Electromagnetic ultrasonic thickness detection method and system based on miniaturized low-power-consumption Internet of Things

Abstract

The invention discloses an electromagnetic ultrasonic thickness detection method and system based on a miniaturized low-power-consumption internet of things, and the method comprises the steps: carrying out the parameter optimization processing through a pulse coding excitation algorithm according to the electromagnetic characteristic parameters of a target material and a preset detection precision requirement, and outputting a low-power-consumption excitation parameter set; driving an electromagnetic ultrasonic transducer to generate a detection signal based on the low-power-consumption excitation parameter set, and outputting a compressed echo signal; performing multi-scale intrinsic mode decomposition on the compressed echo signal, and outputting a thickness correlation characteristic parameter set; inputting the thickness correlation characteristic parameter set into a depth residual error-attention network model, and outputting a thickness prediction result; and performing multi-node data fusion on the thickness prediction result based on an Internet of Things cloud platform to generate a final thickness detection report. According to the embodiment of the invention, high-precision and traceable thickness detection can be realized based on the miniaturized low-power-consumption Internet of Things.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things, and in particular to an electromagnetic ultrasonic thickness detection method and system based on a miniaturized low-power Internet of Things. Background Art

[0002] In the fields of industrial equipment monitoring, pipeline maintenance, and material testing, thickness measurement is a crucial tool for assessing structural integrity and safety. Traditional ultrasonic thickness testing methods typically rely on high-power excitation equipment and complex signal processing systems. These methods suffer from high power consumption, bulk, and poor portability, making them difficult to meet the distributed monitoring needs of the Internet of Things (IoT). While electromagnetic ultrasonic technology (EMAT) is gaining attention due to its non-contact detection advantages, existing EMAT systems still face challenges such as low excitation efficiency, insufficient signal-to-noise ratio, and difficulty in multi-node data coordination. Balancing detection accuracy and energy consumption is a technical challenge, particularly in miniaturized, low-power scenarios. Furthermore, traditional thickness testing methods are poorly adaptable to the electromagnetic properties of materials and lack effective multi-node data fusion and trusted storage mechanisms, resulting in insufficient traceability and reliability of test results. Summary of the Invention

[0003] The purpose of the present invention is to provide an electromagnetic ultrasonic thickness detection method and system based on a miniaturized low-power Internet of Things to address the deficiencies in the prior art and to achieve high-precision and traceable thickness detection based on a miniaturized low-power Internet of Things.

[0004] One embodiment of the present application provides an electromagnetic ultrasonic thickness detection method based on a miniaturized, low-power Internet of Things, the method comprising: According to the electromagnetic properties of the target material and the preset detection accuracy requirements, the pulse coding excitation algorithm is used to optimize the parameters. The pulse duty cycle and carrier frequency are adaptively adjusted to match the material impedance characteristics, and a low-power excitation parameter set is output. Based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal, the reflected echo signal is synchronously collected through the edge computing node, and a time-frequency joint compression algorithm is used to perform dynamic noise reduction and sparse processing, and a compressed echo signal is output; Performing multi-scale eigenmode decomposition on the compressed echo signal to extract characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; Input the thickness-related feature parameter set into the deep residual-attention network model, combine it with the historical data in the material sound velocity database, calculate the dynamic thickness value of the target material, and output the thickness prediction result; Based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction results to generate a final thickness detection report, which is synchronized to the visual interface of the terminal device for three-dimensional thickness distribution display and stored in the encrypted blockchain node to realize the traceability of the detection data.

[0005] Optionally, the method performs parameter optimization processing by a pulse coding excitation algorithm based on the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, and outputs a low-power excitation parameter set by adaptively adjusting the pulse duty cycle and carrier frequency to match the material impedance characteristics, including: According to the electromagnetic permeability and hysteresis loss parameters of the target material, a material impedance spectrum model is constructed to generate a frequency-impedance mapping table; Based on the preset detection accuracy requirements, a genetic algorithm is used to screen the initial pulse parameter combinations that meet the signal-to-noise ratio threshold in the frequency-impedance mapping table to generate a candidate set of duty cycle and carrier frequency; The dynamic impedance matching circuit monitors the surface impedance changes of the material in real time, and uses a closed-loop feedback mechanism to adjust the pulse duty cycle to dynamically adapt the emission energy to the material absorption characteristics. The carrier frequency is optimized by combining the gradient descent method, the effective bandwidth of the excitation signal is maximized under the power consumption constraint, and the optimized low-power excitation parameter set is output.

[0006] Optionally, the driving of the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously collecting the reflected echo signal through the edge computing node, and performing dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm to output a compressed echo signal include: Inputting the excitation parameter set into a programmable pulse generator to generate a modulated electromagnetic ultrasonic excitation waveform to drive the transducer to emit a detection sound wave; The reflected echo signal is captured by the synchronous acquisition module of the edge node, and the direct wave and multipath interference components are separated by sliding window short-time Fourier transform. Based on the instantaneous energy distribution of the signal, a dynamic threshold filter is designed to filter out the environmental electromagnetic noise in real time and obtain the noise-reduced echo signal; Perform wavelet packet decomposition on the denoised echo signal to extract the characteristic frequency band related to the material thickness, and generate compressed time-frequency coefficients through an adaptive quantization algorithm; The compressed sensing algorithm is used to reconstruct the sparse signal and output a compressed echo signal that meets the transmission bandwidth of the Internet of Things.

[0007] Optionally, performing multi-scale intrinsic mode decomposition on the compressed echo signal to extract characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set includes: Adaptive noise-assisted ensemble empirical mode decomposition algorithm is used to decompose the compressed echo signal into six eigenmode components to eliminate the modal aliasing effect; Calculate the energy entropy and kurtosis value of each eigenmode component, and select the components with energy entropy less than 0.5 and kurtosis greater than 3 as thickness-sensitive candidate modes; Perform Hilbert transform on the thickness-sensitive candidate modes to extract the instantaneous phase features and construct the phase-energy joint distribution matrix; The phase-energy joint distribution matrix is ​​input into the kernel principal component analysis algorithm to fuse multimodal features and reduce dimensionality, and output a set of thickness-related feature parameters after dimensionality reduction.

[0008] Optionally, the step of inputting the thickness-related feature parameter set into a deep residual-attention network model, combining historical data in a material sound velocity database, calculating a dynamic thickness value of the target material, and outputting a thickness prediction result includes: The thickness-related feature parameter set is input into the deep residual network, and the weight of each feature is dynamically assigned through the multi-head attention mechanism to generate the initial thickness feature vector; Query the temperature-sound velocity curve of the same type of material in the material sound velocity database, and use cubic spline interpolation to compensate for the sound velocity correction parameters caused by ambient temperature; Design a cross-modal feature fusion layer to perform tensor concatenation of the sound velocity correction parameters and the initial thickness feature vector to generate a fused feature representation; A Bayesian probability module is integrated in the output layer to calculate the confidence interval of the thickness prediction value based on the fused feature representation; The sliding window weighted average algorithm is used to eliminate single measurement noise, and the dynamic thickness change curve is optimized in combination with historical prediction data to output the final thickness prediction result.

[0009] Optionally, the thickness prediction result is subjected to multi-node data fusion based on the IoT cloud platform to generate a final thickness detection report, which is synchronized to a visual interface of a terminal device for three-dimensional thickness distribution display, and stored in an encrypted blockchain node to achieve detection data traceability, including: The thickness prediction results of multiple edge nodes in the IoT cloud platform are aligned through a spatiotemporal registration algorithm, and the Dempster-Shafer evidence theory is used to eliminate spatial measurement conflicts and generate consistent fusion results. The consistency fusion results are input into the 3D surface reconstruction engine, and the iso-thickness line distribution map and thermal map are generated by combining the geometric features of the material surface; Generate the final thickness test report based on the iso-thickness line distribution map and heat map. Use the lightweight blockchain framework to construct a Merkle tree hash for the report and raw data, and store it on the chain through smart contracts. An interactive traceability module is deployed in the visual interface to associate the three-dimensional isopach distribution map with the thickness detection report stored in the blockchain. It supports clicking on any thickness point to query the corresponding original echo signal and processing log.

[0010] Optionally, aligning thickness prediction results of multiple edge nodes in the IoT cloud platform through a spatiotemporal registration algorithm, using Dempster-Shafer evidence theory to eliminate spatial measurement conflicts, and generating consistent fusion results, includes: Based on the time and space stamp information and geographic location coordinates of each edge node, a sliding window dynamic time warping algorithm is used to align the time series of multi-node thickness prediction results, eliminate the timing offset caused by sampling frequency differences, and generate a time-synchronized thickness data matrix. Based on the mapping relationship of the three-dimensional coordinate system of the material surface, the time-synchronized thickness data matrix is ​​input into the spatial registration model, the spatial position deviation between nodes is corrected through the affine transformation matrix, and the spatially aligned multi-node thickness distribution tensor is output; Uncertainty modeling is performed on the spatially aligned multi-node thickness distribution tensor. A trust function is generated based on the historical detection error distribution of each node. Combined with the basic probability distribution rule in the Dempster-Shafer theory, the confidence interval and conflict coefficient matrix of each thickness value are calculated. The conflict coefficient matrix is ​​used to construct the evidence fusion weight table, and the orthogonal sum formula is used to perform joint probability distribution on multi-node evidence to eliminate spatial measurement conflicts. The thickness probability distribution map after consistency fusion is output as the consistency fusion result.

