Edge side multi-modal signal hierarchical encryption and compression transmission method and system, medium

Abstract

This invention provides a method, system, and medium for hierarchical encryption and compression transmission of multimodal signals at the edge. It achieves accurate labeling of the core feature region of a single-cycle signal through deep learning, and constructs a hierarchical encryption strategy based on the characteristics of multimodal physiological signals, ensuring that encryption operations only apply to core privacy regions. The differentiated encryption logic ensures the security of privacy information while significantly reducing the encryption computational cost of edge devices through a non-feature redundancy region encryption-free design, effectively resolving the contradiction between the low efficiency of traditional full encryption and the insufficient security of simplified encryption. Based on periodic similarity analysis to identify signal redundancy, and combined with real-time bandwidth status to dynamically adjust the compression strategy, a precise match between transmission efficiency and signal quality is achieved. The improved time-domain processing method achieves high-precision signal peak labeling and period segmentation, providing reliable input for subsequent feature extraction and reducing the complexity requirements of the deep learning model.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method, system, and medium for hierarchical encryption and compression transmission of multimodal signals at the edge. Background Technology

[0002] With the rapid development of telemedicine, health monitoring, and other fields, edge devices (such as digital docks) are widely used in the real-time acquisition and processing of multimodal physiological signals such as cardiac impulse signals (BCG) and photoplethysmography (PPG). These physiological signals are directly related to user health privacy, and secure data transmission is one of the core requirements. At the same time, edge devices are usually deployed in scenarios with limited bandwidth resources, and bandwidth limitation becomes a key bottleneck for signal transmission. Furthermore, the analysis and diagnosis of physiological signals depend on the integrity and accuracy of the signals themselves, and signal fidelity is a fundamental prerequisite for subsequent clinical applications and health assessments.

[0003] Existing technologies have significant shortcomings in practical applications, failing to effectively balance data security, transmission efficiency, and signal fidelity. Specific deficiencies are as follows:

[0004] (1) The encryption strategy lacks specificity, resulting in an imbalance between security and efficiency: Existing full-signal encryption schemes do not consider the periodicity of physiological signals and the differences in the distribution of privacy features. They use the same encryption strength for redundant parts of the signal that do not involve privacy and core privacy parts, resulting in a large amount of encryption computation and a long time consumption. This not only increases the computing power burden of edge devices, but also significantly increases the amount of encrypted data, further exacerbating the pressure on bandwidth transmission. On the other hand, some simplified encryption schemes only reduce the encryption strength of the signal as a whole, which improves efficiency but sacrifices the security of privacy information.

[0005] (2) Poor adaptability of compression strategies and prominent contradiction between efficiency and fidelity: General compression algorithms do not combine with the real-time bandwidth dynamic adjustment strategy of edge devices. In high bandwidth scenarios, excessive compression may lead to bandwidth waste and loss of key signal features. In low bandwidth scenarios, insufficient compression may lead to low transmission efficiency, affecting the accuracy of subsequent analysis. At the same time, existing compression schemes do not utilize the periodic similarity characteristics of physiological signals and do not perform targeted optimization for repetitive periodic signals, thus limiting compression efficiency.

[0006] (3) Insufficient model lightweight adaptation and difficulty in deploying edge devices: Existing schemes that rely on deep learning to achieve signal feature recognition have complex model layers and a large number of parameters, which require high computing power and storage resources from edge devices and make it difficult to achieve real-time processing; while traditional signal period segmentation and peak labeling methods (such as segmentation methods based on fixed thresholds) have low accuracy and cannot provide reliable feature support for subsequent encryption and compression, resulting in a decline in overall processing performance.

[0007] For example, Chinese patent application CN202510797019.1 belongs to the interdisciplinary field of medical information technology and privacy computing, with a core architecture of a dual-end system of "client-doctor". The client encrypts physiological data according to sensitivity levels (fully homomorphic encryption for highly sensitive data, attribute-based encryption for low-sensitivity data). The doctor's end deploys a zero-knowledge neural network with an improved zkReLU activation function, generating intermediate-layer zero-knowledge proofs through forward propagation. During queries, these proofs are integrated into aggregate proofs and verified to ensure that diagnostic results are compliant and do not leak privacy. Its core objective is to solve the balance between medical data privacy protection and computational efficiency in deep learning scenarios. However, this patent application directly uses a deep learning model to extract required features, without combining specific preprocessing operations based on the periodicity and rhythm of physiological signals in the model implementation, and without simplifying the complexity of subsequent processing models through preprocessing and model fusion methods.

[0008] For example, Chinese patent ZL202411000198.3 belongs to the field of data processing technology. Its core method is to achieve health data management through a process of "data collection - privacy preference analysis - hierarchical encryption - local cloud-based off-site storage - identity verification and selective decryption". Specifically, it first collects physiological and biochemical health data and behavioral lifestyle data, formulates hierarchical encryption strategies based on user privacy preferences and data attack frequency, and then implements local-cloud off-site storage by combining data access frequency and terminal storage capacity. During querying, it verifies permissions through multimodal biometrics and selectively decrypts. Its core goal is to improve the efficiency of health data classification management and retrieval, and reduce the risk of privacy leakage. However, this patent directly uses hierarchical encryption based on lifestyle data and physiological and biochemical data tags, fully encrypting physiological data without analyzing the characteristics of physiological data, distinguishing its specific and non-specific parts, or performing feature analysis, hierarchical encryption, and dynamic compression based on periodic similarity for periodic physiological data. It does not simultaneously consider data security and compression effectiveness.

[0009] Therefore, it is necessary to provide a new approach to solve the aforementioned technical problems. Summary of the Invention

[0010] To achieve the above-mentioned objectives and other advantages of the present invention, the first objective of the present invention is to provide a method for hierarchical encryption and compression transmission of multimodal signals at the edge side, comprising the following steps:

[0011] The multimodal physiological signals are preprocessed, and the preprocessed multimodal signals are adapted and enhanced to obtain peak indexes. Based on the peak indexes, standardized single-cycle signals are formed.

[0012] Spatiotemporal feature extraction and similarity evaluation are performed on the standardized single-cycle signal to obtain feature core region annotation and similarity criteria;

[0013] Based on the annotation of the core feature region, a differentiated strategy is adopted for hierarchical encryption;

[0014] Based on real-time bandwidth detection results and periodic similarity criteria, layered compression is performed on encrypted data.

[0015] Furthermore, the multimodal physiological signals include BCG signals and PPG signals.

[0016] Furthermore, the step of adapting and enhancing the preprocessed multimodal signal includes:

[0017] By adapting and enhancing the multimodal signal, the J peak of the BCG signal and the contraction peak of the PPG signal are accurately located, and the peak index is obtained.

[0018] Furthermore, the step of accurately locating the J-peak of the BCG signal and the contraction peak of the PPG signal through multimodal signal adaptation and enhancement includes:

[0019] For BCG signal adaptation, a 7-point differential operator is used to highlight the performance of J-peak slope acquisition. The calculation formula is as follows:

[0020]

[0021]

[0022] in, For point i after denoising, The signal after BCG differentiation enhancement. This is to enhance the final signal of BCG;

[0023] For PPG signal adaptation, a low-pass smoothing and amplitude normalization enhancement structure is adopted to highlight the shrinkage peak characteristics. The calculation formula is as follows:

[0024]

[0025]

[0026] in, The signal is after PPG low-pass smoothing. This represents the maximum value of the signal after PPG low-pass smoothing. This is the final enhancement signal for PPG.

[0027] Furthermore, the step of accurately locating the J-peak of the BCG signal and the contraction peak of the PPG signal through multimodal signal adaptation and enhancement also includes:

[0028] A threshold coefficient is set based on the differences in characteristics between the two signals to adapt to the peak characteristics of different signals. The calculation formula is as follows:

[0029]

[0030] in, As the initial threshold, This represents the maximum value of the enhanced signal.

