A low-temperature ball valve opening degree self-adaptive control method based on deep learning

Abstract

The present application relates to the technical field of opening degree control, in particular to a low-temperature ball valve opening degree self-adaptive control method based on deep learning, which comprises receiving opening degree adjustment request and generating target opening degree instruction along opening degree track, obtaining time sequence operation data containing current sequence and actual opening degree displacement sequence, and simultaneously calculating dynamic tracking error; inputting pre-trained deep learning perception model to output sediment jam index of current time step in real time; performing opening degree self-adaptive control based on comparison result of the index and preset threshold, and respectively executing additional disturbance amplitude or local track instruction containing reverse rollback and positive high acceleration impact according to different jam grades. The present application effectively overcomes the problem of control response lag under extremely low-temperature complex environment, realizes dynamic online perception, and actively removes solid-state condensation obstacles through multi-stage intervention.

Description

Technical Field

[0001] This invention relates to the field of opening control technology, specifically to a deep learning-based adaptive control method for the opening of a cryogenic ball valve. Background Technology

[0002] In modern petrochemical processes (such as liquefied natural gas and cryogenic ethylene separation), cryogenic ball valves are crucial control components in fluid pipelines. Precise control of their opening degree directly affects the stable operation and energy efficiency of the entire process system. In extremely low-temperature environments (such as -196°C), the physical properties of the fluid medium change complexly. Traditional opening degree control methods based on conventional PID or preset experience often suffer from response lag and low control accuracy, making them unsuitable for complex and variable cryogenic conditions.

[0003] During long-term low-temperature operation, petrochemical media are prone to condensation and crystallization inside the ball valve (such as in the valve seat and flow channel), forming solid deposits. This not only damages the tight fit between the ball and the valve seat, causing internal leakage during shutdown and serious safety accidents, but also significantly increases the valve's opening and closing torque, causing jamming. Ultimately, this results in the actuator being unable to accurately reach its position, causing fluid control distortion. Currently, the petrochemical industry primarily relies on traditional methods to address seal failures caused by such deposits, such as periodic shutdowns for dismantling and cleaning or passively waiting for repairs after a malfunction. Although some equipment has incorporated simple online monitoring, it cannot integrate real-time data to proactively sense the state of the deposits and implement adaptive interventions. To address this, a deep learning-based adaptive control method for the opening of cryogenic ball valves is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based adaptive control method for the opening of a cryogenic ball valve, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based adaptive control method for the opening degree of a cryogenic ball valve includes: Receive an opening adjustment request, acquire the opening trajectory of the target cryogenic ball valve, and generate a target opening command along the opening trajectory; acquire the timing operation data of the target cryogenic ball valve; the timing operation data includes a current sequence and an actual opening displacement sequence; calculate the dynamic tracking error of the actual opening displacement sequence relative to the target opening command; The current sequence and dynamic tracking error are input into a pre-trained deep learning perception model, which outputs the sediment caching index at the current time step. Based on the comparison between the sediment trapping index and a preset threshold, adaptive opening control is performed. When the target opening command is less than or equal to the first threshold, it is output. When the target opening command is between the first and second thresholds, the corresponding disturbance amplitude is extracted, and a disturbance period including a sudden change window and a recovery window is set. Within the sudden change window, a disturbance amplitude is added to the target opening command. Within the recovery window, the added disturbance amplitude is gradually reduced to zero over time steps, and the target opening command is updated and output in a loop. When the target opening command is greater than or equal to the second threshold, the current trajectory is interrupted, and local trajectory commands for reverse retreat and forward high acceleration impact are generated sequentially. After the trapping index falls back, the target opening command is interpolated and transitioned to the opening trajectory.

[0006] Preferably, the opening adjustment request is a control signal that includes a target opening value and a desired adjustment time parameter, the opening trajectory is a smooth desired angular displacement curve that changes continuously with time based on the current actual valve opening, the target opening value, and the desired adjustment time, and the target opening command is a sequence of expected opening angular displacements corresponding to each control time step generated by discretizing and sampling the opening trajectory.

[0007] Preferably, the dynamic tracking error is calculated point by point under a unified control time step, by calculating the real-time position deviation between the expected opening value corresponding to the target opening command and the actual opening value in the actual opening displacement sequence; calculating the rate of change of the real-time position deviation between adjacent control time steps to obtain the speed deviation characterizing the valve motion hysteresis; and fusing the real-time position deviation and the speed deviation in the time dimension to generate the dynamic tracking error.

[0008] Preferably, the deep learning perception model includes a feature extraction module, a cross-modal cross-attention module, and a temporal evolution module; The feature extraction module consists of two parallel one-dimensional convolutional networks, which independently receive the current sequence and the dynamic tracking error, and output the current response features and tracking error features respectively. The cross-modal cross-attention module linearly projects the tracking error features as a query matrix and linearly projects the current response features as a key matrix and a value matrix. It calculates the dynamic attention weights of the tracking error features on the current response features and uses the dynamic attention weights to perform weighted fusion on the current response features to generate abnormal friction resistance features. The time-series evolution module includes a gated loop unit and a multilayer perceptron, which receives the abnormal frictional resistance characteristics and captures the long-term dependent evolution trend, and outputs the sediment sticking index at the current time step through an activation function.

[0009] Preferably, the cross-modal cross-attention module introduces an opening position mask matrix. The specific process includes: dividing the entire valve stroke into a valve seat sealing area at both ends of the stroke and a free adjustment area between them; extracting the real-time opening value from the actual opening displacement sequence, and generating an opening position mask matrix with the same dimension as the product of the query matrix and the key matrix. When the real-time opening value is in the valve seat sealing area, the corresponding element of the opening position mask matrix is ​​set to a decay weight less than 1; when the real-time opening value is in the free adjustment area, the corresponding element is set to 1; the query matrix and the key matrix are multiplied by a dot product to obtain the original attention score matrix, and the original attention score matrix and the opening position mask matrix are multiplied by an element-wise Hadamard product, and the dynamic attention weights are output through a normalization function.

[0010] Preferably, the specific process for determining the mutation window and the recovery window is as follows: The mutation window is calculated by dividing the difference between the sediment scouring index and a first threshold by the difference between a second threshold and the first threshold to obtain a dynamic ratio; the dynamic ratio is multiplied by the total number of time steps of the disturbance period and rounded to obtain the mutation step number, which is a continuous time step interval with the start point of the disturbance period as the starting point and the length as the mutation step number; the recovery window is calculated by subtracting the mutation step number from the total number of time steps of the disturbance period to obtain the recovery step number, which is a continuous time step interval with the end point of the mutation window as the starting point and the length as the recovery step number.

