Method and system for detecting living body in metal loading cabin

Abstract

The invention provides a living body detection method and system in a metal carrying cabin, and relates to the technical field of detection.The method comprises the steps that multi-source heterogeneous sensing data streams in the carrying cabin are synchronously collected through a multi-band microwave radar, an infrared thermal imaging array, a broadband ultrasonic sensor and a high-sensitivity micro-vibration sensor set which are deployed in the metal carrying cabin; according to a multi-source heterogeneous sensing data stream, a self-adaptive signal enhancement algorithm based on frequency domain sub-band decomposition is adopted, reverberation noise and multipath interference caused by metal cavity resonance are suppressed, and a coherence enhanced target signal cluster is output. According to the invention, the accuracy of living body detection is improved.

Description

Technical Field

[0001] The present invention relates to the field of detection technology, and in particular to a method and system for detecting living bodies in a metal carrier cabin. Background Art

[0002] Traditional liveness detection methods are mostly suitable for open or non-metallic environments, but some exhibit limitations within metal cargo compartments. For example, single-sensor detection technologies, such as microwave radar, are susceptible to reverberation noise and multipath interference generated by the resonance of metal cavities, causing the target signal to be submerged in the noise and making it difficult to accurately extract subtle vital signs such as breathing and heartbeat. Infrared thermal imaging technology can distort thermal images due to the strong heat reflection properties of metal surfaces, making it impossible to effectively distinguish live targets from the surrounding environment. Ultrasonic sensors are prone to standing wave interference caused by multiple reflections of sound waves within metal enclosures, leading to misjudgment of the live object's movement status.

[0003] Although multi-sensor fusion liveness detection solutions have emerged in existing technologies, several challenges remain in the unique scenario of metal capsules. For one thing, multi-source heterogeneous sensor data (such as microwave, infrared, ultrasonic, and vibration signals) exhibit varying attenuation characteristics and noise distribution patterns in metallic environments. Traditional signal enhancement algorithms struggle to adaptively suppress environmental noise, resulting in insufficient accuracy in extracting effective target signals. Furthermore, existing feature fusion methods often employ simple weighting or splicing schemes, failing to fully account for the varying reliability of different sensors in metallic environments. Consequently, they fail to highlight the contribution of key vital sign features, resulting in weak discriminative power in the combined feature vector. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for detecting living bodies in a metal carrier cabin, thereby improving the accuracy of living body detection.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] In a first aspect, a method for detecting living bodies in a metal carrier is provided, the method comprising:

[0007] Step 1: A multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier cabin synchronously collect multi-source heterogeneous sensor data streams inside the carrier cabin;

[0008] Step 2: Based on the multi-source heterogeneous sensor data stream, an adaptive signal enhancement algorithm based on frequency domain subband decomposition is used to suppress the reverberation noise and multipath interference caused by the resonance of the metal cavity, and output a coherently enhanced target signal cluster;

[0009] Step 3: Based on the coherently enhanced target signal cluster, the respiratory waveform envelope features, cardiac ballistic spectrum features, and movement pattern time-frequency features of the living target are extracted through time-frequency analysis technology, and a feature fusion method with weighted attention mechanism is used to generate a joint feature vector;

[0010] Step 4: Based on the joint eigenvector, the energy distribution characteristics of the living targets in the cargo hold are analyzed, and a multi-sensor coverage confidence matrix is ​​dynamically constructed. The confidence matrix is ​​subjected to singular value decomposition, and combined with an adaptive swarm division of labor optimization algorithm, the sensor weight distribution parameters are used as the swarm's foraging targets. Each bee represents a set of candidate weight parameters. Leveraging the swarm's division of labor and collaboration, the search strategy is dynamically adjusted based on the energy distribution characteristics of the living targets in the cargo hold. Through iterative optimization, the weight parameters are converged to the final solution, generating the sensor weight distribution parameters.

[0011] Step 5: Based on the sensor weight distribution parameters and the joint feature vector, a metal environment adaptive decision model is constructed to adaptively adjust the sensor sampling strategy and output the probability of living body presence and quantitative indicators of vital signs.

[0012] Furthermore, a multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier can synchronously collect multi-source heterogeneous sensor data streams inside the carrier, including:

[0013] Step 1.1: Generate a three-dimensional coordinate mapping table based on the structural parameters of the metal cargo cabin. Deploy the sensor network in a layout where the microwave radar covers the longitudinal section, the infrared array covers the horizontal section, the ultrasonic sensor covers the corner blind spots, and the micro-vibration sensor group fits the cabin wall.

[0014] In step 1.2, based on the three-dimensional spatial coordinate mapping table, the multi-sensor hardware synchronization trigger mechanism is activated. A global timestamp is generated using a high-precision clock source. This ensures that the microwave radar pulse emission phase, infrared array exposure period, ultrasonic emission sequence, and vibration sampling time are strictly aligned in the time domain to obtain the raw sensor data.

[0015] Step 1.3: Receive the raw sensor data and, based on the sensor position parameters recorded in the 3D spatial coordinate mapping table, add the cabin material attenuation compensation coefficient to the microwave radar signal, the ambient temperature drift correction parameter to the infrared thermal imaging data, the air density compensation factor to the ultrasonic signal, and the mechanical coupling gain coefficient to the micro-vibration signal, to generate a primary data stream with physical compensation.

[0016] In step 1.4, the primary data stream is position-matched according to the three-dimensional spatial coordinate mapping table, where microwave radar data is mapped to the polar coordinate system, infrared data is mapped to the Cartesian coordinate system, ultrasonic data is mapped to the spherical coordinate system, and micro-vibration data is mapped to the structural grid coordinate system, and the multi-source heterogeneous sensor data stream with temporal and spatial correlation is output.

[0017] Furthermore, based on the multi-source heterogeneous sensor data stream, an adaptive signal enhancement algorithm based on frequency domain subband decomposition is used to suppress the reverberation noise and multipath interference caused by the resonance of the metal cavity, and output a coherently enhanced target signal cluster, including:

[0018] Step 2.1: Based on the multi-source heterogeneous sensor data stream, after performing millimeter-wave penetration compensation on the microwave radar signal, the full frequency band is decomposed into multiple sub-bands based on the resonant mode distribution of the metal carrier. Using this decomposition result, the ultrasonic signal is divided into critical frequency bands based on the attenuation coefficient of the sound wave in the metal medium. Furthermore, based on the structural transfer function of the carrier, the micro-vibration signal is grouped into octaves to match the mechanical vibration propagation characteristics.

[0019] In step 2.2, based on the results of the sub-band, frequency band, and group division, a coherence matrix across sensor channels is constructed within each sub-band, and the reverberation noise subspace is separated by eigenvalue decomposition. Sub-band weight coefficients are dynamically generated, with the phase continuity of the target signal cluster between adjacent sub-bands as the constraint.

[0020] Step 2.3: Based on the weighted sub-band signals, extract the time delay characteristic peaks formed by multiple reflections on the metal wall, calculate the anti-phase cancellation beam to cancel the multipath interference; use spatiotemporal adaptive filtering to eliminate the metal surface mirror reflection artifacts; and output the sub-band signals after suppressing the interference.

[0021] In step 2.4, each subband signal is reconstructed into a time domain signal cluster through inverse transformation, and the mutual information entropy value between the reconstructed signal and the original sensor data is calculated; if the entropy value is lower than the dynamic threshold, the coherently enhanced target signal cluster is output; otherwise, the data re-acquisition mechanism is triggered for closed-loop correction.

[0022] Furthermore, based on the coherently enhanced target signal cluster, the respiratory waveform envelope features, cardiac ballistometry features, and motion pattern time-frequency features of the living target are extracted through time-frequency analysis technology, and a feature fusion method with weighted attention mechanism is used to generate a joint feature vector, including:

[0023] Step 3.1: Based on the coherently enhanced target signal cluster, extract the microwave radar and micro-vibration signals. Separate the respiratory harmonic components using time-frequency analysis techniques, and use envelope detection and adaptive baseline drift correction to generate the time-domain periodic eigenvector and frequency-domain main harmonic energy ratio eigenvector of the respiratory waveform envelope.

[0024] Step 3.2: Based on the respiratory feature extraction results and the corresponding signal segments, the heartbeat harmonic reconstruction technique is used to separate the cardiac ballistomic component from the microwave radar signal. Simultaneously, the microvibration signal is frequency-domain compensated in combination with the cabin structure vibration transfer characteristics to extract the fundamental frequency stability eigenvector and harmonic complexity eigenvector of the cardiac ballistomic spectrum.

[0025] Step 3.3: Based on the ballistocardiographic feature extraction results and the processed signal clusters, the infrared thermal imaging signal is dynamically segmented to eliminate metal heat reflection interference and extract the time-frequency feature vector of the hot spot displacement trajectory. Simultaneously, the output ultrasonic signal is analyzed for the short-term zero-crossing rate characteristics of its Doppler frequency shift to generate a motion pattern classification feature vector.

[0026] In step 3.4, a feature quality assessment matrix is ​​constructed based on the respiratory feature vector, cardiac impact feature vector, and motion feature vector. The respiratory feature weight, cardiac impact feature weight, and motion feature weight are dynamically calculated based on the intra-class discreteness of the respiratory feature, the inter-class separability of the cardiac impact feature, and the confidence of the motion feature. The weighted feature vectors are concatenated into a joint feature vector and normalized.

