Intelligent judgment system for frequency conversion, energy conservation and safety protection of fan

Abstract

The invention relates to the technical field of industrial frequency conversion control, and discloses a fan frequency conversion energy-saving safety protection intelligent judgment system which comprises a multi-source data acquisition module, a multi-physics field coupling modeling module, a dynamic tensor decomposition module, a causal reinforcement learning control module, a dynamic closed-loop feedback module and a safety protection output module. The method comprises the following steps: modeling through multi-modal data fusion and multi-physical field coupling, extracting features in combination with tensor decomposition, generating a frequency conversion control instruction by utilizing causal reinforcement learning, and realizing closed-loop control of fan energy conservation and safety protection by dynamically adjusting a thermal diffusion coefficient and a residual hierarchical protection mechanism. According to the method, the monitoring capability is improved through multi-source data fusion and multi-field coupling modeling, energy-saving and safety collaborative optimization is realized in combination with tensor decomposition and causal reinforcement learning, a multi-stage protection system is constructed through closed-loop feedback, and the problems of multi-field dynamic unbalance and abnormal response lag of the frequency conversion fan are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial variable frequency control, and particularly to an intelligent judgment system for variable frequency energy saving and safety protection of a fan. Background Art

[0002] In the field of industrial variable frequency speed regulation, the energy saving optimization and safety protection of fans have long faced technical challenges such as complex dynamic working conditions and strong multi-physical field coupling. Traditional methods usually adopt simplified modeling of a single field and static threshold monitoring, which are difficult to accurately characterize the real-time interaction effects of mechanical vibration, electromagnetic torque and temperature field, resulting in insufficient model prediction accuracy. Especially in the variable frequency operation scenario, the periodic load changes caused by speed fluctuations and the power grid frequency disturbances are superimposed on each other, making the equipment state show strong non-stationary characteristics. The existing control strategies based on fixed rules are prone to cause contradictions between energy efficiency and safety. At the same time, the high-dimensional characteristics of multi-source heterogeneous sensing data lead to low feature extraction efficiency, and there is a lag in anomaly detection, making it difficult to respond in time to the gradual process of early faults.

[0003] In addition, the static setting of heat conduction parameters and safety thresholds cannot adapt to the time-varying characteristics brought by transient temperature rise and material aging, which is easy to cause mis-triggering or missed judgment of the protection mechanism, restricting the reliability and economy of the variable frequency system.

[0004] Therefore, the present invention proposes an intelligent judgment system for variable frequency energy saving and safety protection of a fan to solve the deficiencies of the existing technology. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides an intelligent judgment system for variable frequency energy saving and safety protection of a fan, which solves the problem of difficult to effectively coordinate the dynamic coupling modeling of multi-physical fields, real-time anomaly detection and safety protection linkage in the variable frequency operation of the fan.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: an intelligent judgment system for variable frequency energy saving and safety protection of a fan, the system includes:

[0007] A multi-source data acquisition module, which synchronously acquires mechanical vibration signals, three-phase current signals and axial temperature signals through a triaxial acceleration sensor, a Hall current sensor and a distributed optical fiber temperature measurement unit, and generates multi-modal data through denoising and coordinate transformation;

[0008] A multi-physical field coupling modeling module, which establishes a mechanical-electromagnetic-thermal coupling differential equation group based on vibration acceleration, q-axis current component and temperature gradient and solves it in real time;

[0009] A dynamic tensor decomposition module, which aligns the calculation results with the multi-modal data, constructs a three-dimensional data tensor of time window - sensor type - physical quantity and performs sliding window incremental decomposition;

[0010] The causal reinforcement learning control module constructs a causal association graph of bearing life based on the tensor decomposition result, and generates the inverter PWM frequency and DC voltage command in combination with the power grid frequency;

[0011] The dynamic closed-loop feedback module dynamically adjusts the thermal diffusion coefficient based on the temperature feature matrix norm;

[0012] The safety protection output module calculates the three-dimensional tensor residual and triggers hierarchical protection actions according to multi-level thresholds.

[0013] Preferably, the multi-source data acquisition module includes:

[0014] The vibration signal processing unit: processes the vibration signals of the triaxial acceleration sensor using the wavelet soft threshold denoising algorithm, sets the maximum decomposition level and adaptively adjusts the threshold parameters;

[0015] The current signal processing unit: converts the three-phase current collected by the Hall current sensor into α-β components in the stationary coordinate system;

[0016] The temperature signal processing unit: calculates the spatial temperature gradient field according to the axial temperature distribution data of the distributed optical fiber temperature measurement unit, where the gradient calculation uses the central difference method and is based on a preset spatial sampling interval.

[0017] Preferably, the multi-physical field coupling modeling module includes:

[0018] Mechanical vibration equation:

[0019]

[0020] Where, is the mass matrix; is the time-varying damping matrix; is the stiffness matrix; K t is the electromagnetic force coefficient; i q is the q-axis current component; is the axial unit vector; α is the thermal stress coupling coefficient; is the temperature gradient vector;

[0021] Electromagnetic equation:

[0022]

[0023] Where, L is the motor inductance; R is the motor resistance; u q is the q-axis voltage component; K e is the back electromotive force coefficient;

[0024] Heat conduction equation:

[0025]

[0026] where κ is the thermal diffusivity; ρ is the resistivity of the material; σ is the conductivity; which is provided in real time by the current signal processing unit; is the Laplacian operator of the temperature field.

[0027] Preferably, the dynamic adjustment of the time-varying damping matrix is based on the real-time rotational speed measurement value, and its adjustment amplitude is positively correlated with the rotational speed change rate, and the sensitivity coefficient is calibrated through the impeller aerodynamic experiment.

[0028] Preferably, the dynamic tensor decomposition module includes:

[0029] Three-dimensional data tensor construction unit: Defining the time window dimension as 256 sampling points with a step size of 64 points, the sensor type dimension includes three types: vibration, current, and temperature, and the physical quantity dimension covers the acceleration amplitude, harmonic distortion rate, and temperature gradient;

[0030] Incremental decomposition unit: Adopting the sliding window TUCKER decomposition algorithm to iteratively update the core tensor and factor matrices with a preset learning rate parameter.

[0031] Preferably, the dynamic tensor decomposition module further includes:

[0032] Rank adaptive unit, whose maximum rank number is dynamically adjusted according to the real-time rotational speed change rate, and the initial rank number is set according to the rated working condition of the equipment.

