High-precision liquid level measurement anti-interference calibration method

Abstract

The invention relates to the technical field of industrial process measurement, and discloses a high-precision liquid level measurement anti-interference calibration method, which comprises the following steps of: 1, synchronously acquiring a liquid level signal, a mechanical vibration signal, an electromagnetic field intensity signal and a temperature distribution signal through a multi-mode sensor array to form multi-physical field data; 2, constructing a hypergraph adjacency tensor based on the coupling relationship of the multi-physics field data, and generating a Laplacian matrix constraint; step 3, performing CP tensor decomposition with hypergraph constraint on an original signal tensor formed by the multi-physical field data, and separating a liquid level signal kernel from an interference kernel; and step 4, inputting the decomposed signal kernel characteristics and the sensor impedance parameters into a depth deterministic strategy gradient algorithm to realize hardware acceleration of a processing flow. A multi-physics field coupling relation is constructed through a hypergraph adjacency tensor and a Laplacian matrix, and the limitation of a traditional binary correlation model is broken through.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process measurement, and specifically to a high-precision liquid level measurement anti-interference calibration method. Background Technique

[0002] In industrial fields such as petrochemical and nuclear power generation, high-precision liquid level measurement is a core link to ensure process safety and process control. In a complex industrial environment, the coupling effects of multiple physical fields such as mechanical vibration, electromagnetic interference, and temperature gradient will cause serious distortion of the measurement signal. Traditional single-modal anti-interference methods are difficult to cope with the synergistic effects of multi-source heterogeneous interference, and there is an urgent need to develop intelligent calibration technologies based on multi-physical field coupling modeling.

[0003] Current high-precision liquid level measurement anti-interference calibration methods mainly rely on fixed threshold filtering, frequency domain notch filtering, or off-line calibration compensation. These methods are usually designed based on static interference models, unable to dynamically adapt to time-varying interference characteristics, and there are problems with insufficient model decoupling ability when dealing with multi-physical field cross-coupling. For example, the combined action of vibration and electromagnetic interference will cause non-linear signal aliasing, and existing methods are prone to signal over-attenuation or residual interference leakage due to the lack of explicit modeling of the interference coupling mechanism.

[0004] However, the following bottlenecks generally exist in the prior art: First, the synergistic suppression mechanism of multi-physical field coupling interference is not yet perfect, resulting in limited separation accuracy of liquid level signals and interference components; second, the closed-loop cooperation ability of parameter optimization and hardware execution is insufficient, making it difficult to achieve a balance between anti-interference performance and real-time performance under complex working conditions. The present invention realizes the synergistic suppression and adaptive calibration of multi-modal interference for the first time through the dynamic optimization of hypergraph-constrained tensor decomposition and deep reinforcement learning, providing an innovative solution to the above problems. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a high-precision liquid level measurement anti-interference calibration method, which solves the problems of low signal separation accuracy and lack of dynamic calibration ability caused by insufficient synergistic suppression ability of multi-physical field coupling interference in existing liquid level measurement methods.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A high-precision liquid level measurement anti-interference calibration method, including: Step 1: Synchronously collect liquid level signals, mechanical vibration signals, electromagnetic field intensity signals, and temperature distribution signals through a multi-modal sensor array to form multi-physical field data; Step 2: Construct a hypergraph adjacency tensor based on the coupling relationship of the multi-physical field data, and generate a Laplacian matrix constraint; Step 3: Perform CP tensor decomposition with hypergraph constraints on the original signal tensor formed by the multi-physical field data to separate the liquid level signal core and the interference core; Step 4: Input the decomposed signal kernel features and sensor impedance parameters into the deep deterministic policy gradient algorithm to dynamically optimize the tensor decomposition rank, filter cut-off frequency, and self-healing trigger current; Step 5: Generate deposition current parameters according to the optimization results and repair the sensor sensitive layer through electrochemical deposition; Step 6: Use a reconfigurable computing unit to parallelly execute tensor decomposition, policy inference, and self-healing control operations to achieve hardware acceleration of the processing flow.

[0007] Preferably, the multi-physical field data in Step 1 includes the following types and corresponding sensors: Liquid level signal: collected by a fiber Bragg grating sensor; Mechanical vibration signal: collected by a three-axis MEMS accelerometer; Electromagnetic field intensity signal: collected by an electromagnetic field probe; Temperature distribution signal: collected by a distributed temperature sensor; among them, the synchronous acquisition of the multi-physical field data is realized through a time synchronization protocol and satisfies: The constructed original signal tensor is a four-dimensional tensor , and the dimension definitions are as follows: : number of time slices; : number of spatial sampling points; : number of frequency sub-bands; : number of physical field types.

[0008] Preferably, the construction of the hypergraph adjacency tensor in Step 2 satisfies the following conditions: Define a third-order adjacency tensor: ; Among them is the number of physical field types; the tensor element if and only if there is a coupling interference relationship between the -th, -th, and -th types of physical fields; Generate a hypergraph Laplacian matrix based on the adjacency tensor, and its expression is: ; Among them, is the hyperedge degree diagonal matrix, and its diagonal element .

