Vacuum concentration real-time monitoring and intelligent optimization system

Abstract

The invention provides a vacuum concentration real-time monitoring and intelligent optimization system, which comprises a multi-mode sensing data acquisition module used for acquiring temperature, pressure, material concentration and flow velocity data in a vacuum concentration process in real time; the dynamic parameter optimization algorithm module is used for generating a control instruction based on a nonlinear time-varying model; the multi-modal sensing fusion module is used for integrating multi-source sensing data through a space-time incidence matrix; the real-time error correction module adopts a dual feedback mechanism to compensate system deviation; the intelligent decision engine module is used for executing dynamic priority task allocation and resource scheduling; the gradient compensation controller module adjusts the power and the heating rate of the vacuum pump according to real-time working conditions; and the self-learning knowledge base module is used for storing historical optimization paths and abnormal working condition feature vectors. According to the invention, precise monitoring and intelligent optimization of the vacuum concentration process can be realized, the concentration efficiency is improved, the energy consumption is reduced, the product quality stability is improved, and the adaptability and reliability of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation control and intelligent systems, and more specifically, to a vacuum concentration real-time monitoring and intelligent optimization system. Background Art

[0002] In modern industrial production, vacuum concentration technology is widely used in the pharmaceutical, chemical, and food industries for key process steps such as efficient solvent removal and increased material concentration. Traditional vacuum concentration systems typically rely on fixed process parameters and simple sensor feedback control, achieving the concentration process through preset parameters such as temperature, pressure, and flow rate. However, this fixed parameter control method is difficult to adapt to complex operating conditions, such as fluctuations in material properties and equipment aging, resulting in low concentration efficiency, high energy consumption, and difficulty in ensuring stable product quality.

[0003] In recent years, with the advancement of automation and sensor technologies, some improved vacuum concentration systems have begun to incorporate single-parameter feedback control or simple multi-parameter monitoring capabilities. While these systems have improved the controllability of the concentration process to a certain extent, they still have limitations. For example, they are generally only able to optimize a single parameter, lacking the ability to comprehensively analyze and dynamically optimize multimodal data, and are unable to adjust control strategies in real time to cope with complex operating conditions. Furthermore, the error correction mechanisms of existing systems are relatively simple, making it difficult to effectively compensate for system deviations, resulting in insufficient control accuracy.

[0004] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the existing vacuum concentration system is unable to deeply integrate and dynamically optimize multimodal data, lacks real-time error correction and self-learning capabilities, and is difficult to adapt to complex working conditions. Changes, resulting in low concentration efficiency, high energy consumption, and unstable product quality, cannot meet the requirements of modern industrial production for high precision, high efficiency and intelligence. Summary of the Invention

[0005] The present invention provides a vacuum concentration real-time monitoring and intelligent optimization system, comprising:

[0006] Multimodal sensor data acquisition module, used to obtain real-time temperature, pressure, material concentration and flow rate data during the vacuum concentration process;

[0007] Dynamic parameter optimization algorithm module, which generates control instructions based on nonlinear time-varying models;

[0008] Multimodal perception fusion module, which integrates multi-source sensor data through spatiotemporal correlation matrix;

[0009] Real-time error correction module, using a dual feedback mechanism to compensate for system deviations;

[0010] Intelligent decision engine module, which performs dynamic priority task allocation and resource scheduling;

[0011] Gradient compensation controller module, which adjusts the vacuum pump power and heating rate according to real-time working conditions;

[0012] The self-learning knowledge base module stores historical optimization paths and abnormal operating condition feature vectors.

[0013] Furthermore, the dynamic parameter optimization algorithm module includes the following steps:

[0014] a) Adaptive learning rate adjustment unit, calculates the current learning rate α(t) according to the formula:

[0015]

[0016] in:

[0017] α0 is the initial learning rate; λ is the decay factor, which is used to control the decay speed of the learning rate over time; η is the gain coefficient, which is used to adjust the dynamic adjustment amplitude of the learning rate; Δx i (t) is the instantaneous change of the i-th state variable at time t; n is the total number of state variables;

[0018] b) Dynamic weight factor generation unit, which calculates the weight matrix β(t) by the following formula:

[0019]

[0020] Among them, σ is the sigmoid function, which is used to map the input to the (0,1) interval; w j is the weight coefficient of the jth feature; f j (t) is the eigenvalue of the jth feature at time t; g k (t) is the gradient of the kth constraint at time t; m is the total number of features; k is the index of the constraint;

[0021] c) Gradient compensation amount generating unit, generates compensation amount ΔP(t) according to the formula:

[0022]

[0023] Where L is the loss function used to measure the error of the model; θ is the model parameter; J(t) is the real-time objective function used to optimize the control process; is the gradient of the objective function.

[0024] Furthermore, in the adaptive learning rate adjustment unit:

[0025] The decay factor λ is updated according to the following formula:

[0026]

[0027] Where: min : minimum value of the attenuation factor; λ max : Maximum value of the attenuation factor; E(t): Cumulative error, which represents the cumulative error of the system at time t; ξ: Adjustment coefficient, which is used to control the dynamic change speed of the attenuation factor;

[0028] The gain coefficient η is dynamically adjusted according to the historical optimization trajectory, as shown in the following formula:

[0029]

[0030] Where T is the time window length, which is used to determine the range of historical data; Δx(k) is the change in the state variable at time k.

[0031] Furthermore, the real-time error correction module includes:

[0032] a) Error double verification unit, which determines the effective error amount by the following formula:

[0033]

[0034] Where, E(t) is the current error; is the historical error mean; σ E is the standard deviation of the historical error; is the rate of change of error; θ th is the threshold of the error change rate;

[0035] b) Dynamic compensation strategy unit generates compensation signal according to the following formula:

[0036]

[0037] Among them, K p is the proportional coefficient; K i is the integral coefficient; K d is the differential coefficient; τ is the integration time constant;

[0038] c) The anti-saturation limiting unit performs compensation constraint according to the following formula:

[0039]

[0040] Among them, C max is the maximum value of compensation.

[0041] Furthermore, the time-varying coefficients are updated in the following manner:

[0042] The proportionality factor is:

[0043] K p (t) = μ·(1e-v·t )

[0044] Among them, μ is the adjustment parameter of the proportional coefficient; ν is the time decay factor of the proportional coefficient;

[0045] The integration coefficient is:

[0046]

[0047] Wherein, ρ is the adjustment parameter of the integral coefficient; T is the length of the integral time window;

[0048] The differential coefficient is:

[0049]

[0050] Among them, κ is the adjustment parameter of the differential coefficient; is the rate of change of error.

