Post-disaster life detection and signal identification method based on transfer learning

Abstract

The application discloses a post-disaster life detection and signal identification method based on transfer learning and belongs to the technical field of post-disaster life information detection and search and rescue, and comprises the following steps: S1, multi-source signal acquisition: at least two sensors in ultrawideband radar, audio vibration pickup, infrared thermal imaging and gas sensing are used to acquire mixed life detection signals in a post-disaster ruin environment, wherein the mixed life detection signals comprise one or more of breathing echo, heartbeat micro-vibration, distress sound wave, human infrared hot spot and carbon dioxide concentration gradient; pure life signal prior knowledge accumulated in a source domain scene is used to complete signal residual separation, can accurately filter out clutter interference, mechanical vibration noise and natural wind and rain noise in the post-disaster ruin environment, completely retains weak breathing, low-frequency heartbeat and other life characteristic signals of trapped personnel which are prone to be lost, and effectively solves the actual problems that traditional noise reduction methods are prone to damage weak effective signals and the missing detection rate of deep buried scenes is high.

Description

Technical Field

[0001] This invention belongs to the field of post-disaster life information detection and search and rescue technology, specifically concerning a post-disaster life detection and signal recognition method based on transfer learning. Background Technology

[0002] Post-disaster life detection and signal recognition are core technologies for disaster relief. The core technology is to use electromagnetic waves, sound waves, infrared, optics, gas, electronic signals, AI fusion and other technologies to penetrate obstacles such as rubble, smoke, and darkness to capture weak life signals (breathing, heartbeat, body temperature, sound, movement, mobile phone signal, metabolic gas) of trapped people, so as to achieve rapid location and identification.

[0003] Existing technologies have the following significant drawbacks: First, real-world post-disaster rubble life signal samples are scarce, difficult to collect, and exhibit extremely high scene variability. Traditional deep learning models require massive amounts of disaster scene samples for support, and their generalization ability is extremely poor in small-sample scenarios, making them prone to overfitting and recognition failure. Second, traditional filtering, wavelet denoising, and other mathematical denoising methods are globally uniform denoising, which, while filtering out noise, severely damages low-frequency, weak life signals such as faint breathing, faint heartbeats, and low-pitched cries for help under the rubble, leading to missed detections and misjudgments. Third, conventional intelligent recognition models are only trained for a single fixed scene, and there is a severe domain separation between clean data in non-disaster source domains and dirty data in post-disaster target domains. First, direct transfer of data due to fabric offset will result in a large number of transfer errors, leading to low recognition accuracy and high false positive and false negative rates. Second, traditional multimodal fusion is mostly simple weighted fusion, which does not distinguish the feature weights of different levels and cannot adapt to different occlusion and attenuation conditions in ruins. At the same time, it lacks a cross-modal cross-verification mechanism, and single-modal interference can easily cause false positive and false negative recognition results. Third, existing technologies cannot achieve small-sample adaptive adaptation, dynamic feature optimization, and accurate multi-target differentiation, making it difficult to meet the high-precision, high-stability, and rapid rescue and detection needs in the golden 72 hours after a disaster. Therefore, a post-disaster life detection and signal recognition method based on transfer learning is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a post-disaster life detection and signal recognition method based on transfer learning, so as to solve the problems of strong noise interference, easy loss of weak signals, scarce scene samples, large domain offset error, poor stability of multimodal recognition, and high false negative rate in post-disaster life detection signals.

