A cryogenic liquid cargo pump control and intelligent diagnosis method and system

Abstract

The present application relates to the technical field of cryogenic liquid cargo pumps, and discloses a method and system for controlling and intelligently diagnosing cryogenic liquid cargo pumps. The method comprises the following steps: acquiring an image of the liquid inlet channel of the cryogenic liquid cargo pump and identifying a bright straight line track in the image, wherein an ultrasonic detector and a camera are installed on the liquid inlet channel of the cryogenic liquid cargo pump; performing ultrasonic detection on the liquid inlet channel of the cryogenic liquid cargo pump, and judging whether the object forming the bright straight line track on the image is a gas or a solid based on the detection result; controlling whether the pressure relief valve of the cryogenic liquid cargo pump is activated based on the detection result; acquiring a subsequent detection signal, and controlling the subsequent protection action after the pressure relief valve is activated based on the subsequent detection signal. The present application has the advantages of detecting anomalies earlier, reducing system response time, and improving the safety and operational reliability of the cryogenic liquid cargo pump.

Description

Technical Field

[0001] The present application relates to the technical field of cryogenic liquid pumps, and in particular to a cryogenic liquid cargo pump control and intelligent diagnosis method and system. Background Art

[0002] Cryogenic liquid cargo pumps are specifically designed for transporting cryogenic liquid cargoes (such as liquefied natural gas (LNG), liquefied petroleum gas (LPG), and liquid nitrogen). They are widely used in LNG carriers, cryogenic storage tanks, and gas liquefaction plants. Due to the unique characteristics of cryogenic liquids, these pumps must possess low-temperature resistance, strong sealing, and resistance to cavitation to ensure safe and stable transportation. However, in actual operation, the working state of cryogenic liquid cargo pumps is easily affected by impurities and cavitation, which can cause equipment damage or system failure. Therefore, effective emergency control is particularly important.

[0003] During the operation of cryogenic liquid cargo pumps, tiny solid impurities may be present in the liquid medium. These include particle deposits within the pump system, debris from the inner walls of the pipes, or external particles that enter the liquid due to environmental factors. Once these impurities enter the pump body with the liquid flow, they may become lodged between key components such as the impeller and guide vanes, causing flow anomalies, impeller damage, or even blockage of the pump cavity. Large impurity particles can directly cause physical damage, while smaller particles, due to the high-speed flow and low temperature environment, can quickly cause localized blockages, impacting normal pump operation.

[0004] In addition to the effects of impurities, cavitation is also a common problem in the operation of cryogenic liquid cargo pumps. When the pressure at the pump's suction port falls below the vapor pressure of the liquid being transported, the liquid partially vaporizes, forming bubbles that rapidly burst as the liquid flows into the high-pressure area. During this process, the impact force generated by the collapse of the bubbles can severely erode the impeller and guide components, leading to material fatigue damage and further reducing pump efficiency. In cryogenic environments, cavitation can be more pronounced due to pressure fluctuations caused by temperature changes. This not only increases pump vibration but can also cause rapid wear or fracture of internal components.

[0005] Existing emergency control methods for cryogenic liquid cargo pumps primarily rely on pressure sensors, vibration sensors, and temperature monitoring systems to determine the pump's operating status and trigger appropriate protective measures. For example, flow anomalies are determined by monitoring the pressure difference between the inlet and outlet, or early signs of cavitation are detected through vibration sensors. However, these methods have significant shortcomings when it comes to detecting impurities entering the pump cavity. On the one hand, due to the small size and high flow rate of impurity particles, traditional sensors have difficulty effectively identifying them before they enter the pump cavity, and it is usually necessary to wait until the equipment is damaged before abnormalities can be detected. On the other hand, the vibration characteristics of cavitation often require a long period of data accumulation to effectively distinguish, making real-time detection and rapid response impossible. Summary of the Invention

[0006] In order to detect anomalies earlier, reduce system response time, and improve the safety and operational reliability of cryogenic liquid cargo pumps, the present application provides a cryogenic liquid cargo pump control and intelligent diagnosis method and system.

[0007] In a first aspect, the present application provides a cryogenic liquid cargo pump control and intelligent diagnosis method, which adopts the following technical solutions:

[0008] A cryogenic liquid cargo pump control and intelligent diagnosis method comprises the following steps:

[0009] S1. Acquire an image of the liquid inlet channel of a cryogenic liquid cargo pump and identify a bright straight line track in the image, wherein an ultrasonic detector and a camera are installed on the liquid inlet channel of the cryogenic liquid cargo pump;

[0010] S2. Perform ultrasonic detection on the liquid inlet channel of the cryogenic liquid cargo pump and, based on the detection results, determine whether the object forming a bright straight line on the image is a gas or a solid;

[0011] S3. Control whether the pressure relief valve of the cryogenic liquid cargo pump is activated based on the detection results;

[0012] S4. Obtain a subsequent detection signal, and control a subsequent protection action after the pressure relief valve is activated based on the subsequent detection signal.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Acquire an image of the liquid inlet channel and perform preprocessing, wherein the camera is provided with a preset shooting step and a preset exposure time when shooting;

[0015] S12. Inputting the image into a pre-trained convolutional neural network, wherein the convolution kernel of the convolutional neural network is used to identify straight line features;

[0016] S13. Obtain the detection result output by the convolutional neural network.