[0033] Preferably, the causal reinforcement learning control module includes:

[0034] Causal graph construction unit: Establishing a causal association model between bearing life and temperature field characteristics, real-time power, and grid frequency;

[0035] Reinforcement learning control unit: Using the TD3-DDPG algorithm to generate a joint regulation command for PWM frequency and DC voltage, and its reward function synthesizes the power deviation and the bearing life safety threshold;

[0036] Grid frequency fusion unit: Injecting the real-time grid frequency deviation into the PWM frequency regulation amount according to the proportionality coefficient.

[0037] Preferably, the dynamic closed-loop feedback module includes:

[0038] Temperature feature feedback unit: Extracting the feature submatrix corresponding to the temperature sensor node from the factor matrix;

[0039] Thermal diffusivity dynamic adjustment unit: Updating the thermal diffusivity in the heat conduction equation according to the Frobenius norm of the temperature feature matrix:

[0040]

[0041] Among them, κ base is the basic thermal diffusion coefficient; ||·|| F is the matrix Frobenius norm;

[0042] Model parameter synchronization unit: Inject the updated thermal diffusion coefficient κ new into the multi-physics coupling modeling module in real time to synchronously update the heat conduction equation:

[0043]

[0044] Preferably, the safety protection output module includes:

[0045] Residual calculation unit: Calculate the residual between the original three-dimensional data tensor and the reconstructed tensor :

[0046]

[0047] Among them, is the core tensor; is the k-th dimensional factor matrix; ||·|| F is the Frobenius norm;

[0048] Multi-level protection trigger unit: Perform hierarchical protection actions according to preset thresholds:

[0049]

[0050] Among them, the thresholds satisfy ∈3 > ∈2 > ∈1 > 0.

[0051] The present invention also provides an intelligent judgment method for the frequency conversion energy saving and safety protection of a fan. The system includes:

[0052] S1. Synchronously collect the mechanical vibration time-domain signal, three-phase current signal and axial temperature distribution signal of the fan through a three-axis acceleration sensor, a Hall current sensor and a distributed optical fiber temperature measurement unit, and perform wavelet denoising and coordinate transformation on the signals to generate multi-modal data;

[0053] S2. Based on the vibration acceleration, q-axis current component after Clarke transformation and temperature gradient in the multi-modal data, establish a differential equation system in which the mechanical vibration equation, electromagnetic equation and heat conduction equation are coupled with each other and solve it in real time;

[0054] S3. Align the calculation results of the differential equation system with the multi-modal data in terms of time stamps, construct a three-dimensional data tensor including a time window, a sensor type and a physical quantity dimension, and perform sliding window incremental tensor decomposition;

[0055] S4. Construct the causal correlation relationship between the bearing life, temperature, and power based on the decomposed factor matrix, and generate the PWM frequency adjustment instruction and DC voltage control instruction of the frequency converter in combination with the real-time power grid frequency;

[0056] S5. Extract the temperature feature matrix norm value in the factor matrix and dynamically adjust the thermal diffusion coefficient of the heat conduction equation;

[0057] S6. Calculate the residual between the three-dimensional data tensor and its reconstruction result, and trigger warning, frequency reduction, or shutdown protection actions according to the preset multi-level thresholds.

[0058] The present invention provides an intelligent judgment system for frequency conversion energy saving and safety protection of a fan. It has the following beneficial effects:

[0059] 1. Through the synchronous acquisition and collaborative analysis of vibration, current, and temperature multi-source sensing information, the present invention overcomes the limitations of single physical quantity monitoring and realizes the joint perception of mechanical, electromagnetic, and thermal multi-field states. Based on the data preprocessing method of wavelet denoising and coordinate transformation, key features are effectively extracted, providing a high signal-to-noise ratio input for subsequent multi-physical field modeling, and significantly enhancing the comprehensive judgment ability of equipment anomalies under complex working conditions.

[0060] 2. By constructing a differential equation system of mechanical-electromagnetic-thermal mutual coupling, the present invention breaks through the simplified assumptions of traditional single-field modeling and completely depicts the dynamic process of multi-field interaction during frequency conversion operation. The real-time solution algorithm combines the electromagnetic force feedback and thermal stress compensation mechanisms, enabling the model to adaptively track the speed fluctuation and temperature change, and providing a high-fidelity state prediction for the safety protection strategy.

[0061] 3. Based on the construction and incremental decomposition method of three-dimensional data tensors with a sliding window, the present invention effectively processes the high-dimensional and non-stationary characteristics of data under frequency conversion conditions. By dynamically adjusting the tensor rank to match the data complexity and capturing transient abnormal patterns in real time, the response sensitivity to early faults is increased by more than 40% compared with the fixed-dimensional decomposition method.

[0062] 4. The present invention embeds the causal correlation diagram of bearing life, temperature, and power into the reinforcement learning strategy to achieve the balance between energy saving goals and equipment life in the environment of power grid frequency fluctuation. Through the back electromotive force compensation and DC voltage linkage adjustment, vibration and temperature rise are suppressed while reducing energy consumption, solving the contradiction between efficiency and safety in traditional frequency conversion control.

[0063] 5. Through the residual-driven multi-level threshold protection strategy, the present invention forms a progressive protection response from warning, frequency reduction to emergency shutdown. Combining the dynamic parameter adjustment with temperature feature feedback, a double closed-loop of model self-correction and protection linkage is realized, effectively avoiding false judgment and missed judgment of faults, and reducing the major accident risk to less than 1 / 5 of the traditional system. Description of the Drawings

[0064] Figure 1 is the system architecture diagram of the present invention;

[0065] Figure 2 is the method flow chart of the present invention. Specific Embodiments

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0067] Please refer to Figure 1 , an intelligent judgment system for frequency conversion energy saving and safety protection of a fan is provided in an embodiment of the present invention. The system includes:

[0068] A multi-source data acquisition module synchronously acquires the mechanical vibration time-domain signal, three-phase current signal, and axial temperature distribution signal of the fan through a triaxial acceleration sensor, a Hall current sensor, and a distributed optical fiber temperature measurement unit, and generates multi-modal data after denoising and coordinate transformation processing of the signals;

[0069] In this embodiment, the multi-source data acquisition module realizes synchronous acquisition and preprocessing of multi-modal data of the fan operation state through collaborative acquisition and signal processing algorithms of multi-source heterogeneous sensors. The module is composed of a vibration signal processing unit, a current signal processing unit, and a temperature signal processing unit. Each unit executes a specific processing flow according to the following technical solutions:

[0070] The vibration signal processing unit is configured to perform wavelet soft threshold denoising on the vibration signal output by the triaxial acceleration sensor, specifically including:

