A pressure-adaptive based dynamic sealing control method for cable entry

Abstract

The present application relates to cable entry device sealing control technical field, especially a kind of dynamic sealing control method of cable entry device based on pressure self-adaption, its technical scheme includes: with the pressure difference inside and outside sealing cavity, sealing force, compression displacement and temperature as state variable, establish discrete state space prediction model;Optimal fusion estimation is carried out to multiple-source information using extended Kalman filter;Future pressure difference sequence is predicted using long short-term memory neural network;With the predicted sequence as reference trajectory, construct multi-objective optimization function including state tracking, incremental smoothing, energy consumption suppression and life loss function based on Miner theory, solve the optimal control sequence that meets the constraint through model predictive control rolling optimization, drive the independently regulated sealing force of multiple pressure action units arranged circumferentially;Online identification updates model parameters;Realize the active prediction and self-adaptive adjustment of sealing force to pressure change, improve the sealing reliability and sealing life under variable pressure working condition.

Description

Technical Field

[0001] This invention relates to the field of sealing control technology for cable entry devices, and in particular to a dynamic sealing control method for cable entry devices based on pressure adaptation. Background Technology

[0002] Cable entry devices are critical connecting components when cables pass through equipment enclosures, bulkheads, or pipe walls. They are widely used in explosion-proof electrical equipment, shipboard cable entry, underground coal mine electrical equipment, submarine observation networks, deep-sea submersibles, and petrochemical industries for cable sealing and protection. Their main function is to allow cables to pass through the enclosure wall while reliably sealing and isolating the internal environment from the external environment, preventing harmful media such as water, gas, oil, and dust from seeping into the equipment along the cable's outer wall, thereby ensuring the safe operation of the internal electrical equipment.

[0003] Currently, the sealing structures of cable entry devices are mainly divided into two categories: sealing ring type and filler casting type. Among them, the sealing ring type entry device is the most widely used due to its advantages such as simple structure, convenient installation, and reusability. Its typical structure is a compression nut type or a pressure plate type. By tightening the compression nut, the elastic sealing ring undergoes radial compression deformation, thereby forming a sealing contact surface between the sealing ring and the outer wall of the cable, and between the sealing ring and the inner wall of the device, thus achieving the sealing function. The sealing ring is usually made of elastic materials such as rubber. Under the action of external force, it deforms but its volume remains unchanged. It relies on the contact pressure generated by the compression deformation to seal the gap between the cable and the device.

[0004] However, existing cable entry device sealing technologies have revealed the following significant shortcomings in long-term engineering applications: (I) Sealing performance heavily depends on the sealing ring material, resulting in an extremely high failure rate. The sealing ring is the most critical component in cable entry devices, and its performance directly determines the sealing function. According to relevant statistics, failures caused by the sealing ring account for more than 90% of all failures in explosion-proof cable entry devices. During long-term use, the sealing ring inevitably undergoes aging, loss of elasticity, permanent deformation, and deterioration, leading to insufficient effective dimensions and fitting clearances to meet sealing requirements. Especially under harsh operating conditions such as high temperature, high humidity, and strong corrosion, the performance degradation of the sealing material is accelerated, and its service life is significantly shortened.

[0005] (ii) Fixed sealing force, lacking dynamic adaptability to changes in internal and external pressure. Traditional cable entry devices employ a "one-time installation, fixed pre-tightening" sealing method. Once the sealing ring is compressed to a certain position by the clamping nut, its compression and sealing specific pressure remain fixed. However, in actual working conditions, the internal and external pressure differences faced by the cable entry device are dynamically changing—such as drastic changes in external water pressure during the descent and ascent of deep-sea submersibles, air pressure fluctuations in high-altitude areas, periodic fluctuations in internal pressure of underground pipelines, and the sealing of hull penetrations on ships under different water depth conditions. When the external environmental pressure increases while the sealing specific pressure remains constant, the clamping force on the sealing contact surface may be insufficient to resist the external medium pressure, leading to seal failure. When the external environmental pressure decreases, excessive fixed pre-tightening force will cause the sealing ring to be in an over-compressed state for a long time, accelerating material fatigue and creep relaxation, thus shortening the service life of the seal.

[0006] (III) The aging and creep relaxation of seals cannot be actively compensated. Viscoelastic sealing materials such as rubber will undergo stress relaxation and creep deformation under long-term constant compressive stress, causing the clamping force of the sealing ring on the cable to gradually decrease over time. Current technology lacks active monitoring and compensation methods for this, and can only rely on periodic manual inspection and replacement. This "passive maintenance" mode is not only costly to maintain, but also cannot prevent the safety hazards caused by sudden seal failure—once the performance of the seal degrades below the critical value between two maintenance intervals, it may cause a catastrophic leakage accident.

[0007] (iv) The sealing status is unknowable and uncontrollable, lacking intelligent closed-loop management. Existing cable entry devices are typical "open-loop" sealing systems—after installation, key parameters such as the actual compression of the sealing ring, sealing specific pressure, and material aging degree cannot be obtained in real time. When the sealing performance deteriorates, operators often only passively detect it after a leak occurs, making online monitoring, early warning, and proactive control of the sealing status impossible. Although some products employ passive anti-loosening measures such as anti-loosening nuts and limiting structures, they still cannot solve the fundamental problem of dynamic degradation of sealing performance with operating conditions and time.

[0008] (v) Uneven distribution of sealing force along the cable circumference exacerbates the risk of localized failure. In a clamping nut type structure, the sealing ring is subjected to axial compressive force, which is converted into radial clamping force on the cable through the deformation of the sealing ring material. In actual installation, due to factors such as cable weight, installation posture, and cable eccentricity, the clamping force of the sealing ring on different positions along the cable circumference is often uneven—the lower position bears a greater clamping force, while the upper position has a relatively insufficient clamping force. This uneven circumferential force not only reduces the overall sealing effect but also leads to excessive fatigue of local sealing rings, accelerating failure.

[0009] (vi) The sealing performance assessment cycle is long and costly, making it difficult to guide online optimization. The current explosion-proof performance assessment of cable entry devices is mainly completed through heat and cold resistance tests, clamping tests, sealing tests, and mechanical strength tests, with a test cycle of more than one month. These tests are all offline type tests, which can only verify the initial performance of the product under specific conditions. They cannot reflect the dynamic sealing performance changes of the device during actual service, and cannot provide real-time decision-making basis for online sealing control.

[0010] In summary, existing cable entry device sealing technologies generally suffer from inherent defects such as "passive sealing, fixed pre-tightening, lack of state sensing, lack of dynamic adjustment, and lack of aging compensation," making it difficult to meet the urgent needs for sealing reliability, safety, and intelligence under variable pressure, variable temperature, and long service life conditions such as deep-sea operations, high-voltage power transmission and transformation, coal mine explosion protection, and shipboard passage.

[0011] In view of this, we propose a dynamic sealing control method for cable introduction devices based on pressure adaptation to solve the existing problems. Summary of the Invention

[0012] The purpose of this invention is to provide a dynamic sealing control method for a pressure-adaptive cable introduction device to solve the problems mentioned in the background art.