[0017] Optionally, the convolutional neural network training step in S12 is:

[0018] Acquire pre-training image materials, wherein the pre-training image materials include images of a liquid inlet channel operating normally, images of the liquid inlet channel with smear and marked smear angles, and artificially simulated smear data by simulating smear caused by delayed exposure using different camera parameter settings;

[0019] Perform data preprocessing on the pre-training image materials and form a pre-training set;

[0020] Evenly discretize 0°-180° into 12 main directions to cover possible smear angles;

[0021] Initialize 16 3×3 convolution kernels as direction-sensitive convolution kernels, ensuring that there is at least one filter for each main direction, where each direction corresponds to the line detection capability of a specific angle, and the parameters of the convolution kernel are fixed after being generated by Gaussian weighted line segments;

[0022] Output the direction-sensitive convolution kernel and calculate the variance of the directional response at the same position. If the variance is too large, increase the penalty to reduce directional confusion.

[0023] The pre-training set is batch-inputted into the convolutional neural network for forward propagation;

[0024] Perform loss calculation, where the loss function includes a classification cross entropy loss that measures whether the smear direction classification is correct and a direction response variance penalty that suppresses simultaneous activation of multiple directions;

[0025] Backpropagation and optimization are performed, in which the first layer of direction-sensitive convolution kernels is not updated, and the parameters of subsequent convolution and fully connected layers are updated to optimize the direction classification ability.

[0026] Optionally, the convolutional neural network in S12 includes:

[0027] The direction-sensitive convolution layer sets 16 3×3 direction-sensitive convolution kernels, covering 12 main directions;

[0028] Directional response variance calculation, which is used to calculate the variance of responses in different directions and add a penalty term to force the network to select the most significant direction;

[0029] Subsequent convolution and pooling layers are used to extract high-level features;

[0030] The fully connected layer is used to extract the final smear features and perform classification;

[0031] The loss function is the classification cross entropy loss and the directional response variance penalty.

[0032] Optionally, the S11 includes the following steps:

[0033] Median filtering is performed on the image to remove salt and pepper noise and preserve edge structure;

[0034] The image is processed by background modeling and foreground extraction to extract the dynamically changing parts.

[0035] Optionally, in S1 , when the camera is exposed, a laser pulse is synchronously triggered to irradiate the interior of the liquid inlet channel of the cryogenic liquid cargo pump, wherein the pulse irradiation direction is the same as the shooting direction and is perpendicular to the liquid inlet direction.

[0036] Optionally, the camera is a frost-resistant camera.

[0037] Optionally, the laser pulse emitted by the camera is longer than 100 μs, so that the moving debris forms a bright straight line track in the image.

[0038] In a second aspect, the present application provides a cryogenic liquid cargo pump control and intelligent diagnosis system, which adopts the following technical solutions:

[0039] A cryogenic liquid cargo pump control and intelligent diagnosis system comprises a processor running a program of any one of the cryogenic liquid cargo pump control and intelligent diagnosis methods described above.

[0040] In a third aspect, the present application provides a storage medium that adopts the following technical solution:

[0041] A storage medium storing a program for the cryogenic liquid cargo pump control and intelligent diagnosis method described in any one of the above.

[0042] In summary, this application includes at least one of the following beneficial technical effects:

[0043] By placing a camera and laser pulse illumination within the liquid inlet channel, combined with time-delayed exposure technology, this application can efficiently identify debris trails using a convolutional neural network model before the debris enters the pump chamber, and further confirm its physical properties through ultrasonic detection. Compared to traditional solutions that rely on pressure sensors or vibration detection, this application can achieve early warning before debris enters the pump chamber, thereby providing a longer response time for subsequent protective measures, improving the real-time detection, and reducing misjudgments and missed detections.

[0044] This application combines multi-dimensional signals such as current monitoring, temperature detection, and vibration analysis to establish a graded response mechanism that can dynamically adjust the operating status of the pump according to the degree of debris impact. For example, when the impact of debris is small, the system can only adjust the pump speed or trigger pressure relief; when debris is stuck in the impeller, a speed limit operation + backwash strategy can be implemented; in severe cases, emergency shutdown protection can be triggered. Compared with traditional solutions that rely solely on a single shutdown logic, this application can minimize the damage caused by debris to the pump without affecting the normal operation of the pump, avoid unnecessary shutdowns, and improve the stability and service life of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of the control and intelligent diagnosis method of the cryogenic liquid cargo pump. DETAILED DESCRIPTION

[0046] Embodiments of the present application are described in detail below, examples of which are illustrated in the accompanying drawings.

