Industrial fault real-time diagnosis and disposal system based on RAG and multi-agent cooperation

Abstract

The invention relates to the technical field of industrial fault diagnosis and disposal, and discloses an industrial fault real-time diagnosis and disposal system based on RAG and multi-agent cooperation, and the system comprises a retrieval enhancement generation module which is used for fusing multi-source knowledge, and obtaining environment, geometry, measuring points and equipment constraint parameters; the parameter calculation module is used for calculating medium sound velocity, cavity neck characteristic frequency and propagation time delay according to environment and geometric characteristics; the matrix construction module constructs a lock-in Hankel matrix and a symmetric Toeplitz weight matrix through frequency and time delay, and generates a target high-dimensional matrix; an index and threshold calculation module extracts a boundary curvature index from the high-dimensional matrix and generates a self-baseline threshold to realize fault trend quantification; the command generation module outputs a constrained actuating mechanism action command according to the index and the threshold value to realize real-time intervention; the reset module automatically executes reset when the system restores to be stable, and safe closed loop and dynamic self-repairing in the control process are ensured.

Description

Technical Field

[0001] This invention relates to the field of industrial fault diagnosis and handling technology, and more specifically, to an industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration. Background Technology

[0002] In traditional compressors and fluid transport systems, fault diagnosis relies on indicators such as the amplitude of vibration signals, pressure fluctuations, power changes, or energy ratios in specific frequency bands. These indicators can be used for fault identification under steady-state conditions. However, when the system approaches the surge limit or enters the rotating stall region, the energy of the signal is not steady-state but rather exhibits short-term amplification due to non-normal dynamics and fluid hysteresis effects. Because these changes are extremely rapid in time, the system may have already entered the irreversible oscillation region before the amplitude reaches the preset alarm threshold. Traditional diagnostic methods based on statistical energy or spectral thresholds often lag behind the actual instability process in response. Furthermore, these methods fail to fully utilize the spatiotemporal relationships between multiple measurement points, resulting in an inability to reflect the overall laws governing internal energy transfer and dynamic response, easily leading to misjudgments of stability or delayed intervention.

[0003] In recent years, Retrieval Enhanced Generation (RAG) technology and multi-agent systems have been introduced into the field of industrial fault management to achieve knowledge-driven intelligent diagnosis. RAG can provide contextual assistance to agents through multi-source knowledge retrieval such as text, process drawings, and maintenance manuals; multi-agent systems can collaboratively complete diagnosis, maintenance, control, and safety decisions. However, when dealing with complex fluid dynamic systems, each agent often relies on different types of signal descriptors, such as root mean square vibration, pressure pulsation intensity, and flow deviation, lacking a numerically unified quantity that can represent the true dynamic change trend of the system.

[0004] Because these signals have different dimensions and insufficient time synchronization, it is difficult for agents to form a consistent basis for action. There is also a semantic and practical scale gap between the textual evidence provided by RAG and real-time data, making it difficult to directly map the retrieval results into executable control commands, thus affecting the timeliness and reliability of the entire diagnosis and treatment chain.

[0005] The energy evolution of industrial compression systems under critical conditions does not follow the traditional steady-state characteristic spectrum, but rather exhibits a geometrically rapid energy expansion. This rapid energy expansion phenomenon stems from the non-regularity of the system operators: although the eigenvalues ​​remain in the stable region, the operators have a very strong short-term amplification effect on perturbations. Traditional singular value, power spectrum, or correlation matrix indices cannot reveal this spectrally stable but rapidly increasing energy behavior. When the system is in this short-term energy amplification phase, if multi-agent systems still make decisions based on traditional energy thresholds, they may miss the optimal intervention window due to information delays or disagreements. Summary of the Invention

[0006] This invention provides an industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration, which solves the technical problem of how to construct a quantitative method that can reflect the geometry of unsteady energy changes and serve as the basis for coordinated action of multiple agents.

[0007] This invention provides a real-time industrial fault diagnosis and handling system based on RAG and multi-agent collaboration, comprising: The retrieval enhancement generation module is used to obtain environmental parameters, geometric parameters, measurement point location parameters, equipment characteristic parameters, and actuator constraint parameters; The parameter calculation module is used to calculate the medium characteristic parameters based on environmental parameters, calculate the cavity neck characteristic frequency based on the medium characteristic parameters and geometric parameters, and calculate the propagation delay based on the measurement point location parameters and medium characteristic parameters. The matrix construction module is used to construct a locked Hankel matrix based on the cavity neck characteristic frequency and propagation delay, generate a symmetric Toplitz weighted matrix based on the equipment characteristic parameters, and construct the target high-dimensional matrix by combining it with the identity matrix. The exponent and threshold calculation module is used to calculate the boundary curvature exponent and baseline threshold based on the target high-dimensional matrix. The command generation module is used to generate actuator action commands based on the boundary curvature index and the baseline threshold, and to make the actuator action commands subject to actuator constraint parameters. The reset module is used to perform a reset operation when the boundary curvature index is not higher than the baseline threshold.

[0008] The beneficial effects of this invention are as follows: By dynamically acquiring multi-source knowledge (such as operating condition data, equipment characteristics, and control constraints) through the RAG module, and combining it with multi-agent collaborative decision-making, diagnosis and control can achieve rapid, accurate, and interpretable intelligent responses in complex hydrodynamic environments. This invention uses the boundary curvature index as a unified dynamic quantitative indicator, overcoming the hysteresis defects of traditional methods that rely on energy thresholds or spectral characteristics. It can predict instability trends before the energy reaches the alarm threshold, thus significantly improving the timeliness and accuracy of fault warnings. Furthermore, through a self-baseline threshold adaptive mechanism and an actuator amplitude limiting and reset strategy, stable and safe closed-loop control actions are achieved, reducing erroneous actions and system oscillations. Overall, this system represents a leap from passive alarm to active prediction and intelligent intervention, exhibiting high robustness, strong generalization, and real-time self-healing capabilities. Attached Figure Description

[0009] Figure 1 This is a block diagram of the industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration of the present invention. Detailed Implementation

[0010] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0011] like Figure 1 As shown, the industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration includes: The retrieval enhancement generation module is used to obtain environmental parameters, geometric parameters, measurement point location parameters, equipment characteristic parameters, and actuator constraint parameters; The parameter calculation module is used to calculate the medium characteristic parameters based on environmental parameters, calculate the cavity neck characteristic frequency based on the medium characteristic parameters and geometric parameters, and calculate the propagation delay based on the measurement point location parameters and medium characteristic parameters. The matrix construction module is used to construct a locked Hankel matrix based on the cavity neck characteristic frequency and propagation delay, generate a symmetric Toplitz weighted matrix based on the equipment characteristic parameters, and construct the target high-dimensional matrix by combining it with the identity matrix. The exponent and threshold calculation module is used to calculate the boundary curvature exponent and baseline threshold based on the target high-dimensional matrix. The command generation module is used to generate actuator action commands based on the boundary curvature index and the baseline threshold, and to make the actuator action commands subject to actuator constraint parameters. The reset module is used to perform a reset operation when the boundary curvature index is not higher than the baseline threshold.