[0011] Another embodiment of the present application provides an electromagnetic ultrasonic thickness detection system based on a miniaturized, low-power Internet of Things, the system comprising: A processing module is used to optimize parameters using a pulse coding excitation algorithm based on the electromagnetic properties of the target material and preset detection accuracy requirements, adaptively adjust the pulse duty cycle and carrier frequency to match the material impedance characteristics, and output a low-power excitation parameter set; an acquisition module, configured to drive the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously acquire the reflected echo signal through the edge computing node, perform dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm, and output a compressed echo signal; a decomposition module for performing multi-scale intrinsic mode decomposition on the compressed echo signal, extracting characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; A prediction module is used to input the thickness-related feature parameter set into a deep residual-attention network model, combine it with historical data in a material sound velocity database, calculate the dynamic thickness value of the target material, and output a thickness prediction result; The fusion module is used to perform multi-node data fusion on the thickness prediction results based on the Internet of Things cloud platform, generate a final thickness detection report, synchronize it to the visual interface of the terminal device for three-dimensional thickness distribution display, and store it in the encrypted blockchain node to realize the traceability of the detection data.

[0012] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0013] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0014] Compared with the prior art, the present invention provides an electromagnetic ultrasonic thickness detection method based on a miniaturized, low-power Internet of Things. According to the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, the pulse coding excitation algorithm is used to perform parameter optimization processing to output a low-power excitation parameter set; based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal and output a compressed echo signal; the compressed echo signal is subjected to multi-scale intrinsic mode decomposition to output a thickness-related feature parameter set; the thickness-related feature parameter set is input into a deep residual-attention network model to output a thickness prediction result; and multi-node data fusion is performed on the thickness prediction result based on the Internet of Things cloud platform to generate a final thickness detection report, thereby enabling high-precision and traceable thickness detection based on a miniaturized, low-power Internet of Things. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A hardware structure block diagram of a computer terminal for a miniaturized, low-power Internet of Things-based electromagnetic ultrasonic thickness detection method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a flow chart of an electromagnetic ultrasonic thickness detection method based on a miniaturized, low-power Internet of Things provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of a miniaturized, low-power Internet of Things-based electromagnetic ultrasonic thickness detection system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0017] The embodiment of the present invention first provides an electromagnetic ultrasonic thickness detection method based on a miniaturized low-power Internet of Things. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0018] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a miniaturized, low-power Internet of Things electromagnetic ultrasonic thickness detection method provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0019] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, cause a processor to perform any one of the electromagnetic ultrasonic thickness detection methods based on a miniaturized, low-power Internet of Things.

[0020] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0021] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any electromagnetic ultrasonic thickness detection method based on the miniaturized low-power Internet of Things.

[0022] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0023] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0024] See also Figure 2 , an embodiment of the present invention provides an electromagnetic ultrasonic thickness detection method based on a miniaturized low-power Internet of Things, which may include the following steps: S201, performing parameter optimization processing using a pulse coding excitation algorithm based on the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, adaptively adjusting the pulse duty cycle and carrier frequency to match the material impedance characteristics, and outputting a low-power excitation parameter set; Specifically, a material impedance spectrum model can be constructed based on the electromagnetic permeability and hysteresis loss parameters of the target material to generate a frequency-impedance mapping table; The construction of a material impedance spectrum model is based on a combination of experimental measurements and numerical analysis. First, a vector network analyzer (VNA, such as the Keysight PNA-L series) is used to perform a frequency sweep test on the target material (such as stainless steel or aluminum alloy), covering the effective frequency range of electromagnetic ultrasonic testing (e.g., 100kHz to 10MHz). During the test, the material sample is placed in the detection area of ​​the electromagnetic ultrasonic transducer, and scattering parameters (S parameters) are measured to obtain complex impedance data, including the real part (resistance component, in Ω) and the imaginary part (reactance component, in Ω). For example, the real part of the impedance of a certain aluminum alloy measured at a frequency of 500kHz is 45Ω, and the imaginary part is 110Ω.

[0025] To quantify hysteresis loss parameters, a BH curve analyzer (such as the Magnet-Physik EP-10) is used to measure the material's hysteresis loop and calculate the hysteresis loss power per unit volume (unit: watts per cubic meter, W / m³). For example, the hysteresis loss of a certain type of mild steel at a magnetic field strength of 1 Tesla is 120 W / m³. Combining electromagnetic permeability (measured using the four-probe method, units: Siemens per meter, S / m) with the hysteresis loss parameters, a frequency-domain impedance analytical model of the material is constructed.

[0026] Frequency-impedance mapping table generation process: Discrete frequency points: with a step size of 10kHz, a total of 1000 frequency points from 100kHz to 10MHz; Impedance calculation: Fit the impedance curve with frequency using experimental data. For example, at 1MHz, the real part of the impedance is 50Ω and the imaginary part is 120Ω. Data storage: The impedance value (real part and imaginary part) corresponding to each frequency point is stored in a table to form a frequency-impedance mapping table.

[0027] The example table fragment is shown in Table 1: Table 1 Frequency (MHz) Real part of impedance (Ω) Imaginary impedance (Ω) 0.1 30 80 0.5 45 110 1.0 50 120 Based on the preset detection accuracy requirements, a genetic algorithm is used to screen the initial pulse parameter combinations that meet the signal-to-noise ratio threshold in the frequency-impedance mapping table to generate a candidate set of duty cycle and carrier frequency; A genetic algorithm (GA) is used to search for the optimal pulse parameter combination (duty cycle D, unit: %; carrier frequency f, unit: MHz) in the frequency-impedance mapping table. The goal is to minimize power consumption while meeting detection accuracy (signal-to-noise ratio SNR ≥ 20 dB).

[0028] Algorithm flow: Chromosome encoding: Each chromosome consists of two genes (D, f), where D ranges from 10% to 90% and f ranges from 0.1 to 10 MHz; Initial population: randomly generate 50 parameter combinations, such as (D=30%, f=2MHz); Fitness function: Signal-to-noise ratio (SNR): calculated as the difference between the echo signal power (unit: dBm) and the noise floor (unit: dBm); Power consumption (P): calculated as D × f × V² / R (V is the excitation voltage, in volts; R is the transducer impedance, in Ω); Comprehensive fitness: The weight distribution is SNR accounting for 70% and power consumption accounting for 30%. For example, when SNR = 22dB and P = 85mW, the fitness is 0.7×22 + 0.3×(1 - 85 / 100) = 16.3; Selection and Crossover: Roulette wheel selection: retain the top 20% of individuals in terms of fitness, and select the remaining 80% based on fitness probability; Simulated binary crossover (SBX): with a crossover probability of 80%, generating offspring parameter combinations, such as parent (D=30%, f=2MHz) and (D=40%, f=1.5MHz) generating offspring (D=35%, f=1.75MHz); Mutation: Gaussian mutation was performed on 10% of the individuals, with the standard deviation set to 5% duty cycle and 0.2 MHz frequency.

[0029] Optimization result example: Candidate 1: D = 40%, f = 1.2MHz, SNR = 22dB, power consumption = 85mW; Candidate 2: D=25%, f=2.5MHz, SNR=21dB, power consumption=78mW.

[0030] The dynamic impedance matching circuit monitors the surface impedance changes of the material in real time, and uses a closed-loop feedback mechanism to adjust the pulse duty cycle to dynamically adapt the emission energy to the material absorption characteristics. The dynamic impedance matching circuit consists of the following modules: Impedance detection module: Use a directional coupler (such as the Mini-Circuits ZFDC-20-5) to sample the incident and reflected waves in real time; Calculate the standing wave ratio (SWR) and reflection coefficient Γ. For example, when SWR = 2.0, the corresponding impedance mismatch rate is 33%; Closed-loop feedback control: Proportional-integral (PI) controller: Parameters are set to proportional gain Kp = 0.2% / Ω, integral time constant Ti = 0.05% / (Ω·s); Duty cycle adjustment: When the impedance is detected to drop from 50Ω to 45Ω (ΔZ = 5Ω), the controller outputs a duty cycle adjustment of ΔD = 0.2% × 5 + 0.05% × 5 × 1s = 1.25%, and the duty cycle is adjusted from 40% to 41.25%; Hardware execution: Control the on-time of the power amplifier through a PWM chip (such as UCC28C42) and adjust the duty cycle in real time.

[0031] Dynamic adaptation example effect: When the material impedance decreases due to temperature increase, the duty cycle increases from 40% to 45%, the reflected power decreases by 30%, and the energy transmission efficiency increases from 75% to 85%.

[0032] The carrier frequency is optimized by combining the gradient descent method, the effective bandwidth of the excitation signal is maximized under the power consumption constraint, and the optimized low-power excitation parameter set is output.