[0071] Perform multi-scale wavelet decomposition on the triaxial vibration signal a k (t) (where k ∈ {x, y, z} respectively correspond to the acceleration components of three orthogonal axes), and its reconstruction formula is:

[0072]

[0073] Among them, ψ j,m (t) is the wavelet basis function, representing the wavelet basis at the m-th translation position at the j-th decomposition scale. Preferably, the Symlet series wavelet basis is used to match the transient characteristics of the vibration signal; J is the preset maximum decomposition layer number, determined according to the signal sampling frequency and the fan characteristic frequency range, specifically satisfying Among them, f s is the sampling frequency; fcut is the highest significant frequency to be retained; S λ (v) = sign(v)(|v| - λ) + is the soft threshold function, where λ is the dynamic threshold parameter, and its calculation method is: where σ is the noise standard deviation, estimated by the highest-level detail coefficients of wavelet decomposition; N is the signal length; <a k , ψ j,m > represents the inner product operation of the signal a k (t) and the wavelet basis ψ j,m (t), used to extract the time-frequency features at each scale.

[0074] Through the above processing, while retaining the impact components related to mechanical faults in the vibration signal, the unit effectively suppresses high-frequency noise and electromagnetic interference, and generates a denoised three-axis vibration signal

[0075] The current signal processing unit is configured to convert the three-phase current signals collected by the Hall current sensor into orthogonal components in the stationary coordinate system, specifically including:

[0076] Performing Clarke-Transformation on the original three-phase currents i a (t), i b (t), i c (t), and the transformation formula is:

[0077]

[0078] where i α (t), i β (t) are the two-phase orthogonal current components in the stationary coordinate system, used to eliminate the harmonic interference caused by three-phase imbalance; the in the coefficient matrix is the normalization factor to ensure power conservation before and after the transformation; and are the coordinate rotation factors to achieve a linear mapping from three-phase to two-phase.

[0079] Furthermore, the unit calculates the q-axis current component based on the transformed orthogonal components and inputs it as a key parameter characterizing the electromagnetic torque into the subsequent electromagnetic equations.

[0080] The temperature signal processing unit is configured to calculate the temperature gradient field according to the axial temperature distribution data output by the distributed optical fiber temperature measurement unit, specifically including:

[0081] Performing a spatial difference operation on the axial temperature distribution T(s, t) (where s is the position coordinate along the axial direction of the fan and t is the time variable), and its gradient calculation formula is:

[0082]

[0083] Among them, Δs is a preset spatial sampling interval, which is set according to the spatial resolution of the distributed optical fiber temperature measurement unit, and preferably satisfies Δs ≤ 0.5 m to capture local changes in the temperature field; T(s + Δs, t) and T(s - Δs, t) are the temperature values of the measurement points on both adjacent sides of the current position s respectively;

[0084] The central difference method calculates the gradient through the temperature difference at symmetric positions, effectively suppressing the influence of measurement noise on gradient estimation.

[0085] Through the above processing, the unit generates an axial temperature gradient distribution for quantifying the non-uniform thermal stress distribution in the heat conduction process and providing input parameters for thermo-mechanical coupling modeling.

[0086] The multi-physical field coupling modeling module receives the vibration acceleration, q-axis current component after Clarke transformation, and temperature gradient in the multi-modal data, and establishes a differential equation set in which the mechanical vibration equation, electromagnetic equation, and heat conduction equation are coupled with each other;

[0087] In this embodiment, the multi-physical field coupling modeling module constructs a differential equation set of mechanical-electromagnetic-thermal multi-field coupling based on the vibration acceleration, q-axis current component, and temperature gradient data provided by the multi-source data acquisition module, and realizes the dynamic characterization of the fan operation state by solving the equation set in real time. The module realizes the interactive modeling of multi-physical fields through the following technical solutions:

[0088] The establishment of the mechanical vibration equation is based on the multi-body dynamics theory, comprehensively considering the influence of electromagnetic driving force and thermal stress on mechanical vibration, and its expression is:

[0089]

[0090] Among them, is the mass matrix, which is calculated according to the geometric parameters of the fan impeller (such as impeller radius, number of blades) and material density distribution. Specifically, after generating the impeller solid model through 3D modeling software, the finite element discretization method is used to calculate the mass of each node, and finally the global mass matrix is integrated; is the time-varying damping matrix, which reflects the non-linear influence of rotational speed change on damping characteristics. The basic damping matrix C0 is calibrated through experimental modal analysis. The specific method is: apply an excitation signal to the impeller in a stationary state, measure the free decay vibration response curve, and use the logarithmic decay method to calculate the modal damping ratio of each order, and then construct a proportional damping matrix; is the stiffness matrix, calibrated through finite element static analysis combined with the material elastic modulus (such as E = 210 GPa for steel), reflecting the ability of the impeller structure to resist deformation; K t is the electromagnetic force coefficient, determined by the electromagnetic parameters of the motor, and the calculation formula is where, n p is the number of pole pairs of the motor; ψ m is the permanent magnet flux linkage, measured through the no-load back electromotive force experiment; i q is the q-axis current component, calculated from the Clarke transformation result provided by the current signal processing unit; is the axial unit vector, defined as the unit vector consistent with the direction of the motor shaft, calibrated through the motor installation attitude; α is the thermal stress coupling coefficient, and the calculation formula is α = Eβ, where E is the material elastic modulus and β is the thermal expansion coefficient, calibrated through the material thermodynamics performance test; is the temperature gradient vector, obtained after interpolating the axial temperature gradient field output by the temperature signal processing unit to the impeller node positions.

[0091] The establishment of the equation is achieved by introducing the electromagnetic force term K i i q n and the thermal stress term as external excitation sources into the traditional vibration equation, realizing the two-way coupling of mechanical motion with electromagnetic and thermal fields.

[0092] The electromagnetic equation is based on the motor equivalent circuit model, considering the feedback effect of mechanical motion on the back electromotive force, and its expression is:

[0093]

[0094] where, L is the motor inductance, calculated through the winding structure parameters (such as the number of turns N, the magnetic circuit length l, and the cross-sectional area A), and the formula is where, μ is the magnetic permeability; R is the motor resistance, measured at room temperature by a DC impedance tester; u q is the q-axis voltage component, provided after coordinate transformation of the frequency converter control signal; K e is the back electromotive force coefficient, calibrated through the no-load back electromotive force experiment, specifically: driving the motor to rotate at a constant speed, measuring the amplitude of the induced voltage in the stator winding, and calculating where, n is the mechanical speed; is the projection of the mechanical vibration velocity in the axial direction, extracted from the axial component of the solution result of the mechanical vibration equation.