[0011] Preferably, the first threshold is determined by obtaining the test current sequence and test tracking error of the target cryogenic ball valve operating under a sediment-free baseline state, inputting it into the deep learning perception model to generate a baseline Chase index sequence; calculating the mean and standard deviation of the baseline Chase index sequence, and using the sum of the mean and the standard deviation by a set multiple as the first threshold. The second threshold is determined by obtaining the limit current sequence and limit tracking error of the drive motor corresponding to the target cryogenic ball valve under the critical state of triggering overload protection, inputting them into the deep learning perception model to calculate the limit jamming index; calculating the ratio of the rated maximum continuous operating current of the drive motor to the transient stall current, and multiplying the limit jamming index by the ratio as the second threshold.

[0012] Preferably, the generation and interpolation transition process of the local trajectory instruction specifically includes: recording the actual opening displacement when the current trajectory is interrupted as the interruption point; using the interruption point as the starting point and combining it with the target retreat displacement, generating a reverse retreat local trajectory instruction; after the reverse retreat local trajectory instruction is executed, generating a positive high-acceleration impact local trajectory instruction towards the target breakthrough position greater than the interruption point position; during the execution of the positive high-acceleration impact local trajectory instruction, when the sediment sticking index falls back to less than or equal to a first threshold, using the actual opening and running speed at the current moment as the starting state and the expected opening and expected speed of the opening trajectory in the corresponding future time step as the ending state, using a polynomial interpolation algorithm to generate a smooth transition sequence and output it, until it merges into the opening trajectory.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application obtains time-series operating data containing actual opening displacement and current, and combines it with multi-dimensional dynamic tracking error input into a pre-trained deep learning perception model to output the sediment trapping index, which represents the internal crystallization state, in real time. This effectively overcomes the problem of response lag and regulation distortion of conventional PID control caused by complex changes in fluid properties under extremely low temperature environments, and realizes online dynamic perception of special working conditions.

[0014] 2. This application introduces an opening position mask matrix into the cross-modal attention module of the deep learning perception model, which directly maps the mechanical stroke characteristics of the valve seat sealing area and the free adjustment area unique to the cryogenic ball valve into the attenuation weights on the algorithm side. This accurately removes the characteristic interference caused by the inherent normal large friction force of the valve body in the sealing area, and realizes the effective distinction between conventional mechanical physical resistance and abnormal crystallization jamming state, ensuring the objectivity and accuracy of the state monitoring and diagnosis results.

[0015] 3. This application adopts a multi-level adaptive intervention strategy based on the sediment jamming index. When faced with mechanical obstruction caused by sediment in the flow channel, it can automatically trigger high-frequency additional disturbances, or cut into a local running trajectory containing reverse retreat and forward high acceleration impact when severely jammed. It actively breaks up and breaks through solid condensation barriers by using transient physical kinetic energy. This changes the traditional pipeline's reliance on periodic shutdown cleaning or passive fault repair, and ensures the self-healing capability and stable operation of the fluid control system under extreme conditions. Attached Figure Description

[0016] Fig. 1 This is a schematic diagram of a deep learning-based adaptive control method for the opening of a cryogenic ball valve. Fig. 2 This is a schematic diagram of the dynamic tracking error generation process according to an embodiment of the present invention; Fig. 3This is a schematic diagram of the adaptive control process when the cascade index is greater than the second threshold according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are 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.

[0018] Please see Figs. 1 to 3 This invention provides a deep learning-based adaptive control method for the opening degree of a cryogenic ball valve, the technical solution of which is as follows: A deep learning-based adaptive control method for the opening degree of a cryogenic ball valve includes: Receive an opening adjustment request, acquire the opening trajectory of the target cryogenic ball valve, and generate a target opening command along the opening trajectory; acquire the timing operation data of the target cryogenic ball valve; the timing operation data includes a current sequence and an actual opening displacement sequence; calculate the dynamic tracking error of the actual opening displacement sequence relative to the target opening command; The current sequence and dynamic tracking error are input into a pre-trained deep learning perception model, which outputs the sediment caching index at the current time step. Based on the comparison between the sediment trapping index and a preset threshold, adaptive opening control is performed. When the target opening command is less than or equal to the first threshold, it is output. When the target opening command is between the first and second thresholds, the corresponding disturbance amplitude is extracted, and a disturbance period including a sudden change window and a recovery window is set. Within the sudden change window, a disturbance amplitude is added to the target opening command. Within the recovery window, the added disturbance amplitude is gradually reduced to zero over time steps, and the target opening command is updated and output in a loop. When the target opening command is greater than or equal to the second threshold, the current trajectory is interrupted, and local trajectory commands for reverse retreat and forward high acceleration impact are generated sequentially. After the trapping index falls back, the target opening command is interpolated and transitioned to the opening trajectory.

[0019] Example 1: Receive an opening adjustment request, obtain the opening trajectory of the target cryogenic ball valve, and generate a target opening command along the opening trajectory; The opening adjustment request is a control signal that includes a target opening value and a desired adjustment time parameter. The opening trajectory is a smooth desired angular displacement curve that changes continuously with time, constructed based on the current actual valve opening, the target opening value, and the desired adjustment time. The target opening command is a sequence of expected opening angular displacements corresponding to each control time step, generated by discretizing and sampling the opening trajectory.

[0020] Specifically, the communication processing module of the control system receives a string of control signals sent by the host computer control station or the field operation terminal, parses the control signal through the communication protocol, and extracts two core driving parameters: the target opening value and the desired adjustment time parameter. In this embodiment, the mechanical full-opening displacement of the target cryogenic ball valve is 90°, the target opening value is 72°, and the desired adjustment time parameter is 5s. An absolute encoder, coaxially mounted on the top of the valve stem of the target cryogenic ball valve, collects and reads the initial position data of the valve as its current actual opening degree. In this embodiment, the current actual opening degree of the valve is preset to 0°. To avoid stress damage to the mechanical transmission components due to sudden acceleration changes during start-up and shutdown under cryogenic conditions, a fifth-order polynomial interpolation algorithm is used for trajectory planning. The continuous mathematical model of the smooth desired angular displacement curve with respect to time is defined as follows: ; in, Indicates at any given real-time moment The desired opening angle displacement of the target cryogenic ball valve; This represents the real-time running time variable elapsed since the opening adjustment request was received and the adjustment process was initiated, in units of [unit missing]. ; , , , , , These are all constant term coefficients to be solved in the fifth-order polynomial trajectory planning algorithm, used to shape the shape of the smooth desired angular displacement curve; Set the boundary conditions for the start-up time based on the extracted parameters: when time... =0 At this time, the displacement is equal to the current actual valve opening degree of 0°; at this time, the coefficient is used to determine the displacement. The initial position boundary of the representation is always equal to 0°. To ensure a smooth start, the coefficients are used. The velocity represented by the first derivative is equal to 0, determined by the coefficient. The characteristic acceleration (i.e., the second derivative) is equal to 0; the boundary condition at the arrival time is set: when time... When the desired adjustment time parameter is 5 seconds, the displacement equals the target opening value of 72°, the velocity is 0, and the acceleration is 0. The control system substitutes these six boundary conditions into the continuous mathematical model and its first and second derivative equations to form a system of six linear equations. The polynomial coefficients are then calculated using matrix inversion. to The unique numerical pair is used to solve for the smooth transition coefficient of the higher-order trajectory. , and A smooth expected angular displacement curve function that is continuous throughout the entire time domain and has continuous second derivatives is constructed.