[0047] The present application discloses an embodiment, in which a cryogenic liquid cargo pump control and intelligent diagnosis method includes the following steps S1-S4.

[0048] S1. Acquire an image of a liquid inlet channel of a cryogenic liquid cargo pump and identify a bright straight line track in the image, wherein an ultrasonic detector and a camera are installed on the liquid inlet channel of the cryogenic liquid cargo pump.

[0049] To obtain high-quality images, the camera uses a preset capture step size and exposure time. The capture step size determines the interval between image sampling, ensuring that the complete flow of liquid is captured at different time points, while the exposure time directly affects the formation and clarity of bright lines. For faster-flowing debris, a longer exposure time ensures that the debris tracks are sufficiently distinct. However, if the exposure time is too long, the tracks may overlap, affecting recognition accuracy. Therefore, the exposure time needs to be optimized to meet the following requirements: , where L is the length of the bright line, v is the velocity of the debris, is the exposure time.

[0050] It should be noted that in the prior art, conventional continuous light sources are usually used to expose the liquid inlet channel. However, since the exposure time of continuous light sources is usually 10ms, this will cause the moving debris to form a long smear during the camera imaging process. For example, when the speed of the debris is 5m / s, the distance it moves within the 10ms exposure time is:

[0051]

[0052] However, when the smear length exceeds 10 mm, it is easily confused with the reflected light from the inner wall of the pipe, reducing the accuracy of detection and may cause the convolutional neural network to misidentify the smear direction.

[0053] In contrast, using pulsed laser as an exposure light source can significantly shorten the exposure time, resulting in a shorter trail of debris in motion when the camera images it. The core principle is that pulsed laser can provide high-intensity illumination in a very short time, allowing the sensor to receive sufficient light in a short period of time, thus obtaining a clear image without the need for long exposure times. Under traditional continuous light sources, the exposure time needs to be long enough to ensure sufficient photon accumulation, otherwise the image will be too dark or even impossible to form due to insufficient light. The high-power characteristics of laser pulses enable them to provide a luminous flux equivalent to a long-term exposure of a conventional light source in a very short time (such as 100μs), thereby ensuring image brightness while reducing motion trails.

[0054] If a pulsed laser is used, the exposure time can be shortened to 100 μs. In this case, the distance traveled by the debris during the exposure period is calculated as follows: .

[0055] This shows that compared to the 5cm smear of a continuous light source, the pulsed laser reduces the smear length to just 0.5mm, far below the interference threshold of 10mm. This makes the debris trajectory clearer, avoids confusion with reflections from the pipeline structure, and improves the accuracy and reliability of detection. In addition, the extremely short exposure time of the pulsed laser reduces ambient light interference and the impact of the liquid medium on light propagation, allowing the detection system to achieve stable imaging results even in low-temperature, high-speed flow environments.

[0056] It's important to note that during imaging, the direction of the light source determines the brightness and shadow distribution of the object. If the laser illumination direction aligns with the camera's shooting direction (i.e., a coaxial arrangement), light reflected from the debris surface can directly enter the camera sensor, resulting in a clear, high-contrast bright line in the image. If the laser illumination angle deviates from the camera's optical axis, parts of the debris surface may not be effectively illuminated, resulting in reduced image contrast and unstable smear signals.

[0057] Therefore, in this solution, the laser is irradiated from the same direction as the camera, maximizing light reflection from the debris surface and forming a high-brightness straight-line trailing image. If the laser is parallel to the liquid inlet direction, the debris may be in the shadow area at certain moments, resulting in uneven trailing image brightness, or even undetectable debris. For example, the trailing image may produce a brightness gradient due to uneven light intensity, affecting the convolutional neural network's detection of the bright line direction. If the laser is perpendicular to the liquid inlet direction, all debris can obtain the same lighting conditions, the trailing image brightness is more uniform, and can effectively enhance the recognition ability of the direction-sensitive convolutional neural network.

[0058] Optionally, in this embodiment, the camera is a frost-resistant camera. Frost-resistant cameras employ various anti-frost, anti-fog, and low-temperature protection technologies to ensure proper operation in extreme environments. For example, micro-heating elements on the lens surface keep the lens temperature slightly above the ambient temperature (e.g., maintained at 1°C-3°C) to prevent condensation and the formation of a frost layer. Alternatively, a special hydrophobic nano-coating is applied to the lens surface, making it difficult for moisture to adhere, reducing the likelihood of fog and frost formation. Frost on the lens can cause blurred images or detection failures.

[0059] Specifically, in a certain embodiment, S1 includes the following steps S11-S13.

[0060] S11. Acquire an image of the liquid inlet channel and perform preprocessing, wherein the camera is set with a preset shooting step length and a preset exposure time when shooting.

[0061] Specifically, in a certain embodiment, S11 includes the following steps S111-S112.

[0062] S111. Perform median filtering on the image to remove salt and pepper noise and preserve edge structure.

