change in mass distribution, resulting in an opening pressure deviation.

[0080] In this embodiment, the normal opening and closing state also includes the normal opening and closing state under the state of slight valve disc mass drift. The state of slight valve disc mass drift specifically refers to the small change in mass of the valve disc due to slight fouling, condensate adhesion, or coating thickening, but this change does not cause the opening and closing state characterization 101 to deviate from the normal opening and closing state. When the mass of the valve disc does not exceed the preset slight drift range, and the corresponding opening and closing state characterization 101 does not show an abnormal pressure overshoot state, an abnormal pressure holding state, or an opening pressure deviation state, the single opening and closing event is classified as the normal opening and closing state. The sample of slight mass disturbance with an additional amount not exceeding five percent of the valve seat mass is defined as the normal operating condition sample.

[0081] Furthermore, the processing device determines the fault level 202 based on the preset probability range in which the fault probability value 201 falls. The preset probability range specifically refers to multiple probability ranges set according to on-site maintenance requirements. The fault level 202 specifically refers to the graded diagnosis result formed by the fault abnormal state according to the magnitude of the fault probability value 201. When the fault probability value 201 is in a lower abnormal probability range, the processing device generates a lower level; when the fault probability value 201 is in a medium abnormal probability range, the processing device generates an intermediate level; when the fault probability value 201 is in a higher abnormal probability range, the processing device generates a higher level.

[0082] The fault probability value 201 can be obtained through a probability transformation function or through an equivalent output layer that can output the probability of anomalies; the fault level 202 can be formed based on two or more preset probability intervals; the fault abnormal state can include valve disc jamming fault, sealing failure fault and overpressure opening fault, and other opening and closing abnormal types can be added based on subsequent sample expansion.

[0083] After processing by S1 and S2, an event-level state characterization is obtained for a certain opening and closing event. Based on this characterization, the processing device obtains a high fault probability value 201 and determines that the event belongs to a fault abnormal state. If the opening and closing state characterization 101 of the event shows that the pressure peak during the opening phase is significantly higher than that during the normal opening and closing phase, the device outputs a valve disc sticking fault. If the opening and closing state characterization 101 of the event shows that the opening phase is close to normal but the pressure continues to drop after reseating, the device outputs a sealing failure fault. If the opening and closing state characterization 101 of the event shows that the opening pressure peak and the opening and closing oscillation state deviate from the normal opening and closing state, the device outputs an overpressure opening fault. If the valve disc mass corresponding to a certain opening and closing event only experiences a slight drift and its opening and closing state characterization 101 remains consistent with the normal opening and closing state, the processing device outputs a normal opening and closing state.

[0084] Example 2, refer to Figures 3-7 As an embodiment of the present invention, an intelligent monitoring and fault diagnosis method based on a storage tank venting device is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0085] Under normal operating conditions, a 5m³ pilot-scale storage tank and a pilot-operated breather valve were used to collect data. The feeding process was simulated by filling the tank with air using a variable frequency fan. The internal pressure P (mbar) and temperature T (°C) of the tank were collected at a fixed frequency by a PLC control system.

[0086] Normal operating conditions include different initial pressure levels, different intake flow rates, and different opening and closing frequencies.

[0087] In normal samples, the breathing valve disc opens and resets normally according to the designed pressure threshold, without jamming, seal damage, or additional mass. Each batch of normal experiments is repeatedly sampled multiple times. The intake flow rate signal and valve position height signal are only used to assist in defining the start and end times of breathing events and are not used as Q-input features of the intelligent breathing valve fault diagnosis model.

[0088] In addition, to cover real-world scenarios of slight displacement of valve core mass in industrial settings, normal samples under minute mass disturbances were also collected.

[0089] Specifically, a counterweight not exceeding 5% of the valve seat mass is added to the valve disc to simulate the slight increase in mass of the valve disc due to minor fouling, condensate adhesion, or coating thickening during long-term operation. The state of this type of sample is also defined as a normal operating condition sample.

[0090] On the basis of normal operating conditions, a fault is artificially introduced, such as Figure 3 As shown, the sealing failure simulation simulates the situation where damage to the sealing surface leads to a decrease in the valve's pressure-holding capacity. The sealing failure is achieved by applying different viscous media to the sealing surfaces of the valve disc and valve seat. The viscous media can be slightly gelatinous or viscous paste, used to simulate the situation where wear or damage to the sealing surface during long-term operation leads to a decrease in the valve's pressure-holding capacity. In the experiment, the sealing media directly covers the contact surface between the valve disc and valve seat, ensuring that the pressure cannot be fully restored during the opening and reseating process, thus showing the tail decay characteristic of "normal opening and leakage when closing" in the pressure time sequence curve.