[0021] Based on the operation cycle of the servo drive inside the actuator, the control time step span for discretization sampling is set. In this embodiment, the control time step is preset to 10ms. Discrete sampling is performed at equal intervals according to the 10ms time step, accumulating from time 0s until time 5s, generating a total of 500 discrete timestamps. The microprocessor executes loop calculation logic, sequentially substituting the specific time values ​​of these 500 discrete timestamps into the previously constructed smooth expected angular displacement curve function to calculate a specific opening angle value corresponding to each timestamp. This value is the expected opening angular displacement at that timestamp. The angle values ​​output from these 500 calculations are arranged and combined into a one-dimensional array according to the chronological order to generate a complete expected opening angular displacement sequence. In the real-time control process, every time the timer advances by a 10ms control time step, the control system extracts an expected opening angular displacement value from the expected opening angular displacement sequence according to the array index as the target opening command for the current time. By employing a fifth-order polynomial interpolation algorithm to construct a smooth desired angular displacement curve that continuously varies with time, and strictly constraining the velocity and acceleration at the start and end times to be zero, the sudden acceleration changes generated by the valve at the moment of opening and stopping are eliminated, effectively reducing the impact stress of mechanical transmission components caused by the cold brittleness of materials under extremely low temperature conditions. At the same time, the continuous theoretical trajectory is discretized and sampled at equal intervals into a sequence of expected opening angular displacements, so that the underlying microprocessor only needs to extract the target opening command through simple array indexing in the real-time control process, avoiding the need for real-time calculation of high-order functions in each short control cycle. See Fig. 2 The timing operation data of the target cryogenic ball valve is obtained; the timing operation data includes a current sequence and an actual opening displacement sequence; the dynamic tracking error of the actual opening displacement sequence relative to the target opening command is calculated. The dynamic tracking error is calculated point-by-point under a unified control time step, by determining the real-time position deviation between the expected opening value corresponding to the target opening command and the actual opening value in the actual opening displacement sequence; calculating the rate of change of the real-time position deviation between adjacent control time steps to obtain the speed deviation characterizing the valve motion hysteresis; and fusing the real-time position deviation and the speed deviation in the time dimension to generate the dynamic tracking error.

[0022] Specifically, in this embodiment, a uniform control time step is set. It takes 10ms; in the continuous run of the first... At each control time step, the controller extracts the expected opening value corresponding to the target opening command at that moment. Simultaneously, it receives the actual opening value from the actual opening displacement sequence synchronously collected by the valve displacement sensor. The real-time position deviation at this time step is calculated point by point using the difference operation. The calculation formula is as follows: ; in, Indicates the first Real-time position deviation under each control time step; In obtaining the current number Real-time position deviation of each control time step Then, retrieve the record from the system's data cache at the previous adjacent control time step, i.e., the [previous] control time step. Real-time position deviation of each control time step The first-order backward difference is calculated by taking the real-time position deviation between two adjacent control time steps and dividing it by the unified control time step size. The rate of change of the real-time position deviation is calculated as the velocity deviation characterizing the valve's motion hysteresis, and its calculation formula is as follows: ; in, Indicates the first The speed deviation of each control time step reflects the rate at which the valve's actual movement deviates from the expected trajectory under the influence of low temperature environment and deposits, and is expressed in degrees per second. The sliding time window used to extract time series features is set to a length of 50 control time steps to cover the valve action history within the past 0.5 seconds. After obtaining the real-time position deviation sequence and the velocity deviation sequence within the current time step and the historical window, preset position weight coefficients and velocity weight coefficients are introduced. In this embodiment, the preset position weight coefficient is 0.6, and the preset velocity weight coefficient is 0.4. For each discrete time step within the sliding time window, the entire range from the starting step of the sliding time window to the current time step is traversed, and the fusion node feature value is calculated point by point. The specific calculation method is as follows: for each discrete time step within the sliding time window, the feature value of the fusion node is calculated point by point. The position deviation sequence and velocity deviation sequence are normalized by range, and each is mapped to the dimensionless interval [0,1]. The real-time position deviation at a specific discrete time step is multiplied by the position weight coefficient to obtain a first value, and the velocity deviation at the same discrete time step is multiplied by the velocity weight coefficient to obtain a second value. The first value and the second value are then summed to obtain the feature value of the fusion node corresponding to the discrete time step. All the feature values ​​of the fusion node within the sliding time window are concatenated and stitched together in ascending order of time to construct a sequence vector with a dimension of 50×1 as the dynamic tracking error. By calculating the real-time positional deviation between the expected opening value and the actual opening value point by point, and further obtaining the rate of change of the deviation to obtain the velocity deviation, it is possible not only to accurately quantify the static tracking error of the valve opening at the current control time step, but also to effectively capture the dynamic hysteresis characteristics and error deterioration trend of valve movement under the physical obstruction of low temperature environment and sediment; furthermore, the real-time positional deviation and velocity deviation are deeply fused in the time dimension to generate dynamic tracking error, and a comprehensive feature sequence that takes into account the current spatial displacement state and the historical dynamic evolution trend is constructed. The current sequence and dynamic tracking error are input into a pre-trained deep learning perception model, which outputs the sediment caching index at the current time step. The deep learning perception model includes a feature extraction module, a cross-modal cross-attention module, and a temporal evolution module; The feature extraction module consists of two parallel one-dimensional convolutional networks, which independently receive the current sequence and dynamic tracking error, and output current response features and tracking error features respectively. The cross-modal cross-attention module linearly projects the tracking error features as a query matrix and linearly projects the current response features as a key matrix and a value matrix. It calculates the dynamic attention weights of the tracking error features on the current response features and uses these dynamic attention weights to perform weighted fusion of the current response features to generate abnormal friction resistance features. The temporal evolution module includes a gated recurrent unit and a multilayer perceptron, which receives the abnormal friction resistance features and captures long-term dependent evolution trends. It outputs the sediment sticking index at the current time step through an activation function.