[0051] Furthermore, the multimodal perception fusion module performs the following operations:

[0052] a) Constructing a spatiotemporal correlation matrix:

[0053]

[0054] Among them, x i 、x j is the measurement data of different sensors at time t; Δt is the time lag; N is the length of the time series;

[0055] b) Generate fused feature vector:

[0056]

[0057] Among them, w i is the weight coefficient of principal component analysis (PCA); v j is the weight coefficient of the fast Fourier transform (FFT); n is the number of PCA features; m is the number of FFT features;

[0058] c) Perform abnormal feature suppression:

[0059]

[0060] Among them, ⊙ is the Hadamard product; F is the mean of historical features; ζ is the suppression factor, which is used to adjust the suppression strength of abnormal features; σ is the sigmoid function, which is used to map the input to the (0,1) interval; ReLU is the rectified linear unit, which is used to activate abnormal features.

[0061] Furthermore, the time lag Δt is dynamically adjusted according to the following formula:

[0062]

[0063] Among them, Δt0 is the reference time lag; c i is the differential order weight coefficient; is the i-th order time derivative of the sensor data.

[0064] Furthermore, the self-learning knowledge base module includes:

[0065] a) Optimization path backtracking unit (710), storing the historical optimal parameter set θ*(t) and the corresponding objective function value J*(t);

[0066] b) an abnormal pattern encoding unit (720), which maps the fault features into a 128-dimensional sparse coding vector;

[0067] c) Dynamic retrieval unit (730), matches historical cases according to the following similarity formula:

[0068]

[0069] Among them, Q(t) is the current eigenvector; H k is the characteristic vector of the kth historical case; t k is the time index of the k-th historical case.

[0070] Furthermore, the sparse coding vector is generated by:

[0071] Building an overcomplete dictionary Where d = 64, K = 256;

[0072] Solve for the sparse coefficient vector α:

[0073]

[0074] Among them, x is the input feature vector; λ and μ are sparse constraint coefficients used to control the sparsity degree.

[0075] Furthermore, the gradient compensation controller module performs:

[0076] a) Calculation of vacuum pump power adjustment:

[0077]

[0078] Among them, P v is the vacuum pump power; is the power regulation threshold; To compensate for the rate of change of the signal;

[0079] b) Generation of heating rate compensation:

[0080]

[0081] Among them, τ h is the heating time constant; H is the heating rate; is the symbol of the objective function gradient;

[0082] c) Actuator linkage control:

[0083] U(t)=softmax([ΔP v (t),ΔH(t)])·[P v,max ,H max ] T

[0084] Among them, P v,max is the maximum value of the vacuum pump power; H max is the maximum value of the heating rate; softmax: softmax function, used to normalize the output.

[0085] According to the above-mentioned embodiments of the present invention, there are at least the following beneficial effects: the vacuum concentration real-time monitoring and intelligent optimization system can obtain multi-dimensional data such as temperature, pressure, material concentration and flow rate in the vacuum concentration process in real time through the multi-modal sensor data acquisition module, providing comprehensive data support for precise control. The dynamic parameter optimization algorithm module generates control instructions based on a nonlinear time-varying model. Combined with adaptive learning rate adjustment and dynamic weight factor generation, it can flexibly respond to complex operating conditions and achieve precise parameter optimization. The real-time error correction module uses a dual feedback mechanism to compensate for system deviations, which can effectively improve control accuracy and reduce control deviations caused by sensor errors or operating condition fluctuations. The multi-modal perception fusion module integrates multi-source sensor data through a spatiotemporal correlation matrix and generates a fusion feature vector, which can enhance the system's perception ability and anti-interference performance. In addition, the self-learning knowledge base module stores historical optimization paths and abnormal operating condition feature vectors, which can provide experience references for the system and further enhance the system's intelligence level and adaptability.

[0086] The gradient compensation controller module adjusts the vacuum pump power and heating rate based on real-time operating conditions, enabling dynamic regulation of the equipment's operating status, improving concentration efficiency and reducing energy consumption. The intelligent decision-making engine module performs dynamic priority task allocation and resource scheduling, optimizing system resource utilization and improving overall operational efficiency. Functions such as abnormal feature suppression and sparse coding vector generation further enhance the system's fault diagnosis and handling capabilities, ensuring stable operation under complex operating conditions. Through the synergistic effect of these modules, the system can effectively address existing technological issues and meet the high-precision, high-efficiency, and intelligent requirements of modern industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0088] Figure 1 A schematic structural diagram of a vacuum concentration real-time monitoring and intelligent optimization system provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0089] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0090] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0091] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0092] Reference below Figure 1 , Figure 1 This is a schematic diagram of the structure of a vacuum concentration real-time monitoring and intelligent optimization system provided by one embodiment of the present invention. Figure 1 As shown, a vacuum concentration real-time monitoring and intelligent optimization system 100 includes:

[0093] Multimodal sensor data acquisition module 101, used to obtain temperature, pressure, material concentration and flow rate data during the vacuum concentration process in real time;

[0094] Dynamic parameter optimization algorithm module 102 generates control instructions based on the nonlinear time-varying model;

[0095] The multimodal sensing fusion module 103 integrates multi-source sensor data through a spatiotemporal correlation matrix;

[0096] The real-time error correction module 104 uses a dual feedback mechanism to compensate for system deviations;

[0097] Intelligent decision engine module 105, performs dynamic priority task allocation and resource scheduling;

[0098] Gradient compensation controller module 106 adjusts vacuum pump power and heating rate according to real-time working conditions;

[0099] The self-learning knowledge base module 107 stores historical optimization paths and abnormal operating condition feature vectors.

[0100] It should be noted that this system includes a multimodal sensor data acquisition module, which is responsible for acquiring real-time data on temperature, pressure, material concentration, and flow rate during the vacuum concentration process. The multimodal sensor data acquisition module here refers to a system integrating multiple sensors that can simultaneously monitor multiple physical quantities, providing comprehensive data support for subsequent optimization and control. The dynamic parameter optimization algorithm module generates control instructions based on a nonlinear, time-varying model. This means the system can adapt control strategies based on complex, time-varying models to varying operating conditions. The multimodal perception fusion module integrates multi-source sensor data using a spatiotemporal correlation matrix. Its purpose is to fuse data from different sensors to more accurately reflect the system's operating status. The real-time error correction module utilizes a dual feedback mechanism to compensate for system deviations. This dual feedback mechanism effectively corrects system deviations during operation and improves control accuracy. The intelligent decision engine module performs dynamic priority task allocation and resource scheduling, rationally allocating tasks and resources based on the system's real-time needs and optimizing system efficiency. The gradient compensation controller module adjusts the vacuum pump power and heating rate based on real-time operating conditions to ensure optimal equipment operation. The self-learning knowledge base module stores historical optimization paths and abnormal operating condition feature vectors. By recording historical data and abnormal conditions, the system can continuously learn and optimize its own control strategy.