[0005] To achieve the above objectives, this invention provides a post-disaster life detection and signal identification method based on transfer learning, comprising the following steps: S1. Multi-source signal acquisition: Acquire mixed life detection signals in the post-disaster ruins environment by using at least two of the following sensors: ultra-wideband radar, audio vibration pickup, infrared thermal imaging, and gas sensing. The mixed life detection signals include one or more of the following: respiratory echo, heartbeat micro-vibration, distress call sound wave, human infrared hot spot, and carbon dioxide concentration gradient. S2. Transfer learning residual denoising preprocessing: Introducing a pre-trained transfer denoising residual module to transfer the prior knowledge of pure life signals learned in the source domain scene to the post-disaster target domain, perform residual separation on the collected mixed life detection signals, filter out rubble clutter, mechanical vibration, wind and rain noise interference, and retain weak and effective life signals. S3. Source Domain Model Pre-training: Based on a standard life signal dataset in a non-disaster scenario, a deep learning base model is pre-trained. The base model is used to extract common features of life signals, including respiratory cycle, heart rate, human voice spectrum, temperature threshold, and gas concentration change rate. S4. Adaptive Fine-tuning and Domain Shift Correction in Transfer Learning: Construct a disaster scene dataset in the target domain and adopt a transfer learning strategy of feature transfer + parameter fine-tuning. Freeze the bottom general feature layer of the basic model and only fine-tune the top classification and localization layers with small samples. At the same time, construct a dual discriminator domain adaptive transfer structure to correct the signal intensity distribution shift and the temporal feature distribution shift respectively, reduce the signal distribution difference between the source domain and the target domain, and eliminate scene transfer error. S5. Layered feature adaptive weighting and multimodal signal fusion recognition: Dynamic adaptive weighting is applied to the bottom-level general features, middle-level signal detail features, and top-level scene discrimination features of the fine-tuned model. The feature weights of each layer are adjusted in real time according to the noise intensity of the post-disaster ruins and the degree of signal attenuation. Cross-modal consistency cross-validation is performed on the multi-source detection migration features to eliminate abnormal recognition results caused by single-modal interference. After removing residual interference, the system outputs the determination of the presence of living beings, three-dimensional position coordinates, and vital signs parameters. S6. Result Verification and Output: Integrate optical endoscopy or mobile phone signal-assisted verification to generate a visual recognition result that includes the location, depth, and status of the living organism.

[0006] Furthermore, the non-disaster scenario standard life signal dataset mentioned in S3 includes standard human breathing, heartbeat, sound, infrared, and CO2 signal samples under indoor, outdoor, quiet, and weak interference scenarios. The basic model is a CNN-LSTM hybrid neural network or a Transformer model.

[0007] Furthermore, the transfer learning strategy described in S4 specifically involves using maximum mean difference (MMD) or domain adversarial training (DANN) to reduce the signal distribution difference between the source and target domains, and fine-tuning the sample size to 5%-20% of the source domain sample size.

[0008] Furthermore, in S4, the bottom general feature layer is frozen and the top classification and positioning layer is fine-tuned. The bottom general feature layer is used to extract the general features of the signal in the time domain, frequency domain, and spatial domain, and the top classification and positioning layer is used to adapt to the strong noise, strong attenuation, and multiple occlusion characteristics of post-disaster ruins.

[0009] Furthermore, the multimodal signal fusion recognition described in S5 includes: extracting 0.1-0.5Hz breathing features and 1-2Hz heartbeat features from ultra-wideband radar signals; extracting 200-3000Hz human voice / tapping features from audio signals and calculating the orientation through time difference of location (TDOA); identifying human hotspots at 36±2℃ from infrared signals and detecting abnormal concentration gradients from CO2 signals; and distinguishing between living organisms and animals, mechanical vibrations, and environmental interference through weighted fusion of multimodal features using a transfer learning model.

[0010] Furthermore, the migration noise reduction residual module described in S2 differs from traditional mathematical filtering. It achieves intelligent noise reduction based on prior knowledge of pure life signals in the source domain, removing only noise signals while fully preserving weak breathing and low-frequency life characteristic signals of the heartbeat.

[0011] Furthermore, the cross-modal consistency cross-validation described in S5 specifically involves correcting abnormal results based on the migration features of other effective modalities when any single modal recognition result deviates from the multimodal comprehensive feature threshold, thereby eliminating false positive and false negative recognition errors.

[0012] Furthermore, the dynamic adaptive weighting described in S5 dynamically increases the weight of weak life features and suppresses the weight of invalid noise features to adapt to the complex and ever-changing detection environment after a disaster, taking into account different conditions of rubble obstruction and signal attenuation.

[0013] Furthermore, the auxiliary verification described in S6 is: confirming the life form through mobile phone Bluetooth / Wi-Fi signal detection or fiber optic endoscopic images, with a positioning error of ≤0.5 meters.