[0091] The transparent tape is used as an auxiliary component to fix the valve disc or guide rod, ensuring that the valve disc or guide rod position is stable during the experiment. This prevents external vibrations or slight displacements from affecting the acquisition of the valve disc's opening and closing status, thereby ensuring the accuracy of the pressure and temperature sequence data.

[0092] The valve disc limiting guide rod is used to limit the range of movement of the valve disc during the opening and reseating process, ensuring that the valve disc will not exceed the preset displacement range. At the same time, it ensures the controllability of the valve disc movement during fault simulations such as jamming or sealing failure, so that the experiment can stably reflect the pressure response under different fault conditions.

[0093] Foam rubber is used to simulate the elastic buffering effect between valve discs or valve seats. Especially when simulating sealing failure or valve disc jamming, foam rubber provides slight resistance or contact buffering, making pressure changes more realistic. This allows the model to capture the dynamic characteristics of the valve disc opening, depressurization, and reseating stages, thereby generating reliable data that can be used to train and validate fault diagnosis models.

[0094] like Figure 4 As shown, the valve disc sticking fault simulates the phenomenon of adhesion between the valve disc and the valve seat caused by long-term material deposition (such as heavy oil accumulation and corrosion products). It is achieved by applying different viscous media to the sealing surfaces of the valve disc and the valve seat. The valve disc sticking fault is simulated by applying a substance with adhesive properties between the valve disc and the valve seat. Such substances can be heavy oil deposits, corrosion products, or colloids. It simulates the situation where the valve disc is blocked from opening due to long-term deposition or adhesion. In the experiment, the pressure of the valve disc rises rapidly during the opening stage and a peak overshoot occurs. The pressure difference is significantly higher than that of the normal opening and closing state, which makes the fault have unique characteristics in the pressure time series curve.

[0095] like Figure 5 As shown, the overpressure opening failure is simulated by physically adding calibrated, centrally symmetrical, eccentric counterweights to the valve disc, thereby increasing the mass of the valve disc and the opening pressure. The counterweights are divided into two groups: Centrally symmetrical counterweight: evenly distributed around the center of the valve disc to increase the overall mass of the valve disc and improve the opening pressure.

[0096] Eccentric counterweight: Installed at a position off-center from the valve disc, it is used to change the mass distribution and rotational inertia of the valve core, and further adjust the valve disc opening characteristics.

[0097] To gain a deeper understanding of the differences in the dynamic behavior of the breather valve under different operating conditions and to provide interpretable feature data for the intelligent breather valve fault diagnosis model Q, the collected typical pressure time series curves were visualized and analyzed. Figure 6 The pressure-time series of six typical samples under the same operating conditions are shown, where the green dashed line represents the theoretical opening pressure threshold (12 mbar) of the reference valve disc used.

[0098] The characteristics of each subgraph are analyzed as follows: Figure 6 (a) and Figure 6 (b) shows the typical pressure curve of the normal sample. The main differences between the two are the initial steady-state pressure and the pressure retention characteristics after feeding. Both samples produce breathing oscillations around the 12 mbar threshold during inflation. Due to sensor transmission delay and the initial depressurization rate being lower than the feeding rate, their pressure peaks both exceed 12 mbar.

[0099] Figure 6In the middle (c), the sample with an additional 5% mass is shown, and its curve shape is highly similar to that of the normal sample.

[0100] The above results indicate that whether the instantaneous pressure value exceeds the rated threshold is insufficient to distinguish between normal breathing and fault conditions.

[0101] Figure 6 The middle (d) sample is an overpressure start-up failure sample with an additional 50% mass. The pressure peak is close to 15 mbar, and the oscillation frequency and amplitude are significantly different from the normal sample.

[0102] Figure 6 The middle (e) is a sample of valve disc sticking failure. The pressure rises sharply to an extremely high peak of nearly 30 mbar at about 20 seconds. This extreme overshoot is a unique feature of sticking failure.

[0103] Figure 6 In the middle (f), the failure sample is a sealing failure. Its opening phase is similar to that of a normal valve disc, but after the feeding is completed, the pressure inside the tank continues to drop to atmospheric pressure, showing the tail decay characteristics of "normal opening and leakage when closed". This type of failure is extremely difficult to detect by the threshold method that only monitors the pressure peak during the opening phase.

[0104] The 353 collected samples were used to train and validate the Q-model for the fault diagnosis of the intelligent breathing valve according to the following procedure.