[0101] Specifically, the sensors in the multimodal sensing data acquisition module may include temperature sensors, pressure sensors, concentration sensors, and flow sensors. These sensors are located at key locations within the vacuum concentrator, collecting data in real time and transmitting it to the system. For example, temperature sensors can be thermocouples or RTDs to measure the temperature of the material during the concentration process; pressure sensors can be piezoresistive or capacitive sensors to monitor pressure changes within the system. Material concentration can be measured using a refractometer or conductivity sensor, while flow rate can be obtained using an electromagnetic flowmeter or vortex flowmeter. The nonlinear time-varying model in the dynamic parameter optimization algorithm module refers to a system model whose parameters change with time and operating conditions. This model can more accurately describe the complex dynamic behavior of the vacuum concentrator process. For example, during the concentration process, as material concentration increases, its physical properties may change, leading to changes in heat and mass transfer efficiency. The nonlinear time-varying model can capture these changes and generate control instructions accordingly. The spatiotemporal correlation matrix is the core of the multimodal sensing fusion module. It integrates dispersed data into a coherent whole by calculating the temporal and spatial correlations between different sensor data. For example, when the data from the temperature sensor and pressure sensor show a certain temporal correlation, the system can use the spatiotemporal correlation matrix to determine that this is a normal change caused by the heating process, rather than an abnormal situation. The dual feedback mechanism in the real-time error correction module includes direct feedback and indirect feedback. Direct feedback can be based on the deviation between the sensor measurement value and the set value, while indirect feedback can be based on the long-term trend of the system's operating status. For example, if the system detects a deviation between the temperature sensor's measurement value and the set value, it will immediately make adjustments through direct feedback. At the same time, the system will also determine whether this deviation falls within the normal fluctuation range based on historical data. If it is an abnormal deviation, further correction will be made through indirect feedback.

[0102] Preferably, the sensor layout in the multimodal sensing data acquisition module can be optimized according to the specific vacuum concentration equipment. For example, in large-scale equipment, the number and distribution density of sensors can be increased to more comprehensively monitor the system status. The nonlinear time-varying model in the dynamic parameter optimization algorithm module can be modeled by a machine learning algorithm, such as using a neural network or support vector machine algorithm to train the model based on historical data so that it can more accurately predict the behavior of the system. The construction of the spatiotemporal correlation matrix can introduce more feature extraction methods, such as principal component analysis (PCA) or fast Fourier transform (FFT), to extract key features from the data and further improve the effect of data fusion. The dual feedback mechanism in the real-time error correction module can introduce a fuzzy control algorithm, judge the severity of the error based on fuzzy logic, and adjust the compensation strategy accordingly, so that the system can respond more flexibly to complex working conditions.

[0103] In some embodiments, the dynamic parameter optimization algorithm module includes the following steps:

[0104] a) Adaptive learning rate adjustment unit, calculates the current learning rate α(t) according to the formula:

[0105]

[0106] in:

[0107] α0 is the initial learning rate; λ is the decay factor, which is used to control the decay speed of the learning rate over time; η is the gain coefficient, which is used to adjust the dynamic adjustment amplitude of the learning rate; Δx i (t) is the instantaneous change of the i-th state variable at time t; n is the total number of state variables.

[0108] b) Dynamic weight factor generation unit, which calculates the weight matrix β(t) by the following formula:

[0109]

[0110] Among them, σ is the sigmoid function, which is used to map the input to the (0,1) interval; w j is the weight coefficient of the jth feature; f j (t) is the eigenvalue of the jth feature at time t; g k (t) is the gradient of the kth constraint at time t; m is the total number of features; k is the index of the constraint.

[0111] c) Gradient compensation amount generating unit, generates compensation amount ΔP(t) according to the formula:

[0112]

[0113] Where L is the loss function used to measure the error of the model; θ is the model parameter; J(t) is the real-time objective function used to optimize the control process; is the gradient of the objective function.

[0114] It should be noted that the dynamic parameter optimization algorithm module is one of the core parts of this system. It realizes the optimization control of the vacuum concentration process through the adaptive learning rate adjustment unit, the dynamic weight factor generation unit and the gradient compensation amount generation unit. The function of the adaptive learning rate adjustment unit is to dynamically adjust the learning rate according to the current state of the system so that the algorithm can converge to the optimal solution more efficiently. The dynamic weight factor generation unit dynamically assigns weights according to the gradient of the eigenvalues and constraints, making the optimization process more in line with the actual working conditions. The gradient compensation amount generation unit calculates the gradient of the objective function and generates a compensation amount to correct the model error, thereby improving the control accuracy. These units work together to enable the system to achieve dynamic optimization under complex working conditions.

[0115] Specifically, the learning rate in the adaptive learning rate adjustment unit is a key parameter used to control the convergence speed of the optimization algorithm. The initial learning rate η0 is the learning rate at the beginning of the algorithm, and the attenuation factor α and the gain coefficient β are used to adjust the change of the learning rate over time. The instantaneous change Δxi(t) of the state variable reflects the state change of the system at a certain moment, and the total number of state variables N determines the complexity of the system. The weight matrix W(t) in the dynamic weight factor generation unit is dynamically generated based on the eigenvalues and constraint gradients, where the sigmoid function is used to map the weights to the (0,1) interval to ensure the rationality of the weights. The total number of features M and the index of the constraint K respectively represent the number of features and the type of constraints considered in the optimization process. The loss function L in the gradient compensation generation unit is an indicator to measure the model error, the model parameter θ is the parameter that needs to be adjusted during the optimization process, the real-time objective function f(t) and its gradient It is used to guide the optimization direction.

[0116] Preferably, the attenuation factor α in the adaptive learning rate adjustment unit can be dynamically adjusted according to the accumulated error E(t) of the system to adapt to different working conditions. For example, when the system error is large, the attenuation factor can be appropriately reduced to speed up the adjustment of the learning rate; when the error is small, the attenuation factor can be increased to stabilize the learning rate. The gain coefficient β can be dynamically adjusted according to the historical optimization trajectory, for example, by calculating the mean and variance of the historical state variable changes to determine its value. The weight coefficient wj in the dynamic weight factor generation unit can be adjusted according to the importance and relevance of the features, for example, by optimizing through expert knowledge or machine learning algorithms. The loss function L in the gradient compensation generation unit can use common functions such as mean square error or cross entropy, and the specific choice depends on the optimization objective and system characteristics. In addition, a regularization term can be introduced to prevent overfitting and further improve the robustness of the optimization algorithm.

