Fire-fighting non-destructive detection method based on multi-modal detection

Abstract

The invention discloses a fire-fighting non-destructive detection method based on multi-modal detection, and relates to the technical field of fire-fighting engineering, and the method comprises the following steps: carrying out the rapid scanning of fire-fighting equipment through an ultrasonic array, analyzing and positioning a suspected defect region based on a time domain reflection signal, and generating a three-dimensional coordinate mapping model; and performing multi-angle X-ray projection acquisition on the positioning area, and generating an internal structure cross-sectional image through a three-dimensional reconstruction algorithm. According to the fire-fighting non-destructive detection method provided by the invention, by integrating various technical means such as ultrasonic array scanning, X-ray tomography, thermal imaging analysis and microwave sealing detection, comprehensive and accurate identification of potential defects of fire-fighting equipment is realized, and the dimensionality and precision of defect detection are improved; the structural integrity of the equipment is kept to the maximum extent, extra risks caused by destructive detection are avoided, and the reliability and accuracy of detection are improved through a data fusion algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of fire protection engineering, and more particularly to a non-destructive fire protection detection method based on multi-modal detection. Background Art

[0002] The non-destructive detection technology of fire protection equipment is a key link to ensure the reliability of the fire safety system. Its core goal is to accurately identify potential defects of the equipment while maximizing the structural integrity. With the increasing complexity of modern fire protection equipment, traditional detection methods face severe challenges in terms of material adaptability, defect characterization dimensions, and dynamic monitoring capabilities.

[0003] Currently, the fire protection equipment detection technology generally has problems of multi-dimensional information loss and insufficient dynamic response. Traditional ultrasonic detection is limited by single-frequency scanning and is difficult to balance the identification of surface shallow defects and deep structural anomalies; although X-ray imaging can obtain internal structure information, conventional filtering algorithms are prone to cause edge blurring, affecting the determination accuracy of micro-defects; thermal imaging analysis mostly uses steady-state models and cannot effectively capture the transient temperature change characteristics during equipment operation; while microwave seal detection has limited positioning accuracy due to improper signal attenuation control. The existing technology lacks a deep fusion mechanism for multi-modal data, and the detection results are often independently interpreted, prone to misjudgment and missed detection. In addition, the detection system generally has problems such as redundant hardware architecture and poor environmental adaptability, and it is difficult to meet the rapid response requirements under complex working conditions. In response to this, we propose a non-destructive fire protection detection method based on multi-modal detection. Summary of the Invention

[0004] To solve the above technical problems and provide a non-destructive fire protection detection method based on multi-modal detection, the present technical solution solves the problems that the above ultrasonic detection is restricted by single-frequency scanning and cannot balance the identification of surface or deep defects; the conventional filtering algorithm of X-ray imaging causes edge blurring and reduces the determination accuracy of micro-defects; thermal analysis relies on steady-state models and is difficult to capture the transient temperature change characteristics during equipment operation; microwave seal detection has large positioning errors due to insufficient signal attenuation control, the technical data are independently interpreted without a fusion mechanism, the hardware architecture is redundant and the environmental adaptability is poor, prone to misjudgment and missed detection.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A non-destructive fire protection detection method based on multi-modal detection, comprising the following steps:

[0007] Rapidly scan the fire protection equipment through an ultrasonic array, analyze and locate the suspected defect area based on the time-domain reflection signal, and generate a three-dimensional coordinate mapping model;

[0008] Perform multi-angle X-ray projection acquisition on the positioning area and generate internal structure tomographic images through a three-dimensional reconstruction algorithm;

[0009] Collect surface temperature field distribution data during equipment operation and identify abnormal temperature gradient areas in combination with the unsteady heat conduction model;

[0010] Use a microwave signal source to detect the equipment sealing performance and analyze the leakage point location based on the reflection coefficient;

[0011] Through feature-level fusion and decision-level fusion algorithms, perform multi-modal comprehensive analysis on ultrasonic positioning data, X-ray tomographic images, thermal imaging temperature field data, and microwave detection results, and output a comprehensive defect determination result.