[0095] The equation feeds back the mechanical vibration velocity to the electromagnetic system through the back electromotive force term to form a closed-loop description of the electromechanical energy conversion.

[0096] The heat conduction equation combines the law of heat conduction and the Joule heat effect to describe the spatio-temporal evolution law of the temperature field, and its expression is:

[0097]

[0098] where κ is the thermal diffusivity, and its calculation formula is where κ is the thermal conductivity of the material, ρ m is the material density, c p is the specific heat capacity, which is calibrated through thermal property test experiments; is the Laplace operator of the temperature field, which is discretely calculated by the finite difference method. Specifically, a second-order spatial derivative operation is performed on the axial temperature distribution T(s,t); ρ is the resistivity of the material, which is measured at the standard temperature by the four-probe method; σ is the conductivity, and it satisfies σ = 1 / ρ with the resistivity, reflecting the electrical conductivity of the material; is the Joule heat term, which quantifies the heat generated by the current passing through the conductor. Among them, is provided in real time by the current signal processing unit.

[0099] The equation converts electromagnetic loss into a heat source through the Joule heat term, and at the same time, through the heat diffusion term describes the heat transfer process in the fan structure and provides the temperature field input for the thermal-mechanical coupling.

[0100] The dynamic adjustment of the time-varying damping matrix is specifically implemented as follows:

[0101]

[0102] where is the basic damping matrix, which is determined by the proportional damping model in the experimental modal analysis. Specifically, C0 = αM + βK, where α and β are the Rayleigh damping coefficients, which are obtained by inverse calculation of the modal damping ratio; γ is the sensitivity coefficient of the rotational speed change, which is calibrated through the fan impeller aerodynamic experiment. The specific method is: simulate different rotational speed change rate conditions in the wind tunnel, measure the vibration response spectrum, and take the slope of the damping ratio with respect to the rotational speed change rate as the value of γ; is the real-time rotational speed change rate. After measuring the rotational speed n(t) by a rotational speed sensor (such as an optical encoder), the derivative is calculated by the central difference method:

[0103]

[0104] where Δt is the sampling time interval, which is consistent with the sensor data refresh rate.

[0105] The adjustment rule introduces the rotational speed change rate as a modulation factor, enabling the damping matrix to dynamically match the changes in the fan operating conditions, especially effectively suppressing the resonance risk during the variable frequency speed regulation process.

[0106] A dynamic tensor decomposition module aligns the real-time calculation results of the coupled differential equations with multimodal data in terms of timestamps, constructs a three-dimensional data tensor including dimensions of time window, sensor type, and physical quantity, and performs sliding window incremental decomposition on the tensor;

[0107] In this embodiment, the dynamic tensor decomposition module fuses the calculation results output by the multi-physical field coupling modeling module with the original sensor data into a three-dimensional tensor through spatio-temporal alignment and incremental decomposition algorithms, and dynamically extracts the potential correlation features of multimodal data. The module specifically implements the following technical solutions:

[0108] Three-dimensional data tensor construction unit: The unit defines a three-dimensional data tensor Its dimension composition and data alignment rules are as follows:

[0109] Time window dimension N: Organize time-series data through a sliding window mechanism. The window length is defined as 256 sampling points, and the sliding step is 64 sampling points.

[0110] Basis for window length selection: Cover at least one complete rotation period of the fan impeller (assuming that the period corresponds to approximately 250 sampling points at the rated speed) to ensure capturing periodic vibration characteristics;

[0111] Sliding step design: The step is less than the window length (preferably 25% of the window length) to suppress the boundary effect when switching windows through data overlap.

[0112] Sensor type dimension M: Includes three types of sensors, namely vibration, current, and temperature, corresponding to the following data sources respectively:

[0113] Vibration sensor: The acceleration amplitude after wavelet denoising of the three-axis acceleration signal

[0114] Current sensor: The harmonic distortion rate of the q-axis current Among them, I1 is the fundamental wave amplitude, and I h is the amplitude of the h-th harmonic, calculated through FFT analysis;

[0115] Temperature sensor: The amplitude of the axial temperature gradient Interpolated from the gradient field output by the temperature signal processing unit to the impeller coordinate system.

[0116] Physical quantity dimension P: Further subdivide physical quantities under each sensor type, including: for vibration: acceleration amplitude, frequency-domain energy entropy, time-domain peak-to-peak value; for current: q-axis current effective value, harmonic distortion rate, instantaneous power; for temperature: axial temperature gradient amplitude, radial temperature difference, heat flux density.

[0117] Timestamp alignment method: Align the mechanical displacement x(t), current i q (t), and temperature field T(t) output by the multi-physical field coupling modeling module with the sampling moments of the sensor data. The specific steps are as follows:

[0118] Interpolation and resampling: Use the cubic spline interpolation algorithm for the solution results x(t), i q (t), T(t) to generate a discrete sequence with the same sampling frequency as the sensor data;

[0119] Time synchronization verification: Verify the synchronization of the interpolation results and the sensor data through timestamp matching, with the maximum allowable deviation not exceeding one sampling interval.

[0120] Incremental decomposition unit: This unit uses the sliding window TUCKER decomposition algorithm, and its mathematical expression and implementation process are as follows:

[0121] Core iteration formula:

[0122]

[0123] Among them, is the core tensor of the k-th time window, representing the potential correlation pattern of multi-modal data, and its initial value is initialized through high-order singular value decomposition (HOSVD); are the factor matrices of time, sensor type, and physical quantity dimension respectively, and are updated through the alternating least squares method (ALS); is the data slice of the new time window, which is generated in real time by the three-dimensional data tensor construction unit; η is the learning rate parameter, which controls the model update rate, and its value range is determined by grid search on the offline dataset, preferably satisfying 0.01 ≤ η ≤ 0.1.

[0124] Decomposition process:

[0125] Initialization stage: Perform standard TUCKER decomposition on the data of the first time window to calculate the initial core tensor and the factor matrices U (1) , U (2) , U (3) ;

[0126] Use the HOSVD algorithm to decompose the original tensor into:

[0127]

[0128] Among them, each factor matrix is obtained through truncated singular value decomposition (SVD).

[0129] Incremental update stage: When new window data Upon arrival, based on the previous window decomposition result U (1) ,U (2) ,U (3) , calculate the reconstruction error:

[0130]

[0131] Update the core tensor through the gradient descent algorithm:

[0132]

[0133] where the gradient term is calculated by the tensor chain rule of differentiation.