[0117] In some embodiments, in the adaptive learning rate adjustment unit:

[0118] The decay factor λ is updated according to the following formula:

[0119]

[0120] Where: min : minimum value of the attenuation factor; λ max : The maximum value of the attenuation factor; E(t): The cumulative error, which represents the cumulative error of the system at time t; ξ: The adjustment coefficient, which is used to control the dynamic change speed of the attenuation factor.

[0121] The gain coefficient η is dynamically adjusted according to the historical optimization trajectory, as shown in the following formula:

[0122]

[0123] Where T is the time window length, which is used to determine the range of historical data; Δx(k) is the change in the state variable at time k.

[0124] It should be noted that the decay factor and gain coefficient in the adaptive learning rate adjustment unit are important components of the dynamic parameter optimization algorithm module. Their function is to dynamically adjust the learning rate based on the system's operating state to adapt to varying operating conditions and improve optimization efficiency. The decay factor update formula takes into account the accumulated error and the adjustment coefficient, which is used to control the dynamic change rate of the learning rate over time. The gain coefficient is dynamically adjusted based on the historical optimization trajectory, reflecting the system's sensitivity to changes in state variables. In this way, the learning rate can be adaptively adjusted according to the actual operating conditions of the system, ensuring that the optimization algorithm remains efficient and stable under complex operating conditions.

[0125] Specifically, the update formula of the attenuation factor is

[0126]

[0127] In, α min and α max Respectively represent the minimum and maximum values of the attenuation factor, which are used to limit the range of change of the attenuation factor; E(t) is the cumulative error of the system at time t, reflecting the degree of accumulation of system errors; γ is the adjustment coefficient, which is used to control the speed of dynamic change of the attenuation factor. Dynamic adjustment formula of gain coefficient

[0128]

[0129] In this example, τ represents the time window length, which determines the range of historical data. Δx(k) is the change in the state variable at time k, reflecting the state of the system. By setting these parameters, the attenuation factor and gain coefficient can be dynamically adjusted based on the system's cumulative error and historical state changes, thereby achieving adaptive optimization of the learning rate.

[0130] Preferably, the minimum value of the attenuation factor α min and the maximum value α max It can be set according to the actual needs of the system. For example, when the error is large, α min Set it to a smaller value to speed up the adjustment of the learning rate; when the error is small, you can increase α appropriately. maxTo maintain the stability of the learning rate. The adjustment coefficient γ can be determined through experimentation or empirical formulas, for example, by selecting the optimal value through cross-validation. The time window length τ of the gain coefficient can be adjusted based on the dynamic characteristics of the system. For rapidly changing systems, a shorter time window can be set, while for relatively stable systems, a longer time window can be set. In addition, adaptive adjustment mechanisms can be introduced, such as dynamically adjusting the time window length based on the system's real-time error, to further improve the system's adaptability.

[0131] In some embodiments, the real-time error correction module includes:

[0132] a) Error double verification unit, which determines the effective error amount by the following formula:

[0133]

[0134] Where, E(t) is the current error; is the historical error mean; σ E is the standard deviation of the historical error; is the rate of change of error; θ th is the threshold of the error change rate.

[0135] b) Dynamic compensation strategy unit generates compensation signal according to the following formula:

[0136]

[0137] Among them, K p is the proportional coefficient; K i is the integral coefficient; K d is the differential coefficient; τ is the integration time constant.

[0138] c) The anti-saturation limiting unit performs compensation constraint according to the following formula:

[0139]

[0140] Among them, C max is the maximum value of compensation.

[0141] It should be noted that the real-time error correction module is a key component of the system for improving control accuracy and stability. It uses a dual error verification unit, a dynamic compensation strategy unit, and an anti-saturation limiting unit to achieve real-time monitoring and correction of system errors. The dual error verification unit determines the effective error amount by setting a threshold, avoiding misjudgments caused by sensor noise or abnormal data. The dynamic compensation strategy unit generates a compensation signal based on the error amount to correct system deviations. The anti-saturation limiting unit constrains the compensation amount to prevent excessive compensation signals from causing system instability. Through the synergistic effect of these three units, the system can maintain high-precision control performance under complex operating conditions.

[0142] Specifically, the formula in the error double verification unit

[0143]

[0144] In the equation, Δ(t) represents the current error, σ is the mean of historical errors, which is used to reflect the average level of errors; σ is the standard deviation of historical errors, which is used to measure the degree of error fluctuation. is the rate of change of the error, which is used to judge the dynamic change trend of the error; Δ thresh Is the threshold value of the error change rate, which is used to distinguish normal error changes from abnormal fluctuations. The compensation signal formula in the dynamic compensation strategy unit is

[0145]

[0146] In, K p , K i and K d are the proportional, integral and differential coefficients, respectively, used to adjust the size and dynamic characteristics of the compensation signal; τ is the integral time constant, used to control the response speed of the integral term.

[0147]

[0148] C max It is the maximum value of the compensation amount, which is used to limit the amplitude of the compensation signal to prevent the system from over-compensation.

[0149] Preferably, the threshold Δ in the error double verification unit thresh It can be dynamically adjusted according to the actual operating data of the system. For example, by monitoring the rate of change of the error in real time and combining the statistical characteristics of historical data, it is possible to adaptively adjust Δ thresh The value of , in order to better adapt to the error changes under different working conditions. The proportional coefficient K in the dynamic compensation strategy unit p , integral coefficient K i and differential coefficient K dTuning can be performed through experiments or model-based optimization methods. For example, the Ziegler-Nichols method or genetic algorithm can be used to optimize these parameters to improve the compensation performance of the system. In addition, the integral time constant τ can be adjusted according to the dynamic characteristics of the system. For rapidly changing systems, a shorter integral time constant can be set to speed up the response; for relatively stable systems, the integral time constant can be appropriately increased to improve the compensation accuracy. The maximum compensation value C in the anti-saturation limiting unit max It can be set according to the actual operating range of the system and the characteristics of the equipment. For example, the maximum allowable compensation amount of the system can be determined through experiments to ensure that the system operates within a safe range.