[0105] In one specific embodiment, the following training hyperparameters and network configuration are used: Optimizer: Adam, initial learning rate set to 0.001.

[0106] Batch Size: 32.

[0107] Training epochs: In conjunction with the early stopping strategy, training is terminated if the validation set loss does not decrease for 10 consecutive epochs.

[0108] Learning rate decay: If the validation set loss does not decrease for 5 consecutive rounds, the learning rate decays to 0.5 times its original value.

[0109] Dropout ratio: 0.3 for both CNN layers and fully connected layers, used to suppress overfitting.

[0110] Loss function: Binary Cross Entropy with Logits Loss.

[0111] The first layer of the CNN consists of 32 one-dimensional filters, a kernel size of 5, a stride of 1, and zero padding to preserve the sequence length. It is followed by BatchNorm, ReLU activation, and Dropout (0.3). Then, a max pooling layer is connected with a kernel size of 2 and a stride of 2 to halve the sequence length.

[0112] The second layer of the CNN consists of 64 one-dimensional filters, a kernel size of 3, a stride of 1, and zero padding; followed by BatchNorm, ReLU activation, and Dropout (0.3); without pooling.

[0113] LSTM module: Input dimension is 64 (i.e. the number of output channels of the second layer of CNN), hidden layer dimension is 128, number of layers is 1, and it receives the temporal feature tensor output by CNN.

[0114] Mask-aware weighted global pooling: The mask is updated according to the effective length of the CNN pooling, and the LSTM output is weighted and averaged along the time axis. The weights increase linearly with time to emphasize the convergence and decay characteristics in the later stage of the respiratory event. The mask ensures that only the effective time steps participate in the summation and averaging.

[0115] Fully connected layers: the first layer is 64-dimensional (ReLU+Dropout(0.3)), the second layer is 16-dimensional (ReLU+Dropout(0.3)), and the output layer is 1-dimensional (logit).

[0116] Finally, to objectively evaluate the generalization performance of the intelligent breathing valve fault diagnosis model Q and avoid bias caused by random partitioning, 10-Fold Cross Validation was adopted. All samples were randomly shuffled and divided into 10 subsets. Two subsets were selected as the test set and the remaining eight subsets were selected as the training set. This process was repeated 10 times to ensure that each sample was tested once. See the technical validation results for details.

[0117] Based on the aforementioned pilot-scale platform, 232 fault-type samples and 121 non-fault-type samples were collected. 80% of the data was used to train the intelligent breathing valve fault diagnosis model Q, and 20% of the model was used for validation. The average performance of the intelligent breathing valve fault diagnosis model Q was measured through ten cross-validations.

[0118] Figure 7 The loss curve and accuracy curve of a certain model training and validation process are shown, and the results show that the intelligent breathing valve fault diagnosis model Q did not overfit.

[0119] The diagnostic capability of the classification model is evaluated using three key metrics: F1 score, accuracy (Acc), and wrong predictions*. These metrics reflect, from different perspectives, the ability of the intelligent breathing valve fault diagnosis model Q to correctly distinguish between the normal operating state (Category 0) and the fault state (Category 1) of a pilot-operated breathing valve (PVRV).

[0120] Table 1 presents the statistical metrics obtained by the classification model in ten independent training and inference runs. The model demonstrated consistently high diagnostic accuracy across all runs, with an average accuracy of 0.97 ± 0.02. In terms of category-specific performance, the F1 scores for normal valves (Category 0) and faulty valves (Category 1) were 0.96 ± 0.03 and 0.98 ± 0.03, respectively.

[0121] In Table 1, Round represents the evaluation result round of the model in each round of training or cross-validation, and Mean represents the average and standard deviation of the evaluation metrics for all Round rounds.

[0122] Table 1 Results of the ten-time cross-validation model

[0123] Example 3, an embodiment of the present invention, provides an intelligent monitoring and fault diagnosis system based on a storage tank venting device, including a data acquisition module, a rule determination module, a status characterization module, and a fault diagnosis module.

[0124] The data acquisition module is used to determine the start and end times of a single opening and closing event of the breather valve using at least one of the intake flow signal and the valve position height signal, and to extract the internal pressure sequence Y and the internal temperature sequence W of the storage tank within the start and end times to form the operating status timing data 100 corresponding to the single opening and closing event.

[0125] The rule determination module is used to perform preliminary classification of single opening and closing events according to preset physical rules, directly identify obvious abnormal events, and determine the event handling method.