[0013] To achieve the above objectives, the present invention provides the following technical solution: a dynamic sealing control method for a pressure-adaptive cable introduction device, comprising the following steps: S1. Multi-source sensor information fusion and state-space modeling: using the pressure difference ΔP inside and outside the sealed cavity of the cable entry device and the actual output sealing force F of the sealing actuator. act The compression displacement x of the sealing component and the real-time temperature T inside the sealing cavity are used as state variables to establish a discrete state-space prediction model that reflects the dynamic characteristics of the sealing system: X(k+1) = A·X(k)+ B·u(k) + W·v(k), where X(k) is the state vector at time k, u(k) is the control input, v(k) is the external environmental pressure disturbance, A is the system state transition matrix, B is the control input matrix, and W is the disturbance input matrix; S2. Optimal State Estimation Based on Extended Kalman Filter: Measurement data is collected using pressure sensors, displacement sensors, and temperature sensors located inside and outside the sealed cavity. The extended Kalman filter algorithm is then used to perform optimal state estimation on the state-space prediction model, filtering out sensor noise and outputting the optimal state estimate at time k. Simultaneously, the state estimation covariance matrix is ​​updated online recursively; S3. Future pressure differential prediction based on long short-term memory neural network: The controller retrieves the external environmental pressure value P from the historical time series. 外and the internal sealing cavity pressure value P 内 Input into a pre-trained long short-term memory neural network model, and output the future N. p A time-domain pressure difference prediction sequence {ΔP(k+i|k), i=1, 2, …, N} p}, and use the predicted sequence as the reference input for model predictive control; S4. Model predictive control rolling time-domain multi-objective optimization: at each sampling time k, using the optimal state estimate... Using the pressure difference prediction sequence as the initial condition and the pressure difference prediction sequence as the reference trajectory, in the finite control time domain N c The following multi-objective optimization cost function J(k) is constructed and solved: The first item is the sealing status tracking item, X ref The first term is the reference trajectory for the desired sealed state; the second term is the control increment smoothing term; the third term is the actuator energy consumption suppression term; the fourth term is L. loss (k) is the seal life loss function based on Miner's linear cumulative damage theory, λ is the life loss weight coefficient; Q, R, and S are the corresponding weight matrices; S5. Optimal Control Sequence Output and Execution under Constraints: Under the conditions of satisfying the maximum output force constraint, maximum compression stroke constraint, and lower limit constraint of sealing specific pressure of the sealing actuator, solve for the optimal control sequence U(k) = {u(k), u(k+1), …, u(k+N)} that minimizes the cost function J(k). c -1)}, output the current optimal control quantity u*(k) to multiple independent pressure action units arranged along the circumference of the cable, and each pressure action unit independently performs dynamic adjustment of the sealing force; S6. Feedback Correction and Online Model Parameter Identification: After executing the current control variable, the actual output state value is collected, and the actual state value is compared with the predicted value of the prediction model. The prediction error is calculated and fed back to the controller to perform rolling correction on the predicted output. At the same time, the recursive least squares method is used to identify and update the system state transition matrix A and disturbance input matrix W in the state space prediction model online based on the real-time collected data, so that the model continuously approximates the dynamic characteristics of the actual sealed system.

[0014] Furthermore, in step S4, the seal life loss function L lossThe construction method of (k) is as follows: based on the stress-strain constitutive relationship of the sealing material and the pre-stored fatigue characteristic curve, the cumulative creep damage rate of the sealing component under the current sealing force and temperature conditions is calculated in real time. The instantaneous damage rate in the control time domain is integrated over time to obtain the cumulative life loss prediction in the future control time domain. The controller dynamically adjusts the weight coefficient λ according to the cumulative life loss prediction. When the predicted loss exceeds the preset warning value, λ is adaptively increased to preferentially reduce the amplitude of the sealing force.

[0015] Furthermore, in step S4, the desired sealing state reference trajectory X ref The operating mode is adaptively switched according to the operating condition mode setting, which includes a normal voltage regulation mode, a pressure surge suppression mode, and a low-loss long-life mode; in the pressure surge suppression mode, X ref The target differential pressure value is forcibly set to zero, and the weight of the differential pressure term in the tracking term weight matrix Q is increased to K as in the normal mode. surge times, K surge ≥ 3; In the low-loss, long-life mode, the lifetime loss weighting coefficient λ increases to K in the conventional mode. life times, K life ≥ 2.

[0016] Furthermore, in step S4, the solution of the multi-objective optimization cost function adopts a hybrid optimization strategy combining adaptive genetic algorithm and sequential quadratic programming. Specifically, at each sampling time, the adaptive genetic algorithm is first used to perform coarse-grained optimization in the global solution space to obtain an approximate optimal solution; then, with the approximate optimal solution as the initial iteration point, the sequential quadratic programming algorithm is used to perform a local search to obtain the optimal control sequence that satisfies the constraints; wherein, the crossover probability and mutation probability of the adaptive genetic algorithm are dynamically adjusted according to the fitness concentration of the current generation.

[0017] Furthermore, in step S1, the state-space prediction model also embeds the generalized Maxwell viscoelastic constitutive equation of the sealing material as an intrinsic constraint condition of the model; the controller calculates the real-time stress relaxation of the sealing material based on the real-time temperature T and the current compression displacement x using the generalized Maxwell constitutive equation, and adds a feedforward compensation term u to the control input u(k). ff (k), feedforward compensation term u ff The magnitude of (k) is positively correlated with the real-time stress relaxation amount, so as to actively compensate for the decrease in sealing specific pressure caused by creep relaxation of the sealing material.

[0018] Furthermore, in step S2, while performing state estimation, the extended Kalman filter algorithm calculates the covariance matrix of the innovation sequence and judges in real time whether there is a sudden abnormality or jamming fault in the external environmental pressure sensor based on the statistical characteristics of the innovation sequence. When the mean of the innovation sequence is detected to exceed the preset 3σ confidence interval, the controller determines that the current sensor data is abnormal, automatically reduces the fusion weight of the sensor in the state estimation, and switches to the degradation estimation mode based on the data of other healthy sensors.

[0019] Furthermore, in step S3, the input feature vector of the long short-term memory neural network model includes: the external environmental pressure value, the internal sealing cavity pressure value, the temperature change rate, and the attitude angle information of the cable introduction device at M sampling times prior to the current time; the long short-term memory neural network model is a two-layer stacked structure, with each layer containing L LSTM neurons, and overfitting is prevented through a dropout layer; the training data of this model comes from the actual operation history database of the sealing system and the extreme working condition extended sample set generated based on finite element simulation.

[0020] Further, in step S6, the recursive least squares method calculates the update amount ΔA of the system state transition matrix A at each sampling time, and determines ||ΔA|| F Is it less than the preset convergence accuracy threshold ε, where ||·|| F It is the Frobenius norm; when ||ΔA|| F When the value of the model is less than ε and continues to exceed the preset stable time window, the controller determines that the dynamic characteristics of the current sealing system tend to be stable, the model identification converges, and the currently identified model parameters are stored in non-volatile memory as the reference model parameters under this working condition. When the external environmental pressure undergoes a step change and the change amplitude exceeds the preset amplitude, the controller forcibly resets the covariance matrix of the recursive least squares method to accelerate the convergence speed of the model parameters to the new working condition.

[0021] Furthermore, in step S5, the maximum output force constraint is the upper limit F of the rated output force of each pressure action unit. max The maximum compression stroke constraint is the maximum allowable compression displacement of the seal x. max The lower limit constraint of the sealing specific pressure is calculated in real time based on the current cable outer diameter and sealing cavity size according to the following formula: P sealmin = K safe × (P 外 -P 内 ) max + P margin , where P sealmin K is the lower limit of the dynamic minimum safe sealing pressure. safe For a safety margin factor greater than 1, (P) 外 -P内 ) max P represents the maximum predicted differential pressure value in the differential pressure prediction sequence. margin This is a preset additional sealing safety margin; when solving the optimization problem, the controller transforms the sealing specific pressure lower limit constraint into a hard constraint condition on the control output u(k), ensuring that the actual sealing specific pressure at any time is always not lower than this lower limit value.