[0012] Preferably, the specific implementation process of quickly scanning the fire-fighting equipment through the ultrasonic array is as follows:

[0013] Use an ultrasonic array arranged in a 16×16 matrix to emit pulse signals with a frequency range of 0.5 - 5 MHz, and the pulse width is dynamically adjusted to 100 ns - 1 μs according to the thickness of the measured material;

[0014] The receiving end performs Hilbert transform on the reflected signal to extract the envelope, eliminates high-frequency noise interference, and calculates the signal energy density within the time window. Among them, the calculation formula for the energy density is:

[0015]

[0016] In the formula, E represents the energy density, t1 and t2 represent the start and end times of the defect echo time window, and s(t) represents the time-domain waveform;

[0017] When the ratio of the energy density of the defect area to the energy of the defect-free reference line exceeds the preset threshold, it is determined as a potential defect;

[0018] Through the conversion model from polar coordinates to Cartesian coordinates, convert the ultrasonic propagation time difference into three-dimensional space coordinates, where the calculation formula for the propagation time difference is:

[0019]

[0020] In the formula, Δt represents the propagation time difference, d represents the defect depth, and v represents the sound speed.

[0021] Preferably, the three-dimensional reconstruction process of X-ray tomography includes:

[0022] Apply Hamming window frequency domain filtering to the multi-angle projection data, where the cut-off frequency ω c Is set to 0.8 times the sampling frequency according to the detector resolution to suppress high-frequency noise while retaining image edge information;

[0023] After initially reconstructing the tomographic image through the inverse Radon transform, an algebraic iterative correction algorithm is used to optimize the image resolution, and its attenuation coefficient distribution is updated iteratively each time. The calculation formula for iteratively updating the attenuation coefficient distribution is as follows:

[0024]

[0025] In the formula, μ (k) represents the attenuation coefficient distribution at the k-th iteration, μ (k+1) represents the attenuation coefficient distribution at the (k + 1)-th iteration, λ represents the relaxation factor, with a value of 0.2, which is used to balance the convergence speed and stability, p i represents the i-th projection data, A i represents the projection matrix, represents the transpose of the projection matrix A i of, ||A i || 2 represents the square of the two-norm of the projection matrix A i ;

[0026] Finally, the defect area ratio is calculated by segmenting the image through the Otsu adaptive threshold algorithm. The calculation formula for the defect area ratio is as follows:

[0027] η = N d / N t

[0028] In the formula, η represents the defect area ratio, N d represents the number of defect pixels, N t represents the total number of pixels in the region.

[0029] Preferably, the specific method of thermal imaging-assisted analysis is as follows:

[0030] Establish a non-steady-state heat conduction partial differential equation:

[0031]

[0032] In the formula, ρ represents the material density, c represents the specific heat capacity, k represents the thermal conductivity, T represents the temperature, and q is the internal heat source during equipment operation;

[0033] The equation is discretely solved using the Crank-Nicolson difference scheme, and the time step is set to 1 second to match the sampling frequency of the infrared thermal imager;

[0034] The surface emissivity is corrected through the dual-band temperature measurement method, and the calculation formula is:

[0035] ∈ = (W1 - W2) / (B1(T) - B2(T))

[0036] Wherein, ∈ represents the surface emissivity, W1 represents the measured radiation energy in the 3-5μm band, W2 represents the measured radiation energy in the 8-12μm band, B1(T) represents the Planck blackbody radiation function in the 3-5μm band, and B2(T) represents the Planck blackbody radiation function in the 8-12μm band;

[0037] When the modulus of the temperature gradient exceeds the threshold value T corresponding to the coefficient of thermal expansion of the material g it is determined as an abnormal temperature rise region, where the calculation formula for the modulus of the temperature gradient is:

[0038]

[0039] Wherein, represents the modulus of the temperature gradient, represents the partial derivative of the temperature in the x direction, represents the partial derivative of the temperature in the y direction.