[0009] Preferably, the objective function of the CP tensor decomposition with hypergraph constraints in Step 3 is: ; Among them: is the original signal tensor; , , are the time, space, and frequency factor matrices, respectively; is the hypergraph Laplacian matrix; is the hypergraph constraint coefficient; denotes the outer product of vectors, is the Frobenius norm; The liquid level signal kernel and the interference kernel are separated to satisfy: ; where and are the component classification sets based on hypergraph constraints.

[0010] Preferably, the calculation of the deposition current parameters in step four satisfies the following formula: ; where: is the deposition current density; , represents the deviation of the current impedance of the sensor sensitive layer from the reference impedance; is the initial impedance of the sensor; is the activation energy of the electrochemical deposition reaction; is the Boltzmann constant; is the absolute temperature of the sensitive layer surface; is the deposition rate proportionality coefficient; The electrochemical deposition transports the electrolyte to the surface of the sensitive layer through a microfluidic channel and applies a current density for a duration , until the impedance recovers to .

[0011] Preferably, the dynamic optimization process of the deep deterministic policy gradient algorithm in step four satisfies the following conditions: State space: Signal-to-noise ratio SNR; Current tensor decomposition rank R; Sensor sensitive layer impedance Z; Action space: Tensor decomposition rank adjustment amount ΔR; Filter cut-off frequency fc; Self-healing trigger current Ih; Reward function: ; where: : The number of self-healing triggers per unit time; : The signal processing delay time; is the weight coefficient.

[0012] Preferably, the parallel execution of the reconfigurable computing unit in step six satisfies the following conditions: Supports three operation modes: Mode 1: Execute the alternating least squares algorithm for tensor decomposition to calculate matrix chain multiplication: ; Mode 2: Calculate the output of the fully connected layer of the policy network: ; Mode 3: Solve the partial differential equation of the deposition current: ; Adopt a double-buffered data flow mechanism: While the current frame of data is processed sequentially through Mode 1 to Mode 3, the next frame of data is pre-loaded into the cache through a high-speed interface, where: The current frame of data: denoted as ; The pre-loaded data: denoted as .

[0013] Preferably, step one further includes: The multi-physical field data after synchronous acquisition needs to be processed for cross-domain alignment, including: Perform phase compensation on the vibration signal and the electromagnetic signal, and the compensation formula is: ; Among them, is the clock synchronization error, is the signal center frequency; Perform spatial interpolation on the temperature distribution signal to generate a temperature field matrix that matches the spatial resolution of the liquid level signal.

[0014] Preferably, step five further includes: After the electrochemical deposition is completed, verify the sensor repair effect through the following steps: Collect the impedance of the repaired sensor , and calculate the impedance recovery rate: ; If , it is determined that the repair is successful; otherwise, regenerate the deposition current parameters for secondary repair.

[0015] Preferably, step six further includes: Its interaction with the main control module satisfies: The main control module dynamically switches the operation mode through the configuration register; After the reconfigurable computing unit finishes processing the current frame, it sends an interrupt signal to the main control module and receives the configuration parameters of the next frame; Interrupt response delay time Satisfy: ; Wherein, is the signal processing delay time.

[0016] The present invention provides a high-precision liquid level measurement anti-interference calibration method. It has the following beneficial effects: 1. The present invention constructs a multi-physical field coupling relationship through a hypergraph adjacency tensor and a Laplace matrix, breaks through the limitations of the traditional binary correlation model, explicitly represents the high-order coupling effects of liquid level signals and multi-source interferences such as vibration, electromagnetism, and temperature, enhances the pertinence and accuracy of interference separation from the mechanism level, and solves the problem of signal aliasing in complex industrial scenarios.

[0017] 2. The present invention, through a dynamic parameter adjustment mechanism based on deep reinforcement learning, integrates signal quality, model complexity, and sensor health status into a unified optimization goal, and generates the optimal decomposition rank, filtering frequency, and self-healing current in real time.

[0018] 3. The present invention, through the hardware-level parallel design of a systolic array, a multiplier-accumulator array, and a finite difference engine, combined with a double-buffer data stream and a dynamic resource allocation strategy, realizes the efficient pipelined execution of three core algorithms: tensor decomposition, policy inference, and deposition control, meets the low-latency requirements of high-dimensional data stream processing, and overcomes the computing power bottleneck of traditional serial architectures.