[0012] In one embodiment of the present invention, calculating dielectric characteristic parameters based on environmental parameters and calculating the cavity neck characteristic frequency based on the dielectric characteristic parameters and geometric parameters includes: Environmental parameters include temperature and relative humidity. The medium characteristic parameter is the humid air velocity parameter, which is calculated from the temperature parameter and the relative humidity parameter using the humid air velocity calculation function. The geometric parameters include the neck cross-sectional area parameter, the cavity volume parameter, and the equivalent neck length parameter; The characteristic frequency of the cavity neck is calculated by combining the parameters of the sound velocity in moist air, the cross-sectional area of ​​the neck, the volume of the cavity, the equivalent neck length, and pi.

[0013] Calculate the characteristic frequency of the cavity neck ;in, This represents the speed of sound in moist air. This represents the cross-sectional area parameter of the neck. Indicates the cavity volume parameter, Indicates the equivalent neck length parameter; The input parameters for the function to calculate the speed of sound in moist air include temperature and relative humidity parameters, and the output parameter is the speed of sound in moist air. .

[0014] Temperature and relative humidity parameters reflect the air conditions of the system's environment.

[0015] The medium characteristic parameter refers to the speed of sound in moist air, which is calculated using a moist air speed of sound calculation function. The inputs to the moist air speed of sound calculation function are the temperature and relative humidity parameters mentioned above among the environmental parameters. By quantifying these two environmental parameters through the moist air speed of sound calculation function, the output can characterize the speed of sound propagating in moist air. The moist air speed of sound parameter is an intermediate quantity connecting the environmental state and the system's hydrodynamic characteristics.

[0016] The cavity neck structure is an abstract physical model of pipeline-cavity-throttling components in industrial compression systems and fluid transportation systems. It simulates the dynamic characteristics of fluid (such as humid air) flowing and vibrating between the containment space and the narrow channel. The specific correspondence and composition are as follows: The cavity-neck structure consists of two parts: the cavity and the neck. The cavity corresponds to the cavity components in the system (such as the air storage cavity or fluid buffer cavity in a compression system, whose size is described by the cavity volume parameter). The neck corresponds to the narrow channel components in the system that connect the cavity and the pipeline (such as a throttle valve or a pipeline diameter change section, whose cross-sectional area parameter describes the cross-sectional size of the channel, and the equivalent neck length parameter describes the effective length of the channel).

[0017] When fluid flows in the cavity neck structure, pressure pulsations are generated. The pulsation signals are reflected and superimposed in the cavity neck structure to form resonance, thereby affecting the instability risk of the system (such as surge, rotational stall).

[0018] Neck cross-sectional area parameter (reflects the size of the cross-section of the neck in the cavity-neck structure).

[0019] Cavity volume parameters (reflecting the size of the internal space of the cavity in the cavity neck structure).

[0020] Equivalent neck length parameter (a parameter that characterizes the effective length of the neck after considering the influence of the actual neck structure).

[0021] The combination of neck cross-sectional area parameters, cavity volume parameters, and equivalent neck length parameters characterizes the inherent physical morphological features of the cavity-neck structure.

[0022] The characteristic frequency of the cavity neck is used to quantify the sensitivity of the cavity neck structure to pressure fluctuations at a specific frequency. Specifically, when the frequency of pressure fluctuations within the system approaches the characteristic frequency of the cavity neck, the cavity neck structure will resonate, leading to a significant amplification of the pressure pulsation amplitude. This amplification effect is a precursor to system instability (surge).

[0023] In one embodiment of the present invention, calculating the propagation delay based on the measurement point location parameters and medium characteristic parameters includes: The measuring point location parameter is the distance from the measuring point to the neck reference surface; The propagation delay is calculated by dividing the distance from the measuring point to the neck reference surface by the sound speed in moist air.

[0024] Calculate propagation delay parameters ;in, This represents the distance parameter from the measuring point to the neck reference surface.

[0025] The measurement point location parameter refers to the distance parameter from the measurement point to the neck reference surface. The neck reference surface is a reference plane of the neck component in the cavity neck structure (used to uniformly measure the spatial distance between each measurement point and the core flow area of ​​the cavity neck). The measurement point location parameter quantifies the actual physical distance of each measurement point to this reference surface.

[0026] Because the distances from each measuring point to the core region of the cavity neck (neck reference plane) are different, the time it takes for the same pressure fluctuation signal to reach different measuring points varies (i.e., propagation delay). This time difference will cause the original signals at each measuring point to be out of phase.

[0027] Propagation delay is used to eliminate the spatial phase difference of signals from multiple measurement points, ensuring that the signal reflects the dynamic state of the same cavity neck structure. Specifically, multiple measurement points in the system are distributed in different locations, and the distance from each measurement point to the cavity neck structure is different, which will cause the same pressure fluctuation signal to reach the cavity neck reference surface in different times (i.e., propagation delay). If this difference is not corrected, the signals from multiple measurement points will show phase misalignment due to the time difference, and cannot accurately reflect the resonance and wave characteristics of the cavity neck structure.

[0028] In one embodiment of the present invention, constructing a direction-locked Hankel matrix based on the cavity neck characteristic frequency and propagation delay includes: The characteristic angular frequency of the cavity neck is calculated by combining the characteristic frequency of the cavity neck with pi. The phase compensation parameters are calculated from the characteristic angular frequency parameters of the cavity neck and the propagation delay. The phase pre-compensation calculation of the original signal parameters at the measurement point using phase compensation parameters yields the lock sequence. The lock-in Hankel matrix is ​​formed by stacking time-shifted lock-in sequences of length equal to the window length parameter, the number of measurement points, and the current time count.

[0029] Calculate the characteristic angular frequency parameters of the cavity neck .

[0030] Calculate phase compensation parameters .

[0031] Generate locked sequence ;in, The original signal parameters at the measurement point, Indicates the measurement point number. Represents discrete-time counting. Represents the imaginary unit. Represents the natural base.