[0040] Preferably, the implementation steps of microwave seal detection include:

[0041] Design a circular waveguide structure, the inner diameter of which is:

[0042]

[0043] where D represents the inner diameter of the circular waveguide, λ represents the free space wavelength of the 2.45GHz microwave signal, and ∈ r represents the relative dielectric constant of the packaging material;

[0044] Measure the voltage standing wave ratio (VSWR) through a directional coupler and based on the formula:

[0045] Γ = (Z L - Z0) / (Z L + Z0)

[0046] calculate the reflection coefficient. In the formula, Γ represents the reflection coefficient, Z L represents the load impedance, and Z0 represents the characteristic impedance;

[0047] When |Γ| > 0.3, it is determined that there is a seal leak;

[0048] Use the time domain gate reflection method to locate the leak point, where the distance resolution is calculated according to In the formula, Δd represents the distance resolution, c represents the speed of light, B represents the signal bandwidth, with a value of 200MHz, and ∈ r represents the relative dielectric constant of the packaging material.

[0049] Preferably, the implementation logic of multi-modal data fusion is:

[0050] Normalize the original detection data at the data layer. The calculation formula for normalization is as follows:

[0051] X′ i =(X i -μ i ) / (3σ i )

[0052] In the formula, X′ i represents the normalized data, X i represents the original detection data, μ i represents the mean of each modal data, and σ i represents the standard deviation of each modal data;

[0053] Reduce the dimension through principal component analysis at the feature layer, and retain the principal components with a cumulative variance contribution rate greater than 85% to compress the data dimension;

[0054] Construct a basic probability assignment function using the D-S evidence theory at the decision layer. The weight assignment formula is as follows:

[0055]

[0056] In the formula, w i represents the weight of each modality, σ i represents the measurement error of each modality, which is obtained through historical calibration data, represents the sum of the squares of the measurement errors of all modalities, α represents the noise suppression coefficient, with a value of 0.1, and SNR i represents the signal-to-noise ratio calculated in real time by each sensor.

[0057] Preferably, the environmental protection materials and structural designs adopted include:

[0058] The substrate of the thermal imaging sensor uses a graphene-aluminum nitride composite material with a mass ratio of 3:7 and a thermal conductivity ≥ 180 W / (m·K), reducing the power consumption by 40% compared with traditional ceramic substrates;

[0059] The encapsulation of the microwave transmitter uses a polyether ether ketone (PEEK) composite, with a dielectric constant ∈ r = 3.2 ± 0.2 and a loss tangent value < 0.005, meeting the low-loss transmission requirements of high-frequency signals;

[0060] The X-ray detector is based on a CdTe / CZT heterojunction structure, with a bandgap of 1.44 eV and a carrier mobility > 1000 cm 2 / (V·s), and can achieve a photon absorption efficiency > 90% at room temperature.

[0061] Preferably, it includes a dynamic optimization mechanism:

[0062] Define the comprehensive confidence level as:

[0063] C = 1 - ∏(1 - w i c i )

[0064] In the formula, C represents the comprehensive confidence level, w i represents the weight of each modality, and c i represents the single-modal confidence score;

[0065] When C < 0.8, trigger parameter adaptive adjustment and optimize the ultrasonic frequency f according to the formula In the formula, f new represents the optimized ultrasonic frequency, f old represents the ultrasonic frequency before optimization, and Δf represents the step size, which is dynamically set according to the environmental noise spectrum, represents the partial derivative of the comprehensive confidence level C with respect to the ultrasonic frequency f;

[0066] The detection path planning adopts a genetic algorithm to synchronously optimize the detection accuracy and efficiency through a fitness function, and its fitness function is:

[0067] Fitness = 0.7C + 0.3(1 - t / t max )

[0068] In the formula, Fitness represents the fitness function value, C represents the comprehensive confidence level, t represents the current detection time, and t max represents the maximum allowable detection time;

[0069] The upper limit of the number of iterations is set to 100 times.

