Pipeline condensate control system based on convolutional neural network

Abstract

The application discloses a pipeline condensate water control system based on a convolutional neural network, comprising a multi-source signal acquisition, a condensation state map construction, a physical constraint multi-task convolutional neural network, an active gas injection excitation identification, a residual water amount closed-loop control and a control execution and feedback module. The system estimates water amount, condensation growth rate and pressure difference required for drainage according to pressure, temperature and humidity, flow, liquid level and acoustic vibration data; when the reliability is insufficient, a controlled pressure pulse is applied to correct the water amount, and the residual water amount is updated according to the actual drainage volume and re-estimated, and segmented drainage and feedback correction are executed under the interlocking of pressure and valve position.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for gas pipeline condensate, and more specifically to a pipeline condensate control system based on convolutional neural networks. Background Technology

[0002] Gas pipeline condensate control technology is a technique for sensing, estimating, and controlling the discharge of liquid water generated during gas transmission due to changes in temperature, pressure, and moisture content. Low-pressure coalbed methane pipelines have small pressure margins, making it difficult to accurately determine the water accumulation status, drainage pressure differential, and residual water volume using a single liquid level or timed drainage.

[0003] CN110715176A discloses a gas condensate collection system that collects and transports condensate through a transfer tank, a plunger pump, and a condensate delivery pipeline. However, it mainly relies on a fixed mechanical structure to complete the transfer and does not solve the problems of reliable estimation of condensate volume under multi-source operating conditions, active identification of low-reliability results, prediction of pressure difference required for drainage, and closed-loop update of residual water volume after drainage. It is difficult to take into account the drainage adequacy, gas loss, and pressure safety of low-pressure pipelines.

[0004] To address the aforementioned issues, it is necessary to propose a pipeline condensate control system based on convolutional neural networks. Summary of the Invention

[0005] This invention aims to solve the problems in the existing low-pressure gas pipeline condensate control process, such as insufficient reliability of single liquid level signals, difficulty in comprehensively characterizing pipeline operating status, inability to synchronously predict condensate volume and pressure difference required for drainage, lack of independent verification methods for low-reliability estimation, and the tendency for insufficient drainage, excessive drainage, and gas loss to occur when drainage is carried out at fixed times.

[0006] To address the aforementioned technical problems, this invention provides a pipeline condensate control system based on a convolutional neural network, comprising a multi-source signal acquisition module, a timing alignment and condensate state diagram construction module, a physical constraint multi-task convolutional neural network module, an active aeration excitation identification module, a residual water volume closed-loop control module, and a control execution and feedback module.

[0007] The multi-source signal acquisition module is used to acquire pressure, temperature, humidity, flow rate, liquid level, acoustic, vibration and valve status signals of the condensate tank and related pipelines, and write the raw data, engineering quantity data, timestamp and valid flag into the multi-source sampling frame.

[0008] The time alignment and condensation state diagram construction module is used to perform time alignment on sampling data of different frequencies, and construct a multi-channel condensation state diagram by arranging the fixed range absolute value, robust standardized value, first-order change value, local trend, local energy and data validity in a fixed sensor row order.

[0009] The physical constraint multi-task convolutional neural network module is used to receive the condensation state diagram and equipment static parameters, and output non-excited water volume estimates, condensation growth rate estimates, and drainage pressure difference estimates through time feature branches, physical grouping branches, and shared feature layers; and form the non-excited estimation reliability based on the dispersion of the outputs of multiple sub-models and the mass conservation residuals.

[0010] The active aeration excitation identification module is used to perform boundary isolation on the condensate tank when the non-excitation estimation confidence is insufficient or when preparing to drain water. It applies a pressure pulse constrained by the safety pressure through the aeration proportional valve and forms a pressure decay process through the pressure equalization branch. It constructs an active response map based on the pressure, acoustic, vibration and valve position response during the pressure rise, pressure holding and pressure equalization stages, outputs the active water volume estimate and active estimation confidence, and fuses it with the non-excitation water volume estimate.

[0011] The residual water volume closed-loop control module is used to determine whether drainage is feasible based on the fused water volume, condensation growth rate, drainage pressure difference required, drainage terminal pressure, and equipment safety pressure. When drainage is permitted, a segmented drainage method is adopted, and the residual water volume is updated based on the actual drainage volume, condensation growth during drainage, and the estimated value of the retained verification water volume. A drainage stop signal is generated based on the residual water volume ratio, drainage flow rate, and gas phase penetration signal.

[0012] The control execution and feedback module is used to convert control results into control status words and diagnostic codes, and to perform interlock control through pressure upper limit, valve position feedback, boundary leakage, data validity and emergency stop signals; and to use the reference water volume formed by the cumulative drainage volume and independent level-volume curves to perform feedback correction on the common linear calibration layer after the average output of non-excitation water volume.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a fixed-structure condensation state diagram from pressure, temperature, humidity, flow rate, liquid level, and acoustic vibration data. It simultaneously estimates water volume, growth rate, and drainage pressure difference using a physically constrained multi-task convolutional neural network. When reliability is insufficient, it actively corrects for errors by using isolated gas injection and pressure equalization responses, and updates the residual water volume segment by segment based on the actual drainage volume. This reduces misjudgments caused by single sensor failure and fixed drainage times, minimizes gas loss and pressure disturbances, and improves drainage safety through interlocking of valve position, pressure, and model applicability range. Attached Figure Description

[0014] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a system block diagram of the pipeline condensate control system based on a convolutional neural network proposed in the embodiments of the present invention; Figure 2This is a schematic diagram of the pipeline condensate control system based on convolutional neural networks proposed in the embodiments of the present invention; Figure 3 This is a schematic diagram of the physical constraint multi-task convolutional neural network structure and operation process proposed in the embodiments of the present invention; Figure 4 This is a diagram showing the test results of multi-source data processing and control status of the pipeline condensate control system proposed in the embodiments of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] The following description, using this embodiment, illustrates the data acquisition, model execution, and control signal execution processes of this system. This embodiment does not limit the number of sensors, communication interfaces, or processor models, but the data flow and control logic of each module remain unchanged.

[0017] Please see Figure 1 System block diagram of a pipeline condensate control system based on convolutional neural networks.

[0018] like Figure 1 As shown, the pipeline condensate control system based on convolutional neural networks includes a multi-source signal acquisition module, a timing alignment and condensate state diagram construction module, a physical constraint multi-task convolutional neural network module, an active aeration excitation identification module, a residual water volume closed-loop control module, and a control execution and feedback module. The data frames, state diagrams, estimated parameters, or control state words output by the previous module serve as inputs to the next module.