[0032] Constructing a direction-locked Hankel matrix ;in, This represents the window length parameter. The parameter representing the number of measuring points has the following elements: , Represents the direction-locked Hankel matrix. Line 1 Column elements, This indicates the current time count.

[0033] The characteristic frequency of the cavity neck is the inherent resonant frequency of the cavity neck structure, which needs to be converted into an angular frequency that matches the phase.

[0034] The phase compensation parameter is obtained by multiplying the cavity neck characteristic angular frequency parameter by the propagation delay parameter. Essentially, it quantifies the phase deviation of the original signal at the measurement point caused by the propagation delay, that is, the product of the signal propagation time difference (propagation delay) and the angular frequency, which is exactly equal to the phase lag generated by the signal at the cavity neck characteristic frequency, i.e., the phase compensation parameter.

[0035] Phase compensation parameters are used to pre-compensate the original signal parameters at the measurement points. This means that phase correction is used to cancel the phase deviation caused by the propagation delay, so that the signals at all measurement points are aligned with the phase corresponding to the cavity neck characteristic frequency, thereby generating a lock-in sequence. Locking refers to locking the phase corresponding to the characteristic frequency of the cavity neck. The generated locking sequence can eliminate phase misalignment between multiple measurement points, allowing all measurement point signals to uniformly reflect the dynamic response of the cavity neck structure at its inherent resonant frequency, and avoiding phase confusion from masking the true fluctuation trend of the system.

[0036] Window length parameter: refers to the length of the selected short-time signal segment (i.e., taking the current time and several consecutive signal points before it).

[0037] Quantity of measuring points: refers to the total number of all measuring points in the system.

[0038] Current time count: refers to the time starting point for determining the signal segment (ensuring that the selected signal is current and recent, reflecting the latest dynamics of the system).

[0039] The purpose of the direction-locked Hankel matrix is ​​to integrate multi-dimensional dynamic information. It includes both spatial dimensions (the direction-locked sequence of all measurement points, reflecting the spatial distribution of the signal) and temporal dimensions (time-shifted signals within the window length, reflecting the short-term dynamic changes of the signal). The shift-stacking structure of the Hankel matrix is ​​good at encoding the short-term memory characteristics of the signal and can completely preserve the recent dynamic trends of the system.

[0040] In one embodiment of the present invention, generating a symmetric Toplitz weight matrix based on device characteristic parameters includes: Equipment characteristic parameters include damping ratio parameters; The kernel function parameters are generated from the damping ratio parameter, the cavity neck characteristic angular frequency parameter, the sampling step size parameter, and the time displacement count, where the time displacement count only takes integer values ​​from zero to one less than the window length parameter; The elements of the symmetric Toplitz weight matrix are filled by the kernel function parameters based on the difference between the absolute values ​​of the row counts and column counts, where the row counts and column counts are only integers from one to the window length parameter.

[0041] Calculate kernel function parameters: ;in, This represents the damping ratio parameter. Indicates time displacement count, This represents the sampling step size parameter.

[0042] Construct a symmetric Topletz weight matrix ;in, Describes the symmetric Topletz weight matrix. Line 1 Column elements, Indicates row count and only takes values ​​from 1 to 1. integers, This indicates a column count and only takes values ​​from 1 to... integers, Represents the non-negative integer difference between the row count and the column count.

[0043] The damping ratio parameter is used to quantify the energy dissipation capacity of the cavity neck structure: the larger the damping ratio, the faster the energy decays during vibration or pressure fluctuations, and vice versa. It is used to reflect the dynamic characteristics of the cavity neck structure.

[0044] The sampling step size parameter is used to discretize the continuous-time signal into a digital signal at a time interval, ensuring that the kernel function calculation is adapted to the signal acquisition method of the actual system.

[0045] Time shift count: Used to represent the time difference between the current moment and a historical moment, its value range is limited to an integer from zero to the window length parameter minus one. The window length parameter is the length of the short-time signal segment selected when constructing the matrix later, and the range of the time shift count corresponds to all possible time differences within this segment.

[0046] Row / column count: Used to locate the element positions of the symmetric Toplitz weight matrix. The value range is limited to integers from one to the window length parameter, which corresponds to the range of time displacement count (the absolute difference between the row / column counts exactly covers the range of time displacement count).

[0047] The kernel function parameters are obtained through the collaborative operation of damping ratio parameters, cavity neck characteristic angular frequency parameters, sampling step size parameters, and time displacement counts. Specifically: First, the damping attenuation factor is calculated by multiplying the damping ratio parameter by the time displacement count and the sampling step size parameter. Then, the energy attenuation caused by damping is reflected by exponential calculation. Meanwhile, the vibration phase factor is calculated by multiplying the characteristic angular frequency parameter of the cavity neck by the time displacement count and the sampling step size parameter, and then the inherent resonance characteristics of the cavity neck structure are reflected by cosine operation. The final kernel function parameters are the product of the two parts mentioned above, in order to simulate the dynamic response of the cavity neck structure under the combined influence of damping and inherent resonance.

[0048] The kernel function defines the signal weights at different time shifts. Specifically: Due to the existence of the damping ratio, the kernel function parameters will decrease as the time displacement count increases (reflecting energy decay). That is, the smaller the time difference between the historical signal and the current moment, the greater the weight; the larger the time difference, the smaller the weight due to damping decay. Meanwhile, the cosine term is bound to the characteristic angular frequency of the cavity neck, ensuring that the kernel function parameters only assign effective weights to signals near the cavity neck resonant frequency, and give lower weights to interference signals at non-resonant frequencies.

[0049] The element filling of a symmetric Toplitz weighted matrix follows the rule that the absolute difference between the row counts and column counts matches the time displacement counts: For the element in the p-th row and q-th column of the matrix (p is the row count and q is the column count), first calculate the absolute difference between p and q. This difference is exactly equal to the time displacement count (since p and q are both integers from 1 to the window length, the range of the absolute difference is from 0 to the window length minus one, which is exactly the same as the range of the time displacement count). Then, the kernel function parameter corresponding to the absolute difference is used as the element value of the p-th row and q-th column, and finally a symmetric matrix is ​​formed (since the absolute difference between p and q is equal to the absolute difference between q and p, the matrix satisfies symmetry).