[0019] 4. The present invention, through impedance recovery rate quantization evaluation and a multi-level fault response strategy, monitors the risks of sensor damage and algorithm failure in real time, combines redundant channel switching and online parameter calibration, realizes the fast self-diagnosis and self-recovery of the system's abnormal state, and greatly reduces the downtime probability caused by hardware aging or sudden interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] 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. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to the attached Figure 1, the embodiments of the present invention provide a high-precision liquid level measurement anti-interference calibration method, including: Step 1, synchronously collect liquid level signals, mechanical vibration signals, electromagnetic field intensity signals, and temperature distribution signals through a multi-modal sensor array to form multi-physical field data; Step 2, construct a hypergraph adjacency tensor based on the coupling relationship of the multi-physical field data, and generate a Laplacian matrix constraint; Step 3, perform CP tensor decomposition with hypergraph constraints on the original signal tensor formed by the multi-physical field data to separate the liquid level signal core and the interference core; Step 4, input the decomposed signal core features and sensor impedance parameters into the deep deterministic policy gradient algorithm to dynamically optimize the tensor decomposition rank, filter cut-off frequency, and self-healing trigger current; Step 5, generate deposition current parameters according to the optimization results, and repair the sensor sensitive layer through electrochemical deposition; Step 6, use a reconfigurable computing unit to parallelly execute tensor decomposition, policy inference, and self-healing control operations to achieve hardware acceleration of the processing flow.

[0023] The multi-physical field data in Step 1 includes the following types and corresponding sensors: Liquid level signal: collected by a fiber Bragg grating sensor; Mechanical vibration signal: collected by a three-axis MEMS accelerometer; Electromagnetic field intensity signal: collected by an electromagnetic field probe; Temperature distribution signal: collected by a distributed temperature sensor; among them, the synchronous collection of multi-physical field data is realized through a time synchronization protocol and satisfies: The constructed original signal tensor is a four-dimensional tensor , and the dimension definitions are as follows: : number of time slices; : number of spatial sampling points; : number of frequency sub-bands; : number of physical field types.

[0024] The construction of the hypergraph adjacency tensor in Step 2 satisfies the following conditions: Define a third-order adjacency tensor: ; where is the number of physical field types; the tensor element if and only if there is a coupling interference relationship between the th, th, and th types of physical fields; generate a hypergraph Laplacian matrix , and its expression is: ; Among them, is the hyperedge degree diagonal matrix, and its diagonal elements .

[0025] The objective function of CP tensor decomposition with hypergraph constraints in step three is: ; Among them: is the original signal tensor; , , are the time, space, and frequency factor matrices respectively; is the hypergraph Laplacian matrix; is the hypergraph constraint coefficient; represents the vector outer product, is the Frobenius norm; Liquid level signal kernel and interference kernel satisfy the separation: ; Among them, and are the component classification sets based on hypergraph constraints.

[0026] The calculation of the deposition current parameter in step four satisfies the following formula: ; Among them: is the deposition current density; , representing the deviation between the current impedance of the sensor sensitive layer and the reference impedance; is the initial impedance of the sensor; is the activation energy of the electrochemical deposition reaction; is the Boltzmann constant; is the absolute temperature on the surface of the sensitive layer; is the deposition rate proportionality coefficient; Electrochemical deposition transports the electrolyte to the surface of the sensitive layer through the microfluidic channel and applies a current density for a duration , until the impedance recovers to .

[0027] The dynamic optimization process of the deep deterministic policy gradient algorithm in step four satisfies the following conditions: State space: Signal-to-noise ratio SNR; Current tensor decomposition rank R; Impedance Z of the sensor sensitive layer; Action space: Tensor decomposition rank adjustment amount ΔR; Filter cut-off frequency fc; Self-healing trigger current Ih; Reward function: ; Where: : Number of self-healing triggers per unit time; : Signal processing delay time; is the weight coefficient.

[0028] The parallel execution of the reconfigurable computing unit in step six satisfies the following conditions: Supports three operation modes: Mode 1: Execute the alternating least squares algorithm for tensor decomposition to calculate matrix chain multiplication: ; Mode 2: Calculate the output of the fully connected layer of the policy network: ; Mode 3: Solve the partial differential equation of the deposition current: ; Adopt a double-buffered data flow mechanism: While the current frame of data is processed sequentially through Mode 1 to Mode 3, the next frame of data is pre-loaded into the cache through a high-speed interface, where: Current frame of data: Denote as ; Pre-loaded data: Denote as .

[0029] Step one also includes: The multi-physical field data after synchronous acquisition needs to be processed for cross-domain alignment, including: Perform phase compensation on the vibration signal and the electromagnetic signal, and the compensation formula is: ; Where, is the clock synchronization error, is the signal center frequency; Perform spatial interpolation on the temperature distribution signal to generate a temperature field matrix that matches the spatial resolution of the liquid level signal.

[0030] Step five also includes: After electrochemical deposition, verify the sensor repair effect through the following steps: Collect the impedance of the repaired sensor , and calculate the impedance recovery rate: ; If , it is determined that the repair is successful; otherwise, the deposition current parameters are regenerated for secondary repair.

[0031] Step six also includes: Its interaction with the main control module satisfies: The main control module dynamically switches the operation mode through the configuration register; After the reconfigurable computing unit completes the processing of the current frame, it sends an interrupt signal to the main control module and receives the configuration parameters of the next frame; Interrupt response delay time Satisfies: ; Among them, is the signal processing delay time.