[0027] Furthermore, based on the joint eigenvector, the energy distribution characteristics of the living target in the cabin are analyzed, and a multi-sensor coverage confidence matrix is ​​dynamically constructed, including:

[0028] Step 4.1: Receive the joint feature vector, extract the respiratory waveform envelope features, the ballistocardiogram features, and the time-frequency feature components of the motion pattern, and map these feature components to the three-dimensional discretized grid nodes of the metal cabin to generate feature distribution data with spatial location annotations.

[0029] Step 4.2: Based on the spatially labeled feature distribution data, dynamic thermal data reflecting the spatial energy density of the living target is generated by calculating the temporal stability index and spatial continuity metric of the feature amplitude of each grid node;

[0030] Step 4.3: Calculate the coverage coefficient of each sensor for each spatial grid node using the distribution information of each energy peak area in the dynamic thermal data, combined with the preset sensor location parameters and the physical law of signal attenuation;

[0031] In step 4.4, the capability coefficient is normalized, and the initial coverage confidence matrix is ​​constructed with the spatial grid nodes as row indices and the sensor channels as column indices. The matrix element value represents the trustworthy monitoring weight of the corresponding sensor at a specific location.

[0032] In step 4.5, according to the metal shielding compensation parameters provided by the cabin structure preset database, the initial coverage confidence matrix is ​​corrected for multipath reflection loss, and the optimized multi-sensor coverage confidence matrix is ​​output.

[0033] Furthermore, the confidence matrix is ​​subjected to singular value decomposition, and combined with an adaptive swarm division of labor optimization algorithm, the sensor weight distribution parameters are used as the swarm's foraging target. Each bee represents a set of candidate weight parameters. By utilizing the swarm's division of labor and cooperation, the search strategy is dynamically adjusted in combination with the energy distribution characteristics of the living organisms in the cabin. Through iterative optimization, the weight parameters converge to the final solution, generating the sensor weight distribution parameters, including:

[0034] Step 4.6, perform a singular value decomposition operation on the multi-sensor coverage confidence matrix to extract the principal component vectors representing the sensor spatial coverage characteristics and their corresponding singular value spectrum distributions;

[0035] Step 4.7: Determine the initial search boundary of the sensor weight parameters based on the characteristic peak interval of the singular value spectrum distribution, and map each weight parameter combination to the position coordinates of individual bees in the swarm optimization;

[0036] Step 4.8: Based on the cabin space coverage characteristics reflected by the principal component vector and the spatial gradient variation of the living body energy distribution, the search area is divided into a high-gradient fine search area and a low-gradient wide exploration area to obtain the partitioning results;

[0037] In step 4.9, the roles of the bee colony are dynamically configured according to the partitioning results. Scout bees are deployed to high-gradient areas to perform neighborhood depth optimization; follower bees are deployed to low-gradient areas to perform heat tracking exploration; after each round of iteration, the regional search tasks are redistributed based on the fitness of the weight parameters. When the change amplitude of the weight parameters in consecutive iterations is lower than the preset convergence threshold, the weight distribution parameters corresponding to the current optimal bee position coordinates are extracted.

[0038] Furthermore, based on the sensor weight distribution parameters and joint feature vectors, a metal environment adaptive decision model is constructed to adaptively adjust the sensor sampling strategy and output the probability of living body presence and quantitative indicators of vital signs, including:

[0039] Step 5.1: Receive the sensor weight distribution parameters and inject them into the decision model as the input layer neuron initialization weights, and load the joint feature vector into the model feature input channel;

[0040] Step 5.2: Based on the constructed initialization decision model, the historical operating condition database of the cargo hold is used to construct a training sample set. The hidden layer structure parameters of the model are fine-tuned through the transfer learning mechanism to adapt the model to the characteristic distribution offset in the metal cavity environment, thereby generating a metal environment-adaptive decision model.

[0041] Step 5.3: Based on the metal environment adaptive decision model, the activation state of its hidden layer is monitored in real time. According to the spatial distribution pattern of the activation state, sensor resource allocation instructions are generated. The microwave radar scanning frame rate, infrared array resolution, ultrasonic transmission power, and vibration sensor sensitivity levels are dynamically allocated according to the weight parameter ratio.

[0042] In step 5.4, new sensor data streams are collected according to the adaptive sampling strategy. The real-time data is input into the metal environment adaptive decision model. The model parameters are updated through the online incremental learning mechanism to output the probability of the presence of living organisms in the cabin and the quantitative indicators of vital signs such as respiratory rate, heart rate variability, and body movement intensity.

[0043] Step 5.5: When the probability value of the living body exceeds the preset threshold, the vital sign quantitative indicators corresponding to the probability value are analyzed to generate a three-dimensional parameter analysis report including respiratory rate abnormality, heart rate variability stability, and body movement intensity trend.

[0044] In a second aspect, a liveness detection system in a metal carrier includes:

[0045] The acquisition module is used to synchronously collect multi-source heterogeneous sensor data streams within the metal carrier cabin through a multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier cabin. It is used to suppress the reverberation noise and multipath interference caused by the resonance of the metal cavity and output a coherently enhanced target signal cluster based on the adaptive signal enhancement algorithm based on frequency domain sub-band decomposition based on the multi-source heterogeneous sensor data streams.

[0046] A processing module is used to extract the respiratory waveform envelope features, cardiac ballistometry features, and motion pattern time-frequency features of the living target based on the coherently enhanced target signal cluster through time-frequency analysis technology, and to generate a joint feature vector using a feature fusion method weighted by an attention mechanism;

[0047] The computational module is used to analyze the energy distribution characteristics of living targets within the cargo hold based on the joint eigenvectors and dynamically construct a multi-sensor coverage confidence matrix. The confidence matrix is ​​subjected to singular value decomposition and, in conjunction with an adaptive swarm division-of-labor optimization algorithm, the sensor weight allocation parameters are used as the swarm's foraging targets. Each bee represents a set of candidate weight parameters. Leveraging the swarm's division of labor and collaboration, the search strategy is dynamically adjusted in conjunction with the energy distribution characteristics of living objects within the cargo hold. Through iterative optimization, the weight parameters converge to the final solution, generating the sensor weight allocation parameters.

[0048] The intelligent decision-making module is used to build a metal environment-adaptive decision-making model based on sensor weight allocation parameters and joint feature vectors, adaptively adjust the sensor sampling strategy, and output the probability of living organisms and quantitative indicators of vital signs.

[0049] According to a third aspect, a computing device includes:

[0050] one or more processors;

[0051] The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0052] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0053] The above solution of the present invention includes at least the following beneficial effects:

[0054] Through the coordinated deployment and synchronous data acquisition of a multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensors, and a high-sensitivity micro-vibration sensor array, combined with three-dimensional spatial coordinate mapping and physical compensation mechanisms, the system effectively addresses the limited coverage and insufficient data correlation of a single sensor within the metal carrier, ensuring the spatiotemporal consistency and accuracy of multi-source heterogeneous sensor data streams. An adaptive signal enhancement algorithm based on frequency domain subband decomposition specifically suppresses reverberation noise and multipath interference caused by metal cavity resonance, significantly improving the coherence of the target signal. A feature fusion method weighted by an attention mechanism accurately extracts key vital signs, such as respiration, cardiac arrest, and movement, from the enhanced signal and dynamically assigns weights, enhancing the discernibility of vital sign information. By constructing a multi-sensor coverage confidence matrix and combining singular value decomposition with an adaptive swarm optimization algorithm to iteratively optimize sensor weight parameters, the system achieves intelligent allocation of sensor resources and synergistic efficiency. A metal environment-adaptive decision model adaptively adjusts the sampling strategy, adapts to metal environment feature shifts through transfer learning and online incremental learning, and accurately outputs the probability of living organisms. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a method for detecting living bodies in a metal carrier provided by an embodiment of the present invention.

[0056] Figure 2 Schematic diagram of a liveness detection system in a metal carrier provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0058] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting living bodies in a metal carrier, the method comprising the following steps:

[0059] Step 1: A multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier cabin synchronously collect multi-source heterogeneous sensor data streams inside the carrier cabin;

[0060] Step 2: Based on the multi-source heterogeneous sensor data stream, an adaptive signal enhancement algorithm based on frequency domain subband decomposition is used to suppress the reverberation noise and multipath interference caused by the resonance of the metal cavity, and output a coherently enhanced target signal cluster;

[0061] Step 3: Based on the coherently enhanced target signal cluster, the respiratory waveform envelope features, cardiac ballistic spectrum features, and movement pattern time-frequency features of the living target are extracted through time-frequency analysis technology, and a feature fusion method with weighted attention mechanism is used to generate a joint feature vector;

[0062] Step 4: Based on the joint eigenvector, the energy distribution characteristics of the living targets in the cargo hold are analyzed, and a multi-sensor coverage confidence matrix is ​​dynamically constructed. The confidence matrix is ​​subjected to singular value decomposition, and combined with an adaptive swarm division of labor optimization algorithm, the sensor weight distribution parameters are used as the swarm's foraging targets. Each bee represents a set of candidate weight parameters. Leveraging the swarm's division of labor and collaboration, the search strategy is dynamically adjusted based on the energy distribution characteristics of the living targets in the cargo hold. Through iterative optimization, the weight parameters are converged to the final solution, generating the sensor weight distribution parameters.

[0063] Step 5: Based on the sensor weight distribution parameters and the joint feature vector, a metal environment adaptive decision model is constructed to adaptively adjust the sensor sampling strategy and output the probability of living body presence and quantitative indicators of vital signs.