[0033] The gradient descent method is used to search for the optimal carrier frequency f (in MHz) among the candidate frequencies. The goal is to maximize the effective bandwidth B (in kHz) under the power consumption constraint (P ≤ 100 mW).

[0034] Optimize the process: Objective definition: Maximize bandwidth B while satisfying P = D × f × V² / R ≤ 100mW; Gradient calculation: Experimentally measure the bandwidth changes at different frequencies. For example, when f=1MHz, B=200kHz; when f=1.1MHz, B=220kHz; the gradient ΔB / Δf=20kHz / 0.1MHz=0.2; Iterative updates: Initial frequency f0 = 1.0 MHz, learning rate α = 0.1 MHz; Iteration 1: f1 = 1.0 + 0.1 × 0.2 = 1.02 MHz, B = 204 kHz, P = 82 mW (workable). Iteration 2: f2 = 1.02 + 0.1 × 0.25 = 1.045 MHz, B = 210 kHz, P = 85 mW (workable) Constraint processing: If P exceeds the limit after a certain iteration (for example, P = 104 mW when f = 1.3 MHz), then go back to the previous step and reduce the learning rate α to 0.05 MHz.

[0035] Final output parameter example: Optimal solution: D=40%, f=1.2MHz, B=240kHz, P=96mW.

[0036] By analyzing the target material's electromagnetic properties (such as permeability and magnetic permeability) and the required detection accuracy, an intelligent optimization algorithm dynamically adjusts the excitation signal's duty cycle and carrier frequency. This pulse code excitation algorithm automatically matches the optimal transmission parameters based on the material's impedance characteristics, significantly reducing system power consumption while ensuring detection accuracy. This intelligently adapts the detection system to the material's characteristics, avoiding the energy waste associated with traditional fixed-parameter excitation. It is particularly suitable for the long-term monitoring needs of battery-powered IoT devices.

[0037] S202: driving the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously collecting the reflected echo signal through the edge computing node, and performing dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm to output a compressed echo signal; Specifically, the excitation parameter set can be input into a programmable pulse generator to generate a modulated electromagnetic ultrasonic excitation waveform to drive the transducer to emit a detection sound wave; A programmable pulse generator (such as the Keysight 33500B series) receives an optimized low-power stimulus parameter set, including carrier frequency (50kHz to 5MHz, 1kHz steps), pulse duty cycle (5% to 50%, 1% steps), and pulse width (1μs to 100μs). The FPGA is configured to generate the modulation waveform using a hardware description language (such as Verilog). For example: Carrier frequency: 1MHz is selected to match the acoustic impedance characteristics of carbon steel material (sound velocity 5900m / s); Pulse duty cycle: set to 20% to balance power consumption and penetration depth; Pulse width: 50μs, ensuring that the excitation signal covers the reflection path of the full thickness of the material.

[0038] The excitation waveform drives an electromagnetic ultrasonic transducer (EMAT) through an H-bridge power amplifier (such as the TI DRV8432). The EMAT's coil is designed to generate a perpendicularly incident shear wave. The transducer's impedance is adjusted in real time by a dynamic matching circuit (such as an L-type matching network), ensuring a reflection loss of less than 3dB despite variations in the material's surface impedance (e.g., ±10Ω). For example, when testing aluminum alloy (impedance approximately 2.5Ω), the matching circuit automatically adjusts the inductance to 3μH and the capacitance to 220pF to maximize energy transfer efficiency.

[0039] The reflected echo signal is captured by the synchronous acquisition module of the edge node, and the direct wave and multipath interference components are separated by sliding window short-time Fourier transform. The synchronous acquisition module of an edge computing node (such as the NVIDIA Jetson Nano) captures the echo signal at a sampling rate of 100 MS / s, with a trigger delay set to 2 μs after the end of the excitation signal to avoid excitation crosstalk. The acquired signal is first preprocessed through an anti-aliasing filter (cutoff frequency 10 MHz, 4th-order Butterworth filter).

[0040] Sliding window short-time Fourier transform (STFT) parameter settings: Window type: Hanning window, length 256 points (corresponding to 2.56μs time resolution); Overlap rate: 75%, ensuring time-frequency continuity; Frequency resolution: 3906Hz (100MHz / 256).

[0041] STFT spectrum analysis is used to identify direct waves (main peaks at 1 MHz ± 5 kHz) and multipath interference (e.g., boundary reflections with frequency offsets > 20 kHz). For example, when inspecting an 8 mm steel plate, the direct wave arrives at a time of 5.4 μs (calculated as: thickness 8 mm / (5900 m / s × 2) = 6.78 μs, after correcting for sound velocity errors), while boundary reflections are delayed to 12 μs. Morphological filtering (e.g., opening operations) is used to separate the main lobe and side lobes in the spectrum, generating a time-frequency mask matrix that suppresses multipath interference energy by over 90%.

[0042] Based on the instantaneous energy distribution of the signal, a dynamic threshold filter is designed to filter out the environmental electromagnetic noise in real time and obtain the noise-reduced echo signal; Dynamic threshold filtering is divided into two steps: Instantaneous energy calculation: Use Hilbert transform in the time domain to extract the signal envelope and calculate the energy mean E_avg and standard deviation σ within the window (e.g. 50 sampling points). For example, in a certain period of time, E_avg = 0.5V² and σ = 0.1V²; Threshold setting: dynamic threshold T = E_avg + 3σ (99.7% probability of suppressing Gaussian noise). The interval exceeding T is retained, and the rest is set to zero.

[0043] To address periodic electromagnetic interference in industrial environments (such as 25kHz inverter noise), a comb filter was added, with a frequency domain stopband (24.5kHz to 25.5kHz, attenuation -40dB). Zero-phase distortion filtering was achieved using an FIR filter (order 128, Kaiser β window function = 6). For example, in one test, the signal-to-noise ratio (SNR) improved from 15dB to 28dB after filtering.

[0044] Perform wavelet packet decomposition on the denoised echo signal to extract the characteristic frequency band related to the material thickness, and generate compressed time-frequency coefficients through an adaptive quantization algorithm; Wavelet packet decomposition uses the db4 wavelet basis, with 5 decomposition layers, generating 32 sub-bands (100MHz / 32=3.125MHz bandwidth). Based on the material thickness-frequency relationship model, the characteristic frequency band is selected: Thin materials (<5mm): high-frequency sub-bands (such as sub-bands 28 to 32, corresponding to 78.125MHz to 100MHz); Thick material (>20mm): Low-frequency sub-bands (such as sub-bands 1 to 5, corresponding to 0 to 15.625MHz).

[0045] Adaptively quantize the selected subbands: High energy area: 8-bit quantization (step size 0.1mV) to preserve details; Low energy region: 4-bit quantization (1mV step), redundancy compression.

[0046] For example, if the amplitude distribution of a subband signal is [-50mV, +50mV], non-uniform quantization is used: 0-10mV step size is 2mV (5 levels), 10-50mV step size is 5mV (8 levels), and the total number of bits is reduced by 60%.

[0047] The compressed sensing algorithm is used to reconstruct the sparse signal and output a compressed echo signal that meets the transmission bandwidth of the Internet of Things.

[0048] Compressed Sensing (CS) parameter configuration: Sparse basis: Discrete Cosine Transform (DCT) basis, suitable for the energy concentration characteristics of ultrasonic signals; Observation matrix: Gaussian random matrix (size 128×512, meeting RIP conditions); Reconstruction algorithm: Orthogonal Matching Pursuit (OMP), iteration number 50, residual threshold 1e-4.

[0049] Specific process: Sparse representation: Project the 512-point echo signal x onto the DCT basis to obtain sparse coefficients s (only 12% are non-zero); Observation process: y = Φx, Φ is a 128×512 observation matrix, and the compression rate is 75%; Reconstructed signal: The OMP algorithm recovers s' from y, and the inverse DCT is used to obtain the reconstructed signal x'.

[0050] Performance indicators: Reconstruction error: RMSE < 0.5% (compared with the original signal); Data volume: original signal 512 points × 16 bits = 8KB → compressed signal 128 points × 8 bits = 1KB; Transmission latency: In the NB-IoT network (50kbps), the transmission time is reduced from 128ms to 16ms.

[0051] The resulting compressed echo signal is encapsulated using the LoRaWAN protocol and includes a frame header (device ID, timestamp), a compressed data body (1KB), and a checksum (CRC16), ensuring reliable transmission. For example, in one inspection mission, the compressed signal achieved a transmission success rate of >99.9% within a 10km range, meeting the low-power, wide-area coverage requirements of the Industrial Internet of Things.

[0052] Optimized excitation parameters are used to drive the electromagnetic ultrasonic transducer. Edge nodes synchronously collect echo signals and then process them using an advanced time-frequency joint compression algorithm. This algorithm combines short-time Fourier transforms and wavelet packet analysis to achieve signal sparsification while preserving thickness characteristics. This addresses the issue of limited data transmission bandwidth in IoT scenarios. Intelligent compression reduces data transmission volume while maintaining signal quality, significantly extending the operating time of edge nodes.