[0019] The industrial computer utilizes an edge host with a neural network inference accelerator. A programmable logic controller (PLC) is connected to an electrically operated inlet shut-off valve, outlet shut-off valve, injection proportional valve, drain valve, and equalizing valve. The injection pipeline connects to a depressurized nitrogen source; the same injection medium is used for model training, equipment debugging, and operational control, eliminating the need for compressed air. A mechanical check valve is connected in series downstream of the electrically operated inlet shut-off valve in the inlet pipeline. One end of the equalizing branch connects to the condensate tank's vapor phase space, and the other end connects to the downstream pipeline of the outlet shut-off valve, where an equalizing valve and a flow-limiting orifice plate are installed. The condensate tank's vapor phase space is also connected to a mechanical safety relief branch, where the safety valve's set pressure is higher than... and not higher than All five electric valves are equipped with valve position feedback devices; the industrial computer and programmable logic controller communicate via Modbus TCP, and the sensors are connected to the data acquisition card via 4-20mA, RS485, or pulse interfaces.

[0020] Please see Figure 2 , Figure 2 This is a schematic diagram of a pipeline condensate control system based on a convolutional neural network.

[0021] like Figure 2 As shown, the condensate tank is connected to the air inlet pipe, air outlet pipe, air injection pipe, drainage pipe, pressure equalization branch, and mechanical safety pressure relief branch. Each pipe is equipped with pressure, temperature, flow, and valve position detection components to collect signals and transmit them to the industrial computer and programmable logic controller (PLC). The industrial computer sequentially processes data through a multi-source signal acquisition module, a timing alignment and condensation state diagram construction module, a physical constraint multi-task convolutional neural network module, an active air injection excitation identification module, a residual water volume closed-loop control module, and a control execution and feedback module. It then sends control commands to the air injection proportional valve, drainage valve, pressure equalization valve, electric air inlet shut-off valve, and air outlet shut-off valve.

[0022] The device configuration frame is a fixed-length binary structure containing the device number, protocol version, data length, and cyclic redundancy check code. The frame body is stored in the order shown in Table 1. , , , , , , , , and The receiver extracts 16 bits of data from each field and performs the conversion; if the verification fails, a field is missing, or the conversion result exceeds the nameplate range, the original configuration is not overwritten, and an invalid configuration signal is output.

[0023] Table 1 Device Configuration Frame Field Parsing Rules This embodiment takes , , , , , , , , and .

[0024] The multi-source signal acquisition module collects pressure, temperature, humidity, flow rate, liquid level, acoustics, vibration, and valve status data from the condensate tank, condenser, connecting pipelines, and actuators within the same closed-loop control cycle, and timestamps each sampled value using a unified clock. This module includes corresponding sensors, a data acquisition card, and a clock synchronization unit.

[0025] The pressure sensors output the upstream pressure respectively. Downstream pressure Cylinder pressure Injection pressure and drainage terminal pressure . The sensor is located downstream of the outlet shut-off valve and near the equalizing branch connection. All pressures mentioned above, as well as the equipment's allowable safe pressure, are gauge pressures relative to atmospheric pressure, expressed in kPa. Temperature sensor output. , , and Dew point and humidity sensor output and bidirectional gas flow meter output The first acoustic receiver is installed on the lower outer wall of the condensate tank and outputs... The vibration sensor is installed on the side wall of the condensate tank with its sensing axis perpendicular to the tank wall and outputs a signal. The second acoustic receiver is installed on the outer wall of the drain pipe downstream of the drain valve and is used only for determining gas phase continuity. Liquid level sensor output. .

[0026] It should be noted that the output of the bidirectional gas flow meter >0 indicates that the gas flows from upstream towards the condensate tank. <0 indicates reverse flow; the bidirectional gas flow meter is used to characterize the normal gas delivery flow rate, and is used when the electric intake shut-off valve is open and the cylinder pressure is... Upstream pressure It detects reverse flow in the direction of the mechanical check valve.

[0027] The 4-20mA signal is mapped to the sensor range by subtracting 4mA from the current and then dividing by 16mA; the RS485 register is parsed according to whether it is signed or unsigned and multiplied by the factor. Within each control cycle, the pulse-type drainage flow meter uses a pulse count... Multiply by the single-pulse volume constant Obtain the drainage volume Then calculate the drainage flow rate. Its unit is L / s. Equation (2) uses Equation (5) uses the cumulative volume of the drainage section. The original values, engineered values, conversion parameters, and validity flags are all written into the multi-source sampling frame.

[0028] The sampling frequency is set to the minimum supported level, which is no less than 2.5 times the effective bandwidth of the sensor. In this embodiment, pressure, temperature, humidity, flow rate, liquid level, and valve position are sampled at 10Hz, acoustic frequencies at 4000Hz, and vibration frequencies at 2000Hz. Anti-aliasing filtering is performed before downsampling.

[0029] The multi-source sampling frame stores the device number, sensor number, timestamp, raw value, engineered value, and valid flags. Valid flags include normal, missing, communication timeout, over-range, and abnormal rate of change. The rate of change threshold is the 99.9th percentile of the absolute value of the difference between adjacent normal calibration data. Data exceeding the threshold is only written to the raw area and does not directly enter the numerical channel.

[0030] The temporal alignment and condensation state diagram construction module converts asynchronous, heterogeneous data into a six-channel tensor with a fixed row order and a fixed number of time columns. The six channels represent, in order, the fixed-range absolute value, the robust normalized value, the first-order change, the local trend, the local energy, and the validity. This tensor is a data matrix for two-dimensional convolution processing and is not a pipeline scene image.

[0031] Slow variables are uniformly resampled to 10Hz; acoustic and vibration signals are calculated every 0.1s for root mean square of acoustic amplitude, centroid of acoustic spectrum, root mean square of vibration amplitude, and kurtosis, forming four 10Hz derived signals. Continuous analog quantities are interpolated linearly, while switch quantities, operating mode codes, and validity flags are held at zero order. Interpolation is performed when there is no continuous missing value for more than 0.2s; if the missing value exceeds 0.2s, the median of the valid values ​​in the window is filled and the validity is set to 0; if there are no valid values ​​in the window, the statistics of the most recent valid window are used; if there are no historical windows after startup, the midpoint of the range is filled.

[0032] Sliding window length The timeframe is selected from 8s, 12s, 16s, 24s, and 32s based on the error in water collection; the shorter window is used when the error difference does not exceed 1% of the label range. This embodiment uses... It includes 160 columns at 10Hz.