[0032] The multi-modal sensor array consists of a liquid level sensing module, a mechanical vibration sensing module, an electromagnetic field intensity sensing module, and a temperature distribution sensing module. Among them, the liquid level sensing module uses a fiber Bragg grating sensor to invert the change in liquid level height by measuring the shift of the Bragg wavelength. Preferably, the fiber Bragg grating array is evenly distributed along the axis of the measured container to ensure that the spatial resolution ability meets the requirements of multi-physical field coupling analysis.

[0033] The mechanical vibration sensing module uses a three-axis MEMS accelerometer, and its sensitive axes are respectively aligned with the X / Y / Z axes of the container coordinate system to capture the three-dimensional vibration acceleration time-domain signal. The electromagnetic field intensity sensing module uses a broadband electromagnetic field probe. Preferably, a log-periodic antenna and a radio frequency detection circuit are integrated at the front end of the probe to realize the joint time-frequency domain characterization of broadband electromagnetic interference.

[0034] The temperature distribution sensing module uses a distributed temperature sensor array. Preferably, the sensors are arranged in a grid topology on the surface of the measured container to obtain the spatial gradient information of the temperature field through the resistance-temperature conversion relationship. The output signals of each sensor are transmitted to the main control unit after analog-to-digital conversion.

[0035] To achieve the spatio-temporal alignment of the four types of physical field data, the main control module sends a global clock signal to each sensor based on the time synchronization protocol. Preferably, the IEEE1588 precise time protocol is used for clock synchronization. The main control module serves as the master clock source, and each sensor serves as a slave clock node. The transmission delay is eliminated through periodic clock correction messages, so that the time deviation at the sampling moment of each sensor is less than the preset threshold.

[0036] During the synchronous acquisition process, the master control module sends trigger pulses to the sensor array, and each sensor starts data acquisition based on the same time reference. Preferably, the rising edge of the trigger pulse is aligned with the whole-second boundary of the system clock to reduce the timing error introduced by clock jitter. After the acquisition is completed, the data streams of each sensor are transmitted in parallel to the preprocessing unit through a high-speed interface.

[0037] For the liquid level signal, since the Bragg wavelength of the fiber grating is easily affected by temperature, temperature drift compensation is required. The compensation formula is: ; where is the original wavelength offset, is the thermo-optic coefficient of the fiber material, is the current temperature measurement value, is the reference temperature. Preferably, the ambient temperature at the initial calibration of the sensor is taken. For the mechanical vibration signal, it is necessary to suppress the interference of high-frequency noise on the analysis of low-frequency liquid level fluctuations. Preferably, a Butterworth low-pass filter is used to perform frequency-domain filtering on the vibration signal, and its transfer function is: ; where is the cut-off angular frequency, is the filter order, and the time-domain waveform of the filtered vibration signal is used to construct the hypergraph coupling relationship. For the electric field strength signal, it is necessary to eliminate the power frequency and harmonic interference. Preferably, an adaptive notch filter is used to suppress the interference at specific frequency points, and its weight update formula is: ; where is the filter coefficient vector, is the convergence factor, is the error signal, is the input signal vector. For the temperature distribution signal, it is necessary to achieve spatial resolution matching with the liquid level signal through spatial interpolation. Preferably, the Kriging interpolation algorithm is used to construct a continuous temperature field distribution model, and its interpolation weight calculation formula is: ; where is the covariance matrix between temperature measurement points, is the covariance vector between the point to be interpolated and the measurement points, is the Lagrange multiplier.

[0038] The preprocessed multi-physical field data is formatted and encapsulated according to a unified timestamp, spatial coordinates, and physical quantity units. Preferably, the time series is divided into slices of a fixed length (e.g., each slice contains 1000 sampling points), and the data within each slice is organized as a fourth-order tensor in four dimensions: time, space, frequency, and physical field type. , serving as the input data for subsequent hypergraph modeling.

[0039] Organize the preprocessed multi-physical field data in step one into a four-dimensional raw signal tensor according to the preset dimension rules. . Among them, the time dimension represents the number of time slices for data acquisition. Preferably, each time slice corresponds to continuous sampling data of a fixed duration. For example, time windows are divided at intervals of 1 second. The spatial dimension corresponds to the number of spatial sampling points of the sensor array. Preferably, it is consistent with the axial distribution density of the fiber Bragg grating sensors to ensure the spatial resolution of the liquid level signal.

[0040] The frequency dimension represents the number of frequency sub-bands divided after the signal undergoes time-frequency transformation (such as wavelet transform or short-time Fourier transform). Preferably, the frequency sub-bands are evenly divided on a logarithmic scale to cover the frequency-domain coupling characteristics of multi-physical fields. The physical field type dimension is fixed at 4, corresponding to four types of physical fields: liquid level, vibration, electromagnetic, and temperature, and its order is strictly consistent with the acquisition channels of the sensor module.