[0023] The cross-modal cross-attention module introduces an opening position mask matrix. The specific process includes: dividing the entire valve stroke into a valve seat sealing area at both ends of the stroke and a free adjustment area between them; extracting the real-time opening value from the actual opening displacement sequence to generate an opening position mask matrix with the same dimension as the product of the query matrix and the key matrix. When the real-time opening value is in the valve seat sealing area, the corresponding element of the opening position mask matrix is ​​set to a decay weight less than 1; when the real-time opening value is in the free adjustment area, the corresponding element is set to 1; the query matrix and the key matrix are multiplied by a dot product to obtain the original attention score matrix, and the original attention score matrix and the opening position mask matrix are multiplied by an element-wise Hadamard product, and the dynamic attention weights are output through a normalization function.

[0024] Specifically, in this embodiment, the sliding time window length of the input data is set to 50. Since the dynamic tracking error and the current sequence are perfectly aligned in the time dimension, the input feature dimension of both is 50×1. The feature extraction module is configured with a first one-dimensional convolutional network and a second one-dimensional convolutional network. The first one-dimensional convolutional network is specifically used to process the current sequence. It contains one one-dimensional convolutional layer with a preset kernel size of 3, stride of 1, padding size of 1, and the number of kernels (i.e., the number of output channels) set to 64. When a current sequence with a dimension of 50×1 is input into it, after convolution, a matrix with a dimension of 50×64 is output as the current response feature. The second one-dimensional convolutional network is specifically used to process the dynamic tracking error. Its network structure and hyperparameters are completely consistent with the first one-dimensional convolutional network, i.e., a kernel size of 3, stride of 1, padding size of 1, and number of output channels of 64. When a dynamic tracking error with a dimension of 50×1 is input into it, a matrix with a dimension of 50×64 is output as the tracking error feature. The specific implementation process of the cross-modal cross-attention module, which linearly projects the tracking error features as a query matrix and the current response features as key and value matrices, is as follows: After obtaining the current response features and tracking error features, both with dimensions of 50×64, feature space projection is performed through three independent linear transformation layers. In this embodiment, the weight matrices of the three linear transformation layers are preset to have dimensions of 64×64; the tracking error features are then compared with the preset query weight matrix. Perform matrix multiplication to generate a query matrix with dimensions 50×64. The current response characteristics are respectively compared with a preset bond weight matrix. Sum weight matrix Performing matrix multiplication generates key matrices with dimensions of 50×64. Sum matrix ; The cross-modal cross-attention module introduces an opening position mask matrix; the entire valve stroke is divided into a valve seat sealing area at both ends of the stroke and a free adjustment area in between; the real-time opening value is extracted from the actual opening displacement sequence to generate an opening position mask matrix with the same dimension as the product of the query matrix and the key matrix. The specific implementation process is as follows: In this embodiment, the full stroke range of the target cryogenic ball valve is set to 0° to 90°. The 0° to 5° region (close to 0°) and the 85° to 90° region (close to 90°) together constitute the valve seat sealing area; the 5° to 85° region (between the two) is preset as the free adjustment area. The real-time opening value corresponding to 50 control time steps within the current sliding time window is extracted from the actual opening displacement sequence. Due to the query matrix AND key matrix The transposed product has a dimension of 50×50, therefore we construct an aperture position mask matrix with the same dimension of 50×50. For any element in the mask matrix row index and column indexes The values ​​of are all iterated from 1 to 50, and the assignment logic is as follows: Detect the th... The real-time opening value at each control time step, if the real-time opening value is within the valve seat sealing area, then the element... The value is assigned a preset attenuation weight, preferably within the range of [0.1, 0.5]. In this embodiment, the preset attenuation weight is 0.3; if the first... If the real-time opening value at each control time step is within the free adjustment range, then the element will be... The value is assigned a fixed standard weight, that is, set to 1; query matrix transpose of the key matrix Row matrix multiplication is performed, and the result is divided by a scaling factor. In this embodiment, the scaling factor is preset to 8 (i.e., the square root of the feature dimension 64), thus obtaining the original attention score matrix with a dimension of 50×50. The original attention score matrix With aperture position mask matrix Perform element-wise Hadamard product operations, that is, multiply the elements at corresponding positions directly to generate the corrected attention score matrix. The calculation formula is: ; in, The element-wise Hadamard product operator is represented; a normalized function is used to adjust the modified attention score matrix. Row-wise normalization is performed. In this embodiment, the normalization function is a flexible maximum function, and the output dimension is 50×50 for dynamic attention weights. ; Dynamic weights of attention With the value matrix containing current response characteristic information Perform matrix multiplication. Since the dynamic weights of the attention matrix have a dimension of 50×50, the value matrix... The matrix has a dimension of 50×64. The current response features are multiplied and weighted by tracking error, resulting in an abnormal friction resistance feature with a dimension of 50×64. Each time node in this matrix is ​​deeply integrated with the abnormal friction force features under the current action state.

[0025] The abnormal frictional resistance features are input into the time-series evolution module. In this embodiment, the hidden layer state dimension of the gated recurrent unit is set to 64. The abnormal frictional resistance features are sequentially processed by the gated recurrent unit along the time sequence of 50 control time steps to extract long-term dependencies and capture evolution trends. At the end of the last control time step, the final hidden state vector output by the gated recurrent unit is extracted, with a dimension of 1×64. The final hidden state vector is input into the multilayer perceptron. In this embodiment, the multilayer perceptron includes one fully connected hidden layer and one fully connected output layer. The number of neurons in the fully connected hidden layer is preset to 32, and the number of neurons in the fully connected output layer is preset to 1. After the state vector undergoes linear transformation and weight mapping by the multilayer perceptron, a one-dimensional scalar value is output. The scalar value is input into a preset activation function. In this embodiment, the activation function is the logistic activation function. After mapping by this activation function, the output range is strictly limited to between 0 and 1, outputting the sediment sludge index of the current time step. The pre-training process involves constructing a dataset for model training. Operating data of the target cryogenic ball valve is collected under cryogenic test conditions or historical operating conditions. Historical current sequences and historical dynamic tracking errors, each with a dimension of 50x1, are extracted as input feature samples. Simultaneously, the actual physical frictional resistance torque of the valve at the corresponding moment is measured, and this torque value is linearly mapped to a numerical range of 0 to 1 using a range normalization method. This torque value serves as the true label for the sediment sticking index. The input feature samples are then associated one-to-one with their corresponding true labels to construct a complete dataset. This dataset is then randomly divided into training and validation sets at an 8:2 ratio.