[0033] The sensor row order is fixed as shown in Table 2. When no dew point or humidity sensor is configured, all five numerical channels and the validity channel are written as 0. The operating mode codes 0, 0.25, 0.50, 0.75, and 1.00 represent normal monitoring, isolation preparation, active excitation or gas injection pressurization, segmented drainage, pressure equalization, or pressure maintenance verification, respectively.

[0034] Table 2 Fixed row order of the condensation state diagram For the first Line 1 Project volume The absolute value of the fixed range is formed according to formula (1). and robust standardized values : In the formula, and These are the lower and upper limits of the effective measurement range, respectively, and the requirements are... ; and These are the median and absolute median difference in the current window, respectively. Take half of the minimum resolution of the project volume, but not less than ; To standardize the cutoff boundary; `clip` restricts the input to a given closed interval. Adjacent The difference is written into the first-order change channel, the current value and the forward value. Difference of each sampling point divided by Write into the local trend channel, recently The natural logarithm of the squared mean of the standardized values ​​plus 1 is written into the local energy channel; the validity channel for normal samples is written with 1, and the rest with 0.

[0035] , and Determined from normal calibration data. This embodiment uses... , and Six channels are stacked as Create a condensation state diagram and write the device number, model version, etc. The window ends with a timestamp; a new state diagram is generated after each closed-loop control cycle, and overlapping data is referenced through a circular buffer.

[0036] The physical constraint multi-task convolutional neural network module outputs non-excited water volume. Condensation growth rate Pressure difference required for drainage and non-incentivized credibility Non-excitation operating conditions include normal gas transmission and maintenance verification operating conditions that meet the stability criterion after drainage; This indicates the applicability of the estimated result to automatic control, and is not interpreted as a statistical probability.

[0037] Network input includes Condensation state diagram and 5D equipment static parameters.

[0038] In this embodiment, the 5-dimensional static parameters of the device are, in a fixed order, the effective volume of the condensate tank, the inner diameter of the drain pipe, the equivalent length of the drain pipe, the elevation difference at the drain terminal, and the rated flow coefficient of the drain valve. The designed gas phase volume is only used to calculate the physical consistency residual of boundary leakage during the active gas injection excitation process, and is not input into the non-excitation water volume, condensation growth rate, or the pressure difference prediction network required for drainage.

[0039] Equipment asset documents are structured files containing equipment serial numbers, version numbers, and checksums: Drainage pipe inner diameter The straight pipe length and elevation fields are stored in mm and converted to m by multiplying by 0.001; for elbows and tees, the equivalent length table in the file is looked up according to the component type code, and added to the straight pipe length to obtain the equivalent length of the drainage pipe. Drainage terminal elevation difference It is obtained by subtracting the elevation of the lowest point of the condensate chamber from the elevation of the drainage terminal. Converted to m³ / h for water medium and a pressure difference of 1 bar across the valve; design gas phase volume Subtract from the total contents of the condensate tank The file also stores the injection medium code, rated supply pressure, and installation location numbers of the acoustic and vibration sensors, and verifies them against the training records. If file verification fails or any static parameter exceeds the model training range, an invalid signal indicating the model's applicable range is output, and automatic control is disabled; boundary truncation values ​​are not used for replacement. The temporal feature branch and the five physical grouping branches are pooled into 32-dimensional vectors, which are then concatenated with the 5-dimensional static parameters and passed through a 69-input, 64-output fully connected layer and two 64-unit fully connected residual blocks to form shared features.

[0040] The physical grouping branches handle pressure and flow rate, temperature and humidity, liquid level and drainage, acoustic vibration, and valve status sub-matrices respectively. Each group adopts... Convolution, batch normalization, and ReLU are followed by pooling and concatenation to prevent adjacent rows with no direct physical connection from forming convolutional responses. Shared features are concatenated to the water volume, growth rate, and pressure difference output heads, respectively; the water volume head is multiplied by a sigmoid function. The other two output heads use Softplus.

[0041] Specifically, the pressure and flow group reads rows 1-5, 12, and 14 of Table 2; the temperature and humidity group reads rows 6-11; the liquid level and drainage group reads rows 3, 5, 13, 14, and 22; the acoustic and vibration group reads rows 15-18; and the valve status group reads rows 19-24. Different physical groups allow repeated reading of sensor rows that can simultaneously characterize multiple physical processes. The row numbers and order of each group are written into the model metadata and remain unchanged during operation.

[0042] Please see Figure 3 , Figure 3 A schematic diagram of the structure and operation process of a physically constrained multi-task convolutional neural network.

[0043] like Figure 3As shown, the condensation state diagram is input to the time feature branch and the physical grouping branch respectively. The time feature branch extracts the temporal variation features of pressure, temperature, and flow rate along the time axis, while the physical grouping branch performs convolution and pooling on the pressure and flow rate, temperature and humidity, liquid level and drainage, acoustic vibration, and valve state sub-matrices respectively. After the outputs of the two branches are concatenated with the equipment static parameters, they are processed through a fully connected mapping layer and a fully connected residual block to form shared features, and output the non-excited water volume, condensation growth rate, and drainage pressure difference respectively. The dispersion of water output from multiple sub-models and the mass conservation residuals are used to generate the non-excited estimation credibility. When the credibility is insufficient, the active response diagram is input to the active identification branch, outputting active water volume and active credibility, and is then weighted and fused with the non-excited estimation results.

[0044] Table 3. Examples of Model Structure and Training Parameters Table 3 shows the model structure. Training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and a maximum of 200 epochs. Validation is stopped early if the loss does not decrease for 15 consecutive epochs, and the learning rate is multiplied by 0.5 if it does not decrease for 5 consecutive epochs. Complete experimental events are divided into independent test sets, parameter switching validation sets, confidence calibration sets, and 50% training validation sets at 10%, 10%, 10%, and 70% respectively. The same event cannot cross sets. The confidence calibration set does not participate in the training of any sub-models.

[0045] Training samples include normal gas delivery, staged water addition, active gas injection, actual drainage, and post-drainage verification. The water volume added by the metering pump is used as the calibration water volume label; operational events are tagged with the sum of the cumulative drainage volume and the weighed residual volume before drainage. The level-volume curve saves calibration nodes in ascending order of level; when the level is between adjacent nodes, it is calculated using a piecewise linear method, with the interval using left-closed and right-open, and the last node closed; if the level exceeds the range, no feedback label is generated. The active response diagram uses the actual water volume before excitation as the label. The differential pressure label is when the drain valve is fully open and... continuous When the cycle exceeds the no-load noise limit ; Take the 95th percentile of the time required for the flow rate to reach the final stable value ±5% after the valve is fully open, and divide by... Round up to the nearest integer, and take the 99.9th percentile of the 10-minute data of the off-valve noise limit plus one minimum resolution.