[0014] Furthermore, this method also includes: using a transfer learning model to simultaneously identify multiple target life forms, distinguishing between adults, children, and animals, and outputting the independent location and vital signs parameters of each target.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention adopts a transfer learning residual noise reduction preprocessing method, which is different from the traditional mathematical filtering noise reduction mode. It relies on the prior knowledge of pure life signals accumulated in the source domain scene to complete the signal residual separation. It can accurately filter out clutter interference, mechanical vibration noise and natural wind and rain noise in the post-disaster ruins environment, and completely preserve the life characteristic signals of trapped personnel such as weak breathing and low-frequency heartbeat, which are easily lost. It effectively solves the practical problems of traditional noise reduction methods that easily damage weak effective signals and have a high rate of missed detection in deep buried scenarios.

[0016] (2) This invention uses hierarchical parameter transfer combined with dual discriminator domain offset correction technology. By freezing the general feature layer of the model and only fine-tuning the training of the top identification and positioning layer, the general vital signs features learned by the model are fully preserved. At the same time, the difference in data distribution between the source domain and the post-disaster target domain is reduced by using MMD and DANN algorithms. The scene adaptation can be completed by relying on only 5% to 20% of the small sample data of the source domain sample, which greatly reduces the data cost of model training and scene adaptation, shortens the time for model debugging at the emergency rescue site, and enables the intelligent identification model to quickly adapt to the ruin detection environment with different collapse degrees and different obstruction media, significantly improving the generalization application capability of the model.

[0017] (3) The invention sets up a layered feature adaptive weighting and cross-modal consistency cross-verification mechanism, which can dynamically adjust the feature weights of each layer according to the on-site noise intensity, the thickness of the rubble obstruction and the degree of signal attenuation, actively amplify the weights of effective life features and suppress the weights of environmental interference features; at the same time, it relies on radar, audio, infrared and gas multimodal detection features to achieve mutual verification, and promptly corrects the abnormal identification results caused by the influence of the environment on a single mode, thereby significantly reducing the probability of false positives and false negatives from the algorithm level.

[0018] (4) This invention completes secondary result verification through fiber optic endoscopy and mobile phone electronic signals, which can control the overall detection and positioning error within 0.5 meters. It can accurately output the burial depth, specific location and basic vital signs parameters of trapped personnel, providing accurate and reliable data support for on-site rescue and demolition operations. At the same time, it can simultaneously realize single-target and multi-target life detection, accurately distinguish between adult, child and wild animal target types, and enable rescuers to accurately grasp the number and type of trapped personnel on site, rationally allocate rescue forces, and comprehensively improve the overall search and rescue efficiency after the disaster. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the post-disaster life detection and signal recognition method based on transfer learning of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0021] To keep the drawings concise, each drawing only schematically shows the parts relevant to the invention; these do not represent the actual structure of the product. Furthermore, for ease of understanding, in some drawings, only one of components with the same structure or function is schematically shown, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one," and "several" includes "two" and "more than two."

[0022] In this document, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0024] In the description of this embodiment, terms such as "upper," "lower," "left," and "right" are based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of description and simplification of operation, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0025] Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] like Figure 1 As shown, it illustrates a post-disaster life detection and signal identification method based on transfer learning in one embodiment of the present invention, including the following steps: S1. Multi-source signal acquisition: Acquire mixed life detection signals in the post-disaster ruins environment through at least two of the following sensors: ultra-wideband radar, audio vibration pickup, infrared thermal imaging, and gas sensing. The mixed life detection signals include one or more of the following: respiratory echo, heartbeat micro-vibration, distress call sound wave, human infrared hot spot, and carbon dioxide concentration gradient. S2. Transfer learning residual denoising preprocessing: Introducing a pre-trained transfer denoising residual module to transfer the prior knowledge of pure life signals learned in the source domain scene to the post-disaster target domain, perform residual separation on the collected mixed life detection signals, filter out rubble clutter, mechanical vibration, wind and rain noise interference, and retain weak and effective life signals. S3. Source Domain Model Pre-training: Based on a standard life signal dataset in non-disaster scenarios, a deep learning base model is pre-trained. The base model is used to extract common features of life signals, including respiratory cycle, heart rate, human voice spectrum, temperature threshold, and gas concentration change rate. S4. Adaptive Fine-tuning and Domain Shift Correction in Transfer Learning: Construct a disaster scene dataset in the target domain and adopt a transfer learning strategy of feature transfer + parameter fine-tuning. Freeze the bottom general feature layer of the basic model and only fine-tune the top classification and localization layers with small samples. At the same time, construct a dual discriminator domain adaptive transfer structure to correct the signal intensity distribution shift and the temporal feature distribution shift respectively, reduce the signal distribution difference between the source domain and the target domain, and eliminate scene transfer error. S5. Layered feature adaptive weighting and multimodal signal fusion recognition: Dynamic adaptive weighting is applied to the bottom-level general features, middle-level signal detail features, and top-level scene discrimination features of the fine-tuned model. The feature weights of each layer are adjusted in real time according to the noise intensity of the post-disaster ruins and the degree of signal attenuation. Cross-modal consistency cross-validation is performed on the multi-source detection migration features to eliminate abnormal recognition results caused by single-modal interference. After removing residual interference, the system outputs the determination of the presence of living beings, three-dimensional position coordinates, and vital signs parameters. S6. Result Verification and Output: Integrate optical endoscopy or mobile phone signal-assisted verification to generate a visual recognition result that includes the location, depth, and status of the living organism.