[0070] Preferably, the method for constructing and analyzing the defect database includes:

[0071] Define the multi-modal feature vector as:

[0072]

[0073] In the formula, V represents the multi-modal feature vector, E represents the ultrasonic energy density, Δμ represents the difference in X-ray attenuation coefficients, and |Γ| represents the absolute value of the microwave reflection coefficient;

[0074] Adopt an improved K-means algorithm for clustering analysis, and the distance metric function is:

[0075] d(V i , V j ) = ∑w k |V i,k - V j,k | 1.5

[0076] In the formula, d(V i , Vj ) represents the feature vector V i and V j the distance between, w k represents the feature weight, calculated by the entropy weight method, V i,k represents the feature vector V i the k-th component of, V j,k represents the feature vector V j the k-th component of;

[0077] The feature weight w k is calculated by the entropy weight method to reflect the differences in the importance of each dimension;

[0078] Defect classification is based on the support vector machine (SVM) model, and the Gaussian kernel is selected as the kernel function:

[0079] K(x i , x j ) = exp(-γ||x i - x j || 2 )

[0080] In the formula, K(x i , x j ) represents the value of the Gaussian kernel function, x i represents the sample vector, x j represents the sample vector, γ represents the kernel parameter, ||x i - x j || 2 represents the square of the Euclidean distance between the sample vectors x i and x j ;

[0081] The regularization parameter C and the kernel parameter γ are jointly optimized by grid search and cross-validation.

[0082] Preferably, the hardware architecture of the detection system includes:

[0083] The distributed data acquisition module realizes multi-device clock synchronization based on the IEEE1588 precise time protocol, and the time deviation is controlled within 1 μs;

[0084] The edge computing unit uses FPGA to parallel process ultrasonic time-domain signals and X-ray projection data, and the operation delay is less than 5 ms;

[0085] The 3D visualization engine renders the multi-modal fusion result based on OpenGL, supports the dynamic superposition display of tomographic images, temperature field distributions, and leakage points, and the rendering frame rate ≥ 30 fps;

[0086] The data security module uses the AES-256 encryption algorithm to transmit the detection data, and the key is updated every 24 hours through the Diffie-Hellman protocol.

[0087] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0088] The non-destructive fire detection method proposed by the present invention integrates various technical means such as ultrasonic array scanning, X-ray tomography, thermal imaging analysis, and microwave seal detection, realizing comprehensive and accurate identification of potential defects in fire-fighting equipment, improving the dimension and accuracy of defect detection, and maximizing the structural integrity of the equipment, avoiding additional risks brought by destructive detection. Through the data fusion algorithm, the problem of false judgment and missed detection caused by isolated interpretation of multi-modal data is effectively solved, improving the reliability and accuracy of detection. Through the dynamic optimization mechanism, it can adaptively adjust the detection parameters according to the real-time data feedback during the detection process, further optimizing the detection path, thereby significantly improving the detection efficiency while ensuring the detection accuracy. Brief Description of the Drawings

[0089] Figure 1 It is a flowchart of a non-destructive fire detection method based on multi-modal detection. Detailed Embodiment

[0090] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and those skilled in the art can think of other obvious variations.

[0091] Refer to Figure 1 As shown, the non-destructive fire detection method based on multi-modal detection includes the following implementation steps:

[0092] The first step: Ultrasonic array scanning and defect location

[0093] After the system is started, the ultrasonic array arranged in a 16×16 matrix quickly scans the surface of the fire-fighting equipment. Each probe in the array emits ultrasonic waves with a pulse signal of 0.5 - 5 MHz, and the pulse width is dynamically adjusted within the range of 100 ns to 1 μs according to the thickness of the material to be measured. After the receiving end captures the reflected signal, the signal envelope is extracted through Hilbert transform to eliminate high-frequency noise interference, and the energy density within a specific time window is calculated. When the energy density in the detection area exceeds 1.25 times the defect-free reference value, the system automatically marks it as a suspected defect area. Through the sound velocity model and time difference calculation, the ultrasonic propagation data in the polar coordinate system is converted into a three-dimensional Cartesian coordinate system to generate a spatial mapping model containing the defect position and depth, providing accurate positioning guidance for subsequent detection.