[0150] In some embodiments, the time-varying coefficients are updated as follows:

[0151] The proportionality factor is:

[0152] K p (t) = μ·(1e -v·t )

[0153] Among them, μ is the adjustment parameter of the proportional coefficient; ν is the time decay factor of the proportional coefficient.

[0154] The integration coefficient is:

[0155]

[0156] Where ρ is the adjustment parameter of the integral coefficient; T is the length of the integral time window.

[0157] The differential coefficient is:

[0158]

[0159] Among them, κ is the adjustment parameter of the differential coefficient; is the rate of change of error.

[0160] It should be noted that the proportional coefficient, integral coefficient and differential coefficient are key parameters used to adjust the compensation signal in the real-time error correction module. Its dynamic adjustment mechanism can optimize the compensation strategy according to the real-time state of the system, thereby improving the adaptability and control accuracy of the system. p Instant response for direct adjustment of error, integral coefficient K i Used to eliminate the cumulative effect of system deviation, and the differential coefficient K d The coefficients are used to suppress rapid changes in errors and prevent system oscillation. By dynamically adjusting these coefficients, the system can achieve more accurate error compensation under different working conditions.

[0161] Specifically, the proportionality coefficient K pDynamic adjustment formula

[0162] K p (t) = K p0 ·(1-γ p t)

[0163] In, K p0 is the initial value of the proportional coefficient, which determines the initial response strength of the system to the error; p It is the time decay factor of the proportional coefficient, which is used to control the decay speed of the proportional coefficient over time, thereby avoiding system instability caused by overcompensation. i Dynamic adjustment formula

[0164]

[0165] In, K i0 is the initial value of the integral coefficient, which is used to determine the initial strength of the integral term; β i is the adjustment parameter of the integral coefficient, which is used to adjust the weight of the integral term according to the historical sum of squares of the error; τ i Is the length of the integration time window, which is used to determine the range of historical data considered. Differentiation coefficient K d Dynamic adjustment formula

[0166]

[0167] In, K d0 is the initial value of the differential coefficient, which is used to determine the initial strength of the differential term; It is the rate of change of the error, which is used to adjust the weight of the differential term according to the dynamic change of the error, so as to better suppress the rapid change of the error.

[0168] Preferably, the initial value K of the proportional coefficient p0 It can be set according to the static gain of the system and the desired response speed. For example, in a fast response system, a larger K p0 value to enhance the system's immediate response capability; in a slowly changing system, a smaller K p0 The value should be set to avoid over-adjustment. The adjustment parameter β of the integral coefficient i It can be adjusted by the error characteristics of the system. For example, when the system has a large steady-state error, β can be increased. i To enhance the effect of the integral term. The integral time window length τ i The selection can be made based on the dynamic characteristics of the system. For rapidly changing systems, a shorter τ i To improve the response speed; for slowly changing systems, a longer τ can be selected i To improve the compensation accuracy. The initial value of the differential coefficient K d0It can be set according to the damping characteristics of the system. For example, if the system is prone to oscillation, a larger K d0 In addition, adaptive mechanisms can be introduced, such as dynamically adjusting these parameters according to the real-time performance indicators of the system, to further improve the adaptability and robustness of the system.

[0169] In some embodiments, the multimodal perception fusion module performs the following operations:

[0170] a) Constructing a spatiotemporal correlation matrix:

[0171]

[0172] Among them, x i 、x j is the measurement data of different sensors at time t; Δt is the time lag; N is the length of the time series.

[0173] b) Generate fused feature vector:

[0174]

[0175] Among them, w i is the weight coefficient of principal component analysis (PCA); v j is the weight coefficient of the fast Fourier transform (FFT); n is the number of PCA features; m is the number of FFT features.

[0176] c) Perform abnormal feature suppression:

[0177]

[0178] Among them, ⊙ is the Hadamard product; F is the mean of historical features; ζ is the suppression factor, which is used to adjust the suppression strength of abnormal features; σ is the sigmoid function, which is used to map the input to the (0,1) interval; ReLU is the rectified linear unit, which is used to activate abnormal features.

[0179] It should be noted that the multimodal perception fusion module is a key component of the system for integrating multi-source sensor data, improving data quality and perception capabilities. By constructing a spatiotemporal correlation matrix and generating fusion feature vectors, this module effectively fuses data from different sensors and further enhances system robustness through anomalous feature suppression. The spatiotemporal correlation matrix captures the temporal and spatial correlations between sensor data, while the fusion feature vectors integrate the features of multi-source data, providing a more comprehensive basis for subsequent optimization and decision-making. In this way, the system can more accurately perceive and understand the state of the vacuum concentration process under complex operating conditions.

[0180] Specifically, the construction formula of the spatiotemporal correlation matrix is

[0181]

[0182] In, x i and x j Represents the measurement data of different sensors at time t, which is used to reflect the correlation between sensors; τ is the time lag, which is used to capture the time delay effect of data; T is the length of the time series, which determines the time range considered. The generation formula of the fusion feature vector

[0183]

[0184] In, α i and β j are the weight coefficients of principal component analysis (PCA) and fast Fourier transform (FFT), which are used to adjust the importance of different features; m and n are the number of PCA and FFT features, respectively, reflecting the dimension of the fusion feature. Abnormal feature suppression formula

[0185]

[0186] In the equation, ⊙ represents the Hadamard product, which is used for element-by-element multiplication; σ is the sigmoid function, which is used to map the input to the (0, 1) interval; λ is the suppression factor, which is used to adjust the suppression strength of abnormal features; It is the mean of historical characteristics, which is used to reflect the characteristic level under normal conditions.

[0187] Preferably, the time lag τ can be dynamically adjusted according to the dynamic characteristics of the sensor data. For example, for fast-changing signals, a smaller τ can be set to capture short-term correlations; while for slow-changing signals, τ can be appropriately increased to reflect long-term trends. The weight coefficient α of PCA and FFT is i and β j Dynamic adjustments can be made using optimization algorithms (such as genetic algorithms or particle swarm optimization) to ensure that the fused feature vector better reflects the actual state of the system. Furthermore, the suppression factor λ can be adaptively adjusted based on the system's noise level and the strength of the abnormal features. For example, when the system is noisy, λ can be appropriately increased to enhance the suppression of abnormal features. Alternatively, deep learning methods (such as convolutional neural networks) can be introduced to replace traditional PCA and FFT to further improve the performance and robustness of feature extraction.

[0188] In some embodiments, the time lag Δt is dynamically adjusted according to the following formula:

[0189]

[0190] Among them, Δt0 is the reference time lag; c i is the differential order weight coefficient; is the i-th order time derivative of the sensor data.