[0064] In an embodiment of the present invention, efficient and accurate liveness detection is achieved through multi-dimensional technical means. The collaborative acquisition of multi-source heterogeneous sensors ensures comprehensive coverage of the space within the cabin and the spatiotemporal correlation of data, solving the detection limitations of a single sensor in a metal environment. The signal enhancement algorithm based on frequency domain subband decomposition effectively suppresses the reverberation noise and multipath interference caused by metal resonance, thereby improving the clarity of the target signal. The feature fusion weighted by the attention mechanism enhances the recognition of key liveness features such as breathing, cardiac arrest, and movement, avoiding the interference of invalid features. The dynamic optimization of sensor weights achieves efficient collaboration of multi-source data through the swarm algorithm. The final intelligent decision-making model combines adaptive sampling strategies with online learning to achieve accurate detection of the entire process from liveness recognition to state assessment in the metal cabin.

[0065] In a preferred embodiment of the present invention, the above step 1 may include:

[0066] Step 1.1: Generate a three-dimensional coordinate mapping table based on the structural parameters of the metal cargo cabin. Deploy the sensor network in a layout where the microwave radar covers the longitudinal section, the infrared array covers the horizontal section, the ultrasonic sensor covers the corner blind spots, and the micro-vibration sensor group fits the cabin wall.

[0067] In step 1.2, based on the three-dimensional spatial coordinate mapping table, the multi-sensor hardware synchronization trigger mechanism is activated. A global timestamp is generated using a high-precision clock source. This ensures that the microwave radar pulse emission phase, infrared array exposure period, ultrasonic emission sequence, and vibration sampling time are strictly aligned in the time domain to obtain the raw sensor data.

[0068] Step 1.3: Receive the raw sensor data and, based on the sensor position parameters recorded in the 3D spatial coordinate mapping table, add the cabin material attenuation compensation coefficient to the microwave radar signal, the ambient temperature drift correction parameter to the infrared thermal imaging data, the air density compensation factor to the ultrasonic signal, and the mechanical coupling gain coefficient to the micro-vibration signal, to generate a primary data stream with physical compensation.

[0069] In step 1.4, the primary data stream is position-matched according to the three-dimensional spatial coordinate mapping table, where microwave radar data is mapped to the polar coordinate system, infrared data is mapped to the Cartesian coordinate system, ultrasonic data is mapped to the spherical coordinate system, and micro-vibration data is mapped to the structural grid coordinate system, and the multi-source heterogeneous sensor data stream with temporal and spatial correlation is output.

[0070] In this embodiment of the present invention, scientific layout and precise processing establish a high-quality data foundation for liveness detection within metal cargo cabins. Multiple sensor types are deployed according to the structural characteristics of the cargo cabin, and combined with a three-dimensional coordinate mapping table, they achieve full coverage of the vertical, horizontal, corner, and cabin walls, eliminating detection blind spots. A high-precision clock synchronization mechanism ensures strict time domain alignment among multiple sensors, preventing data time deviations. Physical compensation coefficients such as material attenuation and temperature drift are added to address the characteristics of the metal environment to correct signal distortion. Finally, data from different coordinate systems is uniformly mapped into a spatiotemporal correlation framework, enabling the organic integration of multi-source heterogeneous data. This overall process ensures both comprehensive data collection and spatiotemporal consistency.

[0071] In an embodiment of the present invention, the specific steps include:

[0072] Step 1.1: Based on the structural parameters of the metal cargo cabin, such as length, width, height, and bulkhead thickness, construct a three-dimensional spatial model of the interior of the cargo cabin, divide the space into grid cells, and generate a three-dimensional spatial coordinate mapping table that records the coordinates of each spatial point. According to the principle of "microwave radar covers the longitudinal section, infrared array covers the horizontal section, ultrasonic wave fills the blind spot, and micro-vibration is attached to the bulkhead," calculate the optimal installation coordinates of each sensor (for example, microwave radar is installed on the longitudinal axis of the middle section of the cabin, infrared array is installed at the top center, etc.) and write them into the mapping table.

[0073] Step 1.2: Generate a global time base using a high-precision clock source (such as a 10MHz oven-controlled crystal oscillator) to calculate the triggering time of each sensor:

[0074] The trigger time of the microwave radar pulse transmission phase = global timestamp + radar channel delay compensation value;

[0075] The starting time of the infrared array exposure cycle = global timestamp + infrared detector response delay value;

[0076] The start time of the ultrasonic emission sequence = global timestamp + sound wave propagation warm-up time;

[0077] The sampling time of the micro-vibration sensor = the global timestamp (no additional delay, because the vibration needs to be captured in real time).

[0078] The above delay compensation values ​​are pre-determined through calibration experiments to ensure that the time domain alignment error of each sensor is ≤10 microseconds.

[0079] Step 1.3, cabin material attenuation compensation coefficient (microwave radar): First, multiply the material attenuation constant, the square root of the microwave frequency, the square root of the cabin metal electromagnetic permeability, and the path length of the microwave signal penetrating the metal to obtain a product. Then divide this product by 20 to obtain a new value. Finally, use 10 as the base and this new value as the exponent to perform a power operation. The result of the operation is the cabin material attenuation compensation coefficient.

[0080] Ambient temperature drift correction parameter (infrared thermal imaging): First, calculate the difference between the real-time ambient temperature and the reference temperature. Multiply this difference by the infrared sensor's temperature drift coefficient to obtain a product. Then, add 1 to this product to obtain the final value, which is the ambient temperature drift correction parameter.

[0081] Ultrasonic air density compensation factor: Calculate the real-time air density and divide it by the air density under standard conditions (1.225kg / m³). The quotient is the ultrasonic air density compensation factor.

[0082] Mechanical coupling gain coefficient (micro-vibration sensor): The mechanical coupling gain coefficient is obtained by multiplying the coupling coefficient, the contact area between the sensor and the bulkhead, and the elastic modulus of the coupling material.

[0083] Step 1.4: Map microwave radar data to polar coordinates:

[0084] The coordinates of the polar coordinate system are expressed as (r, θ), where r is the radar-measured target distance after compensation for cabin material attenuation; θ is the radar scanning angle directly obtained by the radar motor angle sensor.

[0085] Infrared thermal imaging data is mapped to a Cartesian coordinate system: Cartesian coordinates are represented by (x, y). x is calculated as (the column number of the target pixel in the infrared array minus the center column number of the array) multiplied by the size of a single pixel.

[0086] y is calculated as: (the row number of the target pixel in the infrared array minus the row number of the center of the array) multiplied by the size of a single pixel;

[0087] The above calculation results need to be combined with the installation position of the infrared array to perform coordinate offset correction, and finally obtain the coordinates in the Cartesian coordinate system.

[0088] Ultrasonic sensor data is mapped to a spherical coordinate system: the coordinates of the spherical coordinate system are represented by (r, φ, θ). Here, r is the ultrasonic distance value after air density compensation; φ is the horizontal rotation angle of the ultrasonic sensor; and θ is the vertical elevation angle of the ultrasonic sensor.

[0089] Mapping micro-vibration sensor data to a structural grid coordinate system: Coordinates in the structural grid coordinate system are represented by (i, j), where i corresponds to the grid row index and j corresponds to the grid column index. The specific coordinate value is determined by the structural grid node index corresponding to the sensor installation location. The signal value collected by the micro-vibration sensor is directly mapped to the grid node corresponding to (i, j).

[0090] In a preferred embodiment of the present invention, the above step 2 may include:

[0091] Step 2.1: Based on the multi-source heterogeneous sensor data stream, after performing millimeter-wave penetration compensation on the microwave radar signal, the full frequency band is decomposed into multiple sub-bands based on the resonant mode distribution of the metal carrier. Using this decomposition result, the ultrasonic signal is divided into critical frequency bands based on the attenuation coefficient of the sound wave in the metal medium. Furthermore, based on the structural transfer function of the carrier, the micro-vibration signal is grouped into octaves to match the mechanical vibration propagation characteristics.

[0092] In step 2.2, based on the results of the sub-band, frequency band, and group division, a coherence matrix across sensor channels is constructed within each sub-band, and the reverberation noise subspace is separated by eigenvalue decomposition. Sub-band weight coefficients are dynamically generated, with the phase continuity of the target signal cluster between adjacent sub-bands as the constraint.

[0093] Step 2.3: Based on the weighted sub-band signals, extract the time delay characteristic peaks formed by multiple reflections on the metal wall, calculate the anti-phase cancellation beam to cancel the multipath interference; use spatiotemporal adaptive filtering to eliminate the metal surface mirror reflection artifacts; and output the sub-band signals after suppressing the interference.

[0094] In step 2.4, each subband signal is reconstructed into a time domain signal cluster through inverse transformation, and the mutual information entropy value between the reconstructed signal and the original sensor data is calculated; if the entropy value is lower than the dynamic threshold, the coherently enhanced target signal cluster is output; otherwise, the data re-acquisition mechanism is triggered for closed-loop correction.

[0095] In an embodiment of the present invention, by specifically addressing problems such as resonant reverberation and multipath interference unique to metal carriers, precise frequency band division is first performed based on the propagation characteristics of different sensor signals in a metal environment (such as microwave resonant modes, ultrasonic attenuation laws, and vibration propagation characteristics) to ensure that signal processing is compatible with the metal cavity environment; then, noise is separated through coherence analysis, and sub-band signals are dynamically weighted using phase continuity to enhance the consistency of the target signal; then, anti-phase cancellation and space-time filtering are used to specifically eliminate multipath reflections and mirror artifacts, thereby suppressing interference from the metal environment; finally, mutual information entropy verification and a closed-loop correction mechanism are used to ensure the reliability of the output signal.