[0046] Table 4. Label Formation Path in Multi-Task Model During training, row 13 of Table 2 was randomly blocked according to the invalidity rate of the liquid level sensor's most recent calibration period, with the blocking rate limited to [0.05, 0.30]. The fixed range absolute value, robust standardized value, change value, local trend, local energy, and validity of the blocked row were all set to 0, so that the model could still estimate water volume using other signals when the liquid level sensor temporarily failed.

[0047] Physically constrained multi-task convolutional neural networks calculate the training loss according to equation (2): In the formula, The sum of normalized Huber losses is calculated based on three terms: water volume, growth rate, and pressure difference; water volume is expressed as... Normalization is performed on the other two items, which are normalized to their respective 99.5% quantiles. The weights of the three items are 1:1:1. The mass conservation term is only used for sample pairs with the same equipment number, consecutive timestamps, valid drainage flow meters, and adjacent states that are under normal monitoring or maintained for verification. When there are active excitation, pressure equalization, fault protection, or different events, this term is set to 0 and does not participate in gradient updates.

[0048] The gradient norm ratio of the shared layer is initialized using the initial supervised loss and the mass conservation loss, and updated according to the median of the gradient norms of the last 5 rounds; if the gradient is lower than the zero-prevention amount, the previous value is used. This embodiment converges after... The safety pressure is not included in the differential pressure regression loss; the model outputs the actual required differential pressure. Whether drainage can be performed is determined by the control module.

[0049] The arithmetic mean of the water volume, growth rate, and pressure difference outputs of the five cross-validation sub-models was obtained. , and And calculate the standard deviation of water volume. The absolute value on the left side of equation (2) of the mass conservation equation is divided by... get ; Credibility calibration sets were established respectively and The empirical cumulative distribution function is calculated according to equation (3). .

[0050] The experience accumulation distribution uses only fault-free events from the confidence calibration set, and retains at most one window for each complete event. Inputs less than the minimum value of the calibration set are assigned a value of 0, those greater than the maximum value are assigned a value of 1, and those equal to the sorted value are assigned the proportion of samples not greater than that value.

[0051] Large water volume error boundary The percentage of independent measurement bias in the fault-free drainage event set for reliability calibration The 95th percentile of the proportion, limited to [0.05, 0.15]; in this embodiment, 0.10 is used. Trigger threshold. Traversing from 0.50 to 0.95, where the high-confidence false positive rate refers to... The water quantity error in the samples that reach the candidate threshold exceeds The sample proportion; after excluding candidate values ​​whose proportion exceeds 2%, the one with the fewest active identification times is selected, which is 0.78 in this embodiment.

[0052] The active gas injection stimulus identification module in When preparing to perform drainage, first close the inlet and outlet air shut-off valves, drain valve, and equalizing valve to form a boundary isolation, then apply a known pressure pulse. The excitation response segment covers the pressurization, target pressure holding, and equalizing decay phases, and is recorded. , The data includes the first acoustic receiver signal, vibration signal, gas injection valve position, and equalization valve position. The baseline before excitation is the 2.5% and 97.5% quantiles of the data after 5 seconds of isolation. The equalization completion threshold is also considered. Take pressure resolution and normal connection conditions The larger of the 95th percentiles, when continuous The equalization valve is closed during each cycle; if the condition is not met after 4 seconds of opening the equalization valve, the active identification is invalid and the equalization timeout diagnosis is written.

[0053] For the tasks of pressure equalization completion, gas phase connection, fault recovery, and valve position stabilization, select the minimum number of confirmations with a false trigger rate not exceeding 1% from 2, 3, 4, and 5 cycles respectively, and take the maximum value among the results of each task as the unified number of confirmations. In this embodiment, it is 3. Valve actuation allowance time. Take the larger of the rated full stroke time and the 95th percentile of the time of 20 actual movements, then multiply by 1.20; exceeding... If the valve is not in position, it is determined to be a malfunction.

[0054] Active incentive candidate conditions also include The problem could be due to a switch from manual to automatic control or continuous failure of the liquid level indicator. Prior to execution, the most recent 10 cycles must be considered. , and If the fluctuation does not exceed the 95th percentile of the stability calibration, the drain valve and equalizing valve are closed and there is no emergency shutdown; if boundary leakage is detected after isolation or the pressure does not stabilize within 4 seconds, the excitation is not executed and the fault protection is entered.

[0055] and The parameters were determined by candidate tests with empty cylinders and water volumes of 10%, 30%, 50%, 70%, and 85%. After whitening the response characteristics, the Mahalanobis distance between the centroids of adjacent water volume classes was calculated. Combinations of excessive peak pressure, abnormal upstream and downstream pressure, and boundary leakage were excluded, and the combination with the largest minimum inter-class distance was selected; in this embodiment, 4.0 kPa and 0.40 s were used.

[0056] controller with To stimulate the target pressure, the injection valve is adjusted every 10 ms according to the minimum stable opening increment in the valve actuator configuration file; in this embodiment, this is taken as 2%. The allowable bandwidth for the target pressure is taken as the minimum resolution of the in-cylinder pressure sensor. The larger of the values. After entering the target zone, close the air injection valve and accumulate the holding time; if the value is below the lower limit of the zone, replenish air; if the value is above the upper limit of the zone, close the air injection valve and briefly open the pressure equalization valve. The accumulated time within the zone reaches... Then it enters the equalization and decay phase. Reaching... Abnormal rate of pressure rise If the pressure continues to increase or a boundary leak is detected, immediately stop the gas injection, activate the equalization function, and output a fault signal.

[0057] Active response diagram Tensor. The first 5 lines are fixed as follows: , , The first-order differential signal, the actual opening of the gas injection proportional valve, and the pressure equalization valve position are followed by 32 rows of the first acoustic receiver frequency band, 32 rows of the vibration frequency band, and 5 rows of isolation lines. Acoustic and vibration signals are subjected to Hann window short-time Fourier transform, with the window length covering two cycles from the lowest analysis frequency and rounded up to the power of 2, and a 50% window shift. The frequency bands from the lowest analysis frequency to the upper limit of the effective bandwidth are divided into 32 bands, and the arithmetic mean of the FFT amplitudes in each band is taken; the amplitudes are then summed. Then take the natural logarithm and standardize it according to the training set statistics.