[0053] S203, performing multi-scale intrinsic mode decomposition on the compressed echo signal, extracting characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; Specifically, an adaptive noise-assisted ensemble empirical mode decomposition algorithm can be used to decompose the compressed echo signal into six eigenmode components to eliminate the modal aliasing effect; Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is an improved EMD algorithm designed to address the modal mixing problem (i.e., different frequency components are mixed in the same mode) in traditional methods. The specific implementation steps are as follows: Noise injection and decomposition initialization: Add Gaussian white noise to the compressed echo signal (for example, sampling rate 100kHz, length 1024 points). The noise amplitude is 10% of the signal amplitude (for example, if the original signal amplitude peak is 5V, the noise standard deviation is 0.5V). Perform 100 noise injection iterations.

[0054] After each iteration, empirical mode decomposition (EMD) is performed on the noisy signal to generate a set of intrinsic mode components (IMFs).

[0055] Mode integration and screening: The IMFs generated by 100 iterations are averaged to obtain the final IMF components. For example, the first IMF (IMF1) corresponds to the highest frequency component in the signal (e.g., 100kHz to 50kHz), IMF2 to the second highest frequency component (50kHz to 25kHz), and so on.

[0056] The decomposition process is forced to terminate by the preset mode number constraint (6 IMFs). If any residual signal energy remains after decomposition (such as low-frequency drift), it is classified as the sixth IMF.

[0057] Modal aliasing elimination: Calculates the spectral overlap of adjacent IMFs. If the overlapping energy between IMF1 and IMF2 in the 25kHz to 50kHz frequency band exceeds 30% (adjustable threshold), modal aliasing is detected.

[0058] The aliased modes are decomposed twice: adaptive noise is re-injected into the overlapping frequency band (with the amplitude reduced to 5%), and the CEEMDAN process is repeated until the spectral overlap of all IMFs is below the threshold.

[0059] The final output of 6 IMF components, for example: IMF1: 80kHz~100kHz (noise-dominated, can be filtered out later); IMF2: 50kHz~80kHz (material surface echo characteristics); IMF3: 25kHz~50kHz (thickness-related main mode); IMF4~IMF6: low-frequency components (reflection from the internal structure of the material).

[0060] Calculate the energy entropy and kurtosis value of each eigenmode component, and select the components with energy entropy less than 0.5 and kurtosis greater than 3 as thickness-sensitive candidate modes; Energy entropy calculation: Sliding window partitioning: Divide the time series of each intrinsic mode component (IMF) into multiple 200ms windows (for example, at a 100kHz sampling rate, each window contains 20,000 sampling points) with a step size of 50ms.

[0061] Window energy statistics: Calculates the signal energy within each window, which is the sum of the squares of the amplitudes of all sampling points (for example, the energy of IMF3 within a window is 5.2 × 10^4).

[0062] Normalization and Entropy Calculation: The energy values ​​of each window are normalized to a probability distribution, and the Shannon Entropy is calculated. For example, the energy distribution of IMF3 is relatively concentrated, with an entropy of 0.32; while the energy of IMF1 is dispersed due to noise interference, with an entropy of 0.89.

[0063] Screening rules: retain modes with entropy values ​​lower than 0.5 and exclude high entropy noise components (such as IMF1).

[0064] Kurtosis value calculation: Kurtosis Definition: Kurtosis reflects the pulse characteristics of a signal. To calculate it, first find the mean (μ) and standard deviation (σ) of the IMF, then calculate the normalized fourth-order central moment. For example, IMF3 has a kurtosis of 4.2, indicating that it contains significant pulse characteristics; IMF2 has a kurtosis of only 2.8 and is therefore excluded.

[0065] Screening rules: retain modes with kurtosis values ​​greater than 3 (such as IMF3 and IMF4) and eliminate low kurtosis components (such as IMF2).

[0066] Sample output: IMF3 (energy entropy 0.32, kurtosis 4.2): retained, thickness-sensitive candidate mode.

[0067] IMF4 (energy entropy 0.45, kurtosis 3.1): retained.

[0068] IMF1 (energy entropy 0.89, kurtosis 1.2): removed.

[0069] Perform Hilbert transform on the thickness-sensitive candidate modes to extract the instantaneous phase features and construct the phase-energy joint distribution matrix; Hilbert transform processing: Analytic signal generation: Perform a Hilbert transform on the candidate modes (such as IMF3 and IMF4) to obtain the analytic signal (the real part is the original signal, and the imaginary part is the transformed result). For example, the imaginary part of the analytic signal of IMF3 at time t = 50ms is -0.8V.

[0070] Instantaneous phase extraction: Calculates the phase angle (in radians). For example, the phase of IMF3 at t=50ms is 1.57 radians (corresponding to 90 degrees).

[0071] Instantaneous amplitude calculation: Combine the real and imaginary parts to calculate the instantaneous amplitude (for example, the amplitude of IMF3 at t=50ms is 2.1V).

[0072] Feature matrix construction: Time segmentation: Divide the timeline into segments at intervals of 10ms (for example, 100 segments are generated when the total duration is 1000ms).

[0073] Feature extraction: Phase standard deviation: The degree of phase fluctuation within each segment (for example, the phase standard deviation of IMF3 in segment 5 is 0.12 radians).

[0074] Energy percentage: The percentage of the square of the instantaneous amplitude in the segment to the total energy (for example, the energy percentage of IMF3 segment 5 is 15%).

[0075] Phase slope: The rate of change of phase over time is calculated by linear regression (e.g., the slope of IMF3 segment 5 is 0.05 rad / ms).

[0076] Matrix concatenation: Merge the features of IMF3 and IMF4 by columns to form a joint matrix of 200 rows × 3 columns (each row corresponds to a time segment).

[0077] Data standardization: Z-score normalization: Normalize each column of data (phase standard deviation, energy fraction, and phase slope) so that its mean is 0 and its standard deviation is 1. For example, after normalization of the phase standard deviation column, the value of segment 5 changes from 0.12 to 1.3.

[0078] Sample output: The joint phase-energy matrix dimension is 200 × 3 (100 segments × 2 modes, 3 features per segment).

[0079] The phase-energy joint distribution matrix is ​​input into the kernel principal component analysis algorithm to fuse multimodal features and reduce dimensionality, and output a set of thickness-related feature parameters after dimensionality reduction.

[0080] Kernel function selection and parameter setting: Gaussian kernel (RBF kernel): used for nonlinear mapping, the kernel width parameter σ is set to 1.5 times the average Euclidean distance between features (for example, when the average distance is 2.5, σ=3.75).

[0081] Dimensionality reduction goal: Reduce the 200×3 matrix to 5 dimensions, retaining 95% of the original information.

[0082] KPCA implementation process: Kernel matrix calculation: Calculate the Gaussian kernel values ​​for all sample pairs, generating a 200 × 200 kernel matrix. For example, the kernel value for sample 1 and sample 2 is 0.82.

[0083] Centering: Center the kernel matrix so that its mean is zero.

[0084] Eigendecomposition: Extract the eigenvectors corresponding to the first five largest eigenvalues ​​(e.g., the eigenvalues ​​are 12.5, 8.3, 5.1, 3.7, and 2.9).

[0085] Projection dimensionality reduction: Project the original data into a 5-dimensional space. For example, the original feature [1.3, 0.8, -0.2] of a certain time segment is mapped to [0.32, -1.2, 0.8, 0.05, -0.4].

[0086] Feature parameter set generation: Timestamp association: Add a timestamp label (such as t=0ms, 10ms, ...,990ms) to each feature vector after dimensionality reduction.

[0087] Data structuring: Generate a 5-dimensional feature vector sequence.

[0088] The compressed signal is subjected to multi-scale decomposition and feature extraction. The intrinsic mode components that are strongly correlated with thickness are separated through ensemble empirical mode decomposition (EEMD). Kernel principal component analysis (KPCA) is then used to fuse multi-dimensional features to eliminate noise and redundant information interference, improve the accuracy and robustness of thickness feature extraction, and provide high-quality input for subsequent thickness calculation.

[0089] S204, inputting the thickness-related feature parameter set into the deep residual-attention network model, combining it with historical data in the material sound velocity database, calculating the dynamic thickness value of the target material, and outputting a thickness prediction result; Specifically, the thickness-related feature parameter set can be input into the deep residual network, and the weights of each feature can be dynamically allocated through the multi-head attention mechanism to generate the initial thickness feature vector; The architecture of the deep residual network (ResNet) is designed to include three residual blocks, each consisting of two convolutional layers (kernel size 3×3, stride 1) and a skip connection. The input layer receives a set of thickness-related feature parameters (including energy entropy, kurtosis, instantaneous phase, and other parameters) with a dimension of 128. After passing through the first convolutional layer (output channels 64, Reluctance Unit (ReLU) activation), it enters the sequence of residual blocks. Batch normalization (BatchNorm) and dropout (dropout rate 0.2) are introduced after the second layer of each residual block to prevent overfitting. For example, when the input feature is the phase-energy distribution matrix of material A, the first residual block may extract low-frequency vibration modes related to thickness, while the third residual block captures high-frequency surface wave features.