[0096] In an embodiment of the present invention, the specific steps include:

[0097] Step 2.1, millimeter wave penetration compensation: First, multiply the millimeter wave penetration attenuation coefficient in metal, the metal bulkhead thickness, and the millimeter wave frequency to obtain the product. Then, divide this product by 20 to obtain a new value. Finally, raise this new value to the power of 10 to obtain the penetration compensation coefficient. Multiply the original radar signal by this penetration compensation coefficient to obtain the compensated signal.

[0098] Microwave sub-band decomposition: The nth-order resonant frequency is calculated by multiplying the positive integer n by the propagation velocity of electromagnetic waves in metals. The product is then divided by (twice the characteristic dimension of the carrier cavity). The quotient is the nth-order resonant frequency. The full microwave frequency band is divided into multiple sub-bands, with adjacent resonant frequencies as the boundaries, so that each sub-band avoids the frequency band where the metal carrier resonant interference is concentrated.

[0099] Ultrasonic critical frequency band division: Multiply the dielectric constant by the kth power of the ultrasonic frequency (k is approximately equal to 2 in metals), and the product obtained is the attenuation coefficient.

[0100] The critical frequency band boundary (f_c) is calculated by dividing the maximum acceptable attenuation value (α_max) by the dielectric constant. The resulting quotient is then raised to the power of k, and the result is the critical frequency band boundary. The frequency range below this critical frequency band boundary is classified as the effective frequency band of ultrasound.

[0101] Micro-vibration octave grouping: Based on the cabin structure transfer function H(f), the first peak frequency of this function is found as the fundamental frequency (f0). Starting from this fundamental frequency, the frequency range of the micro-vibration signal is divided into multiple octave intervals. The upper frequency limit of each interval is twice the lower frequency limit. For example, the first interval is [f0, 2f0], the second interval is [2f0, 4f0], and so on. Each frequency interval is matched to the propagation characteristics of mechanical vibration in the cabin structure.

[0102] Step 2.2, coherence matrix construction: First, calculate the expected cross-correlation of the two sensor channel signals; then calculate the standard deviation of the two channel signals and multiply them; finally, divide the expected cross-correlation by the product of the standard deviations. The result is the (i, j)th element of the coherence matrix, which reflects the correlation between the two channel signals.

[0103] Noise subspace separation: Perform eigenvalue decomposition on the coherence matrix to obtain a series of eigenvalues, which are then sorted from largest to smallest. A threshold is set at 1 / 10 of the maximum eigenvalue. The eigenvectors corresponding to all eigenvalues ​​less than this threshold are selected. The space formed by these eigenvectors is the reverberation noise subspace.

[0104] Subband weight coefficient generation: The method for calculating the phase difference between adjacent subband signals is: take the phase values ​​of the center frequencies of two adjacent subbands, calculate their absolute difference, and obtain the phase difference.

[0105] The subband weight coefficient is calculated by adding 1 to the phase difference to obtain a value. This value is then divided by 1 to obtain the result. The smaller the phase difference, the larger the subband weight coefficient, ensuring that subbands with continuous phases receive higher weights.

[0106] Step 2.3, extracting the time-delay characteristic peak: First, calculate the integral product of the signal and its own signal after a delay of τ. Then, calculate the energy of these two signals, multiply them, and take the square root. Finally, divide the integral product by the square root of the energy product to obtain the cross-correlation function. The delay time τ corresponding to the peak of the cross-correlation function is the time-delay characteristic peak of the metal wall reflection.

[0107] To calculate the inverted beam cancellation, first calculate the product of the imaginary unit j and 2π, the signal frequency f, and the reflection delay τ. Then, minus the sign of this product and raise it to the power of the natural constant e. The result is the inverted beam weight. This weight is added to the original signal to cancel the multipath reflection signal.

[0108] Spatiotemporal adaptive filtering: First, the inverse of the noise covariance matrix is ​​calculated. This inverse matrix is ​​then multiplied by the desired signal direction vector to obtain the filter weight vector. This vector is used to filter artifacts generated by specular reflections on metal surfaces.

[0109] Step 2.4, time domain signal reconstruction: ,in, It is The frequency domain signal of the sub-bands, is the subband index, ranging from 0 to N-1, is the time variable, is an imaginary unit, is a natural constant, is the total number of subbands.

[0110] Mutual information entropy calculation: For all possible signal value probabilities, calculate the product of their logarithms with base 2; add all these products and take the negative value, and the result is the entropy of the signal.

[0111] The method for calculating mutual information entropy is: add the entropy of the original signal and the entropy of the reconstructed signal, and then subtract the joint entropy of the original signal and the reconstructed signal. The result is the mutual information entropy.

[0112] Dynamic Threshold Comparison: First, calculate the mean of the historical mutual information entropy. Then, calculate the standard deviation of the historical mutual information entropy and multiply it by 2. Finally, add the mean and 2 times the standard deviation to obtain the result as the dynamic threshold. The current mutual information entropy is compared with the dynamic threshold. If it is lower than the threshold, a coherent enhancement signal is output; otherwise, the data re-acquisition mechanism is triggered.

[0113] In a preferred embodiment of the present invention, the above step 3 may include:

[0114] Step 3.1: Based on the coherently enhanced target signal cluster, extract the microwave radar and micro-vibration signals. Separate the respiratory harmonic components using time-frequency analysis techniques, and use envelope detection and adaptive baseline drift correction to generate the time-domain periodic eigenvector and frequency-domain main harmonic energy ratio eigenvector of the respiratory waveform envelope.

[0115] Step 3.2: Based on the respiratory feature extraction results and the corresponding signal segments, the heartbeat harmonic reconstruction technique is used to separate the cardiac ballistomic component from the microwave radar signal. Simultaneously, the microvibration signal is frequency-domain compensated in combination with the cabin structure vibration transfer characteristics to extract the fundamental frequency stability eigenvector and harmonic complexity eigenvector of the cardiac ballistomic spectrum.

[0116] Step 3.3: Based on the ballistocardiographic feature extraction results and the processed signal clusters, the infrared thermal imaging signal is dynamically segmented to eliminate metal heat reflection interference and extract the time-frequency feature vector of the hot spot displacement trajectory. Simultaneously, the output ultrasonic signal is analyzed for the short-term zero-crossing rate characteristics of its Doppler frequency shift to generate a motion pattern classification feature vector.

[0117] In step 3.4, a feature quality assessment matrix is ​​constructed based on the respiratory feature vector, cardiac impact feature vector, and motion feature vector. The respiratory feature weight, cardiac impact feature weight, and motion feature weight are dynamically calculated based on the intra-class discreteness of the respiratory feature, the inter-class separability of the cardiac impact feature, and the confidence of the motion feature. The weighted feature vectors are concatenated into a joint feature vector and normalized.

[0118] In an embodiment of the present invention, the interference problem of feature extraction in a metal cabin environment is specifically solved through hierarchical extraction and intelligent fusion of core living features, which has multiple benefits: first, the three core living features of respiration, cardiac arrest, and motion are accurately separated from the enhanced signal, among which the respiration feature is combined with time-frequency analysis and baseline correction to ensure the stability of periodic and harmonic features, the cardiac arrest feature is adapted to the cabin vibration transfer characteristics through frequency domain compensation to extract the fundamental frequency stability and complexity, and the motion feature eliminates metal heat reflection with the help of dynamic region segmentation and captures the motion pattern through the Doppler zero-crossing rate, thereby achieving environmental adaptability and specificity of feature extraction; then, based on the feature quality (intra-class discreteness, inter-class separability, confidence), the attention weight is dynamically allocated to give effective features a higher weight and weaken the influence of noise interference features. The final normalized joint feature vector not only retains the unique information of each feature, but also enhances the overall recognition through fusion.

[0119] In an embodiment of the present invention, the specific steps include:

[0120] In step 3.1, the microwave radar signal and microvibration signal are extracted from the coherently enhanced target signal cluster. Time-frequency analysis techniques (such as short-time Fourier transform) are used to isolate the respiratory-related harmonic components (typically concentrated in the 0.1-0.5 Hz frequency range). Envelope detection is performed on these respiratory harmonic components (preserving the amplitude trend of the signal). Adaptive baseline drift correction is then performed (calculating the mean of the signal within a sliding window as the baseline and subtracting the baseline from the original signal) to obtain a smooth respiratory waveform envelope.

[0121] Feature vector generation:

[0122] Time domain periodic feature vector: Calculate the time difference between adjacent peaks (or valleys) in the respiratory waveform envelope, take the average of multiple cycles as the periodic feature, and then calculate the standard deviation, maximum value, and minimum value of these cycles to form a time domain periodic feature vector.

[0123] Frequency-domain main harmonic energy ratio eigenvector: Calculate the energy of each harmonic in the respiratory harmonics (integral of the signal power in each harmonic frequency band). Calculate the time integral of the instantaneous power of the signal within the frequency band (the square of the signal value) (the integration interval is 10 seconds to ensure that multiple respiratory cycles are covered). Obtain the energy of each harmonic. Take the ratio of the main harmonic energy with the largest energy to the total energy of all harmonics as the main harmonic energy ratio. Then, combine it with the energy ratio of the sub-highest harmonic to the main harmonic to form the frequency-domain eigenvector.