[0031] The threshold is adjusted based on historical effective peak feedback to balance stability and real-time adaptability. The calculation formula is as follows:

[0032]

[0033]

[0034] in, The average amplitude of the first three valid peak values. This represents the original amplitude of the kth effective peak. The updated threshold;

[0035] Using the preset peak interval corresponding to cardiac physiological activity as a constraint, false peaks are eliminated, and the peak intervals are interpolated to complete the peaks.

[0036] Furthermore, the step of forming a standardized single-cycle signal based on the peak index includes:

[0037] Centered on the peak index, the truncation range is set according to the physiological waveform characteristics of the dual signals to ensure complete cycle coverage;

[0038] The point completion formula is obtained by using lightweight linear interpolation:

[0039]

[0040] in, This represents the j-th interpolation point of the c-th single-cycle signal. This is the index of the original signal corresponding to interpolation point j.

[0041] Furthermore, the steps of extracting spatiotemporal features and evaluating similarity of the standardized single-cycle signal include:

[0042] Local spatial feature extraction models are used to adapt local feature extraction for BCG and PPG respectively, and local spatial features are extracted from single-cycle signals.

[0043] Bi-LSTM was used to capture long-term time-dependent data and the physiological rules of the J-wave interval were used to correct and extract BCG time-series features.

[0044] A miniature temporal constraint module is used to suppress artifacts, and TCN is used to capture local temporal trends and extract PPG temporal features.

[0045] A similarity determination mechanism based on spatiotemporal feature fusion provides similarity criteria for dynamic compression.

[0046] Furthermore, the local spatial feature extraction model is implemented through two convolutional blocks, channel attention, global pooling, and fully connected layers; wherein each convolutional block includes a convolutional layer, a BN layer, a ReLU layer, and a pooling layer;

[0047] For BCG signals, details are captured progressively through kernel convolution and key feature channels are enhanced through channel attention, focusing on core local features; the channel attention layer is as follows:

[0048]

[0049] Where L is the length of the time dimension of the feature map after pooling; This is the global average pooling value for the l-th channel; Weights for the fully connected layer of the SE module; For the SE module bias vector; The normalized weights for J-wave amplitude; The feature map is after channel attention weighting;

[0050] For PPG signals, kernel convolution is used to capture peak shapes and spatial attention is used to enhance peak positions, thus clustering core local features; the spatial attention layer is as follows:

[0051]

[0052] Where W is the locally normalized window size; The feature map is locally normalized; C is the number of channels in the feature map. This is the spatial weight coefficient vector; Spatial weight bias; The spatial weight of the j-th time position; The feature map after spatial attention weighting.

[0053] Furthermore, the step of using Bi-LSTM to capture long-range temporal dependencies and correcting for BCG temporal features by J-wave interval physiological rules includes:

[0054] A two-layer bidirectional LSTM was used to capture the correlation between the J-wave interval and the period before and after the J-wave.

[0055] J-wave interval calculation: The duration of the J-wave interval was obtained.

[0056] like If the J-wave interval exceeds the normal heart rate, it is marked as an abnormal cycle;

[0057] LSTM output for abnormal periods The average of the output from the first three normal cycles is used for correction. k is a normal periodic index;

[0058] The corrected hidden state is mapped using global pooling and a fully connected layer, and the output is... ,formula:

[0059] .

[0060] Furthermore, the steps of using a micro-temporal constraint module to suppress artifacts and TCN to capture local temporal trends and extract PPG temporal features include:

[0061] Calculate the fluctuation ratio of the maximum local characteristic values ​​of adjacent periods:

[0062]

[0063] Calculate the characteristic slope of adjacent periods:

[0064]

[0065] like and If the periodicity is 0, it is marked as an artifact period, and the mark matrix is ​​output. ;

[0066] Constructing the causal convolution formula: ;

[0067] For artifact periods where the marker matrix M(t)=1, the TCN output is replaced by a weighted average of the output from the previous normal period: , where (t-1) is the normal period;

[0068] After collaboration, the features are global pooled and fully connected to output the final product. ,formula:

[0069] .

[0070] Furthermore, the similarity determination mechanism based on spatiotemporal feature fusion, which provides similarity criteria for dynamic compression, includes the following steps:

[0071] Based on the rule-based labeling results, the normal periodic features are weighted and enhanced, while the artifact / abnormal periodic features are weighted and weakened.

[0072] BCG Fusion:

[0073]

[0074] in, , During abnormal periods, ;

[0075] PPG Fusion:

[0076]

[0077] in, , During the artifact period, ;

[0078] Based on the signal physiological feature template library, similarity is calculated only for features that conform to the rules; the similarity calculation formula is:

[0079]

[0080] in, Template features;

[0081] Candidate features are input into a 2-layer fully connected network, which outputs a similarity score.

[0082] Furthermore, the step of performing hierarchical encryption based on the annotation of the feature core region and using a differentiated strategy includes:

[0083] Input the annotation results and parse the key parameters, including the feature core region type and the reference point index;

[0084] The sampling point interval of the feature core area is calculated based on the benchmark index, and the range is set in combination with the physiological characteristics differences of multimodal signals;

[0085] Based on the mask output by the feature core region mapping, region segmentation and algorithm matching are performed on multimodal signals.

[0086] Furthermore, for BCG signals, the characteristic core area includes clinical diagnostic core areas such as J wave amplitude, J wave interval, and K wave depth, covering the J wave initiation to the K wave termination, corresponding to the key diagnostic segment of the standardized single-cycle signal; for PPG signals, the characteristic core area includes key areas related to blood oxygen / heart rate calculation such as the systolic peak and dicrotic wave, covering the rising edge of the systolic peak to the end of the dicrotic wave, corresponding to the key hemodynamic characteristic segment;

[0087] The mask based on the feature core region mapping output, the steps for performing region segmentation and algorithm matching on multimodal signals include:

[0088] For BCG signals, the AES-128 simplified algorithm is used in the feature core area, while the original sampling point data is directly output in the non-feature redundant area.

[0089] For PPG signals, the simplified version of the DES algorithm is used in the feature core region, while the original sampling point data is directly output in the non-feature redundant region.

[0090] Furthermore, it also includes streamlining the encryption execution steps:

[0091] The sampling point data of the feature core area is grouped into 16-byte groups to obtain grouped data;

[0092] Perform an XOR operation between the grouped data and the derived round key;

[0093] The BCG signal is processed through 8 rounds of byte substitution, row shifting, column mixing, and round key addition, omitting the 5th and 6th rounds of column mixing operations in traditional AES.

[0094] For PPG signals, perform the first 6 rounds of complete transformation and the last 2 rounds of simplified transformation, omitting the 9-16 rounds of extended permutation and subkey XOR redundancy operation in traditional DES;

[0095] The results of each round of calculations are concatenated to obtain the encrypted core feature data, which is then concatenated with the original data of the non-feature redundant area to form a complete encrypted signal.

[0096] Furthermore, the step of performing layered compression on the encrypted data based on real-time bandwidth detection results and periodic similarity criteria includes:

[0097] Obtain real-time bandwidth values ​​and, in conjunction with clinical data transmission bandwidth requirements, classify and label the real-time bandwidth.

[0098] For low-bandwidth data, similarity criteria are used to accurately distinguish between redundant similar periods and abnormal periods.

[0099] For non-confidential data, the compression ratio is adjusted according to the bandwidth. For encrypted data, secondary compression is performed only on encrypted data with similar periods based on the period similarity criterion when low bandwidth is detected, while encrypted single-period data is kept uncompressed to ensure accuracy.