[0094] The second step: Multi-angle X-ray tomography

[0095] For the suspected area located by ultrasonic waves, the system controls the X-ray source to perform projection acquisitions from 8 symmetric azimuths. After the detector receives the penetration signal, the Hamming window frequency domain filtering technology is used to suppress high-frequency noise and retain key edge information. The three-dimensional tomographic image is reconstructed through the inverse Radon transform and the algebraic iterative optimization algorithm. During the iterative process, the attenuation coefficient distribution is dynamically adjusted to improve the resolution. After imaging is completed, based on the adaptive threshold segmentation technology, the defect boundary is automatically identified, the proportion of the defect area is calculated, and the abnormal area is marked. This process supports real-time display of tomographic slice images, and the operator can interactively adjust the detection angle and parameters through the human-machine interface.

[0096] Step 3: Dynamic thermal field gradient analysis

[0097] Under the condition that the device is powered on and running, a dual-band infrared thermal imager is used to synchronously collect the surface temperature field data. The system establishes an unsteady heat conduction model, solves the dynamic distribution of the temperature field through the time difference algorithm, and corrects the surface emissivity error of the material by using the dual-band radiation characteristics. When it is detected that the modulus of the temperature gradient exceeds the thermal expansion threshold of the material, it is automatically marked as an abnormal temperature rise area. The thermal imaging data is analyzed in real-time association with the device operation parameters (such as current, voltage), which can distinguish the normal working condition heating from the abnormal temperature rise caused by defects and avoid misjudgment.

[0098] Step 4: Microwave sealing detection and positioning

[0099] A 2.45 GHz microwave signal source and a circular waveguide structure are used to scan the sealed interface of the device. The standing wave ratio is measured through a directional coupler, and the existence of leakage points is judged by combining the reflection coefficient analysis. After the leakage signal is detected, the time-domain gate reflection method is started for precise positioning, and a 200 MHz broadband signal is used to achieve millimeter-level spatial resolution. The system automatically generates the three-dimensional coordinates of the leakage points and performs spatial matching with the ultrasonic positioning results to verify the consistency of the defect positions. For complex curved surface structures, the system dynamically adjusts the waveguide angle and the signal incident direction to ensure the detection coverage.

[0100] Step 5: Multimodal data fusion and decision-making

[0101] The data fusion module processes the detection results at three levels: at the data level, the ultrasonic energy density, the X-ray attenuation coefficient, the modulus of the temperature gradient, and the microwave reflection coefficient are normalized to eliminate the dimension difference; at the feature level, the key feature vectors are extracted through principal component analysis, the data dimension is compressed, and more than 85% of the effective information is retained; at the decision-making level, the evidence theory is used to fuse the confidence degrees of each modality, the weight coefficients are dynamically allocated, and the determination results of the defect types (such as surface cracks, internal cavities, sealing failures, etc.) are comprehensively output. The system automatically generates a quantitative report including the defect position, size, and risk level, and supports three-dimensional visualization to display the superposition effect of multimodal data.

[0102] Step 6: Dynamic Optimization and Adaptive Adjustment

[0103] The system is built with an intelligent optimization mechanism to calculate the comprehensive confidence index in real time. When the confidence level is lower than 0.8, parameter adjustment is automatically triggered: dynamically optimize the ultrasonic frequency to match the acoustic properties of the material, use the genetic algorithm to re-plan the detection path, and balance detection accuracy and efficiency. During the detection process, the system continuously collects the ambient noise spectrum and adaptively adjusts the parameters of the signal processing algorithm to ensure detection stability under complex working conditions. For historical detection data, the system automatically updates the defect feature database and optimizes the parameters of the classification model.