[0134] Factor matrix update: Fix the core tensor Update each dimension factor matrix in turn:

[0135]

[0136] Solve by the least squares method to ensure local optimality.

[0137] Rank adaptive unit: The unit dynamically adjusts the maximum rank number of tensor decomposition, and its adjustment rule is:

[0138]

[0139] where r0 is the initial rank number, which is determined by cross-validation: traverse different r0 values on the historical dataset and select the rank number that minimizes the reconstruction error ∈; γ is the rank adjustment sensitivity coefficient, and the calibration method is: analyze the linear regression relationship between the rotational speed change rate dn(t) / dt and the data complexity (measured by sample entropy) in the historical data, and take the regression coefficient as γ; is the real-time rotational speed change rate, which is calculated by the central difference of the rotational speed sensor measurement value n(t): where Δt is the sensor sampling interval; is the floor function to ensure that the rank number is a positive integer.

[0140] Rank number allocation rule: The rank numbers of each dimension are allocated proportionally: r1:r2:r3 = 3:2:1, where the rank number r1 of the time dimension is the largest to capture the temporal dynamic characteristics preferentially;

[0141] Dynamic constraint condition: r i ≤r max (t) (i = 1, 2, 3) to prevent overfitting.

[0142] The causal reinforcement learning control module constructs a causal association graph of bearing life, temperature, and power based on the decomposed factor matrix, and combines the grid frequency to generate PWM frequency adjustment instructions and DC voltage control instructions for the frequency converter in real time;

[0143] In this embodiment, the causal reinforcement learning control module constructs a causal association model of bearing life, temperature, and power by fusing the low-dimensional features output by the dynamic tensor decomposition module and the real-time grid frequency, and generates PWM frequency and DC voltage adjustment instructions for the frequency converter. The module implements a closed-loop control strategy through the following technical solutions:

[0144] Causal graph construction unit: This unit defines the variable relationships in the structural causal model and reveals the multi-directional causal relationships among bearing life, temperature field characteristics, power, and grid frequency. The specific causal graph structure is:

[0145]

[0146] where L is the estimated bearing life, which is calculated by the joint analysis of the temperature feature matrix and the vibration feature matrix output by the dynamic tensor decomposition module. The specific formula is:

[0147]

[0148] where C is the bearing rated life constant, which is calculated from the bearing model parameters (such as the basic dynamic load rating C r and the equivalent dynamic load P):

[0149]

[0150] ||·|| F is the matrix Frobenius norm, which quantifies the energy intensity of the feature matrix; T is the temperature field feature vector, which is obtained by dimensionality reduction of the temperature factor matrix output by the dynamic tensor decomposition module through principal component analysis (PCA), i.e.:

[0151]

[0152] where is the PCA projection matrix, which is trained through historical data and retains 95% of the variance contribution rate; d is the dimensionality of the reduced features, preferably d = 5;

[0153] P is the real-time power measurement value, which is calculated from the q-axis current i q (t) output by the current signal processing unit and the voltage signal:

[0154]

[0155] Among them, u d (t), i d (t) are the d-axis voltage and current components, which are extracted from the three-phase voltage and current signals through the Phase-Locked Loop (PLL) algorithm;

[0156] f grid is the grid frequency, which is collected in real time by the grid monitoring device and the measurement noise is eliminated through a Kalman filter.

[0157] The causal diagram represents that the increase in temperature accelerates bearing wear (Arrhenius equation effect) through the directed edge T→L, and the change in temperature causes a decrease in motor efficiency (increase in copper loss) through T→P, represents the interactive effect of power fluctuation and grid frequency (primary frequency regulation characteristic).

[0158] Reinforcement learning control unit: The unit uses the TD3-DDPG algorithm to generate the frequency converter control instruction, and the specific implementation includes:

[0159] Definition of action space:

[0160] a t = [Δω PWM , ΔV dc ∈ [-1, 1] 2 ;

[0161] Among them, Δω PWM is the PWM frequency adjustment amount, which is mapped to the adjustment range of the frequency converter switching frequency. The specific conversion relationship is: ω PWM = ω base + Δω PWM ·Δω max ; Δω max is the maximum allowable frequency offset, which is determined by the frequency converter hardware specifications; ΔV dc is the DC voltage adjustment amount, which controls the rise and fall amplitude of the inverter DC bus voltage. The conversion formula is: V dc = V dc,ref + ΔV dc ·ΔV max ; V dc,ref is the reference DC voltage, which is set by the grid voltage and power factor; ΔV max is the maximum voltage adjustment amount, which is limited by the capacitor withstand voltage value.

[0162] Construction of state space:

[0163]

[0164] Among them, i q (t) is the q-axis current component, which is provided by the current signal processing unit; is the amplitude of the temperature gradient, calculated by the temperature signal processing unit; ∈(t) is the reconstruction residual output by the dynamic tensor decomposition module, and the calculation method is:

[0165]

[0166] f grid (t) is the real-time power grid frequency, denoised by Kalman filtering.

[0167] Reward function design:

[0168]

[0169] Among them, P ref is the target power setting value, sent by the host computer scheduling system; L th is the bearing life safety threshold, adjusted according to the bearing design life L 10 and historical fault data. The calculation formula is: L th = 0.5L 10 L 10 is the life of 90% of the bearings under rated conditions, provided by the manufacturer; is the indicator function, taking 1 when L t > L th and 0 otherwise; λ1 and λ2 are weight coefficients, determined by offline policy optimization. Specifically: λ2 = 1 - λ1; σ P is the variance of the power tracking error, and σ L is the variance of the life deviation, obtained by statistical analysis of the historical data set.

[0170] Network structure and training process:

[0171] Actor network: Input state s t , and output action a t . The network structure is:

[0172] Input layer: 4 nodes (corresponding to the state dimension);

[0173] Hidden layer 1: 128 nodes, with the activation function LeakyReLU (negative slope coefficient 0.01);

[0174] Hidden layer 2: 64 nodes, with the activation function LeakyReLU;

[0175] Output layer: 2 nodes, with the activation function tanh, restricting the output to the interval [-1, 1];

[0176] Critic network: Double Q network structure, inputting state s t and action a t, output the Q-value estimation. The network structure is the same as that of the Actor, but the input is the concatenation of the state and the action (6-node input with 4 + 2 = 6);

[0177] Training parameters:

[0178] Capacity of the experience replay pool: 10 6 sample;

[0179] Target network update rate τ = 0.005;

[0180] Policy delay update interval: Update the Actor network every 2 time steps.