[0041] Generation of the hypergraph adjacency tensor and determination of coupling relationships: To characterize the high-order coupling relationships between multi-physical fields, a third-order adjacency tensor is defined, and the assignment rule for its element is as follows: if and only if there is a coupling interference between the , , types of physical fields, , otherwise it is 0. Preferably, the determination basis for coupling interference is the correlation analysis result of the preprocessed signal in step one. For example, the mutual information or coherence function is calculated to quantify the field-to-field correlation strength.

[0042] Specifically, there is a mechanical coupling between the liquid level signal (type 1) and the vibration signal (type 2) in the low-frequency band, an electro-hydraulic coupling at specific frequency points between the liquid level signal and the electromagnetic signal (type 3), and a thermal expansion effect coupling between the liquid level signal and the temperature signal (type 4). Therefore, the following non-zero elements need to be explicitly marked in the adjacency tensor: ; The remaining elements are default set to 0. This definition ensures that the hypergraph can capture the multiple coupling patterns of the liquid level signal with at least two types of interfering fields.

[0043] Based on the adjacency tensor Generate the hypergraph Laplacian matrix , and its construction process includes the following steps: First, calculate the hyperedge degree diagonal matrix , whose diagonal elements represent the number of hyperedges participated by the th type of physical field, and the calculation formula is: ; Subsequently, the Laplacian matrix is generated by the following formula: ; where, represents expanding the third-order adjacency tensor into a matrix form according to the first dimension. Preferably, the expansion method adopts mode-1 expansion, that is, the first dimension (physical field type) of the tensor is used as the row index, and the remaining two dimensions are combined as the column index. Correlation design between hypergraph constraints and tensor decomposition: The generated Laplacian matrix will be embedded as a regularization term into the objective function of subsequent tensor decomposition to constrain the sparsity of the factor matrix and the consistency of the physical field coupling relationship. Preferably, the introduction of the Laplacian matrix makes the decomposed liquid level signal kernel far from the interference kernel in the factor space, thereby enhancing the robustness of signal separation.

[0044] Specifically, the Laplacian matrix forces the physical fields belonging to the same coupling relationship to have similar time evolution patterns by penalizing the local non-smoothness of the factor matrix (time dimension factor). For example, the coupling relationship between the liquid level and the vibration signal will be reflected as an increase in the covariance of the corresponding columns in , while the columns corresponding to non-coupled fields remain independent.

[0045] Based on the hypergraph Laplacian matrix generated in the second step, construct the CP tensor decomposition objective function with hypergraph constraints. The objective function consists of a data fitting term and a regularization term, and its mathematical expression is: ; where, the data fitting term measures the reconstruction error between the original signal tensor and the decomposition result, and the regularization term restricts the smoothness of the time factor matrix through hypergraph constraints, forcing the coupled physical fields to have a consistent evolution pattern in the time dimension. The parameter To balance the weights of two items, preferably, its value is determined by cross-validation to ensure the balance between interference suppression and signal fidelity in the decomposition result. Alternating least squares solution of the factor matrix: The alternating least squares method (ALS) is used to iteratively optimize the objective function. Specifically, in each iteration, two factor matrices are fixed and the third factor matrix is updated until convergence. Taking the update of the time factor matrix as an example, its closed-form solution is: ; where represents the mode-1 unfolding matrix of the tensor along the time dimension, represents the Khatri-Rao product, represents the Hadamard product. Preferably, the initial factor matrix is generated by random orthonormalization to avoid local optima.

[0046] Set the residual change rate threshold as the iteration termination condition. Calculate the relative change of the reconstruction error after each iteration: ; When is less than the preset threshold (e.g., ), it is determined to converge and the iteration is terminated. Preferably, to prevent overfitting, the Frobenius norm growth of the factor matrix is monitored simultaneously. If it exceeds twice the initial value, the early stopping mechanism is triggered. Separation rule for the liquid level signal kernel and the interference kernel: Classify the decomposed components according to the hypergraph constraint to generate the liquid level signal kernel and the interference kernel . Define the classification sets and , which satisfy and . Specifically, the basis for component classification is: if the factor vector is highly correlated with the coupling mode of the hypergraph constraint (e.g., its projection energy with exceeds the threshold), then ; otherwise . The mathematical expressions for the liquid level signal kernel and the interference kernel are: ; Preferably, the determination of the classification set can be corrected by combining prior knowledge (such as the frequency band range of the liquid level signal) to improve the separation accuracy.