[0100] Furthermore, the step of accurately distinguishing redundant similar periods from abnormal periods for low-bandwidth data by combining similarity criteria includes:

[0101] Obtain the single-cycle similarity score, calculate the cumulative similarity value of the current cycle and the previous 4 cycles, and make a judgment based on the reverse similarity sequence;

[0102] For similar sequences, select the first period in the sequence as the reference period, and calculate the difference in encrypted data between the current period and the reference period;

[0103] For dissimilar sequences, they are directly marked as independent periods, and the difference value is not calculated;

[0104] The steps of adjusting the compression ratio of non-encrypted data according to bandwidth, and performing secondary compression on encrypted data with similar periods only when low bandwidth is detected, while keeping encrypted single-period data uncompressed to ensure accuracy, include:

[0105] Based on the output data type of hierarchical encryption, two types of compressed objects are divided and matched with lightweight algorithms to adapt to edge computing power;

[0106] Based on bandwidth level and similarity markers, dynamically adjust the compression ratio and perform compression;

[0107] The compressed unencrypted and encrypted data are packaged according to periodic indexes, data types, and data and checksum formats, and sent to the cloud for processing or local archiving.

[0108] The second objective of this invention is to provide an edge-side multimodal signal hierarchical encryption and compression transmission system, which utilizes the aforementioned method and includes a signal preprocessing and period segmentation unit, a spatiotemporal feature extraction unit, a hierarchical encryption unit, and a dynamic compression unit; wherein,

[0109] The signal preprocessing and period segmentation unit is used to preprocess the multimodal physiological signal, adapt and enhance the preprocessed multimodal signal, obtain the peak index, and form a standardized single-period signal based on the peak index.

[0110] The spatiotemporal feature extraction unit is used to extract spatiotemporal features and evaluate similarity of the standardized single-cycle signal to obtain feature core area annotation and similarity criteria.

[0111] The hierarchical encryption unit is used to perform hierarchical encryption based on the annotation of the feature core area and using a differentiated strategy.

[0112] The dynamic compression unit is used to perform layered compression on encrypted data based on real-time bandwidth detection results and periodic similarity criteria.

[0113] A third object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0114] Compared with the prior art, the beneficial effects of the present invention are:

[0115] This invention achieves a balance between security and efficiency through the collaborative approach of feature core area annotation and hierarchical encryption: Deep learning is used to accurately annotate the feature core areas of single-cycle signals, and a hierarchical encryption strategy is constructed by combining the characteristics of multimodal physiological signals, ensuring that encryption operations only apply to core privacy regions. The differentiated encryption logic not only ensures the security of privacy information but also significantly reduces the encryption computational overhead of edge devices through the encryption-free design of non-feature redundant areas, effectively resolving the contradiction between the low efficiency of traditional full encryption and the insufficient security of simplified encryption.

[0116] This invention achieves a balance between transmission efficiency and signal fidelity through periodic similarity determination and dynamic compression: Signal redundancy is identified based on periodic similarity analysis, and the compression strategy is dynamically adjusted according to real-time bandwidth status, achieving a precise match between transmission efficiency and signal quality. In high-bandwidth scenarios, it avoids resource waste caused by excessive compression; in low-bandwidth scenarios, it reduces redundant transmission through similarity period optimization; and through targeted protection of the core feature area, it ensures that key signal features required for subsequent analysis are not lost, thus balancing transmission efficiency and diagnostic reliability.

[0117] This invention achieves lightweight adaptation for edge scenarios through preprocessing and model fusion: the improved temporal processing method enables high-precision signal peak labeling and period segmentation, providing reliable input for subsequent feature extraction and reducing the complexity requirements of deep learning models. Through the fusion design of simplified model structure and optimized preprocessing, the computational and storage requirements of edge devices are reduced, ensuring the feasibility of real-time processing of multimodal signals and adapting to low-resource edge application scenarios.

[0118] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0119] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0120] Figure 1 A schematic diagram illustrating the principle of hierarchical encryption and compression transmission of multimodal signals at the edge.

[0121] Figure 2 Flowchart of a hierarchical encryption and compression transmission method for multimodal signals at the edge;

[0122] Figure 3 This is a flowchart of the preprocessing of multimodal physiological signals.

[0123] Figure 4 Process for multimodal signal adaptation, enhancement, and precise localization of BCG signal J-peak and PPG signal shrinkage peak Figure 1 ;

[0124] Figure 5 Process for multimodal signal adaptation, enhancement, and precise localization of BCG signal J-peak and PPG signal shrinkage peak Figure 2 ;

[0125] Figure 6 A standardized process flow chart for cycle segmentation;

[0126] Figure 7 A flowchart for spatiotemporal feature extraction and similarity assessment;

[0127] Figure 8 A structural diagram of a local spatial feature extraction model;

[0128] Figure 9 Flowchart for local spatial feature extraction;

[0129] Figure 10 This is a structural diagram of the BCG temporal feature extraction model;

[0130] Figure 11 Flowchart for BCG temporal feature extraction;

[0131] Figure 12 Here is a structural diagram of the PPG time-series feature extraction model;

[0132] Figure 13 Flowchart for PPG time series feature extraction;

[0133] Figure 14 Flowchart for spatiotemporal fusion similarity determination;

[0134] Figure 15 Here is a flowchart of the hierarchical encryption process;

[0135] Figure 16 Flowchart for encryption region division and algorithm selection;

[0136] Figure 17 To simplify the encryption execution flowchart;

[0137] Figure 18 This is a flowchart of the dynamic compression process;

[0138] Figure 19 Here is a flowchart for the analysis and determination of similarity criteria;

[0139] Figure 20 Flowchart for matching compressed regions with algorithms;

[0140] Figure 21 Block diagram of a multimodal signal hierarchical encryption and compression transmission system at the edge;

[0141] Figure 22 This is a schematic diagram of a computer device.

[0142] Figure 23 This is a schematic diagram of a computer-readable storage medium. Detailed Implementation

[0143] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0144] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0145] The drawing numbers in this application are only used to distinguish the steps in the scheme and are not used to limit the execution order of the steps. The specific execution order is as described in the specification.

[0146] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0147] This invention is based on a "preprocessing-feature extraction-encryption-compression" architecture to achieve efficient processing of multimodal signals in edge scenarios, while balancing data security, transmission efficiency and signal fidelity.

[0148] Example 1

[0149] A hierarchical encryption and compression transmission method for multimodal signals at the edge, such as Figure 1 , Figure 2 As shown, it includes the following steps:

[0150] S100. Preprocess the multimodal physiological signal, adapt and enhance the preprocessed multimodal signal to obtain the peak index, and form a standardized single-cycle signal based on the peak index.

[0151] In this embodiment, the multimodal physiological signals include BCG signals and PPG signals.

[0152] This embodiment implements preprocessing of multimodal physiological signals (BCG and PPG signals), constructs a lightweight preprocessing module adapted to both BCG and PPG signals, and achieves a unification of high-precision feature anchoring and lightweight processing through multimodal signal adaptation enhancement, dynamic threshold peak detection, and standardized period segmentation. The entire process of the signal preprocessing and period segmentation unit includes preprocessing, dynamic peak detection, and period segmentation.

[0153] In some embodiments, such as Figure 3 As shown, the preprocessing steps for multimodal physiological signals include:

[0154] S110: Perform resampling, baseline drift removal, power frequency interference suppression, and Z-fraction normalization on the sampled original multimodal physiological signals, and output a denoised signal;

[0155] For example, the original signal sampled from 125 to 500 Hz is subjected to 500 Hz resampling, baseline drift removal (moving average filtering), 50 Hz power line interference suppression (IIR notch filtering), and Z-score normalization to output a denoised signal.