[0124] In step 3.2, based on the signal segments corresponding to the extracted respiratory features (excluding periods of strong respiratory interference), the cardiac component is separated from the microwave radar signal using heartbeat harmonic reconstruction technology (retaining the 1-3 Hz frequency band, which is the primary frequency band of the cardiac signal). Taking into account the vibration transmission characteristics of the cabin structure (such as the vibration attenuation coefficient at the cabin wall), the frequency domain amplitude of the microvibration signal is compensated (low-frequency vibrations are multiplied by a smaller coefficient, and high-frequency vibrations are multiplied by a larger coefficient to offset structural transmission losses).

[0125] Fundamental frequency stability eigenvector: Calculate the values ​​of the fundamental frequency (main peak frequency) in the ballistocardiac spectrum within a continuous time window, find the ratio of the standard deviation of these values ​​to the mean (the smaller the ratio, the more stable it is), and then combine it with the fluctuation range of the fundamental frequency to form the fundamental frequency stability eigenvector.

[0126] Harmonic complexity eigenvector: Count the number of harmonics higher than the fundamental frequency in the ballistocardial spectrum, calculate the ratio of each harmonic energy to the fundamental frequency energy, and take the number of harmonics and the total energy ratio with a ratio greater than a threshold (such as 0.1) to form the harmonic complexity eigenvector.

[0127] In step 3.3, based on the living body area corresponding to the extracted cardiac ballistic features, the infrared thermal imaging signal is dynamically segmented (a temperature threshold is set: areas 5-10°C higher than the ambient temperature are identified as living body hot spots) to eliminate high-temperature artifact areas reflected by metal surfaces.

[0128] Time-frequency characteristics of hot spot displacement: Track the changes in the hot spot center coordinates over time, calculate the average velocity, acceleration and frequency components of the displacement trajectory in different time windows (obtained through time-frequency analysis), and combine them into the time-frequency characteristic vector of the hot spot displacement.

[0129] Ultrasonic Doppler shift analysis: Calculate the short-term zero-crossing rate of the Doppler shift component of the ultrasonic signal (the number of times the signal crosses the zero value point per unit time), count the mean, peak, and rate of change of the zero-crossing rate in different time windows, and generate a motion pattern classification feature vector (e.g., low zero-crossing rate for stillness, high zero-crossing rate for movement).

[0130] Step 3.4: Construct a feature quality assessment matrix: the rows of the matrix correspond to the three types of features: respiration, cardiac shock, and motion; the columns correspond to feature quality indicators (intra-class consistency, inter-class discrimination, and noise tolerance); and the matrix elements are the quantitative scores (between 0 and 1) of each feature on the corresponding indicator.

[0131] Attention weight calculation:

[0132] Respiratory feature weight: It is inversely proportional to the intra-class dispersion of the respiratory feature (the variance of the feature vector) (the smaller the dispersion, the greater the weight). The calculation formula is "Respiratory weight = 1 ÷ (1 + intra-class dispersion of respiratory feature)".

[0133] Cardiac shock feature weight: It is proportional to the inter-class separability of the cardiac shock feature (the mean difference of different target features divided by the sum of intra-class variances) (the greater the separability, the greater the weight). The calculation formula is "cardiac shock weight = inter-class separability of cardiac shock ÷ (intra-class dispersion of respiration + inter-class separability of cardiac shock + confidence of motion feature)".

[0134] Motion feature weight: It is proportional to the motion feature confidence (the success rate of feature extraction, such as the accuracy of hot spot recognition). The calculation formula is "motion weight = motion feature confidence ÷ (the sum of the above three indicators)" (the sum of the three weights is 1, which is ensured by normalization).

[0135] Joint eigenvector generation: Multiply the respiratory, cardiac, and motion eigenvectors by their corresponding weights and concatenate them into a total eigenvector. This vector is then normalized (each element is subtracted from the vector mean and divided by the vector standard deviation) to output the final joint eigenvector.

[0136] In a preferred embodiment of the present invention, the above step 4 may include:

[0137] Step 4.1: Receive the joint feature vector, extract the respiratory waveform envelope features, the ballistocardiogram features, and the time-frequency feature components of the motion pattern, and map these feature components to the three-dimensional discretized grid nodes of the metal cabin to generate feature distribution data with spatial location annotations.

[0138] Step 4.2: Based on the spatially labeled feature distribution data, dynamic thermal data reflecting the spatial energy density of the living target is generated by calculating the temporal stability index and spatial continuity metric of the feature amplitude of each grid node;

[0139] Step 4.3: Calculate the coverage coefficient of each sensor for each spatial grid node using the distribution information of each energy peak area in the dynamic thermal data, combined with the preset sensor location parameters and the physical law of signal attenuation;

[0140] In step 4.4, the capability coefficient is normalized, and the initial coverage confidence matrix is ​​constructed with the spatial grid nodes as row indices and the sensor channels as column indices. The matrix element value represents the trustworthy monitoring weight of the corresponding sensor at a specific location.

[0141] In step 4.5, according to the metal shielding compensation parameters provided by the cabin structure preset database, the initial coverage confidence matrix is ​​corrected for multipath reflection loss, and the optimized multi-sensor coverage confidence matrix is ​​output.

[0142] In an embodiment of the present invention, by converting abstract feature vectors into concrete spatial energy distributions, a precise association between living features and the physical space of the cabin is achieved, which has multiple benefits: first, the respiratory, cardiac, and motion features are mapped to a three-dimensional grid, giving the features clear spatial position attributes; then, dynamic thermal data is generated through time domain stability and spatial continuity analysis, which intuitively reflects the energy distribution and change trend of the living target, providing a spatial basis for sensor coverage assessment; then, the sensor coverage capability coefficient for each grid node is calculated and a confidence matrix is ​​constructed to quantify the monitoring efficiency of different sensors in the cabin; finally, the metal occlusion compensation parameter correction matrix is ​​combined to eliminate the interference of metal structures on the sensor signal, making the matrix more suitable for actual monitoring scenarios.

[0143] In an embodiment of the present invention, the specific steps include:

[0144] In step 4.1, three types of feature components are extracted from the joint feature vector according to the preset dimensions: respiratory waveform envelope feature component (corresponding to the time domain period and frequency domain energy ratio features generated in step 3.1), cardiac ballistometry feature component (corresponding to the fundamental frequency stability and harmonic complexity features generated in step 3.2), and motion pattern time-frequency feature component (corresponding to the hot spot displacement and time-frequency features generated in step 3.3).

[0145] Based on the discretized grid of the metal cabin (e.g., a 10cm×10cm×10cm cube grid) and the sensor layout coordinates from step 1.1, map the characteristic components to the corresponding grid nodes:

[0146] Microwave radar-related features (respiration, cardiac shock) are mapped to the grid covered by their longitudinal sections;

[0147] Infrared thermography related features (moving hot spots) are mapped to the grid covered by their horizontal profiles;

[0148] Ultrasonic and microvibration features are mapped to their blind spots and bulkhead grids.

[0149] Finally, the feature distribution data of each grid node is generated, which contains "spatial coordinates (x, y, z) + three types of feature component values".

[0150] In step 4.2, the feature amplitude of each grid node (the weighted sum of the three types of feature components, with the weight being the feature weight in step 3.4) is calculated within a continuous time window (e.g., 5 seconds / window, with a sliding step of 2 seconds): time domain stability index = (standard deviation of feature amplitude within the window) ÷ (mean of feature amplitude within the window). (The smaller the index, the more stable the feature is in time and the more likely it is to come from a living target).

[0151] For each grid node, calculate the difference in feature amplitude between it and the six adjacent grids (front and back, left and right, top and bottom): spatial continuity measure = 1-[(the sum of the absolute differences between the feature amplitudes of adjacent grids and the feature amplitude of the current node) ÷ (6 × the mean of the feature amplitude of the current node)] (the closer the measure is to 1, the more continuous the feature is in space, and the more likely it is the energy distribution of the same living target).

[0152] The energy density value of each grid node = (characteristic amplitude) × (time domain stability index) × (spatial continuity measure). The energy density values ​​of all grid nodes are arranged according to spatial coordinates to form dynamic thermal data updated over time (the higher the energy, the greater the possibility of the existence of living organisms).

[0153] In step 4.3, from the dynamic thermal data, extract regions where the energy density exceeds twice the global mean as energy peak regions and record the coordinates of the grid nodes within these regions. For each sensor (microwave radar, infrared array, ultrasonic, or micro-vibration sensor) and each grid node, combine the sensor location parameters (step 1.1) and the physical laws of signal attenuation to calculate the coverage capability coefficient: Coverage capability coefficient = (maximum monitoring distance of sensor / straight-line distance from sensor to grid node) × (1 - signal attenuation rate) × (peak region weight).

[0154] Signal attenuation rate: For microwave or ultrasonic signals, it is calculated as "distance × medium attenuation coefficient" (e.g. microwave attenuation rate in metal = 0.05 × distance); for infrared signals, it is calculated as "the inverse of the square of the distance"; for micro-vibration signals, it is calculated as "bulkhead transmission loss coefficient × distance";

[0155] Peak region weight: If the grid node belongs to the energy peak region, the weight is 1.2, otherwise it is 1.0 (to enhance the coverage evaluation of high energy regions).

[0156] Step 4.4: For each grid node, normalize the coverage capability coefficients of all sensors to it: Normalization coefficient = (coverage capability coefficient of a sensor) ÷ (sum of coverage capability coefficients of all sensors at the grid node), (ensure that the sum of the sensor coefficients corresponding to each grid node is 1, and the value range is 0-1).

[0157] The three-dimensional grid nodes are used as row indices (numbered in order of spatial coordinates), and the sensor channels are used as column indices (numbered in order of microwave, infrared, ultrasonic, and micro-vibration). The (i, j)th element in the matrix is ​​the normalized coefficient of the jth sensor corresponding to the i-th grid node, that is, the initial coverage confidence matrix (the larger the element value, the higher the monitoring confidence of the sensor for the grid).