[0050] The symmetric Topletz weight matrix is ​​the energy-weighted carrier of the dynamic characteristics of the cavity neck structure, specifically: As a weighting matrix, its elements (kernel function parameters) encode the damping attenuation and resonance characteristics of the cavity neck structure. When combined with the locked Hankel matrix (including short-time locked signals), it can physically adapt and weight the short-time signals in the matrix. This highlights the effective signals at the resonance frequency and in the near time, while suppressing attenuated signals and interference at non-resonance and in the far time. The positional weights of displacement differences at the same time (with the same absolute value difference in row / column counts) are consistent, which conforms to the dynamic law of the invariance of time shift in the cavity neck structure (the energy attenuation and resonance response are consistent under the same time difference), and avoids the weighting bias caused by the matrix structure.

[0051] In one embodiment of the present invention, constructing the target high-dimensional matrix includes: The symmetric Topletz weight matrix is ​​decomposed into an orthogonal eigenvector matrix, an eigenvalue diagonal matrix, and the transpose of the orthogonal eigenvector matrix; The square root of a matrix is ​​obtained from the orthogonal eigenvector matrix, the square root of the eigenvalue diagonal matrix, and the transpose of the orthogonal eigenvector matrix. The square root of the eigenvalue diagonal matrix is ​​formed by taking the square root of each eigenvalue in the eigenvalue diagonal matrix. The dimension of the identity matrix corresponds to the number of measurement points. The dimension of the square root of the matrix corresponds to the window length parameter; The target high-dimensional matrix is ​​formed by performing the Kronecker product operation on the identity matrix and the square root of the matrix, followed by matrix multiplication with the locked Hankel matrix.

[0052] For symmetric Topletz weight matrix .

[0053] Calculate the square root of the matrix based on the eigenvalue decomposition result. .

[0054] in, Describes the symmetric Toplitz weight matrix. Represents an orthogonal eigenvector matrix. Represents an eigenvalue diagonal matrix. This represents the eigenvalue diagonal matrix. The diagonal matrix formed by taking the square root of each eigenvalue. Let denote the square root of the symmetric Topletz weight matrix, and The dimension is .

[0055] Construct the target high-dimensional matrix ; in, Represents the target high-dimensional matrix. The dimension is The identity matrix, This represents the Kronecker product operation. Denotes the square root of a symmetric Topletz weight matrix. This represents a locked Hankel matrix.

[0056] The eigenvalues ​​of the symmetric Toplitz weight matrix are decomposed to obtain an orthogonal eigenvector matrix. The column vectors of the orthogonal eigenvector matrix are the eigenvectors of the weight matrix, and the vectors are mutually orthogonal (uncorrelated), ensuring that the decomposition process does not introduce additional signal interference, and the dimension is consistent with that of the symmetric Toplitz weight matrix.

[0057] The eigenvalue diagonal matrix is ​​obtained by eigenvalue decomposition of the symmetric Topulitz weight matrix. The elements on the diagonal of the eigenvalue diagonal matrix are the eigenvalues ​​of the weight matrix (quantifying the importance of the weights corresponding to each eigenvector), and the off-diagonal elements are 0. The dimension is the same as that of the symmetric Topulitz weight matrix.

[0058] The square root of the eigenvalue diagonal matrix is ​​obtained by taking the square root of each diagonal element (eigenvalue) in the eigenvalue diagonal matrix. The purpose of the square root of the eigenvalue diagonal matrix is ​​to enable subsequent weighted adaptation energy inner product operations, and to avoid signal weighting distortion caused by the original eigenvalues ​​being too large or too small. The dimension of the eigenvalue diagonal matrix is ​​consistent with that of the eigenvalue diagonal matrix.

[0059] The square root of the matrix is ​​obtained by multiplying the orthogonal eigenvector matrix, the diagonal matrix of eigenvalues ​​after taking the square root, and the transpose of the orthogonal eigenvector matrix. The square root of the matrix is ​​the energy-adaptive weight carrier of the symmetric Topletz weight matrix, and the dimension is still window length × window length.

[0060] The dimension of the identity matrix is ​​consistent with the number of measurement points (set as number of measurement points × number of measurement points). The role of the identity matrix is ​​to expand the spatial dimension of the weights, ensuring that the short-time signals of multiple measurement points can be weighted by a unified cavity dynamic weight without changing the relative relationship between the signals of the measurement points.

[0061] Based on the phase-compensated locked sequence, a locked Hankel matrix is ​​constructed to encode a multi-measurement point short-time aligned signal (spatial dimension: number of measurement points; time dimension: window length). The dimension is window length × number of measurement points, which is the dynamic signal carrier of the target high-dimensional matrix.

[0062] Eigendecomposition of the symmetric Topplitz weight matrix: Utilizing the diagonalizability of symmetric matrices, the symmetric Topplitz weight matrix is ​​decomposed into an orthogonal eigenvector matrix × a diagonal matrix of eigenvalues ​​× the transpose of the orthogonal eigenvector matrix. By decomposing the dynamic weights of the weight matrix into eigenvectors (weight directions) and eigenvalues ​​(weight magnitudes), the damping and resonance characteristics of the cavity neck structure are preserved.

[0063] Calculating the square root of a matrix: Multiply the orthogonal eigenvector matrix, the diagonal matrix of the squared eigenvalues, and the transpose of the orthogonal eigenvector matrix sequentially to obtain the square root of the symmetric Topletz weight matrix. The square root is the updated matrix of the weight matrix. When the original weight matrix is ​​used for signal weighting, the signal energy may be distorted due to differences in the magnitude of the eigenvalues. The eigenvalues ​​of the square root are the square roots of the original eigenvalues, making the weighting operation equivalent to an energy inner product based on the cavity neck dynamics, more accurately highlighting the effective signal at the resonant frequency and suppressing interference signals.

[0064] The Kronecker product of the identity matrix and the square root of a matrix: The Kronecker product is performed on the identity matrix (dimension: number of measurement points × number of measurement points) and the square root of a matrix (dimension: window length × window length), resulting in a matrix of (window length × number of measurement points) × (window length × number of measurement points). The purpose of the Kronecker product is to extend the weights of the time dimension to the spatial dimension. That is, the square root of the matrix only encodes the cavity dynamics weights in the time dimension (window length). By performing the Kronecker product with the identity matrix (spatial dimension, number of measurement points), the short-time signal of each measurement point can be weighted with the same time dimension weights. This ensures that the multi-measurement signals retain their individual dynamic characteristics while adhering to a unified cavity dynamics law during the weighting process, avoiding analytical biases caused by inconsistent weights between measurement points.