[0022] Furthermore, in step S4, before generating the dynamic sealing compensation command and driving each pressure action unit to output the sealing force, the controller also performs the following initial off-center load balancing and self-calibration steps: S4a, Initial Compression Displacement Acquisition: The controller reads the initial feedback value of the displacement sensor corresponding to each pressure action unit, and records it as... Let i = 1, 2, ..., N, and calculate the arithmetic mean of all initial feedback values. And determine the maximum initial displacement value. With minimum initial displacement value ; S4b, Off-center load determination: The controller calculates the difference Δx0 between the maximum and minimum initial displacement values. - And compare Δx0 with the preset allowable off-center load deviation threshold Δx allow Compare; when Δx0 < Δx allow When the controller determines that the cable is properly centered within the inlet device, it directly enters the dynamic sealing main control process; when Δx0 ≥ Δx allow When the controller determines that the cable is eccentrically installed in the sealed cavity or is not uniformly compressed due to gravity sag, it proceeds to step S4c. S4c, Self-calibrating equalization control: The controller generates an off-center load equalization pre-adjustment command based on the average value. As the target equilibrium displacement value for each pressure action unit, a proportional-integral-derivative control strategy is adopted, and an equilibrium adjustment amount Δu is independently applied to each pressure action unit. i = K p ·( - ) + K i ·∫( - )dt + K d ·d( - The pressure actuator with an initial compression displacement that is too large actively retracts to release part of the sealing force, while the pressure actuator with an initial compression displacement that is too small actively feeds to increase the sealing force, until the real-time displacement value x of each pressure actuator is reached. i All satisfy |x i- |<ε, where ε is the preset equilibrium convergence precision; S4d, Equalization Reference Lock: When the displacement values ​​of each pressure action unit converge to within the equalization convergence accuracy ε, the controller calibrates the current displacement values ​​of each pressure action unit as the new initial reference state, stores the current output force values ​​of each pressure action unit as the offset compensation amount, and executes the generation and output of dynamic sealing compensation commands with this equalized state as the starting state; at the same time, the controller stores the offset compensation amount in non-volatile memory, and directly calls the offset compensation amount as the initial preload setting reference when the same specification cable is installed again.

[0023] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention establishes a discrete state-space prediction model that includes pressure difference, sealing force, compression displacement, and temperature. Combined with extended Kalman filter optimal state estimation and LSTM neural network future pressure difference prediction, this enables the controller to predict future N pressure fluctuations before they actually occur. p The pressure difference change trend within a time domain is analyzed, and compensation commands are output in advance through model predictive control and rolling time domain optimization. This "prediction-decision-execution" feedforward control mechanism ensures that the sealing force always matches the internal and external pressure difference, achieving a fundamental leap from "passive response after pressure changes" to "active pre-adjustment before pressure changes," and from "passive fixed sealing" to "active predictive sealing." This fundamentally solves the technical problem that the sealing specific pressure cannot be dynamically adjusted according to the operating conditions.

[0024] 2. The cost function constructed in this invention comprises four terms: a sealing state tracking term to ensure sealing safety, a control increment smoothing term to prevent frequent oscillations of the actuator, an actuator energy consumption suppression term to reduce system power consumption, and a sealing component life loss function based on Miner's linear cumulative damage theory. More importantly, it further limits the controller to dynamically adjust the weight coefficient λ based on the cumulative life loss prediction—when the predicted loss exceeds a preset warning value, λ adaptively increases to prioritize reducing the sealing force amplitude. This online dynamic game and trade-off optimization mechanism with three objectives—"safety-energy consumption-life"—maximally delays fatigue and creep damage of the sealing material while ensuring absolute sealing safety, significantly extending the service life of the sealing component. Through multi-objective optimization of the cost function, it achieves multi-dimensional synergistic optimization of sealing safety, energy efficiency, and sealing component lifespan.

[0025] 3. This invention defines three adaptive switching modes for different operating conditions: In the pressure surge suppression mode, the target differential pressure value is forcibly set to zero, and the weight of the differential pressure term is increased to more than three times that of the conventional mode, ensuring that the system eliminates the differential pressure as quickly as possible and guarantees safety; in the low-loss, long-life mode, the life loss weight coefficient is increased to more than twice that of the conventional mode, prioritizing the protection of seals and extending the replacement cycle. This differentiated control strategy of "one machine, multiple strategies" enables the same sealing control system to achieve optimal control in vastly different operating environments, such as deep-sea submersibles, shipboard penetration, coal mine explosion-proof environments, and seabed observation networks. The adaptive switching of operating modes achieves differentiated optimal control strategies for different service scenarios.

[0026] 4. This invention limits the Extended Kalman Filter (EKF) algorithm to simultaneously calculate the covariance matrix of the innovation sequence and perform sensor fault detection based on the 3σ confidence interval criterion while performing state-optimal estimation. When a sensor's data is abnormal, the controller automatically reduces the weight of that sensor in state fusion and switches to a degradation estimation mode based on the data of the remaining healthy sensors. This mechanism enables the system to maintain reliable state awareness and sealing control even when some sensors fail, greatly improving the engineering availability and fault tolerance of the sealing system in harsh industrial environments. Highly robust state awareness in harsh environments is achieved through EKF and sensor fault self-diagnosis.

[0027] 5. This invention embeds the generalized Maxwell viscoelastic constitutive equation of the sealing material into the state-space prediction model as an intrinsic constraint. The controller calculates the real-time stress relaxation of the sealing material online using this constitutive equation based on real-time temperature and current compression displacement, and adds a positively correlated feedforward compensation term to the control input. This mechanism enables the system to proactively "predict" the performance degradation of the sealing material due to aging and increase the sealing force in advance to compensate, rather than passively detecting it only after the sealing performance has decreased to a critical value. This fundamentally overcomes the inherent defect of traditional sealing technology, which is "peak performance upon installation, and only declines thereafter." Through feedforward compensation using the generalized Maxwell viscoelastic constitutive equation, it proactively offsets the sealing performance degradation caused by creep relaxation of the sealing material.

[0028] 6. This invention employs the recursive least squares method to calculate the update amount ΔA of the system state transition matrix A online at each sampling time, and uses the Frobenius norm ||ΔA|| FThe system determines whether the model has converged. Once the model identification shows convergence, the current model parameters are stored in non-volatile memory as the baseline model parameters for that operating condition. When the external environmental pressure undergoes a sudden change, the controller forcibly resets the covariance matrix of the recursive least squares method to accelerate the model's convergence to the new operating condition. This online identification mechanism, which "runs, learns, and evolves simultaneously," ensures that the predictive model always approximates the true dynamic characteristics of the physical system, realizing a digital twin of the sealed system—the mathematical model in the controller is synchronized and mapped in real time with the physical sealed system. This fundamentally guarantees the accuracy of control decisions during long-term service. The online recursive least squares identification achieves the self-evolution and self-correction of the "digital twin" model.

[0029] 7. This invention defines a lower limit constraint for the sealing specific pressure. This safety lower limit is not a fixed value calculated based on the current differential pressure, but rather a dynamically calculated value based on the maximum possible differential pressure in the future. More importantly, this constraint is defined as a hard constraint condition in the MPC optimization solver—regardless of the trade-off between energy consumption optimization and lifespan protection, the optimal control quantity solved must ensure that the actual sealing specific pressure at any given time is never lower than the dynamic minimum safe sealing specific pressure lower limit; otherwise, the solution is directly rejected. This "hard constraint" mechanism ensures that sealing safety has the highest priority at the algorithm architecture level. Under no operating condition can absolute sealing safety be sacrificed for energy saving or protecting the seals. This provides the highest level of safety assurance for applications with extremely stringent requirements for sealing reliability, such as explosion-proof, deep-sea, and nuclear industries. Through the hard constraint of the lower limit of the sealing specific pressure, an absolute leak-proof defense line of "safety first" is established.