[0026] During the model initialization phase, the weights of various network nodes within the deep learning perceptual model are randomly initialized in memory. In this embodiment, the adaptive moment estimation optimization algorithm is preset for model training, the initial learning rate is preset to 0.001, the training batch size for each iteration is preset to 32 samples, the maximum number of training epochs is preset to 100 epochs, and the loss function used to measure prediction error is preset to mean squared error loss function.

[0027] After entering the iterative training phase, for each training round, 32 samples are randomly selected from the training set as a batch, and their corresponding historical current sequences and historical dynamic tracking errors are input into the deep learning perception model in parallel. Forward propagation calculation is performed according to the logic of the aforementioned feature extraction module, cross-modal cross-attention module, and temporal evolution module. Finally, the logistic activation function at the end of the multilayer perceptron outputs the predicted sediment kerat index.

[0028] The mean squared error loss function is used to calculate the squared difference between the 32 predicted sediment kerat indices and their corresponding true labels within each batch, and the average value is taken to obtain the global loss value for the current batch. Based on this loss value, backpropagation is initiated, and the partial derivative gradient of the loss value with respect to each learnable parameter within the model is calculated layer by layer according to the chain rule. The learnable parameters include the kernel weights of the one-dimensional convolutional network, the linear projection weight matrices in the cross-modal attention module, the state transition weights of the gated recurrent unit, and the node weights of the fully connected layers of the multilayer perceptron. After the calculation is completed, the optimization algorithm synchronously fine-tunes and updates all the above learnable parameters according to the gradient descent direction and the dynamically adjusted learning rate.

[0029] At the end of each training epoch, the system inputs all data from the validation set into the currently updated model, performs parameterless forward propagation, and calculates and records the average test loss value of the model on the validation set. This embodiment includes a pre-defined early stopping strategy: if the average test loss value on the validation set does not show a decreasing trend for 10 consecutive training epochs, the system will forcibly interrupt the training process to prevent overfitting of the network to the training data; if this condition is not triggered, the forward and backward propagation processes continue to cycle until the preset maximum number of training epochs (100 epochs) is reached, at which point the system automatically stops. After the training process is completed, the weight parameters of each network layer that currently exhibits the best performance on the validation set are solidified, encapsulated, and persistently saved to obtain the pre-trained deep learning perception model. By designing a deep learning perception model that incorporates feature extraction, cross-modal cross-attention, and temporal evolution, two parallel sets of one-dimensional convolutional networks are used to independently extract local low-level features of the current sequence and dynamic tracking error. Subsequently, through a cross-modal cross-attention mechanism, the tracking error is used as a guide to dynamically optimize and focus on local abrupt changes in the current response that are strongly correlated with abnormal friction, effectively removing background noise under normal operating conditions. Combined with the long-term memory capability of the gated recurrent unit, the evolution trend of weak jamming precursors is effectively captured, thus providing a more reliable feature representation system for the accurate output of sediment jamming index. A mask matrix for valve opening position is introduced into the cross-attention calculation, integrating prior knowledge of the valve's physical structure throughout its entire stroke into the deep learning model. By applying attention attenuation weights to the valve seat sealing area with high inherent friction and maintaining standard weights for the free adjustment area, the risk of the resistance generated by normal mechanical seal compression when the valve is close to the fully open or fully closed position being misjudged as deposit jamming is effectively suppressed. Based on the comparison between the sediment trapping index and a preset threshold, adaptive aperture control is performed. When the trapping index is less than or equal to the first threshold, a target aperture command is output. When the trapping index is between the first and second thresholds, the corresponding disturbance amplitude is extracted, and a disturbance period including a sudden change window and a recovery window is set. Within the sudden change window, a disturbance amplitude is added to the target aperture command. Within the recovery window, the added disturbance amplitude is gradually reduced to zero over time steps, and the target aperture command is updated and output in a loop. When the trapping index is greater than or equal to the second threshold, the current trajectory is interrupted, and local trajectory commands for reverse retreat and forward high acceleration impact are generated sequentially. After the trapping index falls back, the trajectory is interpolated to transition to the aperture trajectory. The specific process for determining the mutation window and the recovery window is as follows: The mutation window is calculated by dividing the difference between the sediment scouring index and a first threshold by the difference between a second threshold and the first threshold to obtain a dynamic ratio; the dynamic ratio is multiplied by the total number of time steps of the disturbance period and rounded to obtain the mutation step number, which is a continuous time step interval with the start point of the disturbance period as the starting point and the length as the mutation step number; the recovery window is calculated by subtracting the mutation step number from the total number of time steps of the disturbance period to obtain the recovery step number, which is a continuous time step interval with the end point of the mutation window as the starting point and the length as the recovery step number.

[0030] Specifically, the sediment sludge index output at the current time step is extracted. First, a subtraction operation is performed to calculate the difference between the sediment sludge index and the dynamically acquired first threshold. Then, the difference between the preset second threshold and the first threshold is calculated. Finally, a division operation is performed, dividing the former difference by the latter difference to obtain the dynamic ratio. The specific calculation formula logic is: the dynamic ratio equals (sediment sludge index minus the first threshold) divided by (second threshold minus the first threshold). The dynamic ratio accurately represents the relative position of the current sludge level within the set warning trigger interval.

[0031] The specific implementation process for calculating the number of mutation steps in the mutation window is as follows: The controller extracts the dynamic ratio calculated in the above steps, multiplies it with the preset total time steps of the disturbance period (set to 100 steps in this embodiment) to obtain the theoretical step length. Since the time steps of the underlying control system must be discrete integers, the controller performs a rounding operation on the multiplication result (specifically, a conventional rounding algorithm is used in this embodiment) to finally calculate the number of mutation steps in integer form.

[0032] The system takes the time step containing the starting point of the currently set disturbance period as the absolute starting point and then extracts a continuous time step interval on the time series with a length exactly equal to the number of mutation steps calculated above. The mutation window is the continuous interval from the starting point of the disturbance period to a number of time steps containing the mutation steps. Within this mutation window interval, the system will strictly execute the control action of adding the set disturbance amplitude to the target opening command.