[0027] The S3 non-disaster scenario standard life signal dataset includes standard human breathing, heartbeat, sound, infrared, and CO2 signal samples in indoor, outdoor, quiet, and low-interference scenarios. The basic model is a CNN-LSTM hybrid neural network or a Transformer model.

[0028] In this embodiment: the basic capabilities of the model are built by relying on standard and clean samples from multiple scenarios. The CNN-LSTM or Transformer model can simultaneously take into account the extraction of the time-frequency spatial features and the temporal periodic features of the signal, ensuring that the model fully learns the inherent and universal laws of human life signals, laying a solid foundation of underlying features for subsequent cross-scenario transfer and adaptation, and avoiding the later recognition deviation caused by insufficient feature extraction capabilities of the basic model.

[0029] The specific transfer learning strategy in S4 is to use maximum mean difference (MMD) or domain adversarial training (DANN) to reduce the signal distribution difference between the source domain and the target domain, and to fine-tune the sample size to 5%-20% of the sample size of the source domain.

[0030] In this embodiment, the MMD and DANN algorithms are used to efficiently bridge the feature distribution gap between clean data in the source domain and noisy data in the post-disaster target domain. Scene adaptation can be completed using only 5% to 20% of the source domain samples, which greatly reduces the difficulty and data cost of collecting labeled samples in post-disaster disaster scenarios. This enables rapid on-site model adaptation under small sample conditions and shortens the model debugging and preparation time before rescue.

[0031] In S4, the bottom general feature layer is frozen and the top classification and positioning layer is fine-tuned. The bottom general feature layer is used to extract general features of the signal in the time domain, frequency domain, and spatial domain, while the top classification and positioning layer is used to adapt to the strong noise, strong attenuation, and multiple occlusion characteristics of post-disaster ruins.

[0032] In this embodiment, the general life-related basic features learned by the model are not lost. The top-level identification and positioning logic is optimized specifically to adapt to the harsh working conditions of strong noise, strong attenuation, and multiple obstructions in the ruins. This not only retains the accurate feature extraction capability obtained from training with massive amounts of standard data, but also quickly adapts to the actual detection environment at the disaster site, taking into account both the model's universality and the adaptability to the on-site scene.

[0033] Multimodal signal fusion recognition in S5 includes: Extract respiratory features (0.1-0.5Hz) and heartbeat features (1-2Hz) from ultra-wideband radar signals; Extract human voice / tapping features from the 200-3000Hz frequency range of the audio signal and calculate the location using Time Difference of Origin (TDOA). It can identify human hotspots at 36±2℃ using infrared signals and detect abnormal concentration gradients in CO2 signals. By using a transfer learning model to weightedly fuse multimodal features, we can distinguish between living organisms and animals, mechanical vibrations, and environmental disturbances.

[0034] In this embodiment, refined feature extraction is performed on signals from different sensors, strictly matching the inherent frequency and physical characteristics of human vital signs. Preliminary screening is completed by relying on fixed feature thresholds, and then accurate differentiation of living organisms, animal bodies, mechanical equipment, and environmental clutter is achieved through model fusion. This effectively reduces misjudgments caused by similar interference signals and improves the professionalism and recognition accuracy of feature identification.

[0035] The S2 migration noise reduction residual module differs from traditional mathematical filtering. It achieves intelligent noise reduction based on prior knowledge of pure life signals in the source domain, removing only noise signals while fully preserving weak breathing and low-frequency life characteristic signals of the heartbeat.