[0090] The Multi-Head Attention (MHA) mechanism uses four independent attention heads, and the query, key, and value matrices of each head are 32×32 in size. The specific operation is as follows: Feature segmentation: Split the 256-dimensional feature vector output by the residual network into four 64-dimensional sub-vectors; Attention calculation: Each head performs self-attention on the sub-vectors, focusing on the correlation between different feature dimensions. For example, head 1 focuses on the correlation between phase change rate and energy entropy, and head 3 analyzes the coupling effect of kurtosis value and temperature drift. Weight fusion: The outputs of each head are integrated through concatenation and a fully connected layer (output dimension 128) to generate an initial thickness feature vector.

[0091] Example of dynamic weight allocation: In a certain detection, the attention weight of the phase feature is 0.6, the energy entropy weight is 0.3, and the kurtosis value weight is 0.1, reflecting the dominant role of phase information on the current material thickness.

[0092] Query the temperature-sound velocity curve of the same type of material in the material sound velocity database, and use cubic spline interpolation to compensate for the sound velocity correction parameters caused by ambient temperature; Database structure and data retrieval The material sound velocity database stores the temperature-sound velocity relationship of different materials. Each record contains the following fields: Material type (e.g., aluminum alloy 6061, stainless steel 304), identified by a unique ID; Temperature sequence (range -20℃~200℃, step size 5℃, such as 25℃, 30℃, 35℃, etc.); Sound velocity (unit: m / s, e.g., aluminum alloy 6061 is 6200 m / s at 25°C and 6185 m / s at 30°C); Data source (laboratory measurement or authoritative literature, with an error range of ±0.5%).

[0093] When the system detects that the current material is aluminum alloy 6061 and the ambient temperature is 38°C, it automatically retrieves the temperature-sound velocity mapping table corresponding to this material from the database. For example, the temperature points stored in the database are 25°C, 30°C, 35°C, and 40°C, and the corresponding sound velocity values ​​are 6200m / s, 6185m / s, 6160m / s, and 6135m / s.

[0094] Cubic spline interpolation calculation Segmented interval division: adjacent temperature points in the database are used as interpolation intervals, for example, 38°C is in the 35°C~40°C interval; Interpolation point generation: A smooth cubic polynomial curve (called a "spline") is constructed within each interval, ensuring continuity of the first-order derivative (slope) and second-order derivative (curvature) at the junction of adjacent intervals. For example, within the 35°C to 40°C interval, a smooth curve is generated using four control points (35°C: 6160m / s, 36°C: 6155m / s, 38°C: target value, 40°C: 6135m / s). Sound speed calculation: Based on the current temperature of 38°C, the sound speed is interpolated from a spline curve in the range of 35°C to 40°C. For example, the calculated output sound speed is 6148 m / s (compared to the linear interpolation of 6145 m / s, the spline interpolation is closer to the actual physical characteristics).

[0095] Dynamic error compensation and verification Kalman filter correction: If the database temperature step size is large (e.g., 5°C), the interpolation results may be biased due to material nonlinearity. The system uses a Kalman filter to integrate real-time calibration data (e.g., the actual sound velocity at 38°C measured by a portable sound velocity meter is 6142 m / s) and dynamically adjust the interpolation parameters, reducing the error from 6 m / s (6148 vs. 6142) to ±1 m / s. Temperature sensitivity analysis: Verify the slope of the sound velocity-temperature curve (for example, the sound velocity temperature coefficient of aluminum alloy 6061 is -2.5 m / s / °C). If the interpolated coefficient exceeds the historical statistical range (for example, -3.0 to -2.0 m / s / °C), an alarm is triggered and the backup interpolation algorithm (such as piecewise linear interpolation) is switched.

[0096] Example process: The testing environment temperature is 38°C and the material is aluminum alloy 6061; Extract data of 35℃ (6160m / s), 40℃ (6135m / s) and adjacent temperature points from the database; Generate a cubic spline curve in the range of 35℃ to 40℃, and interpolate the sound speed at 38℃ to 6148m / s; The portable calibrator measured a sound velocity of 6142 m / s, and the Kalman filter correction interpolation result was 6143 m / s (error ±1 m / s), and the corrected sound velocity parameters were output.

[0097] Design a cross-modal feature fusion layer to perform tensor concatenation of the sound velocity correction parameters and the initial thickness feature vector to generate a fused feature representation; The cross-modal fusion layer adopts a dual-channel architecture: Sound speed parameter expansion: The scalar sound speed value (such as 6172m / s) is mapped to a high-dimensional vector through a fully connected layer (input dimension 1, output dimension 64), and the activation function is LeakyReLU (negative slope 0.01); Feature splicing: The expanded sound velocity vector (64 dimensions) and the initial thickness feature vector (128 dimensions) are spliced ​​along the channel dimension to generate a 192-dimensional fusion feature; Dimensionality reduction: Features are compressed through a fully connected layer (input 192 dimensions, output 128 dimensions) to eliminate redundant information.

[0098] For example, in aluminum alloy detection, the sound velocity correction parameter highlights the influence of temperature on ultrasonic propagation, while the thickness feature vector contains echo phase information. The fused feature can accurately reflect the thickness change after temperature compensation.

[0099] A Bayesian probability module is integrated in the output layer to calculate the confidence interval of the thickness prediction value based on the fused feature representation; The Bayesian probability module uses the Monte Carlo Dropout method to achieve uncertainty estimation: Random forward propagation: During the test phase, 50 forward inferences are performed on the fused features, each time randomly dropping some neurons (Dropout rate 0.1); Statistical distribution modeling: Collect 50 inference results (thickness predictions) and calculate the mean μ and standard deviation σ. For example, if the 50 predictions are [2.1mm, 2.0mm, 2.2mm, ...], μ = 2.05mm and σ = 0.06mm. Confidence interval calculation: At a 95% confidence level (Z = 1.96), the confidence interval is μ ± 1.96σ. In the above example, the confidence interval is 2.05 ± 0.12 mm.

[0100] Anomaly detection: If a predicted value deviates from the mean by more than 3σ (e.g., 2.5mm), it is considered an abnormal measurement and a re-inspection mechanism is triggered.

[0101] The sliding window weighted average algorithm is used to eliminate single measurement noise, and the dynamic thickness change curve is optimized in combination with historical prediction data to output the final thickness prediction result.

[0102] The sliding window weighted average algorithm sets the window size to 5 and the weight coefficient decays exponentially (the most recent data has the highest weight): Weight distribution: , i=1~5, the normalized weights are [0.53, 0.26, 0.13, 0.06, 0.02]; Dynamic Update: Each time a new measurement enters the window, the oldest data is discarded and a weighted average is calculated. For example, if the thickness values ​​in the window are [2.05, 2.08, 2.03, 2.06, 2.04] mm, the weighted average is 2.05 × 0.53 + ... + 2.04 × 0.02 = 2.052 mm.

[0103] Historical data optimization: Trend fitting: A first-order autoregressive model (AR(1)) is used to predict thickness trend. Model parameters are fitted to historical data using the least squares method. Residual compensation: Calculate the residual between the sliding window mean and the AR prediction value. If the residual exceeds the threshold (such as ±0.02mm), adjust the AR model coefficient.

[0104] Final output example: Instantaneous measurement value: 2.05 mm (confidence interval 2.05 ± 0.12 mm); Sliding window mean: 2.052mm; AR predicted value: 2.051mm; Final thickness prediction: 2.051mm (accuracy ±0.05mm).

[0105] A hybrid model combining a deep residual network and an attention mechanism is constructed. Temperature compensation is performed by querying the material sound velocity database, and the Bayesian method is used to output prediction results with confidence intervals. Finally, sliding window smoothing is used to improve dynamic measurement stability and achieve high-precision thickness measurement. Temperature compensation reduces measurement errors, and the confidence interval quantifies the reliability of the results, providing a scientific basis for industrial decision-making.

[0106] S205, based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, which is synchronized to the visual interface of the terminal device for three-dimensional thickness distribution display, and stored in the encrypted blockchain node to realize the traceability of the detection data.

[0107] Specifically, the thickness prediction results of multiple edge nodes in the IoT cloud platform can be aligned through the spatiotemporal registration algorithm, and the Dempster-Shafer evidence theory can be used to eliminate spatial measurement conflicts and generate consistent fusion results. Specifically, the thickness prediction results of multiple edge nodes in the IoT cloud platform are aligned through a spatiotemporal registration algorithm, and the Dempster-Shafer evidence theory is used to eliminate spatial measurement conflicts and generate consistent fusion results, which may include: Based on the time and space stamp information and geographic location coordinates of each edge node, a sliding window dynamic time warping algorithm is used to align the time series of multi-node thickness prediction results, eliminate the timing offset caused by sampling frequency differences, and generate a time-synchronized thickness data matrix. On an IoT cloud platform, multiple edge nodes (e.g., 10 detection terminals deployed at different locations on the material surface) upload thickness prediction results at different sampling frequencies (e.g., 100Hz for node A and 50Hz for node B). The core of the sliding window dynamic time warping (DTW) algorithm is to eliminate phase shifts caused by sampling rate differences by elastically matching time series. The specific operation is as follows: Time window division: Based on the lowest sampling rate (50 Hz), the window length is set to 200 ms (corresponding to 10 sampling points of node B). Node A contains 20 sampling points in this window. Sequence alignment: Calculate the minimum curvature path of the thickness sequence of nodes A and B within the window. For example, the 15th to 20th sampling points of node A correspond to the 8th to 10th points of node B. Dynamic programming is used to find the optimal matching path, and the path cost function is the sum of squared Euclidean distances. Interpolation compensation: perform cubic spline interpolation on the sparse sampling points of node B to generate pseudo sampling points with the same time density as node A (for example, interpolate 10 points of node B to 20 points); Data Matrix Generation: Arrange the aligned time series into a matrix based on spatial position, with rows representing time points (one row every 2 ms), columns representing node numbers (1-10), and element values ​​representing thickness values ​​(in mm). For example, position (50,3) = 2.1 mm in the matrix represents the thickness measurement at the 50th time point (100 ms) and node 3.