[0063] Median filtering is used to remove salt and pepper noise from images, ensuring that the edge structure of the smear is not disturbed. Salt and pepper noise is a randomly distributed black and white noise, often caused by sensor defects, signal interference, or data transmission errors. In an image, this noise appears as isolated extreme pixel values, causing localized black or white spots in the image, which affects subsequent feature extraction. To address this issue, median filtering is used for noise reduction. The core idea of ​​median filtering is to sort the neighborhood window of each pixel and replace the original pixel value with the median value, effectively removing salt and pepper noise while preserving edge details in the image.

[0064] S112. Perform background modeling and foreground extraction processing on the image to extract the dynamically changing part.

[0065] After noise reduction, S112 performs background modeling and foreground extraction to extract moving debris from a complex environment. Since the ambient lighting in the inlet channel of the cryogenic liquid cargo pump may change, or there is fixed structure reflection in the camera's perspective, the background needs to be dynamically updated. The basic idea of ​​background modeling is to establish a background model. , and calculate the current frame image Difference from the background model To extract the foreground: .

[0066] Among them, if If the value is greater than the set threshold T, the pixel is considered a moving target, likely a smear of debris; otherwise, it is considered part of the background. Background modeling can use a mixed Gaussian model. Its basic principle is to model the historical values ​​of each pixel using a multi-Gaussian distribution to adapt to lighting changes and complex background interference. For example, if the background brightness fluctuates over time (such as slight changes in reflectivity caused by pump vibration), a single threshold method may mistakenly interpret these changes as foreground. However, MOG automatically adjusts the background model by statistically analyzing the distribution of background pixels over different time periods, thereby reducing false detections.

[0067] Optionally, the S11 can also use the Radon transform to enhance the line features of the image, making the directional information of the smear more prominent. The Radon transform is a method that projects a two-dimensional image into the angular domain. It can convert the line structure in the image into projected curves in different directions, thereby detecting significant straight line patterns in the image.

[0068] S12. Input the image into a pre-trained convolutional neural network, wherein the convolution kernel of the convolutional neural network is used to identify straight line features.

[0069] Specifically, in one embodiment, the convolutional neural network training steps in S12 are S121-S128.

[0070] S121. Obtain pre-trained image materials, wherein the pre-trained image materials include images of the liquid inlet channel operating normally, images of the liquid inlet channel with smear and marked smear angles, and artificially simulated smear data of smear caused by simulating delayed exposure through different camera parameter settings.

[0071] In this step, a pre-training dataset is constructed. This involves capturing images of the liquid inlet channel under normal operating conditions, labeled images with known smear directions, and artificially simulated smear data. This diversity ensures that the convolutional neural network can learn smear morphologies under different conditions. For example, when collecting real smear data, varying lighting intensities and smear clarity may occur. Therefore, sufficient samples must be collected and tested under various lighting conditions.

[0072] S122. Perform data preprocessing on the pre-training image material and form a pre-training set.

[0073] First, all images need to be normalized to reduce the impact of lighting variations on training results. Typically, the pixel values ​​of the input images are normalized to between [-1, 1] or [0, 1] to improve the convergence speed of the convolutional neural network. Next, the images are resized to ensure that all samples have a consistent input size, typically set to 256*256 or 512*512 pixels. Optionally, additional data augmentation can be performed, such as random cropping, color transformation, and Gaussian blurring, to prevent model overfitting.

[0074] S123. Evenly discretize 0°-180° into 12 main directions to cover possible smear angles.

[0075] S124. Initialize 16 3×3 convolution kernels as direction-sensitive convolution kernels, ensuring that there is at least one filter for each main direction, where each direction corresponds to the straight line detection capability of a specific angle, and the parameters of the convolution kernel are fixed after being generated by Gaussian weighted line segments.

[0076] Initialize the direction-sensitive convolution kernels and evenly discretize the angles from 0° to 180° into 12 principal directions. Since smear appears as straight lines at varying angles in an image, convolutional neural networks (CNNs) are not sufficiently sensitive to directional features. For example, the VGG16 model only achieves 72% accuracy in recognizing 45° diagonal lines (versus >95% for horizontal and vertical lines). Therefore, CNNs must be directionally aware. To this end, the first convolution kernel layer of the CNN uses 16 3×3 convolution kernels, covering 12 different directions, with at least one filter for each direction. The weights of these convolution kernels are not randomly initialized, but rather parameterized using Gaussian weighted line segments, making the filters more sensitive to linear structures in specific directions. For example, the response pattern of a 90° convolution kernel is primarily concentrated in the vertical direction, while the response pattern of a 45° convolution kernel is more biased towards diagonal lines. Multiple convolution kernels can be used for specific directions.

[0077] S125. Output the direction-sensitive convolution kernel and calculate the variance of the directional response at the same position. If the variance is too large, increase the penalty to reduce directional confusion.