[0156] The step of adapting and enhancing the preprocessed multimodal signal includes:

[0157] S120. By adapting and enhancing the multimodal signal, the J peak of the BCG signal and the contraction peak of the PPG signal are accurately located to obtain the peak index.

[0158] This embodiment designs a differentiated enhancement operator to achieve adaptation to BCG and PPG signals. Specifically, as follows: Figure 4 As shown, the step of accurately locating the J-peak of the BCG signal and the contraction peak of the PPG signal through multimodal signal adaptation and enhancement includes:

[0159] Multimodal signal adaptation enhancement includes two parts: BCG and PPG. For BCG signal adaptation, a timing processing algorithm is implemented based on the "differential-squared" core logic. A 7-point differential operator is used to highlight the performance of J-peak slope acquisition. The calculation formula is as follows:

[0160]

[0161]

[0162] in, For point i after denoising, The signal after BCG differentiation enhancement. This is the final enhancement signal for BCG.

[0163] S122. For PPG signal adaptation, a "low-pass smoothing-amplitude normalization" enhancement structure is designed to highlight the shrinkage peak characteristics. The calculation formula is as follows:

[0164]

[0165]

[0166] in, The signal is after PPG low-pass smoothing. This represents the maximum value of the signal after PPG low-pass smoothing. This is the final enhancement signal for PPG.

[0167] To address the shortcomings of traditional time-domain algorithms that use fixed thresholds and have a high false negative rate for amplitude fluctuation signals, dynamic threshold detection is achieved by using differentiated initial thresholds, iterative updates, and physiological constraints.

[0168] In some embodiments, such as Figure 5 As shown, the step of accurately locating the J-peak of the BCG signal and the contraction peak of the PPG signal through multimodal signal adaptation and enhancement further includes:

[0169] Differentiation initial threshold setting, S123: Set threshold coefficients based on the differences in characteristics between the two signals to adapt to the peak characteristics of different signals. The calculation formula is as follows:

[0170]

[0171] To enhance the maximum value of the signal, the BCG waveform is adjusted based on the differences in waveform characteristics between BCG and PPG. =0.45, taking PPG ;

[0172] Iterative updates are implemented, S124, adjusting the threshold based on historical effective peak feedback to balance stability and real-time adaptability. The calculation formula is as follows:

[0173]

[0174]

[0175] in, The average amplitude of the first three valid peak values. This represents the original amplitude of the kth effective peak. To update the threshold, a weighting ratio of 0.6 to 0.4 is used to balance historical experience and real-time signals.

[0176] Physiological constraint supplementation, S125, using the preset peak interval corresponding to cardiac physiological activity as a constraint, eliminates false peaks, and at the same time completes the result by interpolation of the peak interval before and after.

[0177] For example, using the peak interval (167-500 sampling points) corresponding to a heart rate of 60-180 bpm as a constraint, false peaks are eliminated. If the index difference between adjacent peaks exceeds the range of 167-500, it is judged as a false peak and eliminated, while the difference is filled by interpolation of the peak intervals before and after. That is, actual signals of heart rates outside this normal range (167-500) are uniformly marked as risk signals, and all of them need to be uploaded / recorded, without being included in compression considerations.

[0178] This embodiment uses the peak index as a benchmark to extract complete physiological signal cycles in a differentiated manner and unifies the length through interpolation; finally, it outputs a 400-point standardized single-cycle signal and peak index, providing a unified length input for the subsequent feature extraction module and reducing the parameter redundancy of the model in adapting to the input length.

[0179] Specifically, such as Figure 6 As shown, the step of forming a standardized single-cycle signal based on the peak index includes:

[0180] Differentiated truncation: S130, with the peak index as the center, the truncation range is set according to the physiological waveform characteristics of the dual signals to ensure complete cycle coverage;

[0181] The interception range of the BCG signal is: It covers the entire JKT band; the intercept range for PPG signals is... It covers the systolic peak-diastolic trough-diplophonic wave. Among them, This is the peak index marker for the corresponding signal.

[0182] Linear interpolation, S140, uses lightweight linear interpolation to complete the points (e.g., to 400 points), providing a uniform input length for subsequent feature extraction modules and reducing parameter redundancy in model adaptation to input length. The completion formula is:

[0183]

[0184] in, Let j be the j-th interpolation point (j=1,...,400) of the c-th single-cycle signal. This is the index of the original signal corresponding to interpolation point j.

[0185] This embodiment achieves the function of outputting standardized periodic segments of dual signals through differentiated truncation and linear interpolation.

[0186] S200. Perform spatiotemporal feature extraction and similarity evaluation on the standardized single-cycle signal to obtain feature core area annotation and similarity criteria;

[0187] This embodiment designs a spatiotemporal feature extraction unit to extract spatiotemporal features from multimodal physiological signals (BCG and PPG signals). It consists of a local spatial feature extraction module, a differentiated temporal feature modeling module, and a spatiotemporal fusion similarity determination module. This unit extracts spatiotemporal features and evaluates similarity for the BCG / PPG standardized single-cycle signals output by the signal preprocessing and periodic segmentation units. Its core steps include: a dual-branch local feature extraction architecture based on "CNN + attention" designed to address the morphological differences between the two signals (BCG and PPG), solving the problem of low feature recognition in traditional general architectures; constructing differentiated temporal modeling strategies based on the temporal characteristics of the two signals (BCG: Bi-LSTM + J-wave interval rule; PPG: micro-temporal constraint + TCN), solving the problem of poor generalization in deep learning; and proposing a similarity determination mechanism for spatiotemporal feature fusion, providing similarity criteria for subsequent dynamic compression.

[0188] In some embodiments, such as Figure 7 As shown, the steps for extracting spatiotemporal features and evaluating similarity of the standardized single-cycle signal include:

[0189] S210. Local spatial feature extraction models are used to adapt to the local feature extraction of BCG and PPG respectively, and local spatial features are extracted from single-cycle signals.

[0190] To adapt to the morphological differences between BCG and PPG, improve the recognition of local features, and avoid missing key physiological features, this embodiment designs a "CNN + differential attention" structure to adapt to the local feature extraction of BCG and PPG respectively. It extracts 64-dimensional local spatial features from a single-cycle signal. The designed local spatial feature extraction model layer structure is as follows: Figure 8 As shown.

[0191] The local spatial feature extraction model is implemented using two convolutional blocks + channel attention + global pooling + fully connected layers, configured with Same Padding and a convolutional stride of 1. Each convolutional block includes a convolutional layer + a batch normalization (BN) layer + a ReLU layer + a pooling layer. The formulas for each layer structure are explained below:

[0192] Convolutional layers:

[0193]

[0194] in, Let c be the channel index, j be the temporal index, and K be the kernel size. Let C be the weight matrix of the convolutional kernel of the l-th layer, C be the number of input channels, and I be the input features of the convolutional layer. [K / 2] is the bias vector for the l-th convolution layer; [K / 2] is the padding offset to ensure that the output length is equal to the input length.

[0195] BN layer + ReLU layer:

[0196]

[0197]

[0198] in, This is the batch mean of the c-th channel in the l-th layer; Let be the batch variance of the c-th channel in the l-th layer; is the scaling factor for the c-th channel of the l-th layer; is the offset coefficient of the c-th channel in the l-th layer; It is the numerical stability constant; This is the output feature map after ReLU activation.

[0199] Pooling layer:

[0200]

[0201] in, P represents the output feature map of the l-th pooling layer; P is the pooling kernel size. This is a pooled window index with a step size of 2.

[0202] Global pooling layer:

[0203]

[0204] in, is the global pooling output vector; L is the time dimension length of the attention layer output feature map.