[0058] The average of the outputs of the 5 active identification sub-models is obtained. Its standard deviation is Active physical consistency residual Take the target peak pressure deviation and the boundary leakage volume as the percentage The average value of the three terms—proportion, liquid level change ratio before and after excitation, and value truncated to [0,1]—is used. If the bidirectional gas flow meter outputs standard state volumetric flow rate, it is first converted to the volume under cylinder pressure and temperature according to the gas state equation. Active excitation samples are established in the reliability calibration set. and The empirical cumulative distribution is obtained according to the same rule as in equation (3). .

[0059] In the formula, and These are the combined water volume and the reliability of the combined water volume, respectively. Fusion control threshold The credibility calibration set simultaneously has , And formed on samples of actual water volume labels; high confidence false positive rate refers to The water quantity error in the samples that reach the candidate threshold exceeds The sample proportion is determined by excluding candidate values ​​whose proportion exceeds 2%, and the value with the fewest manual confirmations is selected. In this embodiment, the value is 0.80.

[0060] like or After the system equalizes the voltage, it will either require manual confirmation or output a low-confidence alarm; if and Equalize pressure and resume gas supply; when the drainage start ratio is reached, maintain boundary isolation and enter the drainage pressure executability determination.

[0061] The residual water volume closed-loop control module is based on the combined water volume. , Integration credibility growth rate Pressure difference required for drainage Drainage flow rate and drainage terminal pressure Generate target pressure, drainage timing, and stop signal. Residual water volume refers to the estimated volume of liquid water still within the effective volume of the condensate tank after drainage begins.

[0062] when and It then enters the drainage preparation phase. The activation ratio has not been reached and... At that time, according to: Calculate the estimated arrival time when A scheduled drainage signal is generated at the time. No scheduled drainage signal is generated when the value is ≤0. The drainage execution period requires the most recent 10 cycles. , and The fluctuation did not exceed the 95th percentile of the stability calibration, all 5 valves were fault-free, and the manual maintenance position and emergency stop position were not set.

[0063] Close the equalizing valve and the inlet and outlet air shut-off valves during drainage preparation. The candidate target pressure is... Higher than When pressurization and drainage are prohibited, the lower 3 bits enter fault protection 110, and 0x0002 is written to the independent 16-bit diagnostic register. When the candidate value is valid, the air injection valve adjusts the pressure in 2% increments every 10ms; the target pressure allowable bandwidth is the larger of the minimum resolution of the pressure sensor and 0.5% of the target pressure. After entering the target band, the air injection valve is closed and confirmation is performed. One cycle. During drainage, the pressure equalization valve remains closed. When the pressure is below the lower limit of the target zone, replenish the air supply. When the pressure is within the target zone, keep or close the air supply valve. When the pressure is above the upper limit, immediately close the drain valve and the air supply valve before opening the equalizing valve.

[0064] In the formula, ; For the first The drainage volume obtained by accumulating pulses in the drainage section; This is the actual time interval between the last residual water volume update and the current estimated time of maintenance, covering the drainage, pressure equalization, and maintenance processes. and This is the unexcited verification value after equalizing and stabilizing the pressure; If the drain flow meter is invalid or the pulse interval is abnormal, do not execute formula (5), maintain the previous residual water volume, close the drain valve and air injection valve and enter the pressure equalization verification.

[0065] In equation (5), This indicates the reliability of the mass conservation recursion before drainage. When the current credibility of the review is low and the overall credibility is high, Increase and emphasize the recursive calculation of actual drainage volume; when the current verification credibility is high. Decrease and increase The impact.

[0066] The drainage system employs a segmented approach of "drainage—pressure equalization—holding—reassessment." The preset opening degree of the drain valve is selected from candidates of 50%, 75%, and 100%, with the drainage volume variation coefficient not exceeding 5% and the shortest average drainage time. Timing begins once the valve reaches the preset opening degree. The minimum effective opening time of the drain valve is also specified. Select the shortest candidate time that can form no less than 5 effective flow pulses in 3 consecutive cycles. Round up to get After the drainage section is completed, close the drain valve and air injection valve in sequence, and open the pressure equalization valve to meet the requirements. Then close the equalizing valve and hold for 1-4 seconds. A significant rebound is defined as the increase in volume of the subsequent drainage section exceeding the larger of the single-pulse volume and the no-load noise within the corresponding volume of that drainage section. The residual proportion should not exceed [a certain value]. And recently Drainage will cease if the water level does not rise significantly.

[0067] The gas-phase penetration threshold is calibrated by the ratio of the frequency band energy of pure water discharge and gas passing through the drain pipe, as recorded by the second acoustic receiver. The decision metric with the lowest error rate among the two sample types is selected, and the intersection of the two distributions is used as the threshold. When the number of acoustic sub-windows exceeds the threshold, it is determined that the gas phase is connected.

[0068] The control execution and feedback module writes the control results into an 8-bit control status word and transmits the cause codes via an independent 16-bit diagnostic register: 0x0000 Normal, 0x0001 Low model confidence, 0x0002 Insufficient pressure capacity, 0x0003 Valve action timeout, 0x0004 Boundary leakage, 0x0005 Drainage flow meter invalid, 0x0006 Pressure equalization timeout, 0x0007 Drainage terminal pressure sensor invalid, 0x0008 Invalid model applicable range, 0x0009 Data acquisition link failure, 0x000A Invalid device configuration, 0x000B Mechanical check valve malfunction. The industrial computer sends the control status word, diagnostic codes, target pressure, and valve opening; the controller returns the actual status word and valve position.

[0069] Table 5. Bit Field Resolution Rules for Control Status Words The controller first parses bit 7, then bits 5 and 6, and finally the lower 3 bits. Bit 3 uses valid conditions that change with the operating state. The valve is not in... Write 0x0003 when the internal position is reached. When the electric intake shut-off valve opens... And the bidirectional gas flow meter is continuous When reverse flow is detected during a cycle, the mechanical check valve is deemed abnormal and 0x000B is written. When bit7 is set, the air injection valve, drain valve, and equalizing valve are closed by default. The equalizing valve is allowed to be opened briefly only when the shutdown reason is cylinder overpressure, the downstream and equalizing branch are in normal condition, and the downstream isolation signal is not set. In other cases, the mechanical safety pressure relief branch protects against the problem.