[0191] It's important to note that the dynamic adjustment mechanism for time lag is a crucial component of the multimodal perception fusion module. Its purpose is to adaptively adjust the time lag based on the dynamic characteristics of sensor data, thereby better capturing temporal correlations between data. The adjustment formula for time lag τ involves the baseline time lag τ0, the differential order weight coefficient β, and the higher-order time derivatives of the sensor data. These parameters collectively determine the dynamics of the time lag. This allows the system to automatically adjust the time lag based on real-time data changes, thereby improving data fusion accuracy and overall system performance.

[0192] Specifically, the reference time lag τ0 is the initial value of the time lag, which is usually pre-set according to the sampling frequency of the sensor and the dynamic characteristics of the system. For example, in a high-frequency sampling system, the reference time lag can be set to a smaller value to capture rapidly changing signal characteristics; while in a low-frequency sampling system, the value can be appropriately increased. The differential order weight coefficient β is used to adjust the influence of the high-order time derivatives of the sensor data on the time lag. High-order time derivatives of sensor data This reflects the changing trends and acceleration of the data. By incorporating this information into the time lag adjustment formula, the system can better adapt to dynamic changes in the data. For example, when the higher-order derivatives of sensor data are large, indicating that the data is changing rapidly, the time lag can be dynamically reduced through the adjustment formula to more closely track data changes.

[0193] Preferably, the reference time lag τ0 can be determined experimentally or through system identification methods. For example, by analyzing the correlation curve of sensor data and finding the point where the data correlation significantly decreases, an appropriate reference time lag can be determined. The differential order weight coefficient β can be adjusted according to the system noise level and the severity of data changes. In systems with high noise, the value of β can be appropriately reduced to reduce the interference of higher-order derivatives on the time lag adjustment; in systems with drastic data changes, the value of β can be increased to enhance the time lag's ability to respond to dynamic changes. In addition, the time lag adjustment formula can also introduce an adaptive mechanism, such as dynamically optimizing the value of β through a machine learning algorithm, to further improve the system's adaptability and robustness. As an alternative, one can consider introducing time series analysis methods (such as the autoregressive moving average model ARMA) to predict future trends in sensor data and dynamically adjust the time lag accordingly to achieve more accurate capture of time correlation.

[0194] In some embodiments, the self-learning knowledge base module includes:

[0195] a) Optimization path backtracking unit (710), storing the historical optimal parameter set θ*(t) and the corresponding objective function value J*(t);

[0196] b) an abnormal pattern encoding unit (720), which maps the fault features into a 128-dimensional sparse coding vector;

[0197] c) Dynamic retrieval unit (730), matches historical cases according to the following similarity formula:

[0198]

[0199] Among them, Q(t) is the current eigenvector; H k is the characteristic vector of the kth historical case; t k is the time index of the k-th historical case.

[0200] It should be noted that the self-learning knowledge base module is a key part of the system for storing and retrieving historical optimization paths and abnormal operating condition feature vectors. It realizes the effective management and utilization of historical data through the optimization path backtracking unit, the abnormal pattern encoding unit and the dynamic retrieval unit. The optimization path backtracking unit is used to store the historical optimal parameter set and its corresponding objective function value, providing an optimization reference for the system; the abnormal pattern encoding unit maps the fault characteristics into a sparse coding vector, which facilitates the rapid retrieval and identification of abnormal patterns; the dynamic retrieval unit matches historical cases through a similarity formula, providing decision support for the system when facing similar operating conditions. Through the synergistic effect of these units, the system can continuously learn and optimize its own control strategy and improve its adaptability to complex operating conditions.

[0201] Specifically, the historical optimal parameter set θ stored in the optimization path backtracking unit is * (t) and the objective function value f * (t) is the system's optimized parameters and performance indicators during historical operation. These data provide an important reference for dynamic adjustment of the system. The abnormal pattern encoding unit maps the fault characteristics into a 128-dimensional sparse coding vector. This high-dimensional sparse representation can effectively compress the fault feature information while retaining key features for fast retrieval and matching. The similarity formula of the dynamic retrieval unit is:

[0202]

[0203] In this example, F(t) is the current eigenvector, F kis the feature vector of the kth historical case. By calculating the similarity between the current feature and the historical feature, the system can quickly find the historical case that is most similar to the current working condition, thus providing a reference for optimal control. k It is used to introduce the time factor so that the similarity calculation considers not only the similarity of features but also the temporal proximity.

[0204] Preferably, the optimization path backtracking unit can regularly update the stored historical optimal parameters, for example, by retaining the optimization path in the most recent period through a sliding window mechanism to ensure that the system can adapt to the latest changes in working conditions. In the abnormal pattern encoding unit, the dimension of the sparse coding vector can be adjusted according to actual needs. For example, when the feature information is more complex, the dimension can be appropriately increased to improve the distinguishability of the feature. In the similarity formula of the dynamic retrieval unit, the weight of the time factor can be adjusted according to the dynamic characteristics of the system. For example, in a rapidly changing system, the time weight can be increased to pay more attention to recent historical data. In an alternative, deep learning methods (such as autoencoders) can be introduced to generate sparse representations of fault features to further improve the efficiency and accuracy of feature encoding. In addition, the dynamic retrieval unit can be combined with machine learning algorithms (such as support vector machines or nearest neighbor algorithms) to optimize the matching process of historical cases, thereby improving the decision-making accuracy of the system.

[0205] In some embodiments, the sparse code vector is generated by:

[0206] Building an overcomplete dictionary Where d = 64, K = 256;

[0207] Solve for the sparse coefficient vector α:

[0208]

[0209] Among them, x is the input feature vector; λ and μ are sparse constraint coefficients used to control the sparsity degree.

[0210] It's important to note that the generation of sparse coding vectors is a key technical approach used in the self-learning knowledge base module to efficiently represent and retrieve fault signatures. By constructing an overcomplete dictionary and solving for the sparse coefficient vector, the system can map complex fault signatures into a low-dimensional sparse representation, facilitating rapid retrieval and matching of historical cases. This approach not only effectively compresses data but also preserves key information about fault signatures, improving the system's fault diagnosis capabilities and adaptability.