[0205] Fully connected layer:

[0206]

[0207] in, Output vectors for local features; This is the weight matrix of the fully connected layer; This is the bias vector for the fully connected layer.

[0208] like Figure 9 As shown in Figure S211, for the BCG signal, the design of "middle kernel convolution progressively capturing details + channel attention to enhance key feature channels" focuses on core local features;

[0209] S212. For PPG signals, in order to adapt to the spatial characteristics of the wide peak of the PPG contraction peak and the shallow notch of the diphtheria wave, the design of "mid-core convolution to capture peak shape + spatial attention to enhance peak position" is adopted to focus on core local features such as the contraction peak and diphtheria wave.

[0210] The implementation of channel attention and spatial attention is explained below. Table 1 records the specific parameters of the spatial feature extraction models for BCG and PPG signals.

[0211] Channel Attention Layer—BCG Signal Usage:

[0212]

[0213]

[0214]

[0215] Where L is the length of the time dimension of the feature map after pooling; This is the global average pooling value for the l-th channel; Weights for the fully connected layer of the SE module; For the SE module bias vector; The normalized weights for J-wave amplitude; The feature map after channel attention weighting.

[0216] Spatial Attention Layer—PPG Signal Usage:

[0217]

[0218]

[0219]

[0220] Where W is the locally normalized window size; The feature map is locally normalized; C is the number of channels in the feature map. This is the spatial weight coefficient vector; Spatial weight bias; The spatial weight of the j-th time position; The feature map after spatial attention weighting.

[0221] Table 1. Layer parameters used by the local spatial feature extraction model for BCG / PPG signals

[0222]

[0223] S220. Bi-LSTM is used to capture long-term time dependence and J-wave interphase physiological rule constraints to extract BCG time features;

[0224] To address the characteristics of dual-signal timing and physiological cycles, and considering the poor generalization problem inherent in full deep learning, a timing module combining physiological rule constraints and timing detail capture is designed and proposed.

[0225] The core of BCG timing is the J-wave interphase rhythm (premature beats and conduction block manifest as J-wave interphase abrupt changes). It employs a design of "Bi-LSTM capturing long-term timing dependencies + J-wave interphase physiological rule constraint correction" to address the misjudgment problem of abnormal rhythms by LSTM. The BCG timing feature extraction model is as follows: Figure 10 As shown. Specifically, as Figure 11 As shown, the steps for extracting BCG time-series features using Bi-LSTM to capture long-range time-series dependencies and J-wave interphase physiological rule constraints include:

[0226] Bi-LSTM long-range time series modeling: S221, using a 2-layer bidirectional LSTM with 160 hidden units, capturing the correlation before and after the J-wave interval, the core gating formula of Bi-LSTM is explained as follows;

[0227] The Gate of Oblivion (Preserving the Rhythm of History) ;

[0228] Input gate (update current features): , ;

[0229] Cell status update: ;

[0230] Output gate (bidirectional fusion): , Bidirectional output ;

[0231] in, For the local characteristics of the signal in period t, The state was hidden in the previous moment. / For weights / biases, It is Sigmoid. It is the hyperbolic tangent.

[0232] J-wave interval rule constraint correction: J-peak index based on signal preprocessing and period segmentation unit output (Index of J peak in period t), the LSTM output is corrected by constraining physiological rhythms. The specific steps include:

[0233] S222 and J-wave interval calculation: The duration of the J-wave interval was obtained.

[0234] Physiological range constraint: A normal heart rate of 60-180 bpm corresponds to a J-wave interval of 0.33-1.0 s, S223, if If it exceeds this range, it will be marked as an abnormal cycle;

[0235] Feature correction: S224, LSTM output for abnormal periods The average of the output from the first three normal cycles is used for correction. k is a normal periodic index;

[0236] Temporal feature output: S225, the corrected hidden state is processed by global pooling + fully connected mapping, output ,formula: (256-dimensional bidirectional output).

[0237] S230: A miniature temporal constraint module is used to suppress artifacts and TCN is used to capture local temporal trends and extract PPG temporal features;

[0238] The core of PPG time series analysis is amplitude trend and artifact interference (motion artifacts manifest as abrupt amplitude changes). A design employing a "miniature time series constraint module to suppress artifacts + TCN to capture local time series trends" addresses the weakness of pure TCN in resisting artifacts. The PPG time series feature extraction model is as follows: Figure 12 As shown.

[0239] Miniature temporal constraint module (artifact detection and labeling): A parameterless rule module that detects artifacts based on PPG amplitude and slope characteristics. Specifically, such as... Figure 13 As shown, the steps of using a micro-temporal constraint module to suppress artifacts and TCN to capture local temporal trends and extract PPG temporal features include:

[0240] Amplitude fluctuation calculation, S231, calculate the fluctuation ratio of local characteristic maximum values ​​of adjacent periods:

[0241]

[0242] Slope abrupt change calculation, S232, calculate the characteristic slope of adjacent periods:

[0243]

[0244] Artifact marker: S233 and The periodicity of the artifact is marked, and the mark matrix is ​​output. (1 indicates an artifact, 0 indicates normal).

[0245] Local Temporal Enhancement Modeling of TCN: A two-layer causal TCN is adopted, which expands the receptive field through dilated convolution to adapt to the local amplitude trend of PPG (contraction peak-diastolic valley transition). The core design is as follows:

[0246] S234. Construct the causal convolution formula: ;

[0247] Rule and TCN Collaboration: S235, For artifact periods of the marker matrix M(t)=1, the TCN output is replaced by a weighted average of the output of the previous normal period. ;

[0248] Temporal feature output: S236, the collaborative features are processed by global pooling and fully connected to output... formula:

[0249] .

[0250] S240. A similarity determination mechanism based on spatiotemporal feature fusion provides similarity criteria for dynamic compression.

[0251] Specifically, such as Figure 14 As shown, the similarity determination mechanism based on spatiotemporal feature fusion, which provides similarity criteria for dynamic compression, includes the following steps:

[0252] Physiological rule-weighted spatiotemporal fusion: S241, based on the rule labeling results of the differential temporal feature modeling module, normal periodic features are weighted and enhanced, while artifact / abnormal periodic features are weighted and weakened.

[0253] BCG Fusion:

[0254]

[0255] in, , (The local peak shape characteristics of BCG are as important as the J-wave inter-epoch timing); abnormal periodicity ;

[0256] PPG Fusion:

[0257]

[0258] in, , (Local peak shape is as important as temporal trend); during artifact periodization, ;

[0259] Rule-assisted cosine similarity judgment: S242, combining the signal physiological feature template library, calculate similarity only for features that conform to the rules; the similarity calculation formula is:

[0260]

[0261] in, Template features (including normal / abnormal templates); retained Candidate features.

[0262] Fully connected output: S243, candidate feature input, 2-layer fully connected network, output similarity score.

[0263] Hidden layer: , ;

[0264] Output layer: , .

[0265] S300. Based on the annotation of the core feature area, a differentiated strategy is adopted for hierarchical encryption;

[0266] The hierarchical encryption unit, based on the feature core region annotation information output by the aforementioned unit, implements a differentiated hierarchical encryption method adapted to the sensitivity of physiological characteristics of multimodal signals (BCG, PPG) in edge scenarios through precise feature core region mapping and differentiated encryption. Its main components include a feature core region mapping module and a differentiated encryption module. Specifically, the differentiated encryption module employs a high-security simplified encryption algorithm for the BCG feature core region and a medium-security simplified encryption algorithm for the PPG feature core region, directly outputting non-feature redundant regions to save computational power.

[0267] In this embodiment, the core area of ​​privacy-related physiological features is referred to as the feature core area. The feature core area refers to the concentrated area of ​​physiological features in the BCG signal containing JKT wave groups and the PPG signal containing systolic peak-diplophonic wave. It is privacy-sensitive because personal health privacy information can be extracted. On the other hand, the non-feature redundancy area refers to the flat or noisy area in the signal without clear physiological features and does not carry privacy-related information.