[0104] Step 7: System Hardware Cooperative Control

[0105] The distributed acquisition module realizes multi-sensor clock synchronization through the IEEE1588 protocol, and the time deviation is controlled within 1 microsecond. The FPGA edge computing unit processes ultrasonic time-domain signals and X-ray projection data in parallel to complete real-time feature extraction and preliminary analysis. The 3D visualization engine dynamically superimposes and renders data such as tomographic images and temperature fields, supporting interactive operations at 30 frames per second. The data security module adopts a dynamic key management mechanism to update the encryption key every 24 hours to ensure the security of detection data during transmission and storage.

[0106] Step 8: Output and Application of Detection Results

[0107] The system finally outputs three types of results: 1) a 3D atlas of defect distribution, marking the spatial coordinates and types of each defect; 2) a quantitative evaluation report, listing defect sizes, risk levels, and repair suggestions; 3) a detection process data packet, saving the original signals and intermediate processing results for recheck. The detection results are uploaded to the fire equipment management platform through the industrial Internet of Things interface, linked with the equipment life cycle database, and provide data support for predictive maintenance. For key facilities, the system supports regular automatic re-inspection and comparative analysis of historical data to realize the trend prediction of defect development.

[0108] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection required by the present invention is defined by the appended claims and their equivalents.

Claims

1. A non-destructive fire detection method based on multimodal detection, characterized in that It includes the following steps: Quickly scan the fire-fighting equipment through an ultrasonic array, analyze and locate the suspected defect area based on the time-domain reflection signal, and generate a three-dimensional coordinate mapping model; Implement multi-angle X-ray projection acquisition on the located area, and generate an internal structure tomographic image through a three-dimensional reconstruction algorithm; Collect the surface temperature field distribution data under the equipment operating state, and identify the temperature gradient abnormal area in combination with the unsteady heat conduction model; Use a microwave signal source to detect the equipment sealing performance, and analyze the leakage point location based on the reflection coefficient; Through the feature-level fusion and decision-level fusion algorithms, conduct multi-modal comprehensive analysis on the ultrasonic positioning data, X-ray tomographic image, thermal imaging temperature field data and microwave detection results, and output the comprehensive defect determination result.

2. The fire non-destructive detection method based on multi-modal detection according to claim 1, wherein The specific implementation process of quickly scanning the fire-fighting equipment through the ultrasonic array is as follows: Adopt an ultrasonic array arranged in a 16×16 matrix, emit pulse signals with a frequency range of 0.5 - 5 MHz, and dynamically adjust the pulse width to 100 ns - 1 μs according to the thickness of the measured material; The receiving end performs Hilbert transform on the reflection signal to extract the envelope line. After eliminating high-frequency noise interference, calculate the signal energy density within the time window. Among them, the calculation formula for the energy density is: In the formula, E represents the energy density, t1 and t2 represent the start and end moments of the defect echo time window, and s(t) represents the time-domain waveform; When the ratio of the energy density of the defect area to the energy of the defect-free reference line exceeds the preset threshold, it is determined as a potential defect; Through the conversion model from the polar coordinate system to the Cartesian coordinate system, convert the ultrasonic propagation time difference into three-dimensional space coordinates. Among them, the calculation formula for the propagation time difference is: In the formula, Δt represents the propagation time difference, d represents the defect depth, and v represents the sound speed.

3. The non-destructive fire detection method based on multi-modal detection according to claim 1, characterized in that The three-dimensional reconstruction process of X-ray tomography includes: Apply Hamming window frequency domain filtering to multi-angle projection data, where the cut-off frequency ω c is set to 0.8 times the sampling frequency according to the detector resolution, retaining the image edge information while suppressing high-frequency noise; After initially reconstructing the tomographic image through Radon inverse transform, use the algebraic iterative correction algorithm to optimize the image resolution, and update its attenuation coefficient distribution in each iteration. Among them, the calculation formula for updating the attenuation coefficient distribution in iteration is: where μ (k) represents the attenuation coefficient distribution at the k-th iteration, μ (k+1) represents the attenuation coefficient distribution at the (k + 1)-th iteration, λ represents the relaxation factor, with a value of 0.2, used to balance the convergence speed and stability, p i represents the i-th projection data, A i represents the projection matrix, represents the projection matrix A i is the transpose of, ∥A i ∥ 2 represents the square of the two-norm of the projection matrix A i ; Finally, segment the image through the Otsu adaptive threshold algorithm to calculate the proportion of the defect area. Among them, the calculation formula for the proportion of the defect area is: η = N d / N t Where η represents the proportion of the defect area, N d represents the number of defect pixels, and N t represents the total number of pixels in the region.