[0030] 8. Before the dynamic sealing main control process, this invention independently sets up an initial off-center load balancing and self-calibration step: by reading the initial feedback values ​​of the displacement sensors corresponding to each pressure action unit, the degree of off-center load is calculated. When the degree of off-center load exceeds the preset allowable deviation, a proportional-integral-derivative control strategy is used to independently apply a balancing adjustment amount to each pressure action unit, driving the unit with a large displacement to actively retract and the unit with a small displacement to actively advance, until the displacement values ​​of each unit converge to within the balancing convergence accuracy. This intelligent mechanism of "installation-based self-calibration" eliminates the uneven circumferential sealing force distribution caused by cable eccentricity from the source, making the sealing component uniformly stressed in the entire circumference, greatly improving the uniformity and reliability of the seal. At the same time, the offset compensation amount is stored in a non-volatile memory for direct retrieval when subsequent cables of the same specification are installed again, realizing the engineering convenience of "one-time calibration, multiple reuses". Through initial off-center load balancing self-calibration, the uneven circumferential sealing force caused by cable eccentric installation is eliminated. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the workflow of a dynamic sealing control method for a pressure-adaptive cable introduction device according to the present invention. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1 As shown, a dynamic sealing control method for a pressure-adaptive cable entry device includes the following steps: S1. Multi-source sensor information fusion and state-space modeling; S2. Optimal state estimation based on extended Kalman filter; S3. Future pressure differential prediction based on long short-term memory neural network; S4. Model predictive control rolling time-domain multi-objective optimization; S5. Output and execution of the optimal control sequence under constraints; S6. Feedback correction and online identification of model parameters. Example

[0034] This embodiment focuses on the hardware configuration of the dynamic sealing control system for the cable entry device.

[0035] The pressure-adaptive dynamic sealing control method for cable entry devices in this embodiment is applied to a cable entry device, which includes a device body, a sealing cavity disposed inside the device body, a cable passing through the sealing cavity, and multiple pressure action units arranged along the circumference of the cable.

[0036] One end of the device body is the cable inlet, and the other end is the cable outlet. The sealed cavity is located inside the device body. An annular elastic seal (a nitrile rubber sealing ring is used in this embodiment) is installed inside the sealed cavity, and the seal is fitted around the outer circumference of the cable. N pressure actuation units are evenly arranged along the outer circumference of the seal (N=8 in this embodiment, i.e., one unit is arranged every 45° circumferentially). Each pressure actuation unit is an electric actuator (specifically, a lead screw and nut mechanism driven by a micro stepper motor). The output end of each electric actuator presses against the outer circumferential surface of the seal. By independently controlling the feed amount of each electric actuator, the radial compression of the cable by the seal is adjusted, thereby changing the sealing pressure.

[0037] An internal pressure sensor (range 0~10MPa, accuracy ±0.5%FS) and a temperature sensor (PT100 platinum resistance thermometer, accuracy ±0.1℃) are installed inside the sealed cavity. Multiple external environmental pressure sensors (four in this embodiment, located at the top, bottom, left, and right of the device casing) are installed in the external environment. Each pressure actuation unit is equipped with a position sensor (a linear variable differential transformer displacement sensor, range 0~20mm, accuracy ±0.01mm, used in this embodiment) to provide real-time feedback on the compression displacement of each electric actuator. The outputs of all the above sensors are connected to a controller (an industrial-grade ARM Cortex-A9 embedded processor, 1GHz, equipped with a floating-point unit, used in this embodiment), and the controller outputs are connected to the drive circuits of each electric actuator.

[0038] The controller integrates a data storage module (including RAM and Flash non-volatile memory) to store state-space prediction model parameters, LSTM neural network model weights, adaptive learning parameters, and operation log data. The controller is also connected to an early warning module (including an audible and visual alarm and a remote communication interface) to issue alarm signals in case of sealing abnormalities. Example

[0039] This embodiment focuses on the specific implementation of multi-source sensor information fusion and state space modeling.

[0040] The controller operates according to a preset sampling period T. s (In this embodiment, T) s =100ms) Synchronously acquire the output signals of each sensor. Assume that the data acquired at time k includes: the pressure value P inside the sealed cavity. 内 (k) External environmental pressure value P 外 (k) (Take the weighted average of the values ​​collected by the four external sensors as the final P) 外 (k), weights are allocated according to the inverse of the historical variance of each sensor), and the compression displacement x of each pressure action unit. i (k) (i=1,2,…,8), the temperature T(k) inside the sealed cavity.

[0041] The controller calculates the pressure difference ΔP(k) between the inside and outside of the sealed cavity at time k = |P 内 (k)-P 外 (k)|。 Actual output sealing force F of each pressure action unit act(i) (k) Obtained through displacement-force calibration curve conversion—Each electric actuator undergoes calibration tests before leaving the factory to obtain the displacement-output force correspondence curve, which is stored in the controller; the total actual output sealing force F act (k)=ΣF act(i) (k).

[0042] With ΔP, F act Using x (the arithmetic mean of the displacement values ​​of each pressure action unit as the overall compression displacement of the seal) and T as four-dimensional state variables, a discrete state-space prediction model is established: X(k+1) = A·X(k) + B·u(k) + W·v(k) Where X(k)=[ΔP(k), F act (k), x(k), T(k)] T Let u(k) be the state vector at time k; u(k) be the control input at time k (in this embodiment, it is the equivalent control input after weighted summation of the target drive current values ​​of each electric actuator); v(k) be the external environmental pressure disturbance, which is the current external environmental pressure value P. 外 (k) Subtract the external environmental pressure value P from the previous moment 外 The difference of (k-1); A is the 4×4 system state transition matrix, B is the 4×1 control input matrix, and W is the 4×1 disturbance input matrix.

[0043] The initial values ​​of the system state transition matrix A, control input matrix B, and disturbance input matrix W are determined as follows: After the cable entry device is installed, the controller performs an offline system identification—a set of known pressure disturbance signals (in this embodiment, a sinusoidal pressure disturbance with an amplitude of 0.5 MPa and a frequency of 0.1 Hz) is applied into the sealed cavity, the system's input and output response data are collected, and the initial values ​​of A, B, and W are estimated using a subspace identification method. The initial values ​​are then stored in non-volatile memory.

[0044] Furthermore, the state-space prediction model also incorporates the generalized Maxwell viscoelastic constitutive equation for the sealing material as an intrinsic constraint condition. Specifically, the generalized Maxwell model consists of multiple Maxwell elements connected in parallel, and its stress relaxation modulus G(t) is expressed as:

[0045] Among them, G ∞ To balance the modulus, G j Let τ be the elastic modulus of the j-th Maxwell element. j Let be the relaxation time constant of the j-th Maxwell element, m be the number of Maxwell elements (m=5 in this embodiment), and t be the time variable. All parameters are obtained by performing stress relaxation tests on the sealing material samples and fitting the data, and are pre-stored in the controller. The controller calculates the real-time stress relaxation of the sealing material based on the real-time temperature T and the current compression displacement x using this constitutive equation, and adds a feedforward compensation term u to the control input u(k). ff (k). Example

[0046] This embodiment focuses on a specific implementation of optimal state estimation based on extended Kalman filtering.

[0047] Since the signals collected by the sensors inevitably contain measurement noise and environmental interference, this embodiment uses the extended Kalman filter algorithm to perform optimal state estimation on the state-space prediction model.

[0048] The extended Kalman filter algorithm includes a prediction step and an update step: Prediction steps: (k|k-1) = A· (k-1|k-1) + B·u(k-1) + W·v(k-1) P(k|k-1) = A·P(k-1|k-1)·A T + Q ekf in, (k|k-1) is the prior estimate of the state at time k based on time k-1, P(k|k-1) is the prior estimate covariance matrix, and Q... ekf The process noise covariance matrix (in this embodiment, it is determined as diag([0.01, 0.1, 0.001, 0.001]) through offline parameter tuning).

[0049] Update steps: K g (k) = P(k|k-1)·H T ·[H·P(k|k-1)·H T + R ekf ] -1 (k|k) = (k|k-1) + K g (k)·[Z(k) - H· (k|k-1)] P(k|k) = [I - K g [(k)·H]·P(k|k-1) Among them, K g Z(k) is the Kalman gain, H is the observation matrix (in this embodiment, H is the identity matrix I, because all state variables can be measured directly or indirectly), Z(k) is the sensor observation vector at time k, and R ekf To measure the noise covariance matrix (determined through variance analysis of the sensor's static output), P(k|k) is the posterior estimated covariance matrix. (k|k) is the optimal state estimate at time k.