[0065] The logic for calculating the target high-dimensional matrix is ​​as follows: The Kronecker product result is multiplied by a locked-Hankel matrix (with dimensions of window length × number of measurement points) to obtain the target high-dimensional matrix (with dimensions of (window length × number of measurement points) × number of measurement points). This deeply integrates physical weights with dynamic signals. The Kronecker product provides spatially-temporally unified cavity neck dynamic weights, while the locked-Hankel matrix provides short-time phase-aligned signals from multiple measurement points. After multiplication, the target high-dimensional matrix encodes both the damping and resonance characteristics (weights) of the cavity neck structure and contains the system's latest multi-measurement point dynamic response (signals).

[0066] In one embodiment of the present invention, calculating the boundary curvature exponent and the baseline threshold based on the target high-dimensional matrix includes: The Hermitian part matrix is ​​obtained by adding the target high-dimensional matrix and its conjugate transpose and then dividing by two. The oblique Hermitian part matrix is ​​obtained by subtracting the target high-dimensional matrix from its conjugate transpose and then dividing by twice the imaginary unit. The support function is the maximum eigenvalue of the sum of the product of the cosine of the angular parameter and the Hermitian part matrix, and the product of the sine of the angular parameter and the oblique Hermitian part matrix, where the maximum eigenvalue is obtained through the maximum eigenvalue function; The principal eigenvalue is the support function value when the angular parameter is zero, the principal eigenvector is the eigenvector corresponding to the principal eigenvalue of the Hermitian part matrix, and the 2 norm of the principal eigenvector is one. The remaining eigenvalues ​​and eigenvectors are the eigenvalues ​​and eigenvectors in the Hermitian partial matrix other than the principal eigenvalues ​​and principal eigenvectors; The first-order change is obtained by multiplying the conjugate transpose of the principal eigenvector, the oblique Hermitian part matrix, and the principal eigenvector. The second-order change is the negative value of the product of the conjugate transpose matrix, the Hermitian matrix, and the principal eigenvector, plus twice the square of the absolute value of the product of the conjugate transpose matrix, the oblique Hermitian matrix, and the principal eigenvector, divided by the difference between the principal eigenvalue and the corresponding other eigenvalues. The radius of curvature is the sum of the principal eigenvalue and the second-order variation. The spectral norm parameter is the second norm of the target high-dimensional matrix; The boundary curvature exponent is a product of the radius of curvature and the spectral norm parameter. The integer parameter is obtained by dividing the sliding time window length parameter by the sampling step size parameter and then taking the integer part. The baseline threshold is the 95th percentile value of the historical sequence of boundary curvature indices, where the historical sequence of boundary curvature indices is composed of the boundary curvature indices corresponding to the current time count parameter, and the boundary curvature indices obtained by taking integer multiples of the sampling step size parameter and the total number of parameters.

[0067] Calculate the oblique Hermitian partial matrix Calculate the Hermitian partial matrix ;in, Representing the target high-dimensional matrix The conjugate transpose of .

[0068] Define support function Take angle parameters , obtain the principal eigenvalue Main eigenvectors satisfy and ;in, Represents angular parameters, Represents the function with the largest eigenvalue. Represents the principal eigenvalues ​​of the Hermitian partial matrix. This represents the principal eigenvector corresponding to the principal eigenvalue of the Hermitian partial matrix.

[0069] Define the calculation of first-order change ;in, express The conjugate transpose of .

[0070] Define a second-order change that satisfies ;in, Indicates the index of the non-principal eigenvalue of the Hermitian partial matrix. The remaining eigenvalues ​​of the Hermitian partial matrix are represented. This represents the eigenvectors corresponding to the remaining eigenvalues ​​of the Hermitian partial matrix. express The conjugate transpose of .

[0071] Calculate the radius of curvature ;in, Main eigenvalues .

[0072] Define the spectral norm parameter to satisfy .

[0073] Calculate the boundary curvature index .

[0074] Calculating the baseline threshold includes: First define the integer quantity parameter to satisfy ; Calculate the baseline threshold: ; in, This represents the length parameter of the sliding time window. Indicates an integer parameter. Indicates the time counting parameter. This represents the 95th percentile function. This represents the historical sequence index.

[0075] The Hermitian partial matrix is ​​the real part energy information carrier extracted from the target high-dimensional matrix, reflecting the stable energy components of the system dynamics.

[0076] The oblique Hermitian part matrix is ​​the imaginary phase information carrier extracted from the target high-dimensional matrix, reflecting the transient fluctuation components of the system dynamics and corresponding to the phase changes of non-regular characteristics.

[0077] The support function describes the energy distribution limit of the target high-dimensional matrix in different phase directions (angular parameters). Its value is the maximum eigenvalue in a specific direction, which is used to locate the direction in which energy amplification is most significant.

[0078] The principal eigenvalue / principal eigenvector is the largest eigenvalue (principal eigenvalue) and corresponding vector (principal eigenvector) of the support function in the cavity neck phase-lock direction (angular parameter is zero). The principal eigenvector has a 2-norm of 1 (normalization process to avoid magnitude interference) and is a parameter for focusing energy in the cavity neck resonance direction.

[0079] First-order / second-order changes are used to quantify the first-order rate of change and second-order trend of the principal eigenvalue in the phase-locked direction, respectively, reflecting the speed and acceleration of the energy amplification trend and capturing the short-term energy surge characteristics of non-canonical dynamics.

[0080] The radius of curvature is a fundamental parameter that transforms the second-order change of energy into geometric curvature. The smaller the radius of curvature, the steeper (bulging) the energy distribution in the phase-locked direction, and the closer the system is to instability.

[0081] The spectral norm parameter is the L2 norm of the target high-dimensional matrix, which quantifies the overall signal energy scale of the matrix and is used to normalize the boundary curvature exponent, avoiding index distortion caused by differences in absolute energy values.

[0082] The boundary curvature index is a quantitative indicator of failure trend that combines geometric curvature and overall energy. The larger the index, the stronger the near-instability trend of the system.

[0083] The integer parameter is used to determine the size of the time window for the historical data range. It is obtained by rounding down the ratio of the sliding time window length to the sampling step size, ensuring that the baseline reflects the recent stable state.

[0084] Since the baseline threshold is the 95th percentile value of the historical sequence of the boundary curvature index, the early warning benchmark is adaptively set to exclude interference from extreme fluctuations.

[0085] Hermitian partial matrix extracts the real part energy (stable component) of the matrix, while oblique Hermitian partial matrix extracts the imaginary part phase (transient fluctuation component). Together, they decompose the energy and phase information of the target high-dimensional matrix.