[0268] The feature core region mapping module completes the conversion of labeled feature information into signal sampling points, such as... Figure 15 As shown, the specific steps include parsing the annotation information and mapping the coordinates of the sampling points.

[0269] Annotation information analysis: S310, the annotation results output by the input spatiotemporal feature extraction unit, and analysis of key parameters. This includes the feature core region type, reference point index (BCG J-peak index, PPG contraction peak index, and feature core region half-width);

[0270] Sampling point coordinate mapping: S320, calculate the sampling point interval of the feature core area based on the benchmark index, and set the range in combination with the physiological characteristics differences of multimodal signals; for BCG signals, the feature core area includes clinical diagnostic core areas such as J wave amplitude, J wave interval, and K wave depth, covering the J wave initiation to the K wave end, corresponding to the key diagnostic segment of the standardized single-cycle signal; for PPG signals, the feature core area includes key areas related to blood oxygen / heart rate calculation such as systolic peak and dicrotic wave, covering the rising edge of the systolic peak to the end of the dicrotic wave, corresponding to the key hemodynamic feature segment.

[0271] To address the computational overload caused by traditional encryption methods that use uniform encryption strength across all signals, a differentiated encryption logic is designed based on the varying signal sensitivity of multimodal signals. A simplified encryption algorithm is adopted to adapt to edge computing. Specific implementation includes encryption region division and algorithm selection, as well as simplified encryption execution. Figure 16 As shown.

[0272] Encryption region division and algorithm selection: S330, based on the mask output by the feature core region mapping module, performs region division and algorithm matching on multimodal signals.

[0273] S331. For BCG signals, the AES-128 simplified algorithm is used in the feature core area, and the original sampling point data is directly output in the non-feature redundant area.

[0274] S332. For PPG signals, the simplified version of the DES algorithm is used in the feature core area, while the original sampling point data is directly output in the non-feature redundant area.

[0275] In some embodiments, such as Figure 17 As shown, it also includes S340, a simplified encryption execution step:

[0276] Data grouping, S341: Group the sampling point data of the feature core area into 16-byte groups (use zero padding if less than 16 bytes) to obtain grouped data;

[0277] Round key addition, S342, XOR the block data with the derived round key;

[0278] Simplify round operations: S343 performs 8 rounds of "byte substitution-row shifting-column mixing-round key addition" on the BCG signal, omitting the 5th and 6th rounds of column mixing operations in traditional AES;

[0279] S344. For the PPG signal, perform the first 6 rounds of complete transformation and the last 2 rounds of simplified transformation, omitting the 9-16 rounds of extended permutation and subkey XOR redundancy operation in the traditional DES.

[0280] Finally, the encrypted output is processed. In step S345, the results of each round of calculations are concatenated to obtain the encrypted core feature data, which is then concatenated with the original data of the non-feature redundant area to form a complete encrypted signal.

[0281] S400 performs layered compression on encrypted data based on real-time bandwidth detection results and periodic similarity criteria.

[0282] The dynamic compression unit, based on real-time bandwidth and periodic similarity criteria (output of the spatiotemporal feature extraction unit), dynamically compresses encrypted data and outputs the processed data. This achieves a balance between transmission efficiency and signal fidelity in data compression. It accurately identifies redundant periods based on similarity criteria and dynamically adjusts the compression strategy in conjunction with real-time bandwidth. Its main components include a real-time bandwidth detection module, a periodic similarity matching module, and a hierarchical compression execution module, such as... Figure 18 As shown.

[0283] Real-time bandwidth detection module: Collects and classifies link bandwidth in real time. It calls the underlying real-time bandwidth data interface of the edge computing platform system (S410) to obtain real-time bandwidth values. Based on clinical data transmission bandwidth requirements, it classifies and labels the real-time bandwidth (high bandwidth: general compression; medium bandwidth: high compression ratio; low bandwidth: secondary compression).

[0284] Filtering redundant periods and determining the baseline period: S420, for low-bandwidth data, combining the similarity criteria of the spatiotemporal feature extraction unit, accurately distinguishes between "redundant similar periods" and "abnormal periods," providing a redundancy basis for secondary compression under low bandwidth and avoiding over-compression of abnormal periods. Specific implementation steps include similarity criterion parsing and determination, baseline period selection and difference value calculation, such as... Figure 19 As shown.

[0285] The similarity criterion analysis and judgment part uses S421 to obtain the single-cycle similarity score output by the spatiotemporal feature extraction unit, calculates the cumulative similarity value of the current cycle and the previous 4 cycles, and makes a judgment based on the reverse similarity sequence.

[0286] In the section on baseline period selection and difference value calculation, S423, for similar sequences, the first period in the sequence is selected as the baseline period, and the difference value of the encrypted data between the current period and the baseline period is calculated; S424, for dissimilar sequences, they are directly marked as "independent periods" and the difference value is not calculated.

[0287] S430. Adjust the compression ratio of non-confidential data according to bandwidth. For encrypted data, only perform secondary compression on encrypted data with similar periods based on the period similarity criterion when low bandwidth is detected, and keep encrypted single-period data uncompressed to ensure accuracy.

[0288] The compression execution module performs differentiated compression on encrypted and unencrypted data. For unencrypted data, the compression ratio is adjusted according to bandwidth. For encrypted data, secondary compression is performed only when low bandwidth is detected, based on a periodic similarity criterion, while single-period encrypted data remains uncompressed to ensure accuracy. The specific implementation steps include three parts: matching the compression region with the algorithm, dynamically adjusting and executing the compression ratio, and encapsulating the compressed data. Figure 20 As shown.

[0289] The compression region and algorithm matching part, S431, divides the output data type of the hierarchical encryption unit into two types of compression objects and matches them with lightweight algorithms to adapt to edge computing power. The dynamic adjustment and execution part of the compression ratio, S432, dynamically adjusts the compression ratio and executes compression based on bandwidth level and similarity markers. The compressed data encapsulation part, S433, encapsulates the compressed unencrypted and encrypted data in the format of "periodic index + data type + data + checksum" for sending to the cloud for processing or local archiving.

[0290] This invention addresses the challenge of balancing data encryption, transmission bandwidth, and signal fidelity when edge devices process multimodal physiological signals such as BCG and PPG. Specifically, this invention provides a deep learning-based feature determination method: A lightweight deep learning model is constructed to analyze the periodic physiological signal characteristics of BCG and PPG, analyzing the signal periodically and outputting precise annotations of core feature regions and period-period similarity indices. The annotation of core feature regions provides a basis for hierarchical encryption, while the period-period similarity indices provide optimization support for dynamic compression.

[0291] This invention provides a preprocessing-model fusion technique for edge adaptation: an improved temporal processing method (multimodal signal adaptation + dynamic threshold peak detection) to achieve high-precision peak labeling and period segmentation, reducing the dependence of deep learning models on preprocessing, thereby simplifying the model structure (using lightweight structures and reducing convolutional layers) and reducing the resource consumption of edge devices.

[0292] This invention provides a hierarchical encryption strategy design within a period: based on the above-mentioned feature core area annotation results, a hierarchical encryption mechanism is adopted to encrypt the feature core areas in a single-period signal using corresponding encryption algorithms (the feature core area of ​​the BCG signal is encrypted using a simplified AES-128 encryption, and the feature core area of ​​the PPG signal is encrypted using a simplified DES encryption), while non-feature redundant areas (such as the baseline interval of the signal stability) are not encrypted, thereby reducing the amount of encryption computation while ensuring core privacy security.