[0050] The Extended Kalman Filter algorithm calculates the innovation sequence ε(k) = Z(k) - H· while performing state estimation. The covariance matrix of (k|k-1) is used to determine whether there are abrupt changes or jamming faults in each sensor in real time based on the statistical characteristics of the innovation sequence. Specifically, when the mean of the innovation component corresponding to a certain sensor exceeds the preset 3σ confidence interval, the controller determines that the sensor data is abnormal, automatically reduces the fusion weight of the sensor in the observation vector Z(k), and switches to the degradation estimation mode based on the data of the remaining healthy sensors—that is, the abnormal sensor data is removed from the observation vector, the corresponding row in the observation matrix H is set to zero, and the state estimation is performed using the remaining sensor data. Example

[0051] This embodiment focuses on a specific implementation of future pressure differential prediction based on a long short-term memory neural network.

[0052] The controller retrieves the external environmental pressure value P from the historical time series. 外 and the internal sealing cavity pressure value P 内 Specifically, the controller extracts the pressure value sequence from the data storage module based on the M sampling times prior to the current time (M=100 in this embodiment, corresponding to 10 seconds of historical data).

[0053] The input feature vector for constructing the LSTM neural network includes: a sequence of external environmental pressure values ​​{P} from the M sampling times prior to the current time step. 外 (kM), …, P 外 (k-1)}, Internal sealing cavity pressure value sequence {P 内 (kM), …, P 内 (k-1)}, temperature change rate sequence, and attitude angle information of the cable introduction device (provided by an attitude sensor installed on the device housing to determine whether the device is installed horizontally or vertically, thus affecting the effect of gravity on the compression distribution of the seal).

[0054] The LSTM neural network in this embodiment is a two-layer stacked structure, with each layer containing 64 LSTM neurons, and overfitting is prevented through a Dropout layer (dropout rate of 0.2). The network's output layer is a fully connected layer with an output dimension of N. p (Predicted time domain length, N in this embodiment) p =10, meaning the predicted pressure difference change sequence within the next second, due to the sampling period T s =100ms, N p =10 corresponds to predicting the next 10 sampling times.

[0055] The training process of the LSTM neural network is as follows: Historical data of the sealing system during actual operation are collected, including pressure, temperature, and attitude data under various operating conditions (stable operation, sudden pressure rise, sudden pressure drop, temperature changes, etc.), forming a training dataset. Simultaneously, to cover extreme operating conditions, an expanded sample set is generated based on finite element simulation—a finite element model of the cable entry device is built in ANSYS to simulate the sealing system response under extreme pressure shocks (such as a pressure change rate of 20 MPa / s), generating simulation data which is added to the training set. During training, the Adam optimizer is used with a learning rate of 0.001, a batch size of 64, and 200 training epochs, using mean squared error as the loss function. After training, the model weights are permanently stored in the controller's non-volatile memory.

[0056] During online prediction, the controller inputs the current input feature vector into a trained LSTM neural network and outputs the future N... p A time-domain pressure difference prediction sequence {ΔP(k+i|k), i=1, 2, …, N} p The predicted sequence is then used as the reference input for the model's predictive control. Example

[0057] This embodiment focuses on the specific implementation of model predictive control rolling time-domain multi-objective optimization.

[0058] At each sampling time k, the optimal state estimate of the extended Kalman filter output is used. Using (k|k) as the initial condition and the pressure difference prediction sequence output by the LSTM neural network as the reference trajectory, in the finite control time domain N c (N in this embodiment) c =5, meaning the control time domain is 5 sampling periods, corresponding to 500ms, within which the multi-objective optimization cost function is constructed and solved:

[0059] in: The first item is the sealing status tracking item: X ref (k+i) represents the reference trajectory for the desired sealing state. X ref The settings automatically switch according to the operating mode—the system has three preset operating modes: Standard voltage regulation mode: X ref The target differential pressure is set to 0.1 MPa (a small differential pressure is allowed considering the sensor accuracy and steady-state error of the control). The target sealing force and the target compression displacement remain unchanged, and the target temperature is not controlled.

[0060] Stress mutation suppression mode: X refThe target differential pressure value is forcibly set to zero, and the weight of the differential pressure term in the tracking term weight matrix Q is increased to K as in the normal mode. surge times, K surge =3~5 (4 in this embodiment) to force the system to eliminate the pressure difference as quickly as possible. This mode is automatically triggered by the controller based on the maximum rate of change of pressure difference in the LSTM prediction sequence—when the future N... p When the rate of change of differential pressure within a time domain exceeds a preset threshold (0.5 MPa / s in this embodiment), the system automatically switches to this mode.

[0061] Low-loss, long-life mode: The lifetime loss weighting coefficient λ is increased to K in the conventional mode. life times, K life =2~3 (2.5 in this embodiment), prioritizing the protection of seals and extending the replacement cycle. This mode is manually set by the operator according to the equipment service strategy, or automatically recommended by the controller when the cumulative operating time of the seals exceeds 80% of the design life.

[0062] Q is a 4×4 weight matrix for state tracking terms. In this embodiment, under normal mode, Q=diag([10, 1, 0.5,0.1]) (the pressure difference term has the largest weight, and the temperature term has the smallest weight).

[0063] The second term is the control increment smoothing term: Δu(k+i)=u(k+i)-u(k+i-1) is the change in control quantity between adjacent time points, and R is the weight matrix of the control increment smoothing term (R=0.5 in this embodiment). This term is used to prevent the actuator from frequently starting and stopping and oscillating due to drastic changes in control quantity, and to extend the service life of the electric actuator.

[0064] The third term is the energy consumption suppression term for the actuator: S is the weight matrix of the energy consumption suppression term (S=0.1 in this embodiment). This term is used to reduce the continuous output of the actuator in steady state and save energy.

[0065] The fourth term is the seal life loss function: L loss (k) Based on Miner's linear cumulative damage theory. Specifically, the controller pre-stores the stress-strain constitutive relationship and fatigue characteristic curves (SN curves, i.e., stress-life curves, obtained through standard fatigue tests) of the sealing material. At each sampling moment, based on the current sealing force F... act Given temperature T(k), calculate the instantaneous stress amplitude σ of the sealing material. a (k), the fatigue life N corresponding to this stress amplitude is obtained by interpolation of the SN curve. f (σ a If the instantaneous damage rate is dD / dt, then the instantaneous damage rate is 1 / N. f (σ a(Assuming that the damage caused by each stress cycle accumulates linearly). Integrating the instantaneous damage rate over time in the control time domain yields the predicted cumulative lifetime loss in the future control time domain. .

[0066] The controller dynamically adjusts the weighting coefficient λ based on the cumulative lifetime loss prediction: when L loss (k) When it is less than the preset warning value (0.01% of the total design life of the seal in this embodiment), λ remains at the reference value (λ=1); when L loss (k) When the warning value is exceeded, λ increases adaptively (up to 10) to preferentially reduce the sealing force amplitude, thereby reducing the stress amplitude and extending the life of the seal.

[0067] The multi-objective optimization cost function is solved using a hybrid optimization strategy combining adaptive genetic algorithm and sequential quadratic programming. At each sampling time step: Global optimization using an adaptive genetic algorithm: First, coarse-grained optimization is performed within the global solution space. The population size is set to 50, and the chromosomes are encoded as control sequences U(k) = {u(k), u(k+1), …, u(k+N)}. c The real number encoding of} -1). Crossover probability P c And the probability of mutation P m Dynamic adjustment based on the current generation's fitness concentration: when the population fitness variance is small (indicating that the population tends to converge locally), increase P. c and P m To maintain population diversity; when fitness variance is large (indicating population dispersion), reduce P. c and P m To accelerate convergence. After 50 generations of evolution, a near-optimal solution U is obtained. GA (k).