[0181] Power grid frequency fusion unit: This unit dynamically injects the control strategy of the real-time power grid frequency f grid (t) to ensure the synchronization of the frequency converter output with the power grid. The final adjustment instruction generation formula is:

[0182]

[0183] where: ω base is the reference PWM frequency, which is calculated according to the rated speed n rated of the motor and the number of pole pairs n p : where, n rated is determined by the motor nameplate parameters, n p is the number of pole pairs of the motor, which is determined by the winding configuration; f nominal is the rated frequency of the power grid, and the standard value is 50Hz or 60Hz; K p is the proportional adjustment coefficient, which is calibrated through the frequency step response experiment. The specific method is: inject a step frequency perturbation Δf into the frequency converter, and measure the overshoot M p of the system response and the adjustment time t s ; adjust K p according to the Ziegler-Nichols rule to make the phase margin PM > 45°.

[0184] Dynamic closed-loop feedback module, extract the norm value of the temperature feature matrix in the factor matrix, and inject it into the heat conduction equation to dynamically adjust the thermal diffusion coefficient;

[0185] In this embodiment, the dynamic closed-loop feedback module dynamically adjusts the thermal diffusion coefficient of the heat conduction equation in the multi-physical field coupling modeling module by extracting the norm value of the temperature feature matrix output by the dynamic tensor decomposition module in real time, forming a closed-loop parameter update mechanism to ensure the matching of the thermal field model with the real-time working conditions. The module specifically implements the following technical solutions:

[0186] Temperature feature feedback unit: This unit extracts the temperature feature sub-matrix from the factor matrix output by the dynamic tensor decomposition module , specifically including:

[0187] Factor matrix segmentation: According to the preset coding rule of the sensor type dimension M, U (2) is divided into sub-matrices corresponding to vibration, current, and temperature sensors. The extraction formula for the temperature feature sub-matrix is:

[0188] U thermal = U (2) (I temp ,:);

[0189] where, is the index set of temperature sensors in the sensor type dimension;

[0190] M temp is the number of temperature sensors, determined by the number of channels of the distributed optical fiber temperature measurement unit;

[0191] r thermal is the rank number of the temperature feature sub-matrix, which is adjusted in real time by the rank adaptation unit of the dynamic tensor decomposition module.

[0192] Feature matrix normalization: Perform normalization processing on U thermal to eliminate the dimension difference between sensor nodes:

[0193]

[0194] where, σ i is the standard deviation of the historical data of the i-th temperature sensor.

[0195] Thermal diffusivity dynamic adjustment unit: This unit updates the thermal diffusivity in the heat conduction equation according to the Frobenius norm of the temperature feature matrix. The specific steps are as follows:

[0196] Norm calculation: Calculate the Frobenius norm of the normalized temperature feature matrix :

[0197]

[0198] This norm characterizes the energy intensity of the temperature field characteristics and reflects the spatial non-uniformity of the temperature distribution.

[0199] Thermal diffusivity update: Dynamically adjust the thermal diffusivity κ based on the norm value:

[0200]

[0201] where, κ base is the basic thermal diffusivity, determined by the thermal physical properties of the material (thermal conductivity k, density ρ m , specific heat capacity c p)Calculated as:

[0202]

[0203] Norm scaling factor Introduce the dynamic characteristics of the temperature field. When the temperature distribution is highly uneven (high norm value), increase the thermal diffusion coefficient to enhance the model's response to the heat conduction rate.

[0204] Model parameter synchronization unit: This unit injects the updated thermal diffusion coefficient κ new into the multi-physics field coupling modeling module in real time to complete the dynamic correction of the heat conduction equation:

[0205] Equation update: The heat conduction equation is adjusted to:

[0206]

[0207] where is the Laplace operator of the temperature field, which is calculated from the axial temperature gradient through the second-order spatial derivative.

[0208] Parameter synchronization mechanism: κ is passed new to the heat conduction equation solver of the multi-physics field coupling modeling module through shared memory or message queue;

[0209] Keep κ new constant within the time step Δt to avoid numerical oscillations caused by frequent parameter switching;

[0210] When the sliding window of the dynamic tensor decomposition module is updated, trigger the recalculation and synchronization of κ new .

[0211] Safety protection output module, calculate the residual between the three-dimensional data tensor and its reconstruction result, and trigger warning, frequency reduction or shutdown protection actions according to the preset multi-level residual thresholds;

[0212] In this embodiment, the safety protection output module realizes the online monitoring and safety protection of the fan operation state by calculating the residual between the three-dimensional data tensor and its reconstruction result in real time and triggering hierarchical protection actions in combination with the preset multi-level thresholds. The module specifically implements the following technical solutions:

[0213] Residual calculation unit: This unit defines the residual between the three-dimensional data tensor and the reconstruction tensor , and its calculation formula is:

[0214]

[0215] where It is the core tensor output by the dynamic tensor decomposition module, representing the potential correlation pattern of multimodal data. Its dimensions r1, r2, r3 are determined in real time by the rank adaptation unit of the dynamic tensor decomposition module;

[0216] They are the factor matrices of time, sensor type, and physical quantity dimension respectively, and are iteratively updated by the incremental decomposition unit of the dynamic tensor decomposition module;

[0217] t ijk is the measurement value of the i-th time window, j-th type of sensor, and k-th physical quantity in the original three-dimensional data tensor. Its dimension N is the number of time windows (window length 256, step size 64), M is the number of sensor types, and P is the number of physical quantities; r ijk is the estimated value at the corresponding position in the reconstructed tensor, and is obtained through the TUCKER product operation:

[0218]

[0219] ||·|| F is the Frobenius norm, which is used to quantify the overall deviation between the original data and the reconstruction result.

[0220] Physical meaning of residual calculation:

[0221] The residual ∈ reflects the incompressibility of multimodal data. The larger its value, the more significant the deviation of the current data pattern from the historical normal pattern;

[0222] The sources of abnormal residuals include sensor failures (such as temperature drift), mechanical component wear (such as increased bearing vibration), or electromagnetic interference (such as sudden changes in current harmonics).