[0078] In this step, the convolutional neural network calculates the directional response variance and applies a penalty term to reduce the simultaneous activation of multiple directions. The directional response variance is calculated as follows:

[0079]

[0080] Among them, V(x, y) is the directional response variance, Indicates the location The response value in the i-th direction, represents the average response value across all directions, where N is the number of directions. By adding a variance penalty, the convolutional neural network is forced to focus on the most prominent direction, rather than responding to multiple directions simultaneously. This helps improve direction recognition accuracy and reduces the likelihood of misclassification. The core purpose of this step is to prevent the convolutional neural network from making ambiguous decisions when detecting smears, ensuring that the directional response of each pixel is as concentrated as possible in a single primary direction, rather than being activated in multiple directions simultaneously.

[0081] S126. Batch input the pre-training set into the convolutional neural network for forward propagation.

[0082] S127. Perform loss calculation, wherein the loss function includes a classification cross entropy loss for measuring whether the smear direction classification is correct and a directional response variance penalty for suppressing simultaneous activation of multiple directions.

[0083] Data is fed into the convolutional neural network in batches for forward propagation and loss calculation. Forward propagation is the core step in convolutional neural network training, where each input image passes through multiple convolutional layers, pooling layers, and fully connected layers, ultimately outputting a prediction of the smear direction. During loss calculation, categorical cross-entropy loss is used to measure the error between the predicted and true directions. A directional response variance penalty is also incorporated to ensure clearer directional predictions from the convolutional neural network. The loss function can be expressed as:

[0084]

[0085] Where L is the loss function, is the cross entropy loss, V is the direction response variance, This design allows the convolutional neural network to not only focus on classification correctness during optimization, but also ensure that the directional response is as clear as possible, thereby improving detection accuracy.

[0086] S128. Perform backpropagation and optimization, wherein the first layer of direction-sensitive convolution kernels is not updated, and the parameters of subsequent convolution and fully connected layers are updated to optimize the direction classification capability.

[0087] In this step, the convolutional neural network undergoes backpropagation and optimization, adjusting network weights to improve classification capabilities. During the optimization process, the weights of the first layer's direction-sensitive convolution kernels remain fixed, while the parameters of subsequent convolutional and fully connected layers are updated based on error backpropagation. The optimization algorithm typically uses Adam or stochastic gradient descent to accelerate network convergence and improve training stability.

[0088] Specifically, in one embodiment, the convolutional neural network in S12 includes a direction-sensitive convolution layer, a direction-responsive variance calculation, subsequent convolution and pooling layers, a fully connected layer, and a loss function.

[0089] First, 16 3×3 direction-sensitive convolution kernels are initialized in the direction-sensitive convolution layer, covering 12 principal directions (uniformly discretized from 0° to 180°). These kernels are designed to detect line features in different directions within the image, enabling the convolutional neural network to locate possible debris tracks within the input image. Because debris smear typically appears as long, thin bright lines, their direction can be affected by factors such as debris flow velocity, liquid turbulence, and illumination angle. Therefore, the convolutional neural network requires kernels that are sensitive to specific directions to ensure the model can effectively extract line information when detecting debris. For example, for a 90° smear, the kernel might focus on detecting pixel gradient changes in the vertical direction, while for a 45° smear, the kernel would focus more on the diagonal brightness distribution. Convolution kernels for different directions can be generated using Gaussian-weighted line segments. Specifically, within a 3×3 window, a filter is generated along a specific direction based on a Gaussian weight distribution, making it more responsive to features in that direction.

[0090] For example, in a 90° convolution kernel, the filter has the following structure:

[0091]

[0092] This convolution kernel can emphasize vertical edge features while ignoring horizontal or diagonal interference, thereby improving the convolutional neural network's ability to detect vertical smears. Due to this direction-sensitive design, the first convolution layer output of a CNN typically generates responses in multiple directions, some of which may be more significant than others. Therefore, multiple convolution kernels can be set to correspond to these directions.

[0093] The system introduces a variance constraint on the directional response during the directional response variance calculation step. Specifically, this is done by calculating the activation values ​​of the same pixel in different directions and constraining their variances to avoid directional confusion caused by excessively high responses in multiple directions.

[0094] After the direction-sensitive convolution layer and directional variance optimization, the convolutional neural network continues to extract higher-level features through subsequent convolution and pooling layers. At this stage, the network uses a multi-layer standard convolutional neural network structure, including 3×3 or 5×5 ordinary convolution layers to extract more complex local features, and combines them with maximum pooling layers to reduce computational complexity while improving robustness to local noise. The role of maximum pooling is to retain the strongest response in the feature map to reduce information loss. For example, in a 2×2 pooling window, suppose the feature map pixel value of a certain area is:

[0095]

[0096] The output of the maximum pooling is 8 because it is the strongest response value in this area, which can retain the most significant features and reduce the amount of computation. In this way, the convolutional neural network can further learn more advanced smear patterns while ensuring that the output features have stronger directionality and stability.