[0068] Sequential quadratic programming with local fine search: with U GA (k) is the initial iteration point. A sequential quadratic programming algorithm is used to perform a local fine search near the constraint boundary to obtain the optimal control sequence U(k) = {u(k), u(k+1), …, u(k+N)} that satisfies all constraints. c -1)}. The sequential quadratic programming algorithm constructs a quadratic approximation of the Lagrange function and a linear approximation of the constraints in each iteration, solves the quadratic programming subproblem to obtain the search direction, and then performs a one-dimensional search along this direction to determine the step size until the convergence condition is met (in this embodiment, the convergence condition is that the gradient norm is less than 10). -4 ). Example

[0069] This embodiment is a specific implementation method for the output and execution of the optimal control sequence under constraints.

[0070] When solving optimization problems, the controller must satisfy the following three types of constraints: (a) Maximum output force constraint: The upper limit of the rated output force F of each pressure action unit max In this embodiment, the rated output force of each electric actuator is 500N, i.e., 0 ≤ F act(i) ≤ 500N (i=1,2,…,8).

[0071] (ii) Maximum compression stroke constraint: the maximum allowable compression displacement of the seal x max In this embodiment, the maximum allowable compression of the seal is 5mm (exceeding this value will result in permanent deformation or damage to the seal), i.e., 0 ≤ x i ≤ 5mm (i=1,2,…,8).

[0072] (III) Sealing pressure lower limit constraint: This is the most critical safety constraint. The sealing pressure lower limit is calculated in real time based on the current cable outer diameter and sealing cavity size according to the following formula: P sealmin = K safe × (P 外 -P 内 ) max + P margin Among them, P sealmin The minimum dynamic safety sealing pressure lower limit (unit: MPa); K safe As a safety margin factor, it is taken as 1.5 in this embodiment; (P 外 -P 内 ) max P represents the maximum predicted differential pressure (in MPa) in the LSTM differential pressure prediction sequence. margin The preset additional sealing safety margin is 0.2 MPa in this embodiment.

[0073] When solving the optimization problem, the controller transforms the lower limit constraint of the sealing specific pressure into a hard constraint on the control output u(k)—that is, in each optimization iteration, it is necessary to verify whether the predicted value of the sealing specific pressure calculated from the current control quantity satisfies P. seal (k+i|k) ≥ P sealmin (k+i). If not satisfied, the candidate solution is directly rejected (imposing an infinite cost on the solution). Sealing pressure P seal The conversion relationship between the compression displacement x and the pressure ratio is determined by the force-displacement-specific pressure calibration curve obtained from the compression test of the seal and is pre-stored in the controller.

[0074] After obtaining the optimal control sequence U(k), the controller outputs only the optimal control quantity u(k) at the current moment to the eight independent electric actuator drive circuits arranged circumferentially along the cable. Each electric actuator independently adjusts its output force according to its received drive current command, providing differentiated compensation for the sealing gap between the cable and the seal. It should be noted that u*(k) is the equivalent total control quantity. The controller also needs to allocate it to each independent electric actuator according to its current state and bias compensation amount. The allocation principle is: under the premise of meeting the total control quantity requirements, priority is given to increasing the output force of the electric actuator with the smallest current displacement and priority is given to decreasing the output force of the electric actuator with the largest current displacement, so that the displacement of each electric actuator tends to be consistent. Example

[0075] This embodiment is a specific implementation method for feedback correction and online identification of model parameters.

[0076] After executing the current control input, the controller acquires the actual output state values ​​(including actual differential pressure, actual sealing force, actual compression displacement, and actual temperature), compares the actual state values ​​with the predicted values ​​from the state-space prediction model, and calculates the prediction error e(k)=Z. act (k)-H· (k|k-1), where e(k) is the prediction error vector at time k, Z act (k) represents the actual measurement vector at time k. This prediction error is fed back to the controller for rolling correction of the prediction output at the next time step—specifically, in the prediction step at the next time step, an error correction term is added to the predicted value: corr (k+1|k)= (k+1|k)+K corr ·e(k), where K corr To correct the gain matrix (in this embodiment, it is taken as 0.5·I), corr (k+1|k) is the predicted value at time k+1 after error correction.

[0077] Simultaneously, the recursive least squares method is used to identify and update the system state transition matrix A and the disturbance input matrix W in the state-space prediction model online. Specifically, the state-space model is rewritten in linear regression form: X(k+1) = Θ·Φ(k) + ε(k) Where Θ=[A, W] is the parameter matrix to be identified, and Φ(k)=[X(k)] T , v(k)] T Let be the regression vector, and ε(k) be the modeling error. The update formula for the recursive least squares method is: K rls(k) = P rls (k-1)·Φ(k)·[1 + Φ(k) T ·P rls (k-1)·Φ(k)] -1 Θ^(k) = Θ^(k-1) + K rls (k)·[X(k) - Θ^(k-1)·Φ(k)] P rls (k) = [I - K rls (k)·Φ(k) T ]·P rls (k-1) Among them, K rls (k) is the RLS gain, P rls (k) is the covariance matrix (initialized to 10). 4 ·I), Θ^(k) are the parameter estimates at time k.

[0078] At each sampling time, the controller calculates the update amount ΔA = of the system state transition matrix A. (k) - (k-1), and calculate its Frobenius norm. Determine ||ΔA|| F Is it less than the preset convergence accuracy threshold ε (ε=10 in this embodiment)? -4 ). When ||ΔA|| F When the value of ε exceeds the preset stable time window (100 consecutive sampling cycles, or 10 seconds in this embodiment), the controller determines that the dynamic characteristics of the current sealing system tend to be stable, the model identification converges, and the currently identified model parameters are stored in non-volatile memory as the reference model parameters under this working condition.

[0079] When the external environmental pressure undergoes a sudden change and the magnitude of the change exceeds a preset value (1 MPa in this embodiment), the controller determines that the operating conditions have changed drastically and the original model parameters are no longer applicable. At this time, the controller forcibly resets the covariance matrix P of the recursive least squares method. rls The initial value (10) 4 •I), to accelerate the convergence speed of model parameters to new operating conditions. Example

[0080] This embodiment focuses on the specific implementation of the initial off-center load balancing and self-calibration steps.

[0081] After the cable installation is completed and before the dynamic sealing main control process is started, the controller first performs the initial off-center load balancing and self-calibration steps.

[0082] S4a, Initial Compression Displacement Acquisition: The controller reads the initial feedback value of the displacement sensor corresponding to each pressure action unit, and records it as... (i=1,2,…,8). Calculate the arithmetic mean of all initial feedback values. And determine the maximum initial displacement value. With minimum initial displacement value .

[0083] S4b, Off-center load determination: The controller calculates the difference Δx0 between the maximum and minimum initial displacement values. - And compare Δx0 with the preset allowable off-center load deviation threshold Δx allow In this embodiment, the comparison is performed using 0.3 mm. When Δx0 < Δx allow When the controller determines that the cable is properly centered within the inlet device, it directly enters the dynamic sealing main control process (i.e., jumps to step S1). When Δx0 ≥ Δx allow When the controller determines that the cable is eccentrically installed in the sealed cavity or is not uniformly compressed due to gravity sag, it proceeds to step S4c.

[0084] S4c, Self-calibrating equalization control: The controller generates an off-center load equalization pre-adjustment command based on the average value. As the target equilibrium displacement value for each pressure action unit, a proportional-integral-derivative (PID) control strategy is adopted to independently apply an equilibrium adjustment to each pressure action unit: Δu i = K p ·( - ) + K i ·∫( - )dt + K d ·d( - ) / dt Among them, K p K i K d These are the proportional, integral, and differential coefficients (K is taken after tuning in this embodiment). p =2.0, K i =0.5, K d =0.1). This control law drives the pressure actuator with an initial compression displacement that is too large to actively retract and release part of the sealing force, and drives the pressure actuator with an initial compression displacement that is too small to actively feed and increase the sealing force. The controller continues to execute this closed-loop regulation until the real-time displacement value x of each pressure actuator is reached. i All satisfy |x i - |<ε, where ε is the preset equalization convergence accuracy (0.05mm in this embodiment).