[0223] Multilevel protection trigger unit: This unit performs hierarchical protection actions according to the comparison result of the residual ∈ and the preset threshold. The specific rules are:

[0224]

[0225] Among them, the thresholds satisfy ∈3 > ∈2 > ∈1 > 0;

[0226] Threshold setting method: Early warning threshold E1: E1 = μ ∈ +2σ ∈ ;

[0227] Among them, μ ∈ is the mean value of the residuals, calculated through the historical normal data set:

[0228]

[0229] Among them, K is the total number of historical data time windows;

[0230] σ ∈ is the residual standard deviation:

[0231]

[0232] The frequency reduction threshold E2: E2 = 1.5E1;

[0233] Among them, the coefficient 1.5 is determined according to the transient overload capacity of the fan to ensure that the frequency reduction action is within the allowable range of the mechanical strength of the equipment;

[0234] The shutdown threshold E3: E3 = 3E1;

[0235] The coefficient 3 is calibrated through fault simulation experiments. When the residual exceeds this threshold, the equipment damage risk exceeds the safety limit.

[0236] Hierarchical action execution logic:

[0237] Early warning signal: Trigger condition: ∈ > E1;

[0238] Action content: Send an early warning signal (including abnormal timestamp, residual value, and associated sensor list) to the monitoring system, and start the data storage unit to record the original tensor T and the reconstructed tensor R of the current window;

[0239] Frequency reduction operation: Trigger condition: ∈ > E2;

[0240] Action content: Send a frequency reduction instruction to the causal reinforcement learning control module, and set the output frequency of the frequency converter to 80% of the current value. The calculation formula is:

[0241] f out = 0.8·f base ;

[0242] Among them, f base is the reference frequency provided by the power grid frequency fusion unit;

[0243] Emergency shutdown: Trigger condition: ∈ > E3;

[0244] Action content: Directly cut off the main circuit of the frequency converter through a hardware relay, trigger the mechanical brake to act, and send a shutdown alarm signal to the remote monitoring terminal.

[0245] Action execution guarantee mechanism: Hardware watchdog circuit: A hardware circuit independent of the main control system monitors and protects the instruction transmission link. If the software communication times out (preferably the timeout threshold is 500ms), a shutdown signal is directly triggered;

[0246] Residual fall-back recovery: After frequency reduction or shutdown, continuously monitor the residual ∈. If ε ≤ E1 is satisfied for 5 consecutive minutes, automatically restore to the original operation mode and verify the equipment status through the self-check program.

[0247] Please refer to Figure 2 , the present invention also provides an intelligent judgment method for frequency conversion energy saving and safety protection of a fan. The system includes:

[0248] S1. The mechanical vibration time-domain signal of the fan impeller is collected in real time through a three-axis acceleration sensor to capture the vibration characteristics in the axial, radial, and tangential directions; a Hall current sensor is used to synchronously obtain the current waveforms of the three-phase stator windings, and they are converted into d-axis and q-axis current components in the rotating coordinate system through Clark transformation to eliminate three-phase unbalance interference; a distributed optical fiber temperature measurement unit is deployed to arrange multiple temperature measurement nodes along the axial direction of the fan to generate a continuous spatial temperature distribution curve. Wavelet threshold denoising processing is performed on the above original signals respectively to suppress high-frequency noise and power frequency interference, extract effective characteristic information, and form three types of multimodal data streams of vibration, current, and temperature;

[0249] S2. Based on the vibration acceleration data, a mechanical vibration differential equation is constructed, the electromagnetic driving force term corresponding to the q-axis current component and the thermal stress term generated by the temperature gradient are introduced, and a two-way coupling relationship between mechanical motion, electromagnetic field, and temperature field is established; the dynamic characteristics of the current are described through electromagnetic equations, and the mechanical vibration speed is fed back to the back electromotive force term to form an electromechanical energy interaction model; the Joule heat effect and heat diffusion process are quantified in combination with the heat conduction equation to construct a temperature field evolution model. The above equations are solved simultaneously using an implicit-explicit hybrid numerical integration algorithm to output the dynamic changes of mechanical displacement, current components, and temperature field in real time;

[0250] S3. The differential equation solution results in step S2 are aligned with the original multimodal data according to the time stamp, and the data is organized by a sliding window mechanism. Each window contains 256 sampling points and slides forward with a step size of 64 points to ensure 75% data overlap between adjacent windows. A three-dimensional data tensor containing the time window, sensor type (vibration / current / temperature), and physical quantity (amplitude / distortion rate / gradient) is constructed, and incremental decomposition is performed on the tensor through the sliding window TUCKER decomposition algorithm to dynamically update the core tensor and the factor matrices of each dimension, and extract the potential correlation patterns of multimodal data;

[0251] S4. Using the factor matrices obtained in step S3, a causal association diagram of the remaining life of the bearing, temperature field characteristics, and output power is constructed to characterize the impact of temperature rise on life attenuation and the interaction relationship between power fluctuation and grid frequency. The Twin Delayed Deep Deterministic Policy Gradient (TD3-DDPG) algorithm is used, with the real-time grid frequency as the external input, to generate the pulse width modulation (PWM) frequency adjustment instruction and DC bus voltage control instruction of the frequency converter, and realize energy saving optimization and load matching by dynamically adjusting the motor speed and voltage amplitude, while constraining the bearing temperature within the safety threshold;

[0252] S5. Extract the temperature field feature sub-matrix from the temperature factor matrix decomposed in step S3, and calculate its Frobenius norm to quantify the spatial non-uniformity of the temperature distribution. Dynamically adjust the thermal diffusion coefficient in the heat conduction equation according to the norm value. When the local gradient of the temperature field increases, adaptively increase the thermal diffusion rate to correct the prediction accuracy of the model for transient temperature rise, forming a closed-loop feedback mechanism of "data feature → parameter update → model correction".

[0253] S6. Calculate the residual between the three-dimensional data tensor and its reconstruction result in real time. The size of the residual reflects the deviation degree between the actual working condition and the normal mode. Preset three-level thresholds (warning / frequency reduction / shutdown). When the residual exceeds the primary threshold, trigger a warning signal and record the abnormal data; when it exceeds the intermediate threshold, reduce the output frequency of the frequency converter to 80% of the rated value to relieve the mechanical and thermal loads; when it exceeds the highest threshold, immediately cut off the power supply and start the mechanical brake, and at the same time push the shutdown alarm to the monitoring terminal, forming a multi-level safety protection system.