[0108] Technical example: When node A detects that the thickness of a certain point suddenly changes from 2.0 mm to 2.5 mm within 50 ms, while node B only captures 2.3 mm due to its low sampling rate, DTW stretches the time axis of node B to align the mutation time and ensure timing consistency.

[0109] Based on the mapping relationship of the three-dimensional coordinate system of the material surface, the time-synchronized thickness data matrix is ​​input into the spatial registration model, the spatial position deviation between nodes is corrected through the affine transformation matrix, and the spatially aligned multi-node thickness distribution tensor is output; The core of the spatial registration model is to construct an affine transformation matrix to resolve the deviation of the edge node installation position (for example, the actual coordinates of node 5 deviate from the calibration position by 5 cm). Reference coordinate system establishment: With the lower left corner of the material surface as the origin (0,0,0), a right-handed three-dimensional coordinate system is established, and the node calibration coordinates are stored in the cloud platform (for example, node 1 is calibrated to (10cm,20cm,0)); Feature point matching: Select feature points on the material surface (such as welds and holes) and match the actual coordinates with the calibrated coordinates using node detection data. For example, node 1 detects the actual coordinates of a hole as (10.2cm, 19.8cm, 0.1cm), which deviates from the calibrated value (10, 20, 0); Affine transformation solution: Solve the 6-DOF transformation matrix (translation Tx, Ty, Tz; rotation Rx, Ry, Rz) for each node, using least squares optimization, with the objective function being the sum of squares of feature point matching errors. For example, after optimizing the transformation matrix for node 1, the detection coordinates (10.2, 19.8, 0.1) are corrected to (10, 20, 0); Tensor Generation: The corrected thickness data is interpolated onto a 3D grid (1 cm × 1 cm × 1 cm) to generate a 4D tensor of size X × Y × Z × T (X / Y / Z are spatial dimensions, T is the time dimension). For example, the material surface is divided into a 100 × 100 grid, and each grid point stores a sequence of thickness values ​​over time.

[0110] Technical example: A node has a 0.5cm Z-axis coordinate deviation due to tilted installation. Affine transformation corrects this by rotating it -0.5° around the X-axis, allowing the thickness data to be accurately mapped to the material surface.

[0111] Uncertainty modeling is performed on the spatially aligned multi-node thickness distribution tensor. A trust function is generated based on the historical detection error distribution of each node. Combined with the basic probability distribution rule in the Dempster-Shafer theory, the confidence interval and conflict coefficient matrix of each thickness value are calculated. Uncertainty modeling aims to quantify the credibility of multi-node measurement results and solve the problem of conflicting data fusion: Historical error statistics: Analyze the error distribution of each node over the past 1,000 tests. For example, the error of node 3 follows a normal distribution with a mean of 0.05 mm and a standard deviation of 0.02 mm. Trust function construction: Convert the error distribution into a basic probability allocation (BPA). Assume that the true thickness interval is [2.0mm, 2.2mm]. When node 3 measures 2.1mm, its BPA is: m({2.1mm})=0.7 (corresponding to the probability of the error being within ±0.05mm); m({2.0~2.2mm})=0.3 (covering the possible error range); Confidence calculation: For each grid point's thickness value, the BPA of all nodes is combined to calculate the confidence interval. For example, if the BPAs of nodes 1 to 3 at a certain point are 0.6, 0.7, and 0.5, respectively, the confidence interval for the thickness at that point is [2.08mm, 2.12mm] (95% confidence level). Conflict coefficient matrix: Calculates the degree of evidence conflict between nodes. For example, if node 4 has a BPA of m({2.5mm})=0.8 at a certain point, while all other nodes have m({2.1mm})>0.6, the conflict coefficient K=0.8×0.6=0.48. If it exceeds the threshold of 0.3, it is marked as a high-conflict area.

[0112] Technical example: In the edge area of ​​the material, node 7 has a large signal attenuation error (BPA=0.4), which causes a conflict with node 2 in the center area (BPA=0.9). The system automatically reduces the fusion weight of node 7.

[0113] The conflict coefficient matrix is ​​used to construct the evidence fusion weight table, and the orthogonal sum formula is used to perform joint probability distribution on multi-node evidence to eliminate spatial measurement conflicts. The thickness probability distribution map after consistency fusion is output as the consistency fusion result.

[0114] The orthogonal sum formula (Dempster's combination rule) resolves conflicting evidence through weighted fusion: Weight table construction: Assign a weight to each node based on the conflict coefficient K, using the formula weight = 1-K. For example, node 4 in the conflict area has K = 0.48 and weight = 0.52; Evidence synthesis: For each grid point, the BPA is synthesized by weight. For example, the BPA of nodes 1 to 3 are m1, m2, and m3, with weights w1 = 0.8, w2 = 0.9, and w3 = 0.7. After synthesis, m_combined = (w1*m1 + w2*m2 + w3*m3) / (w1+w2+w3); Probability distribution: Convert the synthesized BPA into a probability distribution. For example, if m_combined({2.1mm})=0.75 and m_combined({2.0~2.2mm})=0.25, the thickness probability distribution is 2.1mm (75%) and 2.0~2.2mm (25%). Conflict elimination: For high-conflict areas (K>0.3), the re-detection mechanism is activated, triggering the edge node to perform three redundant measurements of the area, and then re-integrate after updating the BPA.

[0115] Technical example: In the weld area, nodes 5 and 8 conflict due to material heterogeneity (K=0.4). After the system synthesizes the weights, it generates a probability distribution diagram showing that the thickness of this area is 2.15mm±0.03mm (90% confidence level) and marks it as an area requiring manual re-inspection.

[0116] The consistency fusion results are input into the 3D surface reconstruction engine, and the iso-thickness line distribution map and thermal map are generated by combining the geometric features of the material surface; The 3D surface reconstruction engine is based on the Marching Cubes algorithm: Voxelization: Convert the spatially aligned thickness tensor into a voxel grid (resolution 1mm³), storing a thickness value and confidence for each voxel. For example, voxel (10,20,30) has a thickness of 2.5mm and a confidence of 0.85.

[0117] Isosurface extraction: Set a threshold (e.g. thickness = 2.5mm ± 0.1mm) and traverse the voxel grid to generate triangular facets. Algorithm steps: Voxel classification: marking whether the voxel vertex is within the threshold range; Triangulation: Connect vertices according to 15 predefined topologies; Smoothing optimization: Laplace smoothing algorithm is used to eliminate aliasing, and it is iterated 3 times.

[0118] Contour Line Generation: Contour lines are extracted on the XY plane with a line spacing of 0.1mm and color-coded (blue → thin, red → thick). For example, a thickness of 2.3mm is light blue, 2.5mm is green, and 2.7mm is orange.

[0119] Heatmap rendering: Color Mapping: HSV hue maps to thickness (0-5mm corresponds to 0°-240°), and saturation maps to confidence (0-1 corresponds to 0%-100%). For example, a thickness of 2.5mm (H=120°) and a confidence of 0.85 (S=85%) will result in a bright green.

[0120] Lighting model: Phong lighting model enhances the three-dimensional effect, ambient light intensity 0.3, diffuse reflection coefficient 0.7, specular reflection index 20.

[0121] Interactive visualization: Supports rotation, zooming, and cross-sectioning. For example, users can cut a Z=5mm plane to view the internal thickness distribution.

[0122] Example output: The isothickness lines of a metal plate show that the thickness at the edge (X < 50 mm) is 2.3 ± 0.1 mm, and at the center (X = 150 mm) is 2.8 ± 0.15 mm. The red areas in the heat map indicate excessive thickness and require a warning.

[0123] Generate the final thickness test report based on the iso-thickness line distribution map and heat map. Use the lightweight blockchain framework to construct a Merkle tree hash for the report and raw data, and store it on the chain through smart contracts. Thickness test report generation Structured data encapsulation: The test report is divided into three parts, including: Metadata: including detection time (UTC timestamp, accuracy 1ms), material unique identifier (such as UUID format), operator ID (8-digit code) and equipment serial number; Statistical results: Calculate the overall average thickness of the material (e.g. 2.5mm±0.07mm), maximum thickness (3.2mm), minimum thickness (1.8mm) and standard deviation (0.1mm); Abnormal area list: Mark the three-dimensional coordinates of the thickness deviation position (such as X=120mm, Y=80mm, Z=5mm) and the deviation range (such as +10%).