[0085] S4d, Equalization Reference Lock: When the displacement values ​​of each pressure actuator converge to within the equalization convergence accuracy ε, the controller calibrates the current displacement values ​​of each pressure actuator as the new initial reference state, stores the current output force values ​​of each pressure actuator as zero as the bias compensation amount (i.e., the current drive current value of each electric actuator is recorded as "zero bias"), and executes the generation and output of dynamic sealing compensation commands with this equalized state as the starting state. At the same time, the controller stores the bias compensation amount in non-volatile memory, and directly calls this bias compensation amount as the initial preload setting reference when installing cables of the same specification again, without having to repeat the complete self-calibration process. Example

[0086] This embodiment focuses on the integrated workflow applied to deep-sea submersibles. Taking the cable introduction device of a deep-sea submersible (maximum working depth 6000m, corresponding to an external water pressure of about 60MPa) as an example, the complete working process of the present invention is explained.

[0087] Initial installation phase: When the submersible leaves the factory, after the cable is threaded into the inlet device, the controller is powered on and started. First, it executes the initial off-center load balancing and self-calibration steps described in Example 8 to eliminate uneven circumferential sealing force caused by cable eccentricity. After calibration, the controller enters standby mode.

[0088] Descent Phase: The submersible begins its descent from the surface, with the external water pressure gradually increasing from 0.1 MPa. The controller executes steps S1 to S6 in a 100ms cycle. In step S3, the LSTM neural network predicts that the pressure difference will continue to increase within the next second based on the external pressure change trend over the past 10 seconds (pressure continuously rises during descent). In step S4, the MPC controller outputs a boost command in advance based on this prediction. In step S5, each electric actuator synchronously increases the compression of the seal, ensuring that the seal specific pressure remains above P. sealmin (The dynamically calculated value increases synchronously with the increase of the predicted pressure difference). Throughout the entire descent, the sealing force increases steadily with water depth, rather than remaining "fixed" as in traditional technologies, ensuring sealing reliability across the entire depth range.

[0089] Deep-sea cruising phase: The submersible cruises at a depth of 6000m, where the external water pressure is relatively stable. At this point, the LSTM prediction of future pressure difference changes in step S3 tends to zero, and the MPC controller output maintains the current control value in step S4. Since the seals undergo creep relaxation under constant compressive stress, the generalized Maxwell viscoelastic constitutive equation embedded in step S1 calculates the stress relaxation in real time, and a feedforward compensation term u is added to the control input.ff (k) actively increases the sealing force to compensate for the decrease in sealing specific pressure caused by relaxation. At the same time, the recursive least squares method in step S6 continuously identifies the model parameters online, ensuring that the prediction model always approximates the true dynamic characteristics of the sealing system.

[0090] Ascent Phase: As the submersible ascends from 6000m to the surface, the external water pressure gradually decreases. LSTM predicts that the pressure difference will decrease in the future, and the MPC controller outputs a pressure reduction command in advance accordingly. The electric actuators gradually retract, releasing some of the sealing force. This intelligent adjustment of "pressure reduction equals force reduction" avoids the seals being in an over-compressed state for a long time, significantly extending the service life of the seals.

[0091] Abnormal operating condition response: If a sudden pressure shock (such as a sudden pressure change caused by ocean currents) occurs during cruise, and the rate of change of pressure difference in the LSTM prediction sequence exceeds the 0.5 MPa / s threshold, the controller automatically switches to pressure surge suppression mode—the target pressure difference value is forcibly set to zero, and the weight of the pressure difference term is increased to four times that of the normal mode. The system eliminates the pressure difference as quickly as possible to ensure safety. If an external pressure sensor fails, the extended Kalman filter algorithm in step S2 automatically identifies the abnormal sensor by detecting the 3σ confidence interval of the innovation sequence and reduces its weight, switching to degradation estimation mode to ensure that the system can maintain reliable control even if some sensors fail.

[0092] Lifetime Management: The lifetime depletion function L in step S4 throughout the entire service life. loss (k) Continuous accumulation of fatigue damage to the seals. When the accumulated damage exceeds the preset warning value, λ adaptively increases to preferentially reduce the sealing force amplitude and actively slow down the aging rate of the seals. When the compression displacement of a certain electric actuator reaches the maximum stroke (5mm) and the pressure difference has not yet dropped to a safe range, the controller determines that the seal has reached its service life limit and issues an audible and visual alarm and a "Please replace the seal" prompt message through the warning module. Example

[0093] This embodiment focuses on parameter selection and applicable scope.

[0094] The specific parameters in the above embodiments (including but not limited to the sampling period T) s =100ms, prediction time domain N p =10, Control Time Domain N c =5. Number of pressure action units N=8, Safety margin factor K safe =1.5, Additional sealing safety margin P marginThe parameters (e.g., 0.2 MPa, equalization convergence accuracy ε = 0.05 mm) are preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can adapt the above parameters to meet the specific requirements of different application scenarios (e.g., explosion-proof electrical equipment, shipboard penetration, underground coal mine equipment, seabed observation networks, petrochemical pipelines, etc.).

[0095] For example, in applications with rapid pressure changes (such as cable entry devices in aircraft engine compartments), the sampling period can be appropriately reduced (e.g., T). s =20ms) and increase the prediction time domain (e.g., N) p =20); For applications where seal replacement costs are high (such as nodes in deep-sea observation networks), K can be appropriately increased in the low-loss, long-life mode. life A value (e.g., 5) can be used to further extend the life of the seal; for the introduction device of large diameter cables, the number of pressure action units can be appropriately increased (e.g., N=12 or 16) to ensure the uniformity of circumferential force application.

[0096] The above specific embodiments are merely several preferred embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A dynamic sealing control method for a pressure-adaptive cable introduction device, characterized in that the steps include... include: S1. Multi-source sensor information fusion and state-space modeling: using the pressure difference ΔP inside and outside the sealed cavity of the cable entry device and the actual output sealing force F of the sealing actuator. act The compression displacement x of the sealing component and the real-time temperature T inside the sealing cavity are used as state variables to establish a discrete state-space prediction model that reflects the dynamic characteristics of the sealing system: X(k+1) = A·X(k) +B·u(k) + W·v(k), where X(k) is the state vector at time k, u(k) is the control input, v(k) is the external environmental pressure disturbance, A is the system state transition matrix, B is the control input matrix, and W is the disturbance input matrix; S2. Optimal State Estimation Based on Extended Kalman Filter: Measurement data is collected using pressure sensors, displacement sensors, and temperature sensors located inside and outside the sealed cavity. The extended Kalman filter algorithm is then used to perform optimal state estimation on the state-space prediction model, filtering out sensor noise and outputting the optimal state estimate at time k. Simultaneously, the state estimation covariance matrix is ​​updated online recursively; S3. Future pressure differential prediction based on long short-term memory neural network: The controller retrieves the external environmental pressure value P from the historical time series. 外 and the internal sealing cavity pressure value P 内 Input into a pre-trained long short-term memory neural network model, and output the future N. p A time-domain pressure difference prediction sequence {ΔP(k+i|k), i=1, 2, …, N} p }, and use the predicted sequence as the reference input for model predictive control; S4. Model predictive control rolling time-domain multi-objective optimization: at each sampling time k, using the optimal state estimate... Using the pressure difference prediction sequence as the initial condition and the pressure difference prediction sequence as the reference trajectory, in the finite control time domain N c The following multi-objective optimization cost function J(k) is constructed and solved: The first item is the sealing status tracking item, X ref The first term is the reference trajectory for the desired sealed state; the second term is the control increment smoothing term; the third term is the actuator energy consumption suppression term; the fourth term is L. loss (k) is the seal life loss function based on Miner's linear cumulative damage theory, λ is the life loss weight coefficient; Q, R, and S are the corresponding weight matrices; S5. Optimal Control Sequence Output and Execution under Constraints: Under the conditions of satisfying the maximum output force constraint, maximum compression stroke constraint, and lower limit constraint of sealing specific pressure of the sealing actuator, solve for the optimal control sequence U(k) = {u(k), u(k+1), …, u(k+N)} that minimizes the cost function J(k). c -1)}, output the current optimal control quantity u*(k) to multiple independent pressure action units arranged along the circumference of the cable, and each pressure action unit independently performs dynamic adjustment of the sealing force; S6. Feedback Correction and Online Model Parameter Identification: After executing the current control variable, the actual output state value is collected, and the actual state value is compared with the predicted value of the prediction model. The prediction error is calculated and fed back to the controller to perform rolling correction on the predicted output. At the same time, the recursive least squares method is used to identify and update the system state transition matrix A and disturbance input matrix W in the state space prediction model online based on the real-time collected data, so that the model continuously approximates the dynamic characteristics of the actual sealed system.

2. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that, In step S4, the seal life loss function L loss The construction method of (k) is as follows: based on the stress-strain constitutive relationship of the sealing material and the pre-stored fatigue characteristic curve, the cumulative creep damage rate of the sealing component under the current sealing force and temperature conditions is calculated in real time. The instantaneous damage rate in the control time domain is integrated over time to obtain the cumulative life loss prediction in the future control time domain. The controller dynamically adjusts the weight coefficient λ according to the cumulative life loss prediction. When the predicted loss exceeds the preset warning value, λ is adaptively increased to preferentially reduce the amplitude of the sealing force.

3. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that: In step S4, the desired sealing state reference trajectory X ref The operating mode is adaptively switched according to the operating condition mode setting, which includes a normal voltage regulation mode, a pressure surge suppression mode, and a low-loss long-life mode; in the pressure surge suppression mode, X ref The target differential pressure value is forcibly set to zero, and the weight of the differential pressure term in the tracking term weight matrix Q is increased to K as in the normal mode. surge times, K surge ≥ 3; In the low-loss, long-life mode, the lifetime loss weighting coefficient λ increases to K in the conventional mode. life times, K life ≥ 2.

4. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that, In step S4, the solution of the multi-objective optimization cost function adopts a hybrid optimization strategy combining adaptive genetic algorithm and sequential quadratic programming. Specifically, at each sampling time, the adaptive genetic algorithm is first used to perform coarse-grained optimization in the global solution space to obtain an approximate optimal solution; then, with the approximate optimal solution as the initial iteration point, the sequential quadratic programming algorithm is used to perform a local search to obtain the optimal control sequence that satisfies the constraints; wherein, the crossover probability and mutation probability of the adaptive genetic algorithm are dynamically adjusted according to the fitness concentration of the current generation.

5. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that: In step S1, the state-space prediction model also embeds the generalized Maxwell viscoelastic constitutive equation of the sealing material as an intrinsic constraint condition of the model; the controller calculates the real-time stress relaxation of the sealing material based on the real-time temperature T and the current compression displacement x using the generalized Maxwell constitutive equation, and adds a feedforward compensation term u to the control input u(k). ff (k), feedforward compensation term u ff The magnitude of (k) is positively correlated with the real-time stress relaxation amount, so as to actively compensate for the decrease in sealing specific pressure caused by creep relaxation of the sealing material.

6. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that: In step S2, while performing state estimation, the extended Kalman filter algorithm calculates the covariance matrix of the innovation sequence and judges in real time whether there is a sudden abnormality or jamming fault in the external environmental pressure sensor based on the statistical characteristics of the innovation sequence. When the mean of the innovation sequence is detected to exceed the preset 3σ confidence interval, the controller determines that the current sensor data is abnormal, automatically reduces the fusion weight of the sensor in the state estimation, and switches to the degradation estimation mode based on the data of other healthy sensors.

7. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that, In step S3, the input feature vector of the Long Short-Term Memory Neural Network model includes: the external environmental pressure value, the internal sealing cavity pressure value, the temperature change rate, and the attitude angle information of the cable introduction device at M sampling times prior to the current time. The Long Short-Term Memory Neural Network model is a two-layer stacked structure, with each layer containing L LSTM neurons, and overfitting is prevented through a Dropout layer. The training data of this model comes from the actual operation history database of the sealing system and the extreme working condition extended sample set generated based on finite element simulation.

8. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that: In step S6, the recursive least squares method calculates the update amount ΔA of the system state transition matrix A at each sampling time, and determines ||ΔA|| F Is it less than the preset convergence accuracy threshold ε, where ||·|| F It is the Frobenius norm; when ||ΔA|| F When the value is less than ε and continues to exceed the preset stable time window, the controller determines that the dynamic characteristics of the current sealing system tend to be stable, the model identification converges, and the currently identified model parameters are stored in non-volatile memory as the reference model parameters under this working condition. When the external environmental pressure undergoes a step change and the change amplitude exceeds the preset amplitude, the controller forcibly resets the covariance matrix of the recursive least squares method to accelerate the convergence speed of the model parameters to the new working condition.

9. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that: In step S5, the maximum output force constraint is the upper limit F of the rated output force of each pressure action unit. max The maximum compression stroke constraint is the maximum allowable compression displacement of the seal x. max The lower limit constraint of the sealing specific pressure is calculated in real time based on the current cable outer diameter and sealing cavity size according to the following formula: P sealmin = K safe × (P 外 -P 内 ) max + P margin , where P sealmin K is the lower limit of the dynamic minimum safe sealing pressure. safe For a safety margin factor greater than 1, (P) 外 -P 内 ) max P represents the maximum predicted differential pressure value in the differential pressure prediction sequence. margin This is a preset additional sealing safety margin; when solving the optimization problem, the controller transforms the sealing specific pressure lower limit constraint into a hard constraint condition on the control output u(k), ensuring that the actual sealing specific pressure at any time is always not lower than this lower limit value.

10. The dynamic sealing control method for a pressure-adaptive cable introduction device according to claim 1, characterized in that, In step S4, before generating the dynamic sealing compensation command and driving each pressure action unit to output the sealing force, the controller also performs the following initial off-center load balancing and self-calibration steps: S4a, Initial Compression Displacement Acquisition: The controller reads the initial feedback value of the displacement sensor corresponding to each pressure action unit, and records it as... Let i = 1, 2, ..., N, and calculate the arithmetic mean of all initial feedback values. And determine the maximum initial displacement value. With minimum initial displacement value ; S4b, Off-center load determination: The controller calculates the difference Δx0 between the maximum and minimum initial displacement values. - And compare Δx0 with the preset allowable off-center load deviation threshold Δx allow Compare; when Δx0 < Δx allow When the controller determines that the cable is properly centered within the inlet device, it directly enters the dynamic sealing main control process; when Δx0 ≥ Δx allow When the controller determines that the cable is eccentrically installed in the sealed cavity or is not uniformly compressed due to gravity sag, it proceeds to step S4c. S4c, Self-calibrating equalization control: The controller generates an off-center load equalization pre-adjustment command based on the average value. As the target equilibrium displacement value for each pressure action unit, a proportional-integral-derivative control strategy is adopted, and an equilibrium adjustment amount Δu is independently applied to each pressure action unit. i = K p ·( - ) + K i ·∫( - )dt + K d ·d( - The pressure actuator with an initial compression displacement that is too large actively retracts to release part of the sealing force, while the pressure actuator with an initial compression displacement that is too small actively feeds to increase the sealing force, until the real-time displacement value x of each pressure actuator is reached. i All satisfy |x i - | < ε, where ε is the preset equilibrium convergence accuracy; S4d, Equalization Reference Lock: When the displacement values ​​of each pressure action unit converge to within the equalization convergence accuracy ε, the controller calibrates the current displacement values ​​of each pressure action unit as the new initial reference state, stores the current output force values ​​of each pressure action unit as the offset compensation amount, and executes the generation and output of dynamic sealing compensation commands with this equalized state as the starting state; at the same time, the controller stores the offset compensation amount in non-volatile memory, and directly calls the offset compensation amount as the initial preload setting reference when the same specification cable is installed again.

Images