[0047] The state space consists of three key parameters: signal quality, model complexity, and sensor health status. The signal quality is quantified by the signal-to-noise ratio (SNR), and its calculation method is the liquid level signal kernel and the interference kernel Frobenius norm ratio: ; This index reflects the effectiveness of current signal separation. The model complexity is characterized by the rank of tensor decomposition, and its value directly affects the risks of overfitting and underfitting. The health status of the sensor is monitored through the impedance of the sensitive layer. Preferably, the impedance value is obtained by the four-wire measurement method to eliminate the error introduced by contact resistance. Definition of decision variables in the action space: The action space contains three adjustable parameters: the adjustment amount of tensor decomposition rank , the filter cut-off frequency , and the self-healing trigger current . Among them, is used to dynamically adjust the model complexity. Preferably, its value range is , corresponding to the decreasing, maintaining, and increasing operations of the rank respectively. The filter cut-off frequency realizes high-frequency interference suppression through a Butterworth filter, and its value is dynamically optimized according to the time-frequency characteristics of electromagnetic interference. The self-healing trigger current is used to control the initial current of electrochemical deposition. Preferably, its value is positively correlated with the sensor impedance deviation . Multi-objective trade-off design of the reward function: The reward function is used to guide the Deep Deterministic Policy Gradient (DDPG) algorithm to converge to the optimal policy, and its expression is: ; Among them, is the number of self-healing triggers per unit time, which is used to constrain the sensor loss caused by excessive repair; is the signal processing delay time, which is determined by the parallel processing efficiency of the reconfigurable computing unit; is the weight coefficient, satisfying . Preferably, the weight coefficient is dynamically adjusted according to the real-time working conditions. For example, in a strong interference environment, is preferentially increased to ensure signal quality.

[0048] According to the action parameters output by the policy network and the sensor impedance deviation , the deposition current density is generated, and its calculation formula is: ; Among them, is the deposition rate proportionality coefficient, is the activation energy of the electrochemical deposition reaction, is the surface temperature of the sensitive layer, is the Boltzmann constant. Preferably, the activation energy Calibrated by the Arrhenius equation to ensure the adaptive matching of the deposition rate and temperature. Closed-loop execution and verification of electrochemical deposition: After the deposition current parameters are generated, the electrolyte is delivered to the surface of the sensitive layer through the microfluidic channel, and the current density is applied for a preset duration . Preferably, the hydrodynamic characteristics of the microfluidic channel are optimized by the Navier-Stokes equation to ensure that the electrolyte uniformly covers the damaged area. After the deposition is completed, the impedance of the repaired area is collected , and the impedance recovery rate is calculated as follows: ; If , it is determined that the repair is successful and the reference impedance is updated ; otherwise, the deposition parameters are regenerated based on the current impedance deviation for iterative repair.

[0049] The DDPG algorithm includes an Actor network and a Critic network. The Actor network generates action parameters through the policy gradient method, and the Critic network evaluates the action value through the temporal difference error. Preferably, the network parameters are updated using experience replay and target network techniques to improve the training stability. During the online learning process, the real-time collected state-action-reward tuples are stored in the experience pool, and when the data volume reaches the threshold, batch gradient update is triggered The reconfigurable computing unit supports three operation modes, corresponding to the tensor decomposition, policy inference, and deposition control core algorithms defined in claim 7 respectively. The mode switching is completed through the configuration register of the main control module. Preferably, the register bit width corresponds one-to-one with the operation mode encoding. For example, "00" represents mode one, "01" represents mode two, and "10" represents mode three.

[0050] In mode one, the alternating least squares algorithm with hypergraph constraints is executed. The computing unit deploys a systolic array structure for efficient calculation of matrix chain multiplication and regularization terms. Preferably, the row and column scales of the systolic array match the dimensions of the factor matrices to maximize the data reuse rate. The matrix inversion operation is implemented by a coprocessor based on Cholesky decomposition, and its input data path is directly interconnected with the output of the systolic array to reduce the intermediate data transfer delay.

[0051] In mode two, the fully connected layer calculation of the policy network is executed. The computing unit is configured as a parallel multiplier-accumulator array, and the weight matrix of each layer of the neural network is pre-loaded into the on-chip memory. Preferably, the activation function σ(·) is implemented by a piecewise linear approximation circuit to reduce the hardware overhead of non-linear operations. The inter-layer data transfer adopts the ping-pong buffer mechanism. When the calculation result of the current layer is written into buffer A, the next layer reads data from buffer B to ensure the uninterrupted calculation pipeline.

[0052] In Mode 2, the fully connected layer of the execution policy network is calculated. The computing unit is configured as a parallel multiplier-accumulator array, and the weight matrix of each layer of the neural network is pre-loaded into the on-chip memory. Preferably, the activation function is implemented using a piecewise linear approximation circuit to reduce the hardware overhead of non-linear operations. The inter-layer data transfer adopts a ping-pong buffering mechanism. When the calculation result of the current layer is written into buffer A, the next layer reads data from buffer B to ensure the uninterrupted calculation pipeline.

[0053] In Mode 3, the numerical solution of the deposition current partial differential equation is performed. The computing unit is configured as a finite difference engine, and the spatial discretization grid matches the surface topography of the sensitive layer. Preferably, the diffusion term is discretized using a central difference format, and the reaction term realizes non-linear calculation through a look-up table method to balance accuracy and calculation efficiency.