[0086] The angular parameter is zero, locking the inherent resonance direction of the cavity neck. The principal eigenvalue is the maximum energy component in that direction, and the principal eigenvector is the signal mode with concentrated energy. Both focus on the core direction where energy is most easily amplified when the system is near instability, avoiding interference from irrelevant directions.

[0087] The first-order change reflects the initial velocity of the energy amplification trend, while the second-order change reflects the acceleration of energy amplification. That is, when the system approaches instability, the second-order change will increase significantly, reflecting the key characteristic of non-canonical dynamic energy amplification but still stable spectrum, and is the core sensitive item for early warning.

[0088] The radius of curvature converts energy acceleration into geometric curvature (the smaller the radius, the steeper the energy distribution); the spectral norm parameter normalizes the overall energy to avoid the influence of differences in the absolute value of energy under different operating conditions on the judgment; the boundary curvature index combines the two, the larger the index, the more significant the short-term energy amplification in the cavity neck phase-locking direction, the stronger the near-instability trend of the system, which is the core quantitative basis for fault early warning.

[0089] The integer parameter is used to calculate the historical data scale of the baseline. For example, if the sliding time window length is 10 seconds and the sampling step size is 0.1 seconds, then the integer parameter is 100, which means that the boundary curvature index of the last 100 sampling points is taken as the historical sequence to ensure that the baseline reflects the recent stable state rather than long-term outdated data.

[0090] The 95th percentile value means that under recent stable conditions, 95% of the index values ​​are below this threshold, excluding only 5% of extreme fluctuations. This avoids misjudgments caused by changes in operating conditions (such as ambient temperature and load adjustment) due to fixed thresholds (missed judgments due to excessively high thresholds, and misjudgments due to excessively low thresholds), while ensuring that the baseline can adaptively track the stable state of the system and improve the accuracy of early warning.

[0091] In one embodiment of the present invention, generating actuator action commands based on the boundary curvature index and a baseline threshold, and subjecting the actuator action commands to actuator constraint parameters, includes: The actuator constraint parameters include the lower limit of actuator action commands, the upper limit of actuator action commands, and actuator rate constraints; The original increment is obtained by multiplying the preset command gain parameter by the difference between the boundary curvature exponent and the baseline threshold, and then processing it with a non-negative truncation function. The rate constraint increment is the result of the original increment after being restricted. The restriction is that the original increment does not exceed the product of the actuator rate constraint and the sampling step size parameter, and is not lower than the negative value of the product. The first intermediate command is obtained by adding the rate constraint increment to the actuator action command at the previous sampling time; The final action command of the actuator is obtained by the first intermediate command sandwiched between the lower limit of the actuator action command and the upper limit of the actuator action command, that is, the first intermediate command does not exceed the upper limit of the actuator action command and is not lower than the lower limit of the actuator action command.

[0092] Calculate the original increment ;in, The non-negative truncation function satisfies: , For the input variables of the non-negative truncation function, This indicates the command gain parameter.

[0093] Incremental computation rate constraint: ;in, This represents the actuator rate constraint in the actuator constraint parameters.

[0094] Calculate the first intermediate command ;in, This indicates the action command executed at the previous sampling time.

[0095] Calculate the final actuator action command: ;in, Indicates the final action command from the executing agency. This indicates the upper limit of actuator action commands in the actuator constraint parameters. This indicates the lower limit of the actuator action command in the actuator constraint parameters.

[0096] Actuator constraint parameters: key limiting conditions that ensure the safe and smooth operation of the actuator, including three categories: Minimum action command limit of the actuator: The minimum action command that the actuator can output (hardware safety minimum, to prevent equipment damage due to insufficient action); Upper limit of actuator action commands: The maximum action command that the actuator can output (hardware safety limit to prevent excessive action from causing system oscillation). Actuator rate constraint: The maximum allowable change in action of the actuator per unit time (to prevent excessively rapid action from causing imbalance in the system's dynamic response).

[0097] Preset command gain parameter: A coefficient for adjusting the intervention intensity (preset according to equipment characteristics; the greater the gain, the greater the action increment under the same risk, ensuring that the intervention response matches the equipment sensitivity). Sampling step size parameter: The time interval between two calculations of the actuator command (consistent with the system signal acquisition frequency to ensure that the rate constraint is adapted to the actual action time). The action command executed at the previous sampling time: the action instruction output at the last time (serving as the reference for the current command to avoid sudden changes in action); Non-negative truncation function: Only retains the difference when the boundary curvature exponent is higher than the baseline threshold (i.e., only generates action increments when there is a risk of near instability in the system, and does not intervene when there is no risk).

[0098] The initial increment is the theoretical intervention amount based on the level of risk. The increment is only generated when the system has a near-instability risk (exponential overshoot of the baseline). The gain parameter controls the matching relationship between risk and intervention amount (e.g., a larger gain is required for high-risk scenarios to ensure rapid suppression of risk). When there is no risk, the increment is zero to avoid ineffective actions.

[0099] The rate constraint increment is the actual intervention amount that conforms to the actuator's operating speed. Actuators have maximum operating speed limits (such as the maximum opening change of a valve per second). If the initial increment is too large, causing the operating speed to exceed the limit, it can lead to damage to the mechanism or system fluctuations.

[0100] The first intermediate command is a preliminary intervention instruction based on historical actions. It uses the previous command as a benchmark and adds compliant increments to avoid action jumps (such as jumping directly from 0 to the maximum action), ensures smooth dynamic response of the system, and reduces the impact of intervention on system stability.

[0101] The final command is an executable instruction that meets hardware safety requirements. The actuator has hardware action range limitations (the lower limit prevents the valve from closing completely and causing pressure buildup, and the upper limit prevents the valve from opening completely and causing flow control failure).

[0102] In one embodiment of the present invention, performing a reset operation when the boundary curvature index is not higher than a baseline threshold includes: The trigger condition for performing a reset operation is that the boundary curvature index is not higher than the baseline threshold. Determine the lift start time count parameter and the lift start command parameter, wherein the lift start time count parameter is the time count parameter for the first time that the boundary curvature index is higher than the baseline threshold, and the lift start command parameter is the actuator action command corresponding to the lift start time count parameter; Determine the reset gain parameter and the reset termination threshold parameter; Calculate the target difference between the parameters of the actuator action command and the start-up command at the previous sampling time; The original reset increment is obtained by multiplying the reset gain parameter by the negative value of the target difference; The reset increment after rate constraint is the result of the original reset increment after being restricted. The restriction is that the original reset increment does not exceed the product of the actuator rate constraint and the sampling step size parameter, and is not lower than the negative value of the product. The second intermediate command is obtained by adding the reset increment after the rate constraint to the actuator action command at the previous sampling time; The final reset output command is obtained by sandwiching the second intermediate command between the lower limit of the actuator action command and the upper limit of the actuator action command, that is, the second intermediate command does not exceed the upper limit of the actuator action command and is not lower than the lower limit of the actuator action command; The reset termination condition is that the absolute value of the difference between the final reset output command and the start-up command parameter is not greater than the reset termination threshold parameter. When the reset termination condition is met, the reset operation stops.