[0158] In step 4.5, extract the metal shielding compensation parameters corresponding to each grid node from the cabin structure preset database: shielding coefficient (0-1, 1 means no shielding, 0 means complete shielding), multipath reflection loss ratio (0-0.5, such as the corner grid reflection loss is 0.3).

[0159] Each element (i, j) of the initial coverage confidence matrix is ​​corrected: the corrected element value = (initial matrix element value) × (blocking coefficient) × (1-multipath reflection loss ratio), (the sensor weight of the grid node blocked by metal is reduced by the blocking coefficient, and the signal attenuation caused by multipath interference is corrected by the reflection loss ratio, and finally the optimized multi-sensor coverage confidence matrix is ​​output).

[0160] In a preferred embodiment of the present invention, the above step 4 may include:

[0161] Step 4.6, perform a singular value decomposition operation on the multi-sensor coverage confidence matrix to extract the principal component vectors representing the sensor spatial coverage characteristics and their corresponding singular value spectrum distributions;

[0162] Step 4.7: Determine the initial search boundary of the sensor weight parameters based on the characteristic peak interval of the singular value spectrum distribution, and map each weight parameter combination to the position coordinates of individual bees in the swarm optimization;

[0163] Step 4.8: Based on the cabin space coverage characteristics reflected by the principal component vector and the spatial gradient variation of the living body energy distribution, the search area is divided into a high-gradient fine search area and a low-gradient wide exploration area to obtain the partitioning results;

[0164] In step 4.9, the roles of the bee colony are dynamically configured according to the partitioning results. Scout bees are deployed to high-gradient areas to perform neighborhood depth optimization; follower bees are deployed to low-gradient areas to perform heat tracking exploration; after each round of iteration, the regional search tasks are redistributed based on the fitness of the weight parameters. When the change amplitude of the weight parameters in consecutive iterations is lower than the preset convergence threshold, the weight distribution parameters corresponding to the current optimal bee position coordinates are extracted.

[0165] In an embodiment of the present invention, scientific matrix analysis and intelligent optimization strategies are used to achieve efficient and accurate solution of sensor weight parameters; then, a reasonable search boundary is determined based on the characteristic peak of the singular value spectrum to ensure that the search range of the weight parameters fits the actual coverage characteristics and avoids invalid exploration; the search area is divided by combining spatial coverage characteristics with the energy gradient of the living body, so that high-gradient areas (where the energy of the living body changes dramatically and requires key monitoring) adopt fine search, and low-gradient areas (where the energy is flat and can be widely explored) adopt breadth search, thereby balancing optimization accuracy and efficiency; finally, through the dynamic role allocation and iterative fitness adjustment of the bee colony, the scout bees are allowed to deeply cultivate key areas and the follower bees are allowed to expand potential areas, thereby achieving adaptive optimization of the weight parameters and quickly converging to the final solution.

[0166] In an embodiment of the present invention, the specific steps include:

[0167] In step 4.6, perform singular value decomposition on the multi-sensor coverage confidence matrix (set as matrix A, with dimensions of the number of grid nodes × the number of sensor channels). The decomposition result is expressed as follows: matrix A can be decomposed into the product of three matrices, where the first matrix (U) is the left singular vector matrix, the second matrix (Σ) is a diagonal matrix (elements on the diagonal are singular values ​​and are sorted from large to small), and the third matrix (Vᵀ) is the transpose of the right singular vector matrix.

[0168] The principal component vectors are column vectors in the left singular vector matrix (U) whose corresponding singular values ​​are greater than a threshold (threshold = maximum singular value × 0.1). These vectors reflect the main characteristics of the sensor's spatial coverage (such as coverage range and overlapping area). The singular value spectrum distribution is a sequence of all singular values ​​in the diagonal matrix arranged in the original order, where larger singular values ​​correspond to principal components that contribute more to the coverage characteristics.

[0169] In step 4.7, from the singular value spectrum distribution, extract the continuous interval with singular values ​​greater than "maximum singular value × 0.3" as the characteristic peak interval (this interval contains the singular values ​​that play a dominant role in the coverage characteristics). The initial search boundary of the sensor weight parameters (each sensor corresponds to a weight, such as microwave radar weight, infrared array weight, etc.) is determined according to the following rules:

[0170] The minimum value of each weight parameter = the minimum singular value in the characteristic peak interval × 0.5;

[0171] The maximum value of each weight parameter = the maximum singular value in the characteristic peak interval × 1.5;

[0172] Each weight parameter combination (w1, w2, ..., w n , n is the number of sensors) is mapped to the position coordinates of individual bees in the swarm optimization: the coordinate of the bee in the i-th dimension = (the i-th weight parameter value - the minimum value of the parameter) ÷ (the maximum value of the parameter - the minimum value of the parameter) × the search space dimension range (such as 0-100); (the actual value range of the weight parameter is mapped to the search coordinate space of the swarm algorithm to facilitate the algorithm to find the best solution).

[0173] In step 4.8, the spatial gradient of the energy distribution of the living body is calculated as follows: the gradient value of each grid node = (the sum of the absolute differences in energy density between the node and the six adjacent grid nodes) ÷ (6 × the average energy density of the node); (the larger the gradient value, the more dramatic the change in the living body energy in that area, and the more accurate the sensor weighting is required).

[0174] Combined with the sensor coverage dense area reflected by the principal component vector (the area with high principal component vector amplitude), the search area is divided into:

[0175] High gradient fine search area = the area where the gradient value is greater than the mean value of all grid gradients, and this area is the core area covered by the principal component vector;

[0176] Low gradient breadth exploration area = the area where the gradient value is ≤ the mean value of all grid gradients, or the edge area covered by the principal component vector.

[0177] In step 4.9, deploy the bees in a ratio of "60% scouts in high-gradient areas and 60% follower bees in low-gradient areas": the initial positions of the scout bees are randomly distributed within the weight parameter range of the high-gradient area. At each iteration, new candidate parameters are generated in the neighborhood of the current position (e.g., within ±5% of the parameter value). The initial positions of the follower bees are distributed in the low-gradient area based on the high-fitness parameters (the parameters that performed well in the previous iteration). At each iteration, they move toward the parameter position with higher fitness (movement step size = distance from the current position to the high-fitness position × 0.3).

[0178] Fitness function calculation: Weight parameter fitness = 1 ÷ (1 + sensor coverage error), where: Sensor coverage error = (the sum of the squares of the difference between the actual coverage confidence matrix and the coverage matrix weighted by the weight parameter) ÷ (the total number of matrix elements). (The greater the fitness, the more the weight parameter optimizes sensor coverage.)

[0179] After each iteration, the change in weight parameters is calculated: Change = (the sum of the absolute differences between the weight parameter values ​​of the current iteration and the weight parameter values ​​of the previous iteration) ÷ (the total number of weight parameters);

[0180] When the variation amplitude of the three consecutive iterations is less than a preset convergence threshold (e.g., 0.001), the iteration is stopped, and the weight parameter combination corresponding to the position coordinates of the honeybee with the highest fitness is extracted as the final sensor weight distribution parameter.

[0181] In a preferred embodiment of the present application, step 5 can include:

[0182] Step 5.1, receiving the sensor weight distribution parameter, injecting it as the input layer neuron initialization weight into the decision model, and loading the joint feature vector into the model feature input channel;

[0183] Step 5.2, based on the constructed initialization decision model, calling the cabin historical working condition database to construct a training sample set, and fine-tuning the hidden layer structure parameters of the model through a transfer learning mechanism to make the model adapt to the feature distribution offset in the metal cavity environment, generating a metal environment adaptive decision model;

[0184] Step 5.3, based on the metal environment adaptive decision model, real-time monitoring of the activation state of its hidden layer, generating sensor resource allocation instructions according to the spatial distribution pattern of the activation state, and dynamically allocating the microwave radar scan frame rate, infrared array resolution, ultrasonic transmission power, and vibration sensor sensitivity level according to the weight parameter proportion;

[0185] Step 5.4, according to the adaptive sampling strategy, collecting new sensor data streams, inputting real-time data into the metal environment adaptive decision model, updating model parameters through online incremental learning mechanism, and outputting the probability value of the existence of living beings in the cabin and the quantified indicators of vital signs composed of respiratory rate, heart rate variability, and body movement intensity;

[0186] Step 5.5, when the probability value of the existence of living beings exceeds the preset threshold, analyzing the vital sign quantified indicators corresponding to the probability value, and generating a three-dimensional parameter analysis report containing respiratory rate abnormality, heart rate variability stability, and body movement intensity trend.

[0187] In the embodiment of the present application, by constructing an intelligent decision system adapted to the metal environment, the whole process optimization from model initialization to dynamic output of living body detection is realized, the sensor resources (such as radar frame rate, infrared resolution, etc.) are dynamically allocated based on the activation state of the model hidden layer, which optimizes resource consumption while ensuring detection accuracy; through online incremental learning, the model parameters are updated in real time, so that the model can dynamically adapt to the changes of the cabin environment and living body state, ensuring the continuous accuracy of the output living body existence probability and vital sign indicators (respiratory rate, heart rate variability, etc.); finally, when the living body existence probability exceeds the threshold, a three-dimensional analysis report is generated, which intuitively presents the abnormality and trend of vital signs.