[0070] Drainage Incident Log , , , , Five valve positions, residual water volume, and cumulative drainage volume. After drainage, pressure equalization is performed and stabilization is awaited; model verification values ​​are only used for control confirmation. Independent residual water volume is obtained from the level-volume curve. Confirmation is required before resuming gas supply. With the equalizing valve closed, first open the outlet shut-off valve, and wait... continuous After each cycle falls within the 95th percentile of normal fluctuations, the electric intake shut-off valve is opened, and... continuous A cycle greater than the minimum resolution of the flow meter confirms the resumption of positive gas delivery.

[0071] Feedback calibration is only used when there are no safety faults, the drainage flow meter is effective throughout the entire process, the liquid level is valid at the end, and before drainage... The event. The reference water volume is the sum of the cumulative drainage volume and the independent residual water volume. Online correction only applies to the common linear calibration layer after averaging the outputs of the 5 unexcited sub-models: with Obtain the calibrated water volume and limit it to Candidate gain is limited to Actively identify branches, convolutional feature layers, and physical constraint parameters that are not updated online.

[0072] Grand total Candidates are generated after a number of valid events. and Candidate parameters must reduce the mean absolute error of valid events, ensuring that the error of the parameter switching validation set increases by no more than 1%, and without generating out-of-bounds output, before being written into the spare parameter area. Switching occurs when the system is running stably and there are no drainage tasks; three consecutive high-confidence events with a deviation exceeding [a certain threshold] are considered valid parameters. If the validation set error increases by more than 2% relative to before the switch, a rollback will occur. The independent test set will never participate in online calibration or parameter switching decisions.

[0073] Furthermore, the multi-source signal acquisition module uses the device number and closed-loop control cycle number as a joint primary key to write the engineering quantities, original register values, sampling timestamps, and valid flags within the same cycle into the same multi-source sampling frame. Slow variables such as pressure, temperature, humidity, liquid level, and valve position are indexed using a 10Hz timestamp. The original acoustic and vibration sequences are stored separately in a high-speed buffer, and the start and end addresses of the original segments are recorded after the derived features are generated. The industrial computer first checks whether the range, resolution, magnification, and device number in the sensor configuration file are complete, and then performs numerical conversion. If the upper limit of the range is not greater than the lower limit of the range, the magnification is 0, or the device number is inconsistent, the corresponding valid flag is set to invalid configuration, and the data is not used to drive control. The communication timeout threshold is three times the sensor's nominal update cycle; if the continuous timeout exceeds one sliding window, the system outputs a data acquisition link fault and prohibits active identification and automatic drainage. For abnormal rate of change values, the original values ​​are still saved so that maintenance personnel can distinguish between real sudden changes and communication errors; after recovery, continuous... Only when all sampling points meet the range and rate of change conditions will the valid marker be restored to normal.

[0074] During the generation of the condensation state diagram, the industrial computer reads multi-source sampling frames according to the fixed row numbers in Table 2, rather than temporarily sorting them according to the sensor access order. Each row stores the sensor type code, engineering unit code, range upper and lower limits, and channel version number in the model metadata, and these are compared item by item before inference. If any row is missing or the version is inconsistent, only the missing row pattern already included in the training phase is allowed; other sensor data are not allowed to be substituted. The fixed-range absolute value channel is used to retain the absolute positions of liquid level, pressure, and temperature; the robust normalized channel and the variation channel are used to reflect short-term dynamics; and the validity channel distinguishes between the true median and the median filled in due to missing values. The 0.1s sub-windows of acoustic and vibration derived features correspond one-to-one with the 10Hz column. When the sub-window is less than the full length, zero-padding expansion is not used; instead, the corresponding numerical and validity channels are written as 0. After generating the state diagram, the industrial computer checks whether its dimensions are correct. Check whether there are non-finite values ​​and whether the pattern code belongs to the five specified discrete values; if the check fails, do not call the neural network, and retain the previous valid estimate for alarm display, but do not use it to start a new drainage action.

[0075] During model inference, the temporal feature branch performs convolution along the time axis for each sensor row, while the physical grouping branch first copies sub-matrices for pressure and flow, temperature and humidity, liquid level and drainage, acoustic vibration, and valve status groups according to predetermined row numbers, and then performs convolution and pooling respectively. The outputs of the two branches are fused with the equipment static parameters in a fully connected mapping layer to avoid using control thresholds as physical inputs for water volume prediction. Each model file, in addition to network weights, also stores the input row order, static parameter training range, label scale, standardized statistics, empirical cumulative distribution function, and model version number; the industrial computer loads the model only when the above metadata matches the equipment configuration. After the five sub-models are inferred sequentially, the water volume, growth rate, and pressure difference outputs are cached respectively. The mean and standard deviation are calculated only when all five sub-models are completed and no numerical overflow occurs; if any sub-model malfunctions, the current cycle is interrupted. Set to 0 and trigger active candidate identification. Water output is mapped to Sigmoid. The growth rate and differential pressure are kept non-negative by Softplus, but the controller still performs secondary interlocking based on the equipment safety pressure, valve status and model applicability range signals. The neural network output cannot be directly written to the valve drive register.

[0076] Training data is presented in complete trial events as the smallest unit. Verification samples are kept only when the equalizing valve is closed. Furthermore, the acoustic vibration energy is extracted after recovery. Training, 5-fold validation, confidence calibration, parameter switching validation, and independent test sets are divided by event, with adjacent sliding windows from the same event always located in the same set. Before model release, normal gas delivery, low liquid level, high liquid level, maintenance verification, and sensor-missing operating conditions are statistically analyzed respectively; the acceptance boundary in this embodiment is that the average absolute error of water volume under each operating condition does not exceed [a certain value]. Furthermore, the average absolute error of the pressure difference does not exceed 5 kPa, and no data will be released if any operating condition exceeds the limit.

[0077] Before the active air injection excitation begins, the controller closes the equalizing valve, the outlet shut-off valve, and the electric inlet shut-off valve, and confirms boundary isolation using valve position feedback. After isolation is completed, the allowable pressure, acoustic, and vibration ranges are calculated using 5-second baseline data; if... If continuous flow persists beyond one flow resolution, it is considered a boundary leak. During the pressurization phase, the injection valve opening is adjusted every 10 ms, and the injection valve is closed once the pressure reaches the target zone; only when... Gas is supplied only at the minimum stable opening when the pressure falls below the lower bandwidth limit. During the pressure equalization phase, the gas injection valve is closed, and then the pressure equalization valve is reopened. After active identification is completed, proceed as follows... and Choose manual confirmation and restore gas supply or drainage preparation status; do not rely solely on [the following]. Start drainage.