[0097] At the end of the network, after processing by the fully connected layer, all features extracted by the convolutional neural network are mapped to a final classification output of the smear direction, that is, predicting which of the 12 directions the smear in the current image belongs to. The role of the fully connected layer is to integrate all convolutional features to generate a high-dimensional feature representation, which is ultimately converted into a discrete category label. For example, the model may output:

[0098]

[0099] The category corresponding to 0.6 is the direction of the smear predicted by the convolutional neural network. For example, if 0.6 corresponds to 45°, the model will determine that the current smear direction is 45°, and the system can trigger subsequent control logic accordingly, such as adjusting the operating status of the pump or activating the pressure relief mechanism.

[0100] In order to optimize the training effect, the convolutional neural network adopts classification cross entropy loss and direction response variance penalty as the joint loss function.

[0101] S13. Obtain the detection result output by the convolutional neural network.

[0102] After forward propagation calculation, the convolutional neural network finally outputs a probability distribution about the direction of the smear, which is calculated by the Softmax activation function. Each value represents the probability that the input image belongs to a different direction category.

[0103] S2. Perform ultrasonic detection on the liquid inlet channel of the cryogenic liquid cargo pump and, based on the detection results, determine whether the object forming a bright straight line trajectory on the image is a gas or a solid.

[0104] During implementation, ultrasonic detection utilizes the principle of ultrasonic echo reflection, that is, ultrasonic waves of a certain frequency are emitted into the liquid inlet channel, and the properties of the target object are analyzed based on the characteristics of the echo signal. When ultrasonic waves encounter different media, part of the sound wave will be reflected, while the other part will be transmitted or absorbed. For solid impurities, since their density and acoustic impedance are much greater than those of liquids, the reflection coefficient of ultrasonic waves at the interface is higher, so the echo signal is stronger, usually manifesting as a high-amplitude, high-frequency echo pattern. For bubbles, since their density is much lower than that of liquids, ultrasonic waves will be strongly scattered and attenuated after entering the bubbles, resulting in a weak echo signal and a low-frequency oscillation characteristic. Therefore, by analyzing parameters such as the echo amplitude, attenuation characteristics, and time delay, it is possible to effectively distinguish whether the target object is solid debris or bubbles.

[0105] S3. Control whether the pressure relief valve of the cryogenic liquid cargo pump is activated based on the detection results.

[0106] The system first compares the ultrasonic detection results with the convolutional neural network's smear detection results to confirm whether the detected debris is indeed solid foreign matter. If the debris meets the ultrasonic echo intensity threshold and the convolutional neural network detects a significant smear in the corresponding direction, the system preliminarily determines the presence of solid foreign matter that could affect pump operation. In this case, the pressure relief valve does not immediately initiate emergency protection, but instead enters a standby state, effectively preparing for possible debris impacts and minimizing delays in subsequent protective measures.

[0107] At this stage, the system uses a time window decision mechanism to track the movement trajectory of debris and determine its potential impact on the system. Due to the complex fluid environment of the cryogenic liquid cargo pump, the movement trajectory of debris may be affected by factors such as flow rate and pump operation status. Therefore, the system will not take direct measures based on a single debris detection result, but will continuously collect multiple frames of data to determine whether the debris has a stable trajectory pattern. For example, in the time window The system performs a weighted calculation on the debris detection directions of the past n frames to confirm its continued existence and determine whether it is likely to enter the pump cavity: .

[0108] in, is the weighted direction of debris movement, is the detection direction of the i-th frame, is the time decay weight, and the newer detection data has a larger weight.

[0109] Alternatively, in another embodiment, the mode of the debris detection direction in the time window can be taken, and the difference between the number of data corresponding to the mode and the number of other data can be determined to determine whether a false detection has occurred, or to determine the trajectory stability. If the ratio of the number of data corresponding to the mode to the number of other data is greater than a predetermined threshold, the mode is used as the debris movement direction. For example, if the detection direction of the past 5 frames is [45°, 46°, 44°, 45°, 45°], then 45° is the mode, and there are 3 45° and 2 other data. If the ratio of the two is greater than or equal to the preset threshold of 1.5, it can be determined that the stability of the debris in the 45° direction is high and it is not a single false detection, so the pressure relief valve can be triggered. If the detection direction changes greatly, such as [45°, 30°, 50°, 20°, 10°], it may be an optical error or turbulent disturbance. At this time, the system will not immediately trigger the pressure relief, but will enter a delayed monitoring state to reduce the risk of false triggering.

[0110] After confirming the presence of debris and trajectory stability, the system will enter the pressure relief standby mode, which controls the pressure relief valve to enter a fast-actuating state so that immediate action can be taken when necessary. At this time, the pressure relief valve will not be fully opened, but will be in a low-power standby mode and may perform small pressure adjustments to ensure the pressure of the entire system is stable. For example, if the current pressure of the pipeline is , the system may first perform a fine-tuning decompression to release a small pressure change To observe the system's response:

[0111]

[0112] in, is the adjusted pressure, Typically, a small safety value is set to prevent excessive pressure relief from impacting normal pumping operations. This small pressure relief not only prepares for subsequent emergency protection actions but also serves as a proactive test to observe whether debris is being displaced or settled by changes in fluid pressure. If the movement pattern of debris changes significantly after fine-tuning the pressure relief, such as being dispersed by the fluid or settling to a safe area, the system can suspend further action and resume normal monitoring mode.