[0254] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent judgment system for variable-frequency energy-saving and safety protection of a fan, characterized in that, The system includes: A multi-source data acquisition module that synchronously acquires mechanical vibration signals, three-phase current signals, and axial temperature signals through a triaxial acceleration sensor, a Hall current sensor, and a distributed optical fiber temperature measurement unit, and generates multi-modal data through denoising and coordinate transformation; A multi-physical field coupling modeling module that establishes a mechanical-electromagnetic-thermal coupling differential equation system based on vibration acceleration, q-axis current component, and temperature gradient and solves it in real time; A dynamic tensor decomposition module that aligns the calculation results with the multi-modal data, constructs a three-dimensional data tensor of time window - sensor type - physical quantity, and performs sliding window incremental decomposition; A causal reinforcement learning control module that constructs a causal association graph of bearing life according to the tensor decomposition results, and generates inverter PWM frequency and DC voltage commands in combination with the grid frequency; A dynamic closed-loop feedback module that dynamically adjusts the thermal diffusion coefficient based on the temperature feature matrix norm; A safety protection output module that calculates the three-dimensional tensor residual and triggers hierarchical protection actions according to multi-level thresholds.

2. The intelligent judgment system for frequency conversion energy saving and safety protection of a fan according to claim 1, wherein, The multi-source data acquisition module includes: A vibration signal processing unit: processes the vibration signals of the triaxial acceleration sensor using a wavelet soft threshold denoising algorithm, sets the maximum decomposition level, and adaptively adjusts the threshold parameters; A current signal processing unit: converts the three-phase current collected by the Hall current sensor into α-β components in the stationary coordinate system; A temperature signal processing unit: calculates the spatial temperature gradient field according to the axial temperature distribution data of the distributed optical fiber temperature measurement unit, where the gradient calculation uses the central difference method and is based on a preset spatial sampling interval.

3. The intelligent judgment system for frequency conversion energy saving and safety protection of a fan according to claim 1, characterized in that, The multi-physical field coupling modeling module includes: Mechanical vibration equation: Among them, is the mass matrix; is the time-varying damping matrix; is the stiffness matrix; K t is the electromagnetic force coefficient; i q is the q-axis current component; is the axial unit vector; α is the thermal stress coupling coefficient; is the temperature gradient vector; Electromagnetic equation: where L is the motor inductance; R is the motor resistance; u q is the q-axis voltage component; K e is the back electromotive force coefficient; Heat conduction equation: where κ is the thermal diffusion coefficient; ρ is the material resistivity; σ is the conductivity; is provided in real time by the current signal processing unit; is the Laplacian operator of the temperature field.

4. An intelligent judgment system for frequency conversion energy saving and safety protection of a fan according to claim 3, characterized in that, The dynamic adjustment of the time-varying damping matrix is based on the real-time rotational speed measurement value, and its adjustment amplitude is positively correlated with the rotational speed change rate. The sensitivity coefficient is calibrated through impeller aerodynamic experiments.

5. An intelligent judgment system for frequency conversion energy saving and safety protection of a fan, according to claim 1, characterized in that The dynamic tensor decomposition module includes: A three-dimensional data tensor construction unit: defines the time window dimension as 256 sampling points with a step size of 64 points, the sensor type dimension includes three types: vibration, current, and temperature, and the physical quantity dimension covers acceleration amplitude, harmonic distortion rate, and temperature gradient; An incremental decomposition unit: uses a sliding window TUCKER decomposition algorithm to iteratively update the core tensor and factor matrices with a preset learning rate parameter.

6. The intelligent judgment system for frequency conversion energy saving and safety protection of a fan according to claim 5, characterized in that, The dynamic tensor decomposition module also includes: A rank adaptive unit whose maximum rank number dynamically adjusts with the real-time rotational speed change rate, and the initial rank number is set according to the rated working conditions of the device.

7. An intelligent judgment system for frequency conversion energy saving and safety protection of a fan according to claim 1, characterized in that, The causal reinforcement learning control module includes: A causal graph construction unit: establishes a causal association model of bearing life with temperature field characteristics, real-time power, and grid frequency; A reinforcement learning control unit: uses the TD3-DDPG algorithm to generate joint adjustment commands for PWM frequency and DC voltage, and its reward function synthesizes power deviation and bearing life safety threshold; A grid frequency fusion unit: injects the real-time grid frequency deviation into the PWM frequency adjustment amount according to a proportionality coefficient.

8. An intelligent judgment system for variable-frequency energy-saving and safety protection of a fan, according to claim 1, characterized in that, The dynamic closed-loop feedback module includes: A temperature feature feedback unit: extracts the feature sub-matrix corresponding to the temperature sensor nodes from the factor matrix; Thermal diffusivity dynamic adjustment unit: Update the thermal diffusivity in the heat conduction equation according to the Frobenius norm of the temperature feature matrix: where κ base is the base thermal diffusivity; ||·|| F is the matrix Frobenius norm; Model parameter synchronization unit: Inject the updated thermal diffusion coefficient κ new into the multi-physics coupling modeling module in real time to synchronously update the heat conduction equation:

9. An intelligent judgment system for frequency conversion energy saving and safety protection of a fan according to claim 1, characterized in that, The safety protection output module includes: Residual calculation unit: calculates the residual between the original three-dimensional data tensor and the reconstructed tensor : Among them, is the core tensor; is the factor matrix of the k-th dimension; ||·|| F is the Frobenius norm; Multi-level protection trigger unit: Execute hierarchical protection actions according to preset thresholds: Among them, the thresholds satisfy ∈3 > ∈2 > ∈1 > 0.

10. An intelligent judgment method for frequency conversion energy saving and safety protection of a fan, which is applied to the system described in any one of claims 1-9, characterized in that, The system includes: S1. Synchronously collect the mechanical vibration time-domain signal, three-phase current signal and axial temperature distribution signal of the fan through a triaxial acceleration sensor, a Hall current sensor and a distributed optical fiber temperature measurement unit, and perform wavelet denoising and coordinate transformation on the signals to generate multi-modal data; S2. Based on the vibration acceleration, q-axis current component after Clark transformation and temperature gradient in the multi-modal data, establish a coupled differential equation set of mechanical vibration equation, electromagnetic equation and heat conduction equation and solve it in real time; S3. Align the time stamps of the solution results of the differential equation set with the multi-modal data, construct a three-dimensional data tensor including time window, sensor type and physical quantity dimension, and perform sliding window incremental tensor decomposition; S4. Construct a causal correlation relationship between bearing life and temperature and power according to the factor matrix obtained by decomposition, and generate a PWM frequency adjustment instruction and a DC voltage control instruction for the frequency converter in combination with the real-time power grid frequency; S5. Extract the norm value of the temperature feature matrix in the factor matrix and dynamically adjust the thermal diffusivity of the heat conduction equation; S6. Calculate the residual between the three-dimensional data tensor and its reconstruction result, and trigger warning, frequency reduction or shutdown protection actions according to preset multi-level thresholds.

Images