[0078] During segmented drainage, the time before the drain valve reaches the preset opening is not counted in the effective drainage segment; the controller starts accumulating data from the moment the valve reaches the preset opening position. One control cycle. Each drainage section Only pulses that are accumulated after the valve is in position and the flow meter is active are recorded. After pressure equalization is complete, the pressure change range, acoustic energy, and vibration energy are continuously monitored. After each cycle returns to the baseline verification interval, the new state diagram is input into the non-excitation model. If the cumulative drainage volume of the event is greater than... If so, stop automatic drainage and manually verify; Take the cumulative permissible error of the flow meter and The larger value in the range.

[0079] The control status word and diagnostic register are sent by the industrial computer within the same Modbus TCP transaction. A 16-bit application-layer CRC field is appended to the holding register block. The controller performs a CRC16 calculation on the transaction sequence number, status word, diagnostic code, target pressure, and valve opening, and then compares the results. This field is not interpreted as the serial CRC of the Modbus TCP message itself. Status transitions follow the sequence of "command—valve position confirmation—data validity confirmation—next state." When resuming gas supply, first confirm that the equalizing valve is closed and... Then open the exhaust shut-off valve; continuous After each cycle is within the normal fluctuation range, the electric intake shut-off valve is opened, and... continuous A cycle greater than one minimum resolution confirms positive gas delivery.

[0080] During the feedback correction phase, the industrial computer filters out records containing overpressure, boundary leakage, valve timeout, flow meter malfunction, or liquid level exceeding the calibration range. The sum of the cumulative drainage volume and the residual volume obtained from independent level-volume curves is then used as the reference water volume before drainage. The common linear calibration layer is located after the average output of the five unexcited sub-models and before confidence calculation and active fusion, and includes candidate gains. and bias Only change the non-incentivized water volume scale, and When calculating candidate parameters, the most recent... The least squares fit is performed on each valid event; switching parameters is rejected if the error on the validation set increases by more than 1% or if out-of-bounds output is generated.

[0081] The equipment configuration frame, equipment asset file, and model metadata constitute three independent parameter source chains. The equipment configuration frame provides adjustable safety and control parameters during operation; the equipment asset file provides physical attributes such as pipe diameter, equivalent length, elevation, valve flow coefficient, and design gas phase volume; and the model metadata provides input row order, normalization range, label scale, and empirical cumulative distribution function. The industrial computer first performs verification separately, and then performs joint matching based on equipment number, equipment asset version, and applicable equipment type. If only the control threshold changes, the equipment configuration frame can be updated without retraining the model; if the pipe diameter, condensate tank volume, or drainage path structure changes, it is necessary to re-determine whether the static parameters are still within the training range. The minimum and maximum values ​​in the model metadata are formed by training event statistics; values ​​equal to the endpoints are treated as valid, and values ​​exceeding either endpoint are not used for automatic inference. During configuration updates, the controller maintains normal monitoring status, and the air injection valve, drain valve, and equalizing valve are all closed; the working parameter area is only replaced after the new configuration has been read consistently twice and passed verification. This process avoids mixing operating control parameters, equipment physical properties, and model statistical parameters from the same source, and also avoids the situation where some new values ​​and some old values ​​coexist after an external file update fails.

[0082] In case of sensor failure, the system retains the original data and restricts control access. If acoustic or vibration data times out during active identification, the current response graph is deemed invalid and pressure equalization is performed. During drainage... Write 0x0007 when invalid. Write 0x0005 when invalid; write 0x0008, 0x0009, and 0x000A respectively when the model's applicable range is invalid, the data acquisition link is faulty, or the device configuration is invalid. When the valve position feedback and control command are inconsistent, the valve position feedback shall be taken as the actual state; after the fault is recovered, the sensor engineering value must be continuously updated. Each cycle is within the range and rate of change boundaries.

[0083] Using the parameters of this embodiment to illustrate a complete drainage event: the system obtains a blended water volume ratio of 100%. and Then proceed with drainage preparation. Model output. Then, combine it with Add and combine Comparison: Pressure is increased only when the target pressure is effective. The drain valve continues to operate after reaching the preset opening degree. Each cycle, during which according to Deviation replenishment. After the drainage section is completed, close the drain valve and the air injection valve in sequence, and open the pressure equalization valve. Wait... A status diagram for maintaining verification is then generated. The industrial computer uses the actual drainage volume, And verify the estimated amount and substitute it into equation (5) to update the residual water volume; to achieve Then end. When resuming gas supply, open the valves in the order of downstream and then upstream, and so on. continuous A cycle greater than one minimum resolution confirms positive gas delivery.

[0084] Please see Figure 4 , Figure 4 This is a graph showing the test results of multi-source data processing and control status of the pipeline condensate control system.

[0085] like Figure 4 As shown, the system synchronously processes upstream pressure, downstream pressure, cylinder pressure, drainage terminal pressure, and target pressure, and outputs non-excited water volume, actively identified water volume, fused water volume, and reference water volume, respectively. During active identification and segmented drainage, gas flow rate, drainage flow rate, air injection valve, drainage valve, and pressure equalization valve change accordingly according to the control stage. The non-excited reliability, active reliability, and fused reliability are dynamically adjusted based on data validity and active identification results. Simultaneously, diagnostic codes record pressure capability, sensor status, and actuator anomalies. In the figure, the target pressure does not exceed the upper limit of the operating pressure, and the fused water volume gradually decreases with segmented drainage, indicating that the system can complete multi-source data estimation, active correction, safety judgment, and closed-loop update of residual water volume.

[0086] In summary, this system forms a closed loop of "multi-source acquisition - state diagram construction - non-excitation estimation - active identification - safe drainage - independent feedback correction", and restricts the direct driving of the actuator by the neural network through boundary isolation, pressure equalization branches, valve position interlocking, pressure capacity judgment and model applicability range verification.

[0087] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A pipeline condensate control system based on convolutional neural networks, characterized in that, include: The multi-source signal acquisition module is used to collect pressure, temperature, humidity, flow rate, liquid level, acoustics, vibration, and valve status of the condensate tank and related pipelines. The time-series alignment and condensation state diagram construction module is used to perform time-series alignment on sampled data and generate condensation state diagrams containing absolute values, standardized values, changes, trends, energy, and validity. A physical constraint multi-task convolutional neural network module is used to output water volume, condensation growth rate, pressure difference required for drainage, and reliability based on the condensation state diagram. An active aeration excitation identification module is used to isolate the condensate tank before drainage if the reliability is insufficient or before drainage, apply pressure pulses and equalize pressure in a controlled manner, and correct the water volume based on pressure, acoustic and vibration responses; The residual water volume closed-loop control module is used to update the residual water volume by combining the drainage volume and the verification water volume and generate drainage control instructions. The control execution and feedback module is used to execute commands based on the pressure limit and valve position interlock and to correct the water output using effective drainage events.