[0033] The preset total number of time steps for the disturbance period (set to 100 steps in this embodiment) is called, and the integer number of mutation steps calculated in the previous steps is directly subtracted from it. The difference obtained by the subtraction operation is the recovery step number. The specific calculation formula logic is: the recovery step number equals the total number of time steps of the disturbance period minus the number of mutation steps.

[0034] The specific implementation process of defining the recovery window is as follows: the controller takes the next control time step immediately following the end point of the abrupt change window as the starting point, and truncates a continuous time step interval on the time series with a length exactly equal to the number of recovery steps calculated above. Specifically, the recovery window is the continuous interval from the next time step after the end of the abrupt change window to the time step containing the number of recovery steps (i.e., exactly reaching the end point of the total number of steps of the disturbance period).

[0035] By calculating the dynamic ratio of the sediment jamming index within a preset threshold range, the severity of valve jamming can be quantified in real time and accurately, and adaptively converted into a perturbation time allocation ratio. When the jamming is severe, the sudden change window step is automatically extended to apply more sustained perturbation energy, thereby effectively overcoming frictional resistance and breaking or scouring sediment. When the jamming is mild, the sudden change window is shortened and the recovery window is extended accordingly, allowing sufficient time for the additional perturbation amplitude to smoothly decrease to zero. This window allocation strategy based on real-time jamming severity ensures the effectiveness of valve unblocking while minimizing the impact vibration and fatigue wear caused by prolonged hard perturbation on the cryogenic ball valve's mechanical structure.

[0036] The first threshold is determined by obtaining the test current sequence and test tracking error of the target cryogenic ball valve under a sediment-free baseline state, inputting it into the deep learning perception model to generate a baseline Chase index sequence; calculating the mean and standard deviation of the baseline Chase index sequence, and using the sum of the mean and the standard deviation by a set multiple as the first threshold. The second threshold is determined by obtaining the limit current sequence and limit tracking error of the drive motor corresponding to the target cryogenic ball valve under the critical state of triggering overload protection, inputting them into the deep learning perception model to calculate the limit jamming index; calculating the ratio of the rated maximum continuous operating current of the drive motor to the transient stall current, and multiplying the limit jamming index by the ratio as the second threshold.

[0037] Specifically, the target cryogenic ball valve is first cleaned and defrosted, and then calibrated on a test bench at room temperature with no internal media resistance to ensure it is in a strictly deposit-free baseline state. In this baseline state, a standard opening adjustment request is sent to the target cryogenic ball valve, causing it to perform a complete action from fully closed to fully open and back to fully closed. During this operation, the operating current of the drive motor is continuously collected by the controller's built-in current sensor at a sampling frequency of 50Hz, forming the test current sequence containing time-series characteristics. Simultaneously, the actual opening displacement is synchronously collected by the absolute encoder on the valve stem, and the positional deviation from the target opening command is calculated point by point, forming the test tracking error.

[0038] The acquired test current sequence and test tracking error are timestamped and synchronously input into the trained deep learning perception model. The deep learning perception model uses its internal feature extraction module and cross-modal cross-attention module to perform forward propagation calculations, outputting the Chase index at each time step in the sediment-free baseline state. This series of consecutive Chase indices is spliced ​​and recorded in chronological order to generate the baseline Chase index sequence.

[0039] Statistical calculations are performed on all discrete values ​​in the benchmark Caesar index sequence to calculate the arithmetic mean of the entire sequence as the mean, and the root mean square deviation of the sequence data is calculated as the standard deviation. In this embodiment, the set factor in the claim is specifically set to 3, and the mean plus 3 times the standard deviation is used as the first threshold.

[0040] Artificially simulated external mechanical damping is applied to the drive motor of the target cryogenic ball valve on a hardware testing platform. By gradually tightening the simulated load device, the valve's operating resistance continuously increases, while the operating status of the drive motor is continuously monitored. When the motor's operating current climbs to the critical point that is about to trigger the internal thermal protection or overcurrent protection relay circuit of the motor driver, this state is maintained, and it is determined that the overload protection is in a critical state. In this critical state, data within a set time window (set to 2 consecutive seconds in this embodiment) is collected again through a current sensor and a position encoder to generate the limit current sequence and the limit tracking error, respectively. The limiting current sequence and the limiting tracking error are synchronously input into the deep learning perception model for inference calculation. At this time, because the input features contain extremely high current fluctuation characteristics and significant hysteresis tracking error characteristics, the deep learning perception model, after calculation, outputs an extremely high jamming index evaluation value characterizing the valve's near-complete jamming state. This value is recorded as the limiting jamming index. The manufacturer's technical manual for the drive motor corresponding to the target cryogenic ball valve is consulted and extracted. In this specific embodiment, the rated maximum continuous operating current parameter of the drive motor is set to 10A, and the transient stall current parameter of the drive motor in the state of complete rotor lockup is set to 25A. The rated maximum continuous operating current (10A) is divided by the transient stall current (25A) using the division unit inside the controller, and the ratio is calculated to be 0.4. The control system multiplies the limit Chase index output by the deep learning perception model with the obtained ratio, and writes the product result into a non-volatile memory as the second threshold. By setting the first threshold based on statistical regularities under baseline conditions, conventional friction and background noise are filtered out, effectively avoiding false triggering of adaptive control and improving the robustness of jamming detection. On the other hand, by setting the second threshold based on the limit data converted from the motor current ratio, a safety defense line with sufficient lead time is constructed to ensure that the interruption and backoff mechanism can be triggered in time before destructive stall occurs, thereby effectively preventing motor burnout or valve body damage. See Fig. 3 The generation and interpolation transition process of the local trajectory command specifically includes: recording the actual opening displacement when the current trajectory is interrupted as the interruption point; using the interruption point as the starting point and combining it with the target retreat displacement, generating a reverse retreat local trajectory command; after the reverse retreat local trajectory command is executed, generating a positive high-acceleration impact local trajectory command towards the target breakthrough position greater than the interruption point position; during the execution of the positive high-acceleration impact local trajectory command, when the sediment sticking index falls back to less than or equal to the first threshold, using the actual opening and running speed at the current moment as the starting state and the expected opening and expected speed of the opening trajectory in the corresponding future time step as the ending state, using a polynomial interpolation algorithm to generate a smooth transition sequence and output it until it merges into the opening trajectory; Specifically, the feedback data is read in real time, the currently executing target opening command is forcibly interrupted, the actual opening displacement of the target cryogenic ball valve at the moment of command interruption is recorded, and its value is stored in memory as the interruption point; In this embodiment, the target retraction displacement is preset to an opening displacement of 5°. Starting at a 45° opening, the calculated retraction endpoint is 40°. Based on this starting and ending points and the preset reverse running speed, discretization calculations are performed according to the control time step to generate an angular displacement sequence from the 45° opening position to the 40° opening position. This sequence is used as the local trajectory command for the reverse retraction and is issued for execution, causing the ball valve core to reverse and break free from the tightly locked state with the deposits. After the local trajectory command for the reverse retreat is executed, i.e., when the actual opening reaches 40°, the local trajectory command for the positive high-acceleration impact is generated towards the target breakthrough position, which is greater than the interruption point position. In this embodiment, the target breakthrough position is preset to be an opening displacement of 2° beyond the interruption point, i.e., a 47° opening position. The maximum allowable transient acceleration parameter set by the drive motor is read, and a positive motion trajectory is constructed with 40° opening as the starting point and 47° opening as the ending point, using the maximum allowable transient acceleration as the initial acceleration. Discrete sampling is performed according to the control time step to generate the local trajectory command for the positive high-acceleration impact and send it out for execution. The deposits at the valve seat are crushed or pushed away by the instantaneous high-kinetic-energy positive impact. During the execution of the local trajectory command for the positive high-acceleration impact, the sediment jamming index at the current time step is continuously received from the deep learning perception model. Once the sediment jamming index is detected to drop to less than or equal to the first threshold, it is determined that the jamming material has been effectively removed, the execution of the remaining high-acceleration impact commands is immediately stopped, and a smooth transition mechanism is triggered.