[0103] Define the promotion start time count parameter For the first time to meet Time counting parameters; Define the start command parameters for promotion The time count parameter is equal to The execution mechanism's action command at that time.

[0104] Calculate the original reset increment ;in, Reset gain parameters; Calculate the rate-constrained reset increment: ; Calculate the second intermediate command ; Calculate the final reset command: ; When satisfied When this happens, the reset operation is stopped; among them, This represents the reset termination threshold parameter.

[0105] Reset trigger condition: The boundary curvature index is not higher than the baseline threshold (indicating that the risk of near instability of the system has been eliminated and it needs to be restored from the intervention state to the normal state). Increase the start time count parameter: the time count at which the boundary curvature index first exceeds the baseline threshold (i.e., the time reference for the start of risk, used to locate the time node of the normal state). Increase the start command parameter: Increase the actuator action command at the time corresponding to the start time count parameter (i.e., the normal action command before the risk occurs, which is the target benchmark for reset and ensures reset to a stable initial state). Reset gain parameter: A coefficient that adjusts the reset speed (preset value; the higher the gain, the larger the reset increment; it needs to be matched with the sensitivity of the actuator to avoid reset being too slow or too fast). Reset termination threshold parameter: The accuracy standard for determining whether the reset is complete (if the absolute value of the difference is less than this value, it means that the target benchmark has been approached, and the reset should be stopped to avoid repeated adjustments).

[0106] Actuator rate constraint: the maximum allowable change in motion per unit time (to prevent system fluctuations caused by excessively rapid reset actions); Lower / upper limit of actuator action commands: the minimum / maximum action commands allowed by the actuator hardware (to prevent reset from exceeding the hardware safety range); When the boundary curvature exponent is not higher than the baseline threshold, it indicates that the risk of near instability of the system has been eliminated (the exponent returns to the stable range). There is no need to continue to execute the intervention command. The system should be restored to the normal state before the risk occurred (triggering the reset operation) to avoid the intervention command remaining for a long time and causing the system to deviate from the steady state.

[0107] The reset target is the normal command before the risk begins, ensuring that the system returns to a stable initial state after the reset, and avoiding new stability problems caused by excessive or insufficient reset.

[0108] The original reset increment is the theoretical reset adjustment amount, and its direction is always towards the normal command before the risk, ensuring that the reset direction is correct, and the gain avoids the mismatch between the reset speed and the actuator.

[0109] Constraining the reset speed is used to prevent excessively fast reset actions (such as rapid valve closure) from causing sudden changes in system pressure / flow, ensuring a smooth reset process and avoiding new dynamic fluctuations.

[0110] The second intermediate command is based on the previous command and superimposed with a compliant increment to avoid abrupt changes in the reset command (such as jumping directly from a high-intervention command to the target command), thus achieving a smooth transition in the reset process and reducing the impact on the system's steady state.

[0111] First, the second intermediate command is compared with the lower limit of the actuator's action command, and the larger value is taken (to avoid falling below the hardware lower limit); then, the result is compared with the upper limit of the actuator's action command, and the smaller value is taken (to avoid exceeding the hardware upper limit), thus obtaining the final reset output command. In this way, by clamping the upper and lower limits of the hardware, it is ensured that the final reset command is within the safe operating range of the actuator, preventing damage to the mechanism due to reset deviation (such as valve overtravel).

[0112] The reset operation stops when the absolute value of the difference between the final reset output command and the start-up command parameter is less than or equal to the reset termination threshold parameter. A preset accuracy standard is used to determine whether the reset is complete, avoiding continuous adjustments due to minor deviations (such as commands approaching the target but fluctuating repeatedly), ensuring that the system stabilizes in its normal state before any risks arise after the reset.

[0113] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A real-time industrial fault diagnosis and handling system based on RAG and multi-agent collaboration, characterized in that, include: The retrieval enhancement generation module is used to obtain environmental parameters, geometric parameters, measurement point location parameters, equipment characteristic parameters, and actuator constraint parameters; The parameter calculation module is used to calculate the medium characteristic parameters based on environmental parameters, calculate the cavity neck characteristic frequency based on the medium characteristic parameters and geometric parameters, and calculate the propagation delay based on the measurement point location parameters and medium characteristic parameters. The matrix construction module is used to construct a locked Hankel matrix based on the cavity neck characteristic frequency and propagation delay, generate a symmetric Toplitz weighted matrix based on the equipment characteristic parameters, and construct the target high-dimensional matrix by combining it with the identity matrix. The exponent and threshold calculation module is used to calculate the boundary curvature exponent and baseline threshold based on the target high-dimensional matrix. The command generation module is used to generate actuator action commands based on the boundary curvature index and the baseline threshold, and to make the actuator action commands subject to actuator constraint parameters. The reset module is used to perform a reset operation when the boundary curvature index is not higher than the baseline threshold.

2. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 1, characterized in that, The dielectric characteristic parameters are calculated based on environmental parameters, and the cavity neck characteristic frequency is calculated based on the dielectric characteristic parameters and geometric parameters, including: Environmental parameters include temperature and relative humidity. The medium characteristic parameter is the humid air velocity parameter, which is calculated from the temperature parameter and the relative humidity parameter using the humid air velocity calculation function. The geometric parameters include the neck cross-sectional area parameter, the cavity volume parameter, and the equivalent neck length parameter; The characteristic frequency of the cavity neck is calculated by combining the parameters of the sound velocity in moist air, the cross-sectional area of ​​the neck, the volume of the cavity, the equivalent neck length, and pi.

3. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 2, characterized in that, The propagation time delay is calculated based on the measurement point location parameters and medium characteristic parameters, including: The measuring point location parameter is the distance from the measuring point to the neck reference surface; The propagation delay is calculated by dividing the distance from the measuring point to the neck reference surface by the sound speed in moist air.

4. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 3, characterized in that, Construct the direction-locked Hankel matrix based on the cavity neck characteristic frequency and propagation delay, including: The characteristic angular frequency of the cavity neck is calculated by combining the characteristic frequency of the cavity neck with pi. The phase compensation parameters are calculated from the characteristic angular frequency parameters of the cavity neck and the propagation delay. The phase pre-compensation calculation of the original signal parameters at the measurement point using phase compensation parameters yields the lock sequence. The lock-in Hankel matrix is ​​formed by stacking time-shifted lock-in sequences of length equal to the window length parameter, the number of measurement points, and the current time count.

5. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 4, characterized in that, Generate a symmetric Toplitz weight matrix based on the equipment characteristic parameters, including: Equipment characteristic parameters include damping ratio parameters; The kernel function parameters are generated from the damping ratio parameter, the cavity neck characteristic angular frequency parameter, the sampling step size parameter, and the time displacement count, where the time displacement count only takes integer values ​​from zero to one less than the window length parameter; The elements of the symmetric Toplitz weight matrix are filled by the kernel function parameters based on the difference between the absolute values ​​of the row counts and column counts, where the row counts and column counts are only integers from one to the window length parameter.

6. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 5, characterized in that, Constructing the target high-dimensional matrix includes: The symmetric Topletz weight matrix is ​​decomposed into an orthogonal eigenvector matrix, an eigenvalue diagonal matrix, and the transpose of the orthogonal eigenvector matrix; The square root of a matrix is ​​obtained from the orthogonal eigenvector matrix, the square root of the eigenvalue diagonal matrix, and the transpose of the orthogonal eigenvector matrix. The square root of the eigenvalue diagonal matrix is ​​formed by taking the square root of each eigenvalue in the eigenvalue diagonal matrix. The dimension of the identity matrix corresponds to the number of measurement points. The dimension of the square root of the matrix corresponds to the window length parameter; The target high-dimensional matrix is ​​formed by performing the Kronecker product operation on the identity matrix and the square root of the matrix, followed by matrix multiplication with the locked Hankel matrix.

7. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 6, characterized in that, Calculate the boundary curvature exponent and baseline threshold based on the target high-dimensional matrix, including: The Hermitian part matrix is ​​obtained by adding the target high-dimensional matrix and its conjugate transpose and then dividing by two. The oblique Hermitian part matrix is ​​obtained by subtracting the target high-dimensional matrix from its conjugate transpose and then dividing by twice the imaginary unit. The support function is the maximum eigenvalue of the sum of the product of the cosine of the angular parameter and the Hermitian part matrix, and the product of the sine of the angular parameter and the oblique Hermitian part matrix, where the maximum eigenvalue is obtained through the maximum eigenvalue function; The principal eigenvalue is the support function value when the angular parameter is zero, the principal eigenvector is the eigenvector corresponding to the principal eigenvalue of the Hermitian part matrix, and the 2 norm of the principal eigenvector is one. The remaining eigenvalues ​​and eigenvectors are the eigenvalues ​​and eigenvectors in the Hermitian partial matrix other than the principal eigenvalues ​​and principal eigenvectors; The first-order change is obtained by multiplying the conjugate transpose of the principal eigenvector, the oblique Hermitian part matrix, and the principal eigenvector. The second-order change is the negative value of the product of the conjugate transpose matrix, the Hermitian matrix, and the principal eigenvector, plus twice the square of the absolute value of the product of the conjugate transpose matrix, the oblique Hermitian matrix, and the principal eigenvector, divided by the difference between the principal eigenvalue and the corresponding other eigenvalues. The radius of curvature is the sum of the principal eigenvalue and the second-order variation. The spectral norm parameter is the second norm of the target high-dimensional matrix; The boundary curvature exponent is a product of the radius of curvature and the spectral norm parameter. The integer parameter is obtained by dividing the sliding time window length parameter by the sampling step size parameter and then taking the integer part. The baseline threshold is the 95th percentile value of the historical sequence of boundary curvature indices, where the historical sequence of boundary curvature indices is composed of the boundary curvature indices corresponding to the current time count parameter, and the boundary curvature indices obtained by taking integer multiples of the sampling step size parameter and the total number of parameters.

8. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 7, characterized in that, The actuator motion commands are generated based on the boundary curvature index and the baseline threshold, and the actuator motion commands are subject to actuator constraint parameters, including: The actuator constraint parameters include the lower limit of actuator action commands, the upper limit of actuator action commands, and actuator rate constraints; The original increment is obtained by multiplying the preset command gain parameter by the difference between the boundary curvature exponent and the baseline threshold, and then processing it with a non-negative truncation function. The rate constraint increment is the result of the original increment after being restricted. The restriction is that the original increment does not exceed the product of the actuator rate constraint and the sampling step size parameter, and is not lower than the negative value of the product. The first intermediate command is obtained by adding the rate constraint increment to the actuator action command at the previous sampling time; The final action command of the actuator is obtained by the first intermediate command sandwiched between the lower limit of the actuator action command and the upper limit of the actuator action command, that is, the first intermediate command does not exceed the upper limit of the actuator action command and is not lower than the lower limit of the actuator action command.

9. The industrial fault real-time diagnosis and handling system based on RAG and multi-agent collaboration as described in claim 8, characterized in that, A reset operation is performed when the boundary curvature exponent is not higher than the baseline threshold, including: The trigger condition for performing a reset operation is that the boundary curvature index is not higher than the baseline threshold. Determine the lift start time count parameter and the lift start command parameter, wherein the lift start time count parameter is the time count parameter for the first time that the boundary curvature index is higher than the baseline threshold, and the lift start command parameter is the actuator action command corresponding to the lift start time count parameter; Determine the reset gain parameter and the reset termination threshold parameter; Calculate the target difference between the parameters of the actuator action command and the start-up command at the previous sampling time; The original reset increment is obtained by multiplying the reset gain parameter by the negative value of the target difference; The reset increment after rate constraint is the result of the original reset increment after being restricted. The restriction is that the original reset increment does not exceed the product of the actuator rate constraint and the sampling step size parameter, and is not lower than the negative value of the product. The second intermediate command is obtained by adding the reset increment after the rate constraint to the actuator action command at the previous sampling time; The final reset output command is obtained by sandwiching the second intermediate command between the lower limit of the actuator action command and the upper limit of the actuator action command, that is, the second intermediate command does not exceed the upper limit of the actuator action command and is not lower than the lower limit of the actuator action command; The reset termination condition is that the absolute value of the difference between the final reset output command and the start-up command parameter is not greater than the reset termination threshold parameter. When the reset termination condition is met, the reset operation stops.

Images