2. The pipeline condensate control system based on convolutional neural networks according to claim 1, characterized in that, The multi-source signal acquisition module includes an upstream pressure sensor, a downstream pressure sensor, an in-cylinder pressure sensor, an injection pressure sensor, a drainage terminal pressure sensor, a temperature sensor, a humidity or dew point sensor, a bidirectional gas flow meter, a liquid level sensor, a drainage flow meter, a first acoustic receiver installed on the outer wall of the condensate tank, a second acoustic receiver installed on the outer wall of the drainage pipe, a vibration sensor, and a valve position feedback device; the multi-source signal acquisition module writes the device number, sensor number, sampling timestamp, raw value, engineering value, and valid flag into the multi-source sampling frame.

3. The pipeline condensate control system based on convolutional neural networks according to claim 2, characterized in that, The time-series alignment and condensation state diagram construction module resamples the sampled data to a unified time column according to a fixed sensor row order, and generates a fixed-range absolute value channel, a robust standardized value channel, a first-order change channel, a local trend channel, a local energy channel, and an effectiveness channel respectively; continuous analog quantities are aligned by interpolation, valve status and operating mode codes are held at zero order, and the numerical channels of abnormal or missing data are filled with preset values ​​and the effectiveness channel is set to invalid.

4. The pipeline condensate control system based on convolutional neural networks according to claim 3, characterized in that, The physical constraint multi-task convolutional neural network module includes a time feature branch and a physical grouping branch. The physical grouping branch processes pressure and flow data, temperature and humidity data, liquid level and drainage data, sound and vibration data, and valve status data, respectively. The outputs of the two branches are fused with the equipment static parameters, including the effective volume of the condensate tank and the structural parameters of the drainage passage, to form a shared feature. The shared feature is connected to the water output head, the condensation growth rate output head, and the pressure difference output head required for drainage, respectively.

5. The pipeline condensate control system based on convolutional neural networks according to claim 4, characterized in that, The physical constraint multi-task convolutional neural network module uses multiple sub-models with the same structure but different training data. It takes the average value of the water volume, condensation growth rate and drainage pressure difference output of each sub-model. Based on the dispersion of the water volume output and the mass conservation residual formed by the water volume change, condensation growth and drainage volume at adjacent control times, it generates the non-excitation estimation confidence by combining the empirical cumulative distribution function of the confidence calibration data.

6. The pipeline condensate control system based on convolutional neural networks according to claim 1, characterized in that, The condensate tank is connected to an inlet pipe, an outlet pipe, an injection pipe, a drain pipe, and a pressure equalization branch. The inlet pipe and outlet pipe are respectively equipped with an electric inlet shut-off valve and an outlet shut-off valve. The injection pipe is connected to a nitrogen source and is equipped with an injection proportioning valve. The pressure equalization branch is connected to the vapor phase space of the condensate tank and the downstream pipe of the outlet shut-off valve and is equipped with a pressure equalization valve. Before active injection, the electric inlet shut-off valve, outlet shut-off valve, drain valve, and pressure equalization valve are closed. After the pressure is maintained, the injection proportioning valve is closed and the pressure equalization valve is opened. The drainage pipe is equipped with a drainage valve, and the pressure equalization branch is connected to the gas phase space of the condensate tank and the downstream pipe of the gas outlet shut-off valve and is equipped with a pressure equalization valve.

7. The pipeline condensate control system based on convolutional neural networks according to claim 6, characterized in that, The active injection excitation identification module uses the cylinder pressure increment, injection pressure increment, cylinder pressure first-order difference, actual opening of the injection proportional valve, and pressure equalization valve position as pressure and valve position features of the active response diagram, and writes the time-frequency features of acoustic and vibration signals into the corresponding frequency band region. The active identification sub-model outputs active water volume estimates and active estimation confidence based on the active response map, and performs weighted fusion of non-excited water volume estimates and active water volume estimates according to the non-excited estimation confidence and active estimation confidence.

8. The pipeline condensate control system based on convolutional neural networks according to claim 1, characterized in that, The residual water volume closed-loop control module superimposes the drainage terminal pressure with the pressure difference required for drainage to obtain the candidate target pressure, and allows segmented drainage when the candidate target pressure is not higher than the upper limit of the operating pressure. After each drainage segment is completed, the drainage valve and the air injection proportional valve are closed in sequence, and the pressure equalization valve is opened to complete the pressure equalization and form a maintenance verification water volume. The current residual water volume is updated according to the previous residual water volume, the amount of condensate generated between adjacent verification times, the actual drainage volume of the drainage segment, the maintenance verification water volume, and the corresponding confidence level. When the residual water volume reaches the stop condition, the drainage flow rate is lower than the noise threshold, or gas phase penetration is detected, drainage is stopped and the pressure equalization verification state is entered.

9. The pipeline condensate control system based on convolutional neural networks according to claim 8, characterized in that, The control execution and feedback module uses control status words to indicate normal monitoring, isolation preparation, gas injection and pressurization, segmented drainage, pressure equalization verification, gas supply restoration, and fault protection status. It also uses a diagnostic register to indicate low model confidence, insufficient pressure capacity, valve action timeout, boundary leakage, sensor invalidity, pressure equalization timeout, and invalid model applicability range. The control execution and feedback module controls the gas injection proportional valve, drainage valve, and pressure equalization valve according to preset priorities based on emergency stop signals, pressure upper limit signals, valid data signals, and valve position feedback.

10. The pipeline condensate control system based on convolutional neural networks according to claim 1, characterized in that, The control execution and feedback module uses the cumulative drainage volume obtained from the drainage flow meter and the residual water volume after drainage obtained from the level-volume calibration curve to form a reference water volume before drainage. It then corrects the common linear calibration layer set after the average water volume output of multiple non-excitation sub-models. The candidate calibration parameters are only enabled when the candidate calibration parameters reduce the water volume error of the effective drainage event, the error increase of the parameter switching verification data does not exceed the preset value, and the calibration result does not exceed the effective volume of the condensate tank. The parameters of the active gas injection excitation identification module and the convolutional feature extraction layer remain unchanged.

Images