[0113] S4. Obtain a subsequent detection signal, and control a subsequent protection action after the pressure relief valve is activated based on the subsequent detection signal.

[0114] When debris enters the pump chamber, it comes into contact with the impeller or pump body, which may increase the impeller load, change the power demand of the motor, and cause instantaneous current fluctuations. This phenomenon can be detected by the current monitoring module and the threshold value can be set. When the current changes beyond this threshold, it is determined that the pump may be affected by debris:

[0115]

[0116] in, is the current motor current, is the current reference value during normal operation. , it may indicate that the pump chamber is disturbed by debris. In addition, debris may cause the impeller to increase the resistance to rotation, causing the speed to Descending, the system can calculate the slip rate through the speed sensor:

[0117]

[0118] If the slip ratio S is too high, it means that the impeller is subject to greater resistance, further proving that debris may have an impact on the pump. At the same time, the pump body may increase in temperature due to friction from debris or abnormal load. The temperature change of the pump body can be calculated by the following formula:

[0119]

[0120] like , indicating that the friction heat inside the pump chamber has increased significantly, indicating that the debris may have caused friction damage to the pump impeller or pump wall.

[0121] After acquiring these abnormal signals, the system adopts a graded response mechanism to ensure that the pump takes appropriate protective actions under different impact levels:

[0122] 1. Low risk (automatically adjust operating parameters)

[0123] If a slight current fluctuation is detected (such as If the pump temperature and vibration levels remain unchanged (below a set threshold), the system will adjust operating parameters, reducing pump speed and optimizing pressure to minimize the impact of debris on the impeller. Meanwhile, the system will maintain monitoring and wait for the next round of data collection to confirm whether debris is still present.

[0124] 2. Medium risk (partial load reduction + triggered pressure relief)

[0125] If increased current fluctuations accompanied by slight vibration abnormalities are detected, debris may have impacted the pump impeller but has not yet caused serious damage. At this point, the system enters partial load shedding mode, reducing the pump's operating power while triggering the pressure relief valve to release some pressure in an attempt to change the fluid dynamics and dislodge the debris from the impeller. The system closely monitors current, temperature, and vibration trends. If the situation improves, it maintains this mode; otherwise, it enters a higher level of protection.

[0126] 3. High risk (speed limit operation + backwash mechanism)

[0127] If debris causes significant current surges, temperature rises, or pump vibrations, it indicates that the debris may be lodged in the impeller or flow path, affecting normal pump operation. At this point, the system will enter speed-limited operation mode, reducing the speed to a safe range and attempting to trigger the backwash mechanism, which briefly reverses the flow direction of the pump in an attempt to expel the debris from the impeller. The backwash control logic is as follows:

[0128]

[0129] in, is the normal operating pressure, is the reverse flushing pressure, The backwash coefficient is usually set between 0.8 and 1.2 to ensure that the reverse impact force is strong enough but does not damage the pump structure. If the current returns to normal, the temperature drops, and the vibration disappears after backwashing, it means that the debris has been successfully discharged and the system has resumed normal operation.

[0130] 4. Emergency response (shutdown protection)

[0131] If debris causes severe vibration, excessive current fluctuations, and a sharp temperature rise, and all mitigation measures fail, the system enters emergency shutdown mode, immediately closing the fluid inlet and deactivating the pump power source to prevent further damage. The system also logs complete fault data and notifies maintenance personnel for inspection. This mode is typically triggered only in severe cases, such as when debris causes impeller blockage or bearing damage, requiring manual intervention.

[0132] The various processes and procedures described above, such as a cryogenic liquid cargo pump control and intelligent diagnostic method, can be executed by a CPU. For example, in some embodiments, a cryogenic liquid cargo pump control and intelligent diagnostic method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as a memory. In some embodiments, part or all of the computer program can be loaded and / or installed onto a computing device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, the cryogenic liquid cargo pump control and intelligent diagnostic method described above can be implemented.

[0133] The present disclosure relates to methods, computing devices, computer-readable storage media, and / or computer program products. The computer program products may include computer-readable program instructions for executing various aspects of the present disclosure.

[0134] A computer-readable storage medium can be a tangible device that can hold and store instructions for use by an instruction execution device. A computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or raised-in-groove structure on which instructions are stored, and any suitable combination thereof. As used herein, a computer-readable storage medium is not to be construed as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through an electrical wire.

[0135] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge computing devices. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0136] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, the state information of the computer-readable program instructions is used to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), so that the electronic circuit can execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

[0137] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine such that, when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device is generated that implements the functions / actions specified in one or more blocks in the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium, where these instructions cause the computer, programmable data processing device, and / or other device to operate in a specific manner, so that the computer-readable medium storing the instructions comprises an article of manufacture, which includes instructions for implementing various aspects of a multi-sensor fusion-based intelligent cold source disaster monitoring method or multiple actions.