4. The non-destructive fire detection method based on multi-modal detection according to claim 1, characterized in that The specific method of thermal imaging-assisted analysis is: Establish an unsteady heat conduction partial differential equation: In the formula, ρ represents the material density, c represents the specific heat capacity, k represents the thermal conductivity, T represents the temperature, and q is the internal heat source during equipment operation; Use the Crank-Nicolson difference scheme to discretize and solve the equation, and set the time step to 1 second to match the sampling frequency of the infrared thermal imager; Calibrate the surface emissivity through the two-band temperature measurement method, and the calculation formula is: ∈=(W1-W2) / (B1(T)-B2(T)) In the formula, ∈ represents the surface emissivity, W1 represents the measured radiation energy in the 3 - 5 μm band, W2 represents the measured radiation energy in the 8 - 12 μm band, B1(T) represents the Planck blackbody radiation function in the 3 - 5 μm band, and B2(T) represents the Planck blackbody radiation function in the 8 - 12 μm band; When the modulus of the temperature gradient exceeds the threshold T corresponding to the coefficient of thermal expansion of the material g it is determined as an abnormal temperature rise region, where the calculation formula for the modulus of the temperature gradient is: In the formula, represents the modulus of the temperature gradient, represents the partial derivative of the temperature in the x direction, represents the partial derivative of the temperature in the y direction.

5. The fire non-destructive detection method based on multi-modal detection according to claim 1, characterized in that The implementation steps of microwave sealing detection include: Design a circular waveguide structure, and its inner diameter is: where D represents the inner diameter of the ring waveguide, λ represents the free-space wavelength of the 2.45 GHz microwave signal, and ∈ r represents the relative permittivity of the encapsulation material; Measure the voltage standing wave ratio (VSWR) through a directional coupler, and based on the formula: Γ=(Z L -Z0) / (Z L +Z0) Calculate the reflection coefficient, where Γ represents the reflection coefficient, Z L represents the load impedance, and Z0 represents the characteristic impedance; Determine that there is a seal leak when |Γ| > 0.3; The time-domain gate reflection method is used to locate the leakage point. Among them, the distance resolution is calculated according to , where Δd represents the distance resolution, c represents the speed of light, B represents the signal bandwidth, with a value of 200 MHz, and ∈ r represents the relative permittivity of the packaging material.

6. The non-destructive fire detection method based on multi-modal detection according to claim 1, characterized in that The implementation logic of multi-modal data fusion is as follows: Perform normalization processing on the original detection data at the data layer. Among them, the calculation formula for normalization processing is: X′ i =(X i -μ i ) / (3σ i ) where X′ i represents the normalized data, and X i represents the original detected data, μ i represents the mean of each modal data, and σ i represents the standard deviation of each modal data; Reduce the dimension through principal component analysis at the feature layer, and retain the principal components with a cumulative variance contribution rate greater than 85% to compress the data dimension; Adopt the D-S evidence theory to construct a basic probability assignment function at the decision layer, and the weight assignment formula is: where \(w\) i represents the weight of each modality, and \(\sigma\) i represents the measurement error of each modality, which is obtained from historical calibration data. represents the sum of the squares of the measurement errors of all modalities, \(\alpha\) represents the noise suppression coefficient, with a value of 0.1, and SNR i represents the signal-to-noise ratio calculated in real time by each sensor.