[0188] In the embodiment of the present application, the specific steps include:

[0189] In step 5.1, assign sensor weight parameters (e.g., microwave radar weight 0.4, infrared array weight 0.3, etc.) to the connection weights of neurons in the decision model input layer according to sensor type (e.g., the first neuron in the input layer corresponds to microwave radar, and its initial weight is set to 0.4). Input the joint feature vector (containing respiratory, cardiac, and motion features) into the model's feature input channel as the model's initial input data.

[0190] In step 5.2, extract historical feature data from the cargo hold's historical operating condition database for metallic environments (e.g., live body signals under different metal thicknesses and structural configurations) and construct a training sample set with a ratio of "70% normal operating conditions, 30% abnormal operating conditions." Initialize the decision model's input and output layer parameters by freezing them. Iteratively fine-tune the hidden layer parameters using the training sample set (e.g., using a mini-batch gradient descent algorithm with a learning rate of 0.001). After every 50 training batches, evaluate the model's classification accuracy for metallic environment features on the validation set. Fine-tuning is stopped when the accuracy improves by less than 0.5% for three consecutive times, generating a decision model adapted for metallic environments.

[0191] In step 5.3, the activation values ​​of the hidden layer of the metal environment adaptive decision model (such as the output value of the ReLU activation function) are extracted in real time, and the activation intensity is statistically analyzed by grouping according to the spatial region corresponding to the neuron (such as the 1st to 10th neurons corresponding to the front of the cabin).

[0192] If the activation intensity of a certain area is greater than 1.5 times the global average, it is determined to be a key monitoring area, and the sensor resources in the corresponding area are increased according to the weight ratio (for example, the microwave radar scanning frame rate is increased by 20%, and the infrared resolution is increased by 1 level);

[0193] If the activation intensity of a certain area is less than 0.5 times the global mean, it is determined to be a secondary monitoring area and resource allocation is reduced (e.g., ultrasonic transmission power is reduced by 15%);

[0194] The final resource parameters of each sensor = basic configuration parameters × corresponding sensor weight × regional adjustment coefficient.

[0195] In step 5.4, when the probability of the existence of a living organism output by the model is greater than 0.7, the sampling frequency is increased to 10 Hz (originally 5 Hz);

[0196] When the fluctuation range of vital sign indicators is greater than 2 times the historical average, emergency sampling is triggered (the sampling frequency is temporarily increased to 20Hz for 30 seconds).

[0197] The newly collected sensor data is mixed with historical data in a ratio of 1:4. Every time 100 new samples are accumulated, the model parameters are updated (using an incremental learning algorithm and retaining 80% of the historical knowledge weight).

[0198] Indicator output:

[0199] Probability of living body existence: the probability value corresponding to the "living body existence" category of the model output layer normalization function;

[0200] Respiratory rate: the reciprocal of the periodic mean extracted from the respiratory waveform envelope feature;

[0201] Heart rate variation coefficient: the ratio of the standard deviation to the mean in the heart impact spectrum fundamental frequency stability feature;

[0202] Body movement intensity: the hot spot displacement velocity mean in the motion mode time-frequency feature.

[0203] Step 5.5, when the probability of living body existence > 0.8 (preset threshold), trigger the report generation process.

[0204] Respiratory rate abnormality = | current respiratory rate - normal range mean | ÷ normal range standard deviation;

[0205] Heart rate variation stability = 1 - heart rate variation coefficient (the smaller the coefficient, the more stable);

[0206] Body movement intensity trend = current body movement intensity - 5-minute body movement intensity mean.

[0207] A three-dimensional coordinate system is constructed with the respiratory rate abnormality as the X-axis, the heart rate variation stability as the Y-axis, and the body movement intensity trend as the Z-axis; the current index value is mapped into the coordinate system to generate scatter data; according to the scatter position (such as falling in the upper right corner area), automatically label warning information such as "rapid breathing and unstable heart rate", and generate an analysis report containing historical trend comparison.

[0208] As shown in Figure 2 , the embodiment of the present application also provides a living body detection system in a metal cabin, comprising:

[0209] The acquisition module is used for synchronously collecting multi-source heterogeneous sensor data streams in the cabin by deploying a multi-frequency microwave radar, an infrared thermal imaging array, a wideband ultrasonic sensor and a high-sensitivity micro-vibration sensor group inside the metal cabin; and for suppressing reverberation noise and multipath interference caused by metal cavity resonance by using an adaptive signal enhancement algorithm based on frequency domain sub-band decomposition according to the multi-source heterogeneous sensor data streams, and outputting a coherent enhanced target signal cluster;

[0210] The processing module is used for extracting the respiratory waveform envelope feature, the heart impact spectrum feature and the motion mode time-frequency feature of the living body target by time-frequency analysis technology based on the coherent enhanced target signal cluster, and generating a joint feature vector by using a feature fusion method weighted by an attention mechanism;

[0211] The computational module is used to analyze the energy distribution characteristics of living targets within the cargo hold based on the joint eigenvectors and dynamically construct a multi-sensor coverage confidence matrix. The confidence matrix is ​​subjected to singular value decomposition and, in conjunction with an adaptive swarm division-of-labor optimization algorithm, the sensor weight allocation parameters are used as the swarm's foraging targets. Each bee represents a set of candidate weight parameters. Leveraging the swarm's division of labor and collaboration, the search strategy is dynamically adjusted in conjunction with the energy distribution characteristics of living objects within the cargo hold. Through iterative optimization, the weight parameters converge to the final solution, generating the sensor weight allocation parameters.

[0212] The intelligent decision-making module is used to build a metal environment-adaptive decision-making model based on sensor weight allocation parameters and joint feature vectors, adaptively adjust the sensor sampling strategy, and output the probability of living organisms and quantitative indicators of vital signs.

[0213] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for detecting living things in a metal carrier, characterized in that: The method comprises: Step 1: A multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier cabin synchronously collect multi-source heterogeneous sensor data streams inside the carrier cabin; Step 2: Based on the multi-source heterogeneous sensor data stream, an adaptive signal enhancement algorithm based on frequency domain subband decomposition is used to suppress the reverberation noise and multipath interference caused by the resonance of the metal cavity, and output a coherently enhanced target signal cluster; Step 3: Based on the coherently enhanced target signal cluster, the respiratory waveform envelope features, cardiac ballistic spectrum features, and movement pattern time-frequency features of the living target are extracted through time-frequency analysis technology, and a feature fusion method with weighted attention mechanism is used to generate a joint feature vector; Step 4: Based on the joint eigenvector, the energy distribution characteristics of the living targets in the cargo hold are analyzed, and a multi-sensor coverage confidence matrix is ​​dynamically constructed. The confidence matrix is ​​subjected to singular value decomposition, and combined with an adaptive swarm division of labor optimization algorithm, the sensor weight distribution parameters are used as the swarm's foraging targets. Each bee represents a set of candidate weight parameters. Leveraging the swarm's division of labor and collaboration, the search strategy is dynamically adjusted based on the energy distribution characteristics of the living targets in the cargo hold. Through iterative optimization, the weight parameters are converged to the final solution, generating the sensor weight distribution parameters. Step 5: Based on the sensor weight distribution parameters and the joint feature vector, a metal environment adaptive decision model is constructed to adaptively adjust the sensor sampling strategy and output the probability of living body presence and quantitative indicators of vital signs.

2. A method for detecting living bodies in a metal carrier according to claim 1, characterized in that: The multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier synchronously collect multi-source heterogeneous sensor data streams inside the carrier, including: Step 1.1: Generate a three-dimensional coordinate mapping table based on the structural parameters of the metal cargo cabin. Deploy the sensor network in a layout where the microwave radar covers the longitudinal section, the infrared array covers the horizontal section, the ultrasonic sensor covers the corner blind spots, and the micro-vibration sensor group fits the cabin wall. In step 1.2, based on the three-dimensional spatial coordinate mapping table, the multi-sensor hardware synchronization trigger mechanism is activated. A global timestamp is generated using a high-precision clock source. This ensures that the microwave radar pulse emission phase, infrared array exposure period, ultrasonic emission sequence, and vibration sampling time are strictly aligned in the time domain to obtain the raw sensor data. Step 1.3: Receive the raw sensor data and, based on the sensor position parameters recorded in the 3D spatial coordinate mapping table, add the cabin material attenuation compensation coefficient to the microwave radar signal, the ambient temperature drift correction parameter to the infrared thermal imaging data, the air density compensation factor to the ultrasonic signal, and the mechanical coupling gain coefficient to the micro-vibration signal, to generate a primary data stream with physical compensation. In step 1.4, the primary data stream is position-matched according to the three-dimensional spatial coordinate mapping table, where microwave radar data is mapped to the polar coordinate system, infrared data is mapped to the Cartesian coordinate system, ultrasonic data is mapped to the spherical coordinate system, and micro-vibration data is mapped to the structural grid coordinate system, and the multi-source heterogeneous sensor data stream with temporal and spatial correlation is output.