[0138] Computer-readable program instructions can also be loaded onto a computer, other programmable data processing device, or other device, so that a series of operating steps are executed on the computer, other programmable data processing device, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable data processing device, or other device implement an intelligent cold source disaster-causing object monitoring method or multiple actions based on multi-sensor fusion.

[0139] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable other persons skilled in the art to understand the embodiments disclosed herein.

[0140] Based on the same inventive concept, an embodiment of the present invention provides an intelligent terminal, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute a method.

Claims

1. A cryogenic liquid cargo pump control and intelligent diagnosis method, characterized in that: The following steps are involved: S1. Acquire an image of the liquid inlet channel of a cryogenic liquid cargo pump and identify a bright straight line track in the image, wherein an ultrasonic detector and a camera are installed on the liquid inlet channel of the cryogenic liquid cargo pump; S2. Perform ultrasonic detection on the liquid inlet channel of the cryogenic liquid cargo pump and, based on the detection results, determine whether the object forming a bright straight line on the image is a gas or a solid; S3. Control whether the pressure relief valve of the cryogenic liquid cargo pump is activated based on the detection results; S4 obtains subsequent detection signals, and controls subsequent protection actions after the pressure relief valve is activated based on the subsequent detection signals; Said S1 comprises the following steps: S11. Acquire an image of the liquid inlet channel and perform preprocessing, wherein the camera is provided with a preset shooting step and a preset exposure time when shooting; S12. Inputting the image into a pre-trained convolutional neural network, wherein the convolution kernel of the convolutional neural network is used to identify straight line features; S13. Obtaining the detection result output by the convolutional neural network; The training steps of the convolutional neural network in S12 are: S121. Obtain pre-trained image materials, wherein the pre-trained image materials include images of a liquid inlet channel operating normally, images of a liquid inlet channel with smear and smear angles marked, and artificially simulated smear data caused by simulating delayed exposure by setting different camera parameters; S122. Preprocess the pre-training image material and form a pre-training set; S123. Evenly discretize 0°-180° into 12 main directions to cover possible smear angles; S124. Initialize 16 3×3 convolution kernels as direction-sensitive convolution kernels, ensuring that there is at least one filter for each main direction, where each direction corresponds to the line detection capability of a specific angle, and the parameters of the convolution kernels are fixed after being generated by Gaussian weighted line segments; S125. Output the direction-sensitive convolution kernel and calculate the variance of the directional response at the same position. If the variance is too large, increase the penalty to reduce directional confusion. S126. Batch input the pre-training set into the convolutional neural network for forward propagation; S127. Calculate the loss, wherein the loss function includes a classification cross entropy loss for measuring whether the smear direction classification is correct and a directional response variance penalty for suppressing simultaneous activation of multiple directions; S128. Perform backpropagation and optimization, wherein the first layer of direction-sensitive convolution kernels is not updated, and the parameters of subsequent convolution and fully connected layers are updated to optimize the direction classification capability.

2. The cryogenic liquid cargo pump control and intelligent diagnosis method according to claim 1, characterized in that: The S11 includes the following steps: S111. Perform median filtering on the image to remove salt and pepper noise and preserve edge structure; S112. Perform background modeling and foreground extraction processing on the image to extract the dynamically changing part.

3. The cryogenic liquid cargo pump control and intelligent diagnosis method according to claim 2, characterized in that: The convolutional neural network in S12 includes: The direction-sensitive convolution layer sets 16 3×3 direction-sensitive convolution kernels, covering 12 main directions; Directional response variance calculation, which is used to calculate the variance of responses in different directions and add a penalty term to force the network to select the most significant direction; Subsequent convolution and pooling layers are used to extract high-level features; The fully connected layer is used to extract the final smear features and perform classification; The loss function is the classification cross entropy loss and the directional response variance penalty.

4. The cryogenic liquid cargo pump control and intelligent diagnosis method according to claim 3, characterized in that: In the S1 , when the camera is exposed, a laser pulse is synchronously triggered to illuminate the interior of the liquid inlet channel of the cryogenic liquid cargo pump, wherein the pulse irradiation direction is the same as the shooting direction and is perpendicular to the liquid inlet direction.

5. The cryogenic liquid cargo pump control and intelligent diagnosis method according to claim 4, characterized in that: The camera is a frost-resistant camera.

6. The cryogenic liquid cargo pump control and intelligent diagnosis method according to claim 5, characterized in that: The laser pulses emitted by the camera are longer than 100 μs, so that moving debris will form bright straight line tracks in the image.

7. A cryogenic liquid cargo pump control and intelligent diagnosis system, characterized in that: The invention comprises a processor, wherein a program of the cryogenic liquid cargo pump control and intelligent diagnosis method according to any one of claims 1 to 6 is run in the processor.

8. A storage medium, characterized in that: A program for the cryogenic liquid cargo pump control and intelligent diagnosis method according to any one of claims 1 to 6 is stored.

Images