7. The non-destructive fire detection method based on multimodal detection according to claim 1, wherein The environmental protection materials and structural designs adopted include: The substrate of the thermal imaging sensor uses a graphene-aluminum nitride composite material with a mass ratio of 3:7, a thermal conductivity ≥ 180 W / (m·K), and the power consumption is reduced by 40% compared with the traditional ceramic substrate; The microwave transmitter package uses a polyetheretherketone (PEEK) composite, with a dielectric constant ∈ r = 3.2 ± 0.2 and a tangent of the loss angle < 0.005, meeting the requirement for low-loss transmission of high-frequency signals; The X-ray detector is based on the CdTe / CZT heterojunction structure, with a band gap of 1.44 eV and a carrier mobility > 1000 cm 2 / (V·s), and can achieve a photon absorption efficiency of > 90% at room temperature.

8. The non-destructive fire detection method based on multi-modal detection according to claim 1, wherein Include a dynamic optimization mechanism: Define the comprehensive confidence level as: C = 1 - ∏(1 - w i c i ) where C represents the comprehensive confidence level, and w i represents the weight of each modality, and c i represents the single-modal confidence score; When C < 0.8, trigger parameter adaptive adjustment according to the formula Optimize the ultrasonic frequency f, where f new Represents the optimized ultrasonic frequency, f old Represents the ultrasonic frequency before optimization, and Δf represents the step size, which is dynamically set according to the environmental noise spectrum. Represents the partial derivative of the comprehensive confidence C with respect to the ultrasonic frequency f; The detection path planning adopts a genetic algorithm, and synchronously optimizes the detection accuracy and efficiency through a fitness function. The fitness function is: Fitness=0.7C+0.3(1-t / t max ) where Fitness represents the fitness function value, C represents the comprehensive confidence level, t represents the current detection time, and t max represents the maximum allowable detection time; Set the upper limit of the number of iterations to 100 times.

9. The fire non-destructive detection method based on multi-modal detection according to claim 1, characterized in that, The construction and analysis methods of the defect database include: Define the multi-modal feature vector as: In the formula, V represents the multi-modal feature vector, E represents the ultrasonic energy density, Δμ represents the difference in X-ray attenuation coefficients, and |Γ| represents the absolute value of the microwave reflection coefficient; Adopt an improved K-means algorithm for clustering analysis, and the distance metric function is: d(V i ,V j ) = ∑w k |V i,k -V j,k | 1.5 where d(V i , V j ) represents the distance between the feature vectors V i and V j , w k represents the feature weight, calculated by the entropy weight method, V i,k represents the k-th component of the feature vector V i , and V j,k represents the k-th component of the feature vector V j ; Feature weight w k Calculated by the entropy weight method to reflect the differences in the importance of each dimension; Defect classification is based on a support vector machine (SVM) model, and the Gaussian kernel is selected as the kernel function: K(x i ,x j ) = exp(-γ ∥x i - x j ∥ 2 ) where K(x i , x j ) represents the value of the Gaussian kernel function, x i represents the sample vector, x j represents the sample vector, γ represents the kernel parameter, ∥x i - x j ∥ 2 represents the square of the Euclidean distance between the sample vectors x i and x j ; The regularization parameter C and the kernel parameter γ are jointly optimized through grid search and cross-validation.

10. The fire non-destructive detection method based on multi-modal detection according to claim 1, characterized in that, The hardware architecture of the detection system includes: The distributed data acquisition module realizes multi-device clock synchronization based on the IEEE1588 precise time protocol, and the time deviation is controlled within 1 μs; The edge computing unit uses an FPGA to parallel process ultrasonic time-domain signals and X-ray projection data, and the operation delay is less than 5 ms; The 3D visualization engine renders the multi-modal fusion result based on OpenGL, supports the dynamic overlay display of tomographic images, temperature field distributions, and leakage points, and the rendering frame rate ≥ 30 fps; The data security module uses the AES-256 encryption algorithm to transmit detection data, and the encryption key is updated once every 24 hours through the Diffie-Hellman protocol.

Images