[0036] In this embodiment, the traditional one-size-fits-all mathematical filtering method is abandoned because it easily damages weak low-frequency signals. It relies on the prior life knowledge of the source domain to achieve intelligent residual separation. While completely filtering out debris, mechanical vibration, wind and rain noise, it completely preserves the weak breathing and heartbeat low-frequency effective signals of deeply buried and trapped personnel, greatly reducing the probability of missing weak life signals.

[0037] S5 cross-modal consistency cross-validation specifically involves correcting abnormal results based on the migration features of other valid modalities when the recognition result of any single modality deviates from the multimodal comprehensive feature threshold, thereby eliminating false positive and false negative recognition errors.

[0038] In this embodiment, a fault-tolerant identification mechanism is formed by mutual corroboration of multiple source signals. When a single detection mode is affected by environmental interference and produces abnormal results, the results are corrected by relying on the migration features of other effective modes. This mechanism significantly reduces the two core identification errors of false positives and false negatives, thereby improving the reliability of identification results in complex environments.

[0039] The S5 dynamic adaptive weighting dynamically increases the weight of weak life features and suppresses the weight of invalid noise features to adapt to the complex and ever-changing detection environment after a disaster, taking into account different conditions of rubble obstruction and signal attenuation.

[0040] In this embodiment, the feature weights can be autonomously adjusted in real time according to the thickness of the rubble obstruction, the intensity of noise, and the degree of signal attenuation. It actively amplifies the weights of weak life-related effective features and suppresses the weights of useless environmental noise features, allowing the recognition model to dynamically adapt to the changing search and rescue conditions on site, and further improving the recognition stability in extremely complex rubble environments.

[0041] The auxiliary verification in S6 is as follows: confirming the presence of a living being through mobile phone Bluetooth / Wi-Fi signal detection or fiber optic endoscopic images, with a positioning error of ≤0.5 meters.

[0042] In this embodiment, the intelligent identification results are manually reviewed a second time through dual verification of endoscopic visualization and mobile phone electronic signals. At the same time, the search and rescue positioning error is strictly controlled within 0.5 meters, accurately locating the buried position and depth of the trapped personnel, providing accurate data support for subsequent safe demolition and rescue, and meeting the actual construction accuracy requirements of emergency rescue.

[0043] It also includes: using transfer learning models to simultaneously identify multiple target life forms, distinguish between adults, children, and animals, and output the independent location and vital signs parameters of each target.

[0044] In this embodiment, multiple life forms can be detected and identified simultaneously within the ruins area, accurately distinguishing different target types such as adults, children, and wild animals. This effectively eliminates search and rescue interference caused by wild animal activities, allowing rescuers to accurately grasp the number and type of people trapped on site, rationally allocate rescue resources, and comprehensively improve the overall efficiency of post-disaster search and rescue.

[0045] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0046] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for post-disaster life detection and signal recognition based on transfer learning, characterized in that, Includes the following steps: S1. Multi-source signal acquisition: Acquire mixed life detection signals in the post-disaster ruins environment by using at least two of the following sensors: ultra-wideband radar, audio vibration pickup, infrared thermal imaging, and gas sensing. The mixed life detection signals include one or more of the following: respiratory echo, heartbeat micro-vibration, distress call sound wave, human infrared hot spot, and carbon dioxide concentration gradient. S2. Transfer learning residual denoising preprocessing: Introducing a pre-trained transfer denoising residual module to transfer the prior knowledge of pure life signals learned in the source domain scene to the post-disaster target domain, perform residual separation on the collected mixed life detection signals, filter out rubble clutter, mechanical vibration, wind and rain noise interference, and retain weak and effective life signals. S3. Source Domain Model Pre-training: Based on a standard life signal dataset in a non-disaster scenario, a deep learning base model is pre-trained. The base model is used to extract common features of life signals, including respiratory cycle, heart rate, human voice spectrum, temperature threshold, and gas concentration change rate. S4. Adaptive Fine-tuning and Domain Shift Correction in Transfer Learning: Construct a disaster scene dataset in the target domain and adopt a transfer learning strategy of feature transfer + parameter fine-tuning. Freeze the bottom general feature layer of the basic model and only fine-tune the top classification and localization layers with small samples. At the same time, construct a dual discriminator domain adaptive transfer structure to correct the signal intensity distribution shift and the temporal feature distribution shift respectively, reduce the signal distribution difference between the source domain and the target domain, and eliminate scene transfer error. S5. Hierarchical feature adaptive weighting and multimodal signal fusion recognition: Dynamic adaptive weighting is applied to the bottom-level general features, middle-level signal detail features, and top-level scene discrimination features of the fine-tuned model. The feature weights of each layer are adjusted in real time according to the noise intensity of the post-disaster ruins and the degree of signal attenuation. Cross-modal consistency cross-validation is performed on the migration features of multi-source detection to eliminate abnormal identification results caused by single-modal interference. After removing residual interference, the results are output as the existence determination of living organisms, three-dimensional position coordinates, and vital signs parameters. S6. Result Verification and Output: Integrate optical endoscopy or mobile phone signal-assisted verification to generate a visual recognition result that includes the location, depth, and status of the living organism.