[0126] The state representation module is used to standardize and unify the time length of the running state time series data 100, generate a binary mask matrix, and form the start-up and closing process state representation through a one-dimensional convolutional network, a long short-term memory network, and mask-aware weighted global pooling.

[0127] The fault diagnosis module is used to generate a fault probability value 201 based on the state representation of the opening and closing process, and to determine the normal opening and closing state, the fault abnormal state, or the fault level 202 according to the fault probability value 201 and its corresponding state representation of the opening and closing process.

[0128] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements an intelligent monitoring and fault diagnosis method based on a storage tank venting device as proposed in the above embodiment.

[0129] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements an intelligent monitoring and fault diagnosis method based on a storage tank venting device as proposed in the above embodiment.

[0130] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0132] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0133] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0134] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent monitoring and fault diagnosis based on a storage tank venting device, characterized in that, include: Obtain the operating status timing data (100) corresponding to a single opening and closing event of the breathing valve. An opening and closing state representation (101) is generated based on the operating state timing data (100). The opening and closing state representation (101) represents the dynamic state change of the breathing valve from opening to reseating in a single opening and closing event. Based on the opening and closing state characterization (101), the normal opening and closing state and the fault abnormal state of the breathing valve are distinguished, and the fault diagnosis result (200) is generated according to the dynamic state change corresponding to the fault abnormal state. Physical rules are used to pre-screen and classify the pressure sequence and valve position height signal of a single breathing valve opening and closing event. The physical rules include determining that the valve disc is faulty when the pressure sequence exceeds a preset threshold but the valve position height signal is still in the closed state, which is used to determine whether it is a valve disc jamming fault or a valve disc overload fault. When the pressure sequence is below the preset threshold, but the valve position signal is close to the open state, it is determined that the valve guide rod is skewed and the valve disc cannot fall back normally. Event samples that have been pre-screened and classified according to physical rules are input into the intelligent breathing valve fault diagnosis model Q for fine classification; The operating status timing data (100) includes the tank internal pressure sequence Y, the tank internal temperature sequence W, and at least one other sensor sequence related to the opening and closing of the breather valve; One part of the sensor sequence is used for the time period boundary determination of a single opening and closing event, while the other part of the sensor sequence can be used as model input features to participate in the generation of opening and closing state representation and fault diagnosis. The running status timing data (100) is extracted according to the start and end times of a single start-stop event; The start and end times are determined by at least one of the changes in air intake flow rate and valve position height, so that the extracted internal pressure sequence Y and internal temperature sequence W of the storage tank correspond to the actual opening and reseating process of the breather valve. The internal pressure sequence Y and internal temperature sequence W of the storage tank in the operating status time series data (100) are standardized to form a normalized time series input; The normalized time-series input retains the time correspondence between pressure change and temperature change in the same opening and closing event, forming an opening and closing state characterization that distinguishes between normal opening and closing state and fault abnormal state (101). The fault diagnosis result (200) includes the fault probability value (201); The fault probability value (201) is expressed as: ， in, This represents the binary cross-entropy loss value, used to measure the difference between the fault probability output by the Q-classification fault diagnosis model for intelligent breathing valves and the true state label. This represents the total number of samples participating in a single loss calculation. [0,1], Indicates the first The true label of each sample The fault diagnosis model Q of the intelligent breathing valve represents the first... The logit value output by each sample. Represents the Sigmoid function; ， in, This represents the probability value of a fault, that is, the probability that a single start-up or shutdown event belongs to a faulty or abnormal state. This represents the logit value output by the fault diagnosis model Q of the intelligent breathing valve; The on / off state representation (101) includes local waveform state representation (101a) and long-range time sequence state representation (101b). The local waveform state representation (101a) is formed by a one-dimensional convolutional network based on the normalized temporal input; The long-range temporal state representation (101b) is formed by a long short-term memory network based on the local waveform state representation (101a); The long-term temporal state representation (101b) is transformed into an event-level state representation through mask-aware weighted global pooling. The mask-aware weighted global pooling excludes filling time steps based on the binary mask matrix and assigns weights to effective time steps that increase over time, so that the pressure convergence state or pressure decay state in the later stage of a single start-up or shutdown event participates in the event-level state representation.

2. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1, characterized in that: The normalized timing input is length-adapted according to a preset uniform timing length. Normalized timing inputs shorter than the preset uniform timing length are padded with backward padding. The normalized timing input, which is longer than the preset uniform timing length, is truncated from the starting position, and a binary mask matrix is ​​generated based on the effective timing length before adaptation.

3. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 2, characterized in that: The binary mask matrix and the normalized timing input after length adaptation maintain the time position correspondence; When forming the open / closed state representation (101), the valid marker positions in the binary mask matrix participate in feature aggregation, while the filler marker positions are excluded from feature aggregation.

4. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1, characterized in that: Physical rules also include, When the internal pressure of the storage tank is lower than the normal reseating level and the valve position signal shows that the valve disc is still in the open state, the current event is directly determined to be a guide rod misalignment fault, without needing to enter the intelligent breathing valve fault diagnosis model Q. When the tank pressure exceeds the design start-up range and the valve position signal shows that the valve disc is not open, the current event is sent to the intelligent breathing valve fault diagnosis model Q to distinguish between valve disc jamming fault and valve disc overload fault. When the tank pressure reaches the trigger threshold and the valve position signal shows that the valve disc is open normally, the current event is directly determined to be a normal opening and closing state, without the need for intervention from the intelligent breathing valve fault diagnosis model Q.

5. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1, characterized in that: The fault diagnosis result (200) includes the fault probability value (201); The fault probability value (201) is obtained by outputting the event-level state representation through a fully connected network and then transforming it using the Sigmoid function; When the fault probability value (201) is greater than or equal to the preset probability threshold, the corresponding single opening and closing event is determined to be a fault abnormal state. When the fault probability value (201) is less than the preset probability threshold, the corresponding single opening and closing event is determined to be a normal opening and closing state.

6. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1 or 4, characterized in that: The fault / abnormal state includes valve disc sticking fault; The opening and closing process state representation corresponding to the valve disc sticking fault includes the abnormal pressure overshoot state. The abnormal pressure overshoot state is characterized by a significant increase in the pressure peak during the opening phase of the breather valve compared to the normal opening and closing phase.

7. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1 or 5, characterized in that: The fault abnormality includes sealing failure; The state characterization of the opening and closing process corresponding to the sealing failure includes an abnormal pressure holding state. The abnormal pressure condition is characterized by the breather valve being close to normal opening and closing during the opening phase, while the internal pressure of the storage tank continues to drop after the feeding is completed.

8. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1 or 5, characterized in that: The abnormal fault states also include overpressure start-up faults; The state representation of the opening and closing process corresponding to the overpressure opening fault includes the opening pressure offset state. The opening pressure deviation state is manifested as a deviation between the peak opening pressure of the breather valve and the opening and closing oscillation state relative to the normal opening and closing state.

9. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1 or 2, characterized in that: The normal opening and closing state includes the normal opening and closing state under the condition of slight valve disc mass drift. When the valve disc adhesion mass does not exceed the preset micro-drift range, and the corresponding opening and closing process state characterization is consistent with the normal opening and closing state, the corresponding single opening and closing event is classified as the normal opening and closing state.

10. The intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in claim 1 or 5, characterized in that: The fault diagnosis result (200) also includes the fault level (202); The fault level (202) is determined by the preset probability interval where the fault probability value (201) is located, so that the same fault abnormal state forms a graded diagnosis result according to the probability size; The low probability interval in the preset probability interval is defined as the first level, the medium probability interval is defined as the second level, and the high probability interval is defined as the third level. The fault probability value (201) of each start-up and shutdown event is obtained by the event-level state representation through a fully connected network and the output of the Sigmoid function, representing the probability that the current start-up and shutdown event belongs to a fault abnormal state.

11. An intelligent monitoring and fault diagnosis system based on a storage tank venting device, employing the intelligent monitoring and fault diagnosis method based on a storage tank venting device as described in any one of claims 1 to 10, characterized in that: It includes a data acquisition module, a rule determination module, a status representation module, and a fault diagnosis module; The data acquisition module is used to determine the start and end times of a single opening and closing event of the breathing valve using at least one of the intake flow signal and the valve position height signal, and to extract the internal pressure sequence Y and the internal temperature sequence W of the storage tank within the start and end times to form the operating status time sequence data (100) corresponding to the single opening and closing event. The rule determination module is used to perform preliminary classification of single opening and closing events according to preset physical rules, directly identify obvious abnormal events, and determine the event handling method. The state representation module is used to standardize and unify the time length of the running state time series data (100), generate a binary mask matrix, and form the start-up and closing process state representation through a one-dimensional convolutional network, a long short-term memory network and mask-aware weighted global pooling. The fault diagnosis module is used to generate a fault probability value (201) based on the state representation of the opening and closing process, and to determine the normal opening and closing state, fault abnormal state or fault level (202) according to the fault probability value (201) and its corresponding state representation of the opening and closing process.

Images