[0054] Implementation details of the double-buffer data stream mechanism: The double-buffer mechanism realizes real-time processing through a parallel data path and memory bank switching. During the processing of the current frame of data the next frame of data is pre-loaded into the spare memory bank through a high-speed interface. Preferably, the memory bank switching is controlled by a crossbar switch matrix. When the processing of the current frame is completed, the data path is immediately switched to the pre-loaded spare memory bank, and at the same time, the original memory bank is cleared and the loading of the next frame of data is started.

[0055] The dimension of the data frame is strictly consistent with the original signal tensor defined in Step 2 i.e., . Preferably, the time slice length is aligned with the preprocessing segments in Step 1 to avoid boundary effects introduced by cross-frame data stitching.

[0056] Interrupt response and master-slave interaction protocol: After the reconfigurable computing unit completes the processing of the current frame, it sends an interrupt request signal to the main control module. The main control module responds to the interrupt within a preset time. Preferably, the interrupt response delay satisfies where is the signal processing delay time defined in Step 4. The interrupt service routine performs the following operations: Reads the processing results from the reconfigurable unit (such as the decomposed factor matrix, the output value of the policy network, the deposition current distribution); Writes the configuration parameters of the next frame (such as the decomposition rank , the filter cut-off frequency ) into the configuration register; Triggers the memory bank switching instruction to start a new round of data processing.

[0057] Dynamically allocate the resource ratio of computing units according to the real-time load conditions. Preferably, a greedy scheduling algorithm is adopted: when the signal-to-noise ratio (SNR) of the liquid level signal is lower than the threshold, more resources are preferentially allocated to Mode 1 (tensor decomposition); when the sensor impedance deviation exceeds the threshold, resources are preferentially allocated to Mode 3 (deposition control). The resource allocation ratio is implemented through time division slice parameters in the hardware description language. For example, the clock cycle is divided into three segments and allocated to the three operation modes respectively.

[0058] Monitor the abnormal fluctuations of the sensor impedance Z and the tensor decomposition residual ε = ∥X L N∥F in real time. When any of the following conditions is detected, it is determined that the system has a fault and the fault tolerance mechanism is triggered: Impedance mutation: |ΔZ / Z0| > 20%, indicating irreversible damage to the sensitive layer of the sensor or electrolyte leakage; Residual decomposition overrun: ε > 0.1, indicating that tensor decomposition fails or the interference kernel is not fully separated.

[0059] The fault tolerance recovery strategy includes two-level responses: First-level response: Forcefully reset the reconfigurable computing unit, clear the current data cache, and resume the operation from the previous valid state; Second-level response: If the fault still persists after the first-level response, switch to the redundant sensor channel and re-initialize the tensor decomposition parameters (such as resetting the decomposition rank R to the default value). Preferably, the switching of the redundant channel is implemented through a multiplexer, and its control signal is output from the GPIO pin of the main control module.

[0060] During the system operation, dynamically calibrate the policy network parameters of the DDPG algorithm in Step 4 according to the real-time working conditions. Preferably, an online incremental learning strategy is adopted: when it is monitored that the sensor impedance recovery rate ηZ continuously remains lower than the threshold, trigger the gradient update of the policy network parameters. During the update process, freeze the action output of the Actor network and only update the Q-value estimation part of the Critic network to avoid policy oscillation. The learning rate is adaptively adjusted according to the sliding average value of the historical rewards, and its adjustment formula is: αnew = αbase · exp (−β · ∑N−1 ri), where αbase is the base learning rate, β is the attenuation coefficient, and N is the window length of the historical rewards.

[0061] Dynamically adjust the supply voltage and clock frequency of the reconfigurable computing unit according to the operation load. Preferably, the dynamic voltage and frequency scaling (DVFS) technology is adopted: when the system is in a low-load state (such as SNR > 40 dB and ΔZ < 5%), reduce the FPGA core voltage and clock frequency to save power consumption; when a high-load task is detected (such as self-healing trigger or residual overrun), instantaneously increase the supply voltage to the nominal value to ensure computing performance. The voltage regulation is implemented through a PMIC (Power Management Integrated Circuit), and its control signal is sent by the I2C interface of the main control module.

[0062] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A high-precision liquid level measurement anti-interference calibration method, characterized in that, Including: Step 1: Synchronously collect liquid level signals, mechanical vibration signals, electromagnetic field intensity signals, and temperature distribution signals through a multi-modal sensor array to form multi-physical field data; Step 2: Construct a hypergraph adjacency tensor based on the coupling relationship of the multi-physical field data and generate a Laplacian matrix constraint; Step 3: Perform CP tensor decomposition with hypergraph constraints on the original signal tensor formed by the multi-physical field data to separate the liquid level signal core and the interference core; Step 4: Input the decomposed signal core features and sensor impedance parameters into the deep deterministic policy gradient algorithm to dynamically optimize the tensor decomposition rank, filter cut-off frequency, and self-healing trigger current; Step 5: Generate deposition current parameters according to the optimization results and repair the sensor sensitive layer through electrochemical deposition; Step 6: Use a reconfigurable computing unit to parallelly execute tensor decomposition, policy inference, and self-healing control operations to achieve hardware acceleration of the processing flow.