[0293] This invention provides a dynamic compression implementation based on periodic similarity: the edge device detects the bandwidth status in real time and dynamically adjusts the compression strategy based on the periodic similarity index output by deep learning. Under medium to high bandwidth, the core feature area is not compressed, and the non-feature redundant area is compressed according to the bandwidth. Under low bandwidth, only the complete data of the first period is retained for the core feature area with high similarity, and subsequent processing is carried out by indexing and light compression.

[0294] Example 2

[0295] An edge-side multimodal signal hierarchical encryption and compression transmission system applies the above-described method. For a detailed description of the method, please refer to the corresponding description in the above method embodiments; it will not be repeated here. Figure 21 As shown, the system 500 includes a signal preprocessing and period segmentation unit 510, a spatiotemporal feature extraction unit 520, a hierarchical encryption unit 530, and a dynamic compression unit 540; wherein,

[0296] The signal preprocessing and period segmentation unit is used to preprocess the multimodal physiological signal, adapt and enhance the preprocessed multimodal signal, obtain the peak index, and form a standardized single-period signal based on the peak index.

[0297] This embodiment uses multimodal signal adaptation enhancement and dynamic threshold peak detection to perform denoising and standardization on the original BCG and PPG signals, accurately detect the J peak of BCG and the contraction peak of PPG, and output a standardized periodic signal of 400 points / cycle.

[0298] The spatiotemporal feature extraction unit is used to extract spatiotemporal features and evaluate similarity of the standardized single-cycle signal to obtain feature core area annotation and similarity criteria.

[0299] In this embodiment, the spatiotemporal feature extraction unit includes a local spatial feature extraction module, a differentiated temporal feature modeling module, and a spatiotemporal fusion similarity determination module. The local spatial feature extraction module uses a dual-branch architecture of "CNN + differentiated attention" (BCG kernel + channel attention, PPG kernel + spatial attention) to extract 64-dimensional local features; the differentiated temporal feature modeling module uses a hybrid architecture of "Bi-LSTM + J-wave interval rule" (BCG) and "TCN + micro-temporal constraint" (PPG) to extract 128-dimensional temporal features; the spatiotemporal fusion similarity determination module weightedly fuses the spatiotemporal features and outputs the feature core area annotation and similarity criteria.

[0300] The hierarchical encryption unit is used to perform hierarchical encryption based on the annotation of the feature core area using a differentiated strategy; specifically, it uses a differentiated strategy of "BCG feature core area AES-128 simplified version (high density), PPG feature core area DES simplified version (medium density), and non-feature redundant area is not encrypted" for hierarchical encryption.

[0301] The dynamic compression unit performs layered compression on encrypted data based on real-time bandwidth detection results and periodic similarity criteria. Specifically, non-feature redundant regions employ a dynamic compression method, with the compression ratio dynamically adjusted according to bandwidth. Feature core regions are not compressed for a single period, but undergo secondary compression when bandwidth is low and the similarity criteria are met.

[0302] Example 3

[0303] A computer device 600, such as Figure 22 As shown, the system includes a memory 610, a processor 620, and a computer program 630 stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a method for hierarchical encryption and compression transmission of multimodal signals at the edge. For a detailed description of the method, please refer to the corresponding description in the above method embodiments; it will not be repeated here.

[0304] Example 4

[0305] A computer-readable storage medium, such as Figure 23As shown, a computer program is stored thereon. When executed by a processor, the computer program implements the steps of a method for hierarchical encryption and compression transmission of multimodal signals at the edge. For a detailed description of the method, please refer to the corresponding description in the above method embodiments, which will not be repeated here.

[0306] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0307] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

[0308] The apparatus, computer device, and non-volatile computer storage medium and method provided in the embodiments of this specification are corresponding. Therefore, the apparatus, computer device, and non-volatile computer storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding apparatus, computer device, and non-volatile computer storage medium will not be repeated here.

[0309] Those skilled in the art will also know that, besides implementing the controller in the form of purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller take the form of logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices included within it for implementing various functions can also be considered structures within that hardware component. Alternatively, the devices for implementing various functions can be considered as both software units implementing the method and structures within a hardware component.

[0310] The systems, apparatuses, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above apparatuses are described separately by function as various units. Of course, when implementing one or more embodiments of this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0311] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0312] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0313] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0314] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0315] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0316] This specification may be described in the general context of computer-executable instructions, such as program units, that are executed by a computer. Generally, program units include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification may also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program units may reside in local and remote computer storage media, including storage devices.

[0317] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0318] The above description is merely an embodiment of this specification and is not intended to limit the scope of one or more embodiments of this specification. Various modifications and variations can be made to one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of one or more embodiments of this specification.

Claims

1. A method for hierarchical encryption and compression transmission of multimodal signals at the edge, characterized in that, Includes the following steps: The multimodal physiological signals are preprocessed, and the preprocessed multimodal signals are adapted and enhanced to obtain peak indexes. Based on the peak indexes, standardized single-cycle signals are formed. Spatiotemporal feature extraction and similarity evaluation are performed on the standardized single-cycle signal to obtain feature core region annotation and similarity criteria; Based on the annotation of the core feature region, a differentiated strategy is adopted for hierarchical encryption; Based on real-time bandwidth detection results and periodic similarity criteria, layered compression is performed on the encrypted data; The step of hierarchical encryption based on the annotation of the feature core region and the use of a differentiated strategy includes: Input the annotation results and parse the key parameters, including the feature core region type and the reference point index; The sampling point interval of the feature core area is calculated based on the benchmark index, and the range is set in combination with the physiological characteristics differences of multimodal signals; Based on the mask output by the feature core region mapping, region segmentation and algorithm matching are performed on multimodal signals; The step of performing layered compression on encrypted data based on real-time bandwidth detection results and periodic similarity criteria includes: Obtain real-time bandwidth values ​​and, in conjunction with clinical data transmission bandwidth requirements, classify and label the real-time bandwidth. For low-bandwidth data, similarity criteria are used to accurately distinguish between redundant similar periods and abnormal periods. For non-confidential data, the compression ratio is adjusted according to the bandwidth. For encrypted data, secondary compression is performed only on encrypted data with similar periods based on the period similarity criterion when low bandwidth is detected, while encrypted single-period data is kept uncompressed to ensure accuracy.

2. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 1, characterized in that, The multimodal physiological signals include BCG signals and PPG signals.

3. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 2, characterized in that, The step of adapting and enhancing the preprocessed multimodal signal includes: By adapting and enhancing the multimodal signal, the J peak of the BCG signal and the contraction peak of the PPG signal are accurately located, and the peak index is obtained.

4. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 3, characterized in that, The step of accurately locating the J-peak of the BCG signal and the contraction peak of the PPG signal through multimodal signal adaptation and enhancement includes: For BCG signal adaptation, a 7-point differential operator is used to highlight the performance of J-peak slope acquisition. The calculation formula is as follows: in, For point i after denoising, The signal after BCG differentiation enhancement. This is to enhance the final signal of BCG; For PPG signal adaptation, a low-pass smoothing and amplitude normalization enhancement structure is adopted to highlight the shrinkage peak characteristics. The calculation formula is as follows: in, The signal is after PPG low-pass smoothing. This represents the maximum value of the signal after PPG low-pass smoothing. This is the final enhancement signal for PPG.

5. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 4, characterized in that, The step of accurately locating the J-peak of the BCG signal and the contraction peak of the PPG signal through multimodal signal adaptation and enhancement also includes: A threshold coefficient is set based on the differences in characteristics between the two signals to adapt to the peak characteristics of different signals. The calculation formula is as follows: in, As the initial threshold, This represents the maximum value of the enhanced signal. The threshold is adjusted based on historical effective peak feedback to balance stability and real-time adaptability. The calculation formula is as follows: in, The average amplitude of the first three valid peak values. This represents the original amplitude of the kth effective peak. The updated threshold; Using the preset peak interval corresponding to cardiac physiological activity as a constraint, false peaks are eliminated, and the peak intervals are interpolated to complete the peaks.

6. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 2, characterized in that, The step of forming a standardized single-cycle signal based on the peak index includes: Centered on the peak index, the truncation range is set according to the physiological waveform characteristics of the dual signals to ensure complete cycle coverage; The point completion formula is obtained by using lightweight linear interpolation: in, This represents the j-th interpolation point of the c-th single-cycle signal. This is the index of the original signal corresponding to interpolation point j.

7. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 6, characterized in that, The steps for extracting spatiotemporal features and evaluating similarity of the standardized single-cycle signal include: Local spatial feature extraction models are used to adapt local feature extraction for BCG and PPG respectively, and local spatial features are extracted from single-cycle signals. Bi-LSTM was used to capture long-term time-dependent data and the physiological rules of the J-wave interval were used to correct and extract BCG time-series features. A miniature temporal constraint module is used to suppress artifacts, and TCN is used to capture local temporal trends and extract PPG temporal features. A similarity determination mechanism based on spatiotemporal feature fusion provides similarity criteria for dynamic compression.

8. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 7, characterized in that, The local spatial feature extraction model is implemented through two convolutional blocks, channel attention, global pooling, and fully connected layers; each convolutional block includes a convolutional layer, a batch normalization (BN) layer, a ReLU layer, and a pooling layer. For BCG signals, details are captured progressively through kernel convolution and key feature channels are enhanced through channel attention, focusing on core local features; the channel attention layer is as follows: Where L is the length of the time dimension of the feature map after pooling; This is the global average pooling value for the l-th channel; Weights for the fully connected layer of the SE module; For the SE module bias vector; The normalized weights for J-wave amplitude; The feature map is after channel attention weighting; For PPG signals, kernel convolution is used to capture peak shapes and spatial attention is used to enhance peak positions, thus clustering core local features; the spatial attention layer is as follows: Where W is the locally normalized window size; The feature map is locally normalized; C is the number of channels in the feature map. This is the spatial weight coefficient vector; Spatial weight bias; The spatial weight of the j-th time position; The feature map after spatial attention weighting.

9. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 7, characterized in that, The steps for extracting BCG time-series features using Bi-LSTM to capture long-term time-dependent data and correcting for J-wave interphase physiological rules include: A two-layer bidirectional LSTM was used to capture the correlation between the J-wave interval and the period before and after the J-wave. J-wave interval calculation: The duration of the J-wave interval was obtained. like If the J-wave interval exceeds the normal heart rate, it is marked as an abnormal cycle; LSTM output for abnormal periods The average of the output from the first three normal cycles is used for correction. k is a normal periodic index; The corrected hidden state is mapped using global pooling and a fully connected layer, and the output is... ,formula: , .

10. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 7, characterized in that, The steps of using a micro-temporal constraint module to suppress artifacts and TCN to capture local temporal trends and extract PPG temporal features include: Calculate the fluctuation ratio of the maximum local characteristic values ​​of adjacent periods: Calculate the characteristic slope of adjacent periods: like and If the periodicity is 0, it is marked as an artifact period, and the mark matrix is ​​output. ; Constructing the causal convolution formula: ; For artifact periods where the marker matrix M(t)=1, the TCN output is replaced by a weighted average of the output from the previous normal period: , where (t-1) is the normal period; After collaboration, the features are global pooled and fully connected to output the final product. ,formula: ， 。 11. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 7, characterized in that, The similarity determination mechanism based on spatiotemporal feature fusion, which provides similarity criteria for dynamic compression, includes the following steps: Based on the rule-based labeling results, the normal periodic features are weighted and enhanced, while the artifact / abnormal periodic features are weighted and weakened. BCG Fusion: in, , During abnormal periods, ; PPG Fusion: in, , During the artifact period, ; Based on the signal physiological feature template library, similarity is calculated only for features that conform to the rules; the similarity calculation formula is: in, Template features; Candidate features are input into a 2-layer fully connected network, which outputs a similarity score.

12. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 1, characterized in that, For BCG signals, the characteristic core area includes J wave amplitude, J wave interval, and K wave depth, covering the J wave initiation to the K wave termination, corresponding to the key diagnostic segment of the standardized single-cycle signal; for PPG signals, the characteristic core area includes the systolic peak and diabetic wave, covering the rising edge of the systolic peak to the end of the diabetic wave, corresponding to the key hemodynamic characteristic segment. The mask based on the feature core region mapping output, the steps for performing region segmentation and algorithm matching on multimodal signals include: For BCG signals, the AES-128 simplified algorithm is used in the feature core area, while the original sampling point data is directly output in the non-feature redundant area. For PPG signals, the simplified version of the DES algorithm is used in the feature core region, while the original sampling point data is directly output in the non-feature redundant region.

13. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 12, characterized in that, It also includes simplified encryption execution steps: The sampling point data of the feature core area is grouped into 16-byte groups to obtain grouped data; Perform an XOR operation between the grouped data and the derived round key; The BCG signal is processed through 8 rounds of byte substitution, row shifting, column mixing, and round key addition, omitting the 5th and 6th rounds of column mixing operations in traditional AES. For PPG signals, perform the first 6 rounds of complete transformation and the last 2 rounds of simplified transformation, omitting the 9-16 rounds of extended permutation and subkey XOR redundancy operation in traditional DES; The results of each round of calculations are concatenated to obtain the encrypted core feature data, which is then concatenated with the original data of the non-feature redundant area to form a complete encrypted signal.

14. The edge-side multimodal signal hierarchical encryption and compression transmission method as described in claim 1, characterized in that, The steps for accurately distinguishing redundant similar periods from abnormal periods using similarity criteria for low-bandwidth data include: Obtain the single-cycle similarity score, calculate the cumulative similarity value of the current cycle and the previous 4 cycles, and make a judgment based on the reverse similarity sequence; For similar sequences, select the first period in the sequence as the reference period, and calculate the difference in encrypted data between the current period and the reference period; For dissimilar sequences, they are directly marked as independent periods, and the difference value is not calculated; The steps of adjusting the compression ratio of non-encrypted data according to bandwidth, and performing secondary compression on encrypted data with similar periods only when low bandwidth is detected, while keeping encrypted single-period data uncompressed to ensure accuracy, include: Based on the output data type of hierarchical encryption, two types of compressed objects are divided and matched with lightweight algorithms to adapt to edge computing power; Based on bandwidth level and similarity markers, dynamically adjust the compression ratio and perform compression; The compressed unencrypted and encrypted data are packaged according to periodic indexes, data types, and data and checksum formats, and sent to the cloud for processing or local archiving.

15. An edge-side multimodal signal hierarchical encryption and compression transmission system, using the method as described in any one of claims 1 to 14, characterized in that: It includes a signal preprocessing and period segmentation unit, a spatiotemporal feature extraction unit, a hierarchical encryption unit, and a dynamic compression unit; among which, The signal preprocessing and period segmentation unit is used to preprocess the multimodal physiological signal, adapt and enhance the preprocessed multimodal signal, obtain the peak index, and form a standardized single-period signal based on the peak index. The spatiotemporal feature extraction unit is used to extract spatiotemporal features and evaluate similarity of the standardized single-cycle signal to obtain feature core area annotation and similarity criteria. The hierarchical encryption unit is used to perform hierarchical encryption based on the annotation of the feature core area and using a differentiated strategy. The dynamic compression unit is used to perform layered compression on encrypted data based on real-time bandwidth detection results and periodic similarity criteria.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 14.

Images