3. The method for detecting living bodies in a metal carrier according to claim 2, characterized in that: Based on multi-source heterogeneous sensor data streams, an adaptive signal enhancement algorithm based on frequency domain sub-band decomposition is used to suppress reverberation noise and multipath interference caused by metal cavity resonance, and output a coherently enhanced target signal cluster, including: Step 2.1: Based on the multi-source heterogeneous sensor data stream, after performing millimeter-wave penetration compensation on the microwave radar signal, the full frequency band is decomposed into multiple sub-bands based on the resonant mode distribution of the metal carrier. Using this decomposition result, the ultrasonic signal is divided into critical frequency bands based on the attenuation coefficient of the sound wave in the metal medium. Furthermore, based on the structural transfer function of the carrier, the micro-vibration signal is grouped into octaves to match the mechanical vibration propagation characteristics. In step 2.2, based on the results of the sub-band, frequency band, and group division, a coherence matrix across sensor channels is constructed within each sub-band, and the reverberation noise subspace is separated by eigenvalue decomposition. Sub-band weight coefficients are dynamically generated, with the phase continuity of the target signal cluster between adjacent sub-bands as the constraint. Step 2.3: Based on the weighted sub-band signals, extract the time delay characteristic peaks formed by multiple reflections on the metal wall, calculate the anti-phase cancellation beam to cancel the multipath interference; use spatiotemporal adaptive filtering to eliminate the metal surface mirror reflection artifacts; and output the sub-band signals after suppressing the interference. In step 2.4, each subband signal is reconstructed into a time domain signal cluster through inverse transformation, and the mutual information entropy value between the reconstructed signal and the original sensor data is calculated; if the entropy value is lower than the dynamic threshold, the coherently enhanced target signal cluster is output; otherwise, the data re-acquisition mechanism is triggered for closed-loop correction.

4. The method for detecting living bodies in a metal carrier according to claim 3, characterized in that: Based on the coherently enhanced target signal cluster, the respiratory waveform envelope features, cardiac ballistic spectrum features, and motion pattern time-frequency features of the living target are extracted through time-frequency analysis technology. The feature fusion method with weighted attention mechanism is then used to generate a joint feature vector, including: Step 3.1: Based on the coherently enhanced target signal cluster, extract the microwave radar and micro-vibration signals. Separate the respiratory harmonic components using time-frequency analysis techniques, and use envelope detection and adaptive baseline drift correction to generate the time-domain periodic eigenvector and frequency-domain main harmonic energy ratio eigenvector of the respiratory waveform envelope. Step 3.2: Based on the respiratory feature extraction results and the corresponding signal segments, the heartbeat harmonic reconstruction technique is used to separate the cardiac ballistomic component from the microwave radar signal. Simultaneously, the microvibration signal is frequency-domain compensated in combination with the cabin structure vibration transfer characteristics to extract the fundamental frequency stability eigenvector and harmonic complexity eigenvector of the cardiac ballistomic spectrum. Step 3.3: Based on the ballistocardiographic feature extraction results and the processed signal clusters, the infrared thermal imaging signal is dynamically segmented to eliminate metal heat reflection interference and extract the time-frequency feature vector of the hot spot displacement trajectory. Simultaneously, the output ultrasonic signal is analyzed for the short-term zero-crossing rate characteristics of its Doppler frequency shift to generate a motion pattern classification feature vector. In step 3.4, a feature quality assessment matrix is ​​constructed based on the respiratory feature vector, cardiac impact feature vector, and motion feature vector. The respiratory feature weight, cardiac impact feature weight, and motion feature weight are dynamically calculated based on the intra-class discreteness of the respiratory feature, the inter-class separability of the cardiac impact feature, and the confidence of the motion feature. The weighted feature vectors are concatenated into a joint feature vector and normalized.

5. The method for detecting living bodies in a metal carrier according to claim 4, characterized in that: Based on the joint eigenvector, the energy distribution characteristics of the living target in the cabin are analyzed, and the multi-sensor coverage confidence matrix is ​​dynamically constructed, including: Step 4.1: Receive the joint feature vector, extract the respiratory waveform envelope features, the ballistocardiogram features, and the time-frequency feature components of the motion pattern, and map these feature components to the three-dimensional discretized grid nodes of the metal cabin to generate feature distribution data with spatial location annotations. Step 4.2: Based on the spatially labeled feature distribution data, dynamic thermal data reflecting the spatial energy density of the living target is generated by calculating the temporal stability index and spatial continuity metric of the feature amplitude of each grid node; Step 4.3: Calculate the coverage coefficient of each sensor for each spatial grid node using the distribution information of each energy peak area in the dynamic thermal data, combined with the preset sensor location parameters and the physical law of signal attenuation; In step 4.4, the capability coefficient is normalized, and the initial coverage confidence matrix is ​​constructed with the spatial grid nodes as row indices and the sensor channels as column indices. The matrix element value represents the trustworthy monitoring weight of the corresponding sensor at a specific location. In step 4.5, according to the metal shielding compensation parameters provided by the cabin structure preset database, the initial coverage confidence matrix is ​​corrected for multipath reflection loss, and the optimized multi-sensor coverage confidence matrix is ​​output.

6. The method for detecting living bodies in a metal carrier according to claim 5, characterized in that: The confidence matrix is ​​subjected to singular value decomposition and combined with an adaptive bee swarm division of labor optimization algorithm. The sensor weight distribution parameters are used as the swarm's foraging target. Each bee represents a set of candidate weight parameters. The division of labor and cooperation of the swarm is utilized, and the search strategy is dynamically adjusted in combination with the energy distribution characteristics of the living organisms in the carrier. Through iterative optimization, the weight parameters converge to the final solution, generating the sensor weight distribution parameters, including: Step 4.6, perform a singular value decomposition operation on the multi-sensor coverage confidence matrix to extract the principal component vectors representing the sensor spatial coverage characteristics and their corresponding singular value spectrum distributions; Step 4.7: Determine the initial search boundary of the sensor weight parameters based on the characteristic peak interval of the singular value spectrum distribution, and map each weight parameter combination to the position coordinates of individual bees in the swarm optimization; Step 4.8: Based on the cabin space coverage characteristics reflected by the principal component vector and the spatial gradient variation of the living body energy distribution, the search area is divided into a high-gradient fine search area and a low-gradient wide exploration area to obtain the partitioning results; In step 4.9, the roles of the bee colony are dynamically configured according to the partitioning results. Scout bees are deployed to high-gradient areas to perform neighborhood depth optimization; follower bees are deployed to low-gradient areas to perform heat tracking exploration; after each round of iteration, the regional search tasks are redistributed based on the fitness of the weight parameters. When the change amplitude of the weight parameters in consecutive iterations is lower than the preset convergence threshold, the weight distribution parameters corresponding to the current optimal bee position coordinates are extracted.

7. The method for detecting living bodies in a metal carrier according to claim 6, characterized in that: Based on the sensor weight allocation parameters and joint feature vectors, a metal environment adaptive decision model is constructed to adaptively adjust the sensor sampling strategy and output the probability of living body presence and quantitative indicators of vital signs, including: Step 5.1: Receive the sensor weight distribution parameters and inject them into the decision model as the input layer neuron initialization weights, and load the joint feature vector into the model feature input channel; Step 5.2: Based on the constructed initialization decision model, the historical operating condition database of the cargo hold is used to construct a training sample set. The hidden layer structure parameters of the model are fine-tuned through the transfer learning mechanism to adapt the model to the characteristic distribution offset in the metal cavity environment, thereby generating a metal environment-adaptive decision model. Step 5.3: Based on the metal environment adaptive decision model, the activation state of its hidden layer is monitored in real time. According to the spatial distribution pattern of the activation state, sensor resource allocation instructions are generated. The microwave radar scanning frame rate, infrared array resolution, ultrasonic transmission power, and vibration sensor sensitivity levels are dynamically allocated according to the weight parameter ratio. In step 5.4, new sensor data streams are collected according to the adaptive sampling strategy. The real-time data is input into the metal environment adaptive decision model. The model parameters are updated through the online incremental learning mechanism to output the probability of the presence of living organisms in the cabin and the quantitative indicators of vital signs such as respiratory rate, heart rate variability, and body movement intensity. Step 5.5: When the probability value of the living body exceeds the preset threshold, the vital sign quantitative indicators corresponding to the probability value are analyzed to generate a three-dimensional parameter analysis report including respiratory rate abnormality, heart rate variability stability, and body movement intensity trend.

8. A living body detection system in a metal carrier, the system implementing the method according to any one of claims 1 to 7, characterized in that: include: The acquisition module is used to synchronously collect multi-source heterogeneous sensor data streams in the carrier cabin through a multi-band microwave radar, infrared thermal imaging array, broadband ultrasonic sensor, and high-sensitivity micro-vibration sensor group deployed inside the metal carrier cabin; It is used to suppress the reverberation noise and multipath interference caused by the resonance of the metal cavity based on the adaptive signal enhancement algorithm based on the frequency domain sub-band decomposition of multi-source heterogeneous sensor data streams, and output the coherently enhanced target signal cluster; A processing module is used to extract the respiratory waveform envelope features, cardiac ballistometry features, and motion pattern time-frequency features of the living target based on the coherently enhanced target signal cluster through time-frequency analysis technology, and to generate a joint feature vector using a feature fusion method weighted by an attention mechanism; The computational module is used to analyze the energy distribution characteristics of living targets within the cargo hold based on the joint eigenvectors and dynamically construct a multi-sensor coverage confidence matrix. The confidence matrix is ​​subjected to singular value decomposition and, in conjunction with an adaptive swarm division-of-labor optimization algorithm, the sensor weight allocation parameters are used as the swarm's foraging targets. Each bee represents a set of candidate weight parameters. Leveraging the swarm's division of labor and collaboration, the search strategy is dynamically adjusted in conjunction with the energy distribution characteristics of living objects within the cargo hold. Through iterative optimization, the weight parameters converge to the final solution, generating the sensor weight allocation parameters. The intelligent decision-making module is used to build a metal environment-adaptive decision-making model based on sensor weight allocation parameters and joint feature vectors, adaptively adjust the sensor sampling strategy, and output the probability of living organisms and quantitative indicators of vital signs.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images