2. The method for post-disaster life detection and signal recognition based on transfer learning according to claim 1, characterized in that, The standard life signal dataset for non-disaster scenarios described in S3 includes standard human breathing, heartbeat, sound, infrared, and CO2 signal samples under indoor, outdoor, quiet, and low-interference scenarios. The basic model is a CNN-LSTM hybrid neural network or a Transformer model.

3. The method for post-disaster life detection and signal recognition based on transfer learning according to claim 2, characterized in that, The transfer learning strategy described in S4 is as follows: using maximum mean difference (MMD) or domain adversarial training (DANN) to reduce the signal distribution difference between the source domain and the target domain, and fine-tuning the sample size to 5%-20% of the sample size of the source domain.

4. The post-disaster life detection and signal recognition method based on transfer learning according to claim 3, characterized in that, The freezing of the bottom general feature layer and the fine-tuning of the top classification and positioning layer in S4 are described. The bottom general feature layer is used to extract the general features of the signal in the time domain, frequency domain and spatial domain, and the top classification and positioning layer is used to adapt to the strong noise, strong attenuation and multiple occlusion characteristics of post-disaster ruins.

5. The post-disaster life detection and signal recognition method based on transfer learning according to claim 4, characterized in that, The multimodal signal fusion recognition described in S5 includes: Extract respiratory features (0.1-0.5Hz) and heartbeat features (1-2Hz) from ultra-wideband radar signals; Extract human voice / tapping features from the 200-3000Hz frequency range of the audio signal and calculate the location using Time Difference of Origin (TDOA). It can identify human hotspots at 36±2℃ using infrared signals and detect abnormal concentration gradients in CO2 signals. By using a transfer learning model to weightedly fuse multimodal features, we can distinguish between living organisms and animals, mechanical vibrations, and environmental disturbances.

6. The post-disaster life detection and signal recognition method based on transfer learning according to claim 5, characterized in that, The migration noise reduction residual module described in S2 differs from traditional mathematical filtering. It achieves intelligent noise reduction based on prior knowledge of pure life signals in the source domain, removing only noise signals while fully preserving weak breathing and low-frequency life characteristic signals of the heartbeat.

7. The method for post-disaster life detection and signal recognition based on transfer learning according to claim 6, characterized in that, The cross-modal consistency cross-validation described in S5 specifically involves correcting abnormal results based on the migration features of other effective modalities when any single modal recognition result deviates from the multimodal comprehensive feature threshold, thereby eliminating false positive and false negative recognition errors.

8. The post-disaster life detection and signal recognition method based on transfer learning according to claim 7, characterized in that, The dynamic adaptive weighting described in S5 dynamically increases the weight of weak life features and suppresses the weight of invalid noise features to adapt to the complex and ever-changing detection environment after a disaster, taking into account different conditions of rubble obstruction and signal attenuation.

9. The post-disaster life detection and signal recognition method based on transfer learning according to claim 8, characterized in that, The auxiliary verification described in S6 is: confirming the life form through mobile phone Bluetooth / Wi-Fi signal detection or fiber optic endoscope image, with a positioning error of ≤0.5 meters.

10. The post-disaster life detection and signal recognition method based on transfer learning according to claim 9, characterized in that, Also includes: The transfer learning model enables the simultaneous identification of multiple target life forms, distinguishing between adults, children, and animals, and outputting the independent location and vital signs parameters of each target.

Images