[0124] Visual attachments: The distribution map of isopyctic lines was saved as a vector image (SVG format, resolution 300 dpi), with a line spacing of 0.1 mm and a color gradient from blue (thinnest) to red (thickest); The heat map is embedded in interactive links in WebGL format, supports online zooming and rotation, and the color temperature range is set from 2000K (blue) to 6000K (red).

[0125] Blockchain evidence storage and Merkle tree construction Lightweight blockchain framework: Using Hyperledger Fabric 2.3, configured with 4 peer nodes (testing agencies, suppliers, customers, regulators) and 1 orderer node, the consensus algorithm is Raft, and the transaction confirmation delay is less than 2 seconds; Data block and hash calculation: The original data (including echo signals and processing logs) is divided into 1KB blocks, and a SHA-256 hash is calculated for each block (outputting a 64-bit hexadecimal string, such as 0x3a7d...c1f2); When constructing a Merkle tree, the leaf node is the hash of the data block, and the parent node is the concatenation of the hashes of the child nodes, which is then hashed again to generate the root hash (e.g., 0x5b9e...d4a3). Smart contract logic: The deployed chaincode (smart contract) verifies the operator’s digital signature (ECDSA-secp256k1 curve), confirms the permissions, binds the report metadata to the Merkle root hash, and writes it to the blockchain ledger; Each time a certificate is stored, a unique transaction ID (such as TXID=7821) is generated, which is associated with the block height (such as Block#12345) and timestamp.

[0126] Data on-chain and security assurance Encrypted storage: The original data is encrypted with AES-256 and stored in IPFS (InterPlanetary File System), and the key is dynamically managed by the blockchain smart contract; Anti-tampering mechanism: Any data modification will cause the Merkle root hash to change, and blockchain nodes detect inconsistencies through regular verification (every 10 minutes).

[0127] An interactive traceability module is deployed in the visual interface to associate the three-dimensional isopach distribution map with the thickness detection report stored in the blockchain. It supports clicking on any thickness point to query the corresponding original echo signal and processing log.

[0128] Interactive 3D visualization engine Front-end framework: Developed based on the Three.js library, with a rendering frame rate of ≥30fps and support for real-time interaction of tens of thousands of voxels; Coordinate mapping: The vertex coordinates of the 3D model (e.g. X=120mm, Y=80mm, Z=5mm) are aligned with the 3D grid of the test data with an error of <0.01mm; When a click event is triggered, the world coordinates of the click position are obtained through the raycasting algorithm and converted into the material local coordinate system.

[0129] Blockchain data association query Click to trigger the query process: The user clicks a point in the 3D map (e.g., X=120mm, Y=80mm, Z=5mm), and the frontend sends the coordinates to the blockchain gateway (REST API interface); The gateway queries the blockchain ledger and matches the data block hash (e.g., 0x5b9e...d4a3) and the associated transaction ID (TXID=7821) corresponding to the coordinates. Download the encrypted original echo signal (WAV format, sampling rate 1MHz) and processing log (JSON format) from the IPFS distributed storage, decrypt them and return them to the front end.

[0130] Log display content: Collection information: edge node ID (such as Node_05), collection time (2023-10-01 14:05:23.456), signal-to-noise ratio (SNR=32dB); Processing records: time-frequency compression algorithm parameters (wavelet basis = db4, compression ratio = 80%), thickness prediction confidence (0.85); Blockchain evidence: block height (Block#12345), evidence storage time (2023-10-01 14:05:30.120).

[0131] Traceability function expansion and performance optimization Caching mechanism: Frequently used query results (such as frequently detected areas) are cached on edge nodes, reducing response time from 500ms to <100ms. Multi-dimensional association: Clicking on an abnormal area can associate the production batch (such as Batch_B-2032) and process parameters (heat treatment temperature = 650℃±10℃); the historical data comparison function supports sliding timelines (such as comparing the thickness change curves at the same location in September and October of a certain year).

[0132] Multi-node data is integrated on the cloud platform, evidence theory is used to resolve spatial measurement conflicts, visual reports are generated through three-dimensional reconstruction, and blockchain technology is used to achieve data tamper-proofing and full-process traceability, providing globally consistent thickness distribution assessment, and three-dimensional visualization to intuitively display defect locations. Blockchain evidence storage meets industrial testing and certification requirements, and supports quality traceability and responsibility determination.

[0133] It can be seen that according to the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, the pulse code excitation algorithm is used to perform parameter optimization processing, and a low-power excitation parameter set is output; based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal, and a compressed echo signal is output; the compressed echo signal is subjected to multi-scale intrinsic mode decomposition, and a thickness-related feature parameter set is output; the thickness-related feature parameter set is input into the deep residual-attention network model, and a thickness prediction result is output; based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, thereby enabling high-precision and traceable thickness detection based on a miniaturized and low-power Internet of Things.

[0134] Another embodiment of the present invention provides an electromagnetic ultrasonic thickness detection system based on a miniaturized low-power Internet of Things, see Figure 3 , the system may include: Processing module 301 is used to perform parameter optimization processing using a pulse coding excitation algorithm based on the electromagnetic characteristics of the target material and the preset detection accuracy requirements, adaptively adjust the pulse duty cycle and carrier frequency to match the material impedance characteristics, and output a low-power excitation parameter set; An acquisition module 302 is configured to drive the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously acquire the reflected echo signal through the edge computing node, perform dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm, and output a compressed echo signal; A decomposition module 303 is configured to perform multi-scale intrinsic mode decomposition on the compressed echo signal, extract characteristic mode components related to material thickness, fuse multi-modal features through a nonlinear coupling analysis algorithm, and output a thickness-related characteristic parameter set; Prediction module 304, configured to input the thickness-related feature parameter set into a deep residual-attention network model, combine it with historical data in a material sound velocity database, calculate the dynamic thickness value of the target material, and output a thickness prediction result; The fusion module 305 is used to perform multi-node data fusion on the thickness prediction results based on the Internet of Things cloud platform, generate a final thickness detection report, synchronize it to the visual interface of the terminal device for three-dimensional thickness distribution display, and store it in the encrypted blockchain node to realize the traceability of the detection data.

[0135] It can be seen that according to the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, the pulse code excitation algorithm is used to perform parameter optimization processing, and a low-power excitation parameter set is output; based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal, and a compressed echo signal is output; the compressed echo signal is subjected to multi-scale intrinsic mode decomposition, and a thickness-related feature parameter set is output; the thickness-related feature parameter set is input into the deep residual-attention network model, and a thickness prediction result is output; based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, thereby enabling high-precision and traceable thickness detection based on a miniaturized and low-power Internet of Things.

[0136] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0137] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, performing parameter optimization processing using a pulse coding excitation algorithm based on the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, adaptively adjusting the pulse duty cycle and carrier frequency to match the material impedance characteristics, and outputting a low-power excitation parameter set; S202: driving the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously collecting the reflected echo signal through the edge computing node, and performing dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm to output a compressed echo signal; S203, performing multi-scale intrinsic mode decomposition on the compressed echo signal, extracting characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; S204, inputting the thickness-related feature parameter set into the deep residual-attention network model, combining it with historical data in the material sound velocity database, calculating the dynamic thickness value of the target material, and outputting a thickness prediction result; S205, based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, which is synchronized to the visual interface of the terminal device for three-dimensional thickness distribution display, and stored in the encrypted blockchain node to realize the traceability of the detection data.

[0138] It can be seen that according to the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, the pulse code excitation algorithm is used to perform parameter optimization processing, and a low-power excitation parameter set is output; based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal, and a compressed echo signal is output; the compressed echo signal is subjected to multi-scale intrinsic mode decomposition, and a thickness-related feature parameter set is output; the thickness-related feature parameter set is input into the deep residual-attention network model, and a thickness prediction result is output; based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, thereby enabling high-precision and traceable thickness detection based on a miniaturized and low-power Internet of Things.

[0139] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0140] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0141] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, performing parameter optimization processing using a pulse coding excitation algorithm based on the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, adaptively adjusting the pulse duty cycle and carrier frequency to match the material impedance characteristics, and outputting a low-power excitation parameter set; S202: driving the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously collecting the reflected echo signal through the edge computing node, and performing dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm to output a compressed echo signal; S203, performing multi-scale intrinsic mode decomposition on the compressed echo signal, extracting characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; S204, inputting the thickness-related feature parameter set into the deep residual-attention network model, combining it with historical data in the material sound velocity database, calculating the dynamic thickness value of the target material, and outputting a thickness prediction result; S205, based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, which is synchronized to the visual interface of the terminal device for three-dimensional thickness distribution display, and stored in the encrypted blockchain node to realize the traceability of the detection data.

[0142] It can be seen that according to the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, the pulse code excitation algorithm is used to perform parameter optimization processing, and a low-power excitation parameter set is output; based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal, and a compressed echo signal is output; the compressed echo signal is subjected to multi-scale intrinsic mode decomposition, and a thickness-related feature parameter set is output; the thickness-related feature parameter set is input into the deep residual-attention network model, and a thickness prediction result is output; based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction result to generate a final thickness detection report, thereby enabling high-precision and traceable thickness detection based on a miniaturized and low-power Internet of Things.