[0211] Specifically, the overcomplete dictionary Φ∈R d×Kis a matrix, where d represents the dimension of the feature vector, and K is the number of dictionary atoms, which is usually much larger than d to provide richer representation capabilities. For example, when d = 64, the number of dictionary atoms K can be set to 256, which means that the dictionary can provide a variety of possible basis vectors for sparse representation. The solution process of the sparse coefficient vector α is to minimize the objective function

[0212]

[0213] Implementation, where F is the input feature vector, λ and μ are the sparse constraint coefficients used to control L1 and L respectively 2,1 The sparsity of the norm. L1 norm is used to achieve sparsity, while L 2,1 The norm further enhances sparsity while maintaining the structured nature of the features. Through this optimization process, the system is able to generate a sparse coefficient vector, thereby mapping the fault features into a sparse coding vector.

[0214] Preferably, the dimension d and the number of atoms K of the overcomplete dictionary can be adjusted according to the feature complexity and storage requirements of the system. For example, when the feature information is relatively rich, the values of d and K can be appropriately increased to improve the accuracy of the sparse representation. The sparse constraint coefficients λ and μ can be set through cross-validation or experience-based methods to balance sparsity and reconstruction error. For example, when a sparser representation is required, the values of λ and μ can be increased; and when a more accurate reconstruction is required, these coefficients can be appropriately reduced. In an alternative, an online dictionary learning algorithm can be introduced so that the dictionary can be dynamically updated according to the real-time data of the system, so as to better adapt to changes in fault characteristics. In addition, for large-scale data sets, a block sparse coding method can be used to divide the feature vector into multiple sub-blocks for sparse coding to improve computational efficiency and the flexibility of sparse representation.

[0215] In some embodiments, the gradient compensation controller module performs:

[0216] a) Calculation of vacuum pump power adjustment:

[0217]

[0218] Among them, P v is the vacuum pump power; is the power regulation threshold; To compensate for the rate of change of the signal.

[0219] b) Generation of heating rate compensation:

[0220]

[0221] Among them, τ h is the heating time constant; H is the heating rate; is the sign of the objective function gradient.

[0222] c) Actuator linkage control:

[0223] U(t)=softmax([ΔP v (t),ΔH(t)])·[P v,max ,H max ] T

[0224] Among them, P v,max is the maximum value of the vacuum pump power; H max is the maximum value of the heating rate; softmax: softmax function, used to normalize the output.

[0225] It's important to note that the gradient compensation controller module is a key component of the system, dynamically adjusting the vacuum pump power and heating rate based on real-time operating conditions. This module ensures optimal equipment operation by calculating the vacuum pump power adjustment and heating rate compensation, and integrating them with actuator linkage control. The vacuum pump power adjustment is calculated based on the rate of change of the compensation signal, while the heating rate compensation is adjusted based on the sign of the objective function's gradient. Through these operations, the system effectively optimizes equipment operating efficiency, while reducing energy consumption and improving the stability of the concentration process.

[0226] Specifically, the calculation formula for the vacuum pump power regulation is:

[0227]

[0228] In the equation, L is the loss function of the system, which is used to measure the deviation between the current operating state and the target state; P is the vacuum pump power, and ΔP thresh It is the power adjustment threshold, which is used to determine whether the power needs to be adjusted. When the rate of change of the compensation signal ΔP exceeds the threshold ΔP thresh When the system uses the ReLU function to make nonlinear adjustments to the regulation amount to ensure the smoothness of power regulation. The calculation formula of the heating rate compensation amount is:

[0229]

[0230] In the ΔT max is the maximum heating rate, τ h is the heating time constant, which is used to control the rate of change of the heating rate; is the gradient of the objective function, and its sign determines the direction of the heating rate adjustment. Ultimately, the actuator linkage control normalizes the adjustment amount of the vacuum pump power and heating rate through the softmax function to ensure the coordinated operation of each actuator.

[0231] Preferably, the power adjustment threshold ΔP thresh It can be set according to the actual operation requirements of the system and the characteristics of the equipment. For example, in the process of high-precision concentration, a lower threshold can be set to achieve more precise power regulation; while in the process of rapid concentration, the threshold can be appropriately increased to speed up the regulation. h It can be adjusted according to the thermal sensitivity of the material and the dynamic characteristics of the concentration process. For example, for heat-sensitive materials, a larger τ should be selected. h To avoid overheating; for high temperature resistant materials, τ can be reduced h To improve heating efficiency. Alternatively, a dynamic adjustment strategy based on machine learning can be introduced to automatically optimize the power adjustment threshold and heating time constant by analyzing historical operating data. Furthermore, the softmax function in actuator linkage control can be replaced with other normalization methods, such as linear normalization or neural network-based adaptive normalization, to further enhance the system's flexibility and adaptability.

[0232] The aforementioned embodiments of the present invention have the following beneficial effects: Through the multimodal sensor data acquisition module, the system can acquire real-time temperature, pressure, material concentration, and flow rate data during the vacuum concentration process, providing comprehensive data support for precise control. The dynamic parameter optimization algorithm module generates control instructions based on a nonlinear time-varying model. Combined with an adaptive learning rate adjustment unit, a dynamic weight factor generation unit, and a gradient compensation amount generation unit, it can flexibly respond to complex operating conditions and achieve dynamic parameter optimization. The multimodal perception fusion module integrates multi-source sensor data by constructing a spatiotemporal correlation matrix and generates a fused feature vector, enhancing the system's perception capabilities and anti-interference performance. The real-time error correction module uses a dual feedback mechanism to compensate for system deviations, effectively improving control accuracy and reducing control deviations caused by sensor errors or operating condition fluctuations. The gradient compensation controller module adjusts the vacuum pump power and heating rate based on real-time operating conditions, enabling dynamic control of the equipment's operating status, significantly improving concentration efficiency and reducing energy consumption. The self-learning knowledge base module stores historical optimization paths and abnormal operating condition feature vectors, providing empirical references for the system and further enhancing its intelligence and adaptability.

[0233] Furthermore, the attenuation factor and gain coefficient in the adaptive learning rate adjustment unit can be dynamically adjusted based on the accumulated error and historical optimization trajectory, allowing the learning rate to adapt to the system's operating state, further improving the adaptability and convergence speed of the optimization algorithm. The error dual verification unit and dynamic compensation strategy unit in the real-time error correction module accurately determine the effective error amount and generate a compensation signal, ensuring stable system operation under complex operating conditions. The anti-saturation limiting unit constrains the compensation amount to prevent instability caused by overcompensation. The abnormal feature suppression function in the multimodal perception fusion module effectively suppresses the impact of abnormal features, further improving the robustness of the system. The dynamic adjustment of the time lag can be adaptively adjusted based on the time derivative of sensor data, allowing the system to better capture the temporal correlation of the data. The optimization path backtracking unit and abnormal pattern encoding unit in the self-learning knowledge base module store historical optimal parameters and fault characteristics, supporting the system's dynamic retrieval and fault diagnosis. The generation of sparse coding vectors maps fault characteristics into low-dimensional sparse representations, facilitating rapid retrieval and matching of historical cases. The calculation of the vacuum pump power regulation and the generation of the heating rate compensation in the gradient compensation controller module can be dynamically adjusted based on the rate of change of the compensation signal and the sign of the objective function gradient, achieving precise control of the equipment's operating status. The actuator linkage control, which normalizes the output through a softmax function, ensures coordinated system operation under different operating conditions.