[0041] The actual opening degree and current running speed at the current moment are obtained, and these two are used as the starting state. At the same time, in the original opening degree trajectory, the future time step corresponding to a set time interval (in this embodiment, the preset time interval is 0.5 seconds) is extracted, and the expected opening degree and expected speed planned by the opening degree trajectory at the future time step are read, and these two are used as the ending state.

[0042] A smooth transition sequence is generated and output using a polynomial interpolation algorithm. Specifically, this embodiment employs a cubic polynomial interpolation algorithm, using the actual opening and running speed of the starting state, and the expected opening and expected speed of the ending state as four boundary conditions, and substituting them into a system of cubic polynomial equations to solve for the polynomial coefficients. Using the solved polynomials, point-by-point calculations are performed within the time interval between the current time and the future time step to generate a displacement sequence that smoothly connects the starting and ending points, serving as the smooth transition sequence. The smooth transition sequence is then output to the drive motor as a target opening command step by step until the transition sequence is completed, at which point the actual motion state of the target cryogenic ball valve merges back into the original opening trajectory, restoring the normal opening adjustment process.

[0043] By promptly interrupting the current trajectory and executing a reverse retraction when severe jamming is detected, the tight lock between the valve core and the deposits or ice layer is released, while providing necessary acceleration buffer space for subsequent obstacle clearing actions. The generated positive high-acceleration impact local trajectory can utilize the instantaneously accumulated high kinetic energy to generate a powerful impact, effectively crushing or pushing away stubborn jamming objects at the valve seat, avoiding overload burnout of the motor or damage to the mechanical structure due to continuous obstruction. Furthermore, after real-time monitoring of the jamming index decline, a polynomial interpolation algorithm is used to generate a smooth transition sequence, ensuring a seamless and flexible convergence of the valve's motion state from the high-acceleration impact stage to the original target trajectory.

[0044] By inputting time-series operational data and dynamic tracking errors into a pre-trained perception model, a sediment jamming index is output in real time, effectively overcoming the problems of lag and distortion in conventional control response under complex conditions of extremely low temperatures, and realizing online dynamic perception. At the same time, based on this index, a multi-level adaptive intervention strategy is executed, which automatically triggers additional disturbances or local trajectories containing reverse retreat and forward high acceleration impact when facing jamming, actively breaking up and breaking through solid condensation barriers, changing the status quo of traditional passive maintenance.

[0045] Example 2: This embodiment applies a deep learning-based adaptive control method for the opening of a cryogenic ball valve to the regulation of medium flow in a cryogenic separation pipeline for ethylene. Upon receiving an opening adjustment request, the current actual opening of the target cryogenic ball valve is obtained as 10°, the target opening value is 80°, and the desired adjustment time parameter is 4s. A smooth desired angular displacement curve is constructed using a fifth-order polynomial interpolation algorithm, and discretized sampling is performed at 10ms control time steps to generate a total of 400 time steps of the desired opening angular displacement sequence, which serves as the target opening command. Acquire the timing data of the target cryogenic ball valve and calculate the dynamic tracking error. At each 10ms control time step, calculate the real-time position deviation between the target opening command and the actual opening displacement sequence point by point, and calculate the velocity deviation, representing motion hysteresis, through the first-order backward difference between adjacent time steps. Set the sliding time window length to 50, preset the position weight coefficient to 0.5, and the velocity weight coefficient to 0.5. Sum the real-time position deviation and velocity deviation at each time step within the window using weighted summation, and concatenate them in chronological order to generate a 50×1 dynamic tracking error sequence. The current sequence containing 50 consecutive time steps and the dynamic tracking error are input into a pre-trained deep learning perception model, which outputs the sediment trapping index at the current time step. When calculating cross-modal cross-attention, an opening position mask matrix is ​​generated based on the actual opening displacement. The valve seat sealing area from 0° to 5° and from 85° to 90° is assigned a decay weight of 0.2, and the free adjustment area from 5° to 85° is assigned a standard weight of 1. After processing by the feature weighted fusion and temporal evolution module, the logistic activation function at the end of the model outputs a sediment trapping index of 0.88 at the current time step. Based on the comparison between the sediment jamming index and the preset threshold, adaptive control of the opening degree is executed; the first threshold calibrated by the benchmark test is 0.25, and the second threshold calculated by the transient stall current parameter of the motor is 0.80; since the current output sediment jamming index of 0.88 is greater than the second threshold of 0.80, it is determined that a severe solid condensate jamming has occurred, and the current target opening degree trajectory is immediately interrupted. The system sequentially generates local trajectory commands for reverse retreat and forward high-acceleration impact to actively break through obstacles. The controller records the actual opening displacement of 35° when the current trajectory is interrupted as the interruption point. Combined with the preset target retreat displacement of 6°, it generates a local trajectory command for reverse retreat to 29° opening and issues it for execution. After the reverse retreat is completed, it moves towards the target breakthrough position 3° past the interruption point (i.e., 38° opening). Using the maximum allowable transient acceleration set by the drive motor as the initial acceleration, it generates a local trajectory command for forward high-acceleration impact, using the instantaneously accumulated high kinetic energy to crush the deposits at the valve seat. After the jamming index drops, an interpolation transition is performed to smoothly merge back into the original opening trajectory. During the positive high-acceleration impact, the system continuously receives the output of the sensing model. When the jamming index of the sediment rapidly drops to 0.15 (less than the first threshold of 0.25), it is determined that the jamming material has been effectively removed and the impact is stopped. The controller then uses the actual opening of 37° and the running speed at the current moment as the starting state, and the expected opening and expected speed corresponding to the original opening trajectory in the next 0.6 seconds as the ending state. It uses a cubic polynomial interpolation algorithm to generate a smooth transition sequence and outputs it one by one until the actual movement state of the valve seamlessly and flexibly merges back into the original opening trajectory.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based adaptive control method for the opening degree of a cryogenic ball valve, characterized in that, include: Receive an opening adjustment request, obtain the opening trajectory of the target cryogenic ball valve, and generate a target opening command along the opening trajectory; Acquire the timing operation data of the target cryogenic ball valve; The timing operation data includes the current sequence and the actual opening displacement sequence; Calculate the dynamic tracking error of the actual opening displacement sequence relative to the target opening command; The current sequence and dynamic tracking error are input into a pre-trained deep learning perception model, which outputs the sediment choke index at the current time step. Based on the comparison between the sediment tack index and the preset threshold, adaptive control of the opening degree is performed; When the value is less than or equal to the first threshold, the target opening command is output; when the value is between the first and second thresholds, the corresponding disturbance amplitude is extracted, and a disturbance period including a sudden change window and a recovery window is set. Within the sudden change window, the disturbance amplitude is added to the target opening command. Within the recovery window, the additional disturbance amplitude is gradually reduced to zero over time steps, the target opening command is updated and output in a loop; when it is greater than or equal to the second threshold, the current trajectory is interrupted and local trajectory commands for reverse retreat and forward high acceleration impact are generated in sequence, and after the jamming index falls back, the trajectory is interpolated to transition to the opening trajectory.

2. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 1, characterized in that, The opening adjustment request is a control signal that includes a target opening value and a desired adjustment time parameter. The opening trajectory is a smooth desired angular displacement curve that changes continuously with time, constructed based on the current actual valve opening, the target opening value, and the desired adjustment time. The target opening command is a sequence of expected opening angular displacements corresponding to each control time step, generated by discretizing and sampling the opening trajectory.

3. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 1, characterized in that, The dynamic tracking error is the real-time position deviation between the expected opening value corresponding to the target opening command and the actual opening value in the actual opening displacement sequence, calculated point by point under a unified control time step. Calculate the rate of change of the real-time position deviation between adjacent control time steps to obtain the velocity deviation characterizing valve motion hysteresis; fuse the real-time position deviation and the velocity deviation in the time dimension to generate dynamic tracking error.

4. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 1, characterized in that, The deep learning perception model includes a feature extraction module, a cross-modal cross-attention module, and a temporal evolution module; The feature extraction module consists of two parallel one-dimensional convolutional networks, which independently receive the current sequence and dynamic tracking error, and output current response features and tracking error features respectively. The cross-modal cross-attention module linearly projects the tracking error features as a query matrix and linearly projects the current response features as a key matrix and a value matrix. It calculates the dynamic attention weights of the tracking error features on the current response features and uses these dynamic attention weights to perform weighted fusion of the current response features to generate abnormal friction resistance features. The temporal evolution module includes a gated recurrent unit and a multilayer perceptron, which receives the abnormal friction resistance features and captures long-term dependent evolution trends. It outputs the sediment sticking index at the current time step through an activation function.

5. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 4, characterized in that, The cross-modal cross-attention module introduces an opening position mask matrix. The specific process includes: dividing the entire valve stroke into a valve seat sealing area at both ends of the stroke and a free adjustment area between them; extracting the real-time opening value from the actual opening displacement sequence to generate an opening position mask matrix with the same dimension as the product of the query matrix and the key matrix. When the real-time opening value is in the valve seat sealing area, the corresponding element of the opening position mask matrix is ​​set to a decay weight less than 1; when the real-time opening value is in the free adjustment area, the corresponding element is set to 1; the query matrix and the key matrix are multiplied by a dot product to obtain the original attention score matrix, and the original attention score matrix and the opening position mask matrix are multiplied by an element-wise Hadamard product, and the dynamic attention weights are output through a normalization function.

6. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 1, characterized in that, The specific process for determining the mutation window and the recovery window is as follows: The mutation window is calculated by dividing the difference between the sediment scouring index and a first threshold by the difference between a second threshold and the first threshold to obtain a dynamic ratio; the dynamic ratio is multiplied by the total number of time steps of the disturbance period and rounded to obtain the mutation step number, which is a continuous time step interval with the start point of the disturbance period as the starting point and the length as the mutation step number; the recovery window is calculated by subtracting the mutation step number from the total number of time steps of the disturbance period to obtain the recovery step number, which is a continuous time step interval with the end point of the mutation window as the starting point and the length as the recovery step number.

7. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 6, characterized in that, The first threshold is determined by obtaining the test current sequence and test tracking error of the target cryogenic ball valve under a sediment-free baseline state, inputting it into the deep learning perception model to generate a baseline Chase index sequence; calculating the mean and standard deviation of the baseline Chase index sequence, and using the sum of the mean and the standard deviation by a set multiple as the first threshold. The second threshold is determined by obtaining the limit current sequence and limit tracking error of the drive motor corresponding to the target cryogenic ball valve under the critical state of triggering overload protection, inputting it into the deep learning perception model to calculate the limit jamming index; calculating the ratio of the rated maximum continuous operating current of the drive motor to the transient stall current, and multiplying the limit jamming index by the ratio as the second threshold.

8. The method for adaptive control of cryogenic ball valve opening based on deep learning according to claim 1, characterized in that, The generation and interpolation transition process of the local trajectory command specifically includes: recording the actual opening displacement when the current trajectory is interrupted as the interruption point; using the interruption point as the starting point and combining it with the target retreat displacement, generating a reverse retreat local trajectory command; after the reverse retreat local trajectory command is executed, generating a positive high-acceleration impact local trajectory command towards the target breakthrough position greater than the interruption point position; during the execution of the positive high-acceleration impact local trajectory command, when the sediment sticking index falls back to less than or equal to the first threshold, using the actual opening and running speed at the current moment as the starting state and the expected opening and expected speed of the opening trajectory in the corresponding future time step as the ending state, using a polynomial interpolation algorithm to generate a smooth transition sequence and output it until it merges into the opening trajectory.

Images