[0143] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. An electromagnetic ultrasonic thickness detection method based on miniaturized low-power Internet of Things, characterized in that: The method comprises: According to the electromagnetic properties of the target material and the preset detection accuracy requirements, the pulse coding excitation algorithm is used to optimize the parameters. The pulse duty cycle and carrier frequency are adaptively adjusted to match the material impedance characteristics, and a low-power excitation parameter set is output. Based on the low-power excitation parameter set, the electromagnetic ultrasonic transducer is driven to generate a detection signal, the reflected echo signal is synchronously collected through the edge computing node, and a time-frequency joint compression algorithm is used to perform dynamic noise reduction and sparse processing, and a compressed echo signal is output; Performing multi-scale eigenmode decomposition on the compressed echo signal to extract characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; Input the thickness-related feature parameter set into the deep residual-attention network model, combine it with the historical data in the material sound velocity database, calculate the dynamic thickness value of the target material, and output the thickness prediction result; Based on the Internet of Things cloud platform, multi-node data fusion is performed on the thickness prediction results to generate a final thickness detection report, which is synchronized to the visual interface of the terminal device for three-dimensional thickness distribution display and stored in the encrypted blockchain node to realize the traceability of the detection data.

2. The method according to claim 1, characterized in that According to the electromagnetic characteristic parameters of the target material and the preset detection accuracy requirements, the pulse coding excitation algorithm is used to perform parameter optimization processing, and the pulse duty cycle and carrier frequency are adaptively adjusted to match the material impedance characteristics to output a low-power excitation parameter set, including: According to the electromagnetic permeability and hysteresis loss parameters of the target material, a material impedance spectrum model is constructed to generate a frequency-impedance mapping table; Based on the preset detection accuracy requirements, a genetic algorithm is used to screen the initial pulse parameter combinations that meet the signal-to-noise ratio threshold in the frequency-impedance mapping table to generate a candidate set of duty cycle and carrier frequency; The dynamic impedance matching circuit monitors the surface impedance changes of the material in real time, and uses a closed-loop feedback mechanism to adjust the pulse duty cycle to dynamically adapt the emission energy to the material absorption characteristics. The carrier frequency is optimized by combining the gradient descent method, the effective bandwidth of the excitation signal is maximized under the power consumption constraint, and the optimized low-power excitation parameter set is output.

3. The method according to claim 2, characterized in that The method drives the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously collects the reflected echo signal through the edge computing node, and uses a time-frequency joint compression algorithm to perform dynamic noise reduction and sparse processing to output a compressed echo signal, including: Inputting the excitation parameter set into a programmable pulse generator to generate a modulated electromagnetic ultrasonic excitation waveform to drive the transducer to emit a detection sound wave; The reflected echo signal is captured by the synchronous acquisition module of the edge node, and the direct wave and multipath interference components are separated by sliding window short-time Fourier transform. Based on the instantaneous energy distribution of the signal, a dynamic threshold filter is designed to filter out the environmental electromagnetic noise in real time and obtain the noise-reduced echo signal; Perform wavelet packet decomposition on the denoised echo signal to extract the characteristic frequency band related to the material thickness, and generate compressed time-frequency coefficients through an adaptive quantization algorithm; The compressed sensing algorithm is used to reconstruct the sparse signal and output a compressed echo signal that meets the transmission bandwidth of the Internet of Things.

4. The method according to claim 3, characterized in that The method of performing multi-scale eigenmode decomposition on the compressed echo signal, extracting characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set includes: Adaptive noise-assisted ensemble empirical mode decomposition algorithm is used to decompose the compressed echo signal into six eigenmode components to eliminate the modal aliasing effect; Calculate the energy entropy and kurtosis value of each eigenmode component, and select the components with energy entropy less than 0.5 and kurtosis greater than 3 as thickness-sensitive candidate modes; Perform Hilbert transform on the thickness-sensitive candidate modes to extract the instantaneous phase features and construct the phase-energy joint distribution matrix; The phase-energy joint distribution matrix is ​​input into the kernel principal component analysis algorithm to fuse multimodal features and reduce dimensionality, and output a set of thickness-related feature parameters after dimensionality reduction.

5. The method according to claim 4, characterized in that The thickness-related feature parameter set is input into the deep residual-attention network model, combined with the historical data in the material sound velocity database, to calculate the dynamic thickness value of the target material and output the thickness prediction result, including: The thickness-related feature parameter set is input into the deep residual network, and the weight of each feature is dynamically assigned through the multi-head attention mechanism to generate the initial thickness feature vector; Query the temperature-sound velocity curve of the same type of material in the material sound velocity database, and use cubic spline interpolation to compensate for the sound velocity correction parameters caused by ambient temperature; Design a cross-modal feature fusion layer to perform tensor concatenation of the sound velocity correction parameters and the initial thickness feature vector to generate a fused feature representation; A Bayesian probability module is integrated in the output layer to calculate the confidence interval of the thickness prediction value based on the fused feature representation; The sliding window weighted average algorithm is used to eliminate single measurement noise, and the dynamic thickness change curve is optimized in combination with historical prediction data to output the final thickness prediction result.

6. The method according to claim 5, characterized in that The thickness prediction results are subjected to multi-node data fusion based on the IoT cloud platform to generate a final thickness detection report, which is synchronized to the visual interface of the terminal device for three-dimensional thickness distribution display and stored in the encrypted blockchain node to realize detection data traceability, including: The thickness prediction results of multiple edge nodes in the IoT cloud platform are aligned through a spatiotemporal registration algorithm, and the Dempster-Shafer evidence theory is used to eliminate spatial measurement conflicts and generate consistent fusion results. The consistency fusion results are input into the 3D surface reconstruction engine, and the iso-thickness line distribution map and thermal map are generated by combining the geometric features of the material surface; Generate the final thickness test report based on the iso-thickness line distribution map and heat map. Use the lightweight blockchain framework to construct a Merkle tree hash for the report and raw data, and store it on the chain through smart contracts. An interactive traceability module is deployed in the visual interface to associate the three-dimensional isopach distribution map with the thickness detection report stored in the blockchain. It supports clicking on any thickness point to query the corresponding original echo signal and processing log.

7. The method according to claim 6, characterized in that The method aligns the thickness prediction results of multiple edge nodes in the IoT cloud platform through a spatiotemporal registration algorithm, uses the Dempster-Shafer evidence theory to eliminate spatial measurement conflicts, and generates consistent fusion results, including: Based on the time and space stamp information and geographic location coordinates of each edge node, a sliding window dynamic time warping algorithm is used to align the time series of multi-node thickness prediction results, eliminate the timing offset caused by sampling frequency differences, and generate a time-synchronized thickness data matrix. Based on the mapping relationship of the three-dimensional coordinate system of the material surface, the time-synchronized thickness data matrix is ​​input into the spatial registration model, the spatial position deviation between nodes is corrected through the affine transformation matrix, and the spatially aligned multi-node thickness distribution tensor is output; Uncertainty modeling is performed on the spatially aligned multi-node thickness distribution tensor. A trust function is generated based on the historical detection error distribution of each node. Combined with the basic probability distribution rule in the Dempster-Shafer theory, the confidence interval and conflict coefficient matrix of each thickness value are calculated. The conflict coefficient matrix is ​​used to construct the evidence fusion weight table, and the orthogonal sum formula is used to perform joint probability distribution on multi-node evidence to eliminate spatial measurement conflicts. The thickness probability distribution map after consistency fusion is output as the consistency fusion result.

8. An electromagnetic ultrasonic thickness detection system based on a miniaturized low-power Internet of Things, characterized in that: The system comprises: A processing module is used to optimize parameters using a pulse coding excitation algorithm based on the electromagnetic properties of the target material and preset detection accuracy requirements, adaptively adjust the pulse duty cycle and carrier frequency to match the material impedance characteristics, and output a low-power excitation parameter set; an acquisition module, configured to drive the electromagnetic ultrasonic transducer to generate a detection signal based on the low-power excitation parameter set, synchronously acquire the reflected echo signal through the edge computing node, perform dynamic noise reduction and sparsification processing using a time-frequency joint compression algorithm, and output a compressed echo signal; a decomposition module for performing multi-scale intrinsic mode decomposition on the compressed echo signal, extracting characteristic mode components related to material thickness, fusing multi-modal features through a nonlinear coupling analysis algorithm, and outputting a thickness-related characteristic parameter set; A prediction module is used to input the thickness-related feature parameter set into a deep residual-attention network model, combine it with historical data in a material sound velocity database, calculate the dynamic thickness value of the target material, and output a thickness prediction result; The fusion module is used to perform multi-node data fusion on the thickness prediction results based on the Internet of Things cloud platform, generate a final thickness detection report, synchronize it to the visual interface of the terminal device for three-dimensional thickness distribution display, and store it in the encrypted blockchain node to realize the traceability of the detection data.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 7 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 7.

Images