[0234] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, a server, a mobile phone, or a tablet.

[0235] The above descriptions are merely some preferred embodiments of the present invention and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A vacuum concentration real-time monitoring and intelligent optimization system, characterized in that: Includes the following modules: Multimodal sensor data acquisition module, used to obtain real-time temperature, pressure, material concentration and flow rate data during the vacuum concentration process; Dynamic parameter optimization algorithm module, which generates control instructions based on nonlinear time-varying models; Multimodal perception fusion module, which integrates multi-source sensor data through spatiotemporal correlation matrix; Real-time error correction module, using a dual feedback mechanism to compensate for system deviations; Intelligent decision engine module, which performs dynamic priority task allocation and resource scheduling; Gradient compensation controller module, which adjusts the vacuum pump power and heating rate according to real-time working conditions; The self-learning knowledge base module stores historical optimization paths and abnormal operating condition feature vectors.

2. The system according to claim 1, wherein: The dynamic parameter optimization algorithm module includes the following steps: a) Adaptive learning rate adjustment unit, calculates the current learning rate α(t) according to the formula: Among them, α0 is the initial learning rate; λ is the decay factor, which is used to control the decay speed of the learning rate over time; η is the gain coefficient, which is used to adjust the dynamic adjustment amplitude of the learning rate; Δx i (t) is the instantaneous change of the i-th state variable at time t; n is the total number of state variables; b) Dynamic weight factor generation unit, which calculates the weight matrix β(t) by the following formula: Among them, σ is the sigmoid function, which is used to map the input to the (0,1) interval; w j is the weight coefficient of the jth feature; f j (t) is the characteristic value of the jth feature at time t; g k (t) is the gradient of the kth constraint at time t; m is the total number of features; k is the index of the constraint; c) Gradient compensation amount generating unit, generates compensation amount ΔP(t) according to the formula: Where L is the loss function used to measure the error of the model; θ is the model parameter; J(t) is the real-time objective function used to optimize the control process; is the gradient of the objective function.

3. The system according to claim 2, characterized in that In the adaptive learning rate adjustment unit: The decay factor λ is updated according to the following formula: Among them, λ min is the minimum value of the attenuation factor; λ max is the maximum value of the attenuation factor; E(t) is the cumulative error, which represents the cumulative error of the system at time t; ξ is the adjustment coefficient, which is used to control the dynamic change speed of the attenuation factor; The gain coefficient η is dynamically adjusted according to the historical optimization trajectory, as shown in the following formula: Where T is the time window length, which is used to determine the range of historical data; Δx(k) is the change in the state variable at time k.

4. The system according to claim 1, wherein: The real-time error correction module includes: a) Error double verification unit, which determines the effective error amount by the following formula: Where, E(t) is the current error; is the historical error mean; σ E is the standard deviation of the historical error; is the rate of change of error; θ th is the threshold of the error change rate; b) Dynamic compensation strategy unit, generates compensation signal according to the following formula: Among them, K p is the proportional coefficient; K i is the integral coefficient; K d is the differential coefficient; τ is the integration time constant; c) The anti-saturation limiting unit performs compensation constraint according to the following formula: Among them, C max is the maximum value of compensation.

5. The system according to claim 4, characterized in that The time-varying coefficients are updated as follows: The proportionality factor is: K p (t)=μ·(1e -v·t ) Among them, μ is the adjustment parameter of the proportional coefficient; v is the time decay factor of the proportional coefficient; The integration coefficient is: Wherein, ρ is the adjustment parameter of the integral coefficient; T is the length of the integral time window; The differential coefficient is: Among them, κ is the adjustment parameter of the differential coefficient; is the rate of change of error.

6. The system according to claim 1, wherein: The multimodal perception fusion module performs the following operations: a) Constructing a spatiotemporal correlation matrix: Among them, x i 、x j is the measurement data of different sensors at time t; Δt is the time lag; N is the length of the time series; b) Generate fused feature vector: Among them, w i is the weight coefficient of principal component analysis (PCA); v j is the weight coefficient of the fast Fourier transform (FFT); n is the number of PCA features; m is the number of FFT features; c) Perform abnormal feature suppression: Among them, ⊙ is the Hadamard product; is the mean of historical features; ζ is the suppression factor, which is used to adjust the suppression strength of abnormal features; σ is the sigmoid function, which is used to map the input to the (0,1) interval; ReLU is the rectified linear unit, which is used to activate abnormal features.

7. The system according to claim 6, characterized in that The time lag Δt is dynamically adjusted according to the following formula: Among them, Δt0 is the reference time lag; c i is the differential order weight coefficient; is the i-th order time derivative of the sensor data.

8. The system according to claim 1, wherein: The self-learning knowledge base module includes: a) Optimization path backtracking unit (710), storing the historical optimal parameter set θ*(t) and the corresponding objective function value J*(t); b) an abnormal pattern encoding unit (720), which maps the fault features into a 128-dimensional sparse coding vector; c) Dynamic retrieval unit (730), matches historical cases according to the following similarity formula: Among them, Q(t) is the current eigenvector; H k is the characteristic vector of the kth historical case; t k is the time index of the k-th historical case.

9. The system according to claim 8, characterized in that The sparse coding vector is generated by: Building an overcomplete dictionary Where d = 64, K = 256; Solve for the sparse coefficient vector α: Among them, x is the input feature vector; λ and μ are sparse constraint coefficients used to control the sparsity degree.

10. The system according to claim 1, wherein: The gradient compensation controller module performs: a) Calculation of vacuum pump power adjustment: Among them, P v is the vacuum pump power; is the power regulation threshold; To compensate for the rate of change of the signal; b) Generation of heating rate compensation: Among them, τ h is the heating time constant; H is the heating rate; is the symbol of the objective function gradient; c) Actuator linkage control: U(t)=softmax([ΔP v (t),ΔH(t)])·[P v,max ,H max ] T Among them, P v,max is the maximum value of the vacuum pump power; H max is the maximum value of the heating rate; softmax: softmax function, used to normalize the output.

Images