2. The high-precision liquid level measurement anti-interference calibration method according to claim 1, characterized in that The multi-physical field data in Step 1 includes the following types and corresponding sensors: Liquid level signal: Collected by a fiber Bragg grating sensor; Mechanical vibration signal: Collected by a three-axis MEMS accelerometer; Electromagnetic field intensity signal: Collected by an electromagnetic field probe; Temperature distribution signal: Collected by a distributed temperature sensor; where the synchronous collection of the multi-physical field data is achieved through a time synchronization protocol and satisfies: The constructed original signal tensor is a four-dimensional tensor , and the dimensions are defined as follows: : Number of time slices; : Number of spatial sampling points; : Number of frequency sub-bands; : Number of physical field types.

3. The high-precision liquid level measurement anti-interference calibration method according to claim 1, characterized in that, The construction of the hypergraph adjacency tensor in Step 2 satisfies the following conditions: Define a third-order adjacency tensor: ； wherein is the number of physical field types; tensor elements if and only if there is a coupling interference relationship between the -th, -th and -th physical fields; generate a hypergraph Laplacian matrix based on the adjacency tensor , and its expression is: ； Among them, is the hyperedge degree diagonal matrix, and its diagonal elements .

4. The anti-interference calibration method for high-precision liquid level measurement according to claim 1, characterized in that, The objective function of the CP tensor decomposition with hypergraph constraints in Step 3 is: ； Wherein: is the original signal tensor; , , are the time, space and frequency factor matrices respectively; is the hypergraph Laplacian matrix; is the hypergraph constraint coefficient; Denotes the vector outer product, is the Frobenius norm; The liquid level signal core and the interference core are separated to satisfy: ； Among them, and are sets of component classifications based on hypergraph constraints.

5. The high-precision liquid level measurement anti-interference calibration method according to claim 1, characterized in that The calculation of the deposition current parameters in Step 4 satisfies the following formula: ； Wherein: is the deposition current density; , represents the deviation of the current impedance of the sensor sensitive layer from the reference impedance; is the initial impedance of the sensor; is the activation energy of the electrochemical deposition reaction; is the Boltzmann constant; is the absolute temperature on the surface of the sensitive layer; is the deposition rate proportionality coefficient; The electrochemical deposition transports the electrolyte to the surface of the sensitive layer through a microfluidic channel and applies a current density for a duration of , until the impedance recovers to .

6. The anti-interference calibration method for high-precision liquid level measurement according to claim 1, characterized in that, The dynamic optimization process of the deep deterministic policy gradient algorithm in Step 4 satisfies the following conditions: State space: Signal-to-noise ratio SNR; Current tensor decomposition rank R; Sensor sensitive layer impedance Z; Action space: Tensor decomposition rank adjustment amount ΔR; Filter cut-off frequency fc; Self-healing trigger current Ih; Reward function: ； Wherein: : The number of self-healing trigger times per unit time; : Signal processing delay time; is the weight coefficient.

7. The anti-interference calibration method for high-precision liquid level measurement according to claim 1, characterized in that The parallel execution of the reconfigurable computing unit in Step 6 satisfies the following conditions: Supports three operation modes: Mode 1: Execute the alternating least squares algorithm for tensor decomposition and calculate matrix chain multiplication: ； Mode 2: Calculate the output of the fully connected layer of the policy network; ； Mode 3: Solve the partial differential equation of the deposition current; ； Adopt a double-buffer data flow mechanism: While the current frame of data is processed sequentially through Mode 1 to Mode 3, the next frame of data is pre-loaded into the cache through a high-speed interface, where: Current frame data: denoted as ; Preloaded data: denoted as .

8. The high-precision liquid level measurement anti-interference calibration method according to claim 1, wherein Step 1 further includes: The multi-physical field data after synchronous collection needs to be processed for cross-domain alignment, including: Perform phase compensation on the vibration signal and the electromagnetic signal, and the compensation formula is: ； Among them, is the clock synchronization error, is the signal center frequency; Perform spatial interpolation on the temperature distribution signal to generate a temperature field matrix that matches the spatial resolution of the liquid level signal.

9. The high-precision liquid level measurement anti-interference calibration method according to claim 1, characterized in that Step 5 further includes: After the electrochemical deposition is completed, verify the sensor repair effect through the following steps: Collect the impedance of the repaired sensor , and calculate the impedance recovery rate: ； If , it is determined that the repair is successful; otherwise, the deposition current parameters are regenerated for secondary repair.

10. The high-precision liquid level measurement anti-interference calibration method according to claim 1, characterized in that, Step 6 further includes: Its interaction with the main control module satisfies: The main control module dynamically switches the operation mode through the configuration register; After the reconfigurable computing unit completes the processing of the current frame, it sends an interrupt signal to the main control module and receives the configuration parameters of the next frame; Interrupt response latency time Satisfy: ； Among them, is the signal processing delay time.

Images